text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
1
The Augustin Capacity and Center
Barış Nakiboğlu
[email protected]
Abstract
arXiv:1803.07937v1 [cs.IT] 21 Mar 2018
The existence of a unique Augustin mean is established for any positive order and probability mass function on the input set.
The Augustin mean is shown to be the unique fixed point of an operator defined in terms of the order and the input distribution. The
Augustin information is shown to be continuously differentiable in the order. For any channel and any convex constraint set with
finite Augustin capacity, the existence of a unique Augustin center and associated van Erven-Harremoës bound are established.
The Augustin-Legendre (A-L) information, capacity, center, and radius are introduced and the latter three is proved to be equal
to the corresponding Rényi-Gallager quantities. The equality of the A-L capacity to the A-L radius for arbitrary channels and the
existence of a unique A-L center for channels with finite A-L capacity are established. For all interior points of the feasible set
of cost constraints, the cost constrained Augustin capacity and center are expressed in terms the A-L capacity and center. Certain
shift invariant families of probability measures and certain Gaussian channels are analyzed as examples.
C ONTENTS
I
Introduction
I-A
Notational Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I-B
Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I-C
Previous Work and Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
3
4
4
II
Preliminaries
II-A
The Rényi Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
II-B
Tilted Probability Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
II-C
The Conditional Rényi Divergence and Tilted Channel . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
6
7
8
III
The Augustin Information
III-A
Existence of a Unique Augustin Mean . . . . . . . . . . . . .
III-B
Augustin Information as a Function of the Input Distribution
III-C
Augustin Information as a Function of the Order . . . . . . .
III-D
Augustin Information vs Rényi Information . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
10
11
12
The Augustin Capacity
IV-A
Existence of a Unique Augustin Center . . . . . . . . . .
IV-B
Augustin Capacity and Center as a Function of the Order
IV-C
Convex Hulls of Constraints and Product Constraints . . .
IV-D
Augustin Capacity vs Rényi Capacity . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
13
14
15
15
The Cost Constrained Problem
V-A
The Cost Constrained Augustin Capacity and Center
V-B
The Augustin-Legendre Information Measures . . . .
V-C
The Rényi-Gallager Information Measures . . . . . .
V-D
Information Measures for Transition Probabilities . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
16
16
17
19
21
VI
Examples
VI-A
Shift Invariant Families . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-B
Gaussian Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
24
25
VII
Discussion
28
IV
V
Appendix
A
B
C
D
E
References
.
.
.
.
.
.
.
.
.
.
.
.
Proof of Lemmas on the Analyticity of the Rényi Divergence
Proofs of Lemmas on the Augustin Information . . . . . . . .
Augustin’s Proof of Lemma 13-(c) . . . . . . . . . . . . . . .
Proofs of Lemmas on the Augustin Capacity . . . . . . . . .
Proofs of Lemmas on the Cost Constrained Problem . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
28
28
30
39
40
45
53
I. I NTRODUCTION
The mutual information, which is sometimes called the Shannon information, is a pivotal quantity in the analysis of various
information transmission problems. It is defined without referring to an optimization problem; but it satisfies the following two
identities given in terms of the Kullback-Leibler divergence
I(p;W) = inf q∈P(Y) D(p⊛W k p ⊗ q)
X
= inf q∈P(Y)
p(x )D(W (x )k q)
x
(1)
(2)
where P(Y) is the set of all probability measures on the output space (Y, Y), p is a probability mass function that is positive
only on a finite subset of the input set X, and W is a function of the form W : X → P(Y). Either of the expressions on
the right hand side can be taken as the definition of the mutual information. One can define the order α Rényi information
via these expressions by replacing the Kullback-Leibler divergence with the order α Rényi divergence. Since the order one
Rényi divergence is the Kullback-Leibler divergence, the order one Rényi information is equal to the mutual information for
both definitions. However, for other orders these two definitions are not equivalent to the definition of the mutual information
or to one another, as pointed out by Csiszár [1]. The generalization associated with the expression in (1) is called the order
α Rényi information and denoted by Iαg (p;W). The generalization associated with the expression in (2), on the other hand,
is called the order α Augustin information and denoted by Iα (p;W). In accordance with the convention for the constrained
Shannon capacity, the order α Augustin capacity for the constraint set A is defined as supp∈A Iα (p;W).
The Augustin information emerges naturally in the expression of the sphere packing exponent for constant composition
codes on discrete1 stationary product channels, as recently pointed out by2 Dalai [3], [4]. This observation was made implicit
by Csiszár and Körner in [6, p. 172], as well. The cost constrained Augustin capacity plays an analogous role in the expression
of the sphere packing exponent for the cost constrained codes on (possibly non-stationary) product channels, as pointed out by
Augustin in [2, §36], provided that the cost function is additive. We have extended this observation to the case of unbounded cost
functions with better prefactors, i.e. approximation error terms, in [7]. More generally, both Augustin information and capacity
are important in characterizing the optimal performance for various channel coding problems. We discuss this operational
significance more substantially in the concurrent manuscript [8].
Our main aim in the current manuscript is to analyze the Augustin information and capacity, purely as measure theoretic
concepts. Throughout the manuscript, we will refrain from referring to the channel coding problem or any operational
significance because we believe the Augustin information and capacity can and should be understood as measure theoretic
concepts first. The operational significance of the Augustin information and capacity can be established afterwards using
information theoretic techniques together with the results of the measure theoretic analysis, as we do in [8].
All of the previous works studying the Augustin information or capacity, except Augustin’s [2], assume the output set Y of
the channel W to be a finite set [1], [3], [4], [9], [10], [11], [12]. This, however, is a major drawback because the finite output
set assumption is not satisfied by many analytically interesting and practically important models, such as the Gaussian and
Poisson channels. We pursue our analysis on a more general model and assume3 the output space (Y, Y) to be a measurable
space composed of an output set Y and a σ-algebra of its subsets Y. Our analysis of the Augustin information and capacity
in this general framework is built around two fundamental concepts,
P the Augustin mean and center. P
Recall that the mutual information is defined as I(p;W) ,
x p(x )W (x ). Hence
x p(x )D(W (x )k q1,p ) where q1,p =
the infimum in (2) is achieved by q1,p . Furthermore, one can confirm by substitution that
X
p(x )D(W (x )k q) = I(p;W) + D(q1,p k q)
∀q ∈ P(Y).
x
Thus q1,p is the only probability measure achieving the infimum in (2) because the Kullback-Leibler divergence is positive
for distinct probability measures. A similar relation
P holds for other orders, as well: for any α in ℜ+ there exists a unique
probability measure qα,p satisfying Iα (p;W) = x p(x )Dα (W (x )k qα,p ). We call the probability measure qα,p , the order
α Augustin mean. In [2, Lemma 34.2], Augustin established the existence of a unique qα,p for α’s in (0, 1] and derived
certain important characteristics of qα,p that are the corner stones of the analysis of the Augustin information and capacity.
We establish analogous relations for orders greater than one in §III, see Lemma 13-(d).
Kemperman proved the equality of the (unconstrained) Shannon capacity to the Shannon radius4 and the existence of a
unique Shannon center for any channel of the form W : X → P(Y) in [13]. Using the ideas that are already present in
Kemperman’s proof one can establish a similar result for the constrained Shannon capacity provided that the constrained set
A is convex, see [14, Thm. 2]: For any channel W of the form W : X → P(Y) and convex constraint set A,
X
supp∈A I(p;W) = inf q∈P(Y) supp∈A
p(x )D(W (x )k q) .
(3)
x
1 This
observation holds for all stationary channels, not just the discrete ones, as pointed out by Augustin in [2, Remark 36.7-(i)].
[3] and [4], relying on [5], Dalai made this observation for the constant composition codes on the classical quantum channels, as well.
3 In §V-D, we have additional hypotheses; but those assumptions are satisfied by essentially all models of interest, as well.
4 Shannon radius is defined as inf
q∈P(Y) supx ∈X D(W (x )k q).
2 In
2
Furthermore, if the Shannon capacity for constraint set A is finite, then there exists a unique probability measure q1,W ,A ,
called the Shannon center for constraint set A such that
X
supp∈A I(p;W) = supp∈A
p(x )D(W (x )k q1,W ,A ) .
(4)
x
Considering (2), one can interpret (4) as a minimax theorem. The name center is reminiscent of the name of the corresponding
quantity in the unconstrained case, discussed in [13]. Augustin proved an analogous result for Iα (p;W) assuming α to be
an order in (0, 1] and A to be a constraint set determined by cost constraints, see [2, Lemma 34.7]. We prove an analogous
proposition for Iα (p;W) for any α in ℜ+ and convex constraint set A in §IV, see Theorem 1. We call the corresponding
probability measure qα,W ,A the order α Augustin center for constraint set A.
The constraint sets determined by the cost constraints are frequently encountered while employing the Augustin capacity
in the analysis of channel coding problems. One can apply the convex conjugation techniques to provide an alternative
characterization of the cost constrained Augustin capacity and center. Augustin did so in [2, §35] using a concept that was
previously employed by Gallager [15, pp. 13-15], [16, §7.3] in discrete channels and by Gallager [15, pp. 15,16], [16, §§7.4,7.5]
and Ebert [17] in Gaussian channels. We call this concept the Rényi-Gallager information and analyze it in V-C. In [2, §35],
Augustin assumed neither a specific noise model nor the finiteness of the output set. Nonetheless, Gaussian channels are not
subsumed by Augustin’s model because Augustin assumed a bounded cost function in [2, §35]. When compared with the
application of convex conjugation techniques for the cost constrained Shannon capacity provided by Csiszár and Körner in
[6, Ch. 8], Augustin’s analysis in [2, §35] based on the Rényi-Gallager information seems rather convoluted. We provide a
more standard approach which can be seen as a generalization of [6, Ch. 8] in V-B via a new concept, which we call the
Augustin-Legendre information. We show the equivalence of these two approaches using minimax theorems similar to the one
described above for the constrained Augustin capacity.
Some of the most important observations we present in this manuscript have already been derived previously in [2, §§33-35],
[18], [19], [11]. We refer to the original works establishing these observations when we state them with in the body of the
paper. In order delineate our main contributions in the context of these works, we provide a tally in §I-C. Before doing that,
we describe our notional conventions in I-A and our model in I-B.
A. Notational Conventions
P
For any two vectors µ and q in ℜℓ their inner product, denoted by µ · q, is ℓı=1 µı q ı . For any ℓ ∈ Z+ , ℓ dimensional vector
whose all entries are one is denoted by 1, the dimension ℓ will be clear from the context. We denote the closure, interior, and
convex hull of a set S by clS, intS, and chS, respectively; the relevant topology or vector space structure will be evident
from the context.
For any set Y, we denote the set of all subsets of Y, i.e. the power set of Y, by 2 Y , the set of all probability measures on
+
finite subsets of Y by P(Y), and the set of all non-zero finite measures with the same property by M (X). We call the set of
all y’s in Y for which p(y) > 0 the support of p.
On a measurable space (Y, Y), we denote the set of all finite signed measures by M(Y), the set of all finite measures by
+
+
M0 (Y), the set of all non-zero finite measures by M (Y), and the set of all probability measures by P(Y). Let µ and q be
two measures on the measurable space (Y, Y). Then µ is absolutely continuous with respect to q, i.e. µ≺q, iff µ(E) = 0 for
any E ∈ Y such that q(E) = 0; µ and q are equivalent, i.e. µ ∼ q iff µ≺q and q≺µ; µ and q are singular, i.e. µ ⊥ q, iff
∃E ∈ Y such that µ(E) = q(Y \ E) = 0. Furthermore, a set of measures W on (Y, Y) is absolutely continuous with respect to
q, i.e. W≺q, iff w ≺q for all w ∈ W and uniformly absolutely continuous with respect to q, i.e. W≺uni q, iff for every ǫ > 0
there exists a δ > 0 such that w (E) < ǫ for all w ∈ W provided that q(E) < δ.
R
R
We denote the integral of a measurable function f with respect to the measure µ by fRµ(dy ) orR f (y)µ(dy ). If the integral
is on the real line and if it is with respect to the Lebesgue measure, we denote it by f dy or f (y)dy , as well. If µ is a
probability measure, then we also call the integral of f with respect µ the expectation of f or the expected value of f and
denote it by Eµ [f ] or Eµ [f (Y)].
Our notation will be overloaded for certain symbols; however, the relations represented by these symbols
will be clear
P
from the context. We use ℏ(·) to denote both the Shannon entropy and the binary entropy: ℏ(p) , − y p(y) ln p(y) for
all p ∈ P(Y) and ℏ(z ) , − z ln z − (1 − z ) ln(1 − z ) for all z ∈ [0, 1]. We denote the product of topologies [20, p. 38],
σ-algebras [20, p. 118], and measures [20, Thm. 4.4.4] by ⊗. We denote the Cartesian product of sets [20, p. 38] by ×. We
use |·| to denote the absolute value of real numbers and the size of sets. The sign ≤ stands for the usual less than or equal to
relation for real numbers and the corresponding point-wise inequity for functions and vectors. For two measures µ and q on
the measurable space (Y, Y), µ ≤ q iff µ(E) ≤ q(E) for all E ∈ Y.
For a, b ∈ ℜ, a ∧ b is the minimum of a and b. For f : Y → ℜ and g : Y → ℜ, the function f ∧ g is the pointwise
dµ
dq
minimum of f and g. For µ, q ∈ M(Y), µ ∧ q is the unique measure satisfying dµ∧q
dν = dν ∧ dν ν-a.e. for any ν satisfying
µ≺ν and q≺ν. For a collection F of real valued functions ∧f ∈F f is the pointwise infimum of f ’s in F, which is an extended
real valued function. For a collection of measures U ⊂ M(Y) satisfying w ≤ u for all u ∈ U for some w ∈ P(Y), ∧u∈U u
is the infimum of U with respect to the partial order ≤. There exists a unique infimum measure by [21, Thm. 4.7.5]. We use
the symbol ∨ analogously to ∧ but we represent maxima and suprema with it, rather than minima and infima.
3
B. Channel Model
A channel W is a function from the input set X to the set of all probability measures on the output space (Y, Y):
W : X → P(Y)
(5)
Y is called the output set and Y is called the σ-algebra of the output events. We denote the set of all channels from the
input set X to the output space (Y, Y) by P(Y|X). For any p ∈ P(X) and W ∈ P(Y|X), p⊛W is the probability measure
whose marginal on X is p and whose conditional distribution given x is W (x ). Until §V-D, we confine our discussion to the
input distributions in P(X) and avoid the subtleties related to measurability and integrability. The more general case of input
distributions in P(X ) are considered5 in §V-D.
A channel W is called a discrete channel if both X and Y are finite sets. A channel W : X → P(Y) is a product channel
for a finite index set T, iff there exist channels Wt : Xt → P(Yt ) for t ∈ T satisfying
Y⊗
Wt (xt )
∀x ∈ X
(6)
W (x ) =
t∈T
Q×
Q×
Q⊗
where X = t∈T Xt , Y = t∈T Yt , and Y = t∈T Yt . A product channel is stationary iff all Wt ’s are identical. If X ⊂
Q⊗
t∈T Xt , then we call W a memoryless channel.
For any ℓ ∈ Z+ , an ℓ dimensional cost function ρ is a function from the input set to ℜℓ that is bounded from below:
ρ : X → ℜℓ≥z
z ∈ ℜ.
6
We assume without loss of generality that
inf x ∈X ρı (x ) = 0
∀ı ∈ {1, . . . , ℓ}.
We denote the set of all cost constraints that can be satisfied by some member of X by Γρex and the set of all cost constraints
that can be satisfied by some member of P(X) by Γρ :
Γρex , {̺ ∈ ℜℓ≥0 : ∃x ∈ X s.t. ρ(x ) ≤ ̺}
X
p(x )ρ(x ) ≤ ̺}.
Γρ , {̺ ∈ ℜℓ≥0 : ∃p ∈ P(X) s.t.
(7)
(8)
x
Then both Γρex and Γρ have non-empty interiors and Γρ is the convex hull of Γρex , i.e. Γρ = chΓρex .
A cost function ρ for a product channel W is said to be additive iff there exists a ρt : Xt → ℜℓ≥0 for each t ∈ T such that
X
ρ(x ) =
ρt (xt )
∀x ∈ X.
(9)
t∈T
C. Previous Work and Main Contributions
The following list includes our main contributions that are critical for a thorough understanding of the Augustin information
measures and related previous results. For our other contributions related previous results are mentioned after their statement
within the body of the paper.
(I) For α ∈ (0, 1) Augustin, [2, Lemma 34.2], claims to establish not only the existence of a unique probability measure
qα,p satisfying Iα (p;W) = Dα (W k qα,p | p) but also the following characteristics of qα,p in terms of the operator Tα,p (·)
which is determined by α and p and well-defined for all q with finite Dα (W k q| p), see (33) for the exact definition:
• Tα,p (qα,p ) = qα,p and qα,p ∼ q1,p .
• If q1,p ≺q and Tα,p (q) = q, then qα,p = q.
• lim→∞ qα,p − Tα,p (q1,p ) = 0.
7
• Dα (W k q| p) ≥ Iα (p;W) + Dα (qα,p k q) for all q ∈ P(Y).
We can not verify the correctness of Augustin’s proof of [2, Lemma 34.2]; we discuss our reservations for Augustin’s
proof in Appendix C. Lemma 13-(c) is proved8 relying on the ideas employed in Augustin’s proof of [2, Lemma 34.2].
Lemma 13-(c) implies all assertions of [2, Lemma 34.2] except for lim→∞ qα,p − Tα,p (q1,p ) = 0, Lemma 13-(c)
g
establishes lim→∞ qα,p − Tα,p (qα,p
) = 0 instead. Unlike [2, Lemma 34.2], Lemma 13-(c) also bounds Dα (W k q| p)
from above. This bound is new to the best of our knowledge. The following inequality summarizes the upper and lower
bounds on Dα (W k q| p) established in Lemma 13-(c,d):
D1∨α (qα,p k q) ≥ Dα (W k q| p) − Iα (p;W) ≥ D1∧α (qα,p k q)
∀q ∈ P(Y).
(10)
5 The structure described in (5) is not sufficient on its own to ensure the existence of a unique p⊛W with the desired properties for all p ∈ P(X ), in
general. The existence of a unique p⊛W is guaranteed for all p ∈ P(X ), if W is a transition probability from (X, X ) to (Y, Y), i.e. a member of P(Y|X )
rather than P(Y|X).
6 Augustin [2, §33] has an additional hypothesis: ∨
x ∈X ρ(x ) ≤ 1. This hypothesis, however, excludes certain important cases.
2
7 To be precise [2, Lemma 34.2] asserts the inequality D (W k q| p) ≥ I (p;W) + α kq
α
α
α,p − qk rather than the one given above. But Augustin proves
2
the inequality given above first and then uses Pinsker’s inequality to establish the one given in [2, Lemma 34.2].
8 One can prove Lemma 13-(d) using the ideas employed in the proof of Lemma 13-(c), as well.
4
For finite Y case the existence of a q in P(Y) satisfying both q ∼ q1,p and Tα,p (q) = q have been discussed by other
authors. We make a brief digression to point out the discussion of the aforementioned existence result in these works.
• While deriving the sphere packing bound for the constant composition codes on discrete stationary product channels,
Fano implicitly asserts for each α in (0, 1) the existence of a fixed point that is equivalent to q1,p , see [22, §9.2,
(9.24) & p. 292]. Fano, however, does not explain why such a fixed point must exist and does not elaborate on its
uniqueness or relation to qα,p in [22, §9.2].
• While establishing the equivalence of his expression for the sphere packing exponent in finite Y case to the one
provided by Fano in [22], Haroutunian proved for each α in (0, 1) the existence of a fixed point that is equivalent
to q1,p , [18, (16)-(19)].
• While discussing the random coding bounds for discrete stationary product channels, Poltyrev makes an observation
that is equivalent to asserting for each α in [1/2, 1) the existence of a fixed point that is equivalent to q1,p , see [19,
(3.15), (3.16) and Thm. 3.2]. Poltyrev, however, does not formulate his observations as a fixed point property.
In our understanding, the main conceptual contribution of [2, Lemma 34.2] is the characterization of the Augustin mean
as a fixed point that is equivalent to q1,p . Bounds such as the one given in (10) follow from this observation via Jensen’s
inequality.
(II) For α ∈ (1, ∞), Lemma 13-(d) establishes the existence of a unique Augustin mean qα,p and proves that it satisfies (10)
as well as the following two assertions:
• Tα,p (qα,p ) = qα,p and qα,p ∼ q1,p .
• If Tα,p (q) = q, then qα,p = q.
(III) Iα (p;W) is a continuously differentiable function of α from ℜ+ to [0, ℏ(p)] by Lemma 17-(e).
(IV) The following minimax identity is established in Theorem 1 for any convex constraint set A
supp∈A inf q∈P(Y) Dα (W k q| p) = inf q∈P(Y) supp∈A Dα (W k q| p) .
Theorem 1 establishes the existence of a unique Augustin center, qα,W ,A , for any convex A with finite Augustin
capacity and the convergence of {qα,p (ı) }ı∈Z+ to qα,W ,A in total variation topology for any {p (ı) }ı∈Z+ ⊂ A satisfying
limı→∞ Iα p (ı) ;W = Cα,W ,A . Augustin proved this result only for α’s in (0, 1] and the constraint sets determined by
cost constraints, see [2, Lemma 34.7]. For A = P(X) case similar results were proved by Csiszár [1, Proposition 1]
assuming both X and Y are finite sets and by van Erven and Harremoës [9, Thm. 34] assuming Y is a finite set.
(V) The following bound in terms of the Augustin capacity and center established in Lemma 21 is new to the best of our
knowledge
supp∈A Dα (W k q| p) ≥ Cα,W,A + Dα∧1 (qα,W,A k q)
∀q ∈ P(Y).
A similar bound has been conjectured by van Erven and Harremoës in [9]. For the Rényi capacity and center, we have
proved that conjecture and extended it to the constrained case in [14, Lemmas 19 & 25].
(VI) The Augustin-Legendre information Iαλ (p;W), defined as Iα (p;W) − λ · Ep [ρ], as well as the resulting capacity, center,
and radius are new concepts that have not been studied before, except for α = 1 case. Thus, formally speaking, all of
the propositions in §V-B are new. The analysis presented in §V-B is a standard application of the convex conjugation
techniques to characterize the cost constrained Augustin capacity and center. A similar analysis for α = 1 case is
provided by Csiszár and Körner in [6, Ch. 8] under discrete W and single cost constraint hypotheses. The most important
conclusions of the analysis presented in §V-B are the followings:
λ
λ
λ
ℓ
• Cα,W , defined as supp∈P(X) Iα (p;W), satisfies Cα,W = sup̺≥0 Cα,W ,̺ − λ · ̺ for all λ ∈ ℜ≥0 by (90).
λ
• Cα,W,̺ = inf λ≥0 Cα,W + λ · ̺ for all ̺ ∈ intΓρ and the infimum is achieved by all λ in some a non-empty convex
compact set whenever Cα,W,̺ is finite by Lemma 29.
λ
λ
λ
• Cα,W = Sα,W where Sα,W is defined as inf q∈P(Y) supx X Dα (W (x )k q) − λ · ρ(x ) by Theorem 2.
λ
λ
λ
λ
• If Cα,W < ∞, then there exists a unique A-L center qα,W satisfying Cα,W = supx ∈X Dα W (x )k qα,W − λ · ρ(x ).
λ
λ
by
= 0 for all {p (ı) }ı∈Z+ ⊂ P(X) satisfying limı→∞ Iαλ p (ı) ;W = Cα,W
Furthermore, limı→∞ qα,p − qα,W
Theorem 2.
ℓ
λ
λ
• If Cα,W,̺ = Cα,W + λ · ̺ < ∞ for a λ ∈ ℜ≥0 , then qα,W,̺ = qα,W by Lemma 31.
P
λ
ℓ
λ
• If W is a product channel with an additive cost function, then Cα,W =
t∈T Cα,Wt for all λ ∈ ℜ≥0 , α ∈ ℜ+ and
Q
⊗
λ
λ
whenever it exists qα,W
by Lemma 32.
is equal to t∈T qα,W
t
gλ
(VII) The Rényi -Gallager information Iα (p;W) is, in a sense, a generalization of the Rényi information Iαg (p;W) with a
Lagrange multiplier because Iαg0 (p;W) = Iαg (p;W). This quantity was first employed by Gallager in [15] by a different
parametrization and scaling; later considered by Arimoto [23, §IV], Augustin [2], Ebert [17], Oohama [24], [25], and
Vazquez-Vilar, Martinez, and Fàbregas [26] with various parametrizations, scalings, and names. We chose the scaling
and parametrization that is consistent with the ones for Augustin-Legendre information. The most important conclusions
of our analysis in V-C are the followings.
5
gλ
gλ
λ
Cα,W
= Sα,W
by Theorem 3, where Cα,W
is defined as supp∈P(X) Iαgλ (p;W).
(ı)
λ
gλ
λ
λ
g
• If Cα,W < ∞ and limı→∞ Iα p ;W = Cα,W , then limı→∞ qα,p − qα,W = 0 by Theorem 3.
λ
λ
• supx ∈X Dα (W (x )k q) − λ · ρ(x ) ≥ Cα,W + Dα qα,W q for all q ∈ P(Y) by Lemma 35.
For the case when both α ∈ (0, 1) and ∨x ∈X ρ(x ) ≤ 1, Theorem 3 is implied by [2, Lemma 35.2] but Lemma 35 is
new to the best of our knowledge.
While pursuing a similar analysis in [2, §35], Augustin assumed the cost function to be bounded. This hypothesis,
however, excludes certain important and interesting cases such as the Gaussian channels. The issue here is not a matter
of rescaling: certain conclusions of Augustin’s analysis are not correct for the case with unbounded cost functions. We
do not assume the cost function to be bounded. Thus our model subsumes not only Augustin’s in [2, §35] but also other
previously considered models, which were either discrete [15, pp. 13-15], [16, §7.3], [23, §IV], [25], [26] or Gaussian
[15, pp. 15,16], [16, §§7.4,7.5],[17].
(VIII) Especially for channels with uncountable input sets the Shannon information and capacity is often defined via the
probability measures on the input space (X, X ), rather than the probability mass functions on the input set X. In §V-D,
we discuss how and under which conditions one can make such a generalization for Augustin information measures.
The most important conclusions of our analysis are the followings:
• If W is a transition probability, W ∈ P(Y|X ) and Y is countably generated, then
– Iα (p;W) is well defined for all α ∈ ℜ+ and p ∈ P(X ) by (128), (129), and Lemma 37
– Iαλ (p;W) is well defined for all α ∈ ℜ+ , p ∈ P(X ), and λ ∈ ℜℓ≥0 by (130) provided that ρ is X -measurable.
• If W ∈ P(Y|X ), X is countably separated, Y is countably generated, and ρ is X -measurable, then
λ
– Cα,W
= supp∈A∗ Iαλ (p;W) for all λ in ℜℓ≥0 by Theorem 4 where A∗ is defined as {p ∈ P(X ) : Ep [ρ] ∈ ℜℓ≥0 }.
λ
λ
λ
p − λ · Ep [ρ] by Theorem 4.
– If Cα,W
< ∞ for a λ in ℜℓ≥0 , then Cα,W
= supp∈A∗ Dα W k qα,W
– Cα,W ,̺ = supp∈A(̺) Iα (p;W) for all ̺ in intΓρ by Theorem 5. where A(̺) is defined as {p ∈ P(X ) : Ep [ρ] ≤ ̺}.
– If Cα,W ,̺ < ∞ for a ̺ in intΓρ , then Cα,W ,̺ = supp∈A(̺) Dα (W k qα,W ,̺ | p) by Theorem 5.
Thus the A-L capacity and center as well as the cost constrained Augustin capacity and center defined via probability
mass functions are equal to the corresponding quantities that might be defined via probability measures on (X, X ),
provided that X is countably separated and Y is countably generated.
•
II. P RELIMINARIES
A. The Rényi Divergence
+
Definition 1. For any α ∈ ℜ+ and w , q ∈ M (Y); then the order α Rényi divergence between w and q is
( 1
R dw α dq 1−α
( dν ) ( dν ) i ν(dy) α 6= 1
α−1 ln
h
Dα (w k q) , R dw
dq
dw
α=1
dν ln dν − ln dν ν(dy)
(11)
where ν is any measure satisfying w ≺ν and q≺ν.
Customarily, the Rényi divergence is defined for pairs of probability measures —see [9] and [27] for example— rather than
pairs of nonzero finite measures. We adopt this slightly more general definition because it allows us to use the Rényi divergence
to express certain observations more succinctly, see Lemma 1 in the following and §V-C. However, for probability measures
Definition 1 is equivalent to usual definition employed in [9] by [9, Thm. 5].
Lemma 1 ([14, Lemma 8]). Let α be a positive real number and w , q, v be non-zero finite measures on (Y, Y).
• If v ≤ q, then Dα (w k q) ≤ Dα (w k v ).
• If q = γv for a γ ∈ ℜ+ , then Dα (w k q) = Dα (w k v ) − ln γ.
If both arguments of the Rényi divergence are probability measures, then it is positive, unless the arguments are equal to
one another by Lemma 2. For orders in (0, 1] this inequality is called the Pinsker’s inequality, [28], [29].
Lemma 2 ([9, Thm. 3, Thm. 31]). For any α ∈ ℜ+ , probability measure w and q on (Y, Y)
Dα (w k q) ≥
1∧α
2 kw
2
− qk .
(12)
As a function of its arguments, the order α Rényi divergence is continuous for the total variation topology provided that
α ∈ (0, 1). For arbitrary orders we only have lower semicontinuity, but that holds even for the topology of setwise convergence.
Lemma 3 ([9, Thm 15]). For any α ∈ ℜ+ , Dα (w k q) is a lower semicontinuous function of the pair of probability measures
(w , q) in the topology of setwise convergence.
Lemma 4 ([9, Thm 17]). For any α ∈ (0, 1), Dα (w k q) is a uniformly continuous function of the pair of probability measures
(w , q) in the total variation topology.
6
The Rényi divergence is convex in its second argument for all positive orders, jointly convex in its arguments for positive
orders that are not greater than one, and jointly quasi-convex in its arguments for all positive orders.
Lemma 5 ( [9, Thm. 12]). For all α ∈ ℜ+ , w , q0 , q1 ∈ P(Y), β ∈ (0, 1), and ν satisfying (q0 + q1 )≺ν,
Dα (w k βq1 + (1 − β)q0 ) ≤ βDα (w k q1 ) + (1 − β)Dα (w k q0 ) .
Furthermore, the equality holds iff
dq1
dν
=
dq0
dν
(13)
w -almost surely.
Lemma 6 ([9, Thm. 11]). For all α ∈ (0, 1], w0 , w1 , q0 , q1 ∈ P(Y), β ∈ (0, 1), and ν satisfying (w0 + w1 + q0 + q1 )≺ν,
Dα (βw1 + (1 − β)w0 k βq1 + (1 − β)q0 ) ≤ βDα (w1 k q1 ) + (1 − β)Dα (w0 k q0 ) .
Furthermore, for α = 1 the equality holds iff
and Dα (w1 k q1 ) = Dα (w0 k q0 ).
dw0 dq1
dν dν
=
dw1 dq0
dν dν
and for α ∈ (0, 1) the equality holds iff
(14)
dw0 dq1
dν dν
=
dw1 dq0
dν dν
Lemma 7 ([9, Thm. 13]). For all α ∈ ℜ+ , w0 , w1 , q0 , q1 ∈ P(Y), and β ∈ (0, 1)
Dα (βw1 + (1 − β)w0 k βq1 + (1 − β)q0 ) ≤ Dα (w1 k q1 ) ∨ Dα (w0 k q0 ) .
(15)
Lemma 8 ([9, Thm. 3, Thm. 7]). For all w , q ∈ P(Y), Dα (w k q) is an nondecreasing and lower semicontinuous function of
α on ℜ+ that is continuous on (0, (1 ∨ χw,q )] where χw,q , sup{α : Dα (w k q) < ∞}.
If G is a sub-σ-algebra of Y, for any w and q in P(Y) the identities w|G (E) = w (E) for all E ∈ G and q|G (E) = q(E) for all
E ∈ G uniquely define probability measures w|G and q|G on (Y, G). We denote Dα w|G q|G by DαG (w k q).
Lemma 9 ([9, Thm. 21]). Let Y1 ⊂ Y2 ⊂ · · · ⊂ Y be an increasing family of σ−algebras, and let Y∞ = σ(∪∞
ı=1 Yı ) be the
smallest σ-algebra containing them. Then for any order α ∈ ℜ+
limı→∞ DαYı (w k q) = DαY∞ (w k q) .
B. Tilted Probability Measure
Definition 2. For any α ∈ ℜ+ and w , q ∈ P(Y) satisfying Dα (w k q) < ∞, the order α tilted probability measure wαq is
q
dwα
dν
α dq 1−α
, e(1−α)Dα(wkq) ( dw
.
dν ) ( dν )
(16)
w1q
Note that
= w for any q satisfying D1 (w k q) < ∞. For other orders one can confirm the following identity simply by
substitution: if Dα (w k q) < ∞, then for any v ∈ P(Y) satisfying D1 (v k w ) < ∞ and D1 (v k q) < ∞ also satisfies
q
1
1−α D1 (v k wα )
+ Dα (w k q) =
α
1−α D1 (v k w )
+ D1 (v k q) .
This identity is used to derive the following variational characterization of the Rényi divergence for orders other than one.
Lemma 10 ([9, Thm. 30]). For any w , q ∈ P(Y)
(
α
inf v ∈P(Y) 1−α
D1 (v k w ) + D1 (v k q)
Dα (w k q) =
α
supv ∈P(Y) 1−α D1 (v k w ) + D1 (v k q)
α ∈ (0, 1)
α ∈ (1, ∞)
(17)
α
where 1−α
D1 (v k w ) + D1 (v k q) stands for −∞ when α ∈ (1, ∞) and D1 (v k w ) = D1 (v k q) = ∞. Furthermore, if Dα (w k q)
is finite and either α ∈ (0, 1) or D1 (wαq k w ) < ∞, then
Dα (w k q) =
q
α
1−α D1 (wα k w )
+ D1 (wαq k q) .
(18)
We have observed in Lemma 8 that Dα (w k q) is continuous in α on the closure of the interval that it is finite. Lemma 11,
in the following, establishes not only the continuity but also the analyticity of Dα (w k q) on the interior of the interval that
Dα (w k q) is finite. Lemma 11 also establishes the analyticity, and hence finiteness, of D1 (wαq k w ) and D1 (wαq k q) on the same
interval. This allows us to assert the validity of (18) on the same interval as a result of Lemma 11:
Dα (w k q) =
q
α
1−α D1 (wα k w )
+ D1 (wαq k q)
∀α ∈ (0, χw,q ).
(19)
Lemma 11. For any w , q ∈ P(Y) satisfying χw,q > 0, for χw,q , sup{α : Dα (w k q) < ∞}, Dα (w k q), D1 (wαq k w ), and
D1 (wαq k q) are analytic functions of α on (0, χw,q ). Furthermore,
P
h
ı iı
t
κ
(−1)
(−1)κ−t +1 P (1 +2 +···+t −1)! Q
dq
dw
q
ln
E
−
ln
φ 6= 1
κ!
κ−t +1
w
(φ−1)
!
!...
!
ı!
dν
dν
1 2
t
φ
ı=1
t=0
Jt
∂ κ Dα(wkq)
(20)
=
h
i
κ
ı
ı
∂α
P −(1 +2 +···+κ+1 −1)! κ+1
Q (−1)
α=φ
dq
dw
ln
φ
=
1
E
−
ln
κ!
w
1 !2 !...κ+1 !
ı!
dν
dν
Jκ+1
ı=1
7
where Jt , {(1 , 2 , . . . , t ) : ı ∈ Z≥0 ∀ı and 11 + 22 + . . . + t t = t }.
Lemma 11 is new to the best of our knowledge; it is proved in Appendix A using standard results on the continuity and
differentiability of parametric integrals and Faà di Bruno formula for derivatives of compositions of smooth functions.
The expression given in (20), for the derivatives of Dα (w k q) with respect to α implies the the following more succinct
forms for the first two derivatives.
1
α 6= 1
(α−1)2 D1 (wαq k w )
∂
2
,
(21)
D
(w
k
q)
=
α
∂α
21 Ew ln dw
α=1
dq − D1 (w k q)
2
i2
h
q
q
q
dwα
dwα
dwα
1
q
q
− 2 ln dw − Ewα ln dw
α 6= 1
(α−1)3 Ewα ln dw
∂2
D
(w
k
q)
=
.
(22)
2
α
3
∂α
P p(x )
α=1
x 3 Ew ln dw
dq − D1 (w k q)
Analyticity of Dα (w k q) on (0, χw,q ) implies that for any φ ∈ (0, χw,q ) there exits an open interval containing φ on which
Dα (w k q) is equal to the power series determined by the derivatives of Dα (w k q) at α = φ. If we have a finite collection
of pairs of probability measures {(wı , qı )}ı∈I , then for any φ that is in (0, χwı ,qı ) for all ı ∈ I there exists an open interval
containing φ on which each Dα (wı k qı ) is equal to the power series determined by the derivatives of Dα (wı k qı ) at α = φ.
When the collection of pairs of probability measures is infinite, then there might not be an open interval containing φ that
is contained in all (0, χwı ,qı )’s. Lemma 12, in the following, implies the existence of such an interval when Dβ (wı k qı ) is
uniformly bounded for a β > φ for all ı ∈ I. In addition, Lemma 12 implies uniform approximation error terms, over all ı ∈ I,
for the power series on that interval.
Lemma 12. For any γ, φ, β ∈ ℜ+ satisfying φ ∈ (0, β) and w , q ∈ P(Y) satisfying Dβ (w k q) ≤ γ,
(
(κ + 1)!τ κ+1 φ 6= 1
∂ κ Dα(wkq)
≤
∂ακ
α=φ
κ!τ κ+1
φ=1
κ+1
h
i
κ
1
τ |η−φ| κ +
Xκ−1 (η−φ)ı ∂ ı D (wkq)
φ=
6 1
1−|η−φ|τ
α
≤ τ1−|η−φ|τ
Dη (w k q) −
∀η : |η − φ| ≤
ı
κ+1
κ
ı!
∂α
|η−φ|
ı=0
α=φ
φ=1
1−|η−φ|τ
where
τ ,
(
1+e(1∨β)γ
1
|φ−1| ∨ φ∧(β−φ)
1+eβγ
1∧(β−1)
φ 6= 1
φ=1
.
(23)
1
τ
(24)
(25)
Lemma 12 is new to the best of our knowledge; it is proved in Appendix A using (20) together with the elementary properties
of the real analytic functions and power series.
C. The Conditional Rényi Divergence and Tilted Channel
The Augustin information is defined in terms of the Rényi divergence and its analysis rely on the concept of tilted
probability measure. In order to express the frequently used expressions more succinctly we define the concepts of conditional
Rényi divergence and the tilted channel.
Definition 3. For any α ∈ ℜ+ , W : X → P(Y), Q : X → P(Y), and p ∈ P(X) the order α conditional Rényi divergence for
input distribution p is
X
Dα (W k Q | p) ,
p(x )Dα (W (x )k Q (x )) .
(26)
x ∈X
If ∃q ∈ P(Y) such that Q (x ) = q for all x ∈ X, then we denote Dα (W k Q | p) by Dα (W k q| p).
Remark 1. In [30] and [12], Dα (W k Q | p) stands for Dα (p⊛W k p⊛Q). For α = 1 convention of [30] and [12] is equivalent
to ours; for α 6= 1, however,Pit is not. If either α = 1 or Dα (W (x )k Q (x )) has theP
same value for all x ’s with positive p(x ),
then Dα (p⊛W k p⊛Q) =P x p(x )Dα (W (x )k Q (x )), else Dα (p⊛W k p⊛Q) < x p(x )Dα (W (x )k Q (x )) for α ∈ (0, 1)
and Dα (p⊛W k p⊛Q) > x p(x )Dα (W (x )k Q (x )) for α ∈ (1, ∞). The inequalities follow from the Jensen’s inequality and
the concavity of the natural logarithm function.
Definition 4. For any α ∈ ℜ+ , W : X → P(Y) and Q : X → P(Y), the order α tilted channel WαQ is a function from
{x : Dα (W (x )k Q (x )) < ∞} to P(Y) given by
dWαQ (x )
dν
) 1−α
, e(1−α)Dα(W (x )kQ(x )) ( dWdν(x ) )α ( dQ(x
.
dν )
If ∃q ∈ P(Y) such that Q (x ) = q for all x ∈ X, then we denote
8
WαQ
by
Wαq .
(27)
III. T HE AUGUSTIN I NFORMATION
The main aim of this section is to introduce the concepts of Augustin information and mean. We define the order α Augustin
information for input distribution p and establish the existence of a unique Augustin mean for any input distribution p and
positive finite order α in §III-A. After that we analyze the Augustin information, first as a function of the input distribution for
a given order in §III-B and then as a function of the order for a given input distribution in §III-C. We conclude our discussion
by comparing the Augustin information with the Rényi information and characterizing each quantity in terms of the other in
§III-D. Some of the most important observations about the Augustin information and mean were first reported by Augustin in
[2, §34] for orders not exceeding one. This is why we suggest naming these concepts after him. Proof of the lemmas presented
in this section are provided in Appendix B.
A. Existence of a Unique Augustin Mean
Definition 5. For any α ∈ ℜ+ , W : X → P(Y), and p ∈ P(X) the order α Augustin information for the input distribution p is
Iα (p;W) , inf q∈P(Y) Dα (W k q| p) .
(28)
One can confirm by substitution that
D1 (W k q| p) = D1 (W k q1,p | p) + Dα (q1,p k q)
where
q1,p ,
Then Lemma 2 and (28) imply
X
x
∀q ∈ P(Y)
p(x )W (x ).
(29)
(30)
I1 (p;W) = D1 (W k q1,p | p) .
Thus the order one Augustin information has a closed form expression, which is equal to the mutual information. For other
orders, however, Augustin information does not have a closed form expression. Nonetheless, Lemma 13, presented in the
following, establishes the existence of a unique probability measure qα,p satisfying Iα (p;W) = Dα (W k qα,p | p) for9 any
positive order α and input distribution p. Furthermore, parts (c) and (d) of Lemma 13 present an alternative characterization
of qα,p by showing that qα,p is the unique fixed point of the operator Tα,p (·) satisfying q1,p ≺qα,p . Lemma 13-(e) provides an
alternative characterization of the Augustin information for orders other than one.10
+
Definition 6. For any α ∈ ℜ+ , W : X → P(Y), and p ∈ M (X)
• The order α mean measure for input distribution p is given by
α i α1
hX
dµα,p
dW (x )
,
p(x
)
dν
dν
x
P
where ν is any measure for which ( x p(x )W (x ))≺ν.
• The order α Rényi mean for input distribution p is given by
g
qα,p
,
•
(31)
µα,p
kµα,p k .
(32)
The order α Augustin operator for input distribution p, i.e. Tα,p (·) : Qα,p → P(Y), is given by
X
Tα,p (q) ,
p(x )Wαq (x )
∀q ∈ Qα,p
x
(33)
+
where Qα,p , {q ∈ M (Y) : Dα (W k q| p) < ∞} and the tilted channel Wαq is defined in (27). Furthermore, T1α,p (q) =
ı
Tα,p (q) and Tı+1
α,p (q) , Tα,p (Tα,p (q)) for any positive integer ı.
Lemma 13. Let W be a channel of the form W : X → P(Y) and p be an input distribution in P(X).
(a) Iα (p;W) ≤ Dα (W k q1,p | p) ≤ ℏ(p) < ∞ for all α ∈ ℜ+ where q1,p is defined in (30).
(b) I1 (p;W) = D1 (W k q1,p | p). Furthermore,
D1 (W k q| p) − I1 (p;W) = D1 (q1,p k q)
∀q ∈ P(Y).
(34)
9 This is rather easy to prove when Y is a finite set. Uniqueness of q
α,p follows from the strict convexity of the Rényi divergence in its second argument
described in Lemma 5. For finite Y’s P(Y) is compact and the existence of qα,p follows from the lower semicontinuity of the Rényi divergence in its second
argument, which follows from Lemma 3, and the extreme value theorem for the lower semicontinuous functions [31, Ch3§12.2]. For channels with arbitrary
output spaces, however, P(Y) is not compact; thus we can not invoke the extreme value theorem to establish the existence of qα,p .
10 This alternative characterization is employed in proving the equivalence of two definitions of the sphere packing exponent, see [8].
9
(c) If α ∈ (0, 1), then ∃!qα,p such that Iα (p;W) = Dα (W k qα,p | p). Furthermore,
lim qα,p −
→∞
Tα,p (qα,p )
g
Tα,p (qα,p
)
= qα,p ,
(35)
= 0,
(36)
D1 (qα,p k q) ≥ Dα (W k q| p) − Iα (p;W) ≥ Dα (qα,p k q)
∀q ∈ P(Y),
(37)
and qα,p ∼ q1,p . In addition,11 if q1,p ≺q and Tα,p (q) = q, then qα,p = q.
(d) If α ∈ (1, ∞), then ∃!qα,p such that Iα (p;W) = Dα (W k qα,p | p). Furthermore,
Tα,p (qα,p ) = qα,p ,
Dα (qα,p k q) ≥ Dα (W k q| p) − Iα (p;W) ≥ D1 (qα,p k q)
(38)
∀q ∈ P(Y),
(39)
and qα,p ∼ q1,p . In addition, if Tα,p (q) = q, then qα,p = q.
(e) If α ∈ ℜ+ \ {1}, then
qα,p
α
k W | p) + I1 (p;Wαqα,p)
1−α D1 (Wα
(
α
D1 (V k W | p) + I1 (p;V)
inf V ∈P(Y|X) 1−α
=
α
supV ∈P(Y|X) 1−α D1 (V k W | p) + I1 (p;V)
Iα (p;W) =
=
α
1−α
inf V ∈P(Y|X)
α ∈ (0, 1)
α ∈ (1, ∞)
1−α
D1 (V k W | p) + α I1 (p;V) .
(40)
(41)
Part (a) is proved using Lemma 1; Iα (p;W) ≤ ℏ(p) was proved by Csiszár through a different argument in [1, (24)]. Part
(b), which is well known, is proved by substitution. Part (c) is due to12 Augustin [2, Lemma 34.2]. Part (d) is new to the best
our knowledge. Part (e) was proved for the finite Y case by Csiszár [1, (A24), (A27)].
Definition 7. For any α ∈ ℜ+ , W : X → P(Y), p ∈ P(X) the unique probability measure qα,p on (Y, Y) satisfying
Iα (p;W) = Dα (W k qα,p | p) is called the order α Augustin mean for the input distribution p.
Lemma 2 and (34), (37), (39), imply the following bound, which is analogous to [32, Thm. 3.1] of Csiszár.
q
α(p;W)
≥ kqα,p − qk
∀q ∈ P(Y), ∀α ∈ ℜ+ .
2 Dα(W kq|p)−I
α∧1
The Augustin information and mean have closed form expression only for α = 1; for other orders we do not have a closed
form expression. However, the fixed point property Tα,p (qα,p ) = qα,p established in Lemma 13-(c,d) and the definition of
Tα,p (·) given in equation (33) imply the following identity for the Augustin mean:
hX
α
i1
dqα,p
dW (x )
(1−α)Dα(W (x )kqα,p ) α
=
p(x
)
e
∀ν : q1,p ≺ν.
(42)
dν
dν
x
In §III-C, we use this identity in lieu of a closed form expression while analyzing Iα (p;W) and qα,p as a function of α.
Q⊗
Lemma 14. If W is a product channel for finite index set T and input distribution p satisfies p = t∈T pt for some pt ∈ P(Xt ),
then
X
Iα (p;W) =
Iα (pt ;Wt)
α ∈ ℜ+ ,
(43)
t∈T
Y⊗
α ∈ ℜ+ .
(44)
qα,pt
qα,p =
t∈T
B. Augustin Information as a Function of the Input Distribution
The order α Augustin information for input distribution p is defined as the infimum of a family of conditional Rényi divergences, which are linear in p. Then the Augustin information is concave in p, because pointwise infimum of a family of
concave functions is concave. Lemma 15 strengthens this observation using Lemma 13.
Lemma 15. For any α ∈ ℜ+ and W : X → P(Y), Iα (p;W) is a concave function of p satisfying
Iα (pβ ;W) ≥ βIα (p1 ;W) + (1 − β)Iα (p0 ;W) + βDα∧1 qα,p1 k qα,pβ + (1 − β)Dα∧1 qα,p0 k qα,pβ
Iα (pβ ;W) ≤ βIα (p1 ;W) + (1 − β)Iα (p0 ;W) + βDα∨1 qα,p1 k qα,pβ + (1 − β)Dα∨1 qα,p0 k qα,pβ
Iα (pβ ;W) ≤ βIα (p1 ;W) + (1 − β)Iα (p0 ;W) + ℏ(β) − Dα∧1 qα,pβ βqα,p1 + (1 − β)qα,p0
(45)
(46)
(47)
11 Note that T
α,p (q) = q, on its own, does not imply qα,p = q for when α is less than one. Consider for example a binary symmetric channel and let q
be the probability measure that puts all of its probability to one of the output letters. Then Tα,p (q) = q, but qα,p 6= q, for all p ∈ P(X) and α ∈ (0, 1).
g
g
12 To be precise [2, Lemma 34.2] does not include the assertion D (q
1 α,p k q) ≥ Dα (W k q| p) − Iα (p;W) and claims (36) for q1,p instead of qα,p . We
cannot verify the correctness of Augustin’s proof of [2, Lemma 34.2], see Appendix C for a more detailed discussion.
10
where pβ = βp1 + (1 − β)p0 for all p0 , p1 ∈ P(X) and β ∈ [0, 1].
Lemma 15 implies that for any positive order α and channel W , the order α Augustin information Iα (p;W) is a continuous
function of the input distribution p iff supp∈P(X) Iα (p;W) is finite.13 Furthermore, if supp∈P(X) Iη (p;W) is finite for an η ∈ ℜ+
then {Iα (p;W)}α∈(0,η] is uniformly equicontinuous in p on P(X).
In order to see why the finiteness of supp∈P(X) Iα (p;W) is necessary for the continuity, note that (45) imply that
|Iα (pβ ;W) − Iα (p0 ;W)| ≥ β|Iα (p1 ;W) − Iα (p0 ;W)|.
(48)
On the other hand |pβ − p0 | ≤ 2β. Thus if there exists a {pı }ı∈Z+ ⊂ P(X) such that limı↑∞ Iα (pı ;W) = ∞ then Iα (p;W) is
discontinuous at every p in P(X)
The converse statement, i.e. the sufficiency, can be established together with the equicontinuity. For any p1 , p2 ∈ P(X) such
that p1 6= p2 let s∧ , s1 , and s0 be
s∧ =
s1 =
s0 =
p1 ∧p0
kp1 ∧p0 k ,
p1 −p1 ∧p0
1−kp1 ∧p0 k ,
p0 −p1 ∧p0
1−kp1 ∧p0 k .
Then s∧ , s1 , s0 ∈ P(X) and s1 ⊥ s0 . Furthermore,
p1 = ( 2−kp21 −p0 k )s∧ +
p0 = ( 2−kp21 −p0 k )s∧ +
Thus as a result of Lemmas 2 and 15 we have
0k
+
Iα (p0 ;W)−Iα (p1 ;W) ≤ ℏ kp1 −p
2
0k
≤ ℏ kp1 −p
+
2
Thus
|Iα (p0 ;W)−Iα (p1 ;W)| ≤ ℏ
kp1 −p0 k
2
kp1−p0 k
(Iα (s0 ;W)
2
kp1 −p0 k
s1 ,
2
kp1 −p0 k
s0 .
2
− Iα (s1 ;W))
kp1−p0 k
Iα (s0 ;W)
2
+
kp1−p0 k
2
supp∈P(X) Iη (p;W)
∀p1 , p0 ∈ P(X), α ∈ ℜ+ .
(49)
∀p1 , p0 ∈ P(X), α ∈ (0, η].
C. Augustin Information as a Function of the Order
The main goal of this subsection is to characterize the behavior of the Augustin information as a function of the order for
a given input distribution. Lemma 16 presents preliminary observations that facilitate the analysis of Augustin information as
a function of the order; results of this analysis are presented in Lemma 17.
Lemma 16. For any channel W of the form W : X → P(Y) and input distribution p ∈ P(X),
1
(a) Dα (W (x )k qα,p ) ≤ ln p(x
),
1
α∧1
(b) [p(x )]
W (x ) ≤ qα,p ,
dq
1
≤ |α−1|
(c) ln dqα,p
α ln minx :p(x )>0 p(x ) .
1,p
Bounds given in Lemma 16 follow from (42) with elementary manipulations.
Lemma 17. For any channel W of the form W : X → P(Y) and input distribution p ∈ P(X),
(a) (α − 1)Iα (p;W) is a convex function of α from ℜ+ to [−ℏ(p) , ∞).
+
(b) 1−α
α Iα (p;W) is a nonincreasing and continuous function of α from ℜ to ℜ.
(c) Iα (p;W) is a nondecreasing and continuous function of α from ℜ+ to [0, ℏ(p)].
dq
(d) {ln dqα,p
}y∈Y is an equicontinuous family of functions of α on ℜ+ .
1,p
(e) Iα (p;W) is a continuously differentiable function of α from ℜ+ to [0, ℏ(p)] such that
∂
∂α Iα (p;W) α=φ
∂
∂α Dα (W k qφ,p | p) α=φ
1
α 6= 1
(φ−1)2 D1 (Wφ,p k W | p)
2
= P p(x )
.
(x )
x 2 EW (x ) ln dW
α=1
dq1,p − D1 (W (x )k q1,p )
=
(50)
(51)
Parts (a), (b), and (c) of Lemma 17 have been reported by Augustin in [2, Lemma 34.3] for orders between zero and one.
Parts (d) and (e) of Lemma 17 are new to the best of our knowledge. Lemma 17 is primarily about the Augustin information
13 The Rényi information, discussed in III-D, has already shown to satisfy analogous relations, see [14, Lemma 16-(d,e)]. The only substantial subtlety
is that for orders in (0, 1) the Rényi information is a continuous function of p even when the corresponding capacity expression is infinite because the
Rényi information is quasi-concave rather than concave in p for orders in (0, 1), see [14, Lemma 6-(a)].
11
dq
}y∈Y as a function of the
as a function of the order for a given input distribution. Part (d), i.e. the equicontinuity of {ln dqα,p
1,p
order, is derived as a necessary tool for establishing the continuity of the derivative of the Augustin information, i.e. part (e).
Note that Lemma 16-(c) has already established the equicontinuity at α = 1.
D. Augustin Information vs Rényi Information
The Augustin information is not the only parametric family of information that has been defined using the Rényi divergence;
there are other parametric families of information defined using the Rényi divergence. The Rényi information, defined first by
Gallager14 [15] and then by Sibson [33], is arguably the most prominent one among them because of its operational significance
established by Gallager [15].
Definition 8. For any α ∈ ℜ+ , W : X → P(Y), and p ∈ P(X) the order α Rényi information for the input distribution p is
Iαg (p;W) , inf q∈P(Y) Dα (p⊛W k p ⊗ q) .
As noted by Sibson [33], one can confirm by substitution that
g
g
+ Dα qα,p
q
Dα (p⊛W k p ⊗ q) = Dα p⊛W k p ⊗ qα,p
∀p ∈ P(X), q ∈ P(Y), α ∈ ℜ+ .
(52)
(53)
g
where qα,p is the Rényi mean defined in (32). Then using Lemma 2 we can conclude that
g
∀p ∈ P(X), α ∈ ℜ+ .
Iαg (p;W) = Dα p⊛W k p ⊗ qα,p
g
g
∀p ∈ P(X), q ∈ P(Y), α ∈ ℜ+ .
Dα (p⊛W k p ⊗ q) = Iα (p;W) + Dα qα,p q
(54)
(55)
For orders other than one the closed form expression given in (54) is equal to the following expression, which is sometimes
taken as the definition of the Rényi information,
Iαg (p;W) =
α
α−1
ln kµα,p k.
α ∈ ℜ+ \ {1}.
(56)
Note that unlike the order α Augustin mean, the order α Rényi mean has a closed form expression for orders other than one,
as well. Furthermore, the inequalities given in equations (34), (37), (39) of Lemma 13 are replaced by the equality given in
(55). A discussion of the Rényi information similar to the one we have presented in this section for the Augustin information
can be found in [14].
The order one Rényi information is equal to the order one Augustin information for all input distributions. For other
orders such an equality does not hold for arbitrary input distributions. However, it is possible to characterize the Augustin
information and the Rényi information in terms of one another through appropriate variational forms. Characterizing the
Augustin information in a variational form in terms of the Rényi information is especially useful, because the Augustin
information does not have a closed form expression whereas the Rényi information does.
Lemma 18. Let W be a channel of the form W : X → P(Y) and p be an input distribution in P(X).
(a) Let uα,p ∈ P(X) be uα,p (x ) =
p(x )e(1−α)Dα(W (x )kqα,p )
P
(1−α)Dα(W (x̃ )kqα,p )
x̃ p(x̃ )e
for all x ; then
1
Iα (p;W) = Iαg (uα,p ;W) + α−1
D1 (pk uα,p )
(
1
D1 (pk u) α ∈ (0, 1)
supu∈P(X) Iαg (u;W) + α−1
.
=
1
g
inf u∈P(X) Iα (u;W) + α−1 D1 (pk u) α ∈ (1, ∞)
(57)
(58)
g
(b) Let aα,p ∈ P(X) be aα,p (x ) =
p(x )e(α−1)Dα(W (x )kqα,p )
g
P
(α−1)Dα(W (x̃ )kqα,p )
x̃ p(x̃ )e
for all x ; then
1
D1 (aα,p k p)
Iαg (p;W) = Iα (aα,p ;W) − α−1
(
1
D1 (ak p)
inf a∈P(X) Iα (a;W) − α−1
=
1
D1 (ak p)
supa∈P(X) Iα (a;W) − α−1
(59)
α ∈ (0, 1)
.
α ∈ (1, ∞)
(60)
Lemma 18-(a) was first proved by Poltyrev, [19, Thm. 3.4], in a slightly different form, for α ∈ [1/2, 1) case, assuming
finite Y. Equation (60) of Lemma 18-(b) was first proved by Shayevitz, [11, Thm. 1], for finite Y case. Shayevitz, however,
neither gave the expression for the optimal aα,p , nor asserted its existence in [11]. Hence equation (59) of Lemma 18-(b) is
not reported in [11].
The following inequalities follow from either part of Lemma 18 by considering either a = p or u = p point in the variational
characterizations. Evidently, these inequalities can also be obtained using the Jensen’s inequality and the concavity of the natural
logarithm function.
Iα (p;W) ≥ Iαg (p;W)
Iα (p;W) ≤ Iαg (p;W)
14 Gallager
α ∈ (0, 1]
α ∈ [1, ∞).
uses a different parametrization and confines his discussion to α ∈ (0, 1) case.
12
(61)
(62)
IV. T HE AUGUSTIN C APACITY
In the previous section we have defined and analyzed the Augustin information and mean; our main aim in this section is
doing the same for the Augustin capacity and center. In §IV-A, we establish the existence of a unique Augustin center for
all convex constraint sets with finite Augustin capacity and investigate the implications of the existence of an Augustin center
for a given order and constraint set. In §IV-B, we analyze the Augustin capacity and center as a function of the order for a
given constraint set. In §IV-C, we bound the Augustin capacity of the convex hull of a collection of constraint sets on a given
channel in terms of the Augustin capacities of individual constraint sets and determine the Augustin capacity of products of
constraint sets on the product channels. Proofs of the propositions presented in this section can be found in Appendix D.
Augustin provided a presentation similar to the current section in [2, §§33,34] and derived many of the key results —such
as the existence of unique Augustin center and its continuity as a function of order, see [2, Lemmas 34.6, 34.7, 34.8]— for
orders not exceeding one. Augustin, however, defines capacity and center only for the subsets of P(X) defined through cost
constraints. We investigate that important special case more closely in §V.
A. Existence of a Unique Augustin Center
Definition 9. For any α ∈ ℜ+ , W : X → P(Y), and A ⊂ P(X), the order α Augustin capacity of W for constraint set A is
Cα,W,A , supp∈A Iα (p;W) .
(63)
When the constraint set A is the whole P(X), we denote the order α Augustin capacity by Cα,W , i.e. Cα,W , Cα,W,P(X) .
Using the definition of the Augustin information Iα (p;W) given in (28) we get the following expression for Cα,W,A
Cα,W,A = supp∈A inf q∈P(Y) Dα (W k q| p) .
(64)
Theorem 1 in the following demonstrates that at lest for convex A’s one can exchange the order of the supremum and infimum
without changing the value in the above expression.
Theorem 1. For any order α ∈ ℜ+ , channel W of the form W : X → P(Y), and convex constraint set A ⊂ P(X)
supp∈A inf q∈P(Y) Dα (W k q| p) = inf q∈P(Y) supp∈A Dα (W k q| p) .
(65)
If the expression on the left hand side of (65) is finite, i.e. if Cα,W ,A ∈ ℜ, then ∃!qα,W,A ∈ P(Y), called the order α Augustin
center of W for the constraint set A, satisfying
Cα,W,A = supp∈A Dα (W k qα,W,A | p) .
(66)
Furthermore, for every sequence of input distributions {p (ı) }ı∈Z+ ⊂ A satisfying limı→∞ Iα p (ı) ;W = Cα,W,A , the corresponding sequence of order α Augustin means {qα,p (ı) }ı∈Z+ is a Cauchy sequence for the total variation metric on P(Y) and
qα,W,A is the unique limit point of that Cauchy sequence.
In order to prove Theorem 1, we follow the program put forward by Kemperman [13] for establishing a similar result for
α = 1 and A = P(X) case. We first state and prove Theorem 1 assuming that the input set is finite. Then we generalize the
result to the case with arbitrary input sets. In the case when X is a finite set we can also assert the existence of an optimal
input distribution for which the Augustin information is equal to the Augustin capacity.
Lemma 19. For any order α ∈ ℜ+ , channel W of the form W : X → P(Y) with a finite input set X, and closed convex
constraint set A ⊂ P(X), there exists pe ∈ A such that Iα (e
p ;W) = Cα,W,A and ∃!qα,W,A ∈ P(Y) satisfying
Dα (W k qα,W,A | p) ≤ Cα,W,A
∀p ∈ A.
(67)
Furthermore, qα,ep = qα,W,A for all pe ∈ A such that Iα (e
p ;W) = Cα,W,A .
Proofs of Lemma 19 and Theorem 1 are presented in Appendix D. If A is P(X), then the expression on the right hand side
of (66), is equal to the Rényi radius Sα,W defined in the following. Thus Theorem 1 implies Cα,W = Sα,W .
Definition 10. For any α ∈ ℜ+ and W : X → P(Y), the order α Rényi radius of W is
Sα,W , inf q∈P(Y) supx ∈X Dα (W (x )k q) .
(68)
Theorem 1 asserts the existence of a unique order α Augustin center for convex constraint sets with finite Augustin capacity.
However, a probability measure qα,W,A satisfying (66), i.e. an order α Augustin center, can in principle exist even for nonconvex constraint sets.
Definition 11. A constraint set A for the channel W : X → P(Y) has an order α Augustin center iff ∃q ∈ P(Y) such that
supp∈A Dα (W k q| p) = Cα,W,A .
13
(69)
If Cα,W,A is infinite, then all probability measures on the output space satisfy (69) as a result of (64) and max-min inequality.
Thus for constraint sets with infinite order α Augustin capacity all probability measures on the output space are order α Augustin
centers. On the other hand, some constraint sets do not have any order α Augustin center. Consider for example p1 and p2
satisfying qα,p1 6= qα,p2 and Iα (p1 ;W) = Iα (p2 ;W). Then (69) is not satisfied by any probability measure for A = {p1 , p2 }
and A does not have an order α Augustin center. Lemma 20 asserts that if Augustin center exists for a constraint set with
finite Augustin capacity, then the Augustin center is unique.
Lemma 20. Let A ⊂ P(X) be a constraint set satisfying Cα,W,A
∈ ℜ≥0 , and qα,W,A be a probability measure satisfying (69).
Then for every {p (ı) }ı∈Z+ ⊂ A satisfying limı→∞ Iα p (ı) ;W = Cα,W,A the sequence of order α Augustin means {qα,p (ı) }ı∈Z+
is a Cauchy sequence with the limit point qα,W,A and the order α Augustin center qα,W,A is unique.
For any A that has an order α Augustin center and a finite Cα,W,A , Lemma 13-(b,c,d) and Lemma 20 imply that
Cα,W,A − Iα (p;W) ≥ Dα∧1 (qα,p k qα,W,A )
∀p ∈ A.
(70)
Lemma 13-(b,c,d) and Lemma 20 can also be used establish a lower bound on supp∈A Dα (W k q| p) in terms of the Augustin
capacity and center.
Lemma 21. For any constraint set A that has an order α Augustin center and a finite Cα,W,A we have
supp∈A Dα (W k q| p) ≥ Cα,W,A + Dα∧1 (qα,W,A k q)
∀q ∈ P(Y).
(71)
Note that the form of the lower bound given in (71) is, in a sense, analogous to the ones given in (34), (37), (39). The bound
given in (71) is a van Erven-Harremoës bound15 for α ∈ (0, 1], but it is not a van Erven-Harremoës bound for α ∈ (1, ∞)
because we have a D1 (qα,W,A k q) term rather than a Dα (qα,W,A k q) term for α ∈ (1, ∞).
For order other than one, using Csiszár’s form for the Augustin information given in (41) and the definition of the Augustin
capacity we obtain the following expressions:
(
α
D1 (V k W | p) + I1 (p;V) α ∈ (0, 1)
supp∈A inf V ∈P(Y|X) 1−α
.
(72)
Cα,W,A =
α
D1 (V k W | p) + I1 (p;V) α ∈ (1, ∞)
supp∈A supV ∈P(Y|X) 1−α
Then
Cα,W,A = supV ∈P(Y|X) supp∈A
α
1−α D1 (V k W | p)
∀α ∈ (1, ∞).
+ I1 (p;V)
For α ∈ (0, 1), if the constraint set A has an order α Augustin center, e.g. when A is convex, then one can exchange the order
of infimum and supremum and replace the supremum with a maximum as a result of Lemma 20. In other words, for orders
between zero and one Lemma 20 implies the minimax identity for the expression given in (72).
Lemma 22. For any α ∈ (0, 1), if the constraint set A for the channel W : X → P(Y) has an order α Augustin center, then
Cα,W,A = inf V ∈P(Y|X) supp∈A
q
α
1−α D1 (V k W | p)
+ I1 (p;V) .
(73)
If Cα,W,A is finite, then Wα α,W,A satisfies
Cα,W,A = supp∈A
α
1−α D1
q
q
Wα α,W,A W p + I1 p;Wα α,W,A .
(74)
In [34], Blahut proved a similar result assuming both X and Y are finite sets and A = P(X). Even under those assumptions
Blahut’s result [34, Thm. 16] imply (73) and (74) for all orders in (0, 1) only when Cα,W is a differentiable function of the
order. In particular, Blahut was motivated by the expression for the sphere packing exponent; consequently, [34, Thm. 16] is
stated in terms of an optimal input distribution at a given rate R ∈ (C0,W , C1,W ) and the corresponding optimal order α∗ (R).
B. Augustin Capacity and Center as a Function of the Order
Lemma 23. For any channel W of the form W : X → P(Y) and constraint set A ⊂ P(X),
(a) Cα,W,A is a nondecreasing and lower semicontinuous function of α on ℜ+ .
16
(b) 1−α
(0, 1).
α Cα,W,A is a nonincreasing and continuous function of α on
(c) (α − 1)Cα,W,A is a convex function of α on (1, ∞).
(d) Cα,W,A is nondecreasing and continuous in α on (0, 1] and (1, χW ,A ] where χW ,A , sup{φ : Cφ,W,A ∈ ℜ≥0 }.
(e) If supp∈A Iφg (p;W) ∈ ℜ≥0 for a φ > 1, then Cα,W,A is nondecreasing and continuous in α on (0, (1 ∨ χW ,A )].
15 In
[9] van Erven and Harremoës have conjectured that the inequality supx ∈X Dα (W (x )k q) ≥ Cα,W + Dα qα,W q holds for all q ∈ P(Y).
Van Erven and Harremoës have also proved the bound for the case when α = ∞, assuming that Y is countable [9, Thm. 37]. We have confirmed van
Erven-Harremoës conjecture in [14, Lemma 19] and generalized it to the convex constrained case for the Rényi capacity and center in [14, Lemma 25]. See
§IV-D for a brief discussion of the Rényi capacity and center; a more comprehensive discussion can be found in [14].
16 We exclude α = 1 case because we do not want to assume C
1,W,A to be finite.
14
The continuity results presented in parts (d) and (e) are somewhat unsatisfactory. One would like to either establish the
continuity of Cα,W,A from the right at α = 1 whenever Cφ,W,A is finite for a φ > 1 or provide a channel W and a constraint
set A for which Cφ,W,A is finite for a φ > 1 and limα↓1 Cα,W,A > C1,W,A . We could not do either. Instead we establish the
continuity of Cα,W,A from the right at α = 1 assuming that supp∈A Iφg (p;W) is finite for a φ > 1.
If Cφ,W is finite, then supp∈A Iφg (p;W) is finite for all A ⊂ P(X) by (52) and Theorem 1. Thus Cα,W,A is nondecreasing
and continuous in α on (0, χW ,A ] for all A ⊂ P(X), provided that Cφ,W is finite for a φ > 1.
Lemma 21 allows us to use the continuity of Cα,W,A in α and Lemma 2 to establish the continuity of qα,W,A in α for the
total variation topology on P(Y).
Lemma 24. For any η ∈ ℜ+ , W : X → P(Y), and convex A ⊂ P(X) such that Cη,W,A ∈ ℜ+ ,
Dα∧1 (qα,W,A k qφ,W ,A ) ≤ Cφ,W ,A − Cα,W,A
∀α, φ such that 0 < α < φ ≤ η.
(75)
Consequently, if Cα,W,A is continuous in α on I for some I ⊂ (0, η], then qα,W,A : I → P(Y) is continuous in α on I for
the total variation topology on P(Y).
C. Convex Hulls of Constraints and Product Constraints
In the following we consider two kinds of frequently encountered constraint sets that are described in terms of simpler
constraint sets. Lemma 25 considers convex hull of a family constraint sets and bounds the Augustin capacity for the convex
hull in terms of the Augustin capacities of the individual constraint sets. Lemma 26 considers a product channel for a constraint
set that is the product of convex constraints sets on the component channels and shows that its Augustin capacity has an additive
form and its Augustin center has a product form.
Lemma 25. Let α be a positive real, W be a channel of the form W : X → P(Y), and A(ı) be a constraint set that has an
order α Augustin center and a finite Cα,W ,A(ı) for all ı ∈ T. Then
X
C
supı∈T Cα,W,A(ı) ≤ Cα,W,A ≤ ln
(76)
e α,W,A(ı)
ı∈T
(ı)
where A is the convex hull of the union, i.e. A = ch(∪ı∈T A ). Furthermore,
• Cα,W,A(ı) = Cα,W,A ⇔ supp∈A Dα W k qα,W,A(ı) p ≤ Cα,W,A(ı) ⇒ qα,W,A = qα,W,A(ı) .
P
P
−C
C
Cα,W,A(ı)
⇔ qα,W,A(ı) ⊥ qα,W,A() ∀ı 6= ⇒ qα,W,A = ı∈T e α,W,A(ı) α,W,A qα,W,A(ı) .
• Cα,W,A = ln
ı∈T e
Note that if A(ı) is convex and Cα,W ,A(ı) is finite, then A(ı) has a unique order α Augustin center by Theorem 1.
Lemma 26. Let α be a positive real number, W : X → P(Y) be a product channel for a finite index set T, and At ⊂ P(Xt )
be constrain sets with order α Augustin center for all t ∈ T. Then
X
Cα,W,chA = Cα,W,A =
(77)
Cα,Wt ,At .
t∈T
where chA is the convex hull of A and A =
Q×
t∈T At .
Furthermore, if Cα,W,A ∈ ℜ+ , then qα,W,chA = qα,W,A =
Q⊗
t∈T qα,Wt ,At .
The constraint set A is not convex whenever chA 6= A; nonetheless, A is guaranteed to have an order α Augustin center.
D. Augustin Capacity vs Rényi Capacity
Using the Rényi information instead of the Augustin information, one can define the Rényi capacity, as follows.
Definition 12. For any α ∈ ℜ+ , W : X → P(Y), and A ⊂ P(X) the order α Rényi capacity of W for constraint set A is
g
Cα,W,A
, supp∈A Iαg (p;W) .
(78)
g
g
g
.
When the constraint set A is the whole P(X), we denote the order α Rényi capacity by Cα,W
, i.e. Cα,W
, Cα,W,P(X)
g
Since I1 (p;W) = I1g (p;W), C1,W
,A = C1,W ,A by definition. We cannot say the same for other orders; by (61), (62) we have
g
Cα,W
,A ≤ Cα,W,A
α ∈ (0, 1],
g
Cα,W ,A ≥ Cα,W,A
α ∈ [1, ∞).
(79)
(80)
As a result of definitions of the Rényi information and capacity we have
g
Cα,W,A
= supp∈A inf q∈P(Y) Dα (p⊛W k p ⊗ q) .
The Rényi capacity satisfies a minimax theorem, [14, Thm. 2], similar to Theorem 1: For any convex constraint set A ⊂ P(X)
supp∈A inf q∈P(Y) Dα (p⊛W k p ⊗ q) = inf q∈P(Y) supp∈A Dα (p⊛W k p ⊗ q) .
15
g
g
If Cα,W
,A is finite, then ∃!qα,W ,A ∈ P(Y), the order α Rényi center W for the constraint set A, satisfying
g
g
Cα,W,A
= supp∈A Dα p⊛W k p ⊗ qα,W
,A .
g
g
Thus for A = P(X) case, the Rényi capacity equals to the Rényi radius. Hence Cα,W
= Cα,W and qα,W
= qα,W by Theorem
1. The other observations presented in this section have their counter parts for the Rényi capacity and center. Compare for
example Lemma 21 and [14, Lemma 25], Lemma 25 and [14, Lemma 26], or Lemma 26 and [14, Lemma 27].
V. T HE C OST C ONSTRAINED P ROBLEM
In the previous section, we have defined the Augustin capacity for arbitrary constraint sets and proved the existence of a
unique Augustin center for any convex constraint set with finite Augustin capacity. The convex constraint sets of interest are
often defined via the cost constraints; the main aim of this section is to investigate this important special case more closely. In
§V-A we investigate the immediate consequences of the definition of the cost constrained Augustin capacity and ramifications
of the analysis presented in the previous section. In §V-B we define and analyze the Augustin-Legendre (A-L) information,
capacity, radius, and center. The discussion in §V-B is a generalization of certain parts of the analysis presented by Csiszár and
Körner in [6, Ch. 8] for the supremum of the mutual information for discrete channels with single cost constraint, i.e. α = 1,
|X| < ∞, |Y| < ∞, ℓ = 1 case. In §V-C we define and analyze the Rényi-Gallager (R-G) information, mean, capacity, radius,
and center. The most important conclusion of our analysis in §V-C is the equality of the A-L capacity and center to the R-G
capacity and center. In §V-D, we demonstrate how the results presented in §V-A, §V-B, and §V-C can be used to determine
the Augustin capacity and center of a transition probability with cost constraints. Proofs of the propositions presented in §V-A,
§V-B, and §V-C can be found is Appendix E.
Augustin presented a discussion of the cost constrained capacity Cα,W,̺ in [2, §34] for the case when the cost function ρ
is a bounded function of the form ρ : X → [0, 1]ℓ and the order α is in (0, 1]. In [2, §35], Augustin also analyzed quantities
closely related to the R-G information and capacity. The quantities analyzed by Augustin in [2, §35] have first appeared in
Gallager’s error exponents analysis for cost constrained channels [15, §6],[16, §7.3,§7.4,§7.5]. Unlike Augustin, Gallager did
not assume ρ to be bounded; but Gallager confined his analysis to the case when there is a single cost constraint, i.e. ℓ = 1
case and refrained from defining the R-G capacity as a quantity that is of interest on its own right. Other authors studying
cost constrained problems, [23, §IV], [25], [24], [26], have considered the R-G information and capacity, as well. Yet to the
best of our knowledge for orders other than one none of the A-L information measures, which are obtained through a more
direct application of convex conjugation, have been studied before.
A. The Cost Constrained Augustin Capacity and Center
We denote the set of all probability mass functions satisfying a cost constraint ̺ by A(̺), i.e.
A(̺) , {p ∈ P(X) : Ep [ρ] ≤ ̺}.
(81)
A(̺) 6= ∅ iff ̺ ∈ Γρ for where Γρ is the set of all feasible cost constraints defined in (8). A(̺) is nondecreasing in ̺, i.e. if
̺1 ≤ ̺2 , then A(̺1 ) ⊂ A(̺2 ). We define the order α Augustin capacity of W for the cost constraint ̺ as
(
supp∈A(̺) Iα (p;W) if ̺ ∈ Γρ
∀α ∈ ℜ+ .
(82)
Cα,W,̺ ,
−∞
if ̺ ∈ ℜℓ≥0 \ Γρ
We defined Cα,W,̺ for ̺’s that are not feasible in order to be able to use standard results without modifications. Since A(̺) is
a convex set, Theorem 1 holds for A(̺). We denote17 the order α Augustin center of W for the cost constraint ̺ by qα,W,̺ .
For a given order α, the Augustin capacity Cα,W,̺ is a concave function of the cost constraint ̺. Hence, if it is finite at
an interior point of Γρ , then it is a continuous function of the constraint that lies below the tangent planes drawn at interior
points of Γ̺ . Lemma 27, in the following, summarizes these observations.
Lemma 27. Let W be a channel of the form W : X → P(Y) with the cost function ρ of the form ρ : X → ℜℓ≥0 .
(a) For any α ∈ ℜ+ , Cα,W,̺ is a nondecreasing and concave function of ̺ on ℜℓ≥0 , which is either infinite on every point in
intΓρ or finite and continuous on intΓρ .
(b) If Cα,W,̺ is finite on intΓρ for an α ∈ ℜ+ , then for every ̺ ∈ intΓρ there exists a λα,W,̺ ∈ ℜℓ≥0 such that
Cα,W,̺ + λα,W,̺ · (˜
̺ − ̺) ≥ Cα,W,̺˜
∀˜
̺ ∈ ℜℓ≥0
(83)
Furthermore, the set of all such λα,W,̺ ’s is convex and compact.
(c) Either Cα,W,̺ = ∞ for all (α, ̺) ∈ (0, 1)×intΓρ or Cα,W,̺ and qα,W,̺ are continuous in (α, ̺) on (0, 1)×intΓρ for the
total variation topology on P(Y).
17 This slight abuse of notation —which can be avoided by using C
α,W,A(̺) and qα,W,A(̺) instead of Cα,W,̺ and qα,W,̺ — provides brevity without
leading to any notational ambiguity.
16
If the cost function for a product channel is additive, then the cost constrained Agustin capacity of the product channel is
equal to the supremum of the sum of the cost constraint capacities of the component channels over all feasible cost allocations.
Furthermore, if there exists an optimal cost allocation, then the Augustin center of the product channel is a product measure.
Lemma 28, in the following, states these observations formally.
LemmaP
28. Let W : X → P(Y) be a product channel for a finite index set T with a cost function ρ : X → ℜℓ≥0 satisfying
ρ(x ) = t∈T ρt (xt ) for some ρt : Xt → ℜℓ≥0 . Then18
nX
o
X
Cα,W,̺ = sup
Cα,Wt ,̺t :
̺t ≤ ̺, ̺t ∈ ℜℓ≥0
∀̺ ∈ ℜℓ≥0 , α ∈ ℜ+ .
(84)
t∈T
t∈T
Q⊗
Cα,Wt ,̺t , then qα,W,̺ = t∈T qα,Wt ,̺t .
P
̺
whenever W is
Since the Augustin capacity is concave in the cost constraint by Lemma 27-(a), Cα,W,̺ = t∈T Cα,Wt , |T|
stationary and ρt = ρ1 for all t ∈ T. Alternatively, if Γρt ’s are closed and Cα,Wt ,̺ ’s are upper semicontinuous functions of ̺
19
on Γρt ’s, then we can use
P the extreme value theorem for the upper semicontinuous functions to establish the existence of a
{̺t }t∈T s.t. Cα,W,̺ = t∈T Cα,Wt ,̺t . However, such an existence assertion does not hold in general, see Example 3.
If Cα,W,̺ ∈ ℜ≥0 for an α ∈ ℜ+ and there exists a {̺t }t∈T such that Cα,W,̺ =
P
t∈T
B. The Augustin-Legendre Information Measures
The cost constrained Augustin capacity Cα,W,̺ and center qα,W,̺ can be characterized using convex conjugation, as well.
In this part of the paper, we introduce and analyze the concepts of the Augustin-Legendre information, capacity, center, and
radius in order to obtain a more complete understanding of this characterization. The current method seems to us to be the
standard application of the convex conjugation technique to characterize the cost constrained Augustin capacity. Yet, it is not
the customary method. Starting with the seminal work of Gallager [15], a more ad hoc method based on the Rényi information
became the customary way to apply Lagrange multipliers techniques to characterize the Augustin capacity, see [2, §35], [25],
[24]. We discuss that approach in §V-C. Theorem 2 presented in the following and Theorem 3 presented in §V-C establish
the equivalence of these two approaches by establishing the equality of the Augustin-Legendre capacity and center to the
Rényi-Gallager capacity and center.
Definition 13. For any α ∈ ℜ+ , channel W of the form W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , p ∈ P(X), and
λ ∈ ℜℓ≥0 , the order α Augustin-Legendre information for input distribution p and Lagrange multiplier λ is
Iαλ (p;W) , Iα (p;W) − λ · Ep [ρ] .
(85)
Note that as an immediate consequence of the definition of the A-L information we have
inf λ≥0 Iαλ (p;W) + λ · ̺ = ξα,p (̺)
(86)
where ξα,p (·) : ℜℓ≥0 → [−∞, ∞) is defined as
(
Iα (p;W)
ξα,p (̺) ,
−∞
̺ ≥ Ep [ρ]
.
else
(87)
Then the Augustin-Legendre information Iαλ (p;W) can also be expressed as20
Iαλ (p;W) = sup̺≥0 ξα,p (̺) − λ · ̺
(88)
Definition 14. For any α ∈ ℜ+ , channel W of the form W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and λ ∈ ℜℓ≥0 the
order α Augustin-Legendre (A-L) capacity for Lagrange multiplier λ is
λ
Cα,W
, supp∈P(X) Iαλ (p;W) .
(89)
Then as a result of (87) and (88) we have
λ
Cα,W
= sup̺≥0 Cα,W ,̺ − λ · ̺
∀λ ∈ ℜℓ≥0 .
(90)
∀̺ ∈ ℜℓ≥0 .
(91)
Hence, using max-min inequality we can conclude that
λ
Cα,W,̺ ≤ inf λ≥0 Cα,W
+λ·̺
18 If
P
Cα,Wt ,̺t = −∞ for any t ∈ T, then t ∈T Cα,Wt ,̺t stands for −∞; even if one or more of other Cα,Wt ,̺t ’s are equal to ∞.
P
P
19 Consider the function f (̺ , . . . , ̺
1
|T| ) which is equal to
t ∈T ̺t ≤ ̺ and ̺t ∈ Γρt for all t ∈ T and which is equal to −∞
t ∈T Cα,Wt ,̺t if
otherwise. We choose a large enough but bounded set using the vector ̺ to obtain a compact set for the supremum.
−γ
∗
∗ : (−∞, 0]ℓ → ℜ are defined as f (̺) = −ξ
∗
20 Note that if f : ℜℓ
α,p (̺) and f (γ) = Iα (p;W), then f is the convex conjugate,
≥0 → (−∞, ∞] and f
λ
i.e. Legendre transform, of the convex function f . This is why we call Iα (p;W) the Augustin-Legendre information.
17
λ
λ
Then Cα,W,̺ < ∞ for all ̺ ∈ ℜℓ≥0 provided that Cα,W
< ∞ for a λ ∈ ℜ≥0 . But Cα,W
= ∞ might hold for λ small enough
ℓ
21
even when Cα,W,̺ < ∞ for all ̺ ∈ ℜ≥0 , see Example 1.
The inequality given in (91) is an equality for many cases of interest as demonstrated by the following lemma. However,
the inequality given in (91) is not an equality in general, see Example 2.
Lemma 29. Let α ∈ ℜ+ and W be a channel of the form W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 . Then
λ−ǫ1
λ
(a) Cα,W
is convex, nonincreasing, and lower semicontinuous in λ on ℜℓ≥0 and continuous in λ on {λ : ∃ǫ > 0 s.t. Cα,W
< ∞}.
λ
(b) If X is a finite set, then Cα,W,̺ = inf λ≥0 Cα,W + λ · ̺.
λ
(c) If ̺ ∈ intΓρ , then Cα,W,̺ = inf λ≥0 Cα,W
+ λ · ̺. If in addition Cα,W,̺ < ∞, then there exists a non-empty convex,
λα,W,̺
+ λα,W,̺ · ̺.
compact set of λα,W,̺ ’s satisfying (83) and Cα,W,̺ = Cα,W
λ
λ
for all
(d) If Cα,W,̺ is finite and Cα,W,̺ = Cα,W
+ λ · ̺ some ̺ ∈ Γρ and λ ∈ ℜℓ≥0 , then limı→∞ Iαλ p (ı) ;W = Cα,W
(ı)
(ı)
{p }ı∈Z+ ∈ A(̺) s.t. limı→∞ Iα p ;W = Cα,W,̺ .
λ
λ
Using the definitions of Iα (p;W), Iαλ (p;W), and Cα,W
given in (28) (85), (89) we get the following expression for Cα,W
.
λ
Cα,W
= supp∈P(X) inf q∈P(Y) Dα (W k q| p) − λ · Ep [ρ] .
(92)
The A-L capacity satisfies a minimax theorem similar to the one satisfied by the Augustin capacity, which allows us to
assert the existence of a unique A-L center whenever the A-L capacity is finite.
Theorem 2. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and Lagrange multiplier λ ∈ ℜℓ≥0
supp∈P(X) inf q∈P(Y) Dα (W k q| p) − λ · Ep [ρ] = inf q∈P(Y) supp∈P(X) Dα (W k q| p) − λ · Ep [ρ]
= inf q∈P(Y) supx ∈X Dα (W (x )k q) − λ · ρ(x ).
(93)
(94)
λ
λ
If the expression on the left hand side of (93) is finite, i.e. if Cα,W
< ∞, then ∃!qα,W
∈ P(Y), called the order α AugustinLegendre center of W for the Lagrange multiplier λ, satisfying
λ
λ
Cα,W
= supp∈P(X) Dα W k qα,W
p − λ · Ep [ρ]
(95)
λ
(96)
= supx ∈X Dα W (x )k qα,W − λ · ρ(x ).
λ
Furthermore, for every sequence of input distributions {p (ı) }ı∈Z+ ⊂ P(X) such that limı→∞ Iαλ p (ı) ;W = Cα,W
, the
corresponding sequence of order α Augustin means {qα,p (ı) }ı∈Z+ is a Cauchy sequence for the total variation metric on
λ
P(Y) and qα,W
is the unique limit point of that Cauchy sequence.
Note that Theorem 2 for λ = 0 is nothing but Theorem 1 for A = P(X). The proof of Theorem 2 is very similar to that of
Theorem 1, as well; it employs Lemma 30, presented in the following, instead of Lemma 19. Note that, Lemma 30 for λ = 0
is nothing but Lemma 19 for A = P(X), as well.
Lemma 30. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 for a finite input set X, and
λ
λ
p ;W) = Cα,W
and ∃!qα,W
∈ P(Y) satisfying
Lagrange multiplier λ ∈ ℜℓ≥0 , there exists a pe ∈ P(X) such that Iαλ (e
λ
λ
Dα W k qα,W
p − λ · Ep [ρ] ≤ Cα,W
∀p ∈ P(X).
(97)
λ
λ
Furthermore, qα,ep = qα,W
for all e
p ∈ P(X) such that Iαλ (e
p ;W) = Cα,W
.
Note that the expression on the left hand side of equation (93) is nothing but the A-L capacity. Thus Theorem 2 is establishes
the equality of the A-L capacity to the A-L radius defined in the following.
Definition 15. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and λ ∈ ℜℓ≥0 , the order α
Augustin-Legendre radius of W for Lagrange multiplier λ is
λ
Sα,W
, inf q∈P(Y) supx ∈X Dα (W (x )k q) − λ · ρ(x ).
λ
is finite, then Lemma 13-(b,c,d), Theorem 2 and the definition of Iαλ (p;W) given in (85) imply that
If Cα,W
λ
λ
∀p ∈ P(X).
Cα,W
− Iαλ (p;W) ≥ Dα∧1 qα,p k qα,W
(98)
(99)
Using Lemma 13 and Theorem 2 one can also establish a bound similar to the one given in Lemma 21. However, we will not
do so here because one can obtain a slightly stronger results, using the characterization of the A-L capacity and center via
R-L capacity and center presented in §V-C, see Lemma 35 and the ensuing discussion.
21 Augustin [2, §33-§35] considered the case when the cost function ρ is a bounded function of the form ρ : X → [0, 1]ℓ . In that case C λ
α,W < ∞ for all
0
λ
λ ∈ ℜℓ≥0 provided that Cα,W,̺ < ∞ for a ̺ ∈ intΓρ because Cα,W ,1 < ∞ by Lemma 13-(b) and Cα,W ,1 = Cα,W and Cα,W
= Cα,W for
≤ Cα,W
all λ ∈ ℜℓ≥0 by definition.
18
As a result of Lemma 29-(c) we know that if ̺ ∈ intΓρ and Cα,W ,̺ is finite, then there exists at least one λα,W ,̺ for
λα,W ,̺
+ λα,W ,̺ · ̺. Lemma 31, given in the following, asserts that for any such Lagrange multiplier the
which Cα,W ,̺ = Cα,W
corresponding order α A-L center should be equal to the order α Augustin center for the cost constraint ̺. Thus if there are
multiple λα,W ,̺ they all have the same order α A-L center.
Lemma 31. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and a cost constraint ̺ ∈ Γρ such
λ
λ
that Cα,W,̺ < ∞, if Cα,W,̺ = Cα,W
+ λ · ̺ for a λ ∈ ℜℓ≥0 , then qα,W,̺ = qα,W
.
The Augustin capacity has an additive form and the Augustin center has a multiplicative form, whenever it exists, for product
constraints on product channels by Lemma 26. The additive cost constraints, however, are not product constraints. In order
to calculate the cost constrained Augustin capacity for product channels with additive cost functions, we need to optimize
over the feasible allocations of the cost over the component channels by Lemma 28. In addition we can express the cost
constrained Augustin center of the product channel as the product of the cost constrained Augustin centers of the components
channels, using Lemma 28, only when there exists a feasible allocation of the power that achieves the optimum value. For the
A-L capacity and center, on the other hand, we have a considerably neater picture: For product channels with additive cost
functions the A-L capacity is additive and the A-L center is multiplicative, whenever it exists.
Lemma 32. Let W
for a finite index set T with an additive cost function ρ : X → ℜℓ≥0
P : X → P(Y) be a product channel
ℓ
satisfying ρ(x ) = t∈T ρt (xt ) for some ρt : Xt → ℜ≥0 . Then
X
λ
λ
Cα,W
=
∀λ ∈ ℜℓ≥0 , α ∈ ℜ+ .
Cα,W
(100)
t
t∈T
Furthermore, if
λ
Cα,W
< ∞, then
λ
qα,W
=
Q⊗
λ
t∈T qα,Wt .
C. The Rényi-Gallager Information Measures
In §V-B we have characterized the cost constrained Augustin capacity and center in terms of the A-L capacity and center,
which is defined as the supremum of the A-L information. Another definition of an information with a Lagrange multiplier
was tacitly proposed by Gallager in [15, (103) and (116)]. Augustin showed at least for the case when the cost function is
bounded the cost constrained Augustin capacity can be characterized by the supremum of this information [2, Lemmas 35.4(b)
and 35.8(b)]. We call the supremum of this information the R-G capacity. The main aim of this subsection is establishing the
equality of the A-L capacity and center to the R-G capacity and center. We will also derive a van Erven-Harremoës bound for
the A-L capacity and center and use it to derive the continuity of the A-L center as function of the Lagrange multiplier λ.
Definition 16. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , p ∈ P(X), and λ ∈ ℜℓ≥0 the
order α Rényi-Gallager (R-G) information for input distribution p and Lagrange multiplier λ is
(
1−α
α λ·ρ
p
⊗
q
α ∈ ℜ+ \ {1}
p⊛W
e
inf
D
q∈P(Y) α
.
(101)
Iαgλ (p;W) ,
inf q∈P(Y) D1 (p⊛W k p ⊗ q) − λ · Ep [ρ] α = 1
If λ is a vector of zeros, then the R-G information is the Rényi information; similar to the Rényi information, the R-G
information has a closed from expression, described in terms of a mean achieving the infimum in its definition.
Definition 17. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , p ∈ P(X), and λ ∈ ℜℓ≥0 , the
order α power mean for the input distribution p and Lagrange multiplier λ is
α i α1
hX
dµλ
α,p
(1−α)λ·ρ(x ) dW (x )
,
p(x
)e
.
(102)
dν
dν
x
The order α Rényi-Gallager (R-G) mean for the input distribution p and Lagrange multiplier λ is
gλ
qα,p
,
µλ
α,p
kµλα,p k
.
(103)
gλ
Both µλα,p and qα,p
depend on the Lagrange multiplier λ for α ∈ ℜ+ \{1}. Furthermore, one can confirm by substitution that
1−α
1−α
gλ
gλ
+ Dα qα,p
q
α ∈ ℜ+ \ {1}.
(104)
Dα p⊛W e α λ·ρ p ⊗ q = Dα p⊛W e α λ·ρ p ⊗ qα,W
Then as a result of Lemma 2 we have
1−α
gλ
Iαgλ (p;W) = Dα p⊛W e α λ·ρ p ⊗ qα,W
=
α
α−1
ln
µλα,p
.
19
(105)
α ∈ ℜ+ \ {1}.
(106)
gλ
Neither µλ1,p nor q1,p
depends on the Lagrange multiplier λ. In addition, one can confirm by substitution that
gλ
gλ
q .
− λ · Ep [ρ] + D1 q1,p
D1 (p⊛W k p ⊗ q) − λ · Ep [ρ] = D1 p⊛W k p ⊗ q1,W
Then using Lemma 2 we can conclude that
gλ
− λ · Ep [ρ] .
I1gλ (p;W) = D1 p⊛W k p ⊗ q1,W
(107)
(108)
Using the definitions of the A-L information and the R-G information given in (85) and (101) together with the Jensen’s
inequality and the concavity of the natural logarithm function we get
Iαλ (p;W) ≥ Iαgλ (p;W)
Iαλ (p;W)
α ∈ (0, 1]
gλ
≤ Iα (p;W)
α ∈ [1, ∞).
(109)
(110)
It is possible to strengthen this relation by expressing the A-L information and the R-G information in terms of one another
as follows.
Lemma 33. Let W be a channel of the form W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , p be an input distribution
in P(X) and λ be a Lagrange multiplier in ℜℓ≥0 .
λ
λ
(a) Let uα,p
∈ P(X) be uα,p
(x ) =
p(x )e(1−α)Dα(W (x )kqα,p )+(α−1)λ·ρ(x )
P
(1−α)Dα(W (x̃ )kqα,p )+(α−1)λ·ρ(x )
x̃ p(x̃ )e
for all x . Then
1
Iαλ (p;W) = Iαgλ (uα,p ;W) + α−1
D1 (pk uα,p )
(
1
D1 (pk u)
supu∈P(X) Iαgλ (u;W) + α−1
=
1
inf u∈P(X) Iαgλ (u;W) + α−1 D1 (pk u)
λ
λ
∈ P(X) be aα,p
(x ) =
(b) Let aα,p
gλ
α(
α,p )
p(x )e
gλ
P
(α−1)Dα(W (x̃ )kqα,p )+(1−α)λ·ρ(x )
x̃ p(x̃ )e
(α−1)D
W (x )kq
(111)
α ∈ (0, 1)
.
α ∈ (1, ∞)
(112)
+(1−α)λ·ρ(x )
for all x . Then
λ
λ
1
D1 aα,p
p
;W − α−1
Iαgλ (p;W) = Iαλ aα,p
(
1
D1 (ak p)
inf a∈P(X) Iαλ (a;W) − α−1
=
1
λ
supa∈P(X) Iα (a;W) − α−1 D1 (ak p)
(113)
α ∈ (0, 1)
.
α ∈ (1, ∞)
(114)
Lemma 33 for λ = 0 is Lemma 18, which was discussed by Poltyrev [19] and Shayevitz [11] before. Since A-L information
is a concept first introduced in [7], for positive λ’s these relations have not been reported before, i.e. Lemma 33 is new.
Definition 18. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and λ ∈ ℜℓ≥0 , the order α
Rényi-Gallager (R-G) capacity for Lagrange multiplier λ is
gλ
Cα,W
, supp∈P(X) Iαgλ (p;W) .
gλ
Using the definition of Iαgλ (p;W) given in (101) we get the following expression for Cα,W
(
1−α
α ∈ ℜ+ \ {1}
supp∈P(X) inf q∈P(Y) Dα p⊛W e α λ·ρ p ⊗ q
gλ
.
Cα,W =
supp∈P(X) inf q∈P(Y) Dα (p⊛W k p ⊗ q) − λ · Ep [ρ] α = 1
(115)
(116)
The R-G capacity satisfies a minimax theorem similar to the one satisfied by The A-L capacity, i.e. Theorem 2. Since both
the statement and proof of the minimax theorems are identical for the order one A-L capacity and the order one R-G capacity,
we dropped the statement of the order one case from Theorem 3.
Theorem 3. For any α ∈ ℜ+ \{1}, channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and Lagrange multiplier λ ∈ ℜℓ≥0
1−α
1−α
supp∈P(X) inf q∈P(Y) Dα p⊛W e α λ·ρ p ⊗ q = inf q∈P(Y) supp∈P(X) Dα p⊛W e α λ·ρ p ⊗ q
(117)
= inf q∈P(Y) supx ∈X Dα (W (x )k q) − λ · ρ(x ).
(118)
gλ
gλ
If the expression on the left hand side of (117) is finite, i.e. if Cα,W
∈ ℜ, then ∃!qα,W
∈ P(Y), called the order α RényiGallager center of W for the Lagrange multiplier λ, satisfying
1−α
gλ
gλ
(119)
Cα,W
= supp∈P(X) Dα p⊛W e α λ·ρ p ⊗ qα,W
gλ
− λ · ρ(x ).
(120)
= supx ∈X Dα W (x )k qα,W
20
gλ
, correFurthermore, for every sequence of input distributions {p (ı) }ı∈Z+ ⊂ P(X) such that limı→∞ Iαgλ p (ı) ;W = Cα,W
gλ
sponding sequence of the order α Rényi-Gallager means {qα,p (ı) }ı∈Z+ is a Cauchy sequence for the total variation metric on
gλ
P(Y) and qα,W
is the unique limit point of that Cauchy sequence.
Proof of Theorem 3 is very similar to the proofs of Theorem 1 and Theorem 2. It relies on Lemma 34, given in the following,
instead of Lemma 19 or Lemma 30.
Lemma 34. For any α ∈ ℜ+ \ {1}, channel W : X → P(Y) with cost function ρ : X → ℜℓ≥0 for a finite input set X, and
λ
λ
Lagrange multiplier λ ∈ ℜℓ≥0 , there exists a pe ∈ P(X) such that Iαλ (e
p ;W) = Cα,W
and ∃!qα,W
∈ P(Y) satisfying
1−α
gλ
gλ
≤ Cα,W
∀p ∈ P(X).
(121)
Dα p⊛W e α λ·ρ p ⊗ qα,W
gλ
gλ
gλ
Furthermore, qα,e
p ∈ P(X) such that Iαgλ (e
p ;W) = Cα,W
.
p = qα,W for all e
The expression on the left hand side of (117) is the R-G capacity where as the expression in (118) is the A-L radius defined
in (98). Thus Theorem 2 and Theorem 3 implies that
gλ
λ
λ
Cα,W
= Sα,W
= Cα,W
∀α ∈ ℜ+
(122)
λ
Furthermore, whenever Cα,W
is finite the unique R-G center described in (120) is equal to the unique A-L center described
in (96) by Theorem 2 and Theorem 3, as well.
gλ
λ
qα,W
= qα,W
λ
∀α ∈ ℜ+ s.t. Cα,W
< ∞.
(123)
In order to avoid using multiple names for the same quantity, we will state our propositions in terms of the A-L capacity and
center in the rest of the paper.
λ
If Cα,W
is finite, then (104), (105), and Theorem 3 for α ∈ ℜ+ \{1} and (107), (108) and Theorem 2 for α = 1 imply that
gλ
λ
λ
Cα,W
− Iαgλ (p;W) ≥ Dα qα,p
qα,W
∀p ∈ P(X).
(124)
Using the same observations, we can prove a van Erven-Harremoës bound for the A-L capacity, as well.
Lemma 35. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and Lagrange multiplier λ ∈ ℜℓ≥0
λ
satisfying Cα,W
<∞
λ
λ
supx ∈X Dα (W (x )k q) − λ · ρ(x ) ≥ Cα,W
+ Dα qα,W
q
∀q ∈ P(Y).
(125)
One can prove
result
using Lemma 13 and Theorem 2. The right most term of the resulting bound is
a similar, but weaker,
λ
λ
q .
q rather than Dα qα,W
Dα∧1 qα,W
λ
Lemma 35 and the continuity of the A-L capacity Cα,W
as a function of λ, established in Lemma 29-(a), imply the continuity
λ
of the A-L center qα,W in λ for the total variation topology on P(Y) via Lemma 2.
Lemma 36. For any α ∈ ℜ+ , channel W : X → P(Y) with a cost function ρ : X → ℜℓ≥0 , and Lagrange multiplier λ0 ∈ ℜℓ≥0
λ0
satisfying Cα,W
< ∞,
λ2
λ1
λ1
λ2
Dα qα,W
qα,W
≤ Cα,W
− Cα,W
∀λ1 , λ2 ∈ ℜℓ≥0 such that λ0 ≤ λ1 ≤ λ2 .
(126)
λ−ǫ1
λ
Furthermore qα,W
is continuous in λ on {λ : ∃ǫ > 0 s.t. Cα,W
< ∞} for the total variation topology on P(Y).
D. Information Measures for Transition Probabilities
We have defined the conditional Rényi divergence, the Augustin information, the A-L information, and the R-G information,
only for input distributions in P(X), i.e. for probability mass functions that are zero in all but finite number of elements of
X. In many practically relevant and analytically interesting models, however, the input set X is an uncountably infinite set
equipped with a σ-algebra X . The Gaussian channels —possibly with multiple input and output antennas and fading— and the
Poisson channels are among the most prominent examples of such models. For such models, it is often desirable to extend the
definitions of the Augustin information and the A-L information from P(X) to P(X ). For instance, in the additive Gaussian
channels described in Examples 4 and 5, the equality Iα (p;W) = Cα,W ,̺ is not satisfied by any probability mass function p
satisfying the cost constraint; but it is satisfied by the zero mean Gaussian distribution with variance ̺.
In the following, we will first show that if Y is a countably generated σ-algebra, then one can generalize the definitions of
the conditional Rényi divergence, the Augustin information, and the A-L information from P(X) to P(X ) provided that W
and Q are not only functions from X to P(Y) but also transition probabilities from (X, X ) to (Y, Y). After that we will show
that if in addition X is countably separated, then the supremum of A-L information Iαλ (p;W) over P(X ) is equal to the A-L
λ
, see Theorem 4. This will imply that the cost constrained Augustin capacity Cα,W ,̺ , which is defined in (82),
radius Sα,W
21
is equal to the supremum of the Augustin information Iα (p;W) over members of P(X ) satisfying Ep [ρ] ≤ ̺, as well, at least
for the cost constraints that are in the interior of the set of all feasible constraints, see Theorem 5.
Let us first recall the definition of transition probability. We adopt the definition provided by Bogachev [21, 10.7.1] with a
minor modification: we use W (E|x ) instead of W (x |E).
Definition 19. Let (X, X ) and (Y, Y) be measurable spaces. Then a function W : Y × X → [0, 1] is called a transition
probability (a stochastic kernel / a Markov kernel) from (X, X ) to (Y, Y) if it satisfies the following two conditions.
(i) For all x ∈ X, the function W (·|x ) : Y → [0, 1] is a probability measure on (Y, Y).
(ii) For all E ∈ Y, the function W (E|·) : X → [0, 1] is a (X , B([0, 1]))-measurable function.
We denote the set of all transition probabilities from (X, X ) to (Y, Y) by P(Y|X ) with the tacit understanding that X and
Y will be clear from the context. If W satisfies (i), then W : X → P(Y) is a channel, i.e. W is a member of P(Y|X), even if
W does not satisfy (ii). Hence P(Y|X ) ⊂ P(Y|X). Inspired by this observation, we denote the probability measure W (·|x )
by W (x ) whenever it is notationally convenient and unambiguous.
In order to extend the definition of the conditional Rényi divergence from P(X) to P(X ), we need to ensure the X −measurability
of Dα (W (x )k Q (x )) on X and replace the sum in (26) with an integral. If (X, τ ) is a topological space and X is the associated
Borel σ-algebra, then one can establish the measurability by first establishing the continuity. Such a continuity result holds if
)
both dWdν(x ) and dQ(x
are continuous in x for ν-almost every y for some probability measure ν for which (W (x ) + Q (x ))≺ν
dν
for all x ∈ X. This is a rather strong hypothesis on W and Q . On the other hand, if W and Q are transition probabilities
from (X, X ) to (Y, Y) for a countably generated Y, then the desired measurability follows from the elementary properties of
the measurable functions and Lemma 9.
Lemma 37. For any α ∈ ℜ+ , countable generated σ-algebra Y of subsets of Y, and W , Q ∈ P(Y|X ) the function
Dα (W (·)k Q (·)) : X → [0, ∞] is X −measurable.
Proof of Lemma 37. There exists {Eı }ı∈Z+ ⊂ Y such that Y = σ({Eı : ı ∈ Z+ }) because Y is countably generated σ−algebra.
Let Yı be
Yı , σ({E1 , . . . , Eı })
Then Y1 ⊂ Y2 ⊂ · · · ⊂ Y, Y =
σ(∪∞
ı=1 Yı ),
ı ∈ Z+ .
and Lemma 9 implies that
Dα (W (x )k Q (x )) = limı→∞ DαYı (W (x )k Q (x ))
∀x ∈ X.
(127)
On the other hand Yı is finite set for all ı ∈ Z+ . Thus for all ı ∈ Z+ there exists a Yı measurable finite partition Eı of Y. Thus
as a result of the definition of the Rényi divergence given in (11) we have
(
P
α
1−α
1
α ∈ ℜ+ \ {1}
E∈Eı (W (E|x )) (Q (E|x ))
Yı
α−1 ln
.
Dα (W (x )k Q (x )) = P
W (E|x )
α=1
E∈Eı W (E|x ) ln Q(E|x )
Then DαYı (W (x )k Q (x )) is a X -measurable function for any ı ∈ Z+ by [21, Thm. 2.1.5-(i-iv)] and [21, Remark 2.1.6] because
W (E|x ) and Q (E|x ) are X -measurable for all E ∈ Eı by the hypothesis of the lemma. Then Dα (W (x )k Q (x )) is X -measurable
as a result of (127), [21, Thm. 2.1.5-(v)], and [21, Remark 2.1.6].
Definition 20. For any α ∈ ℜ+ , countable generated σ-algebra Y of subsets of Y, W ∈ P(Y|X ) and p ∈ P(X ) the order α
conditional Rényi divergence for the input distribution p is
Z
Dα (W k Q | p) ,
Dα (W (x )k Q (x )) p(dx ).
(128)
If ∃q ∈ P(Y) such that Q (x ) = q for p−a.s., then we denote Dα (W k Q | p) by Dα (W k q| p).
Then one can define the Augustin information and the A-L information for all p in P(X ), provided that W is in P(Y|X )
for a countably generated Y and ρ is a X measurable function.
Definition 21. For any α ∈ ℜ+ , countable generated σ-algebra Y of subsets of Y, W ∈ P(Y|X ), and p ∈ P(X ) the order α
Augustin information for the input distribution p is
Iα (p;W) , inf q∈P(Y) Dα (W k q| p) .
(129)
Furthermore, for any X -measurable cost function ρ : X → ℜℓ≥0 and λ ∈ ℜℓ≥0 the order α Augustin-Legendre information for
input distribution p and Lagrange multiplier λ is defined as
Iαλ (p;W) , Iα (p;W) − λ · Ep [ρ]
with the understanding that if λ · Ep [ρ] = ∞, then Iαλ (p;W) = −∞.
22
(130)
Although we have included λ · Ep [ρ] = ∞ case in the formal definition of the A-L information, we will only be interested
in p’s for which Ep [ρ] is finite. We define A∗ to be the set of all such p’s:
A∗ = {p ∈ P(X ) : Ep [ρ] ∈ ℜℓ≥0 }.
(131)
For an arbitrary σ-algebra X the singletons, i.e. sets with only one element, are not necessarily measurable sets; thus P(X)
is not necessarily a subset of A∗ . If X is countably separated, then the singletons are in X by [21, Thm. 6.5.7], P(X) ⊂ A∗
λ
λ
and supp∈A∗ Iαλ (p;W) ≥ Cα,W
. The reverse inequality follows from Theorem 2 and we have supp∈A∗ Iαλ (p;W) = Cα,W
.
Theorem 4 states these observations formally together with the ones about the A-L center through a minimax theorem.
Theorem 4. Let X be a countably separated σ-algebra, Y be a countably generated σ-algebra, W be a transition probability
from (X, X ) to (Y, Y), ρ : X → ℜℓ≥0 be a X -measurable cost function, and α ∈ ℜ+ . Then for all λ ∈ ℜℓ≥0 we have
supp∈A∗ inf q∈P(Y) Dα (W k q| p) − λ · Ep [ρ] = inf q∈P(Y) supp∈A∗ Dα (W k q| p) − λ · Ep [ρ]
= inf q∈P(Y) supx ∈X Dα (W (x )k q) − λ · ρ(x )
(132)
(133)
λ
where A∗ is defined in (131). If the expression on the left hand side of (132) is finite, then ∃!qα,W
∈ P(Y), called the order
α Augustin-Legendre center of W for the Lagrange multiplier λ, satisfying
λ
supp∈A∗ inf q∈P(Y) Dα (W k q| p) − λ · Ep [ρ] = supp∈A∗ Dα W k qα,W
p − λ · Ep [ρ]
(134)
λ
= supx ∈X Dα W (x )k qα,W − λ · ρ(x ).
(135)
Proof of Theorem 4. Since P(X) ⊂ A∗ as a result of max-min inequality we have
supp∈P(X) inf q∈P(Y) Dα (W k q| p) − λ · Ep [ρ] ≤ supp∈A∗ inf q∈P(Y) Dα (W k q| p) − λ · Ep [ρ]
≤ inf q∈P(Y) supp∈A∗ Dα (W k q| p) − λ · Ep [ρ]
= inf q∈P(Y) supx ∈X Dα (W (x )k q) − λ · ρ(x ).
λ
λ
The expression on the left hand side is Cα,W
by (92). Thus (132) and (133), hold when ever Cα,W
is infinite.
λ
λ
If Cα,W is finite, then there exist a unique qα,W
λ
λ
− λ · ρ(x ) = Cα,W
.
supx ∈X Dα W (x )k qα,W
λ
by Theorem 2. Consequently, (132), (133), (134), and (135) hold when Cα,W
is finite.
Let A(̺) be the subset P(X ) composed of the probability measures satisfying the cost constraint ̺,
A(̺) , {p ∈ P(X ) : Ep [ρ] ≤ ̺}.
(136)
Then A(̺) ⊂ A(̺) and supp∈A(̺) Iα (p;W) ≥ Cα,W ,̺ whenever X is countably separated. For the cost constraints in intΓρ
reverse inequality holds as a result of Lemma 29-(c) and Theorem 4 and we have supp∈A(̺) Iα (p;W) = Cα,W ,̺ . Theorem 5
states these observations formally together with the ones about the Augustin center through a minimax theorem.
Theorem 5. Let X be a countably separated σ-algebra, Y be a countably generated σ-algebra, W be a transition probability
from (X, X ) to (Y, Y), ρ : X → ℜℓ≥0 be a X -measurable cost function, and α ∈ ℜ+ . For any ̺ ∈ intΓρ we have
supp∈A(̺) inf q∈P(Y) Dα (W k q| p) = inf q∈P(Y) supp∈A(̺) Dα (W k q| p)
= Cα,W ,̺
(137)
(138)
where Cα,W ,̺ is defined in (82). If Cα,W ,̺ ∈ ℜ≥0 , then ∃!qα,W ,̺ ∈ P(Y), called the order α Augustin center of W for the
cost constraint ̺, satisfying
Cα,W ,̺ = supp∈A(̺) Dα (W k qα,W ,̺ | p)
= supp∈A(̺) Dα (W k qα,W ,̺ | p) .
λ
λ
Furthermore, qα,W ,̺ = qα,W
for all λ ∈ ℜℓ≥0 satisfying Cα,W ,̺ = Cα,W
+ λ · ̺.
Proof of Theorem 5. Since A(̺) ⊂ A(̺) as a result of max-min inequality we have
supp∈A(̺) inf q∈P(Y) Dα (W k q| p) ≤ supp∈A(̺) inf q∈P(Y) Dα (W k q| p)
≤ inf q∈P(Y) supp∈A(̺) Dα (W k q| p) .
23
(139)
(140)
Thus both (137) and (138) hold whenever Cα,W ,̺ = ∞ by (64). On the other hand, as a result of Theorem 4 for any λ with
λ
λ
finite Cα,W
there exists a unique qα,W
satisfying (96). Thus we have,
λ
inf q∈P(Y) supp∈A(̺) Dα (W k q| p) ≤ supp∈A(̺) Dα W k qα,W
p
λ
≤ supp∈A(̺) Dα W k qα,W
p − λ · Ep [ρ] + λ · ̺
λ
≤ Cα,W
+ λ · ̺.
λ
Furthermore, if Cα,W ,̺ ∈ ℜ, then there exists at least one λ ∈ ℜℓ≥0 satisfying Cα,W ,̺ = Cα,W
+ λ · ̺ by Lemma 29-(c).
λ
λ
Then (137) and (138) hold when Cα,W ,̺ ∈ ℜ and (139) holds for qα,W ,̺ = qα,W provided that Cα,W ,̺ = Cα,W
+ λ · ̺.
λ
On the other hand qα,W ,̺ is a probability measure satisfying (140) by Theorem 1 and qα,W ,̺ = qα,W for all λ satisfying
λ
Cα,W ,̺ = Cα,W
+ λ · ̺ by Lemma 31.
The countable separability of X and countable generatedness of Y are fairly mild assumptions satisfied by most transition
probabilities considered in practice. Hence, Theorems 4 and 5 provide further justification for studying the relatively simple
case of probability mass functions, first.
The existence of an input distribution p satisfying both Ep [ρ] ≤ ̺ and Iα (p;W) = Cα,W ,̺ is immaterial to the existence
λ
λ
of a unique qα,W ,̺ or its characterization through qα,W
for λ’s satisfying Cα,W ,̺ = Cα,W
+ λ · ̺ by Lemma 29-(c,d) and
Theorem 5. Although one can prove the existence of such a p for certain special cases such an input distribution does not exist
in general. Thus, we believe, it is preferable to separate the issue of the existence of an optimal input distribution from the
λ
λ
discussion of Cα,W ,̺ and qα,W ,̺ and their characterization via Cα,W
and qα,W
. That, however, is not the standard practice,
[35, Thm. 1].
We have assumed Y to be countably generated in order to ensure that the conditional Rényi divergence used in (129)
is well-defined. In order define the Rényi information, however, we do not need to assume Y to be countably generated;
the transition probability structure is sufficient. Recall that if W ∈ P(Y|X ), then for any p ∈ P(X ) there exists a unique
probability measure probability measure p⊛W on (X × Y, X ⊗ Y) such that
Z
W (Ey |x )p(dx ).
∀Ex ∈ X , Ey ∈ Y
p⊛W (Ex × Ey ) =
Ex
by [21, Thm. 10.7.2.]. Thus Iα (p;W) is well defined for any W ∈ P(Y|X ) and p ∈ P(X ).
Unfortunately, the situation is not nearly as simple for the R-G information. In order define the R-G information using a
1−α
similar approach one first shows that W e α λ·ρ a transition kernel —rather than a transition probability (i.e. Markov kernel)—
1−α
and proceeds with establishing the existence a unique measure p⊛W e α λ·ρ for all p in P(Y). For orders greater than one,
resulting measure is a sub-probability measure and one can use (101) as the definition of the R-G information. For orders
1−α
between zero and one, on the other hand, p⊛W e α λ·ρ is a σ-finite measure for all p’s in P(X ), but it is not necessarily
a finite measure for all p’s in P(X ). Thus for orders between zero and one, one can use (101) as the definition of the R-G
information, only after extending the definition of the Rényi divergence to σ-finite measures.
g
VI. E XAMPLES
In this section, we will first demonstrate certain subtleties that we have pointed out in the earlier sections. After that we
will study various Gaussian channels and obtain closed form expressions for the Augustin capacity and center.
A. Shift Invariant Families
Example 1 (A Channel with an Affine Capacity). Let the channel W : ℜ≥0 → P(B([0, 1))) and the associated cost function
ρ : ℜ≥0 → ℜ≥0 be
dW (x )
dν
= f⌊x ⌋ (y − x − ⌊y − x ⌋),
ρ(x ) = ⌊x ⌋
where ν is the Lebesgue measure on [0, 1) and fı ’s are given by
fı (y) = eı+1 1{y∈[0,e−ı−1 )}
∀ı ∈ Z≥0 .
Let uı be uniform distribution on [ı, ı + 1); then one can confirm by substitution that Tα,uı (u0 ) = u0 . Then using the Jensen’s
inequality together with the fixed point property we get22
Dα (W k q| uı ) ≥ Dα (W k u0 | uı ) + Dα∧1 (u0 k q) .
22 See
the derivations for (37) and (39) of Lemma 13-(c,d) given in Appendix B.
24
Thus u0 is the unique order α Augustin mean and Iα (uı ;W) = Dα (W k u0 | uı ), i.e. Iα (uı ;W) = ı + 1 and qα,uı = u0 , for all
ı ∈ Z+ and α ∈ ℜ+ . Then using Euı [ρ] = ı, we can conclude that Cα,W ,̺ ≥ (̺ + 1) because Cα,W ,̺ is concave in ̺ by
Lemma 27-(a). One the other hand one can confirm by substitution that
Dα (W k u0 | p) = Ep [ρ] + 1.
(141)
Thus Iα (p;W) ≤ (̺ + 1) for any p satisfying the cost constraint ̺. Hence,
Cα,W ,̺ = ̺ + 1,
qα,W ,̺ = u0 .
Then as a result of (90) we have
λ
Cα,W
=
(
∞ λ ∈ [0, 1)
.
1 λ ∈ [1, ∞)
λ
Then using (141) and Theorem 4 we can conclude that qα,W
= u0 for all λ ∈ [ln 2, ∞).
Example 2 (A Channel with a Non-Upper Semicontinuous Capacity). Let the channel W : ℜ → P(B([0, 1))) and the
associated cost function ρ : ℜ → ℜ≥0 be
dW (x )
dν
= f⌊x ⌋ (y − x − ⌊y − x ⌋)
(
⌊x ⌋ x ≥ 0
ρ(x ) =
.
2⌊x ⌋ x < 0
where ν is the Lebesgue measure on [0, 1) and fı :∈ [0, 1) → ℜ≥0 are given by
ı+1
2 1{y∈[0,2−ı−1 )} ı > 0
fı (y) = 3/21{y∈[0,2/3)}
ı=0.
21{y∈[0,1/2)}
ı<0
Following an analysis similar to the one described above we can conclude that
(
(̺ + 1) ln 2 ̺ > 0
,
Cα,W ,̺ =
̺=0
ln 32
(
∞ λ ∈ [0, ln 2)
λ
Cα,W
=
.
ln 2 λ ∈ [ln 2, ∞)
λ
Hence Cα,W ,̺ 6= inf λ≥0 Cα,W
+ λ · ̺ for ̺ = 0.
Example 3 (A Product Channel). Let W1 and W2 be the channels describe in Examples 1 and 2 and ρ1 and ρ2 be the
associated cost functions. Let W be the product of these two channels with the additive cost function ̺, i.e.
W (x1 , x2 ) = W1 (x1 ) ⊗ W2 (x2 ),
ρ(x1 , x2 ) = ρ1 (x1 ) + ρ2 (x2 ).
Then as a result of Lemma 28
Cα,W ,̺ =
(
̺ + 1 + ln 2
1 + ln 32
̺>0
.
̺=0
Note that for positive values of ̺ there does not exist any (̺1 , ̺2 ) pair satisfying both Cα,W ,̺ = Cα,W1 ,̺1 + Cα,W2 ,̺2 and
the cost constraint ̺1 + ̺2 ≤ ̺ at the same time.
B. Gaussian Channels
In the following, we denote the zero mean Gaussian probability measure on B(ℜ) with variance σ 2 by ϕσ2 . With a slight
abuse of notation, we denote corresponding probability density function by the same symbol:
ϕσ2 (x ) =
x2
√ 1 e− 2σ2
2πσ
∀x ∈ ℜ.
(142)
We use the Gaussian channels and the corresponding transition probabilities interchangeably. These channels and the corresponding transition probabilities have the same capacity and center expressions by Theorems 4 and 5.
25
Example 4 (Scalar Gaussian Channel). Let W be the scalar Gaussian channel with noise variance σ 2 and the associated cost
function ρ be the quadratic one, i.e.
Z
W (E|x ) =
ϕσ2 (y − x )dy
∀E ∈ B(ℜ),
E
ρ(x ) = x 2
∀x ∈ ℜ.
In the following we will first calculate the Augustin information for the zero mean Gaussian distribution with variance ̺, i.e.
for ϕ̺ . Then we demonstrate that this value is the greatest among all input distributions satisfying the cost constraint ̺. This
will imply Cα,W ,̺ = Iα (ϕ̺ ;W) and qα,W ,̺ = qα,ϕ̺ . Resulting expressions are given in (143) and (144) in terms of the θα,σ,̺
defined in (145). Then we derive the closed form expressions for the A-L capacity and center given in (150) and (151) —the
λ
latter in terms of the θα,σ
defined in (152). After that we show thatzero mean
Gaussian distributions are optimal for the A-L
λ
λ
= qα,ϕ̺λ for ̺λα,σ defined in (153).
capacity as well; in particular, we will demonstrate that Cα,W
= Iαλ ϕ̺λα,σ ;W and qα,W
α,σ
α
(1−α)
σ
α̺
1
√ θα,σ,̺
α ∈ ℜ+ \ {1}
2
2 ) + α−1 ln
2
αθα,σ,̺
+(1−α)σ2
Cα,W ,̺ = 2(αθα,σ,̺ +(1−α)σ
,
(143)
1 ln 1 + ̺
α=1
σ2
2
qα,W ,̺ = ϕθα,σ,̺ ,
̺
2
−
Dα (W (x )k ϕθ2 ) =
2
θα,σ,̺
, σ2 +
One can confirm by substitution that
ϕ
2
σ
2α
q
+ ( ̺2 −
(144)
σ 2
2α )
αx 2
2(αθ 2 +(1−α)σ2 )
σ
2
2
+x −θ
2θ 2
2
2
+
+ ̺σ 2 .
1
α−1
(145)
θ α σ(1−α)
αθ 2 +(1−α)σ2
ln √
α ∈ ℜ+ \ {1}
+ ln σθ
α=1
.
(146)
Then the order α tilted channel Wα θ2 defined in (27) is a Gaussian channel as well:
Z
ϕ
αθ 2
Wα θ2 (E|x ) =
ϕ
y − αθ2 +(1−α)σ
dy.
σ 2 θ2
2x
E
αθ2 +(1−α)σ2
Then Tα,p (q) is a zero mean Gaussian probability measures whenever both p and q are so. In particular,
Tα,ϕ̺ (ϕθ2 ) = ϕ
σ2 θ 2
αθ 2
( αθ 2 +(1−α)σ2 )2 ̺+ αθ 2 +(1−α)σ2
.
(147)
Consequently, if ϕθ2 is a fixed point of Tα,ϕ̺ (·), then θ satisfies the following equality
θ2 θ4 − θ2 ̺ + 2 − α1 σ 2 + 1 − α1 σ 4 = 0.
The variance θα,σ,̺ given in (145) is the only root of the above equation that is greater than σ 2 . Furthermore, using (147) one
2
2
2
can confirm that Tα,ϕ̺ (ϕθα,σ,̺
) = ϕθα,σ,̺
, i.e. ϕθα,σ,̺
is a fixed point of Tα,ϕ̺ (·). Then using the Jensen’s inequality together
with the fixed point property we get23
2
2
q
∀q ∈ P(B(ℜ)).
ϕ̺ + D1∧α ϕθα,σ,̺
Dα (W k q| ϕ̺ ) ≥ Dα W k ϕθα,σ,̺
2
2
2
ϕ̺ .
Thus ϕθα,σ,̺
is the order α Augustin mean for the input distribution ϕ̺ , i.e. qα,ϕ̺ = ϕθα,σ,̺
and Iα (ϕ̺ ;W) = Dα W k ϕθα,σ,̺
On the other hand as a result of (146) for any p ∈ P(B(ℜ)) we have
α(E [ρ]−̺)
2
p = 2(αθα,σ,̺p+(1−α)σ2 ) + Iα (ϕ̺ ;W)
Dα W k ϕθα,σ,̺
∀p ∈ P(B(ℜ)).
(148)
Then Iα (p;W) ≤ Iα (ϕ̺ ;W) for all p satisfying Ep [ρ] ≤ ̺, i.e. the cost constraint. Consequently, Cα,W ,̺ = Iα (ϕ̺ ;W) and
2
qα,W ,̺ = qα,ϕ̺ . The expression given in (144) follows from qα,ϕ̺ = ϕθα,σ,̺
. The expression given in (143) is confirmed by
substitution using (146).
The bound given in (148) can also be used to bound the A-L information. Let λα,σ,̺ be
λα,σ,̺ ,
α̺+σ2 +
α
.
(α̺−σ2 )2 +4̺α2 σ2
√
(149)
Note that λα,σ,̺ is a continuous monotonically decreasing function of ̺ such that lim̺↑∞ λα,σ,̺ = 0 and lim̺↓0 λα,σ,̺ = 2σα2 .
λ
λ
Furthermore, Iα α,σ,̺ (p;W) ≤ Iα α,σ,̺ (ϕ̺ ;W) for all p on B(ℜ) as a result of (148) and the definition of the A-L information.
λα,σ,̺
λα,σ,̺
λα,σ,̺
2
ϕ̺ and
= qα,ϕ̺ by Theorem 4. Using Iα (ϕ̺ ;W) = Dα W k ϕθα,σ,̺
(ϕ̺ ;W) and qα,W
Consequently, Cα,W = Iα
23 See
the derivations for (37) and (39) of Lemma 13-(c,d) given in Appendix B.
26
2
qα,ϕ̺ = ϕθα,σ,̺
, one can confirm the following closed form expressions for the A-L capacity and center for λ ∈ (0, 2σα2 ]. The
closed form expressions for λ ∈ ( 2σα2 , ∞) follows from (146).
q
q
1
2σ2 λ
α
α−1 2σ2 λ
ln
+
−
ln
λ ∈ (0, 2σα2 ), α ∈ ℜ+ \ {1}
α−1
α
α
α
α
√
λ
(150)
Cα,W = σ 2 λ − ln 2eσ 2 λ
λ ∈ (0, 2σα2 ), α = 1
0
λ ∈ [ α , ∞)
2σ2
λ
qα,W
= ϕθα,σ
λ
λ
θα,σ
, σ2 +
(151)
1
2λ
−
2
σ
α
+
(152)
λ
λ
α,σ,̺
α,σ,̺
, as expected.
+ λα,σ,̺ ̺ and qα,W ,̺ = qα,W
Note that Cα,W ,̺ = Cα,W
λ
On the other hand, the expression for Cα,W
is a continuously differentiable function of λ such that
λ
d
dλ Cα,W
λ
Cα,W
The value of λ minimizing
̺λα,σ for which λ is optimal.
2
α−2σ λ
= − 2λ(α+(α−1)2σ
2 λ) 1{λ≤
d
λ
dλ Cα,W
+ λ̺ is the one satisfying
λ
Then Cα,W
}.
= −̺. Then for each λ we can define a cost constraint
|α−2σ2 λ|
̺λα,σ ,
α
2σ2
+
(153)
2λ(α+2σ2 λ(α−1)) .
λ
+ λ̺λα,σ = Cα,W ,̺λα,σ by construction and Cα,W ,̺λα,σ = Iα ϕ̺λα,σ ;W by (143).
= Iαλ ϕ̺λα,σ ;W because Cα,W
Example 5 (Parallel Gaussian Channels). Let W be the product of scalar Gaussian channels Wı with noise variance σı for
ı ∈ {1, . . . , n} and the ρ be the additive cost function, i.e.
Z hY
i
n
ϕσı2 (yı − xı ) dy
∀E ∈ B(ℜn ),
W (E|x ) =
ı=1
E
Xn
∀x ∈ ℜn .
xı2
ρ(x ) = x T x =
ı=1
As a result of Lemma 28, the cost constrained Augustin capacity of W can be calculated using the following optimization
Cα,W ,̺ = sup̺1 ,...,̺n :Pı ̺ı ≤̺ Cα,Wı ,̺ı .
One can determine the optimal allocation of cost among different component channels using the above optimization and the
closed form expressions given in (143). However, the more insightful and easier way to determine the cost constrained Augustin
capacity of the parallel Gaussian channels is through the A-L capacity. Note that the A-L capacity and center of W can be
written in terms of the corresponding quantities for the component channel using Lemma 32:
Xn
λ
Cλ ,
Cα,W
=
ı=1 α,Wı
Y⊗
λ
λ
.
qα,W
=
qα,W
ı
ı
λ
λ
λ
Note that Cα,W
’s are continuously differentiable in λ by (150). Then Cα,W
is continuously differentiable and Cα,W
+ λ̺ is
ı
d
λ
minimized at λα,W ,̺ satisfying dλ Cα,W
=
−̺,
i.e.
λ
satisfying
α,W ,̺
λ=λα,W ,̺
Xn
=̺
̺λα,σı
ı=1
λ=λα,W ,̺
Furthermore, the above mentioned minimum is equal to the cost constrained Augustin capacity by Lemma 29-(c), i.e.
λ
α,W ,̺
+ λα,W ,̺ ̺.
Cα,W ,̺ = Cα,W
λ
. Thus
Recall that we have already established in the previous example that Iαλ ϕ̺λα,σı ;Wı = Cα,W
ı
Xn
λ
.
Iαλ ϕ̺λα,σ ;Wı = Cα,W
ı=1
ı
Then we can write the cost constrained Augustin capacity and center in the following parametric form in terms of λα,W ,̺
Xκ
(154)
Cα,Wı ,̺α,ı
Cα,W ,̺ =
Yκı=1
(155)
ϕθα,ı
qα,W ,̺ =
ı=1
σı2
α
+
θα,ı =
σı2
̺α,ı =
+
α−2σı2 λα,W ,̺
2λα,W ,̺ (α+2σı2 λα,W ,̺ (α−1))
1
2λα,W ,̺
+
−
|
|
27
(156)
(157)
The value of λα,W ,̺ is determined using
P
ı
̺α,ı = ̺. Note that each ̺α,ı is nonincreasing λα,W ,̺ .
The cost constrained Augustin capacity and center vector Gaussian channel with multiple input and output antennas can be
analyzed with a similar approach with the help of singular value decomposition.
VII. D ISCUSSION
Similar to the Rényi information, the Augustin information is a generalization of the mutual information relying on the
Rényi divergence. Unlike the order α Rényi information, however, the the order α Augustin information does not have a
closed form expression, except for the order one case. This makes it harder to prove certain properties of the Augustin
information such as its continuous differentiability as a function of the order α, the existence of a unique order α Augustin
mean qα,p or the bounds given in (10). However, once the essentials about the Augustin information is established, the analysis
of the Augustin capacity is rather straightforward and very similar to the analogous analysis for the Rényi capacity, presented
in [14].
The convex conjugation techniques have been applied to the calculation of the cost constrained Augustin capacity through
the quantity Iαgλ (p;W), which we have called the R-G information. Although such an approach can successfully characterize
the cost constrained Augustin capacity via the R-G capacity; it is non-standard and somewhat convoluted. A more standard
approach, based on the concept of A-L information Iαλ (p;W), is presented in V-B. Although the A-L information has not been
used before, the resulting capacity is identical to the one associated with the R-G information. In this light the optimality of
the approach based on the R-G information seems more intuitive.
Our analysis of the Augustin information and capacity was primarily motivated by their operational significance in the
channel coding problem, [2], [7]. We investigate that operational significance more closely and derive a sphere packing bound
with a polynomial prefactor in [8].
A PPENDIX
A. Proof of Lemmas on the Analyticity of the Rényi Divergence
Proof of Lemma 11. Let g(α) and f (α, y) be
g(α) ,
f (α, y) ,
Z
dw α
dν
dq
dν
1−α
ν(dy ),
(158)
1−α
dw α dq
dν
(159)
dν
where ν is any reference measure satisfying w ≺ν and q≺ν. Note that
Dα (w k q) =
1
α−1
α ∈ ℜ+ \ {1}.
ln g(α)
Furthermore g(α) does not depend on the choice of ν, but f (α, y) does.
κ
dq
∂κ
dw
∂ακ f (α, y) = f (α, y) ln dν − ln dν
∀κ ∈ Z≥0 .
(160)
(161)
Then using the inequality z ln z ≥ −1/e we get
∂κ
∂ακ f
dq
κ
κ κ
∀κ ∈ Z≥0 , φ ∈ (α, ∞).
(162)
) 1{ dw ≤ dq } dν
+ f (φ, y)( (φ−α)e
)κ 1{ dw > dq }
(α, y) ≤ ( αe
dν
dν
dν
dν
R
On the other hand |f (φ, y)|ν(dy ) = e(φ−1)Dφ(wkq) and for all α ∈ (0, χw,q ) there exists a φ ∈ (α, χw,q ) with finite Dφ (w k q).
Then as a result of [21, Corollary 2.8.7-(ii)], g(α) is an infinitely differentiable function of α on (0, χw,q ) such that
Z
∂κ
∂κ
g(α)
=
∀κ ∈ Z≥0 .
(163)
∂ακ
∂ακ f (α, y) ν(dy )
Consequently, if χw,q > 1, then
D1 (w k q) =
∂
∂α
ln g(α)
α=1
.
(164)
Using (162) and (163) we get
κ κ
−( αe
) ≤
∂κ
∂ακ g(α)
κ
)κ
≤ g(φ)( (φ−α)e
∀κ ∈ Z≥0 , φ ∈ (α, χw,q ).
√
√
Then using the Stirling’s approximation for the factorial function, i.e. 2πκ(κ/e)κ ≤ κ! ≤ e κ(κ/e)κ , we get
∂κ
∂ακ g(α)
κ
κ κ
) ∨ g(φ)( (φ−α)e
)κ
≤ ( αe
g(φ)
1
κ!
≤ √2πκ
ακ ∨ (φ−α)κ
∀κ ∈ Z≥0 , φ ∈ (α, χw,q ),
∀κ ∈ Z+ , φ ∈ (α, χw,q ).
28
(165)
Thus g(α) is not only infinitely differentiable but also analytic in α on (0, χw,q ) by [36, Proposition 1.2.12]. On the other
hand g(α) ∈ ℜ+ for all α ∈ (0, χw,q ) because g(α) = e(α−1)Dα(wkq) by (160) and Dα (w k q) ∈ ℜ≥0 by Lemmas 2, 8 and
the definition of χw,q . Thus ln g(α) is analytic in α on (0, χw,q ) because composition of analytic functions is analytic by [36,
Proposition 1.4.2]. Then Dα (w k q) is analytic in α on (0, χw,q ) \ {1} because the quotient of analytic functions is analytic at
points with open neighborhoods on which the function in the denominator is non-zero by [36, Proposition 1.1.12].
Now we proceed with establishing the analyticity of Dα (w k q) at α = 1 for χw,q > 1 case. Since ln g(α) is analytic in α
on (0, χw,q ) we can write ln g(α) as a convergent power series around any point in (0, χw,q ) for some neighborhood. Thus,
there exists a δ > 0 for which the following two identities hold for all η ∈ (1 − δ, 1 + δ)
X∞ |η−1|ı ı
∂
ı!
∂αı ln g(α) α=1 < ∞,
Xı=0
∞ (η−1)ı
∂ı
ı!
∂αı ln g(α) α=1 = ln g(η).
ı=0
Then using ln g(1) = 0 together with (160) and (164) we get
X∞ (η−1)ı−1 ı
∂
Dη (w k q) = D1 (w k q) +
ı!
∂αı ln g(α)
ı=2
α=1
∀η ∈ (1 − δ, 1 + δ).
(166)
Then Dη (w k q) is analytic on (1 − δ, 1 + δ) by [36, Corollary 1.2.4] because it is equal to a function defined by a convergent
power series.
The convergent power series given in (166) determines the derivatives of Dα (w k q) at α = 1 by [36, Corollary 1.1.16]:
∂κ
∂ακ Dα (w k q) α=1
=
1
∂ κ+1
κ+1 ∂ακ+1
ln g(α)
α=1
κ ∈ Z+ .
(167)
Using (160) together with the elementary rules of differentiation we can express the derivatives of Dα (w k q) in terms of the
derivatives of ln g(α) for other orders in (0, χw,q ), as well
Xκ
∂κ
κ!
1
∂t
∂ κ−t
D
(w
k
q)
=
ln
g(α)
∂ακ α
∂ακ−t α−1 α=φ
∂αt
α=φ
t=0 t!(κ−t)!
α=φ
Xκ
κ−t
t
∂
κ! (−1)
κ ∈ Z+ , φ ∈ (0, χw,q ) \ {1}.
(168)
=
t! (φ−1)κ−t +1 ∂αt ln g(α)
t=0
α=φ
On the other hand by Faà di Bruno formula for derivatives of the composition of smooth functions [36, Thm. 1.3.2] we have
1
2
t
X
1 ∂1
∂t
1 ∂2
∂ 1 +2 +···+t
t!
1 ∂t
ln
τ
ln
g(α)
=
·
·
·
1 +2 +···+t
1 g(α)
2 g(α)
t g(α)
∂αt
!
!...
!
1!
∂α
2!
∂α
t!
∂α
∂τ
t
Jt 1 2
τ =g(α)
1
2
t
X
(−1)(
+
+···+
−1)!
1 ∂1
∂t
1 ∂2
t!
1
2
t
· · · t!1 ∂α
.
=
t g(α)
1 !2 !...t ! (−g(α))1 +2 +···+t
1! ∂α1 g(α)
2! ∂α2 g(α)
Jt
Then using (158), (159), (161), and (163) we get
t t
X (−1)( + +···+ −1)! (−1)
1
(−1)
dq
dw
dw
dw
1
2
t
∂t
q
q
ln dν − ln dν
.
···
∂αt ln g(α) = t !
1 !2 !...t !
1! Ewα ln dν − ln dν
t! Ewα
Jt
(169)
The expression given in (20) for κth derivative of Dα (w k q)with respect to α follows from (167), (168), (169).
In order to prove the analyticity of D1 (wαq k w ) and D1 (wαq k q), first note that as a result of (21), which follows from (20),
we have
∂
D1 wφq w = (φ − 1)2 ∂α
Dα (w k q) α=φ
∀φ ∈ (0, χw,q ).
(170)
∂
q
Since
Dα (wk q) is analytic in α on (0, χw,q ), so is ∂α Dα (w k q). Hence, D1 (wα k w ) is analytic in α on (0, χw,q ). Since
q
D1 wφ w is analytic in φ on (0, χw,q ) it is finite on (0, χw,q ). Thus (18) holds for all α in (0, χw,q ) and (170) implies
∂
Dα (w k q) α=φ
∀φ ∈ (0, χw,q ).
D1 wφq q = Dα (w k q) − φ(1 − φ) ∂α
Thus D1 (wαq k q) is an analytic function of α on (0, χw,q ), as well.
Proof of Lemma 12. Using the inequality z ln z ≥ −1/e we get
dq
α dq 1−α
ln dw
( dw
dν ) ( dν )
dν − ln dν
ı
dq
β dq 1−β
ı ı
ı
≤ ( αe
) 1{ dw ≤ dq } dν
+ ( (β−α)e
)ı 1{ dw > dq } ( dw
∀ı ∈ Z≥0 .
dν ) ( dν )
dν
dν
dν
dν
√
√
Using the Stirling’s approximation for the factorial function, i.e. 2πκ(κ/e)κ ≤ κ! ≤ e κ(κ/e)κ , we get
h
ıi
α(wkq ) ı!
dq
1
e(1−α)D
e(β−1)Dβ(wkq )
√
≤
∀ı ∈ Z+ .
Ewαq ln dw
−
ln
+
dν
dν
αı
(β−α)ı
2πı
29
Then using Dβ (w k q) ≤ γ together with Lemma 8 we get
h
ıi
dq
e(1∨β)γ
≤ ı! (α∧(β−α))
−
ln
Ewαq ln dw
ı
dν
dν
∀ı ∈ Z+ .
Then using (20) we get,
κ
∂ Dα(wkq)
∂ακ
α=φ
P
On the other hand Jκ
which implies (23).
≤
P
κ
κ!
t=0
1
1
|φ−1|κ−t +1 (α∧(β−φ))t
1
κ! (1∧(β−1))
κ+1
P
(1 +2 +···+κ )! 1 +2 +···+κ
ξ
1 !2 !...κ !
κ
∂ Dα(wkq)
∂ακ
α=φ
Jκ+1
P
Jt
1 +2 +···+t
e(1∨β)γ
(1∨β)γ 1 +2 +···+κ+1
(1 +2 +···+t −1)!
1 !2 !...t !
(1 +2 +···+κ+1 −1)!
1 !2 !...κ+1 !
e
φ 6= 1
.
φ=1
= ξ(1 + ξ)κ−1 by [36, Thm. 1.4.1]. Thus we get the following inequality,
κ
t
P
1+e(1∨β)γ
1
κ!
κ−t +1
φ∧(β−φ)
|φ−1|
t=0
≤
κ+1
(1∨β)γ
κ! 1+e
1∧(β−1)
φ 6= 1
.
φ=1
As a result of [36, Corollaries 1.2.4 and 1.2.5] the following equality holds on the open interval in which the power series on
the right hand side is convergent,
X∞ (η−φ)
∂
Dη (w k q) =
(171)
!
∂α Dα (w k q) α=φ .
=0
Note that as a result of (23) we have
lim supκ→∞
r
κ
1
∂κ
κ! ∂ακ Dα (w k q) α=φ
≤ τ.
Thus radius of convergence of the power series on the right hand side of (171) is at least
Hadamard formula. Thus for all η ∈ (φ − τ1 , φ + τ1 ) using (23) and (171) we get
Dη (w k q) −
Using identities
P∞
ı=0
ξı =
1
1−ξ
Dη (w k q) −
Xκ−1
and
Xκ−1
(η−φ)ı ∂ ı Dα(wkq)
ı!
∂αı
ı=0
α=φ
P∞
ı=0 (ı
+ 1)ξ ı =
1
(1−ξ)2
(η−φ)ı ∂ ı Dα(wkq)
ı!
∂αı
ı=0
α=φ
≤
X∞
ı=κ
1
τ
by [36, Lemma 1.1.8], i.e. by
|η − φ|ı τ ı+1 1 + ı1{φ6=1} .
for |ξ| < 1 we get,
≤
τ κ+1 |η−φ|κ
1−|η−φ|τ
h
1+ κ−1+
1
1−|η−φ|τ
i
1{φ6=1} .
B. Proofs of Lemmas on the Augustin Information
Proof of Lemma 13.
(13-a) Iα (p;W) ≤ Dα (W k q| p) for all q ∈ P(Y) by definition. On the other hand, P
Dα (W (x )k q1,p ) ≤ − ln p(x ) for all x with
positive p(x ) by Lemma 1 because p(x )W (x ) ≤ q1,p . Hence, Iα (p;W) ≤ − x p(x ) ln p(x ).
(13-b) Note that as a result of Lemma 2 and (29),
2
D1 (W k q| p) = D1 (W k q1,p | p) + 21 kq1,p − qk
∀q ∈ P(Y).
Then q1,p is the unique probability measure satisfying I1 (p;W) = D1 (W k q1,p | p). Equation (34) follows immediately.
(13-c) Let S and ς be
ς , minx :p(x )>0 p(x ),
o
n
1−α
g
+
S , s ∈ M (X) : ς1{p(x )>0} ≤ s(x ) ≤ e ς Dα(W kqα,p |p ) 1{p(x )>0} ∀x ∈ X .
The following seven statements collectively imply part (c).
(c-i) If q1,p ≺u and Tα,p (u) = u, then Dα (W k u| p) = Iα (p;W), (35) and (37) hold for qα,p = u, and qα,p is unique:
Note that Tα,p (u) = u and q1,p ≺u imply
i1
α
hX
dW (x )
(1−α)Dα(W (x )ku) α
du
∀ν : q1,p ≺ν.
=
e
p(x
)
dν
dν
x
Then one can confirm by substitution that
X
1
Dα (uk q) = α−1
ln
p(x )e(α−1)(Dα(W (x )kq)−Dα(W (x )ku))
x
30
∀q ∈ P(Y).
Then Jensen’s inequality and convexity of the exponential function imply
Dα (uk q) ≤ Dα (W k q| p) − Dα (W k u| p)
∀q ∈ P(Y).
(172)
Then u is the unique probability measure satisfying Iα (p;W) = Dα (W k u| p) by Lemma 2. Consequently (35) and
the lower bound given in (37) hold.
In order to establish the upper bound given in (37) for q ∈ Qα,p note that
Z
X
1
∼ 1−α
p(x ) ln ( dq
Wαu (x )(dy)
∀q ∈ Qα,p
(173)
Dα (W k q| p) − Dα (W k u| p) = α−1
du )
x
where q∼ is the component of q that is absolutely continuous in u.
On the other hand using the Jensen’s inequality and concavity of the natural logarithm function we get
Z
Z
X
X
dq∼ 1−α
u
1
1
∼ 1−α
p(x ) ln ( du )
p(x ) ln( dq
Wα (x )(dy) ≤ α−1
Wαu (x )(dy)
α−1
du )
x
x
Z
du
)Tα,p (u)(dy).
= ln( dq
∼
(174)
Since Tα,p (u) = u by hypothesis using (173) and (174) we get
D1 (uk q) ≥ Dα (W k q| p) − Dα (W k u| p)
∀q ∈ Qα,p .
In order to establish the upper bound given in (37) for q ∈
/ Qα,p we need to make the following additional
observation. If q ∈
/ Qα,p , then there exists an x for which p(x ) > 0 and W (x ) ⊥ q. As a result there exists
an event E ∈ Y such that such that u(E) > 0 and q(E) = 0 because W (x )≺q1,p and q1,p ≺u. Consequently
D1 (uk q) = ∞ and the upper bound
/ Qα,p , as well.
in equation (37) holds for q ∈
(c-ii) Dα (W k q| p)−Dα W k Tα,p (q) p ≥ D1 Tα,p (q) q for all q ∈ Qα,p : Note that Tα,p (q)≺q for all q ∈ Qα,p by
definition. Then
Z
X
dTα,p (q) 1−α
1
Dα (W k q| p) − Dα W k Tα,p (q) p = 1−α
Wαq (x )(dy)
p(x ) ln
dq
x
Z
X
dTα,p (q) 1−α
1
p(x ) ln
≥ 1−α
Wαq (x )(dy)
dq
x
(175)
= D1 Tα,p (q) q .
The inequality follows from the Jensen’s inequality and the concavity of the natural logarithm function.
+
g
(c-iii) {Tıα,p (qα,p
)}ı∈Z+ is totally bounded for total variation metric on M (Y): For any q ∈ Qα,p , as a result of
definitions of Tα,p (·) and µα,p we have
dTα,p (q)
dν
=(
dµα,s α dq 1−α
dν ) ( dν )
g
where s(x ) = p(x )e(1−α)Dα(W (x )kq) . Furthermore, if Dα (W k q| p) ≤ Dα W k qα,p
p , then s ∈ S.
g
In addition qα,p
is equal to µα,s for an s ∈ S. In particular
g
qα,p
= µα,s0
where s0 = kµα,p k p. One can confirm by substitution that kµα,p k
= e(1−α)Dα(p⊛W kp⊗qα,p ) . Furthermore,
g
g
g
p by the Jensen’s inequality
Dα p⊛W k p ⊗ qα,p ≥ 0 by Lemma 2 and Dα p⊛W k p ⊗ qα,p ≤ Dα W k qα,p
and the concavity of the natural logarithmfunction. Thus s0 ∈ S.
g
g
g
On the other hand, Dα W k Tıα,p (qα,p
) p ≤ Dα W k qα,p
p for all ı ≥ Z+ . Thus we can write Tıα,p (qα,p
) in
terms of the elements of µα,S as follows:
Y⊗
g
ı−
dTıα,p (qα,p
)
dµα,s
dµα,s0 (1−α)ı
( dν )α(1−α)
= ( dν
)
dν
−α
−α
g
∈{1,...,ı}
g
(1−α)Dα(W (x )kT−1
α,p (qα,p ))
where s (x ) = p(x )e
.
g
In order to prove that {Tıα,p (qα,p
)}ı∈Z+ is totally bounded, we prove that a superset of it, i.e. B defined in the
following, is totally bounded.
Y⊗
+
ı−
dµα,s
dµα,s0 (1−α)ı
db
Bı , b ∈ M (Y) : dν
( dν )α(1−α)
= ( dν
)
for some s ∈ S
(176)
∈{1,...,ı}
B , ∪ı∈Z+ Bı .
(177)
31
Let us denote the number of x ’s with p(x ) > 0 by κ. Then S is isometric to a cube in24 ℜκ. We divide each side of
the cube into 2n equal length intervals. Thus S is composed of 2κn sub-cubes. Furthermore, µα,s ≤ µα,es if s ≤ e
s
by definition. Thus, for any s ∈ S we have
α1
1−α D W kq g
α(
α,p |p ) −ς
ς
µα,⌊s⌋n
µα,⌊s⌋n ≤ µα,s ≤ 1 + e
ς2n
where ⌊s⌋n is the corner of the sub-cube s is in that satisfies ⌊s⌋n ≤ s̃ for all s̃ in the sub-cube.
In order to approximate members of Bı one can use the preceding discretization on each s given in definition Bı .
Thus we have 2ıκn point set Kı,n such that:
α1
1−α D W kq g
α(
α,p |p )
e ς
−ς
µ.
∀b ∈ Bı ∃µ ∈ Kı,n such that µ ≤ b ≤ 1 +
ς2n
One can use the points of Kı,n to approximate the points in ∪t>ı Bt , as well. We apply the approximation with
the sub-cubes described above for the last ı − 1 components of b, i.e. for ı − 1 s ’s with the largest indices. The
remaining component of µ is set to the minimum element of25 µα,S . Then
∀b ∈ ∪t>ı Bt ∃µ ∈ Kı,n such that µ ≤ b ≤ 1 +
Let Kn be Kn = ∪∈{0,...,n} K,n . Then
∀b ∈ B ∃µ ∈ Kn such that kb −µk ≤ 1 +
e
1−α D W kq g
α(
α,p |p )
ς
−ς
ς2n
1−α D W kq g
α(
α,p |p ) −ς
e ς
ς2n
α1
ı
1−(1−α)
α
e
1−α D W kq g
α(
α,p |p )
ς
1−α D W kq g
α(
α,p |p )
e ς
ς
ς
n
(1−α)
α
ı
(1−α)
α
µ.
−1sups∈S kµα,s k.
Note that sups∈S kµα,s k is finite an its coefficient converges to zero as n diverges. Furthermore, Kn is a finite
g
set for any n. Thus B is totally bounded. As a result every subset of B, and hence {Tıα,p (qα,p
)}ı∈Z+ , is totally
bounded.
ı() g
ı() g
g
)}∈Z+ satisfying lim→∞ Tα,p (qα,p
(c-iv) {Tıα,p (qα,p
)}ı∈Z+ has a subsequence {Tα,p (qα,p
)−u = 0 for a u ∼ q1,p :
The existence of a limit point u and convergent subsequence follows from the compactness of the completion of
g
g
)}ı∈Z+ is totally bounded.
{Tıα,p (qα,p
)}ı∈Z+ . Completion is compact by [37, Thm 45.1] because {Tıα,p (qα,p
ı
g
g
Note that Tα,p (qα,p )≺q1,p because qα,p ∼ q1,p . Then u≺q1,p because any probability measure that is not absolute
g
continuous in q1,p is outside the closure of {Tıα,p (qα,p
)}ı∈Z+ .
g
On the other hand, µα,p ≤ qα,p by definition because kµα,p k ≤ 1. Furthermore, for any q ∈ Qα,p we have
X
dTα,p (q)
1−α
≥
p(x )( dWdν(x ) )α ( dq
dν
dν )
=(
x
dµα,p α dq 1−α
.
dν ) ( dν )
g
) for all ı ∈ Z+ . Hence µα,p ≤ u,
Hence, if µα,p ≤ q, then µα,p ≤ Tα,p (q). Consequently, µα,p ≤ Tıα,p (qα,p
ı
g
because otherwise u can not be in the closure of {Tα,p (qα,p )}ı∈Z+ . Then q1,p ≺u because q1,p ∼ µα,p .
(c-v) Tα,p (·) : Qα,p → P(Y) is continuous if both Qα,p and P(Y) have the total variation topology: First, note that
(z + t )1−α − z 1−α is a monotonically decreasing function of z on ℜ≥0 for fixed t ∈ ℜ≥0 and α ∈ (0, 1). Then for
any x with positive p(x ) and q1 , q2 ∈ Qα,p as a result of Holder’s inequality we have
Z
α 1−α
α 1−α
1−α
dW (x )
dW (x )
dq1
dq2
−
ν(dy) ≤ kq1 − q2 k
.
dν
dν
dν
dν
+
Hence e(α−1)Dα(W (x )kq) Wαq (x ) is a continuous function of q from Qα,p to M (Y) for the total variation topology.
Then Wαq (x ) is a continuous function of q for the total variation topology, as well, because Dα (W (x )k q) is
continuous in q for the total variation topology by Lemma 4. Thus Tα,p (·) : Qα,p → P(Y) is continuous.
ı() g
(c-vi) The limit point of the convergent subsequence {Tα,p (qα,p
)}∈Z+ is a fixed point of Tα,p (·), i.e. Tα,p (u) = u: the
non-negativity of the Rényi divergence for probability measures and (175) we get
X
g
g
g
)) p
Dα W k qα,p
p ≥
) p − Dα W k Tα,p (Tıα,p (qα,p
Dα W k Tıα,p (qα,p
ı∈Z+
X
g
g
≥
) .
D1 Tα,p (Tıα,p (qα,p
)) Tıα,p (qα,p
ı∈Z+
P
κ
assume ℜκ has the metric d : ℜκ × ℜκ → ℜ≥0 given by d(z , z̃ ) = κ
t =1 |zt − z̃t | for all z , z̃ ∈ ℜ .
25 Such a minimum element might not exist for an arbitrary set of measures, but for the image of S it exists: the minimum element is the image of the
minimum point of S.
24 We
32
ı() g
ı() g
g
g
)) Tα,p (qα,p
) = 0.
)) Tıα,p (qα,p
) = 0. Hence lim→∞ D1 Tα,p (Tα,p (qα,p
Then lim→∞ D1 Tα,p (Tıα,p (qα,p
On the other hand, D1 Tα,p (q) q is lower semicontinuous in q for the total variation topology because the
Rényi divergence is lower semicontinuous in its arguments by Lemma 3 and Tα,p (·) is continuous in total variation
ı() g
) converges to u in total variation topology as diverges.
topology. Then D1 Tα,p (u) u = 0 because Tα,p (qα,p
Thus Tα,p (u) = u as a result of Lemma 2.
(c-vii) qα,p satisfies (36): Recall that Dα (w k q) is continuous in q for the total variation topology by Lemma 4. Furthermore,
ı() g
lim→∞ Tα,p (qα,p
)−qα,p = 0, and Dα (W k qα,p | p) = Iα (p;W). Then
g
lim→∞ Dα W k Tı()
α,p (qα,p ) p = Iα (p;W) .
g
g
) p ≥ Dα W k Tı+t
On the other hand Dα W k Tıα,p (qα,p
α,p (qα,p ) p ≥ Iα (p;W) for all t ∈ Z+ by (175) and the
definition of the Augustin information. Thus
g
limı→∞ Dα W k Tıα,p (qα,p
) p = Iα (p;W) .
Then as a result of (37), which is implied by the assertions we have already established, we have
g
) = 0.
limı→∞ Dα qα,p k Tıα,p (qα,p
g
Then limı→∞ qα,p − Tıα,p (qα,p
) = 0 as a result of Lemma 2.
(13-d) Let the function f (·) and the set of channels U be
f (V ) ,
α
1−α D1 (V k W | p)
Y×
U , {V ∈
x :p(x )>0
∀V ∈ U,
+ I1 (p;V)
P(Y) : D1 (V k W | p) < ∞}.
The following six statements collectively imply part (d).
(d-i) If Tα,p (u) = u, then Dα (W k u| p) = Iα (p;W), (38) and (39) hold for qα,p = u, qα,p is unique and qα,p ∼ q1,p :
Z
X
1
∼ 1−α
Dα (W k q| p) − Dα (W k u| p) = α−1
p(x ) ln ( dq
Wαu (x )(dy)
∀q ∈ Qα,p
(178)
du )
x
where q∼ is the component of q that is absolutely continuous in u.
On the other hand using the Jensen’s inequality and concavity of the natural logarithm function we get
Z
Z
X
X
dq∼ 1−α
u
1
1
∼ 1−α
Wα (x )(dy) ≥ α−1
Wαu (x )(dy)
p(x ) ln ( du )
p(x ) ln( dq
α−1
du )
x
x
Z
du
)Tα,p (u)(dy).
= ln( dq
∼
(179)
Since Tα,p (u) = u by the hypothesis using (178) and (179) we get
Dα (W k q| p) − Dα (W k u| p) ≥ D1 (uk q)
∀q ∈ Qα,p .
(180)
D1 (uk q) > 0 for all q ∈ P(Y) \ {u} by Lemma 2 and Dα (W k q| p) = ∞ for q ∈
/ Qα,p by definition. Then u
is the unique probability measure satisfying Iα (p;W) = Dα (W k u| p) and (38) holds. In addition q1,p ≺u because
otherwise Dα (W k u| p) would have been infinite. Furthermore, u≺q1,p because D1 (uk q1,p ) is finite by (180) and
part (a).
The lower bound given in (39) holds for q ∈ Qα,p by (180) and for q ∈
/ Qα,p by definition. In order to establish
the upper bound given in (37), note that Tα,p (u) = u implies
i α1
α
hX
du
p(x ) dWdν(x ) e(1−α)Dα(W (x )ku)
∀ν : q1,p ≺ν.
dν =
x
Then one can confirm by substitution that
X
1
ln
p(x )e(α−1)(Dα(W (x )kq)−Dα(W (x )ku))
Dα (uk q) = α−1
x
∀q ∈ P(Y).
Then Jensen’s inequality and convexity of the exponential function imply
Dα (uk q) ≥ Dα (W k q| p) − Dα (W k u| p)
∀q ∈ P(Y).
(d-ii) f (·) : U → ℜ is concave and upper semicontinuous on U for the topology of setwise convergence:26 Using the
definition of the tilted channel given in (27), the identity given in (34), and the joint convexity of the order one
26 The set U is a subset of the Cartesian product of a finite number of copies of P(Y). What we mean by topology of setwise convergence on U is the
product topology obtained by assuming topology of setwise convergence on each component of the Cartesian product. We employ this terminology in the rest
of the proof without explicitly mentioning it.
33
Rényi divergence in its arguments, i.e. Lemma 6, we can write f (V ) as the sum of three finite terms as follows
for all V ∈ U:
X
1
(181)
p(x )V (x ) q1,p + Dα (W k q1,p | p) .
D1 (V k Wαq1,p | p) − D1
f (V ) = 1−α
x
Then f (·) is a concave because the order one Rényi divergence is convex in its first argument by Lemma 6.
Similarly, f (·) is upper semicontinuity for the topology of setwise convergence, because Rényi divergence is lower
semicontinuous in its first argument for the topology of setwise convergence by Lemma 3.
(d-iii) U′ , {V ∈ U : maxx :p(x )>0 p(x )D1 (V (x )kW (x )) ≤ α−1
α ℏ(p)} is compact for the topology of setwise convergence: For any v ∈ P(Y) and w ∈ P(Y), the identity z ln z ≥ − 1/e implies that
Z
dv
dv +
w (dy) ≤ D1 (v k w ) + 1/e.
dw ln dw
dv
Then for any γ ∈ ℜ+ and w ∈ P(Y), the set of Radon-Nikodym derivatives { dw
}v :D1(v kw)≤γ is uniformly
integrable because of the necessary and sufficient condition for the uniform integrability given by de la Vallee
Poussin [21, Thm. 4.5.9]. Hence, {v ∈ P(Y) : D1 (v k w ) ≤ γ}≺uni w . Then27 {v ∈ P(Y) : D1 (v k w ) ≤ γ} has
compact closure in the topology of setwise convergence by [21, Thm. 4.7.25]. On the hand the set {v ∈ P(Y) :
D1 (v k w ) ≤ γ} is closed, i.e. it is equal to its closure, because Rényi divergence is lower semicontinuous in its
arguments for the topology of setwise convergence by Lemma 3. Hence {v ∈ P(Y) : D1 (v k w ) ≤ γ} is compact
in the topology of setwise convergence for any γ ∈ ℜ+ and w ∈ P(Y). Then U′ is compact in the topology of
setwise convergence because product of finite number of compact sets is compact by [37, Thm. 26.7].
(d-iv) ∃U∗ ∈ U′ s.t. f (U∗ ) = supV ∈U f (V ): Note that W ∈ U and f (W ) = I1 (p;W). Furthermore, I1 (p;W) ≥ 0 by
Lemma 2 and part (b). On the other hand, if p(x )Dα (V (x )k W (x )) > α−1
α ℏ(p) for an x , then f (V ) < 0 because
Dα (V (x )k W (x )) ≥ 0 by Lemma 2 and I1 (p;V) ≤ ℏ(p) by part (a). Thus,
supV ∈U f (V ) = supV ∈U′ f (V ).
On the other hand, ∃U∗ such that f (U∗ ) = supV ∈U′ f (V ) by the extreme value theorem for the upper semicontinuous functions [31, Ch3§12.2] because U′ is compact and f (·) is upper semicontinuous for the topology of setwise
convergence.
P
(d-v) f (U∗ ) = Dα (W k u∗ | p) where u∗ ,
x p(x )U∗ (x ): As a result of Lemma 10 we have
Dα (W k u∗ | p) = supV ∈U
α
1−α D1 (V k W | p)
+ D1 (V k u∗ | p) .
(182)
α
D1 (U∗ k W | p) + D1 (U∗ k u∗ | p) = f (U∗ ) because I1 (p;U∗) = D1 (U∗ k u∗ | p) by part (b).
On the other hand 1−α
Then Dα (W k u∗ | p) ≥ f (U∗ ) is evident by (182).
In order to prove Dα (W k u∗ | p) ≤ f (U∗ ), let us consider a V ∈ U and define V (ı) and q (ı) for each ı ∈ Z+ as
V (ı) ,
q (ı) ,
1
ı−1
ı U∗ + ı V ,
X
ı−1
1
ı u∗ + ı
x
p(x )V (x ).
As a result of the decomposition given in (181) we have
1
f (V (ı) ) = 1−α
D1 V (ı) Wαq1,p p − D1 q (ı) q1,p + Dα (W k q1,p | p) .
Then using the Jensen’s inequality and convexity of the order one Rényi divergence in its first argument established
in Lemma 6 we get
ı−1
(ı)
q1,p
q1,p
1
1
f (V (ı) ) ≥ 1−α
q
q
|
p)
+
|
p)
−
D
D
(U
k
W
D
(V
k
W
1,p + Dα (W k q1,p | p)
1
1
∗
α
α
ı
ı 1
h
i
(ı)
1
α
= ı−1
q1,p
ı [f (U∗ ) + D1 (u∗ k q1,p )] + ı 1−α D1 (V k W | p) + D1 (V k q1,p | p) − D1 q
i
h
h
i
α
= ı−1
D1 (V k W | p) + D1 V k q (ı) p .
f (U∗ ) + D1 u∗ k q (ı) + 1ı 1−α
ı
Then using f (U∗ ) = supV ∈U f (V ) ≥ f (V (ı) ) and D1 u∗ k q (ı) ≥ 0 we get
α
f (U∗ ) ≥ 1−α
∀ı ∈ Z+ .
D1 (V k W | p) + D1 V k q (ı) p
On the other hand, limı→∞ D1 V k q (ı) p ≥ D1 (V k u∗ | p) because limı→∞ q (ı) − u∗ = 0 by construction and
the Rényi divergence is lower semicontinuous in its second argument for the topology of setwise convergence by
Lemma 3. Then
f (U∗ ) ≥
27 Note
α
1−α D1 (V k W | p)
+ D1 (V k u∗ | p)
that {v ∈ P(Y) : D1 (v k w ) ≤ γ} is bounded in variation norm by definition.
34
∀V ∈ U.
Hence, f (U∗ ) ≥ Dα (W k u∗ | p) by (182).
(d-vi) Tα,p (u∗ ) = u∗ and U∗ (x ) = Wαu∗ (x ) for all x such that p(x ) > 0: Note that Dα (W k u∗ | p) < ∞ by part (d-v)
because f (U∗ ) ≤ ℏ(p) < ∞ by definition. Consequently, we can define the tilted probability measure Wαu∗ (x )
for each x such that p(x ) > 0. Using the definitions of f (·) and the tilted channel Wαq together with the identity
I1 (p;U∗) = D1 (U∗ k u∗ | p), which follows from part (b), we get
f (U∗ ) = Dα (W k u∗ | p) +
u∗
1
1−α D1 (U∗ k Wα | p) .
u∗
Since f (U∗ ) = Dα (W k u∗ | p) by part (d-v) we get D1 (U∗ k Wαu∗ | p) = 0. Hence
P U∗ (x ) = Wα (x ) for all x
such that p(x ) > 0 by Lemma 2. As a result Tα,p (u∗ ) = u∗ because u∗ =
x p(x )U∗ (x ) and Tα,p (u∗ ) =
P
u∗
p(x
)W
(x
)
by
definition.
α
x
(13-e) We prove the statement for α ∈ (0, 1) and α ∈ (1, ∞) cases separately,
P
qα,p
(x ) = qα,p by parts (c). Then as a result of (34) we have
• Recall that
x p(x )Wα
I1 (p;Wαqα,p) = D1 (Wαqα,p k qα,p | p) .
(183)
Then (40) follows from Lemma 10 and Iα (p;W) = Dα (W k qα,p | p) for α ∈ (0, 1).
On the other hand as a result the definition of the Augustin information, and Lemma 10 we have
α
1−α D1 (V k W | p)
α
inf V ∈P(Y|X) inf q∈P(Y) 1−α
D1 (V k W | p)
Iα (p;W) = inf q∈P(Y) inf V ∈P(Y|X)
=
•
+ D1 (V k q| p)
+ D1 (V k q| p) .
Then (41) follows from the definition of the order one Augustin information.
q
Note that for α ∈ (1, ∞) identity given in (40) is nothing but f (Wα α,p ) = Iα (p;W) which is already established in
the proof of part (d). Similarly (41) is equivalent to supV ∈P(Y|Y) f (V ) = Iα (p;W) which is established in the proof
of part (d).
Proof of Lemma 14. For α = 1 one can confirm
Q⊗ (43) and (44) by substitution. For any α ∈ ℜ+ \ {1} one can show by
substitution that the probability measure q = t∈T qα,pt is a fixed point of Tα,p (·), i.e. Tα,p (q) = q. Furthermore, q1,p ≺q
because q1,pt ≺qα,pt for each t ∈ T. Thus (44) follows from Lemma 13-(c,d).
On the other hand, the following identity can be confirmed by substitution
X
Y⊗
qα,pt p =
Dα (Wt k qα,pt | pt ) .
Dα W k
t∈T
t∈T
Then (43) follows from Lemma 13-(c,d) and (44).
Proof of Lemma 15. Note that Dα (W k q| p) is linear and hence concave in p for any q by definition. Then Iα (p;W) is concave
in p because pointwise infimum of a family of concave functions is concave. Furthermore by Lemma 13-(b,c,d), ∃!qα,pβ such
that Dα W k qα,pβ pβ = Iα (pβ ;W). In addition,
Dα W k qα,pβ pβ = βDα W k qα,pβ p1 + (1 − β)Dα W k qα,pβ p0 .
Then equation (45) and (46) are obtained by bounding Dα W k qα,pβ p1 and Dα W k qα,pβ p0 using Lemma 13-(b,c,d).
On the other hand, using Lemma 1 we get
Dα (W k βqα,p1 + (1 − β)qα,p0 | pβ ) = βDα (W k βqα,p1 + (1 − β)qα,p0 | p1 ) + (1 − β)Dα (W k βqα,p1 + (1 − β)qα,p0 | p0 )
≤ βDα (W k qα,p1 | p1 ) − β ln β + (1 − β)Dα (W k qα,p0 | p0 ) − (1 − β) ln(1 − β)
= βIα (p1 ;W) + (1 − β)Iα (p0 ;W) + ℏ(β) .
Then equation (47) follows from the lower bound on Dα (W k q) p given in Lemma 13-(b,c,d).
Proof of Lemma 16. Note that as result of (42) we have,
1
[p(x )] α e
1−α
α Dα(W (x )kqα,p )
W (x ) ≤ qα,p
∀x ∈ X.
(16-a) Using Lemma 1 and (184) we get
1−α
1
Dα (W (x )k qα,p ) ≤ Dα W (x )k [p(x )] α e α Dα(W (x )kqα,p ) W (x )
=
1
α
1
ln p(x
) +
α−1
α Dα (W (x )k qα,p ) .
(16-b) We use Dα (W (x )k qα,p ) > 0 and (184) for α ∈ (0, 1] case and part (a) and (184) for α ∈ (1, ∞) case.
(16-c) We prove α ∈ (0, 1) case and α ∈ (1, ∞) case separately.
35
(184)
•
α
1
α ∈ (0, 1): First use Jensen’s inequality, i.e. E[ξ α ] ≤ E[ξ] , in (42); then invoke Dα (W (x )k qα,p ) ≤ ln p(x
):
X
1−α
dqα,p
(x )
α Dα(W (x )kqα,p )
p(x ) dW
dq1,p ≤
dq1,p e
x
α−1
≤ (minx :p(x )>0 p(x )) α .
P
P
Recall that if ξ(x ) ≥ 0 for all x , then x [ξ(x )]α ≥ [ x ξ(x )]α . Using Dα (W (x )k qα,p ) ≥ 0 we get
X
1
dqα,p
(x ) ( 1−α
α )Dα(W (x )kqα,p )
(p(x )) α dW
dq1,p ≥
dq1,p e
x
≥ (minx :p(x )>0 p(x ))
•
1−α
α
.
α
1
α ∈ (1, ∞): First use Jensen’s inequality, i.e. E[ξ α ] ≥ E[ξ] , in (42), then invoke Dα (W (x )k qα,p ) ≤ ln p(x
):
X
dqα,p
(x ) 1−α
α Dα(W (x )kqα,p )
p(x ) dW
dq1,p ≥
dq1,p e
x
α−1
≥ (minx :p(x )>0 p(x )) α .
P
P
Recall that if ξ(x ) ≥ 0 for all x , then x [ξ(x )]α ≤ [ x ξ(x )]α . Using Dα (W (x )k qα,p ) ≥ 0 we get
X
1
dqα,p
(x ) ( 1−α
α )Dα(W (x )kqα,p )
≤
(p(x )) α dW
dq1,p
dq1,p e
x
≤ (minx :p(x )>0 p(x ))
1−α
α
.
Proof of Lemma 17.
(17-a) Note that D1 (V k W | p) ≥ 0 by Lemma 2 and I1 (p;V) ∈ [0, ℏ(p)] by Lemma 13-(a). Then using D1 (W k W | p) = 0 for
α = 1 case and invoking Lemma 13-(e) for α ∈ ℜ+ \ {1} case we get
(α − 1)Iα (p;W) = supV ∈P(Y|X) (α − 1)I1 (p;V) − αD1 (V k W | p) .
Then (α − 1)Iα (p;W) is convex in α because pointwise supremum of a family of linear/convex functions is convex.
On the other hand, using V = W we can deduce that,
(α − 1)Iα (p;W) ≥ (α − 1)I1 (p;W) .
In addition I1 (p;W) ∈ [0, ℏ(p)] by Lemma 13-(a). Thus (α − 1)Iα (p;W) ≥ −ℏ(p).
(17-b) Since (α − 1)Iα (p;W) is finite and convex in α on ℜ+ , it is continuous on ℜ+ by [20, Thm. 6.3.3]. Then
continuous in α on ℜ+ , as well. Furthermore,
1−α
α Iα (p;W)
= inf V ∈P(Y|X)
1−α
α I1 (p;V)
1−α
α Iα (p;W)
is
+ D1 (V k W | p) .
Then 1−α
α Iα (p;W) is nonincreasing in α because infimum of a family of nonincreasing
1−α
I
(p;V)
+ D1 (V k W | p) is nonincreasing in α because I1 (p;V) is nonnegative.
α 1
functions is deceasing. Note that
(17-c) Iα (p;W) is nondecreasing in α because the pointwise infimum of a family of nondecreasing functions is nondecreasing
and the Rényi divergence is nondecreasing in its order by Lemma 8.
Since (α − 1)Iα (p;W) is finite and convex in α on ℜ+ , it is continuous on ℜ+ by [20, Thm. 6.3.3]. Then Iα (p;W) is
continuous on (0, 1) and (1, ∞). In order extend the continuity to ℜ+ we need to prove that Iα (p;W) is continuous at
α = 1. Note that as a result of the definition of the Augustin information we have Iα (p;W) ≤ Dα (W k q1,p | p) for all
α ∈ ℜ+ . Since Iα (p;W) is nondecreasing in α we have
I1 (p;W) ≤ Iα (p;W) ≤ Dα (W k q1,p | p)
− |1−α|
α
Since qα,p ≤ (minx :p(x )>0 p(x ))
∀α ∈ (1, ∞).
(185)
q1,p by Lemma 16-(c), using Lemma 1 we get
Dα (W k qα,p | p) ≥ Dα (W k q1,p | p) +
1−α
α
ln(minx :p(x )>0 p(x ))
∀α ∈ (0, 1).
Recall that Dα (W k qα,p | p) = Iα (p;W) by Lemma 13-(c) and Iα (p;W) is nondecreasing in α. Thus we have
Dα (W k q1,p | p) +
1−α
α
ln(minx :p(x )>0 p(x )) ≤ Iα (p;W) ≤ I1 (p;W)
∀α ∈ (0, 1).
(186)
On the other hand, Dφ (W k q1,p | p) < ∞ for any φ ∈ ℜ+ by Lemma 13-(a). Then Dα (W k q1,p | p) is continuous in α by
Lemma 8. Furthermore, D1 (W k q1,p | p) = I1 (p;W) by Lemma 13-(b). Then
limα→1 Dα (W k q1,p | p) = I1 (p;W) .
Then the continuity of Iα (p;W) at α = 1 follows from (185) and (186).
36
(17-d) Let τx (α) be τx (α) ,
α−1
α Dα (W (x )k qα,p ).
ln
dqη,p
dq1,p
=
1
η
Then we can rewrite (42) as follows
φη
φ
X
dW (x )
(1−φ)Dφ(W (x )kqφ,p )
eη(τx (φ)−τx (η))
e
ln
p(x )
dq1,p
x
Let us assume without loss of generality that φ > η. Then using the Jensen’s inequality we get
ln
On the other hand, using the fact that
ln
dqη,p
dq1,p
≥ ln
dqφ,p
dq1,p
dqη,p
dq1,p
P
dqφ,p
dq1,p
≤ ln
+ maxx :p(x )>0 (τx (φ) − τx (η)).
η
P
φ for non-negative ξ(x ) we get
x ξ(x )]
x [ξ(x )] ≥ [
(187)
η
φ
+ (1 − φη ) ln(minx :p(x )>0 p(x )) + minx :p(x )>0 (τx (φ) − τx (η)).
(188)
dq
}y∈Y is equicontinuous in α by (187) and (188). On the other hand,
If {τx (α)}x :p(x )>0 is equicontinuous in α, then {ln dqα,p
1,p
there are only finitely many x ’s with positive p(x ). Thus {τx (α)}x :p(x )>0 is equicontinuous if each τx (α) is continuous.
We are left with establishing the continuity of τx (α).
Let gx (·) : [η, φ] → ℜ+ , fx (·, ·) : [η, φ] × Y → ℜ≥0 , and s· : [η, φ] → P(Y) be
Z
gx (α) ,
fx (α, y)ν(dy ),
α
dsα 1−α
fx (α, y) , dWdν(x )
,
dν
sα ,
φ−α
φ−η qη,p
+
α−η
φ−η qφ,p .
Then fx (α, y) is differentiable in α and its derivative can be bounded using Lemma 16-(b) and the identity τ ln τ1 ≤ 1e :
α
α
h
i
dqη,p
dW (x )
dW (x )
dW (x )
dsα 1−α
dsα −α 1−α dqφ,p
∂
f
(α,
y)
=
−
+
ln
x
∂α
dν
dν
dsα
dν
dν
φ−η
dν
dν
h
i
dsβ
dqφ,p
dqη,p
β
1
1
∂
1
+ ( 1 1 )β |1−β|
1 ) ln
1
∂α fx (α, y) α=β ≤ dν
βe + (
φ−η
dν + dν
p(x ) η∧1
p(x ) η∧1
p(x ) η∧1
dqφ,p
dqη,p
1+φ
φ
1
1
1
ln
∀β ∈ [η, φ].
+ φ−η
≤ dν + dν
1 )
1
ηe + (
p(x ) η∧1
p(x ) η∧1
The expression on the right hand side is ν−integrable. Thus as a result of [21, Corollary 2.8.7] we have
Z
∂
∂
g
(α)
=
x
∂α
∂α fx (α, y)ν(dy ).
∂
∂
gx (α) is continuous by [21, Corollary 2.8.7] because ∂α
fx (α, y) is continuous in α. Then ln gαx (α) is a
Furthermore, ∂α
continuous function on [η, φ] that is continuously differentiable on (η, φ). Then, as a result of mean value theorem [38,
Thm. 5.10] we have
ln gx (φ)
φ
−
ln gx (η)
η
Using Lemma 16-(b) and the identity τ ln τ1 ≤
∂ ln gx (α)
∂α
α
α=β
=
≤
− ln gx (β)
β2
ln gx (β)
β2
+
+
1
e
≤ (φ − η) supβ∈(η,φ)
we get
Z
1
βgx (β)
1
αgx (β)
∂ ln gx (α)
∂α
α
α=β
.
∂
∂α fx (α, y) α=β
1
βe
+(
1
1
p(x ) η∧1
ν(dy)
1
)β ln
1
p(x ) η∧1
+
|1−β|
φ−η kqφ,p
(189)
− qη,p k .
(190)
We bound ln gx (β) using Lemma 16-(b) and the convexity of the Rényi divergence in its second argument, i.e. Lemma 5,
|ln gx (β)| = |(β − 1)Dβ (W (x )k sβ )|
≤
φ∨1
η∧1
1
ln p(x
)
∀β ∈ [η, φ].
The Rényi divergence is nondecreasing in its order by part (c). Thus Lemmas 2 and 13 imply that
q
2
kqφ,p − qη,p k ≤ η∧1
(Iφ (p;W) − Iη (p;W)).
(191)
(192)
On the other hand, one can confirm by substitution that
τx (φ) − τx (η) =
ln gx (φ)
φ
−
ln gx (η)
.
η
(193)
Then the continuity of τx (α) in α is implied by (189), (190), (191), (192), (193) and the continuity of the Augustin
information in the order established in part (c).
37
(17-e) Dη (W k qφ,p | p) ≥ Iη (p;W) for any φ ∈ ℜ+ and η ∈ ℜ+ by the definition of the Augustin information. Then the
differentiability of Dα (W k q| p) in α established in Lemma 8 implies that,
limη↓φ
Iη(p;W)−Iφ(p;W)
η−φ
Dη(W kqφ,p |p)−Dφ(W kqφ,p |p)
η−φ
≤ limη↓φ
=
I (p;W)−I (p;W)
limη↑φ η η−φφ
≥
=
∂
∂α Dα (W k qφ,p | p) α=φ
D (W kqφ,p |p)−Dφ(W kqφ,p |p)
limη↓φ η
η−φ
∂
D
(W
k
q
|
p)
.
φ,p
∂α α
α=φ
(194)
(195)
Similarly,28 Dφ (W k qη,p | p) ≥ Iφ (p;W) for any φ ∈ ℜ+ and η ∈ ℜ+ by the definition of the Augustin information. Hence,
Iη(p;W)−Iφ(p;W)
η−φ
Iη(p;W)−Iφ(p;W)
limη↑φ
η−φ
limη↓φ
Dη(W kqη,p |p)−Dφ(W kqη,p |p)
η−φ
Dη(W kqη,p |p)−Dφ(W kqη,p |p)
limη↑φ
.
η−φ
≥ limη↓φ
(196)
≤
(197)
For any δ ∈ (0, φ) by Lemma 16-(b) and Lemma 1 we have
Dα (W (x )k qη,p ) ≤
1
(φ−δ)∧1
1
ln p(x
)
∀η : |η − φ| < δ, α ∈ ℜ+ .
(198)
Then as a result of Lemma 12 there exists29 a Kφ,p > 0 such that for η close enough to φ we have
We show in the following that
∂
∂α Dα (W k qη,p | p) α=φ
limη→β
≤ Kφ,p |η − φ|2 .
∂
∂α Dα (W k qη,p | p) α=φ
Dη (W k qη,p | p) − Dφ (W k qη,p | p) − (η − φ)
(199)
is a continuous function of η, i.e.
∂
∂α Dα (W k qη,p | p) α=φ
∂
∂α Dα (W k qβ,p | p) α=φ
=
.
(200)
Using (199) and (200) we get
limη→φ
Dη(W kqη,p |p)−Dφ(W kqη,p |p)
η−φ
=
∂
∂α Dα (W k qφ,p | p) α=φ
.
(201)
Differentiability of the Augustin information and (50) follow from (194), (195), (196), (197), and (201).
∂
In order to establish the continuity of ∂α
Dα (W k qη,p | p) α=φ in η, i.e. (200), let us first recall that the expression for the
derivative of the Rényi divergence given in (21):
q
R dWφqη,p (x )
P
dWφη,p (x )
1
ln dW (x )
q1,p (dy)
φ 6= 1
(φ−1)2 x p(x )
dq1,p
∂
.
2
∂α Dα (W k qη,p | p) α=φ = P
(x )
p(x ) R dW (x )
ln dW
q1,p (dy) − [D1 (W (x )k qη,p )]2
φ=1
x 2
dq1,p
dqη,p
Recall that,
Dφ (W (x )k qη,p ) =
R dW (x ) φ dqη,p 1−φ
q1,p (dy)
dq1,p
dq1,p
dqη,p
dW (x )
ln dq1,p − ln dq1,p q1,p (dy)
1
φ−1 ln
R
dW (x )
dq1,p
φ 6= 1
.
φ=1
Then Dφ (W (x )k qη,p ) is continuous in η for any φ ∈ ℜ+ by [21, Corollary 2.8.7-(i)] because {ln
tinuous function of η by part (d) and ln
q
ln
dWφη,p (x )
dq1,p
dqη,p
dq1,p
≤
|η−1|
η
1
is equicon-
ln minx :p(x )>0 p(x ) by Lemma 16-(c). On the other hand,
(x )
= φ ln dW
dq1,p + (1 − φ) ln
q
dWφη,p (x )
dqη,p
dq1,p }y∈Y
dqη,p
dq1,p
+ (1 − φ)Dφ (W (x )k qη,p ) .
dq
}y∈Y is equicontinuous and Dφ (W (x )k qη,p ) is continuous
Thus { dq1,p }y∈Y is equicontinuous in η because {ln dqη,p
1,p
in η. Furthermore, using Lemma 1, Lemma 16-(b), and the identity τ ln τ1 ≤ 1e we obtain the following bounds
!
q
dWφη,p (x )
dq1,p
q
ln
dWφη,p (x )
dW (x )
ln
dWφη,p (x )
dW (x )
q
dWφη,p (x )
dq1,p
≤
dW (x )
dq1,p
≤
dW (x )
dq1,p 1
q
1
qη,p
(x )
dW
φ
{ dW
≤1}
(x )
qη,p
(x )
dW
φ
{ dW
≤1}
(x )
1
e
1
e
+
+1
1−φ
1−φ
qη,p
(x )
dW
φ
{ dW
>1}
(x )
dqη,p
dq1,p 1
28 For
(p(x )) η∧1 ln(p(x )) η∧1
qη,p
(x )
dW
φ
{ dW
>1}
(x )
1−φ
φe
−1
(p(x )) η∧1
if φ ∈ [1, ∞)
(202)
if φ ∈ (0, 1).
(203)
φ ∈ (1, ∞) we can use (41) of Lemma 13-(e) to establish reverse inequalities for (194) and (195).
that, this is not just the Taylor expansion of Dα (w k q) around α = φ for a given (w , q) pair. Lemma 11 allows us to apply the Taylor expansion
for a family of (w , q) pairs around α = φ simultaneously if we can bound Dβ (w k q) uniformly for all (w , q)’s for some β > φ.
29 Note
38
q
dWφη,p (x )
dq1,p
q
ln
dWφη,p (x )
dW (x )
is bounded from above by a q1,p -integrable function by Lemma 16-(c) for any
q
closed interval [a, b] ⊂ ℜ+ . Thus D1 Wφ η,p W p is a continuous function of η by [21, Corollary 2.8.7-(i)] for all
Then
η∈[a,b]
∂
∂
φ ∈ ℜ+ . Then ∂α
Dα (W k qη,p | p) α=φ is continuous in η for φ ∈ ℜ+ \ 1. The continuity of ∂α
Dα (W k qη,p | p) α=1 in
η follows from the continuity of D1 (W (x )k qη,p ) and [21, Corollary 2.8.7-(i)] via the following bound, which can be
established using the identity τ (ln τ )2 1{τ ∈(0,1]} ≤ e42 and Lemma 16-(b),
2
dq
dW (x )
dW (x )
dW (x ) ln p(x ) 2
4
ln
q1,p (dy) ≤ dqη,p
( η∧1 ) .
2 + dq
dq1,p
dqη,p
1,p e
1,p
Now we are left with establishing the continuity of the derivative of the Augustin information. Since {ln
equicontinuous in α by part (d), for any ǫ > 0 there exists a δ such that
e−ǫ qφ,p ≤ qη,p ≤ eǫ qφ,p
dqα,p
dq1,p }y∈Y
is
∀η : |η − φ| < δ.
1
On the other hand (p(x )) φ∧1 W (x ) ≤ qφ,p by Lemma 16-(b). Then as a result of Lemma 1
Dz (W (x )k qη,p ) ≤
1
φ∧1
1
ln p(x
) +ǫ
∀η : |φ − η| < δ, ∀z ∈ ℜ+ .
Hence, Lemma 12 implies the existence of a τ ∈ ℜ+ that does not depend on η such that
∂ κ Dα(W kqη,p |p)
∂ακ
Then lim supκ→∞
1 ∂κ
κ! ∂ακ
∂Dα(W kqη,p |p)
∂α
α=φ
≤ (κ + 1)!τ κ+1
∀η : |φ − η| < δ.
1/κ
∂
≤ τ . Thus
∂α Dα (W k qη,p | p) α=φ
φ is at least τ1 for all η ∈ [φ− δ, φ+ δ] by
(204)
the radius of convergence of the Taylor’s expansion of
around α =
Hadamard’s formula [36, Lemma 1.1.8]. Furthermore,
we can use (204) to bound higher order derivatives:
X∞ |β−φ|ı ∂ ı+1 D (W kq |p)
∂Dα(W kqη,p |p)
∂Dα(W kqη,p |p)
α
η,p
≤
−
∂α
∂α
ı!
∂αı+1
ı=1
α=β
α=φ
α=φ
X∞
ı ı
2
2
≤τ
ı + 3ı + 2 |β − φ| τ
∀η : |η − φ| ≤ δ, ∀β : |β − φ| < τ1 .
ı=1
Using identities
P∞
ı=1
ı
ξ =
∂Dα(W kqη,p |p)
∂α
α=β
ξ
1−ξ ,
−
P∞
ı=1
ı
ıξ =
ξ
(1−ξ)2
∂Dα(W kqη,p |p)
∂α
α=φ
and
P∞
ı=1
ı2 ξ ı =
(1+ξ)ξ
(1−ξ)3
2
|β−φ|) +2
≤ τ 3 |β − φ| (1−τ
(1−τ |β−φ|)3
for |ξ| < 1 we get,
∀η : |η − φ| ≤ δ, ∀β : |β − φ| < τ1 .
Then using (50) we get
∂
∂α Iα (p;W) α=η
∂
Iα (p;W) α=η
η→φ ∂α
Hence, lim
−
∂Dα(W kqη,p |p)
∂α
= lim
η→φ
2
α=φ
|η−φ|) +2
≤ τ 3 |η − φ| (1−τ
(1−τ |η−φ|)3
∂Dα(W kqη,p |p)
,
∂α
α=φ
∀η : |η − φ| ≤ δ ∧ τ1 .
if the latter limit exists. However, we have already established
the existence of that limit in order to calculate the derivative of the Augustin information: it is equal to
Thus the Augustin information is continuously differentiable in the order.
∂
∂α Iα (p;W) α=φ .
Proof of Lemma 18. Lemma 18 is nothing but Lemma 33 for the case when λ is a vector of zeros. Thus we do not present
a separate proof for Lemma 18, see the proof of Lemma 33.
C. Augustin’s Proof of Lemma 13-(c)
We have employed the relative compactness in the total variation topology for proving Lemma 13-(c) because we wanted
to assert qα,p ∼ q1,p , the convergence described in (36), and the inequality given in (37). Establishing the existence of a
unique Augustin mean together with identity given in (35) is considerably easier. It can be done using the concept of relative
compactness in the topology of setwise convergence, as demonstrated by Augustin in [2]. Augustin also presents a proof of
Lemma 13-(c), we discuss the caveats in his argument in the following.
Let us first establish the existence of a unique Augustin mean. First, we establish
(175) as we have done in the current proof.
g
Then we consider the set Q′ , {q ∈ P(Y) : Dα (W k q| p) < Dα W k qα,p
p }. Note that Iα (p;W) = inf q∈Tα,p (Q′ ) Dα (W k q| p)
because of the definition of Q′ and (175). Furthermore, using Holder’s inequality one can show that
g
Dα(W kqα,p |p )
Tα,p (q)(E) ≤ e minx :p(x )>0 p(x ) (µα,p (E))α
39
∀E ∈ Y.
g
Thus Tα,p (Q′ )≺uni qα,p
and Tα,p (Q′ ) has compact closure in the topology of setwise convergence by a version of DunfordPettis theorem [21, 4.7.25]. On the other hand, Dα (W k q| p) is lower-semicontinuous in q for the topology of setwise convergence because Dα (W (x )k q) is, by Lemma 3. Then there exists a qα,p in the closure of Tα,p (Q′ ) for the topology of setwise
convergence such that Dα (W k qα,p | p) = inf q∈P(Y) Dα (W k q| p) by the extreme value theorem for lower semicontinuous
functions [31, Ch3§12.2]. The uniqueness of qα,p follows from the strict convexity of the Rényi divergence in its second
argument described in Lemma 5.
This construction establishes certain additional properties of the Augustin mean, as well. Note that qα,p ≺q1,p because qα,p
is in the closure of Tα,p (Q′ ) for the topology of setwise convergence. In addition, Tα,p (qα,p ) = qα,p because of Lemma 2 and
(175). Furthermore, any q satisfying Tα,p (q) = q and q1,p ≺q is equal to qα,p because of the argument presented in step (c-i)
of the proof of Lemma 13-(c). These observations, with minor differences, exist in Augustin’s proof of [2, Lemma 34.2].
Above discussion establishes Lemma 13-(c) except for three assertions:
(i) q1,p ≺qα,p ,
(ii) the identity given in (36),
(iii) the inequality given in (37).
g
g
g
)}ı∈Z+ has a subsequence
Note that Tα,p (Q′ )≺uni qα,p
and {Tıα,p (qα,p
)}ı∈Z+ ⊂ Tα,p (Q′ ). Then by [21, Thm. 4.7.25], {Tıα,p (qα,p
ı() g
ı
g
{Tα,p (qα,p )}∈Z+ converging to a q ∈ cl({Tα,p (qα,p )}ı∈Z+ ) where both the convergence and the closure are for the topology
of setwise convergence. Furthermore, q ∼ q1,p because of the arguments used in step (c-iv) of the proof of Lemma 13-(c). There
g
)}ı∈Z+
are two ways one can prove remaining assertions of Lemma 13-(c) without using the totally boundedness of {Tıα,p (qα,p
established in step (c-iii) of the proof of Lemma 13-(c):
•
If one can show that
ı()
g
dTα,p
(qα,p
)
dq1,p
dq
dq1,p in measure
ı() g
Tα,p (qα,p
)−q =
converges to
theorem [21, 4.5.4] one would have lim→∞
and (c-iv) and proceed with the step (c-v) of the proof.
q1,p , then because of the Lebesgue-Vitali convergence
0, established step (c-iv). Thus one can skip steps (c-iii)
dTı() (q g )
α,p
If one can show that the limit point q of the subsequence α,p
is a fixed point of Tα,p (·), then one would have
dq1,p
a statement equivalent to step (c-vi). Thus one can skip steps (c-iii), (c-iv), and (c-vi) and proceed with the step (c-vii)
after deriving (c-v).
We proved Lemma 13-(c) using the concept of totally boundedness because we could not find an easy way to establish either
the convergence in measure property or the fixed point property mentioned in the preceding discussion. However, we do know
that both properties hold. The convergence in measure holds because of the only if part of the Lebesgue-Vitali convergence
g
theorem [21, 4.5.4]. The fixed point property holds because {Tıα,p (qα,p
)}ı∈Z+ has a unique limit point both in total variation
topology and in topology of setwise convergence by (36).
ı()
While proving [2, Lemma 34.2], after establishing the weak convergence of {Tα,p (q1,p )}∈Z+ to q, Augustin asserts that
•
ı()
lim→∞ Tα,p (q1,p ) − q = 0. This is the first one of our two major reservations for Augustin’s proof of [2, Lemma 34.2].
Note that convergence in the topology of setwise convergence and weak convergence are one and the same thing for sequences
of measures by [21, Corollary 4.7.26]. Furthermore, convergence in the topology of setwise convergence does not imply
convergence in total variation topology.30 Thus we don’t know how one can justify such an assertion.
In order to prove [2, Lemma 34.2], Augustin establishes the totally boundedness of {Tıα,p (q1,p )}ı∈Z+ for the total variation
metric. In that proof Augustin asserts that Tıα,p (q1,p ) is in B , defined in equation (176), for some ∈ Z+ . We don’t know
g
whether such an assertion is correct or not. But we know that Tıα,p (qα,p
) is in Bı . Thus one can fix this problem easily.
A more important problem stems from Augustin’s obliviousness about the infiniteness of the set of positive integers. Either
in his discussion or in his equations there is no evidence suggesting that he makes a distinction of cases for approximating
g
g
{Ttα,p (qα,p
)}t≤ı and {Ttα,p (qα,p
)}t>ı . This is our other major reservation for Augustin’s proof of [2, Lemma 34.2].
D. Proofs of Lemmas on the Augustin Capacity
Proof of Lemma 19.
(i) ∀α ∈ ℜ+ ∃e
p ∈ P(X) s.t. Iα (e
p ;W) = Cα,W,A : Since ℏ(p) ≤ ln |X| for all p ∈ P(X), (49) and Lemma 13-(a) imply that
2k
2k
.
(205)
ln |X| + ℏ kp1 −p
|Iα (p2 ;W) − Iα (p1 ;W)| ≤ kp1 −p
2
2
Hence, Iα (p;W) is continuous in p on P(X). On the other hand, P(X) is compact because X is a finite set. Then A is
compact because any closed subset of a compact set is compact, [37, Thm. 26.2]. Then there exists a pe ∈ A such that
Iα (e
p ;W) = supp∈A Iα (p;W) by the extreme value theorem,31 [37, 27.4].
30 Consider, for example, the sequence of measure on the unit interval whose Radon-Nikodym derivatives with respect to the Lebesgue measure is given
by {(1 + cos(πz ))}∈Z+ . This set of probability measures converges to the Lebesgue measure on every measurable set, but not in total variation.
31 We do not need to establish the continuity of I (p;W) in p; the upper semicontinuity is sufficient as a result of [31, Ch3§12.2]. On the other hand
α
Iα (p;W) is upper semicontinuous in p because it is the infimum of a family of linear functions.
40
(ii) If α ∈ ℜ+ and Iα (e
p ;W) = Cα,W,A , then Dα (W k qα,ep | p) ≤ Cα,W,A for all p ∈ A: Let p be any member of A and p (ı)
1
ı−1
be ı pe + ı p for ı ∈ Z+ . Then p (ı) ∈ A because A is convex. Furthermore, by Lemma 13-(b,c,d) we have
1
p
+
W
k
q
W
k
q
D
D
Iα p (ı) ;W = ı−1
(ı) e
(ı) p
α
α
α,p
α,p
ı
ı
h
i
ı−1
≥ ı Iα (e
∀ı ∈ Z+ .
p ;W) + Dα∧1 qα,ep k qα,p (ı) + 1ı Dα W k qα,p (ı) p
Using Iα p (ı) ;W ≤ Cα,W,A , Iα (e
p ;W) = Cα,W,A , and Dα∧1 qα,ep k qα,p (ı) ≥ 0 we get
Cα,W,A ≥ Dα W k qα,p (ı) p
∀ı ∈ Z+ .
(206)
On the other hand, using Iα p (ı) ;W ≤ Cα,W,A , Iα (e
p ;W) = Cα,W,A and Dα W k qα,p (ı) p ≥ 0 we get
ı−1
1
∀ı ∈ Z+ .
q
k
q
C
≥
D
(ı)
α,W,A
α∧1
α,e
p
α,p
ı
ı
Then using Lemma 2 we get
q
2
1
α∧1 ı−1 Cα,W,A
∀ı ∈ Z+ .
≥ qα,ep − qα,p (ı)
Thus qα,p (ı) converges to qα,ep in the total variation topology and hence in the topology of setwise convergence. Since
the Rényi divergence is lower semicontinuous in the topology of setwise convergence by Lemma 3, we have
(207)
lim inf ı→∞ Dα W k qα,p (ı) p ≥ Dα (W k qα,ep | p)
Then the inequality Dα (W k qα,ep | p) ≤ Cα,W,A follows from (206) and (207).
(iii) If α ∈ ℜ+ , then ∃!qα,W,A ∈ P(Y) satisfying (67) such that qα,p = qα,W,A for all p ∈ A satisfying Iα (p;W) = Cα,W,A :
If Iα (p;W) = Cα,W,A for a p ∈ A such that qα,p 6= qα,ep , then Dα (W k qα,ep | p) > Cα,W,A by Lemma 13-(b,c,d) and
Lemma 2. This is a contradiction because we have already established that Dα (W k qα,ep | p) ≤ Cα,W,A for all p ∈ A.
Then qα,p = qα,ep for any p ∈ A satisfying Iα (p;W) = Cα,W,A .
Proof of Theorem 1. The right hand side of (65) is an upper bound on the left hand side because of max-min inequality.
Furthermore, the left hand side of (65) is equal to Cα,W,A by (64). Thus when Cα,W,A is infinite (65) holds trivially. When
Cα,W,A is finite (65) follows from (66) and max-min inequality. Thus we can assume Cα,W,A to be finite and prove the claims
about qα,W,A in order to prove the theorem.
(i) If Cα,W,A ∈ ℜ≥0 and limı→∞ Iα p (ı) ;W = Cα,W,A for a {p (ı) }ı∈Z+ ⊂ A, then {qα,p (ı) }ı∈Z+ is a Cauchy sequence in
P(Y) for the total variation metric: For any sequence of members of A satisfying limı→∞ Iα p (ı) ;W = Cα,W,A , let
{A(ı) }ı∈Z+ be a nested sequence of closed, convex, subsets of A defined as follows,
A(ı) , ch(∪ı=1 {p () }).
Furthermore, each A(ı) can be interpreted as a constraint set for a W (ı) with a finite input set X(ı) defined as follows,
X(ı) , {x ∈ X : ∃ ∈ {1, . . . , ı} such that p () (x ) > 0}.
With a slight abuse of notation we use the symbol A(ı) not only for a subset of P(X) but also for the corresponding
subset of P(X(ı) ). For any ı ∈ Z+ , there exists a unique qα,W (ı),A(ı) satisfying equation (67) by Lemma 19. Furthermore,
A() ⊂ A(ı) for any ı, ∈ Z+ such that ≤ ı. In order to bound qα,p () − qα,p (ı)
the triangle inequality for qα,p () , qα,p (ı) and qα,W (ı),A(ı)
for positive integers < ı we use
qα,p () − qα,p (ı) ≤ qα,p () − qα,W (ı),A(ı) + qα,p (ı) − qα,W (ı),A(ı) .
Let us proceed with bounding qα,p () − qα,W (ı),A(ı) and qα,p (ı) − qα,W (ı),A(ı)
r
(a)
qα,p () − qα,W (ı),A(ı) ≤
(b) q
2
α∧1 Dα∧1
from above.
qα,p () qα,W (ı),A(ı)
q
Dα W k qα,W (ı),A(ı) p () − Iα p () ;W (ı)
q
(c) q
2
≤ α∧1
Cα,W (ı) ,A(ı) − Iα p () ;W (ı)
q
(d) q
2
≤ α∧1
Cα,W,A − Iα p () ;W
≤
2
α∧1
41
(208)
()
(ı)
where (a) follows from Lemma 2, (b) follows from Lemma 13-(b,c,d), (c) follows from
19
Lemma
because p ∈ A , and
()
(ı)
()
(d) follows from the identities Cα,W (ı) ,A(ı) = Cα,W,A(ı) ≤ Cα,W,A and Iα p ;W
= Iα p ;W . We can obtain a similar
bound on qα,p (ı) − qα,W (ı),A(ı) . Then {qα,p (ı) } is a Cauchy sequence by (208) because lim Iα p () ;W = Cα,W,A .
→∞
(ii) If Cα,W,A ∈ ℜ≥0 , then ∃!qα,W,A ∈ P(Y) satisfying limı→∞ qα,W,A −qα,p (ı) = 0 for all {p (ı) }ı∈Z+ ⊂ A such that
limı→∞ Iα p (ı) ;W = Cα,W,A : Note that M(Y) is a complete metric space for the total variation metric, i.e. every
Cauchy sequence has a unique limit point in M(Y), because M(Y) is a Banach space for the total variation topology
[21, Thm. 4.6.1]. Then {qα,p (ı) }ı∈Z+ has a unique limit point qα,p ∗ in M(Y). Since P(Y) is a closed set for the total
variation topology and ∪ı∈Z+ qα,p (ı) ⊂ P(Y), then qα,p ∗ ∈ P(Y) by [37, Thm 2.1.3].
We have established the existence of a unique limit point qα,p ∗ , for any {p (ı) }ı∈Z+ ⊂ A satisfying limı→∞ Iα p (ı) ;W =
p (ı) }ı∈Z+ ⊂ A satisfying limı→∞ Iα pe(ı) ;W =
Cα,W,A . This, however, implies limı→∞ qα,ep (ı) − qα,p ∗ = 0 for any {e
(ı)
(ı)
p (ı) }ı∈Z+ ⊂ A
p }ı∈Z+ to obtain a new sequence {b
Cα,W,A because we can interleave
the elements of {p }ı∈Z+ and {e
(ı)
satisfying limı→∞ Iα pb ;W = Cα,W,A for which {qα,bp (ı) } is a Cauchy sequence. Then qα,W ,A = qα,p ∗ .
(ı)
(iii) qα,W,A satisfies the equality given
in (66): For any p ∈(ı)A, let us consider any sequence {p }ı∈Z+ ⊂ A satisfying
(ı)
(1)
p = p and limı→∞ Iα p ;W = Cα,W,A . Then p ∈ A for all ı ∈ Z+ . Using Lemma 19 we get
Dα W k qα,W (ı),A(ı) p ≤ Cα,W (ı) ,A(ı)
∀ı ∈ Z+ .
(209)
Since W (ı) has a finite input set, ∃e
p (ı) ∈ A(ı) satisfying Iα pe(ı) ;W (ı) = Cα,W (ı) ,A(ı)
and qα,ep (ı) = qα,W (ı),A(ı) by
(ı)
(ı)
≥ Iα p (ı) ;W (ı) and consequently limı→∞ Iα pe(ı) ;W = Cα,W,A . We have already
Lemma 19. Then Iα pe ;W
established that for such a sequence qα,ep (ı) → qα,W,A in the total variation topology, and hence in the topology of
setwise convergence. Then the lower semicontinuity of the Rényi divergence in its arguments for the topology of setwise
convergence, i.e. Lemma 3, the identity Cα,W (ı) ,A(ı) = Cα,W,A(ı) ≤ Cα,W,A and (209) imply
Dα (W k qα,W,A | p) ≤ Cα,W,A
∀p ∈ A.
On the other hand Dα (W k qα,W,A | p) ≥ Iα (p;W) and supp∈A Iα (p;W) = Cα,W,A by definition. Thus (66) holds.
Proof of Lemma 20. As a result of Lemma 13 we have
Dα∧1 (qα,p k qα,W,A ) ≤ Cα,W,A − Iα (p;W)
Then as a result of Lemma 2,
kqα,p − qα,W,A k ≤
q
∀p ∈ A.
2(Cα,W,A −Iα(p;W))
α∧1
∀p ∈ A.
Thus {qα,p (ı) }ı∈Z+ is a Cauchy sequence with the limit point qα,W,A for any sequence of input distributions {p (ı) }ı∈Z+ ⊂ A
satisfying limı→∞ Iα p (ı) ;W = Cα,W,A .
Proof of Lemma 21. As a result Lemma 13 we have,
supp̃∈A Dα (W k q| p̃) ≥ Dα (W k q| p)
∀p ∈ A
≥ Iα (p;W) + Dα∧1 (qα,p k q)
∀p ∈ A.
(210)
Let {p (ı) }ı∈Z+ ⊂ A be a sequence such that limı→∞ Iα p (ı) ;W = Cα,W,A . Then {qα,p (ı) } → qα,W,A in total variation
topology and hence in the topology of set wise convergence by Lemma 20. On the other hand, the Rényi divergence is lower
semicontinuous in its arguments for the topology of setwise convergence by Lemma 3. Then
i
h
(211)
lim inf ı→∞ Iα p (ı) ;W + Dα∧1 qα,p (ı) q ≥ Cα,W,A + Dα∧1 (qα,W,A k q) .
(71) follows from (210) and (211).
Proof of Lemma 22. Note that as a result of (72) and max-min inequality we have
Cα,W,A ≤ inf V ∈P(Y|X) supp∈A
α
1−α D1 (V k W | p)
+ I1 (p;V) .
Hence, (73) holds trivially whenever Cα,W,A = ∞ and (74) implies (73) whenever Cα,W,A ∈ ℜ≥0 .
In order to establish (74) assuming Cα,W,A ∈ ℜ≥0 , first note that whenever Cα,W,A ∈ ℜ≥0 there exists a unique qα,W ,A
satisfying (69) by Lemma 20. Then as a result of Definitions 1, 2, 3, 4 and Lemma 10 we have
q
q
α
D1 Wα α,W,A W p + D1 Wα α,W,A qα,W,A p .
Dα (W k qα,W,A | p) = 1−α
42
Then using Lemma 13-(b) and Lemma 2 we get,
X
q
q
q
Wα α,W,A W p + I1 p;Wα α,W,A + D1
p(x )Wα α,W,A (x ) qα,W,A
x
q
q
α
≥ 1−α
D1 Wα α,W,A W p + I1 p;Wα α,W,A
Dα (W k qα,W,A | p) =
α
1−α D1
∀p ∈ A.
Thus (69) and Lemma 20 implies that
Cα,W,A ≥ supp∈A
α
1−α D1
q
q
Wα α,W,A W p + I1 p;Wα α,W,A .
On the other hand, using the Csiszár’s form for the Augustin information given in (41) we get,
q
qα,W,A
α
α
W p + I1 p;Wα α,W,A ≥ inf 1−α
D1 (V k W | p) + I1 (p;V)
1−α D1 Wα
(212)
V
= Iα (p;W)
∀p ∈ P(X).
Then (74) follows from the definition of Cα,W,A and (212).
Proof of Lemma 23.
(23-a) Cα,W,A is nondecreasing and lower semicontinuous because it is the pointwise supremum of Iα (p;W) for p ∈ A and
Iα (p;W) is nondecreasing and continuous in α by Lemma 17-(c).
(23-b) 1−α
α Iα (p;W) is nonincreasing and continuous in α on ℜ+ for all p ∈ P(X) by Lemma 17-(b). Furthermore,
1−α
α Cα,W,A
= supp∈A
1−α
α Iα (p;W)
∀α ∈ (0, 1).
Then 1−α
α Cα,W,A is nonincreasing and lower semicontinuous in α on (0, 1) because the pointwise supremum of a family
of nonincreasing (lower semicontinuous) functions is nonincreasing (lower semicontinuous). Thus 1−α
α Cα,W,A and Cα,W,A
are both continuous from the right on (0, 1). On the other hand Cα,W,A and 1−α
α Cα,W,A are both continuous from the
left on (0, 1) because Cα,W,A is nondecreasing and lower semicontinuous on (0, 1) by part (a). Consequently, Cα,W,A
and 1−α
α Cα,W,A are both continuous on (0, 1).
(23-c) (α − 1)Iα (p;W) is convex in α on ℜ+ by Lemma 17-(a). Furthermore,
(α − 1)Cα,W,A = supp∈A (α − 1)Iα (p;W)
∀α ∈ (1, ∞).
Then (α−1)Cα,W,A is convex in α because the pointwise supremum of a family of convex functions is convex.
(23-d) Cα,W,A is continuous in α on (0, 1) by part (b). Furthermore, Cα,W,A is continuous from the left because it is nondecreasing
and lower semicontinuous. Thus Cα,W,A is continuous in α on (0, 1]. If χW ,A = 1 we are done.
If χW ,A > 1, then (α − 1)Cα,W,A is finite and convex in α on [1, χW ,A ) by part (c) and the definition of χW ,A . Then
(α − 1)Cα,W,A is continuous in α on (1, χW ,A ) by [20, Thm. 6.3.3]. The continuity of (α − 1)Cα,W,A on (1, χW ,A )
implies the continuity of Cα,W,A on (1, χW ,A ). Furthermore, Cα,W,A is continuous from the left because Cα,W,A is
nondecreasing and lower semicontinuous. Hence, Cα,W,A is continuous in α on (1, χW ,A ], as well.
(23-e) As a result of part (d), we only need to prove the continuity of Cα,W,A from the right at α = 1 when χW ,A > 1. As a
result of [14, (30)] we have
Iαg (p;W) ≤ I1g (p;W) +
g
8(α−1) η−1
η Iη(p;W)
ǫ2 e2 e
∀ǫ ∈ (0, η−1
η ), α ∈ [1, (1 − ǫ)η].
On the other hand I1 (p;W) = I1g (p;W) and Iα (p;W) ≤ Iαg (p;W) for α > 1. Then,
Cα,W,A ≤ C1,W,A +
g
8(α−1) η−1
η supp∈A Iη (p;W)
ǫ2 e2 e
∀ǫ ∈ (0, η−1
η ), α ∈ [1, (1 − ǫ)η].
Thus Cα,W,A is continuous at α = 1 from the right because Cα,W,A ≥ C1,W,A .
Proof of Lemma 24. Note that Cα,W,A is finite for all α ∈ (0, η] by Lemma 23 Then there exists a unique order α Augustin
center, qα,W,A , for all α ∈ (0, η] by Theorem 1. We apply Lemma 21 for q = qφ,W ,A to get
supp∈A Dα (W k qφ,W ,A | p) ≥ Cα,W,A + Dα∧1 (qα,W,A k qφ,W ,A ) .
(213)
Note that Dα (W k qφ,W ,A | p) is nondecreasing in α for all p ∈ A, because Dα (W (x )k qφ,W ,A ) is, by Lemma 8. Then,
Dφ (W k qφ,W ,A | p) ≥ Dα (W k qφ,W ,A | p)
∀p ∈ A, φ ∈ [α, η].
(214)
∀φ ∈ (0, η].
(215)
On the other hand, by (66) of Theorem 1 we have
supp∈A Dφ (W k qφ,W ,A | p) = Cφ,W ,A
(75) follows from (213), (214), and (215).
43
Using Lemma 2 together with (75) we get
q
2
(Cφ,W,A − Cα,W,A )
kqφ,W ,A − qα,W,A k ≤ α∧1
∀α, φ such that 0 < α < φ ≤ η.
(216)
Then the continuity qα,W,A in α for the total variation topology on P(Y) follows from the continuity Cα,W,A in α on I.
Proof of Lemma 25. We analyze the two inequalities in (76) separately.
(ı)
• supı∈T Cα,W,A(ı) ≤ Cα,W,A : Note that Cα,W,A(ı) ≤ Cα,W,A by definition because A
⊂ A. Thus Cα,W,A is bounded
from below by supı∈T Cα,W,A(ı) , as well.
– If Cα,W,A(ı) = Cα,W,A , then qα,W,A = qα,W,A(ı) because using Theorem 1 for (a), Lemma 21 for (b) and Lemma 2
for (c) we get
(a)
Cα,W,A ≥ supp∈A(ı) Dα (W k qα,W,A | p)
(b)
≥ Cα,W,A(ı) + Dα∧1 qα,W,A(ı) qα,W,A
(c)
≥ Cα,W,A(ı) +
•
α∧1
2
qα,W,A(ı) − qα,W,A
2
.
If Cα,W,A(ı) = Cα,W,A and qα,W,A(ı) = qα,W,A , then supp∈A Dα W k qα,W,A(ı) p ≤ Cα,W,A(ı) by Theorem 1.
– If supp∈A Dα W k qα,A(ı) p ≤ Cα,W,A(ı) , then Cα,W,A ≤ Cα,W,A(ı) by (64) and Theorem 1. On the other hand,
Cα,W,A ≥ Cα,W,A(ı) because A(ı) ⊂ A. Hence, Cα,W,A = Cα,W,A(ı) .
P
P
C
C
Cα,W,A ≤ ln ı∈T e α,W,A(ı) : If T is an infinite set, then the inequality holds trivially because ln ı∈T e α,W,A(ı) = ∞.
W
C
Thus we assume T to be a finite set for the rest of the proof. Let µ be µ = ı∈T e α,W,A(ı) qα,W,A(ı) . Then as a result of
Lemma 1 we have
µ
supp∈A(ı) Dα W k kµk
p = supp∈A(ı) Dα (W k µ| p) + ln kµk
≤ supp∈A(ı) Dα W k qα,W,A(ı) p − Cα,W,A(ı) + ln kµk
∀ı ∈ T.
Since supp∈A(ı) Dα W k qα,W,A(ı) p = Cα,W,A(ı) by hypothesis, we have
µ
p ≤ ln kµk
∀ı ∈ T.
supp∈A(ı) Dα W k kµk
Then using (64) and Theorem 1 we get
µ
p
Cα,W,A ≤ supp∈A Dα W k kµk
µ
p
= supı∈T supp∈A(ı) Dα W k kµk
≤ ln kµk
X
≤ ln
ı∈T
e
Cα,W,A(ı)
.
P
P
C
C
– If qα,W,A(ı) and qα,W,A() are not singular for some ı 6= , then kµk < ı∈T e α,W,A(ı) . Thus Cα,W,A < ln ı∈T e α,W,A(ı) .
P
Cα,W,A(ı)
, then qα,W,A(ı) ⊥ qα,W,A() for all ı 6= .
Consequently, if Cα,W,A = ln ı∈T e
P|T|+1
– If qα,W,A(ı) ⊥ qα,W,A() for all ı 6= , then any s ∈ P(Y)
P can be written as s = ı=1 sı where sı are finite
measures such that sı ≺qα,W,A(ı) for ı ∈ T and s|T|+1 ⊥ ( ı∈T qα,W,A(ı) ) by the Lebesgue decomposition theorem
[20, 5.5.3]. Using Lemmas 1, 2, and 21 we get
supp∈A(ı) Dα (W k s| p) ≥ Cα,W,A(ı) + Dα∧1 qα,W,A(ı) s
= Cα,W,A(ı) + Dα∧1 qα,W,A(ı) kssıı k − ln ksı k
≥ Cα,W,A(ı) − ln ksı k.
supp∈A Dα (W k s| p) = maxı∈T supp∈A(ı) Dα (W k s| p) because Dα (W k s| p) is linear in p and A = ch(∪ı∈T A(ı) ).
P|T|
Then using ı=1 ksı k ≤ ksk = 1 we get,
C
α,W,A(ı)
supp∈A Dα (W k s| p) ≥ maxı∈T ln e ksı k
X
C
≥ ln
e α,W,A(ı)
ı∈T
∀s ∈ P(Y).
P
C
Then Cα,W,A ≥ ln ı∈T e α,W,A(ı) by (64) and Theorem 1. Since we have already established the reverse inequality,
P
C
we have Cα,W,A = ln ı∈T e α,W,A(ı) .
44
P
Cα,W,A(ı)
. One can confirm by
We have proved that if qα,W,A(ı) ⊥ qα,W,A() for all ı 6= , then Cα,W ,A =
ı∈T e
P
−C
+C
substitution that supp∈A(ı) Dα (W k s| p) ≤ Cα,W ,A for all ı ∈ T for s = ı∈T e α,W ,A α,W,A(ı) qα,W,A(ı) . On the other
hand, supp∈A Dα (W k s| p) = maxı∈T supp∈A(ı) Dα (W k s| p) because Dα (W k s| p) is linear in p. Then s is the unique
order α Augustin center by Theorem 1.
Proof of Lemma 26. Let α be any fixed positive real order. Then as a result of Lemma 14 we have
X
Cα,W,A ≥
Cα,Wt ,At .
t∈T
(217)
On the other hand, Cα,W,chA ≥ Cα,W,A because A ⊂ chA. Then
X
Cα,W,chA ≥
(218)
Cα,Wt ,At .
t∈T
P
We proceed with proving Cα,W,chA ≤ t∈T Cα,Wt ,At . If there exists a t ∈ T such that Cα,Wt ,At = ∞, then the inequality
holds trivially. Else Cα,Wt ,At ∈ ℜ≥0 for all t ∈ T and by Lemma 20 there exists a unique qα,Wt ,At for each t ∈ T such that
Let q be q =
Hence,
Q⊗
∀pt ∈ At .
Dα (Wt k qα,Wt ,At | pt ) ≤ Cα,Wt ,At
t∈T qα,Wt ,At .
Then as a result of Tonelli-Fubini theorem [20, 4.4.5] we have
X
Dα (W (x )k q) =
∀x ∈ X.
Dα (Wt (x )k qα,Wt ,At )
t∈T
Dα (W k q| p) ≤
X
t∈T
Cα,Wt ,At
Since Dα (W k q| p) is linear in p the preceding inequality implies
X
Dα (W k q| p) ≤
Cα,Wt ,At
t∈T
Since Dα (W k q| p) ≥ Iα (p;W) by P
definition, (217) and (219) imply Cα,W,A =
(218) and (220) imply Cα,W,chA = t∈T Cα,Wt ,At and qα,W,chA = q.
P
t∈T
∀p ∈ A.
(219)
∀p ∈ chA.
(220)
Cα,Wt ,At and qα,W,A = q. Similarly,
E. Proofs of Lemmas on the Cost Constrained Problem
Proof of Lemma 27.
(27-a) If ̺1 ≤ ̺2 , then Cα,W,̺1 ≤ Cα,W,̺2 because A(̺1 ) ⊂ A(̺2 ). Thus Cα,W,̺ is nondecreasing in ̺.
Let ̺β = β̺1 + (1 − β)̺0 ; then (βp1 + (1 − β)p0 ) ∈ A(̺β ) for any p1 ∈ A(̺1 ) and p0 ∈ A(̺0 ). Hence, using the
concavity of the Augustin information in its input distribution established in Lemma 15 we get
Cα,W,̺β ≥ supp1 ∈A(̺1 ),p0 ∈A(̺0 ) Iα (βp1 + (1 − β)p0 ;W)
≥ supp1 ∈A(̺1 ),p0 ∈A(̺0 ) βIα (p1 ;W) + (1 − β)Iα (p0 ;W)
= βCα,W,̺1 + (1 − β)Cα,W,̺0 .
Thus Cα,W,̺ is concave in ̺.
Now let us proceed by proving that if Cα,W,̺0 = ∞ for a ̺0 ∈ intΓρ ; then Cα,W,̺ = ∞ for all ̺ ∈ intΓρ . Note that
any point ̺ in intΓρ can be written as ̺ = β̺1 + (1 − β)̺0 for some β ∈ (0, 1) and ̺1 ∈ intΓρ because intΓρ is a
convex open subset ℜℓ . Then Cα,W,̺ = ∞ follows from the concavity of Cα,W,̺ .
If Cα,W,̺ is finite on intΓρ , then Cα,W,̺ is continuous on intΓρ by [20, Thm. 6.3.4] because intΓρ is a convex open
subset ℜℓ and (−Cα,W,̺ ) is a convex function on intΓρ .
(27-b) Let us extend the definition of Cα,W,̺ from ℜℓ≥0 to ℜℓ by setting Cα,W,̺ to −∞ for all ̺ ∈ ℜℓ \ ℜℓ≥0 . Then (−Cα,W,̺) is a
proper convex function, i.e. (−Cα,W,̺) : ℜℓ → (−∞, ∞] is a convex function and ∃̺ such that (−Cα,W,̺) < ∞. Furthermore,
intΓρ is also the interior of the effective domain of the extended function. Hence the sub-differential ∂(−Cα,W,̺ ) is nonempty and compact by [39, Proposition 4.4.2]. Then (83) follows from the fact that the epigraph of a convex function lies
above the tangent planes drawn at any point. The non-negativity of the components of λα,W,̺ follows from the monotonicity
of Cα,W,̺ in ̺.
(27-c) If Cα̃,W ,̺˜ = ∞ for a α̃ ∈ (0, 1) and ̺˜ ∈ intΓρ , then Cα,W,̺˜ = ∞ for all α ∈ (0, 1) by Lemma 23-(a,b). Therefore,
Cα,W,̺ = ∞ for all α ∈ (0, 1) and ̺ ∈ intΓρ by part (a).
In order to prove the continuity when Cα,W,̺ is finite, note that as a result of the triangle inequality we have
|Cα1 ,W,̺1 − Cα2 ,W,̺2 | ≤ |Cα1 ,W,̺1 − Cα1 ,W,̺2 | + |Cα1 ,W,̺2 − Cα2 ,W,̺2 |.
45
The first term converges to zero as ̺2 → ̺1 as a result of the continuity of the Augustin capacity in the constraint
established in part (a). The second term converges to zero as α2 → α1 because of (221) established in the following. Thus
Cα,W,̺ is continuous in the pair (α, ̺).
Using the monotonicity of Cα,W,̺ and 1−α
α Cα,W,̺ established in Lemma 23-(a) and Lemma 23-(b) we get
|α2 −α1 |
(α1 ∧α2 )(1−α1 ∨α2 ) Cα1 ,W,̺2 .
|Cα1 ,W,̺2 − Cα2 ,W,̺2 | ≤
Thus using (83) to bound Cα1 ,W,̺2 we get
|Cα1 ,W,̺2 − Cα2 ,W,̺2 | ≤
|α2 −α1 |
(α1 ∧α2 )(1−α1 ∨α2 ) (Cα1 ,W,̺1
+ |λα1 ,W,̺1 · (̺2 − ̺1 )|+ ).
(221)
In order to prove the continuity of the Augustin center, note that by the triangle inequality we have
kqα1 ,W,̺1 − qα2 ,W,̺2 k ≤ kqα1 ,W,̺1 − qα1 ,W,̺2 k + kqα1 ,W,̺2 − qα2 ,W,̺2 k.
Using first Lemmas 2, 24, and then (221) we get
r
2|Cα1 ,W,̺2 −Cα2 ,W,̺2 |
kqα1 ,W,̺2 − qα2 ,W,̺2 k ≤
α1 ∧α2
q
+
2 −α1 |
(Cα1 ,W,̺1 + |λα1 ,W,̺1 · (̺2 − ̺1 )| ).
≤ (α1 ∧α2|α
2
2 ) (1−α1 ∨α2 )
(222)
(223)
In order to bound kqα1 ,W,̺1 − qα1 ,W,̺2 k, we use triangle inequality once more
kqα1 ,W,̺1 − qα1 ,W,̺2 k ≤ kqα1 ,W,̺1 − qα1 ,W ,̺∨ k + kqα1 ,W ,̺∨ − qα1 ,W,̺2 k
where ̺∨ = ̺1 ∨ ̺2 , i.e. ̺ı∨ = ̺ı1 ∨ ̺ı2 for each ı ∈ {1, . . . , ℓ}.
On the other hand by Lemma 21 we have
supp∈A(̺1 ) Dα1 (W k qα1 ,W ,̺∨ | p) ≥ Cα1 ,W,̺1 + D1∧α1 (qα1 ,W,̺1 k qα1 ,W ,̺∨ ) .
Since A(̺1 ) ⊂ A(̺∨ ), using Theorem 1 and Lemma 2 we get,
Cα1 ,W ,̺∨ − Cα1 ,W,̺1 ≥
α1
2 kqα1 ,W,̺1
2
− qα1 ,W ,̺∨ k .
Repeating the same analysis for kqα1 ,W,̺2 − qα1 ,W ,̺∨ k and bounding Cα1 ,W ,̺∨ using (83) we get
q
q q
2
kqα1 ,W,̺1 − qα1 ,W,̺2 k ≤ α1
Cα1 ,W,̺1 − Cα1 ,W,̺2 + λα1 ,W,̺1 · (̺∨ − ̺1 ) + λα1 ,W,̺1 · (̺∨ − ̺1 )
(224)
The continuity of qα,W,̺ for the total variation topology on P(Y) follows from (222), (223), (224) and the continuity of
Augustin capacity as a function of the constraint established in part (a).
Proof of Lemma 28. Let B(̺) be
n
o
̺t ≤ ̺, ̺t ∈ Γρt .
t∈T
P
Note that if B(̺) = ∅, then ̺ ∈
/ Γρ and Cα,W ,̺ = −∞. On the other hand, t∈T Cα,Wt ,̺t is minus infinity for (̺1 , . . . , ̺|T| )’s
that are outside B(̺) by the convention stated in the lemma. Thus (84) holds for ̺ ∈ ℜℓ≥0 \Γρ case and the constraints ̺t ∈ ℜℓ≥0
can be replaced by ̺t ∈ Γρt for ̺ ∈ Γρ case in (84).
Furthermore, as a result of Lemma 26 for any (̺1 , . . . , ̺|T| ) ∈ B(̺) we have
X
Cα,W ,Q× At (̺t ) =
Cα,Wt ,̺t .
t ∈T
t∈T
Q
On the other hand ( ×
t∈T At (̺t )) ⊂ A(̺) for any (̺1 , . . . , ̺|T| ) ∈ B(̺). Thus as a result of Lemma 25 we have
nX
o
X
Cα,W ,̺ ≥ sup
Cα,Wt ,̺t :
̺t ≤ ̺, ̺t ∈ Γρt .
B(̺) ,
(̺1 , . . . , ̺|T| ) :
X
t∈T
t∈T
For deriving the reverse inequality note that for all p ∈ P(X) and t ∈ T there is a corresponding pt ∈ P(Xt ), which is simply
the probability measure generated by the Xt components of x ’s with p(x ) > 0. Furthermore,
X
Y⊗
qα,pt p =
Dα (Wt k qα,pt | pt ) .
Dα W k
t∈T
t∈T
P
P
Thus Iα (p;W) ≤ t∈T Iα (pt ;Wt) by Lemma 13. On the other hand, Ep [ρ] = t∈T Ept [ρt ] and Ept [ρt ] ∈ Γρt . Hence,
X
Iα (pt ;Wt)
supp:Ep[ρ]≤̺ Iα (p;W) ≤ supp1 ,...,p|T| :Pt ∈T Ept[ρt ]≤̺
t∈T
nX
o
X
= sup
Cα,Wt ,̺t :
̺t ≤ ̺, ̺t ∈ Γρt .
t∈T
46
t∈T
Q×
Q×
Thus (84) holds. In addition, A(̺) can be interpreted as the union of ( t∈T At (̺t )) and A(̺) \ ( t∈T At (̺t )). Therefore, if
Q⊗
P
there exists a {̺t }t∈T such that Cα,W,̺ = t∈T Cα,Wt ,̺t and Cα,W,̺ < ∞, then qα,W,̺ = t∈T qα,Wt,̺t by Lemma 25.
Proof of Lemma 29.
λ
λ
(a) Cα,W
is convex, nonincreasing, and lower semicontinuous in λ because Cα,W
is the pointwise supremum of such functions
as a result of (90).
λ
λ
Since Cα,W
is convex it is continuous on the interior of {λ ∈ ℜℓ≥0 : Cα,W
< ∞} by [20, Thm. 6.3.4]. The interior of
λ−ǫ1
ℓ
λ
ℓ
λ
{λ ∈ ℜ≥0 : Cα,W < ∞} is {λ ∈ ℜ≥0 : ∃ǫ > 0 s.t. Cα,W < ∞} because Cα,W
is nonincreasing in λ.
(b) Note that Cα,W,̺ = supp∈P(X) ξα,p (̺) as a result of (82) and (87). Then as a result of (86), we have
Cα,W,̺ = supp∈P(X) inf λ≥0 Iαλ (p;W) + λ · ̺
∀̺ ∈ ℜℓ≥0 .
(225)
If X is finite, then P(X) is compact. Furthermore, using (205) together with triangle inequality we get
kp1 −p2 k
2k
2k
Iαλ (p2 ;W) − Iαλ (p1 ;W) ≤ kp1 −p
+ kp1 −p
ln
|X|
+
ℏ
maxx ∈X λ · ρ(x ).
2
2
2
Then Iαλ (p;W) + λ · ̺ is continuous in p on P(X). On the other hand, Iαλ (p;W) + λ · ̺ is concave in p by Lemma 15 and
convex and continuous in λ. Thus we can change the order of the infimum and supremum in (225) —using the Sion’s
λ
minimax theorem, [40], [41]— and Cα,W,̺ = inf λ≥0 Cα,W
+ λ · ̺ by (89).
λ
(c) If ̺ ∈ intΓρ and Cα,W,̺ is infinite, then Cα,W,̺ = inf λ≥0 Cα,W
+ λ · ̺ follows from (91) trivially.
If ̺ ∈ intΓρ and Cα,W,̺ is finite, then there exists a non-empty, convex, and compact set of λα,W,̺ ’s satisfying (83) by
Lemma 27-(b). Furthermore, for any λα,W,̺ satisfying (83) as a result of (90) we have
λ
α,W,̺
= sup̺≥0
Cα,W ,̺˜ − λα,W,̺ · ̺˜
Cα,W
˜
≤ Cα,W ,̺ − λα,W,̺ · ̺.
λ
α,W,̺
+ λα,W,̺ · ̺ by (91).
Then Cα,W ,̺ = Cα,W
λ
λ
λ
. Furthermore,
(d) Note that Iα (p;W) ≤ Cα,W by definition. Hence lim supı→∞ Iαλ p (ı) ;W ≤ Cα,W
Iαλ (p;W) ≥ Iα (p;W) − λ · ̺
∀p ∈ A(̺).
λ
λ
λ (ı)
λ (ı)
.
Thus lim inf ı→∞ Iα p ;W ≥ Cα,W,̺ − λ · ̺ = Cα,W , as well. Thus limı→∞ Iα p ;W = Cα,W
Proof of Lemma 30.
λ
(i) ∀α ∈ ℜ+ ∃e
p ∈ P(X) such that Iαλ (e
p ;W) = Cα,W
: Using (205) together with triangle inequality we get
2k
2k
2k
Iαλ (p2 ;W) − Iαλ (p1 ;W) ≤ kp1 −p
+ kp1 −p
ln |X| + ℏ kp1 −p
maxx ∈X λ · ρ(x )
2
2
2
Then Iαλ (p;W) is continuous in p on P(X). On the other hand P(X) is compact because X is a finite set. Then there
exists a e
p ∈ P(X) such that Iαλ (e
p ;W) = supp∈P(X) Iαλ (p;W) by the extreme value theorem, [37, 27.4].
λ
λ
λ
(ii) If α ∈ ℜ+ and Iα (e
p ;W) = Cα,W , then Dα (W k qα,ep | p) − λ · Ep [ρ] ≤ Cα,W
for all p ∈ P(X): Let p be any member of
1
ı−1
(ı)
P(X) and p be ı pe + ı p for ı ∈ Z+ . Then by Lemma 13
i
i
h
h
Dα W k qα,p (ı) pe − λ · Eep [ρ] + 1ı Dα W k qα,p (ı) p − λ · Ep [ρ]
Iαλ p (ı) ;W = ı−1
ı
h
h
i
i
λ
1
≥ ı−1
I
(e
p
;W)
+
D
D
−
λ
·
E
[ρ]
∀ı ∈ Z+ .
q
k
q
W
k
q
+
(ı)
(ı) p
α∧1
α
p
α,e
p
α,p
α,p
α
ı
ı
λ
λ
, Iαλ (e
p ;W) = Cα,W
, and Dα∧1 qα,ep k qα,p (ı) ≥ 0 we get
Then using Iαλ p (ı) ;W ≤ Cα,W
λ
(226)
Cα,W
≥ Dα W k qα,p (ı) p − λ · Ep [ρ] .
λ
λ
On the other hand, using Iαλ p (ı) ;W ≤ Cα,W
, Iαλ (e
p ;W) = Cα,W
and Dα W k qα,p (ı) p ≥ 0 we get
λ
Cα,W
+λ·Ep[ρ]
ı−1
∀ı ∈ Z+ .
q
k
q
≥
D
(ı)
α∧1
α,e
p
α,p
ı
ı
Then using Lemma 2 we get
r
λ
2 Cα,W +λ·Ep[ρ]
α∧1
ı−1
≥ qα,ep − qα,p (ı)
47
∀ı ∈ Z+ .
Thus qα,p (ı) converges to qα,ep in the total variation topology and hence in the topology of setwise convergence. Since
the Rényi divergence is lower semicontinuous in the topology of setwise convergence by Lemma 3, we have
(227)
lim inf ı→∞ Dα W k qα,p (ı) p ≥ Dα (W k qα,ep | p) .
λ
Then the inequality Dα (W k qα,ep | p) − λ · Ep [ρ] ≤ Cα,W
follows from (226) and (227).
λ
λ
λ
(iii) If α ∈ ℜ+ , then ∃!qα,W ∈ P(Y) satisfying (97) such that qα,p = qα,W
for all p ∈ P(X) satisfying Iαλ (p;W) = Cα,W
: If
λ
λ
λ
Iα (p;W) = Cα,W for a p ∈ P(X) such that qα,p 6= qα,ep , then Dα (W k qα,ep | p) − λ · Ep [̺] > Cα,W by Lemma 13-(b,c,d)
λ
and Lemma 2. This is a contradiction because we have already established that Dα (W k qα,ep | p) − λ · Ep [̺] ≤ Cα,W
for
λ
λ
all p ∈ P(X). Then qα,p = qα,ep for any p ∈ P(X) satisfying Iα (p;W) = Cα,W .
Proof of Theorem 2. First note that (93) implies (94) and (95) implies (96). Furthermore, the left hand side of (93) is equal
λ
λ
λ
to Cα,W
by (92). Thus when Cα,W
is infinite (93) holds trivially by max-min inequality. When Cα,W
is finite (93) follows
λ
λ
from (95) and max-min inequality. Thus we can assume Cα,W to be finite and prove the claims about qα,W
, in order to prove
the theorem.
λ
λ
, then {qα,p (ı) }ı∈Z+ is a Cauchy sequence in P(Y) for the total variation
(i) If Cα,W
< ∞ and limı→∞ Iαλ p (ı) ;W = Cα,W
λ
(ı)
, let us consider a sequence of
metric: For any sequence {p }ı∈Z+ ⊂ P(X) satisfying limı→∞ Iαλ p (ı) ;W = Cα,W
(ı)
(ı)
channels {W }ı∈Z+ whose input sets {X }ı∈Z+ form a nested sequence of finite subsets of X defined as follows,
X(ı) , {x ∈ X : ∃ ∈ {1, . . . , ı} such that p () (x ) > 0}.
λ
()
Then for any ı ∈ Z+ , there exists a unique qα,W
) ⊂ P(X(ı) ) for any
(ı) satisfying (97) by Lemma 30. Furthermore, P(X
ı, ∈ Z+ such that ≤ ı. In order to bound qα,p () − qα,p (ı)
λ
for qα,p () , qα,p (ı) and qα,W
(ı)
for positive integers < ı we use the triangle inequality
λ
λ
qα,p () − qα,p (ı) ≤ qα,p () − qα,W
+ qα,p (ı) − qα,W
(ı)
(ı) .
(228)
λ
λ
Let us proceed with bounding qα,p () − qα,W
and qα,p (ı) − qα,W
from above.
(ı)
(ı)
r
(a)
λ
2
λ
qα,p () − qα,W
≤
(ı)
α∧1 Dα∧1 qα,p () qα,W (ı)
r
q
(b)
≤
2
α∧1
≤
2
α∧1
(c) q
λ
() − I p () ;W (ı)
Dα W k qα,W
α
(ı) p
q
λ
λ () ;W (ı)
Cα,W
(ı) − Iα p
q
(d) q
2
λ
≤ α∧1
Cα,W
− Iαλ p () ;W
where (a) follows from Lemma 2, (b) follows from
13-(b,c,d),
(c) follows from Lemma 30 because p () ∈ P(X(ı) ),
Lemma
λ
λ
(ı)
λ ()
λ ()
= Iα p ;W and Cα,W
and (d) follows from the identities Iα p ;W
(ı) ≤ Cα,W . We can obtain a similar bound
λ
λ ()
λ
on qα,p (ı) − qα,W
;W = Cα,W
.
(ı) . Then {qα,p (ı) } is a Cauchy sequence as a result of (228) because lim→∞ Iα p
λ
λ
λ
(ii) If Cα,W
< ∞, then ∃!qα,W
∈ P(Y) satisfying limı→∞ qα,W
− qα,p (ı) = 0 for all {p (ı) }ı∈Z+ ⊂ P(X) such that
λ
: Note that M(Y) is a complete metric space for the total variation metric because M(Y)
limı→∞ Iαλ p (ı) ;W = Cα,W
is a Banach space for the total variation topology [21, Thm. 4.6.1]. Then {qα,p (ı) }ı∈Z+ has a unique limit point qα,p ∗
in M(Y). Since P(Y) is a closed set for the total variation topology and ∪ı∈Z+ qα,p (ı) ⊂ P(Y), then qα,p∗ ∈ P(Y), by
[37, Thm. 2.1.3].
We have established
of a unique limit point qα,p∗ , for any sequence {p (ı) }ı∈Z+ ⊂ P(X) satisfying
the existence
λ
λ (ı)
p (ı) }ı∈Z+ satisfying
limı→∞ Iα p ;W = Cα,W . This, however, implies limı→∞ qα,ep (ı) − qα,p∗ = 0 for any {e
(ı)
λ
(ı)
λ (ı)
p }ı∈Z+ to obtain a new sep ;W = Cα,W because we can interleave
the elements of {p }ı∈Z+ and {e
limı→∞ Iα e
λ
λ
quence {b
p (ı) }ı∈Z+ satisfying limı→∞ Iαλ pb(ı) ;W = Cα,W
for which {qα,bp (ı) } is a Cauchy sequence. Then qα,W
= qα,p ∗
λ
(ı)
(iii) qα,W satisfies the equality
in (95): For any p ∈ P(X), let us consider any sequence {p }ı∈Z+ satisfying p (1) = p
given
λ
λ (ı)
and limı→∞ Iα p ;W = Cα,W . Then p ∈ P(X(ı) ) for all ı ∈ Z+ . Using Lemma 30 we get
λ
∀ı ∈ Z+ .
(229)
Dα W k qα,W (ı) p − λ · Ep [ρ] ≤ Cα,W
(ı)
λ
λ
Since X(ı) is a finite set, ∃e
p (ı) ∈ P(X(ı) ) satisfying Iαλ pe(ı) ;W (ı) = Cα,W
(ı) and qα,e
p (ı) = qα,W (ı) by Lemma 30. Then
λ
(ı)
λ (ı)
(ı)
λ (ı)
λ (ı)
and consequently limı→∞ Iα pe ;W = Cα,W . We have already established that for such
p ;W
≥ Iα p ;W
Iα e
48
λ
a sequence qα,ep (ı) → qα,W
in the total variation topology, and hence in the topology of setwise convergence. Then the
lower semicontinuity of the Rényi divergence in its arguments for the topology of setwise convergence, i.e. Lemma 3,
λ
λ
the identity Cα,W
(ı) ≤ Cα,W , and (229) imply that
λ
λ
Dα W k qα,W
p − λ · Ep [ρ] ≤ Cα,W
∀p ∈ P(X).
λ
λ
p − λ · Ep [ρ] ≥ Iαλ (p;W) and Cα,W
= supp∈P(X) Iαλ (p;W) by the definitions of
On the other hand Dα W k qα,W
λ
λ
Iα (p;W), Iα (p;W), and Cα,W . Thus (95) holds.
Proof of Lemma 31. Let {p (ı) }ı∈Z+ ⊂ A(̺) be such that limı→∞ Iα p (ı) ;W = Cα,W,̺ . Then {qα,p (ı) }ı∈Z+ is a Cauchy
λ
by Lemma 29-(d). Then
sequence with the limit point qα,W,̺ by Theorem 1. On the other hand, limı→∞ Iαλ p (ı) ;W = Cα,W
λ
λ
{qα,p (ı) }ı∈Z+ is a Cauchy sequence with the limit point qα,W by Theorem 2. Hence qα,W,̺ = qα,W .
Proof of Lemma 32. For all p ∈ P(X) and t ∈ T there is a corresponding pt ∈ P(Xt ), which is simply the Xt marginal of p.
Furthermore,
X
Y⊗
qα,pt p =
Dα (Wt k qα,pt | pt )
Dα W k
t∈T
Xt∈T
Ep [ρ] =
Ept [ρt ] .
t∈T
P
Thus Iαλ (p;W) ≤ t∈T Iαλ (pt ;Wt) by Lemma 13 and the definition of the A-L information given in (85). Hence,
X
supp∈P(X) Iαλ (p;W) = supp1 ∈P(X1 )...,p|T| ∈P(X|T| )
Iαλ (pt ;Wt) .
t∈T
In order to establish
λ
qα,W
=
Q⊗
λ
t∈T qα,Wt ,
one can confirm by substitution that
Q⊗
λ
t∈T qα,Wt
satisfies (96)
Proof of Lemma 33.
(a) Note that as a result of Lemma 13-(c,d) and the definition of Iαλ (p;W) given in (85) we have
X
λ
D1 pk uα,p
= (α − 1)Iαλ (p;W) + ln
p(x̃ )e(1−α)Dα(W (x̃ )kqα,p )+(α−1)λ·ρ(x ) .
x̃
(230)
On the other hand as a result of (42), (102), and (106)
Z X
1/α
λ
λ
α
uα,p
(x )e(1−α)λ·ρ(x ) ( dWdν(x ) )α
Iαgλ uα,p
ln
ν(dy )
;W = α−1
x
Z
X
dqα,p
1
α
ln
p(x̃ )e(1−α)Dα(W (x̃ )kqα,p )+(α−1)λ·ρ(x ) .
= α−1
dν ν(dy) − α−1 ln
x̃
Then (111) follows from (230)
• Let us proceed with establishing (112) for α ∈ (0, 1). In order to do that we prove,
Iαgλ (u;W) +
1
α−1 D1 (pk u)
≤ Iαλ (p;W)
∀u ∈ P(X).
Preceding inequality together with (111) imply (112) for α ∈ (0, 1). Note that the inequality holds trivially when
p⊀u because D1 (pk u) is infinite in that case. Thus we are left with p≺u case. On the other hand, any u ∈ P(X)
can be written as u = uac + us where uac ≺p and us ⊥ p. Then
(i)
1−α
1
1
Iαgλ (u;W) + α−1
D1 (pk u) ≤ Dα u⊛W e α λ·ρ u ⊗ qα,p + α−1
D1 (pk u)
h
i
X
(ii)
1
1
≤ α−1
ln
D1 (pk uac )
uac (x )e(α−1)Dα(W (x )kqα,p )+(1−α)λ·ρ(x ) + α−1
x
i
h
X
(iii) 1
1
ac (x ) (α−1)Dα(W (x )kqα,p )+(1−α)λ·ρ(x )
+ α−1
= α−1 ln
D1 (pk uac )
p(x ) up(x
) e
x
i
h
X
(iv)
1
ac (x ) (α−1)Dα(W (x )kqα,p )+(1−α)λ·ρ(x )
1
+ α−1
D1 (pk uac )
p(x ) ln up(x
≤ α−1
) e
x
(v)
=
•
Iαλ (p;W) .
where (i) follows from (101), (ii) follows from (11) and the monotonicity of the natural logarithm function, (iii)
follows from uac ∼ p which holds because p≺u, (iv) follows from the Jensen’s inequality and the concavity of the
natural logarithm function, (v) follows from Lemma 13 and the definition of Iαλ (p;W) given in (85).
In order to prove (112) for α ∈ (1, ∞), we prove the following identity
Iαgλ (u;W) +
1
α−1 D1 (pk u)
≥ Iαλ (p;W)
49
∀u ∈ P(X).
The analysis is pretty much the same, except for the direction of the inequalities.
(b) Note that the order α R-G mean for the input distribution p and Lagrange multiplier λ is a fixed point of the order α
λ
Augustin operator for the input distribution aα,p
, i.e.
gλ
dTα,a λ (qα,p
)
α,p
dν
=
X
dq gλ
α,p 1−α (1−α)Dα(W (x )kqα,p )
)
e
a λ (x )( dWdν(x ) )α ( dν
x α,p
X
gλ
dqα,p
)1−α
= (α−1)I1gλ(p;W)
p(x )e(1−α)λ·ρ(x ) ( dWdν(x ) )α ( dν
=
gλ
x
α
e
gλ
dqα,p
dν .
gλ
λ
λ
;W = Dα W k qα,p
Consequently Iα aα,p
aα,p
by Lemma 13-(c,d). Then
•
gλ
(α−1)Dα(W (x )kqα,p )+(1−α)λ·ρ(x )
X λ
λ
1
D1 aα,p
p =
aα,p (x ) ln Pp(x )e (α−1)D W (x̃ )kq gλ +(1−α)λ·ρ(x̃ ) p(x
)
α(
α,p )
x
x̃ p(x̃ )e
X
gλ
λ
;W − ln
p(x̃ )e(α−1)Dα(W (x̃ )kqα,p )+(1−α)λ·ρ(x )
= (α − 1)Iαλ aα,p
x̃
λ
;W − Iαgλ (p;W)].
= (α − 1)[Iαλ aα,p
Let us proceed with establishing (114) for α ∈ (0, 1). In order to do that we prove,
Iαλ (a;W) −
1
α−1 D1 (ak p)
≥ Iαgλ (p;W)
∀a ∈ P(X).
Preceding inequality together with (113) imply (114) for α ∈ (0, 1). Note that the inequality holds trivially when
a⊀p because D1 (ak p) is infinite in that case. Thus we are left with a≺p case. On the other hand, for any a ∈ P(X),
p can be written as p = pac + ps where pac ≺a and ps ⊥ a. Then
Iαλ (a;W) −
(i)
1
1
α−1 D1 (ak p) = Dα (W k qα,a | a) − λ · Ea [ρ] − α−1 D1 (ak pac )
i
h
(ii) 1 X
ac (x ) (α−1)Dα(W (x )kqα,a )+(1−α)λ·ρ(x )
= α−1
e
a(x ) ln pa(x
)
x
(iii)
≥
(iv)
≥
(v)
1
α−1
ln
1
α−1
ln
X
X
x
x
pac (x )e(α−1)Dα(W (x )kqα,a )+(1−α)λ·ρ(x )
p(x )e(α−1)Dα(W (x )kqα,a )+(1−α)λ·ρ(x )
≥ Iαgλ (p;W) .
•
where (i) follows from (11) and Lemma 13, (ii) follows from pac ∼ a which holds because a≺p, (iii) follows from
the Jensen’s inequality and the concavity of the natural logarithm function, (iv) follows from the monotonicity of the
natural logarithm function, (v) follows from (11) and (101).
In order to prove (114) for α ∈ (1, ∞), we prove the following identity
Iαλ (a;W) −
1
α−1 D1 (ak p)
≤ Iαgλ (p;W)
∀a ∈ P(X).
The analysis is pretty much the same, except for the direction of the inequalities.
Proof of Lemma 34.
gλ
(i) ∃e
p ∈ P(X) such that Iαgλ (e
p ;W) = Cα,W
: Note that P(X) is compact because X is a finite set. If Iαgλ (p;W) is continuous
in p the existence of e
p follows from the extreme value theorem, [37, 27.4]. Thus we are left with establishing the
continuity.
First note that for any p1 and p0 there exist mutually singular probability mass functions s1 , s0 , and s∧ satisfying
0k
. Then applying first (104) and (105) we get.
p1 = (1 − δ)s∧ + δs1 and p0 = (1 − δ)s∧ + δs0 where δ = kp1 −p
2
i
h
λ
λ
gλ
λ
λ
gλ
1
Iαgλ (p1 ;W) = α−1
(231)
ln (1 − δ)e(α−1)[Iα (s∧ ;W)+Dα(qα,s∧ kqα,p1 )] + δe(α−1)[Iα (s1 ;W)+Dα(qα,s1 kqα,p1 )]
Since Iαgλ (s∧ ;W) ≤ Iαg (s∧ ;W) ≤ ln |X|, using Lemma 2 we get
Iαgλ (p1 ;W) ≥ Iαgλ (s∧ ;W) +
1
1
α−1
1
h
i
ln (1 − δ) + δe(1−α) ln |X|
On the other hand (1 − δ) α µλα,s∧ ≤ µλα,p1 and δ α µλα,s1 ≤ µλα,p1 by (102). Then using (103) and Lemma 1 we get
λ
λ
gλ
gλ
1
≤ α1 ln 1−δ
qα,p
Dα qα,s
− α−1
1
∧
α (Iα (s∧ ;W) − Iα (p1 ;W)),
gλ
gλ
λ
λ
≤ α1 ln δ1 − α−1
qα,p
Dα qα,s
α (Iα (s1 ;W) − Iα (p1 ;W)).
1
1
50
(232)
Since Iαgλ (s1 ;W) ≤ Iαg (s1 ;W) ≤ ln |X| using (231) and we get
Iαgλ (p1 ;W) ≤ Iαgλ (s∧ ;W) +
Using (232) and (233) we get
Iαgλ (p1 ;W) − Iαgλ (p2 ;W) ≤
α
α−1
α
α−1
h
i
α−1
1
1
ln (1 − δ) α + δ α e α ln |X| .
h
i
α−1
1
1
ln (1 − δ) α + δ α e α ln |X| −
1
α−1
(233)
i
h
ln (1 − δ) + δe(1−α) ln |X| .
Then Iαgλ (p;W) is continuous in p.
1−α
gλ
gλ
gλ
(ii) If Iαgλ (e
p ;W) = Cα,W
, then Dα p⊛W e α λ·ρ p ⊗ qα,e
≤ Cα,W
for all p ∈ P(X): Let e
p ∈ P(X) be such that
p
gλ
Iαgλ (e
p ;W) = Cα,W
, p be any member of P(X) and p (ı) be ı−1
p + 1ı p for ı ∈ Z+ . Then
ı e
#
"
gλ
gλ
gλ
q gλ
(α−1) Iαgλ(e
p ;W)+Dα qα,e
(α−1) Iαgλ(p;W)+Dα qα,p
kqα,p
p k α,p (ı)
(ı)
gλ
(ı)
1
1
ı−1
.
+ ıe
Iα p ;W = α−1 ln ı e
gλ
gλ
gλ
and Dα qα,e
Then using Iαgλ p (ı) ;W ≤ Cα,W
p qα,p (ı) ≥ 0 we get
gλ
gλ
gλ
≤ Cα,W
qα,p
Iαgλ (p;W) + Dα qα,p
(ı)
∀ı ∈ Z+ .
(234)
gλ
gλ
gλ
gλ
≥ 0 we get
qα,p
, Iαgλ (e
p ;W) = Cα,W
, Iαgλ (p;W) ≥ 0, and Dα qα,p
On the other hand using Iαgλ p (ı) ;W ≤ Cα,W
(ı)
#
"
gλ
q gλ
(α−1)Dα qα,e
gλ
p k α,p (ı)
1
0 ≥ α−1
+ 1ı e−(α−1)Cα,W
∀ı ∈ Z+ .
ln ı−1
e
ı
Thus
gλ
gλ
≤ 0.
q
lim supı→∞ Dα qα,e
(ı)
p
α,p
gλ
gλ
Then as a result of Lemma 2, qα,p
(ı) converges to qα,e
p in the total variation topology and hence in the topology of setwise
convergence. Since the Rényi divergence is lower semicontinuous in the topology of setwise convergence by Lemma 3,
we have
gλ
gλ
gλ
gλ
(235)
qα,e
Dα qα,p
p ≤ limı→∞ Dα qα,p qα,p (ı) .
1−α
gλ
gλ
Equations (104), (105), (234), (235) imply that Dα p⊛W e α λ·ρ p ⊗ qα,e
p ≤ Cα,W for all p ∈ P(X).
gλ
gλ
gλ
gλ
gλ
(iii) ∃!qα,W
∈ P(Y) satisfying (121) such that qα,e
e with Iαgλ (e
p ;W) = Cα,W
: If Iαgλ (p;W) = Cα,W
for a
p = qα,W for all p
p ∈ P(X), then as a result of (104), (105), and Lemma 2 we have
2
1−α
gλ
gλ
gλ
gλ
α∧1
Dα p⊛W e α λ·ρ p ⊗ qα,e
qα,p
− qα,e
.
p ≥ Cα,W + 2
p
1−α
gλ
gλ
gλ
gλ
Then qα,p
= qα,p̃
because we have already established that Dα p⊛W e α λ·ρ p ⊗ qα,e
p ≤ Cα,W for any p ∈ P(X).
Proof of Theorem 3. Note that (117) implies (118) and (119) implies (120). Furthermore, the left hand side of (117) is equal
gλ
gλ
gλ
to Cα,W
by (116). Thus when Cα,W
is infinite (117) hold trivially by max-min inequality. When Cα,W
is finite (117) follows
gλ
gλ
from (119) and max-min inequality. Thus we can assume Cα,W to be finite and prove the claims about qα,W
, in order to
prove the theorem.
gλ
gλ
gλ
, then {qα,p
(i) If Cα,W
< ∞ and limı→∞ Iαgλ p (ı) ;W = Cα,W
(ı) }ı∈Z+ is a Cauchy sequence in P(Y) for the total variation
gλ
(ı)
, let us consider a sequence of
metric: For any sequence {p }ı∈Z+ ⊂ P(X) satisfying limı→∞ Iαgλ p (ı) ;W = Cα,W
(ı)
(ı)
channels {W }ı∈Z+ whose input sets {X }ı∈Z+ form a nested sequence of finite subsets of X defined as follows,
X(ı) , {x ∈ X : ∃ ∈ {1, . . . , ı} such that p () (x ) > 0}.
gλ
()
Then for any ı ∈ Z+ , there exists a unique qα,W
) ⊂ P(X(ı) )
(ı) satisfying (121) by Lemma 34. Furthermore, P(X
gλ
gλ
for any ı, ∈ Z+ such that ≤ ı. In order to bound qα,p
() − qα,p (ı)
inequality for
λ
qα,p
() ,
gλ
for positive integers < ı we use the triangle
gλ
qα,p (ı) and qα,W (ı)
gλ
gλ
gλ
gλ
gλ
gλ
qα,p
≤ qα,p
+ qα,p
() − qα,p (ı)
() − qα,W (ı)
(ı) − qα,W (ı) .
51
(236)
gλ
gλ
Let us proceed with bounding qα,p
() − qα,W (ı)
gλ
gλ
qα,p
() − qα,W (ı)
gλ
gλ
and qα,p
(ı) − qα,W (ı) .
r
(a)
gλ
gλ
2
qα,W
Dα∧1 qα,p
≤ α∧1
()
(ı)
r
q
(b)
≤
2
α∧1
≤
2
α∧1
(c) q
gλ
gλ
qα,W
Dα qα,p
()
(ı)
q
gλ
gλ
Cα,W
p () ;W (ı)
(ı) − Iα
q
(d) q
gλ
2
Cα,W
− Iαgλ p () ;W
≤ α∧1
where (a) follows from Lemma 2, (b) follows from
from (104), (105), and Lemma34 because
Lemma 8, (c) follows
gλ
gλ
p () ∈ P(X(ı) ), and (d) follows from Iαgλ p () ;W (ı) = Iαgλ p () ;W and Cα,W
(ı) ≤ Cα,W . We can obtain a similar bound
gλ
gλ
gλ
gλ
gλ
on qα,p
p (ı) ;W = Cα,W
.
(ı) −qα,W (ı) . Then {qα,p (ı) } is a Cauchy sequence as a result of (236) because lim Iα
ı→∞
gλ
gλ
gλ
gλ
− qα,p
= 0 for all {p (ı) }ı∈Z+ ⊂ P(X) such
(ii) If Cα,W
< ∞, then ∃! qα,W
∈ P(Y) satisfying limı→∞ qα,W
(ı)
gλ
: Note that M(Y) is a complete metric space for the total variation metric. Then
that limı→∞ Iαgλ p (ı) ;W = Cα,W
gλ
gλ
{qα,p (ı) }ı∈Z+ has a unique limit point qα,W
in M(Y). Since P(Y) is a closed set for the total variation topology and
gλ
gλ
∪ı∈Z+ qα,p
(ı) ⊂ P(Y), then qα,W ∈ P(Y), by [37, Thm. 2.1.3].
gλ
We have established the existence of a unique limit point qα,W
, for any sequence {p (ı) }ı∈Z+ ⊂ P(X) satisfying
gλ
gλ
gλ
. This, however, implies limı→∞ qα,e
limı→∞ Iαgλ p (ı) ;W = Cα,W
− qα,p∗
= 0 for any {e
p (ı) }ı∈Z+ satisfying
p (ı)
gλ
p (ı) }ı∈Z+ to obtain a new sebecause we can interleave the elements of {p (ı) }ı∈Z+ and {e
limı→∞ Iαgλ pe(ı) ;W = Cα,W
gλ
gλ
gλ
gλ
gλ
(ı)
(ı)
quence {b
p }ı∈Z+ satisfying limı→∞ Iα pb ;W = Cα,W for which {qα,bp (ı) } is a Cauchy sequence. Then qα,W
= qα,p
∗
gλ
(1)
(ı)
(iii) qα,W satisfies the equality given in (119): For any p ∈ P(X), let us consider any sequence {p }ı∈Z+ satisfying p = p
gλ
. Then p ∈ P(X(ı) ) for all ı ∈ Z+ . Using Lemma 34 we get
and limı→∞ Iαgλ p (ı) ;W = Cα,W
1−α
gλ
gλ
≤ Cα,W
∀ı ∈ Z+ .
(237)
Dα p⊛W e α λ·ρ p ⊗ qα,W
(ı)
(ı)
gλ
gλ
gλ
Since X(ı) is a finite set, ∃e
p (ı) ∈ P(X(ı) ) satisfying Iαgλ pe(ı) ;W (ı) = Cα,W
= qα,W
(ı) and qα,e
(ı) by Lemma 34. Then
p (ı)
gλ
(ı)
(ı)
(ı)
gλ
(ı)
(ı)
gλ
gλ
and consequently limı→∞ Iα pe ;W = Cα,W . We have already established that for
≥ Iα p ;W
Iα pe ;W
gλ
gλ
such a sequence qα,e
→
q
in
the total variation topology, and hence in the topology of setwise convergence. Then
α,W
p (ı)
gλ
gλ
the lower semicontinuity of the Rényi divergence, i.e. Lemma 3, the identity Cα,W
(ı) ≤ Cα,W imply that
1−α
gλ
gλ
Dα p⊛W e α λ·ρ p ⊗ qα,W
≤ Cα,W
∀p ∈ P(X).
1−α
gλ
gλ
On the other hand Dα p⊛W e α λ·ρ p ⊗ qα,W
≥ Iαgλ (p;W) and Cα,W
= supp∈P(X) Iαgλ (p;W) by definitions of
gλ
Iαgλ (p;W) and Cα,W
. Thus (119) holds.
Proof of Lemma 35. Let us first consider the case α ∈ ℜ+ \ {1}. As a result of (104) and (105) we have,
1−α
supx ∈X Dα (W (x )k q) − λ · ρ(x ) = supx ∈X Dα W (x )e α λ·ρ(x ) q
1−α
≥ Dα p⊛W e α λ·ρ p ⊗ q
gλ
≥ Iαgλ (p;W) + Dα qα,p
q
∀p ∈ P(X).
(238)
gλ
gλ
Let {p (ı) }ı∈Z+ be a sequence of elements of P(X) such that limı→∞ Iαgλ p (ı) ;W = Cα,W
. Then the sequence {qα,p
(ı) }ı∈Z+
gλ
gλ
gλ
is a Cauchy sequence with the unique limit point qα,W
by Theorem 3. Since {qα,p
(ı) } → qα,W in total variation topology,
same convergence holds in the topology of setwise convergence. On the other hand, the order α Rényi divergence is lower
semicontinuous for the topology of setwise convergence by Lemma 3. Thus we have
i
h
gλ
gλ
gλ
q
q
.
(239)
+
D
q
≥
C
lim inf ı→∞ Iαgλ p (ı) ;W + Dα qα,p
α
(ı)
α,W
α,W
gλ
gλ
λ
λ
(238) and (239) imply (125) for α ∈ ℜ+ \{1} because Cα,W
= Cα,W
= qα,W
by (122) and qα,W
by (123).
For α = 1 case, as a result of Lemma 13 we have,
supx ∈X D1 (W (x )k q) − λ · ρ(x ) ≥ I1λ (p;W) + D1 (q1,p k q)
52
∀p ∈ P(X).
(240)
Repeating the argument leading to (239) and invoking Theorem 2, rather than Theorem 3, we get
i
h
λ
λ
+ D1 q1,W
q .
lim inf ı→∞ I1λ p (ı) ;W + D1 q1,p (ı) q ≥ C1,W
(241)
(240) and (241) imply (125) for α = 1 case.
λ2
λ1
λ0
λ
Proof of Lemma 36. Since Cα,W
is nonincreasing in λ by Lemma 29-(a), Cα,W
≤ Cα,W
≤ Cα,W
< ∞. We apply Lemma
λ1
35 for λ = λ2 and q = qα,W and use the fact that 0 ≤ ρ(x ) for all x ∈ X to obtain
λ1
λ2
λ1
λ2
− λ2 · ρ(x )
+ Cα,W
≤ supx ∈X Dα W (x )k qα,W
qα,W
Dα qα,W
λ1
− λ1 · ρ(x ).
≤ supx ∈X Dα W (x )k qα,W
Then (126) follows from (96) of Theorem 2.
λ−ǫ1
For any two point λ1 and λ2 in {λ : ∃ǫ > 0 s.t. Cα,W
< ∞}, not necessarily satisfying λ1 ≤ λ2 , let λ∨ be λ1 ∨ λ2 , i.e.
ı
ı
ı
λ∨ = λ1 ∨ λ2 for all ı ∈ {1, . . . , ℓ}. Then as a result of the triangle inequality we have
λ1
λ2
qα,W
− qα,W
λ1
λ∨
λ∨
λ2
≤ qα,W
− qα,W
+ qα,W
− qα,W
.
On the other hand, as a result of Lemma 2 and (126) we have,
q q
λ1
λ∨
λ1
λ∨
qα,W
− qα,W
= α2 Cα,W
− Cα,W
,
q q
λ2
λ∨
λ∨
λ2
− Cα,W
.
qα,W
= α2 Cα,W
− qα,W
(242)
(243)
(244)
λ−ǫ1
λ
Then continuity of qα,W
in λ on {λ : ∃ǫ > 0 s.t. Cα,W
< ∞} for the total variation topology on P(Y) follows from (242),
λ−ǫ1
λ
(243), (244), and the continuity of Cα,W in λ on {λ : ∃ǫ > 0 s.t. Cα,W
< ∞} established in Lemma 29-(a).
ACKNOWLEDGMENT
Author would like to thank Fatma Nakiboğlu and Mehmet Nakiboğlu for their hospitality; this work simply would not have
been possible without it. Author would like to thank Marco Dalai for informing him about Fano’s implicit assertion of the
fixed point property in [22] and Gonzalo Vazquez-Vilar for pointing out Poltyrev’s paper [19] on the random coding bound.
R EFERENCES
[1] I. Csiszár. Generalized cutoff rates and Rényi’s information measures. IEEE Transactions on Information Theory, 41(1):26 –34, Jan 1995.
[2] Udo Augustin. Noisy Channels. Habilitation thesis, Universitat Erlangen, 852 Erlangen, West Germany, 1978. (https://goo.gl/8wupWM).
[3] M. Dalai. Classical and classical-quantum sphere packing bounds: Rényi vs Kullback and Leibler. In International Zurich Seminar on Communications,
pages 198–202, 2016.
[4] M. Dalai. Some remarks on classical and classical-quantum sphere packing bounds: Renyi vs. kullbackleibler. Entropy, 19(7):355, 2017.
[5] M. Dalai and A. Winter. Constant compositions in the sphere packing bound for classical-quantum channels. IEEE Transactions on Information Theory,
63(9):5603–5617, Sept 2017.
[6] Imre Csiszár and János Körner. Information theory: coding theorems for discrete memoryless systems. Cambridge University Press, Cambridge, UK,
2011.
[7] B. Nakiboğlu. The Augustin center and the sphere packing bound for memoryless channels. In 2017 IEEE International Symposium on Information
Theory (ISIT), pages 1401–1405, Aachen, Germany, June 2017. (http://arxiv.org/abs/1701.06610).
[8] B. Nakiboğlu. The sphere packing bound for memoryless channels. in preparation, 2018.
[9] T. van Erven and P. Harremoës. Rényi divergence and Kullback-Leibler divergence. IEEE Transactions on Information Theory, 60(7):3797–3820, July
2014.
[10] O. Shayevitz. A note on a characterization of Rényi measures and its relation to composite hypothesis testing. ArXiv e-prints, 1012.4401, 2010.
[11] O. Shayevitz. On Rényi measures and hypothesis testing. In Information Theory Proceedings (ISIT), 2011 IEEE International Symposium on, pages
894–898, Austin, Texas, July 2011.
[12] S. Verdú. α-mutual information. In 2015 Information Theory and Applications Workshop, 2015.
[13] J. H. B. Kemperman. On the Shannon capacity of an arbitrary channel. Indagationes Mathematicae (Proceedings), 77(2):101–115, 1974.
[14] B. Nakiboğlu. The Rényi capacity and center. ArXiv e-prints, 1608.02424v4, 2018.
[15] R. G. Gallager. A simple derivation of the coding theorem and some applications. IEEE Transactions on Information Theory, 11(1):3–18, Jan. 1965.
[16] Robert G. Gallager. Information theory and reliable communication. John Wiley & Sons, Inc., New York, NY, 1968.
[17] Paul Michael Ebert. Error Bounds For Parallel Communication Channels. Technical report 448, Research Laboratory of Electronics at Massachusetts
Institute of Technology, Cambridge, MA, 1966. (https://hdl.handle.net/1721.1/4295).
[18] E. A. Haroutunian. Estimates of the error probability exponent for a semicontinuous memoryless channel. Problems of Information Transmission,
4(4):37–48, 1968.
[19] G. Sh. Poltyrev. Random coding bounds for discrete memoryless channels. Problems of Information Transmission, 18(1):9–21, 1982.
[20] Richard M. Dudley. Real analysis and probability, volume 74. Cambridge University Press, New York, NY, 2002.
[21] Vladimir I. Bogachev. Measure Theory. Springer-Verlag, Berlin Heidelberg, 2007.
[22] Roberto M. Fano. Transmission of Information: A Statistical Theory of Communications. M.I.T. Press, Cambridge, MA, 1961.
[23] S. Arimoto. Computation of random coding exponent functions. IEEE Transactions on Information Theory, 22(6):665–671, Nov 1976.
[24] Y. Oohama. The optimal exponent function for the additive white gaussian noise channel at rates above the capacity. In 2017 IEEE International
Symposium on Information Theory (ISIT), pages 1053–1057, Aachen, Germany, June 2017. (http://arxiv.org/abs/1701.06357).
[25] Y. Oohama. Exponent function for stationary memoryless channels with input cost at rates above the capacity. ArXiv e-prints, 1701.06545, 2017.
53
[26] G. Vazquez-Vilar, A. Martinez, and A. G. i Fabregas. A derivation of the cost-constrained sphere-packing exponent. In 2015 IEEE International
Symposium on Information Theory, pages 929–933, Hong Kong, P.R.C., June 2015.
[27] A. Rényi. On measures of entropy and information. In Proceedings of the 4th Berkeley Symposium on Mathematics, Statistics and Probability, pages
547–561, 1960.
[28] I. Csiszár. Information-type measures of difference of probability distributions and indirect observations. Studia Sci. Math. Hungar., 2:299–318, 1967.
[29] G. L. Gilardoni. On Pinsker’s and Vajda’s type inequalities for Csiszár’s f -divergences. IEEE Transactions on Information Theory, 56(11):5377–5386,
Nov 2010.
[30] Y. Polyanskiy and S. Verdú. Arimoto channel coding converse and Rényi divergence. In Communication, Control, and Computing (Allerton), 2010 48th
Annual Allerton Conference on, pages 1327 –1333, Oct 2010.
[31] A. N. Kolmogorov and S. V. Fomin. Introductory Real Analysis. Dover Books on Mathematics. Dover Publications, New York, NY, 1975.
[32] I. Csiszár. A class of measures of informativity of observation channels. Periodica Mathematica Hungarica, 2(1-4):191–213, 1972.
[33] R. Sibson. Information radius. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 14:149–160, 1969.
[34] R. E. Blahut. Hypothesis testing and information theory. IEEE Transactions on Information Theory, 20(4):405–417, Jul 1974.
[35] V. Kostina and S. Verdú. Channels with cost constraints: Strong converse and dispersion. IEEE Transactions on Information Theory, 61(5):2415–2429,
May 2015.
[36] S.G. Krantz and H.R. Parks. A Primer of Real Analytic Functions. Birkhäuser, Boston, 2002.
[37] James R. Munkres. Topology. Prentice Hall Inc., Upper Saddle River, NJ 07458, 2000.
[38] Walter Rudin. Principles of Mathematical Analysis. McGraw-Hill, New York, NY, 1976.
[39] Dimitri P. Bertsekas, Angelia Nedić, and Asuman E. Ozdaglar. Convex Analysis and Optimization. Athena Scientific, Belmont, Massachusetts, 2003.
[40] H. Komiya. Elementary proof for Sion’s minimax theorem. Kodai Mathematical Journal, 11(1):5–7, 1988.
[41] Maurice Sion. On general minimax theorems. Pacific Journal of Mathematics, 8(1):171–176, 1958.
54
| 7cs.IT
|
arXiv:1802.03045v2 [math.GR] 13 Feb 2018
ALGORITHMIC ASPECTS OF BRANCHED COVERINGS III/V.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT
SEQUENCE
LAURENT BARTHOLDI AND DZMITRY DUDKO
r ý be a Thurston map and let M pfrq be its mapping
Abstract. Let fr: pS 2 , Aq
r of maps obtained by pre- and post-composing fr
class biset: isotopy classes rel A
r Let A Ď A
r be an fr-invariant subset, and
by the mapping class group of pS 2 , Aq.
let f : pS 2 , Aq ý be the induced map. We give an analogue of the Birman short
r is an iterated
exact sequence: just as the mapping class group ModpS 2 , Aq
extension of ModpS 2 , Aq by fundamental groups of punctured spheres, M pfrq
is an iterated extension of M pf q by the dynamical biset of f .
Thurston equivalence of Thurston maps classically reduces to a conjugacy
problem in mapping class bisets. Our short exact sequence of mapping class
bisets allows us to reduce in polynomial time the conjugacy problem in M pfrq
to that in M pf q. In case fr is geometric (either expanding or doubly covered
by a hyperbolic torus endomorphism) we show that the dynamical biset Bpf q
r is a complete conjugacy
together with a “portrait of bisets” induced by A
r
invariant of f .
Along the way, we give a complete description of bisets of p2, 2, 2, 2q-maps
as a crossed product of bisets of torus endomorphisms by the cyclic group of
order 2, and we show that non-cyclic orbisphere bisets have no automorphism.
We finally give explicit, efficient algorithms that solve the conjugacy and
centralizer problems for bisets of expanding or torus maps.
1. Introduction
A Thurston
is a branched covering f : S 2 ý of the sphere whose post-critical
Ť map
n
set Pf :“ ně1 f pcritical points of f q is finite.
Extending [18], we developed in [2–5] an algebraic machinery that parallels the
topological theory of Thurston maps: one considers the orbisphere pS 2 , Pf , ordf q,
with ordf : Pf Ñ t2, 3, . . . , 8u defined by
ordf ppq “ l. c. m.tdegq pf n q | n ě 0, f n pqq “ pu,
and the orbisphere fundamental group G “ π1 pS 2 , Pf , ord, ˚q, which has one generator of order ordf ppq per point p P Pf , and one relation. Then f is encoded in
the structure of a G-G-biset Bpf q: a set with commuting left and right actions of
G. By [5, Theorem 7.9] (see also Corollary 3.7), the isomorphism class of Bpf q is
a complete invariant of f up to isotopy.
There is a missing element to this description, that of “extra marked points”. In
the process of a decomposition of spheres into smaller pieces, one is led to consider
Thurston maps with extra marked points, such as periodic cycles or preimages of
Date: February 8, 2018.
Partially supported by ANR grant ANR-14-ACHN-0018-01, DFG grant BA4197/6-1 and ERC
grant “HOLOGRAM”.
1
2
LAURENT BARTHOLDI AND DZMITRY DUDKO
post-critical points. They can be added to Pf , but they have order 1 under ordf
so are invisible in π1 pS 2 , Pf , ordf q. The orbisphere orders can be made artificially
larger, but then other properties, such as the characterization of expanding maps
as those having a “contracting” biset (see [4, Theorem A]) are lost.
We resolve this issue, in this article, by introducing portraits of bisets and exhibiting their algorithmic properties. As an outcome, the conjugacy and centralizer
problems for Thurston maps with extra marked points reduces to that of the underlying map with only Pf marked.
This allows us, in particular, to understand algorithmically maps doubly covered
by torus endomorphisms.
r Ą
ordq and a sub1.1. Maps and bisets. The natural setting is an orbisphere pS 2 , A,
2
Ą
r
orbisphere pS , A, ordq; namely one has A Ď A and ordpaq | ordpaq for all a P A.
r։
r Ą
There is a corresponding morphism of fundamental groups π1 pS 2 , A,
ord, ˚q “: G
2
G :“ π1 pS , A, ord, ˚q, called a forgetful morphism. We considered in [5, §7] the
inessential case in which ordpaq ą 1 ô Ą
ordpaq ą 1, so that only the type of
singularities are changed by the forgetful functor. Here we are more interested in
r
the essential case, in which A Ř A.
In this Introduction, we restrict ourselves to self-maps of orbispheres; the general non-dynamical case is covered in §2. An orbisphere self-map f : pS 2 , A, ordq ý
is a branched covering f : S 2 ý with ordppq degp pf q | ordpf ppqq for all p P S 2 .
Given a map f : pS 2 , Aq ý, there may exist different orbisphere structures on
pS 2 , Aq turning f into an orbisphere self-map; in particular, the maximal one,
in which ordpaq “ 8 for every a P A, and the minimal one, in which ordpaq “
l. c. m.tdegpf n @xq | n ě 0, x P f ´n paqu. The self-map f is encoded by the biset
(1)
Bpf q :“ tβ : r0, 1s Ñ S 2 zA | βp0q “ ˚ “ f pβp1qqu{«A,ord ,
where «A,ord denotes homotopy rel pA, ordq, and the commuting left and right
π1 pS 2 , A, ord, ˚q-actions are given respectively by concatenation of paths or their
appropriate f -lift.
Bisets form a convenient category with products, detailed in [3]. A self-conjugacy
of a G-G-biset B is a pair of maps pφ : G ý, β : B ýq with βphbgq “ φphqβpbqφpgq;
and an automorphism of B is a self-conjugacy with φ “ 1. Cyclic bisets (that of
p t0, 8uq ý) have a special status; for the others,
maps z ÞÑ z d : pC,
Proposition A (= Corollary 4.30). If
AutpBq “ 1.
G BG
is a non-cyclic orbisphere biset then
r “ A \ tdu, obtained by adding a point to the orbisphere pS 2 , A, ordq,
Consider A
r Ą
resulting in an orbisphere pS 2 , A,
ordq. There is a short exact sequence, called the
Birman exact sequence,
(2)
r Ñ ModpS 2 , Aq Ñ 1
1 Ñ π1 pS 2 zA, dq Ñ ModpS 2 , Aq
(note that the orbisphere structure is ignored.) Let f : pS 2 , A, ordq ý be an orbisphere map, and recall that its mapping class biset is the set
M pf q :“ tm1 f m2 | m1 , m2 P ModpS 2 , Aqu{«A ,
with natural left and right actions of ModpS 2 , Aq. Assume first that the extra point
rĄ
r Ą
d is fixed by f , and write fr: pS 2 , A,
ordq ý for the map f acting on pS 2 , A,
ordq.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
3
There is then a natural map M pfrq Ñ M pf q, and we will see that it is an extension
of bisets (see Definition 2.10):
Theorem B (= Corollary 2.23). Subordinate to the exact sequence of groups (2),
there is a short exact sequence of bisets
˙
ˆ ğ
Bpf 1 q
ãÑ M pfrq ։ M pf q
1
π1 pS 2 zA,dq
f 1 PMpf q
1
2
π1 pS 2 zA,dq
where Bpf q denotes the biset of f : pS , Aq ý rel the base point d.
r “ A\tperiodic pointsu: for a cycle of
The statement can be easily extended to A
length ℓ, the fibres in the short exact sequence are of the form Bpf qℓ with left- and
right-action twisted along an ℓ-cycle. This case is at the heart of our reduction of the
conjugacy problem from M pfrq to M pf q. Approximately the same picture applies
if D contains preimages of points in A, but there are subtle complications which
are taken care of in §2, see Theorem 2.13. In essence, the presence of preperiodic
points imposes a finite index condition on centralizers, and splits conjugacy classes
into finitely many pieces, see the remark after Lemma 4.24.
1.2. Portraits of bisets. Portraits of bisets emerge from a simple remark: fixed
points of f naturally yield conjugacy classes in Bpf q. Indeed if f ppq “ p then choose
a path ℓ : r0, 1s Ñ S 2 zA from ˚ to p, and consider cp :“ ℓ#ℓ´1 Òpf P Bpf q, which
is well-defined up to conjugation by G, namely a different choice of ℓ would yield
g ´1 cp g for some g P G. Conversely, if f expands a metric, then every conjugacy
class in Bpf q corresponds to a unique repelling f -fixed point.
r ý
A portrait of bisets in Bpf q, see Definition 2.15, consists of a map f˚ : A
extending f åA ; a collection of peripheral subgroups pGa qaPAr of G: they are such that
every Ga is cyclic and generated by a “lollipop” around a; and a collection of cyclic
Ga -Gf˚ paq -bisets pBa qaPAr . Two portraits of bisets pGa , Ba qaPAr and pG1a , Ba1 qaPAr
r ý are conjugate if there exist pℓa q r in GAr such
parameterized by the same f˚ : A
aPA
1
´1 1
that Ga “ ℓ´1
a Ga ℓa and Ba “ ℓa Ba ℓf˚ paq . The set of self-conjugacies of a portrait
is called its centralizer.
r every biset admits a unique minimal portrait up to conjugacy,
In case A “ A,
which may be understood geometrically as follows. Consider a branched covering
f : pS 2 , Aq ý. For every a P A choose a small disk neighbourhood Da of it; up
to isotopy we may assume that f : Da ztau Ñ Df paq ztf paqu is a covering. A choice
of embeddings π1 pDa ztauq ãÑ π1 pS 2 , Aq yields a family pGa qaPA of peripheral subgroups; and the corresponding embeddings Bpf : Da ztau Ñ Df paq ztf paquq ãÑ Bpf q
yields a minimal portrait of bisets pGa , Ba qaPA in Bpf q.
r “ A \ te1 , . . . , en u with pe1 , . . . , en q a periodic cycle, the bisets Bei
In case A
consist of single points, and almost coincide with Ishii and Smillie’s notion of homotopy pseudo-orbits, see [13] and §4.2: imagine that pe1 , . . . , en q Ă S 2 is almost a
periodic cycle, in that f pei q is so close to ei`1 that there is a well-defined “shortest”
path ℓi from the first to the second, indices being read modulo n. Choose for each
ei
i a path mi from ˚ to ei . Set then Bi :“ tmi #pℓi`1 #m´1
i`1 qÒf u, the portrait of
bisets encoding pe1 , . . . , en q.
Theorem B is proven via portraits of bisets. There is a natural forgetful intertwiner of bisets
2 r Ą
2
r
r
(3)
r B r :“ Bpf : pS , A, ordq ýq Ñ Bpf : pS , A, ordq ýq “: G BG
G
G
4
LAURENT BARTHOLDI AND DZMITRY DUDKO
r b r G. Every portrait in B,
r for example
given by b ÞÑ 1 b b b 1 where B – G bGr B
G
its minimal one, induces a portrait in B via the forgetful map. Let us denote by
r is
ModpS 2 , Aq the pure mapping class group of S 2 zA. A class m P ModpS 2 , Aq
r
called knitting if the image of m is trivial in ModpS 2 , A \ tauq for every a P AzA.
We prove:
Theorem C (= Theorem 2.19). Let G BG be an orbisphere biset with portrait
r ։ G and f˚ : A
r ý and deg : A
r Ñ N be compatible extensions.
f˚ : A ý, and let G
There is then a bijection between, on the one hand, conjugacy classes of portraits
of bisets pGa , Ba qaPAr in B parameterized by f˚ and deg and, on the other hand,
r G-bisets
r
r ։ G considered up to composition with the
Gprojecting to B under G
biset of a knitting element. This bijection maps every minimal portrait of bisets of
r to pGa , Ba q r .
B
aPA
1.3. Geometric maps. A homeomorphism f : pS 2 , Aq ý is geometric if f is either
of finite order (f n “ 1 for some n ą 0) or pseudo-Anosov (there are two transverse
measured foliations preserved by f such that one foliation is expanded by f while
another is contracted). In both cases, f preserves a geometric structure on S 2 , and
every surface homeomorphism decomposes, up to isotopy, into geometric pieces,
see [25].
Consider now a non-invertible sphere map f : pS 2 , Aq ý, and let A8 Ď A denote the forward orbit of the periodic critical points of f . The map f is Böttcher
expanding if there exists a metric on S 2 zA8 that is expanded by f , and such that
f is locally conjugate to z ÞÑ z dega pf q at every a P A8 . The map f is geometric if
f is either
tExpu Böttcher expanding; or
tGTor/2u a quotient of a torus endomorphism z ÞÑ M z`q : R2 {Z2 ý by the involution
z ÞÑ ´z, for a 2 ˆ 2 matrix M whose eigenvalues are different from ˘1.
The two cases are not mutually exclusive. A map f P tGTor/2u is expanding if
and only if the absolute values of the eigenvalues of M are greater than 1. Note
also that if f is non-invertible and covered by a torus endomorphism then either
f P tExpu or the minimal orbisphere of f satisfies #Pf “ 4 and ordf ” 2.
In that last case, we show that G BG is a crossed product of an Abelian biset
with an order-2 group. Let us fix a p2, 2, 2, 2q-orbisphere pS 2 , A, ordq and let us
set G :“ π1 pS 2 , A, ordq – Z2 ¸ t˘1u. We may identify A with the set of all
order-2 conjugacy classes of G. By Euler characteristic, every branched covering
f : pS 2 , A, ordq ý is a self-covering. Therefore, the biset of f is right principal.
We denote by Mat`
2 pZq the set of 2 ˆ 2 integer matrices M with detpM q ą 0.
2
For a matrix M P Mat`
2 pZq and a vector v P Z there is an injective endomorphism
v
2
M : Z ¸ t˘1u ý given by the following “crossed product” structure:
(4)
M v pn, 1q “ pM n, 1q and M v pn, ´1q “ pM n ` v, ´1q.
Furthermore, Z2 ¸ t˘1u has exactly 4 conjugacy classes of order 2, which we denote
by A. Then M v induces a map pM v q˚ : A ý on these conjugacy classes. We write
BM v “ BM ¸ t˘1u the crossed product decomposition of the biset of M v .
Proposition D (= Propositions 4.5 and 4.6). The biset of M v from (4) is an
orbisphere biset, and conversely every p2, 2, 2, 2q-orbisphere biset B is of the form
2
B “ BM v for some M P Mat`
2 pZq and some v P Z . Two bisets BM v and BN w
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
5
are isomorphic if and only if M “ ˘N and v ” w pmod 2Z2 q. The biset BM v is
geometric if and only if both eigenvalues of M are different from ˘1.
A distinguished property of a geometric map is rigidity: two geometric maps
are Thurston equivalent (namely, conjugate up to isotopy) if and only if they are
topologically conjugate.
An orbisphere biset G BG is geometric if it is the biset of a geometric map,
and tGTor/2u and tExpu bisets are defined similarly. If B is a geometric biset,
then by rigidity there is a map fB : pS 2 , A, ordq ý, unique up to conjugacy, with
BpfB q – B.
r r Ñ G BG is a forgetful intertwiner as in (3), then
If G BG is geometric and Gr B
G
r
elements of AzA
(which a priori do not belong to any sphere) can be interpreted
r is itself geometric,
dynamically as extra marked points on S 2 zA. More precisely, if B
2
and B is the biset of the geometric map f : pS , Aq ý, then there is an embedding
r in S 2 as an f -invariant set, unique unless G is cyclic, in such a manner that
of A
r
r ýq.
B is isomorphic to Bpf : pS 2 , Aq
Since furthermore geometric maps have only finitely many periodic points of
given period, we obtain a good understanding of conjugacy and centralizers of
geometric bisets:
r Ñ G be a forgetful morphism of groups and
Theorem E (= Theorem 4.41). Let G
let
r r Ñ G BG and r B
r1 Ñ GB1G
r BG
r
G
G G
r is geometbe two forgetful biset morphisms as in (3). Suppose furthermore that B
1
1
ric of degree ą 1. Denote by pGa , Ba qaPAr and pGa , Ba qaPAr the portraits of bisets
r and B
r 1 in B and B 1 respectively.
induced by B
1
r B
r are conjugate under ModpGq
r if and only if there exists φ P ModpGq
Then B,
such that B φ – B 1 and the portraits pGφa , Baφ qaPAr and pG1a , Ba1 qaPAr are conjugate.
Furthermore, the centralizer of the portrait pGa , Ba qaPAr is trivial, and the cenr of B
r is isomorphic, via the forgetful map ModpGq
r Ñ ModpGq, to
tralizer ZpBq
ˇ φ φ
(
φ P ZpBq ˇ pGa , Ba qaPAr „ pGa , Ba qaPAr
and is a finite-index subgroup of ZpBq.
Let us call an orbisphere map f : pS 2 , A, ordq ý weakly geometric if its minimal
quotient on pS 2 , Pf , ordf q is geometric; an orbisphere biset B is weakly geometric if
its minimal quotient orbisphere biset is geometric. In Theorem 4.35 we characterize
weakly geometric maps as those decomposing as a tuning by homeomorphisms:
starting from a geometric map, some points are blown up to disks which are mapped
to each other by homeomorphisms.
1.4. Algorithms. An essential virtue of the portraits of bisets introduced above
is that they are readily usable in algorithms. Previous articles in the series already
highlighted the algorithmic aspects of bisets; let us recall the following.
From the definition of orbisphere bisets in [5, Definition 2.6], it is clearly decidable whether a given biset H BG is an orbisphere biset; the groups G and
H may be algorithmically identified with orbisphere groups π1 pS 2 , A, ord, ˚q and
π1 pS 2 , C, ord, :q respectively, and the induced map f˚ : C Ñ A is computable. In
particular, if G BG is a G-G-biset, then the dynamical map f˚ : A ý is computable,
6
LAURENT BARTHOLDI AND DZMITRY DUDKO
the minimal orbisphere quotient G :“ π1 pS 2 , Pf , ordf , ˚q is computable, and the
induced G-G-biset B is computable. It is also easy (see [4, §5]) to determine from
an orbisphere biset whether it is tGTor/2u (and then to determine an affine map
M z ` q covering it) or tExpu.
We shall show that recognizing conjugacy of portraits is decidable, and give
efficient (see below) algorithms proving it, as follows:
Algorithm 1.1 (= Algorithms 5.6 and 5.10). Given a minimal geometric orbirÑA
r of the dynamics of B on its peripheral
sphere biset G BG , an extension f˚ : A
classes, and two portraits of bisets pGa , Ba qaPAr and pG1a , Ba1 qaPAr with dynamics f˚ ,
Decide whether pGa , Ba qaPAr and pG1a , Ba1 qaPAr are conjugate, and Compute the
centralizer of pGa , Ba qaPAr , which is a finite abelian group.
Algorithm 1.2 (= Algorithms 5.7 and 5.11). Given a minimal geometric orr Ñ A
r of the dynamics of B on its
bisphere biset G BG and an extension f˚ : A
peripheral classes,
Produce a list of representatives of all conjugacy classes of portraits of bisets
pGa , Ba qaPAr in B with dynamics f˚ .
Thurston equivalence to a map f : pS 2 , Aq ý reduces to the conjugacy problem
in the mapping class biset M pf q. Let X beŤa basis of M pf q and let N be a finite
generating set of ModpS 2 , Aq; so M pf q “ ně0 N n X. We call an algorithm with
input in M pf q ˆ M pf q efficient if for f, g P N n X the running time of the algorithm
is bounded by a polynomial in n.
We deduce that conjugacy and centralizer problems are decidable for geometric
maps, as long as they are decidable on their minimal orbisphere quotients:
Corollary F (= Algorithm 5.12). There is an efficient algorithm with oracle that,
given two orbisphere maps f, g by their bisets and such that f is geometric, decides
whether f, g are conjugate, and computes the centralizer of f .
The oracle must answer, given two geometric orbisphere maps f, g on their minimal orbisphere pS 2 , Pf , ordf q respectively pS 2 , Pg , ordg q, whether they are conjugate
and what the centralizer of f is.
Algorithms for the oracle itself will be described in details in the last article of
the series [6]. Furthermore, we have the following oracle-free result, proven in §5.3:
Corollary G. There is an efficient algorithm that decides whether a rational map
is equivalent to a given twist of itself, when only extra marked points are twisted.
1.5. Historical remarks: Thurston equivalence and its complexity. The
conjugacy problem is known to be solvable in mapping class groups [11]. The
state of the art is based on the Nielsen-Thurston classification: decompose maps
along their canonical multicurve; then a complete conjugacy invariant of the map
is given by the combinatorics of the decomposition, the conjugacy classes of return
maps, and rotation parameters along the multicurve. For general surfaces, the
cost of computing the decomposition is at most exponential time in n, see [14, 24],
and so is the cost of comparing the pseudo-Anosov return maps [17]. Margalit,
Yurttas, Strenner recently announced polynomial-time algorithms for all the above.
At all rates, for punctured spheres the cost of computing the decomposition is
polynomial [7].
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
7
Kevin Pilgrim developed, in [20], a theory of decompositions of Thurston maps
extending the Nielsen-Thurston decomposition of homeomorphisms. It is a fundamental ingredient in understanding Thurston equivalence of maps, without any
claims on complexity.
In [1], a general strategy is developed, along with computations of the mapping
class biset for the three degree-2 polynomials with three finite post-critical points
and period respectively 1, 2, 3. The Douady-Hubbard “twisted rabbit problem” is
solved in this manner (it asks to determine for n P Z the conjugacy class of r ˝ tn
with rpzq « z 2 ´ 0.12256 ` 0.74486i the “rabbit” polynomial and tpzq the Dehn
twist about the rabbit’s ears). Russell Lodge computed in [16] the solution to an
array of other “twisted rabbit” problems, by finding explicitly the mapping class
biset structure.
If M pf q is a contracting biset, then its conjugacy problem may be solved in
quasi-linear time. In terms of a twist parameter such as the n above, the solution
has Oplog nq complexity. However, this bound is not uniform, in that it requires
e.g. the computation of the nucleus of M pf q, which cannot a priori be bounded.
Nekrashevych showed in [19, Theorem 7.2] that the mapping class biset of a hyperbolic polynomial is contracting. Conversely, if M pf q contains an obstructed map
then it is not contracting.
In the case of polynomials, however, even more is possible: bisets of polynomials admit a particularly nice form by placing the basepoint close to infinity (this
description goes hand-in-hand with Poirier’s “supporting rays” [21]). As a consequence, the “spider algorithm” of Hubbard and Schleicher [12] can be implemented
directly at the level of bisets and yields an efficient algorithm, also in practice since
it does not require the computation of M pf q’s nucleus. This algorithm will be
described in [6].
Selinger and Yampolsky showed in [23] that the canonical decomposition is computable. In this manner, they solve the conjugacy problem for maps whose canonical
decomposition has only rational maps with hyperbolic orbifold: a complete conjugacy invariant of the map is given by the combinatorics of the decomposition, the
conjugacy classes of its return maps, and rotation parameters along the canonical
obstruction.
We showed in [5] that the conjugacy problem is decidable in general: a complete
conjugacy invariant of a Thurston map is given by the combinatorics of the decomposition, together with conjugacy classes of return maps, rotation parameters
along the canonical obstruction, together with the induced action of the centralizer
groups of return maps.
We finally mention a different path towards understanding Thurston maps, in
the case of maps with four post-critical points: the “nearly Euclidean Thurston
maps” from [8]. There, the restriction on the size of the post-critical set implies
that the maps may be efficiently encoded via linear algebra; and as a consequence,
conjugacy of NET maps is efficiently decidable. We are not aware of any direct
connections between their work and ours.
1.6. Notations. Throughout the text, some letters keep the same meaning and are
not always repeated in statements. The symbols A, C, D, E denote finite subsets of
the topological sphere S 2 . There is a sphere map fr: pS 2 , C \ Eq Ñ pS 2 , A \ Dq,
which restricts to a sphere map f : pS 2 , Cq Ñ pS 2 , Aq. Implicit in the definition, we
r “ C \E and A
r “ A\D. If there
have f pCqYtcritical values of f u Ď A. We write C
8
LAURENT BARTHOLDI AND DZMITRY DUDKO
are, furthermore, orbisphere structures on the involved spheres, we denote them by
pS 2 , A, ordq etc., with the same symbol ‘ord’. We also abbreviate X :“ pS 2 , A, ordq
and Y :“ pS 2 , C, ordq.
For sphere maps f0 , f1 : pS 2 , Cq Ñ pS 2 , Aq, we mean by f0 «C f1 that f0 and f1
are isotopic, namely there is a path pft qtPr0,1s of sphere maps ft : pS 2 , Cq Ñ pS 2 , Aq
connecting them.
For γ a path and f a (branched) covering, we denote by γÒxf the unique f -lift of
γ that starts at the preimage x of γp0q.
We denote by f åZ the restriction of a function f to a subset Z of its domain.
Finally, for a set Z we denote by Z Ó. the group of all permutations of Z.
2. Forgetful maps
We recall a minimal amount of information from [5]: a marked orbisphere is
pS 2 , A, ordq for a finite subset A Ă S 2 and a map ord : A Ñ t2, 3, . . . , 8u, extended to S 2 by ordpS 2 zAq ” 1. For a choice of basepoint ˚ P S 2 zA, its orbispace
fundamental group is generated by “lollipop” loops pγa qaPA based at ˚ that each
encircle once counterclockwise a single point of A, and with A “ ta1 , . . . , an u has
presentation
(5)
nq
1q
, γa1 ¨ ¨ ¨ γan y.
, . . . , γaordpa
G “ π1 pS 2 , A, ord, ˚q “ xγa1 , . . . , γan | γaordpa
n
1
Abstractly, i.e. without reference to a sphere, an orbisphere group is a group G as
in (5) together with the conjugacy classes Γ1 , . . . , Γn of γa1 , . . . , γan respectively.
An orbisphere map f : pS 2 , C, ordC q Ñ pS 2 , A, ordA q between orbispheres is
an orientation-preserving branched covering between the underlying spheres, with
f pCq Y tcritical values of f u Ď A, and with ordC ppq degp pf q | ordA pf ppqq for all
p P S2.
To avoid special cases, we make, throughout this article except in §6.3, the assumption
(6)
#A ě 3
and
#C ě 3.
For #A “ 2, many things go wrong: one must require ord to be constant; the fundamental group has a non-trivial centre; and the degree-d self-covering of pS 2 , A, ordq
has an extra symmetry of order d ´ 1. All our statements can be modified to take
into account this special case, see §6.3.
Fix basepoints ˚ P S 2 zA and : P S 2 zC. The orbisphere biset of an orbisphere
map f : pS 2 , C, ordC q Ñ pS 2 , A, ordA q is the π1 pS 2 , C, ordC , :q-π1 pS 2 , A, ordA , ˚qbiset
Bpf q “ tβ : r0, 1s Ñ S 2 zC | βp0q “ :, f pβp1qq “ ˚u { «C,ordC ,
with ‘«C,ord ’ denoting homotopy in the orbispace pS 2 , C, ordq. An orbisphere biset
H BG can also be defined purely algebraically, see [5, §7 and Definition 2.6]. By
[5, Theorem 7.9], there is an orbisphere map f : pS 2 , C, ordC q Ñ pS 2 , A, ordA q,
unique up to isotopy, such that B is isomorphic to Bpf q. We denote by B˚ : C Ñ A
the induced map on the peripheral conjugacy classes of H and G – they are identified
with the associated punctures.
Ą be orbispheres, and suppose ordpaq | Ą
Let pS 2 , A, ordq and pS 2 , A \ D, ordq
ordpaq
for all a P A. We then have a natural forgetful homomorphism FD : π1 pS 2 , A \
D, Ą
ord, ˚q Ñ π1 pS 2 , A, ord, ˚q given by γa ÞÑ γa for a P A and γd ÞÑ 1 for d P D.
We write the forgetful map pS 2 , A \ D, Ą
ordq 99K pS 2 , A, ordq with a dashed arrow,
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
9
because even though FD is a genuine group homomorphism, the corresponding map
between orbispheres is only densely defined. Note, however, that its inverse is a
genuine orbisphere map.
Consider forgetful maps pS 2 , C \E, Ą
ordq 99K pS 2 , C, ordq and pS 2 , A\D, Ą
ordq 99K
2
pS , A, ordq; in the sequel we shall keep the notations
X :“ pS 2 , A, ordq,
r :“ A \ D,
A
Y :“ pS 2 , C, ordq,
r :“ C \ E.
C
Ą Ñ pS 2 , A,
r ordq
r Ą
Let fr: pS 2 , C,
ordq be an orbispace map, with such that fr restricts
to an orbisphere map f : pS 2 , C, ordq Ñ pS 2 , A, ordq. We thus have
Ą
r ordq
pS 2 , C,
(7)
FE
X “ pS 2 , C, ordq
fr
f
rĄ
pS 2 , A,
ordq
FD
pS 2 , A, ordq “ Y.
(In particular, frpCq Ď A and A contains all critical values of fr.) We are concerned,
in this section, with the relationship between Bpfrq and Bpf q; we shall show that
Bpfrq may be encoded by Bpf q and a portrait of bisets, and that this encoding is
unique up to a certain equivalence. The algebraic counterpart of (7) is
(8)
Ą ˚q “: G
r
r Ą
r ord,
r :“ π1 pS 2 , C,
H
ord, :q í Bpfrq ì π1 pS 2 , A,
FE
FE,D
FD
H :“ π1 pS 2 , C, ord, :q í Bpf q ì π1 pS 2 , A, ord, ˚q “: G,
where we denote by
(9)
r rG
FE,D : Bpfrq Ñ Bpf q – H bH
Ă Bpf q bG
the natural map given by b ÞÑ 1 b b b 1.
2.1. Braid groups and knitting equivalence. We recall that there is no difference between ModpS 2 , Aq and ModpXq, see [5, §7]. We shall define a subgroup
r which
ModpX|Dq intermediate between ModpXq – ModpS 2 , Aq and ModpS 2 , Aq
r
will be useful to relate Bpf q and Bpf q.
Definition 2.1 (Pure braid group). Let D be a finite set on a finitely punctured
sphere S 2 zA. The pure braid group BraidpS 2 zA, Dq is the set of continuous motions
m : r0, 1s ˆ D Ñ S 2 zA considered up to isotopy so that
‚ mpt, ´q : D ãÑ S 2 zA is an inclusion for every t P r0, 1s;
‚ mp0, ´q “ mp1, ´q “ 1 åD .
The product in BraidpS 2 zA, Dq is concatenation of motions, and the inverse is
time reversal.
△
Note that in the special case D “ t˚u we have BraidpS 2 zA, t˚uq “ π1 pS 2 zA, ˚q.
Theorem 2.2 (Birman). For every m P BraidpS 2 zA, Dq there is a unique mapping
class pushpmq P kerpModpS 2 , A\Dq Ñ ModpS 2 , Aqq such that pushpmq is isotopic
rel A to the identity via an isotopy moving D along m´1 . The map
(10)
push : BraidpS 2 zA, Dq Ñ kerpModpS 2 , A \ Dq Ñ ModpS 2 , Aqq
is an isomorphism (it would be merely an epimorphism if #A ď 2).
10
LAURENT BARTHOLDI AND DZMITRY DUDKO
From now on we identify BraidpS 2 zA, Dq with its image under (10).
Definition 2.3 (Knitting group). Let X “ pS 2 , A, ordq be an orbisphere and let D
be a finite subset of S 2 zA. The knitting braid group knBraidpX, Dq is the kernel
of the forgetful morphism
ź
ED : BraidpS 2 zA, Dq Ñ
π1 pX, dq;
dPD
2
it is the set of D-strand braids in S zA all of whose strands are homotopically
trivial in X.
△
In case X “ pS 2 , Aq, knitting elements are the “p#D ´ 1q-decomposable braids”
from [15].
Lemma 2.4. The knitting group knBraidpX, Dq is a normal subgroup of ModpS 2 , A \ Dq.
Proof. We show that for every m P knBraidpX, Dq and h P ModpS 2 , A \ Dq
we have h´1 mh P knBraidpX, Dq. Indeed, h restricts to an orbisphere map
h : X ý fixing D pointwise. Thus mpd, ´q is a trivial loop in π1 pX, dq if and
only if hpmpd, ´qq is a trivial loop in π1 pX, dq for every d P D.
Define
(11)
ModpX|Dq :“ ModpS 2 , A \ Dq{knBraidpX, Dq.
With G “ π1 pX, ˚q, we write ModpG|Dq “ ModpX|Dq and ModpGq “ ModpXq.
We interpret elements of ModpGq as outer automorphisms of G, as mapping classes
and as biprincipal bisets. We also introduce the notation
ź
π1 pX, Dq :“
π1 pX, dq – GD .
dPD
Note the following four exact sequences:
knBraidpX, Dq
BraidpS 2 zA, Dq
ED
(12)
r “ ModpS 2 , A \ Dq
ModpGq
π1 pX, Dq;
push
ModpX|Dq
FD
FD
ModpGq “ ModpS 2 , Aq
exactness follows by definition except surjectivity in the top sequence. Given a
sequence of loops pγd qdPD P π1 pX, Dq, we may isotope them slightly to obtain
γd ptq ‰ γd1 ptq for all t P r0, 1s and all d ‰ d1 . Then bpd, tq :“ γd ptq is a braid in
S 2 zA and defines via ‘push’ an element of BraidpS 2 zA, Dq mapping to pγd q.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
11
2.2. fr-impure mapping class groups. As in (7), consider an orbisphere map
fr: pS 2 , C\E, Ą
ordq Ñ pS 2 , A\D, Ą
ordq that projects to f : pS 2 , C, ordq Ñ pS 2 , A, ordq.
We will enlarge the groups in (12) to “fr-impure mapping class groups” so that exact
sequences analogous to (16) hold.
Let Mod˚ pS 2 , C \ Eq be the group of homeomorphisms m : pS 2 , C \ Eq ý
considered up to isotopy rel C \ E such that m åC is the identity and for every
e P E we have frpmpeqq “ frpeq; i.e., m may permute points in fr´1 pfrpeqqzC. There
is a natural forgetful morphism
FE : Mod˚ pS 2 , C \ Eq Ñ ModpS 2 , Cq.
As in Definition 2.1, the braid group Braid˚ pS 2 zC, Eq is the set of continuous
motions m : r0, 1s ˆ E Ñ S 2 zC considered up to isotopy so that
‚ mpt, ´q : E ãÑ S 2 zC is an inclusion for every t P r0, 1s;
‚ mp0, ´q “ 1 åE ;
‚ mp1, eq P fr´1 pfrpeqq for every e P E.
Every m P Braid˚ pS 2 zC, Eq induces a permutation πm : e ÞÑ mp1, eq of E. The
pure braid group consists of those permutations with πm “ 1. The product in
Braid˚ pS 2 zC, Eq is m ¨ m1 “ m#pm1 ˝ p1 ˆ πm qq, with as usual ‘#’ standing
for concatenation of motions. Birman’s theorem (a slight generalization of Theorem 2.2) still holds: the group Braid˚ pS 2 zC, Eq is isomorphic to the kernel of
Mod˚ pS 2 , C \ Eq Ñ ModpS 2 , Cq via the push operator.
Let π1˚ pY, Eq be the group of motions m : r0, 1s ˆ E Ñ Y “ pS 2 , C, ordq, considered up to homotopy, such that
‚ mp0, ´q “ 1 åE ;
‚ mp1, eq P fr´1 pfrpeqq for every e P E;
here m, m1 : E ãÑ Y are homotopic if mp´, eq and m1 p´, eq are homotopic curves
(relative to their endpoints) in Y for all e P E. The product in π1˚ pY, Eq is again
m ¨ m1 “ m#pm1 ˝ p1 ˆ πm qq. We have
ź
ź
pfr´1 pdq X EqÓ.,
(13)
π1˚ pY, Eq –
π1 pY, eq ¸
ePE
dPfrpEq
the isomorphism mapping m to its restrictions mp´, eq and its permutation πm .
Lemma 2.5. The following sequence is exact:
(14)
knBraidpY, Eq
Braid˚ pS 2 zC, Eq
EE
π1˚ pY, Eq,
where EE is the natural forgetful morphism.
Proof. Suppose EE pmq “ EE pm1 q. Then m´1 m1 is pure so m´1 m1 P knBraidpY, Eq.
The converse is also obvious.
The same argument as in Lemma 2.4 shows that the knitting group knBraidpY, Eq
is a normal subgroup of Mod˚ pS 2 , C \ Eq. We may thus define
(15)
Mod˚ pY |Eq :“ Mod˚ pS 2 , C \ Eq{knBraidpY, Eq.
12
LAURENT BARTHOLDI AND DZMITRY DUDKO
As in (12) we have the following exact sequences
knBraidpY, Eq
Braid˚ pS 2 zC, Eq
EE
(16)
Mod˚ pS 2 , C \ Eq
π1˚ pY, Eq.
push
Mod˚ pY |Eq
FE
FE
ModpHq “ ModpS 2 , Cq
2.3. Branched coverings. Recall that, for orbisphere maps f0 , f1 : pS 2 , C, ordq Ñ
pS 2 , A, ordq we write f0 «C f1 , and call them isotopic, if there is a path pft qtPr0,1s
of orbisphere maps ft : pS 2 , C, ordq Ñ pS 2 , A, ordq. Equivalently,
Lemma 2.6. f0 «C f1 if and only if hf0 “ f1 for a homeomorphism h : pS 2 , Cq ý
that is isotopic to the identity.
Proof. If there exists an isotopy pht q witnessing 1 «C h, then ft :“ ht f0 witnesses
f0 «C f1 . Conversely, since all critical values of ft are frozen in A, the set ft´1 pyq
moves homeomorphically for every y P S 2 (equivalently, no critical points collide).
Therefore, we may factor ft “ ht f0 , with ht pzq the trajectory of z P ft´1 pyq; this
defines an isotopy from 1 to h :“ h1 .
Consider orbisphere maps fr, r
g : pS 2 , C \ E, Ą
ordq Ñ pS 2 , A\ D, Ą
ordq as in (7). We
write fr «C|E r
g, and call fr, r
g knitting-equivalent, if fr “ hr
g for a homeomorphism
h : pS 2 , C \ Eq ý in knBraidpY, Eq; we have
fr «C\E r
g ùñ fr «C|E gr ùñ fr «C r
g.
For m P BraidpS 2 zA, Dq we define its pullback pfrq˚ m : r0, 1s ˆ E Ñ S 2 zC by
#
mp´, frpeqqÒefr if frpeq P D,
˚
ppfrq mqp´, eq :“
e
if frpeq P A.
This defines a motion of E; note that pfrq˚ mp1, eq need not equal e:
Lemma 2.7. If pushpmq P BraidpS 2 zA, Dq, then pushppfrq˚ mq defines an element
of Braid˚ pS 2 zC, Eq and we have the following commutative diagram:
(17)
pS 2 , C \ Eq
fr
pS 2 , A \ Dq
pushppfrq˚ mq
pS 2 , C \ Eq
pushpmq
pS 2 , A \ Dq.
«C\E
fr
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
13
Proof. Let us discuss in more detail the operator ‘push’. Consider a simple arc
γ : r0, 1s Ñ S 2 zA and let U Ă S 2 zA be a small disk neighborhood of γ. We can
define (in a non-unique way) a homeomorphism pushpγq : pS 2 , Aq ý that maps γp0q
to γp1q and is identity away from U. Let U1 , U2 , . . . , Ud be the preimages of U under
fr, where d “ degpfrq. Each Ui contains a preimage γi of γ. Let pushpγi q : Ui ý be
the lift of pushpγq åU under fr: Ui Ñ U, extended by the identity on S 2 zUi . Then
(18)
pushpγ1 qpushpγ2 q ¨ ¨ ¨ pushpγd q ¨ fr “ fr ¨ pushpγq.
For m P BraidpS 2 zA, Dq, we can define pushpmq as a composition of pushes
along finitely many simple arcs βi . Using (18) we lift all pushpβi q through fr;
considering the equation rel C \ E we obtain (17).
We note that pushpmq does not necessarily lift to a ‘push’ if m moves a critical
value. Indeed if γ is a simple loop then pushpγq is the quotient of two Dehn twists
about the boundary curves of an annulus surrounding γ, see [9, §4.2.2]; however
the lift of γ will not be a union of simple closed curves if γ contains a critical value;
an annulus around γ will not lift to an annulus, but rather to a more complicated
surface F ; and the quotient of Dehn twists about boundary components of F will
not be a quotient of Dehn twists about boundary components of annuli.
Proposition 2.8. Let fr: pS 2 , C \ E, Ą
ordq Ñ pS 2 , A \ D, Ą
ordq be an orbisphere map
as in (7). Then every element in knBraidpX, Dq lifts through fr to an element of
knBraidpY, Eq.
If gr : pS 2 , C \ E, Ą
ordq Ñ pS 2 , A\ D, Ą
ordq is another orbisphere map, then fr «C|E
g if and only if fr «C hr
r
gk for some h P knBraidpY, Eq, k P knBraidpX, Dq.
Proof. Consider h P knBraidpX, Dq. By Theorem 2.2 we may write h “ pushpbq.
Since bp´, dq is homotopically trivial in X for every d P D, the curve bp´, frpeqqÒefr
ends at e for all e P E with frpeq P D, because this curve is in Y . Therefore,
pfrq˚ bp1, ´q “ 1 åE , and pushppfrq˚ bq P BraidpS 2 zC, Eq. Since the lifts bp´, frpeqqÒer
f
are homotopically trivial in Y for all e P E, we have pushppfrq˚ bq P knBraidpY, Eq
and Lemma 2.7 concludes the first claim.
The second claim is a direct consequence of the first.
As a consequence, we may detail a little bit more the map ED in (12). Choose
for every d P D a path ℓd in S 2 zA from ˚ to d. This path defines an isomorphism
π1 pX, dq Ñ π1 pX, ˚q “ G by γ ÞÑ ℓd #γ#ℓ´1
d . We thus have a map
(19)
ED : BraidpS 2 zA, Dq ։ GD ,
m ÞÑ pℓd #ppush´1 pmq åd q#ℓ´1
d qdPD ,
and kerpED q “ knBraidpX, Dq.
2.4. Mapping class bisets. We introduce some notation parallel to that in (12)
r Ñ H and FD : G
r Ñ G be forgetful
and (16) for mapping class bisets. Let FE : H
r
morphisms of orbisphere groups as in (8), and let B be an orbisphere biset. Let
r r G be the induced H-G-biset. We have forgetful morphisms of
B :“ H bH
Ă B bG
r Ñ ModpGq and FE : ModpHq
r Ñ ModpHq. Corresponding
groups FD : ModpGq
mapping class bisets are written respectively, with fr and f the orbisphere maps
14
LAURENT BARTHOLDI AND DZMITRY DUDKO
r and B,
associated with B
r “ M pfrq – tn b B
r b m | n P ModpHq,
r m P ModpGqu{–
r
M pBq
r m P ModpS 2 , Aqu{«
r
“ tnfrm | n P ModpS 2 , Cq,
r,
C
M pBq “ M pf q – tn b B b m | n P ModpHq, m P ModpGqu{–
“ tnf m | n P ModpS 2 , Cq, m P ModpS 2 , Aqu{«C
together with the natural forgetful intertwiner
(20)
r Ñ M pBq,
FE,D : M pBq
r 1 Ñ FE,D pB 1 q “ H b Ă B
r 1 b r G.
B
G
H
We may also define the following mapping class biset, sometimes larger than
r assume first that Ą
M pBq:
ord is constant on E, possibly 8, and set
(21)
ˇ
+
#
ˇ
r 1 q P M pBq
F
p
B
ˇ
E,D
1
˚ r
˚ r
r an Hr G-orbisphere
r
biset ˇ
{–,
M pBq “ M pf q :“ B
r 1 q˚ “ B
r˚
ˇ
pB
r 1 q˚ : C
rÑA
r denotes the induced map on marked conjugacy classes. It is
where pB
˚
2 r
r
r 1 P M ˚ pBq
r for
an Mod pS , Cq-ModpS 2 , Aq-biset;
note indeed that we have n b B
˚
1
˚ r
2 r
r
r
B P M pBq and n P Mod pS , Cq, because FE,D pn b Bq “ FE pnq b B. Again
there is a natural forgetful intertwiner
(22)
FE,D :
r M
Mod˚ pS 2 ,Cq
˚
pfrqModpS 2 ,Aq
r Ñ ModpS 2 ,Cq M pf qModpS 2 ,Aq .
r on M pfrq does not necessarily extend
We note that the left action of ModpS 2 , Cq
˚
2 r
to an action of Mod pS , Cq on M pfrq, because the result of the action is in general
in M ˚ pf q and not in M pf q, see Example 6.2.
In case Ą
ord is not constant on E, we should be careful, because permutation
r invariant; rather, the image of H
r under such a
of points in E does not leave H
r
r and
permutation gives an orbisphere group isomorphic to H. However, M ˚ pBq
˚
2 r
Mod pS , Cq do not depend on the orbisphere structure, so the definition may be
r and G
r replaced by orbisphere groups with larger orders.
applied with H
Let us call the set of extra marked points E saturated if
fr´1 pfrpEqq Ď C \ E.
Lemma 2.9.
(1) The mapping class biset M ˚ pfrq is left-free.
(2) Suppose that E is saturated and that g0 , g1 : pS 2 , C \ E, Ą
ordq Ñ pS 2 , A \
Ą
D, ordq are orbisphere maps coinciding on C \ E and such that FE,D pg0 q
and FE,D pg1 q are isotopic through maps pS 2 , Cq Ñ pS 2 , Aq. Then g0 , g1 P
M ˚ pfrq, and there is an m P Braid˚ pS 2 zC, Eq such that mg0 “ g1 holds in
M ˚ pfrq.
(3) If E is saturated, then
(23)
M ˚ pfrq “ tmfrn | m P Mod˚ pS 2 , C \ Eq, n P ModpS 2 , A \ Dqu{«Cr .
Proof. The proof of the first claim follows the lines of [5, Proposition 6.4]: suppose
r Ñ pS 2 , Aq
r for some m P Mod˚ pS 2 , C\
that g, mg are isotopic through maps pS 2 , Cq
˚ r
Eq and g P M pf q. By Lemma 2.6, we may assume g “ mg as maps; then the
homeomorphism m is a deck transformation of the covering induced by g, so m has
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
15
finite order because degpgq ă 8. Recall that #C ě 3 by our standing assumption (6). Since m fixes at least 3 points in C and m has finite order, we deduce
that m is the identity. This shows that M ˚ pfrq is left-free.
For the second claim, let pgt : pS 2 , Cq Ñ pS 2 , AqqtPr0,1s be an isotopy between g0
and g1 . By Lemma 2.6, we may write gt at mt g0 for mt : pS 2 , Cq ý. Then m1
preserves E because E is saturated, so m1 P Braid˚ pS 2 zC, Eq as required.
The third claim directly follows from the second.
Definition 2.10 (Extensions of bisets, see [5, Definition 5.2]). Let G1 BG2 be a
G1 -G2 -biset and let N1 , N2 be normal subgroups of G1 and G2 respectively, so that
for i “ 1, 2 we have short exact sequences
(24)
1
Ni
π
Gi
Qi
1.
If the quotient Q1 -Q2 -biset N1 zB{N2 , consisting of connected components of N1 BN2 ,
is left-free, then the sequence
(25)
N 1 BN 2
G 1 BG 2
π
Q1 pN1 zB{N2 qQ2
is called an extension of left-free bisets.
△
r ։ H and G
r ։ G be surjective
Definition 2.11 (Inert biset morphism). Let H
r
r
r
group homomorphisms, and let B be a left-free H-G-biset. Recall that the tenr r G is isomorphic to the double quotient kerpH
r Ñ
sor product B :“ H bH
Ă B bG
r kerpG
r Ñ Gq with natural H-G-actions. The natural map F : B
r Ñ B is
HqzB{
r
called inert if B is a left-free biset and the natural map t¨u bGr B Ñ t¨u bG B is a
r In other words,
bijection. In particular, B has the same number of left orbits as B.
r
r
assuming that the groups H and H have similarly-written generators and so do G
r
and G, the wreath recursions of B and B are identical.
rker ãÑ ĂB
r r ։ H BG the kernel is a
Yet said differently, in the extension ker B
H G
r“H
r and G “ H so that the bisets can
disjoint union of left-principal bisets. If G
r
rÑGÑ
be iterated, then B Ñ B is inert precisely when we have a factorization G
r
IMGGr pBq, the latter groupŮbeing the quotient of G by the kernel of the right action
on the rooted tree t¨u bH ně0 B bn , see [5, §7.3].
△
Define
r
M ˚ pB|E, Dq :“ knBraidpY, EqzM ˚ pBq{knBraidpX,
Dq;
(26)
this is naturally a Mod˚ pY |Eq-ModpX|Dq-biset, and Proposition 2.8 implies in
particular that it is left-free:
Proposition 2.12. The natural forgetful maps
r
r Ñ
Ă M pBq
ModpHq
ModpGq
ModpH|Eq M pB|E, DqModpG|Dq
and
Mod˚ pS 2 ,C\Eq M
are inert.
˚
E Ó.
˚
˚
r
pBq
ModpS 2 ,A\Dq Ñ Mod˚ pY |Eq M pB|E, DqModpX|Dq
Let
denote the group of all permutations t : E ý such that frptpeqq “ frpeq.
˚
˚
We denote by H E ¸E Ó. the semidirect product where E Ó. acts on H E by permuting
˚
coordinates; compare with (13). We have π1˚ pY, Eq – H E ¸ E Ó. .
16
LAURENT BARTHOLDI AND DZMITRY DUDKO
We denote by
Braid˚ pY,Eq M
˚ r
˚
˚
r
pBq
BraidpX,Dq and
π1˚ pY,Eq M
˚
pB|E, Dqπ1 pX,Dq the
restrictions of M pBq and M pB|E, Dq to braid and fundamental groups.
Theorem 2.13. If E is saturated, then the following sequences are extensions of
bisets:
(27)
r
M ˚ pBq
knBraidpY,Eq
knBraidpX,Dq
Braid˚ pY,Eq M
˚
r
pBq
BraidpX,Dq
EE,D
r
M ˚ pBq
FE,D
π1˚ pY,Eq M
pB|E, Dqπ1 pX,Dq
P
˚
M pB|E, Dq
FE,D
M pBq
˚
˚
H E ¸E Ó.
!
–
pB 1 P M pBq, pBc1 qcPCr q
pBc1 qcPCr is a portrait in B 1
)
.
GD
(The isomorphism on the right is the topic of Theorem 2.19, and will be proven
there.)
´1
r
pBq is a connected subbiset of Braid˚ pY,Eq M ˚ pBq
Proof. By Lemma 2.9(2), FE,D
BraidpX,Dq ;
thus the central-to-left sequence is an extension of bisets. Exactness of other se
quences follows from Proposition 2.12.
r by M pBq
r in (27),
Note that, if E were not saturated or if we replaced M ˚ pBq
then we wouldn’t have exact sequences of bisets anymore, because the fibres of
r ։ M pBq wouldn’t have to be connected; see Example 6.1. The failure of
M ˚ pBq
transitivity is described by Lemma 2.20. There are similar exact sequences in case
f˚ : E Ñ D is a bijection, see Theorem 2.22.
2.5. Portraits of bisets. First, a portrait of groups amounts to a choice of representative in each peripheral conjugacy class of an orbisphere group:
Definition 2.14 (Portraits of groups). Let G be an orbisphere group with marked
r be a finite set containing A. A portrait of
conjugacy classes pΓa qaPA and let A
groups pGa qaPAr in G is a collection of cyclic subgroups Ga ď G so that
#
xgy for some g P Γa , if a P A,
Ga “
x1y otherwise.
r “ A, then pGa qaPA is called a minimal portrait of groups.
If A
△
Definition 2.15 (Portraits of bisets). Let H, G be orbisphere groups with peripheral classes indexed by C, A respectively, let H BG be an orbisphere biset, and let
f˚ : C Ñ A be its portrait. We also have a “degree” map deg : C Ñ N. A portrait
of bisets relative to these data consists of
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
17
‚ portraits of groups pHc qcPCr in H and pGa qaPAr in G;
rÑA
r of f˚ and deg : C
r Ñ N of deg;
‚ extensions f˚ : C
‚ a collection pBc qcPCr of subbisets of B such that every Bc is an Hc -Gf˚ pcq biset that is right-transitive and left-free of degree degpcq, and such that if
f˚ pcq “ f˚ pc1 q and HBc “ HBc1 qua subsets of B then c “ c1 .
r “ C and A
r “ A, then pBc qcPC is called a minimal portrait of bisets.
If C
△
r
Note in particular that if c P CzC
then Hc is trivial and the subbiset Bc consists
of degpcq elements permuted by Gf˚ pcq . If moreover f˚ pcq R A, then degpcq “ 1.
For simplicity the portrait of bisets is sometimes simply written pBc qcPCr , the other
data f˚ , deg, pHc qcPCr , pGa qaPAr being implicit.
Here is a brief motivation for the definition. Consider an expanding Thurston
map f and its associated biset B. Recall from [18] that bounded sequences in B
represent points in the Julia set of f ; in particular constant sequences represent
fixed points of f and vice versa. On the other hand, fixed Fatou components
of f are represented by local subbisets of B with same cyclic group acting on
the left and on the right. All of these are instances of portraits of bisets in B.
Furthermore, (pre)periodic Julia or Fatou points are represented by portraits of
bisets with (pre)periodic map f˚ . E.g., closures of Fatou components intersect if
and only if they admit portraits that intersect; and similarly for inclusion in the
closure of a Fatou component of a (pre)periodic point in the Julia set.
r Ñ G be a marked forgetful morphism of orbisphere groups as in (8).
Let FD : G
r
For all a P A choose a small disk neighbourhood Ua Q a that avoids all other points
r so that π1 pUa zAq
r – Z. We call a curve γ close to a if γ Ă Ua .
in A,
r a q r in G
r is uniquely characLemma 2.16. Every minimal portrait of groups pG
aPA
terized by a family of paths pℓa qaPAr with
(28)
r
ℓa : r0, 1s Ñ S 2 rel A,
r for t ă 1, ℓa p0q “ ˚ and ℓa p1q “ a,
ℓa ptq R A
so that for any sufficiently small ǫ ą 0
(29)
r a “ ℓa år0,1´ǫs #αa #pℓa år0,1´ǫs q´1 | αa is close to a
G
(
r
rel A.
Conversely, every collection of curves (28) defines by (29) a minimal portrait of
groups for every sufficiently small ǫ ą 0.
Proof. This follows immediately from the definition of “lollipop” generators, see (5).
ra is self-normalizing in G:
r conjugating G
ra by an element
It follows that every G
r
g R Ga amounts to precomposing the path ℓa with g, resulting in a different path.
Lemma 2.17. Let H BG be an orbisphere biset. Then every pair of minimal portraits of groups pHc qcPC in H and pGa qaPA in G can be uniquely completed to a
minimal portrait of bisets pBc qcPC in B.
As a consequence, the minimal portrait of bisets is unique up to conjugacy. We
note that the lemma is also true in case H BG is a cyclic biset, namely if H, G – Z;
in this case all Bc are equal to B.
Proof. Since by Assumption (6) the sets A and C contain at least 3 elements, H
and G are non-cyclic, and in particular have trivial centre.
18
LAURENT BARTHOLDI AND DZMITRY DUDKO
Let f˚ : C Ñ A be the portrait of B. Choose generators ga P Ga and hc P Hc .
Recall from [3, §2.6] or [5, Definition 2.5] that there are bc P B and kc P H such
degpcq
that kc´1 hc kc bc “ bc gf pcq for all c P C. Set then Bc :“ Hc kc bc Gf pcq , and note that
these subbisets satisfy Definition 2.15.
Suppose next that pBc1 qcPC is another portrait of bisets, and choose elements
1
bc P Bc1 . Then by [5, Definition 2.6(SB3 )] the conjugacy class ∆c appears exactly
once among the lifts of Γf pcq , so Hb1c X Bc ‰ H, so we may choose kc P H with
kc b1c P Bc . Then kc Bc1 “ kc b1c Gf pcq “ Bc Gf pcq “ Bc . We have kc Hc “ Hc kc , so
kc P Hc because Hc is self-normalizing in H, and therefore Bc “ Bc1 .
r r Ñ H BG of orbisphere bisets,
Consider next a forgetful morphism FE,D : H
Ă BG
rc q r be the minimal portrait of bisets given by Lemma 2.17.
and let pB
cPC
Lemma 2.18. Let pmc qcPCr and pℓa qaPAr be the paths (see Lemma 2.16) associr c q r and pG
ra q r . The portrait
ated with the respective portraits of bisets pH
cPC
aPA
rc qq r admits the following description: for every sufficiently small ǫ ą 0,
pFE,D pB
cPC
!
)
β p1q
(30)
Bc “ mc år0,1´ǫs #βc #pℓf˚ pcq år0,1´ǫs q´1 Òf c | βc is close to c rel C.
rc by the intertwiner FE,D (see (9)), it suffices
Proof. Since Bc is obtained from B
to consider the case E “ D “ H; and in that case, by Lemma 2.17 it suffices to
note that Bc is indeed an Hc -Gf˚ pcq -biset.
rÑA
r be an abstract portrait extending
Let H BG be an orbisphere biset; let f˚ : C
r
B˚ ; let deg : C Ñ N be an extension of degB : C Ñ N; and let pBc qcPCr be a portrait
of bisets in B with portraits of groups pHc qcPCr in H and pGa qaPAr in G. A congruence
of portraits is defined by a choice of phc qcPCr in H and pga qaPAr in G, and modifies
the portrait of bisets by replacing
Hc ù h´1
c Hc h c ,
Bc ù h´1
c Bc gf˚ pcq ,
Ga ù ga´1 Ga ga .
By Lemma 2.17, any two minimal portraits of bisets are congruent.
2.6. Main result. Consider an orbisphere map f : pS 2 , C, ordq Ñ pS 2 , A, ordq. We
r Ą
rĄ
call it compatible with FE : pS 2 , C,
ordq 99K pS 2 , C, ordq and FD : pS 2 , A,
ordq 99K
r Ñ A
r and deg : C
r Ñ N if degpeq “ 1 for all e P E with
pS 2 , A, ordq and f˚ : C
Ą degpcq | Ą
r and tdegpf @f ´1 paqqu “
f˚ peq P D, and ordpcq
ordpf˚ pcqq for all c P C,
´1
r Ą
ordq Ñ
tdegpf˚ paqu for all a P A. Equivalently, there is an orbisphere map fr: pS 2 , C,
2 r Ą
2
2
pS , A, ordq that can be isotoped within maps pS , C, ordq Ñ pS , A, ordq to a map
r Ñ N and f˚ : C
r Ñ N are induced by fr.
making (7) commute and such that deg : C
r Ñ H, FD : G
r Ñ G,
Compatibility of an orbisphere biset H BG with FE : H
r Ñ A,
r and deg : C
r Ñ N is defined similarly, and is equivalent to the existence
f˚ : C
r making (8) commute.
of a biset B
We are now ready to relate the mapping class bisets M ˚ to portraits of bisets:
r Ñ H and FD : G
r Ñ G be forgetful morphisms of
Theorem 2.19. Let FE : H
orbisphere groups as in (8), and let H BG be an orbisphere biset compatible with
r Ñ A,
r and deg : C
r Ñ N.
FE , FD , f ˚ : C
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
19
Then for every portrait of bisets pBc qcPCr in B parameterized by f˚ and deg there
r r mapping to B under FE,D with a minimal portrait
exists an orbisphere biset H
Ă BG
mapping to pBc qcPCr .
r is unique up to pre- and post-composition with bisets of knitting
Furthermore, B
mapping classes and we have a congruence of bisets (see (19))
(31)
π1˚ pY,Eq M
˚
pB|E, Dqπ1 pX,Dq
P
˚
H E ¸E Ó.
pB 1 P M pBq, pBc1 qcPCr q | B 1 P M pBq, pBc1 qcPCr a portrait in B 1
r 1 q “ pFE,D pB
r 1 q, induced portrait of B
r 1 q.
given by PpB
(
GD
The H E -GD -action on tpB 1 , pBc1 qcPCr qu is given by
phc qcPE pB 1 , pBc1 qcPCr qpgd qdPD “ pB 1 , phc Bc1 gf˚ pcq qcPCr q,
with the understanding that hc “ 1 if c R E and gf˚ pcq “ 1 if f˚ pcq R D, and the
˚
action of E Ó. is given by permutation of the portrait of bisets.
r “ G),
r Theorem 2.19 proves
In the dynamical situation (i.e. when H “ G and H
Theorem C.
˚
Proof of Theorem 2.19. Clearly, (31) is an intertwiner: firstly, the actions of E Ó.
are compatible; we may ignore them in the sequel. Secondly, let pBc1 qcPCr be the
r 1 . For e P E we may write
induced portrait of bisets of B
Be1 “ tme #p´1 | p : r0, 1s Ñ Y, p´1 p0q “ e, f ˝ p “ ℓf˚ peq u,
see (30). Consider ψ P ModpY |Eq; the action of ψ on Be1 replaces me by ψ˝me ; and
if EE pψq “ phe qePE then ψ ˝ me “ he me by the very definition of ‘push’ and (19).
The same argument applies to the right action.
Let us now show that P is a congruence. Since P is an intertwiner between leftfree bisets with isomorphic acting groups, it is sufficient to show that P preserves
left orbits.
We proceed by adding new points to D and E. If E “ H, then the forgetful
maps M ˚ pB|E, Dq Ñ M pBq and
(
pB 1 P M pBq, pBc1 qcPCr q | B 1 P M pBq, pBc1 qcPCr a portrait in B 1 Ñ M pBq
are bijections and the claim follows. Therefore, it is sufficient to assume that
D “ H.
By [5, Theorem 7.9], the biset B may be written as Bpf q for a branched covering
f : pS 2 , C, ordq Ñ pS 2 , A, ordq, unique up to isotopy rel C. We lift f to a branched
r ` be its biset and let pB
rc` qcPf ´1 pAq be
covering f ` : pS 2 , f ´1 pAqq Ñ pS 2 , Aq. Let B
its minimal portrait of bisets, which is unique by Lemma 2.17.
Let us show that for every portrait of bisets pBc qcPCr in B there is an orbisphere
r whose minimal portrait of bisets maps to pBc q. This B
r will be of the form
biset B
`
2
´1
r
Bpmq b Ff ´1 pAqzC,H
p
B
q
for
a
homeomorphism
m
:
pS
,
f
pAqq
ý.
r
rc` in t¨u b Ă B
r ` – t¨u bH B, and compare
First, consider the images of all B
H
r
them to the images of all Bc . The condition that as c1 ranges over f˚´1 pf˚ pcqq the
20
LAURENT BARTHOLDI AND DZMITRY DUDKO
Bc1 lie in different H-orbits of B lets us select which preimages of A correspond
r and thus produces a well-defined map C
r Ñ f ´1 pAq.
to the marked points in C,
r to f ´1 pAq in the specified manner;
Let m1 be an isotopy of pS 2 , Cq which maps C
1 `
2 r
2
r 0 and portrait of bisets pB
r0 q r ;
then m f : pS , C, ordq Ñ pS , A, ordq has biset B
c cPC
0
0
r q Ď HBc for all c P C,
r so we may write hc FE,H pB
r q “ Bc for
and FE,H pB
c
c
some hc P H. (We recall that Bc consists of dpcq elements permuted by Gf˚ pcq ,
r 0 q “ Bc if and only if
where dpcq is the local degree of f at c. We have hc FE,H pB
c
ś
0
0
r
r q X Bc ‰ H.) We set B
r“
pushph
q
B
,
and
note that the minimal
hc FE,H pB
c
c
cPE
r
portrait of bisets of B maps to pBc qcPCr under FE,H .
The only choice involved is that of a mapping class that yields pushphc q when
restricted to C Y tcu, namely that of knitting mapping classes.
2.7. Fibre bisets. Consider an orbisphere map fr: pS 2 , C \ E, Ą
ordq Ñ pS 2 , A \
Ą
D, ordq as above and define the saturation of E as
ğ
˘
`
s :“
E
fr´1 frpeq zC.
ePE
Ą be an orbisphere map and
Lemma 2.20. Let fr: pS 2 , C \E, Ą
ordq Ñ pS 2 , A\D, ordq
2
s
s
let m : pS , C \ Eq ý be a homeomorphism such that m åC “ 1 and for every e P E
˚
2
r
r
we have f pmpeqq “ f peq. If the isotopy class of m is not in Mod pS , C \ Eq,
s
namely if m moves at least one point in E to EzE,
then mfr ffC\E fr.
˚ r
s ý as above
For every r
g P M pf q there is a homeomorphism m : pS 2 , C \ Eq
r
r
r
s
(i.e. m åC “ 1 and f pmpeqq “ f peq for e P E) such that r
g «C\E mf .
s does not act on M ˚ pfrq: there are orbisphere maps
Note that Mod˚ pS 2 , C \ Eq
2
fr1 , fr2 : pS , C \ E, Ą
ordq Ñ pS 2 , A \ D, Ą
ordq such that fr1 «C\E fr2 but mfr1 ffC\E
2
r
s ý as above.
mf2 for a homeomorphism m : pS , C \ Eq
s is saturated, by Lemma 2.9(2) there is an
Proof. Suppose mfr «C\E fr. Since E
2
s
s
n : pS , C \ Eq ý with n åC\E “ 1 and frpnpeqq “ frpeq for e P EzE
such that
r
r
nmf «C\Es f . This contradicts Lemma 2.9(1): the biset
s Ñ pS 2 , A \ Dqq
(32)
M ˚ pfr: pS 2 , C \ Eq
is left-free while nm ‰ 1.
The second claim follows from Lemma 2.9(2) applied to the biset (32).
We are interested in the fibre biset
F ´1 pB 1 qGD
H E E,D
under the forgetful map
r
FE,D : M pB|E, Dq Ñ M pBq. For every a P A define Ea :“ E X f˚´1 paq, where
rÑA
r is the portrait.
f˚ : C
˚
Proposition 2.21. We have
ź
ź
´1
˚
1
˚
1
(33) H E FE,D
pB 1 qGD –
H Ea M pB |Ea , Hq1 ˆ
H Ed M pB |Ed , tduqGtdu .
aPA
1
dPD
Suppose that B is the biset of g : pS , C, ordq Ñ pS 2 , A, ordq; then
The biset
2
˚
1
Ea
ˆ tι : Ea ãÑ g ´1 paqzCu.
H Ea M pB |Ea , Hq1 – H
˚
1
H Ed M pB |Ed , tduqGtdu is congruent to the biset
tpbe qePEd P B 1Ed | Hbe ‰ Hbe1 if e ‰ e1 u
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
21
endowed with the actions
phe qePEd ¨ pbe qePEd ¨ g “ phe be gqePEd .
´1
pB 1 qGD is isomorphic to the biset of
Proof. By Theorem 2.19, the biset H E FE,D
1
portraits pBc qcPCr in B . Recall that pBc qcPC is fixed (by Lemma 2.17) while pBe qePE
varies; this shows (33).
The second claim follows from Lemma 2.20.
For d P D and e P Ed we have Be “ tbe u; i.e. the choice of pBe qePEd is equivalent
to the choice of pbe qePEd P B 1Ed subject to the condition stated above.
Note that H Ea M ˚ pB 1 |Ea , Hq1 will not be transitive, as soon as there are at least
two maps ι : Ea Ñ g ´1 paqzC.
Let us define
r b ModpX|Dq.
M pB|E, Dq :“ ModpY |Eq b M pBq
Theorem 2.22. Suppose f˚ pEq Ă D and, moreover, that f˚ : E Ñ D is a bijection.
We then have a congruence
` ´1
˘
(34)
FE,D pB 1 q GD Ñ pH B 1 G qE
HE
rc q r in B 1 to pbe qePE where B
re “ tbe u. The group GD is
mapping the portrait pB
ePC
identified with GE via the bijection f˚ : E Ñ D.
r “ M pBq,
r M ˚ pB|E, Dq “ M pB|E, Dq, and exact sequences
Moreover, M ˚ pBq
similar to (27) hold:
(35)
r
M pBq
knBraidpY,Eq
knBraidpX,Dq
r
BraidpY,Eq M pBqBraidpX,Dq
EE,D
r
M pBq
π1 pY,Eq M pB|E, Dqπ1 pX,Dq
P
M pB|E, Dq
FE,D
FE,D
HE
M pBq
!
–
pB 1 P M pBq, pBc1 qcPCr q
pBc1 qcPCr is a portrait in B 1
)
The bottom sequence in (35) can be written (using (34)) as
(36)
Ů
E
B 1 PB
pB 1 q
M pB|E, Dq
M pBq.
Proof. The first claim follows from Proposition 2.21. Since pB 1 qE is a transitive
biset, we obtain M pB|E, Dq “ M ˚ pB|E, Dq and the exact sequences (36) and (35)
hold because the fibres are connected.
.
GD
22
LAURENT BARTHOLDI AND DZMITRY DUDKO
Corollary 2.23. Let f : pS 2 , A, ordq ý be an orbisphere map and let D “
be a finite union of periodic cycles Di of f . Then (36) takes the form
ğ ğ
(37)
Bpgqbp#Di q ãÑ M pf |D, Dq ։ M pf q.
Ů
iPI
Di
gPMpf q iPI
2.8. Centralizers of portraits. Let us now consider the dynamical situation:
r “ H;
r we abbreviate FD “ FD,D and M pB|Dq “ M pB|D, Dq.
H “ G and G
r we denote by ZpBq
r ď ModpGq
r the centralizer of
Given an orbisphere biset B
r
r
r
r in
B in M pBq and by ZpB|Dq ď ModpG|Dq the centralizer of the image of B
M pB|Dq, see Theorem 2.19. We have a natural forgetful map
r Ñ ZpB|Dq
r
ZpBq
(38)
which is, in general, neither injective nor surjective. However, we will show in
r is geometric non-invertible.
Corollary 4.28 that (38) is an isomorphism if B
Recall from (12) the short exact sequence π1 pX, Dq ãÑ ModpG|Dq ։ ModpGq.
We have the corresponding sequence for centralizers:
(39)
1
r
ZpB|Dq
X π1 pX, Dq
r
ZpB|Dq
ZpBq.
r is geometric non-invertible, then ZpB|Dq
r
r
If B
X π1 pX, Dq is trivial, so ZpB|Dq
–
r is naturally a subgroup of ZpBq, and in Theorem 4.41 we will show that it
ZpBq
has finite index.
Definition 2.24. The relative centralizer ZD ppBa qaPAr q of a portrait of bisets
pBa qaPAr is the set of pgd q P GD such that
Bd “ gd´1 Bd gf˚ pdq for all d P D,
with the understanding that gf˚ pdq “ 1 if f˚ pdq R D.
△
We remark that we could also have defined the “full” normalizer, consisting of
r
r and its subgroup
all pga q P GA with Ggaa “ Ga and Ba “ ga´1 Ba gf˚ paq for all a P A,
´1
the “full” centralizer, in which ga centralizes Ga and ga ¨ ´ ¨ gf˚ paq is the identity
on Ba ; but ś
we will make no use of these notions. The “full” normalizer is the direct
product of aPA Ga and the relative centralizer.
We also note that, if pgd qdPD belongs to the relative centralizer of pBa qaPAr and
f n pdq P A for some n P N, then gd “ 1.
Applying Theorem 2.19 to the dynamical situation we obtain:
r r ։ G BG be a forgetful morphism and let pBa q r be
Proposition 2.25. Let Gr B
G
aPA
the induced portrait of bisets in B. Then any choice of isomorphisms π1 pX, dq – G
gives an isomorphism
–
r
ZpB|Dq
X π1 pX, Dq ÝÑ ZD ppBa qaPAr q.
3. G-spaces
We start by general considerations. Let Y be a right H-space, and let X be a
right G-space. For every map f : Y Ñ X there exists a natural H-G-biset M pf q,
defined by
(40)
M pf q :“ thf g :“ f p´ ¨ hq ¨ g | h P H, g P Gu,
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
23
namely the set of maps Y Ñ X obtained by precomposing with the H-action and
post-composing with the G-action. Note that M pf q is right-free if the action of G
is free on X. We have a natural G-equivariant map Y bH M pf q Ñ X given by
evaluation: y b b ÞÑ bpyq. Define
Hf :“ th P H | Dg P G with hf “ f g in Mf u
the stabilizer in H of f G P M pf q{G. Then f descends to a continuous map
f : Y {Hf Ñ X{G.
Lemma 3.1. Suppose that X and Y are simply connected and that G, H act freely
with discrete orbits. Then M pf q is isomorphic to the biset of the correspondence
f
Y {H Ð Y {Hf Ñ X{G as defined in [3].
Proof. Let us define the following subbiset of M pf q:
(41)
M 1 pf q :“ tf p´ ¨ hq ¨ g | h P Hf , g P Gu.
Since Y, X are simply connected, the biset of f : Y {Hf Ñ X{G is isomorphic to
M 1 pf q. The isomorphism is explicit: choose basepoints : P Y and ˚ P X so that
π1 pY {H, :Hq – H and π1 pX{G, ˚Gq – G. Given b P Bpf q, represent it as a
path β : r0, 1s Ñ X{G with βp0q “ f p:Hf q and βp1q “ ˚G, and lift it to a path
β : r0, 1s Ñ X. We have βp0q “ f p:qh and βp1q “ ˚g for some h P Hf , g P G, and
we map b P Bpf q to h´1 ¨ f ¨ g P M 1 pf q. This map is a bijection because both Bpf q
and M 1 pf q are right-free. We finally have
M pf q – H bHf M 1 pf q – BpY {H Ð Y {Hf Ñ X{Gq.
In case the actions of G, H are discrete but not free, there still is a surjective
morphism BpX{G Ø Y {Hq ։ M pf q, when X{G and Y {H are treated as orbispaces.
Example 3.2 (Modular correspondence). The mapping class biset M pf q from §2.4
is isomorphic to the biset of the associated correspondence on Moduli space, see [5,
Proposition 8.1]. Indeed, M pf q is identified with
(42)
tσn ˝ σf ˝ σm | n P ModpS 2 , Cq, m P ModpS 2 , Aqu,
where σm : TA ý, σn : TC ý, and σf : TA Ñ TC are the pulled-back actions
between Teichmüller spaces of m, n, f respectively. By Lemma 3.1, the biset (42)
is isomorphic to the biset of the modular correspondence
σf
i
MC ÐÝ Wf ÝÑ MA .
3.1. Universal covers. Let us now generalize Lemma 3.1 by dropping the requirement that X, Y be simply connected. Choose basepoints :, ˚, write Q “ π1 pY, :q
r “ π1 pY {H, :Hq and G
r “ π1 pX{G, ˚Gq; so we have exact
and P “ π1 pX, ˚q and H
sequences
(43)
π
rÑ
1ÑQÑH
H Ñ 1,
π
rÑ
1ÑP ÑG
G Ñ 1.
The universal cover of X may be defined as
r :“ tβ : r0, 1s Ñ X | βp1q “ ˚u{«;
X
it has a natural basepoint r̊ given by the constant path at ˚, and admits a right
r Ñ X is a covering,
P -action by right-concatenation of loops at ˚. The projection X
24
LAURENT BARTHOLDI AND DZMITRY DUDKO
r _ the left P -set structure on X,
r with
and is given by β ÞÑ βp0q. We denote by X
´1 _
_
´1
_ :“
r with tβ | β P Xu
r and
pβ ¨ g q . We may naturally identify X
action g ¨ β
its natural left P -action.
r Yr of X, Y respectively and a lift fr of f .
Let us consider the universal covers X,
We have the following situation, with the acting groups represented on the right,
and omitted in the left column:
Yr
fr
Yr {Q – Y
f
r ìG
r
X
XìG
Y {Hf
f
r – Y {H
Yr {H
X{G
r f be the full preimage of Hf in H.
r Note that H
r f is the stabilizer of frG
r in
Let H
r
r
r
r
r
r
r
M pf q{G: we have hf P f G if and only if πphqf P f G by the unique lifting property.
By Lemma 3.1, we have
r Ð Yr {H
r f Ñ X{
r Gq
r – BpY {H Ð Y {Hf Ñ X{Gq,
M pfrq – BpYr {H
and we shall see that M pfrq is an extension of bisets. We have a natural map
r g P G.
r
π : M pfrq Ñ M pf q, given by hfrg ÞÑ πphqf πpgq for all h P H,
Lemma 3.3. There is a short exact sequence of bisets
(44)
r
Q M pf qP
M pfrq
π
M pf q,
in which every fibre π ´1 phf gq is isomorphic to a twisted form Bphf gq of the biset
of f .
r with actions of
Proof. This follows immediately from Lemma 3.1 applied to Yr , X
Q, P respectively.
Let us now assume that the short exact sequences (43) are split, so π1 pX{Gq –
P ¸ G and π1 pX{Hq – Q ¸ H. We shall see that the sequence (44) is split, and
that some additional structure on M pf q and Bpf q allow the extension M pf q –
BpY {H Ø X{Gq to be reconstructed as a crossed product.
The splitting of the map π : π1 pX{G, ˚Gq Ñ G means that there is a family
tαg ugPG of curves in X such that αg connects ˚g to ˚ and αg1 g2 « pαg1 g2 q#αg2 .
Similarly there is a family tβh uhPH of curves in Y such that βh connects :h to :
and βh1 h2 « pβh1 h2 q#βh2 .
For every h P Hf there is a unique element of G, written hf P G, such that
h ¨ f “ f ¨ hf in M 1 pf q. For every h P Hf and every b P Bpf q, represented as a path
b : r0, 1s Ñ X with bp0q “ f p:q and bp1q “ ˚, define
bh :“ pf ˝ βh´1 q#pb ¨ hf q#αhf .
f
We clearly have pq ¨ b ¨ pqh “ q h ¨ bh ¨ ph , so Hf acts on Bpf q by congruences. We
´1
convert that right action to a left action by h b :“ bh .
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
25
For every c P M 1 pf q and every p P P , write c “ f g and define c p :“ g p “ gpg ´1 .
We clearly have cg p “ c pg pq.
Lemma 3.4. The biset of f : Y {Hf Ñ X{G is isomorphic to the crossed product
r f – Q ¸ Hf and
Bpf q ¸ M 1 pf q, which is Bpf q ˆ M 1 pf q as a set, with actions of H
r – P ¸ G given by
G
As a consequence,
pq, hq ¨ pb, cq ¨ pp, gq “ pq ¨ h pb ¨ c pq, h ¨ c ¨ gq.
f
r b Ă Bpf q ¸ M 1 pf q.
BpY {H Ð Y {Hf Ñ X{Gq – H
Hf
Proof. This is almost a tautology. The short exact sequence (44) splits, with section
r H
r on X,
r Yr can be identified with
h ¨ f ¨ g ÞÑ p1, hq ¨ fr ¨ p1, gq, and the actions of G,
concatenation of lifts of the paths αg , βh .
3.2. Self-similarity of G-spaces. We return in more detail the situation in which
r so as to keep Y, X for the original
Y, X are universal covers; we rename them to Yr , X
space.
Consider two pointed spaces pY, :q and pX, ˚q with H :“ π1 pY, :q and G :“
π1 pX, ˚q, and let f : Y Ñ X, be a continuous map. Recall that its biset is defined
by
(45)
H Bpf qG
:“ tβ : r0, 1s Ñ X | βp0q “ f p:q, βp1q “ ˚u{«,
with the natural actions by left- and right-concatenation. We thus naturally have
r with corresponding right actions, and left action given by composing
Bpf q Ď X,
loops via f .
The map f : Y Ñ X naturally lifts to a map Yr Ñ Y Ñ X, and every choice
of β P Bpf q defines uniquely, by the lifting property of coverings, a lift frβ : Yr Ñ
r with the property that r: ÞÑ β. Furthermore, we have the natural identities
X
r
fβ p´¨hq¨g “ frhβg , so that the biset Bpf q as defined in (45) is canonically isomorphic
to every biset M pfrβ q as defined in (40), when an arbitrary β P Bpf q is chosen, and
r of f is chosen.
to Bpfrq, when an arbitrary lift fr: Yr Ñ X
r choosing fr “ 1
If f : Y Ñ X is a covering, then we may assume Yr “ X;
r
gives the simple description Bpf q “ H GG . Recall that the biset of a correspondence pf, iq : Y Ð Z Ñ X is defined as Bpf, iq “ Bpiq_ b Bpf q. In the case
of a covering correspondence, in which f is a covering, we therefore arrive at
Bpf, iq “ Bpiq_ bπ1 pZq G. We shall give now a more explicit description of this
biset using covering spaces, just as we had in [3, (15)]
Bpf, iq “ tpδ : r0, 1s Ñ Y, p P Zq | δp0q “ :, δp1q “ ippq, f ppq “ ˚u{«
and the special case, when i : Z Ñ Y is injective, of
Bpf, iq “ tδ : r0, 1s Ñ Y | δp0q “ :, f pi´1 pδp1qqq “ ˚u{«.
Still assuming that f : Z Ñ X is a covering map, define
rH :“ tpδ, pq P Yr ˆ Z | ippq “ δp0qu,
(46)
Z
the fibred product of Yr with Z above Y , see Diagram (48) left. (Note that ZrH is
not the universal cover of Z.) In case i : Z Ñ Y is injective, we may write
rH “ tδ P Yr | δp0q P ipZqu,
Z
26
LAURENT BARTHOLDI AND DZMITRY DUDKO
rH is the full preimage of ipZq under the covering map Yr Ñ Y .
so Z
Proposition 3.5. If pf, iq is a covering correspondence, the following map defines
an isomorphism of H-spaces:
Φ:
H Bpf, iqG
r _ Ñ pZrH q_ given by Φppδ, pq b α_ q “ ppδ#pi ˝ α´1 Òp qq´1 , pq.
bX
f
For every b “ pδ, pq P Bpf, iq the map
r _ Ñ pZ
rH q_ given by α_ ÞÑ Φpb b α_ q
frb´1 : X
is the unique lift of the inverse of the correspondence Y Ð Z Ñ X mapping r̊ to b;
we have the equivariance properties
´1
frhbg
“ h ¨ frb´1 pg ¨ ´q.
(47)
The inverses and contragredients may seem unnatural in the statements above;
but are necessary for the actions to be on the right sides, and are justified by the
fact that we construct a lift of the inverse of the correspondence, rather than the
correspondence itself:
(48)
Yr _
ri
i
Y
pZrH q_
frb´1
Z
f
r_
X
X
rH ,
Proof. It is obvious that Φ is H-equivariant, and it is surjective: given pδ, pq P Z
p
p
´1
´1
´1
r with f ppq “ αp0q, and write δ “ pα Ò q #α Ò #δ,
choose a path α P X
f
f
expressing in this manner pδ, pq_ “ Φpppα´1 Òpf #δq´1 , α´1 Òpf p1qq b α_ q.
If Φppδ, pq b α_ q “ Φppδ 1 , p1 q b pα1 q_ q, then α and α1 start at the same point,
so α “ α1 g for some g P G, and we have then pδ, pq “ pδ 1 , p1 qg ´1 so pδ, pq b α_ “
pδ 1 , p1 q b pα1 q_ and Φ is injective.
It is easy to see that frb´1 is a lift of f ´1 . Conversely, every lift fr´1 of f ´1
maps r̊ to an element b P Bpf, iq because fr´1 pr̊_ q ends at an f -preimage of ˚ by
construction; and then fr´1 “ frb´1 by unicity of lifts. Finally, equivariance follows
´1
from frb´1 pgα_ q “ Φpb b gα_ q “ Φpbg b α_ q “ frbg
pα_ q.
r ì G is a
3.3. Planar discontinuous groups. A planar discontinuous group X
r for
group acting properly discontinuously on a plane, which will be denoted by X:
every bounded set V the set tg P G | V g X V ‰ Hu is finite.
Let X :“ pS 2 , A, ordq be an orbifold with non-negative Euler characteristic,
consider ˚ P S 2 zA a base-point, and write G :“ π1 pX, ˚q. Then the universal cover
rG of X is a plane endowed with a properly discontinuous action of G. We denote
X
r Ñ X the covering map.
by π : X
By the classification of surfaces, there are only two cases to consider: X “ C
with G a lattice in the affine group tz ÞÑ az ` b | a, b P C, |a| “ 1u, and X “ H the
upper half plane, with G a lattice in PSL2 pRq.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
27
Consider another orbisphere Y :“ pS 2 , C, ordq and a branched covering f : Y Ñ
X, and fix basepoints : P Y and ˚ P X with corresponding fundamental groups
H “ π1 pY, :q and G “ π1 pX, ˚q. Let H Bpf qG be the biset of f . As usual, we
view f as a correspondence Y Ð Z Ñ X with Z “ pS 2 , f ´1 pAq, ordq and i a
homeomorphism S 2 Ñ S 2 mapping a subset of f ´1 pAq to C. The fibred product
rH constructed in (46) is naturally a subset of the plane Yr , with orbispace points
Z
and punctures at all H-orbits of f ´1 pAqzC. We need the following classical result.
Theorem 3.6 (Baer [26, Theorem 5.14.1]). Every orientation-preserving homeomorphism of a plane commuting with a properly discontinuous group action is
isotopic to the identity along an isotopy commuting with the action.
Let us reprove [5, Theorem 7.9] using our language of group actions:
Corollary 3.7. Let two orbisphere maps f, g : pS 2 , C, ordq Ñ pS 2 , A, ordq have
isomorphic orbisphere bisets. Then f «C g.
In other words, the orbisphere biset of f : pS 2 , C, ordq Ñ pS 2 , A, ordq is a complete invariant of f up to isotopy rel C.
Proof. Let us write X “ pS 2 , A, ordq and Y “ pS 2 , C, ordq and G “ π1 pX, ˚q
and Y “ π1 pY, :q. We may represent f, g respectively by covering pairs pf, iq
and pg, iq, with coverings f, g : pS 2 , P, ordq Ñ pS 2 , A, ordq and i : pS 2 , P, ordq Ñ
rH its fibred product
pS 2 , C, ordq. Let us furthermore write Z “ pS 2 , P, ordq and Z
with Yr . Identifying Bpf, iq and Bpg, iq, choose b P Bpf, iq “ Bpg, iq, and let
r _ Ñ pZ
rH q_
frb´1 , r
gb´1 : X
be the corresponding lifts as in Proposition 3.5.
Since i is injective, the map pZrH q_ Ñ Z is a covering, so frb´1 and grb´1 are
coverings. We may therefore consider their quotient frb ˝ grb´1 , which is a wellr _ ý, and is normalized to preserve the base point r̊. By (47) it is
defined map pXq
a homeomorphism commuting with the action of G.
By Theorem 3.6 there is an isotopy pr
h´1
b,t qtPr0,1s of maps satisfying (47) between
´1
r
1 and fb ˝ grb .
Since the set of fixed points of G is discrete, r
h´1 preserves all fixed points of G
b,t
and therefore projects to an isotopy pht qtPr0,1s in X. We have h0 “ 1 and h1 ˝g “ f ,
so the maps f and g are isotopic rel B.
4. Geometric maps
Let M be a matrix with integer entries and with detpM q ą 1. We call M
exceptional if one of the eigenvalues of M lies in p´1, 1q; so the dynamical system
M : R2 ý has one attracting and one repelling direction.
For r P R2 , denote by xZ2 , ´z ` ry the group of affine transformations of R2
generated by translations z ÞÑ z ` v with v P Z2 and the involution z ÞÑ ´z ` r.
The quotient R2 {xZ2 , ´z `ry is a topological sphere, with cone singularities of angle
π at the four images of 21 pr ` Z2 q.
We call a branched covering f : S 2 ý a geometric exceptional map if f : S 2 ý
is conjugate to a quotient of an exceptional linear map M : R2 ý under the action
of xZ2 , ´z ` ry for some r P R2 ; in particular, p1 ´ M qr P Z2 . A Thurston map
f : pS 2 , A, ordq ý is called geometric if the underlying branched covering f : S 2 ý
28
LAURENT BARTHOLDI AND DZMITRY DUDKO
is either Böttcher expanding (see [4, Definition 4.1]; there exists a metric on S 2 that
is expanded everywhere except at critical periodic cycles), a geometric exceptional
map, or a pseudo-Anosov homeomorphism. We refer to the first two types as noninvertible geometric maps.
We may consider more generally affine maps z ÞÑ M z ` q on R2 , and then their
quotient by the group xZ2 , ´zy; the map z ÞÑ M z on R2 {xZ2 , ´z ` ry is converted
to that form by setting q “ pM ´ 1qr. Conversely, if 1 is not an eigenvalue of M ,
then we can always convert an affine map into a linear one, at the cost of replacing
the reflection ´z by ´z ` r in the acting group.
Lemma 4.1. Let M : R2 ý be exceptional. Then for every bounded set D Ă R2
containing p0, 0q there is an n ą 0 such that for every m ě n we have M ´m pDq X
Z2 “ tp0, 0qu. Moreover, n “ npDq can be taken with npDq ď log diampDq.
Proof. Let λ1 , λ2 be the eigenvalues of M , and let e1 and e2 be unit-normed eigenvectors associated with λ1 and λ2 . It is sufficient to assume that D is a parallelogram centered at p0, 0q with sides parallel to e1 and e2 :
(49)
D “ tv P R2 | v “ t1 e1 ` t2 v2 with |t1 | ď x and |t2 | ď yu.
In particular, the area of D is comparable to xy. Then M ´n pDq is again a parallelogram centered at p0, 0q with sides parallel to e1 and e2 .
We claim that there is δ ą 0 such that if D is a parallelogram of the form (49)
with areapDq ă δ, then D X Z2 “ tp0, 0qu. This will prove the lemma because
areapM ´n pDqq “ areapDq{pdet M qn and log diampDq — log areapDq.
Without loss of generality we assume x ą y, so D is close to Re1 . Let µ1 be the
slope of Re1 . Since M is exceptional, the numbers λ1 , λ2 , µ1 are quadratic irrational,
so they are not well approximated by rational numbers: there is a positive constant
C such that |µ1 ´ pq | ą qC2 for all pq P Q.
Suppose that D contains a non-zero integer point w “ pp, qq; so x ě |q|. Also,
w is close to Re1 , and in particular q ‰ 0 if δ is sufficiently small; we may suppose
q ą 0. It also follows that µ1 is close to pq . The distance from w to Re1 is
dpw, Re1 q ď y —
On the other hand,
Combining, we get
areapDq
.
x
ˇ
p ˇˇ
ˇ
dpw, Re1 q “ |w| sin ?pw, e1 q ě q ˇµ1 ´ ˇ.
q
ˇ
areapDq
δ
p ˇˇ dpw, Re1 q
ˇ
ď
ď 2,
ˇµ1 ´ ˇ ď
q
q
q¨x
q
and the claim follows for δ ăă C.
2
2
Corollary 4.2. Let f : pS , Aq ý be a geometric exceptional map, let p P S zA be
a periodic point with period n, and let γ P π1 pS 2 zA, pq be a loop starting and ending
at p. Let |γ| be the length of γ with respect to the Euclidean metric of the minimal
orbifold structure pPf , ordf q. If γ is trivial rel pPf , ordf q and m ą log |γ|, then the
lift γÒpf nm is a trivial loop rel A.
Proof. By passing to an iterate, we may assume that p is a fixed point of f . Since
f is geometric exceptional, we have a branched covering map π : R2 Ñ S 2 under
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
29
which f lifts to an exceptional linear map M , and we may assume πp0, 0q “ p. By
q
assumption, γÒp0,0q
is a trivial loop in R2 . By Lemma 4.1, the sets M ´m pγÒp0,0q
π
π
´1
and π pAq are disjoint for m ą log |γ|; hence γÒpf nm is a trivial loop rel A.
Geometric exceptional maps all admit a minimal orbifold modeled on R2 {xZ2 , ´zy,
which has cone singularities of angle π at the images of t0, 12 u ˆ t0, 12 u. We consider
this class in a little more detail:
4.1. p2, 2, 2, 2q-maps. A branched covering f : pS 2 , Pf , ordf q ý is of type p2, 2, 2, 2q
if |Pf | “ 4 and ordf pxq “ 2 for every x P Pf . In this case, f is isotopic to a quotient
of an affine map z ÞÑ M z ` q under the action of xZ2 , ´zy, see Proposition 4.6(B).
Lemma 4.3. Let M be a matrix with integer entries and detpM q ą 1. Denote by
λ1 and λ2 the eigenvalues of M , ordered as |λ2 | ě |λ1 | ą 0. Then the following are
all mutually exclusive possibilities:
‚ if M is exceptional, that is 0 ă |λ1 | ă 1 ă |λ2 | and λ1 , λ2 P R, then
M : R2 ý is expanding in the direction of the eigenvector corresponding to
λ2 and M : R2 ý is contracting in the direction of the eigenvector corresponding to λ1 ;
‚ if λ1 P t˘1u, then M 2 : R2 ý preserves the rational line tz P R2 | M z “ zu;
‚ the map M : R2 ý is expanding in all other cases; that is λ1 “ λ2 R R, or
λ1 , λ2 P R but |λ1 |, |λ2 | ą 1.
Proof. If M ’s eigenvalues are non-real, then λ1 “ λ2 and |λ1 | “ |λ2 |, so M is
expanding. If λ1 and λ2 are real, then they are of the same sign and their product
is greater than 1. The lemma follows.
The following lemma follows from [23, Main Theorem II]:
Lemma 4.4. If f : pS 2 , Pf , ordf q ý is doubly covered by a torus endomorphism
z ÞÑ M z ` q, then f is geometric if and only if f is Levy-free.
Proof. We consider the exclusive cases of Lemma 4.3. In the first case, f is geometric and z ÞÑ M z ` q preserves transverse irrational laminations (given by the
eigenvectors of M ), so z ÞÑ M z ` q admits no Levy cycle and a fortiori neither
does f .
In the second case, f is not geometric and admits as Levy cycle the projection
of the 1-eigenspace of M .
In the third case, f is expanding, and admits no Levy cycle by [4, Theorem A].
We shall mainly study p2, 2, 2, 2q-maps algebraically: we write the torus as R2 {Z2
and the p2, 2, 2, 2q-orbisphere as R2 {xZ2 , ´zy. Its fundamental group G is identified
with xZ2 , ´zy – Z2 ¸ t˘1u. The orbifold points are the images on the orbisphere
of t0, 21 u ˆ t0, 12 u. We start with some basic properties of G:
Proposition 4.5.
(50)
(A) Every injective endomorphism of G is of the form
v
M : pn, 1q ÞÑ pM n, 1q and pn, ´1q ÞÑ pM n ` v, ´1q
for some v P Z2 and some non-degenerate matrix M with integer entries.
We have
N w ˝ M v “ pN M qw`N v .
30
LAURENT BARTHOLDI AND DZMITRY DUDKO
There are precisely 4 order-2 conjugacy classes in G, each of the form
pa, ´1qG “ tpa ` 2w, ´1q | w P Z2 u for some a P t0, 1u ˆ t0, 1u Ă Z2 .
The set of order-2 conjugacy classes of G is preserved by M v .
(B) The automorphism group of G is tM v | det M “ ˘1u and is naturally
identified with Z2 ¸ GL2 pZq. The inner automorphisms of G are identified
with 2Z2 ¸ t˘1u, and the outer automorphism group of G is identified
with pZ{2Zq2 ¸ PGL2 pZq. The index-2 subgroup of positively oriented outer
automorphisms is Out` pGq “ pZ{2Zq2 ¸ PSL2 pZq.
(C) The modular group ModpGq is free of rank 2, and we have
ModpGq “ tM v | detpM q “ 1, M ” 1 pmod 2q, v P 2Z2 u{p˘1q2Z
2
– tM 0 | M ” 1 pmod 2qu{t˘1u.
(D) Two bisets BM v and BN w are isomorphic if and only if M “ ˘N and
pM v q˚ “ pN w q˚ as maps on order-2 conjugacy classes, if and only if M “
˘N and v ” w pmod 2Z2 q.
We remark that (50) also follows from Lemma 3.4.
Proof. (A). It is easy to check that M v defines an injective endomorphism. Conversely, let M 1 : G Ñ G be an injective endomorphism. Then M 1 p0, ´1q “ pv, ´1q
for some v P Z2 because all pw, ´1q have order 2. On the other hand, M 1 åZ2 ˆt1u “
M åZ2 for a non-degenerate matrix M with integer entries because M 1 is injective.
It easily follows that M 1 “ M v as given by (50).
The claims on composition and order-2 conjugacy classes of G follow from direct
computation.
2
(B). Follows directly from (A) and the identification of G with t˘1uZ .
`
(C). We use (B); the modular group of G is the subgroup of Out pGq “ pZ{2q2 ¸
PSL2 pZq that fixes the order-2 conjugacy classes. The action of pZ{2Zq2 on these
classes is simply transitive, so the set of order-2 classes may be put in bijection with
pZ{2q2 under the correspondence pa ` 2Z2 , ´1q Ø a; then the action of PSL2 pZq
on order-2 conjugacy classes is identified with the natural linear action (noting
that ´1 acts trivially mod 2). It follows that ModpGq is the congruence subgroup
tM P PSL2 pZq | M ” 1 pmod ˘ 2qu, and it is classical that it is a free group of
rank 2.
(D). The bisets BM v and BN w are isomorphic if and only the maps M v , N w are
conjugate by an inner automorphism; so the claimed description follows from (B).
We turn to p2, 2, 2, 2q-maps, and their description in terms of the above; we use
throughout G “ Z2 ¸ t˘1u:
Proposition 4.6. Let f be a p2, 2, 2, 2q-map with biset
Then
G BG
“ Bpf, Pf , ordf q.
(A) The biset B is right principal, and for any choice of b0 P B there exists an
injective endomorphism M v of G satisfying gb0 “ b0 M v pgq for all g P G.
(B) The map f is Thurston equivalent to a quotient of z ÞÑ M z ` 21 v : R2 ý
under the action of G.
(C) The map f is Levy obstructed if and only if M has an eigenvalue in t˘1u.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
31
(D) If f is not Levy obstructed then for every b0 P B, writing M v as in (A)
and (50), the map f is Thurston equivalent to the quotient of z ÞÑ M z : Z2 ý
by the action of xZ2 , z ÞÑ ´z ` ry í R2 for a vector r P R2 satisfying
M r “ r ` v. The G-equivariant map of Proposition 4.17 takes the form
(51)
Φ:
G BG
b xZ2 ,´z`ry R2 Ñ xZ2 ,´z`ry R2 ,
b0 b z ÞÑ M ´1 z.
Proof. (A). Since f is a self-covering of orbifolds, G BG is right principal. The claim
then follows from Proposition 4.5(A).
(B). We claim that the quotient map, denoted by f¯, has a biset isomorphic to
G BG . Indeed, for g P G it suffices to verify that
pM z ` 21 vq ˝ gpzq “ M v pgq ˝ pM z ` 12 vq.
If g “ pt, 1q then both parts are M z ` M t ` 21 v, and if g “ p0, ´1q then both parts
are ´pM z ` 12 vq. Therefore, f and f¯ have isomorphic orbisphere bisets because
marked conjugacy classes are preserved automatically by Proposition 4.5 (A). By
Corollary 3.7 the maps f and f¯ are isotopic.
(C). Suppose that M has an eigenvalue in t1, ´1u; let w be the eigenvector of M
associated with this eigenvalue. Consider the foliation Fw of R2 parallel to w. Then
Fw is invariant under z ÞÑ M z ` 21 v as well as under the action of xZ2 , z ÞÑ ´zy.
Therefore, the quotient map has invariant foliation F w “ Fw {G. There are two
leaves in F w connecting points in pairs in the post-critical set of the quotient map;
any other leave is a Levy cycle.
(D). Since detpM ´ 1q ‰ 0, there is a unique r such that M r “ r ` v. It is easy
to see (as in (C)) that the quotient map in (E) has a biset isomorphic to G BG . By
Corollary 3.7 the quotient map is conjugate to f , and (51) is immediate.
Corollary 4.7. Let f be a p2, 2, 2, 2q-map. Then its biset Bpf : pS 2 , Pf , ordf q ýq
is of the form BM v for an endomorphism M v of G :“ Z2 ¸ t˘1u, namely it is G
qua right G-set, with left action given by g ¨ h “ M v pgqh.
Let us also recall how the biset of a p2, 2, 2, 2q-map is converted to the form
BM v . First, the fundamental group G is identified with Z2 ¸ t˘1u. The group
G has a unique subgroup H of index 2 that is isomorphic to Z2 , so H is easy to
find. The restriction of the biset to H yields a 2 ˆ 2-integer matrix M ; and the
translation part v is found by tracking the peripheral conjugacy classes. Note that
this procedure applies as well to non-invertible maps as to homeomorphisms; and
that the map is orientation-preserving precisely if detpM q ą 0.
4.2. Homotopy pseudo-orbits and shadowing. Let f : S 2 ý be a self-map,
and let I be an index set together with a index map also written f : I ý. An
I-poly-orbit pxi qiPI is a collection of points in S 2 such that f pxi q “ xf piq . If all
points xi are distinct, then pxi qiPI is an I-orbit.
Thus a poly-orbit differs from an orbit only in that it allows repetitions. Every
poly-orbit has a unique associated orbit, whose index set is obtained from I by
identifying i and j whenever xi “ xj . Note that we allow I “ N, I “ Z and
I “ Z{nZ as index sets, with f piq “ i ` 1, as well as I “ t0, . . . , m, . . . , m ` nu with
f piq “ i ` 1 except f pm ` nq “ m, an orbit with period n and preperiod m.
We shall consider a homotopical weakening of the notion of orbits. Our treatment
differs from [13] in a subtle manner (see below); recall that β «A,ord γ means that
the curves β, γ are homotopic in the orbispace pS 2 , A, ordq:
32
LAURENT BARTHOLDI AND DZMITRY DUDKO
Definition 4.8 (Homotopy pseudo-orbits). Let f : pS 2 , A, ordq ý be an orbisphere
self-map, and let I be a finite index set together with an index map also written
f : I ý.
An I-homotopy pseudo-orbit is a collection of paths
pβi qiPI with βi : r0, 1s Ñ S 2 zA satisfying βf piq p0q “ f pβi p1qq.
Two homotopy pseudo-orbits pβi qiPI and pβi1 qiPI are homotopic, written pβi qiPI «A,ord
pβi1 qiPI , if βi «A,ord βi1 for all i P I. In particular, βi p0q “ βi1 p0q and βi p1q “ βi1 p1q.
Two homotopy pseudo-orbits pβi qiPI and pγi qiPI are conjugate, written pβi qiPI „
pγi qiPI , if there is a collection of paths pℓi qiPI with
ℓi p0q “ βi p0q,
β p1q
ℓi p1q “ γi p0q and βi #ℓf piq Òf i
«A,ord ℓi #γi .
Poly-orbits are special cases of homotopy pseudo-orbits, in which the paths βi are
all constant.
△
Remark 4.9. A homotopy pseudo orbit can also be defined for an infinite I with
the assumption that pℓi qiPI and pβi qiPI in Definition 4.8 have uniformly bounded
length. Then Theorem 4.22 still holds.
In a homotopy pseudo-orbit, the curves pβi qiPI encode homotopically the difference between xi :“ βi p0q and a choice of preimage of xf piq . Note that Ishii-Smillie
define in [13, Definition 6.1] homotopy pseudo-orbits by paths βi connecting f pxi q
to xf piq ; their βi may be uniquely lifted to paths βi from xi to an f -preimage of
xf piq as in Definition 4.8, but our definition does not reduce to theirs, since our βi
are defined rel pA, ordq, not rel pf ´1 pAq, f ˚ pordqq.
βi
yi ¨
f
xf piq ¨
xi ¨
ℓi
βf piq
f
γi
f
¨ xf 2 piq
ℓf piq
yf piq ¨
γf piq
f
ℓf 2 piq
¨ yf 2 piq
Choose a length metric on S 2 , and define the distance between homotopy pseudoorbits by
`
˘
(52)
d pβi qiPI , pγi qiPI :“ max dpγi ptq, βi ptqq.
iPI,tPr0,1s
The following result states that the set of conjugacy classes of homotopy pseudoorbits is discrete:
Lemma 4.10 (Discreteness). Let pβi qiPI be a homotopy pseudo-orbit of f : pS 2 , A, ordq ý.
Then there is an ε ą 0 such every homotopy pseudo-orbit at distance less than ε
from pβi qiPI is conjugate to it.
Proof. Set ε “ mina‰bPA dpa, bq. Consider a homotopy pseudo-orbit pγi qiPI at distance δ ă ε from pβi qiPI . Connect βi p0q to γi p0q by a path ℓi of length at most δ.
Since SzA is locally contractible space, the curve ℓi is unique up to homotopy and
β p1q
βi #ℓf piq Òf i «A,ord ℓi #γi .
We recall that A8 denotes the union of all periodic cycles containing critical
points.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
33
Definition 4.11 (Shadowing). A homotopy pseudo-orbit pβi qPI shadows rel pA, ordq
a poly-orbit ppi qiPI in SzA8 if they are conjugate; namely if there are curves ℓi connecting βi p0q to pi that lie in S 2 zA except possibly their endpoints and such that
for every i P I we have
β p1q
ℓi «pA,ordq βi #ℓf piq Òf i
.
△
Lemma 4.12. The homotopy pseudo-orbit pβi qPI shadows the poly-orbit ppi qiPI if
and only if every neighbourhood pUi qiPI of ppi qiPI contains a homotopy pseudo-orbit
pβi1 qiPI conjugate to pβi qiPI ; namely, βi1 Ă Ui Q pi for all i P I.
Proof. If pβi q can be conjugated into a small enough neighbourhood of ppi qiPI , then
it is conjugate to ppi qiPI by Lemma 4.10. The converse is obvious.
Proposition 4.13. Suppose f : pS 2 , A, ordq ý is a geometric non-invertible map.
Then a periodic pseudo-orbit pβi qiPI shadows ppi qiPI rel pA, ordq if and only if pβi qiPI
shadows ppi qiPI rel pPf , ordf q.
Proof. Clearly if pβi qiPI shadows an orbit ppi qiPI rel pA, ordq then pβi qiPI shadows ppi qiPI rel pPf , ordf q. Conversely, suppose pβi qiPI shadows an orbit ppi qiPI rel
β p1q
pPf , ordf q, so there are paths ℓi connecting βi p0q to pi with ℓi «Pf ,ordf βi #ℓf piq Òf i .
β p1q
i
Thus ℓ´1
is a loop which is trivial rel pPf , ordf q, but may not be
i #βi #ℓf piq Òf
trivial rel A.
Consider the following pullback iteration. Set pβi0 qiPI :“ pβi qiPI and pℓ0i qiPI :“
0
pℓi qiPI . Define
`
`
βin´1 p1q ˘
βi p1q ˘
.
and pℓni qiPI :“ ℓn´1
Ò
pβin qiPI :“ βfn´1
f piq Òf
piq f
iPI
iPI
β p1q
Clearly the pβin qiPI are all conjugate. Observe that pℓni q´1 #βi #ℓnfpiq Òf i
passing through pi and
β p1q
pℓni q´1 #βi #ℓnfpiq Òf i
is a degree-1 preimage of
β p1q
If f is expanding, then the diameter of pℓni q´1 #βi #ℓnfpiq Òf i
is a loop
βi p1q
pℓn´1
q´1 #βi #ℓn´1
.
i
f piq Òf
tends to 0 expo-
β p1q
pℓni q´1 #βi #ℓnfpiq Òf i
nentially fast, hence
is trivial rel pA, ordq for all sufficiently
large n, and the claim follows.
β p1q
If f is exceptional, then pℓni q´1 #βi #ℓnfpiq Òf i is trivial rel pA, ordq for all sufficiently large n by Corollary 4.2.
At one extreme, homotopy pseudo-orbits include poly-orbits, represented as constant paths. At the other extreme, homotopy pseudo-orbits may be assumed to
consist of paths all starting at the basepoint ˚. As such, these paths represent
elements of the biset Bpf q of f , see 1.
4.3. Symbolic orbits. We shall be interested in marking periodic orbits of regular
points. These are conveniently encoded in the following simplification of portraits
of bisets (in which the subbisets are singletons and therefore represented simply as
elements):
Definition 4.14 (Symbolic orbits). Let I be a finite index set with self-map
f : I ý, and let B be a G-biset. An I-symbolic orbit is a sequence pbi qiPI of
elements of B, and two I-symbolic orbits pbi qiPI and pci qiPI are conjugate if there
exists a sequence pgi qiPI in G with gi ci “ bi gf piq for all i P I.
△
34
LAURENT BARTHOLDI AND DZMITRY DUDKO
Lemma 4.15. Every homotopy pseudo-orbit can be conjugated to a symbolic orbit,
unique up to conjugacy, in Bpf, A, ordq.
Proof. Given pβi qiPI a homotopy pseudo-orbit, choose paths ℓi from βi p0q to ˚
βi p1q
and define γi “ ℓ´1
. Then γi P Bpf, A, ordq and pβi qiPI „ pγi qiPI .
i #βi #ℓf piq Òf
Furthermore another choice of paths pℓ1i qiPI differs from pℓi qiPI by ℓ1i “ ℓi gi for some
β p1q
gi , so the symbolic orbits pγi qiPI and ppℓ1i q´1 #βi #ℓ1f piq Òf i qiPI are conjugate.
Let f : pS 2 , Pf , ordf q ý be an expanding map, and let G BG be its biset. Recall
that the Julia set J pf q of f is the accumulation set of preimages of a generic point
˚,
č ď
J pf q :“
f ´m p˚q.
ně0 měn
Every bounded sequence b0 b1 ¨ ¨ ¨ P B b8 defines an element of J pf q as follows:
c
p1q
set c0 “ b0 and ci “ ci´1 #bi Òfi´1
for all i ě 1; then limnÑ8 cn p1q converges
i
to a point ρpb0 b1 ¨ ¨ ¨ q P J pf q. The following proposition directly follows from the
definition:
Proposition 4.16 (Expanding case). Suppose f : pS 2 , Pf , ordf q ý is an expanding
map with orbisphere biset G BG , and let pbi qiPI be a finite symbolic orbit. Let Σ be
a generating set of G BG containing all bi and let ρ : Σ`8 Ñ J pf q be the symbolic
encoding defined above. Then pbi qiPI shadows pρpbi bf piq qbf 2 piq . . . qiPI .
This proposition is useful to solve shadowing problems (namely, determining
when two symbolic orbits shadow the same poly-orbit) using language of automata.
It also follows from the proposition that every orbit homotopy pseudo-orbit shadows a unique orbit in J pf q.
Proposition 4.17 (Shadowing and universal covers). Let f : pS 2 , A, ordq ý be an
r Ñ pS 2 , A, ordq be the universal covering
orbisphere map with biset G BG , let π : G U
2
r be the G equivariant map defined
r
map of pS , A, ordq, and let Φ : G BG b G U ãÑ G U
by Proposition 3.5.
r such that π and Φ extend to continuous
r ` of G U
Then there is a completion G U
`
2
8
r ` with the following property:
r ` Ñ GU
r Ñ S zA and Φ : G BG b G U
maps π : G U
a symbolic orbit pbi qiPI shadows an orbit pxi qiPI in S 2 zA8 if and only if
r ` with πpr
Φpbi b x
rf piq q “ x
ri for some x
ri P U
xi q “ xi .
r` “ U
r.
Furthermore, if ordpaq ă 8 for every a P AzA8 then we may take U
r ` , add to U
r all limit points of parabolic elements corresponding
Proof. To define U
to small loops around punctures a P AzA8 with ordpaq “ 8. The extension of π
r ` zU
r.
and Φ is given by continuity; π is a branched covering with branch locus U
If pbi qiPI shadows pxi qiPI , then there are curves
x
ri : r0, 1s Ñ pS 2 , A, ordq with x
ri p0q “ ˚ and x
ri p1q “ xi
b p1q
such that x
r´1
xf piq Òfi
is a homotopically trivial loop. This exactly means
i #bi #r
that Φpbi b x
rf piq q “ x
ri . Conversely, if Φpbi b x
rf piq q “ x
ri for all i P I then pbi qiPI
shadows pπpr
xi qqiPI .
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
35
Proposition 4.18 (Shadowing in the p2, 2, 2, 2q case). Using notations of Proposition 4.6, suppose f is a p2, 2, 2, 2q geometric non-invertible map and
Φ:
G BG
b xZ2 ,´z`ry R2 Ñ xZ2 ,´z`ry R2 : pb0 , zq ÞÑ M ´1 z
is as in Proposition 4.6(D). Then for every symbolic orbit pbi qiPI of G BG there is
a unique collection pri qiPI of points in R2 such that ri “ Φpbi b rf piq q. The points
ri are solutions of linear equations (53).
By Proposition 4.17, the image of pri qiPI under the projection π : R2 Ñ R2 {G «
pS , Pf , ordf q is the unique poly-orbit shadowed by pbi qiPI .
The proof will use the following easy fact.
2
Lemma 4.19. If M is non-invertible geometric, then | detpM n ` ǫIq| ě 1 for every
n ě 1 and every ǫ P t˘1u.
Proof. If λ1 , λ2 are M ’s eigenvalues, then Lemma 4.3 gives
| detpM n ` ǫ1q| “ |pλn1 ` ǫqpλn2 ` ǫq| ě 1.
2
Proof of Proposition 4.18. Write bi “ b0 gi . Recall that every gi acts on R as
z ÞÑ ǫi z ` ti with ǫi P t˘1u. We get the following system of linear equations:
M ´1 pǫi rf piq ` ti q “ ri ,
i P I.
By splitting f : I ý into grand orbits and eliminating variables we arrive at equations of the form
(53)
p1 ` θM n qri “ t with θ P t˘1u
and n the period of an orbit and t P R2 some parameters depending on ǫi and ti .
By Lemma 4.19, the system (53) has a unique solution.
Consider an I-symbolic orbit pbi qiPI shadowing a poly-orbit pxi qiPI . This means
(see Definition 4.11) that there are curves pℓi qiPI connecting ˚ to xi such that
bi p1q
ℓ´1
is a trivial loop rel pA, ordq. The local group Gxi ď π1 pX, ˚q “ G
i #bi #ℓf piq Òf
consists of loops of the form
(54)
ℓi r0, 1 ´ εs#α#ℓ´1
i r0, 1 ´ εs,
α is close to xi .
If xi R A, then Gxi is a trivial group; otherwise Gxi is an abelian group of
order ordpxi q. If pA, ordq “ pPf , ordf q, then Gxi is a finite abelian group. Clearly,
pgi bi hi qi also shadows xi for all gi P Gxi and all hi P Gxi`1 . Conversely:
Lemma 4.20. Let f : pS 2 , A, ordq ý be a geometric non-invertible map. Suppose
that the symbolic orbits pbi qiPI and pci qiPI shadow pxi qi . Let Gxi be the local group
associated with pbi qi . Then there are hi P Gxf piq such that pci qiPI is conjugate to
pbi hi qiPI .
If I consists only of periodic indices, then pci qiPI is conjugate to pgi bi qiPI for
some gi P Gxi .
b p1q
i
Proof. By conjugating pci qiPI we can assume that ℓ´1
is a trivial loop; i.e.
i ci ℓi Òf
local groups associated with pci qiPI coincide with whose associated with pbi qiPI . It
is now routine to check that pbi qiPI and pci qiPI differ by the action of local groups.
Indeed, bi is of the form
β p1q
ℓi r0, 1 ´ εs#βi #pℓf piq r0, 1 ´ εsq´1 Òf i
,
βi is close to xi ,
36
LAURENT BARTHOLDI AND DZMITRY DUDKO
and, similarly, ci is of the form
γ p1q
ℓi r0, 1 ´ εs#γi #pℓf piq r0, 1 ´ εsq´1 Òf i
,
γi is close to xi .
Let αi be the image of βi´1 #γi and define hi to be ℓf piq r0, 1 ´ εs#αi#ℓ´1
f piq r0, 1 ´ εs;
compare with (54). Then ci “ bi hi for all i P I.
If I has only periodic indices, then βi´1 and γi end at the same point and we set
αi :“ γi #βi´1 and gi :“ ℓi r0, 1 ´ εs#αi #ℓ´1
i r0, 1 ´ εs. We obtain ci “ gi bi for all
i P I.
Corollary 4.21. Let f : pS 2 , Pf , ordf q ý be a geometric non-invertible map. For
every poly-orbit there are only finitely many symbolic orbits that shadow it. Therefore, there are only finitely many conjugacy classes of portraits in Bpf q.
Theorem 4.22. Let f : pS 2 , A, ordq ý be a non-invertible geometric map. Then
the shadowing operation defines a map from conjugacy classes of symbolic finite
orbits onto poly-orbits in S 2 zA8 . If pA, ordq “ pPf , ordf q, then the shadowing map
is finite-to-one.
Proof. Follows from Propositions 4.13, 4.16, 4.18 and Corollary 4.21.
4.4. From symbolic orbits to portraits of bisets. The centralizer of a symbolic
finite orbit pbi qiPI is the set of pgi qiPI P GI such that gi bi “ bi gf piq for all i P I.
Lemma 4.23. If G BG is the biset of a geometric non-invertible map f and pbi qiPI is
a symbolic
finite orbit shadowing a poly-orbit pxi qiPI , then its centralizer is contained
ś
in iPI Gxi , where Gxi are the local groups associated with pbi qiPI .
Proof. By Definition 4.11 there are curves pℓi qiPI connecting ˚ to xi such that
bi p1q
ℓ´1
is a trivial loop rel pA, ordq. Suppose that pgi qiPI P GI centralizes
i #bi #ℓi Òf
gf piq Òxf i is isotopic to gri rel pA, ordq. By the
pbi qiPI . Set gri :“ ℓ´1
i #gi #ℓi . Then r
expanding property of f , orś
Corollary 4.2 if f is exceptional, gri is trivial rel pA, ordq.
This implies that pgi qiPI P iPI Gxi .
r “
Consider a portrait of bisets pGa , Ba qaPAr in G BG . We may decompose A
n
A \ F \ I with f˚ pF q Ď A for n " 0 and f˚ pIq Ď I. Then for every i P I the group
Gi is trivial and Bi “ tbi u is a singleton. We obtain the symbolic orbit pbi qiPI
which is the essential part of pGa , Ba qAr :
Lemma 4.24. The relative centralizer ZD ppGa , Ba qaPAr q (see §2.8) is isomorphic
to the centralizer of pbi qiPI via the forgetful map pgd qdPD Ñ pgi qiPI .
Let pGa , Ba qaPAr and pG1a , Ba1 qaPAr be two portraits of bisets with associated symbolic orbits pbi qiPI and pb1i qiPI . Assume that Ga “ G1a and Ba “ Ba1 for all a P A.
Then pGa , Ba qaPAr and pG1a , Ba1 qaPAr are conjugate if and only if
(1) pbi qiPI and pb1i qiPI are conjugate; and
(2) G bGd Bd “ G bGd Bd1 for every d P F .
This reduces the conjugacy problem of portrait of bisets to the conjugacy problem of symbolic orbits; indeed Condition (2) is easily checkable. Note that every
r imposes a finite condition on conjugacy and centralizers: for
preperiodic point in A
points attracted to A, Condition (2) imposes a congruence condition modulo the
action on t¨u bG B on the conjugator; for preperiodic points in I, conjugacy again
amounts (by Condition (1) and Definition 4.14) to a congruence condition modulo
the action on t¨u bG B.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
37
Proof. If d P D with f˚ pdq R D, then from gd Bd “ Bd gf˚ pdq “ Bd follows that
gd “ 1. Induction on the escaping time to A gives gd “ 1 for all d P F .
Similarly, if d P D with f˚ pdq R D, then there is a gd P G with gd Bd “ Bd1 “
1
Bd gf˚ pdq if and only if G bGd Bd “ G bGd Bd1 . By induction on the escaping
time, pGa , Ba qaPA\F and pG1a , Ba1 qaPA\F are conjugate if and only if Condition (2)
holds.
We are now ready to show that a geometric map, equipped with a portrait of
bisets, yields a sphere map — possibly with some points infinitesimally close to each
other. A blowup of a two-dimensional sphere is a topological sphere Sr2 equipped
with a monotone map Sr2 Ñ S 2 , namely a continuous map under which preimages
of connected sets are connected. We remark that arbitrary countable subsets of S 2
may be blown up into disks.
r be
Proposition 4.25. Let f : pS 2 , Aq ý be a non-invertible geometric map, let A
r
a set containing a copy of A, let f˚ : A ý be a symbolic map which coincides with
r and let pBa q r be a portrait of bisets in Bpf q.
f on the subset A Ď A,
aPA
r Ñ S 2 extending the identity on A, and a
Then there exists a unique map e : A
2
2
r
blowup b : S Ñ S , with the following properties. The locus of non-injectivity of e
is disjoint from A8 , and b blows up precisely at the grand orbits of
r : x “ epaq “ epa1 qu
tx | Da ‰ a1 P A
r Ă Sr2
replacing points by disks on which the metric is identically 0 and all points in A
2
2 r
r
r
r
are disjoint. The maps f˚ : A ý and f : pS , Aq ý extend to a map f : pS , Aq ý,
which is semiconjugate to f via b, and whose minimal portrait of bisets projects to
pBa qaPAr .
r “ A \ I \ J with
Proof. Let us set G :“ π1 pS 2 zA, ˚q. We may decompose A
f˚ pIq “ I and f˚n pJq Ď A \ I for n " 0. On A, we naturally define e as the identity.
If i P I, then i is f˚ -periodic and the bisets Bi , Bf˚ piq , . . . determine a homotopy
pseudo-orbit which shadows a unique periodic poly-orbit in S 2 , by Theorem 4.22.
Thus e is uniquely defined.
We now blow up the grand orbits of all points in S 2 which are the image of more
than one point in A \ I under e, replacing them by a disk on which the metric
is identically 0. We inject A \ I arbitrarily in the blown-up sphere Sr2 , and now
identify A \ I with its image in the blowup.
Let us next extend f to a self-map fr of Sr2 so that pBa qaPA\I is the induced
portrait of bisets. We first do it by arbitrarily mapping the disks to each other by
homeomorphisms restricting to f˚ on A \ I. Let pBa1 qaPA\I be the projection of
the minimal portrait of bisets of fr via π1 pSr2 , A \ Iq Ñ G. Consider i P I. By
Lemma 4.20 (the periodic case), we can assume that bi “ hi b1i with hi P Gepiq .
(Note that if epiq R A, then hi “ 1 and bi “ b1i .) Let m1 P BraidpSr2 zA, Iq be a
preimage of phi qiPI under EI (see (19)) and set m :“ pushpm1 q; it is defined up to
pre-composing with a knitting element. Since hi P Gepiq , we can assume that m
is identity away from the blown up disks. Then pBa qaPA\I is a portrait of bisets
induced by mfr.
Consider next j P J and assume that f˚ pjq has already been defined. The biset
Bj , and more precisely already its image in t¨u bG Bpf q, see Definition 2.15(B),
38
LAURENT BARTHOLDI AND DZMITRY DUDKO
determines the correct fr-preimage of f˚ pjq P Sr2 that corresponds to j, and thus
determines epjq uniquely.
4.5. Promotions of geometric maps. Let X :“ pS 2 , A, ordq be an orbisphere,
and consider a subset D Ă S 2 zA. Recall from (11) the quotient ModpX|Dq of
mapping classes of pS 2 , A \ Dq by knitting elements; there, we called two maps
f, g : pX, Dq ý knitting-equivalent, written f «A|D g, if they differ by a knitting
element in knBraidpX, Dq. We shall show that knitting-equivalent rigid maps are
conjugate rel A \ D:
Theorem 4.26 (Promotion). Suppose f, g : pX, Dq ý are orbisphere maps, and
assume that either
‚ g is geometric non-invertible, or
‚ g m pDq Ď A for some m ě 0.
Then every conjugacy h P ModpX|Dq between f and g rel A|D lifts to a unique
conjugacy r
h P ModpS 2 , A \ Dq between f and g rel A \ D such that r
h «A|D h.
Proof. We begin by a
Lemma 4.27. Under the assumption on g, consider that b P BraidpS 2 zA, Dq. If
for every m ě 1 the element b is liftable through g m , then pg ˚ qm b is trivial for all
m ą log |g|, where | | is the word metric.
Proof. Consider first the case that g is expanding. Then the lengths of curves
pg ˚ qm pbqp´, aq tend to 0 as m Ñ 8, so pg ˚ qm pbq “ 1 for sufficiently large m. If
g is exceptional, then Corollary 4.2 replaces the expanding argument. If finally
g m pEq Ď A, then pg ˚ qm pbq “ 1 for the same m.
We resume the proof. Let r
h0 P pModpS 2 , A \ Dq be any preimage of h under
2
the forgetful map ModpS , A \ Dq Ñ ModpX|Dq. Then r
h0 f r
h´1
«A\D h´1
0
1 g
´1
r
r
r
r
for some h1 P knBraidpX, Dq. Setting h1 :“ h1 h0 we get h1 f h1 «A\D h´1
2 g,
´1
is
the
lift
of
h
through
g.
Continuing
this
process
we
eventually
get
where h´1
2
1
´1
´1
r
r
hm f hm «A\D g because a the corresponding lift of h1 is trivial by Lemma 4.27.
We have shown the existence of r
h.
If r
h1 is another promotion of h, then r
h1 and r
h differ by a knitting element
h1 “
commuting with f . By Lemma 4.27, that knitting element is trivial, hence r
r
h.
Corollary 4.28. Let ModpS 2 ,A\Dq M pf qModpS 2 ,A\Dq Ñ ModpX|Dq M pf |DqModpX|Dq
be the inert map forgetting the action of knitting elements (see Proposition 2.12).
Suppose that either f is geometric non-invertible or f m pDq Ď A for some m ě 0.
Then f, g are conjugate in M pf q if and only if their images are conjugate in
M pf |Dq. Moreover, the centralizers Zpf q and Zpf |Dq (see §2.8) are naturally
isomorphic via the projection ModpS 2 , A \ Dq Ñ ModpX|Dq.
4.6. Automorphisms of bisets. Recall that the automorphism group of a biset
H BG is the set AutpBq of maps τ : B ý satisfying τ phbgq “ hτ pbqg for all h P
H, g P G.
Proposition 4.29. If H BG is a left-free, right-transitive biset and H is centreless,
then AutpBq acts freely on B, so B{ AutpBq is also a left-free, right-transitive HG-biset.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
39
Proof. AutpBq acts by permutations on t¨u bH B, and commutes with the right
G-action, which is transitive, so AutpBq{ kerpactionq acts freely. Consider now
τ P AutpBq that acts trivially on t¨u bH B, and consider b P B. We have τ pbq “ tb
for some t P H, and for all h P H there exists g P G with hb “ bg, because B is
right-transitive. We have
htb “ hτ pbq “ τ phbq “ τ pbgq “ τ pbqg “ tbg “ thb,
so t P ZpHq “ 1 and therefore τ “ 1. It follows that AutpBq acts freely on B.
Corollary 4.30 (No biset automorphisms). For an orbisphere biset
is non-cyclic, we have AutpBq “ 1.
In the dynamical situation
G BG
H BG
with H
is a cyclic biset if and only if G is a cyclic group.
Proof. Since B is an orbisphere biset, it is left-free and right-transitive; and since it
is not cyclic, H is a non-abelian orbisphere group and in particular is centreless, so
Proposition 4.29 applies. Now B cannot have a proper quotient, because conjugacy
classes of elements of H appear exactly once in wreath recursions of peripheral
elements, by [5, Definition 2.6(SB3 ) of orbisphere bisets].
Corollary 4.31. Suppose f, g : pS 2 , C, ordq Ñ pS 2 , A, ordq are isotopic orbisphere
maps. Let Bpf q and Bpgq be the bisets of f and g with respect to the same base
points : P S 2 zC and ˚ P S 2 zA. Then there is a unique isomorphism between Bpf q
and Bpgq.
Proof. Consider an isotopy pft qtPr0,1s rel A from f to g. It induces a continuous motion Bpft , A, ord, ˚q of the bisets of ft . Therefore, all Bpft , A, ord, ˚q are isomorphic.
This shows existence, the uniqueness follows from Corollary 4.30.
Theorem 4.32 (Rigidity). Suppose f : pS 2 , A, ordq ý is a geometric map. If
h : pS 2 , Aq ý commutes with f , is the identity in some neighbourhood of A8 , and
is isotopic to 1 rel A, then h “ 1.
Proof. Fix a basepoint ˚ P S 2 zA and the fundamental group G “ π1 pS 2 zA, ˚q.
Since h is isotopic to the identity rel A, there is a path ℓ : r0, 1s Ñ S 2 zA from ˚ to
hp˚q such that γ « ℓ#ph ˝ γq#ℓ´1 for all γ P G. Define then
h˚ :
G Bpf qG
ý,
hpβp1qq
h˚ pβq :“ ℓ#ph ˝ βq#pℓ´1 qÒf
.
Since h commutes with f , this defines an automorphism of Bpf q. By Corollary 4.30,
it is the identity on Bpf q. It also fixes, therefore, all conjugacy classes in Bpf qbn
for all n P N. By Theorem 4.22, these are in bijection with periodic orbits of f . It
follows that h is the identity on all periodic points of f , and therefore on f ’s Julia
set. Since furthermore h is the identity near A8 , and every point of S 2 zA either
belongs to J pf q or gets attracted to A8 , we get h “ 1 everywhere.
4.7. Weakly geometric maps. We are now ready to show that maps whose restriction to their minimal orbisphere is geometric are of a particularly simple form.
Definition 4.33 (Tunings). Let f : pS 2 , Aq ý be a Thurston map. A tuning
multicurve for a f is an f -invariant1 multicurve C such that f ´1 pC q does not
contain nested components rel f ´1 pAq; namely, the adjacency graph of components
of S 2 zf ´1 pC q is a star.
1In the sense that f ´1 pC q equals C rel A.
40
LAURENT BARTHOLDI AND DZMITRY DUDKO
A tuning is an amalgam of orbisphere bisets in which the graph of bisets is a star.
Equivalently, it is an amalgam of Thurston maps along a tuning multicurve.
△
Every tuning has a central map, corresponding to the centre of the star, as well
as satellite maps, corresponding to its leaves. Furthermore, unless the star has two
vertices, its central map is uniquely determined.
Definition 4.34 (Weakly geometric maps). A tuning by homeomorphisms is a
tuning in which all satellite maps are homeomorphisms.
A map is weakly geometric if it is a (possibly trivial) tuning by homeomorphisms
in such a manner that the central map is isotopic to a geometric map.
△
Theorem 4.35. Let f : pS 2 , Aq ý be a non-invertible Thurston map. Then f is
weakly geometric if and only if its restriction f : pS 2 , Pf , ordf q ý is isotopic to a
geometric map.
Proof. If f is weakly geometric, then its central map f is isotopic to a geometric
map, and f is isotopic to f rel Pf .
Conversely, assume that the restriction f : pS 2 , Pf , ordf q ý of f is geometric,
and consider its portrait of bisets pBa qaPA induced from the minimal portrait of
bisets of f . By Proposition 4.25, there exists a blown-up sphere Sr2 and extension
fr of f , such that the portraits of bisets of f and fr are conjugate portraits of bisets
in Bpf q. The tuning multicurve we seek is the boundary of the blowup disks.
By Theorem 2.19, the bisets Bpf q and Bpfrq differ by a knitting element, so
r
we may write fr “ mf for some m P knBraidpS 2 zA, AzAq.
By Proposition 2.8,
the class m is liftable arbitrarily often through f , and by Lemma 4.27 its lift
becomes eventually trivial; in other words, we have a relation n´1 frn “ mf for
some m with trivial image in BraidpS 2 zAq, and therefore n is a product of mapping
classes in the infinitesimal disks. Its restriction to these disks yields the required
homeomorphisms of the tuning.
We now turn to the algebraic aspects of geometric maps.
Definition 4.36. An orbisphere biset is geometric if it is the biset of a geometric
map. A biset B is weakly geometric if its projection B to its minimal orbisphere
group is geometric.
△
Corollary 4.37. A Thurston map f is [weakly] geometric if and only if its biset
is [weakly] geometric.
The following result was already obtained by Selinger and Yampolsky in the case
of torus maps [23]:
Corollary 4.38. A non-invertible Thurston map is isotopic to a geometric map if
and only if it is Levy-free.
Proof. Let f be a non-invertible Thurston map. If f admits a Levy cycle, then
either f shrinks no metric under which a curve in the Levy cycle is geodesic, or f
is doubly covered by a torus endomorphism with eigenvalue ˘1; in all cases, f is
not geometric.
Conversely, if f is Levy-free, then its restriction f to pS 2 , Pf , ordf q is still Levyfree. By [4, Theorem A] (for tExpu maps) or Lemma 4.4 (for tGTor/2u maps) the
map f is geometric. By Proposition 4.25, there is a map e : A Ñ S 2 whose image is
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
41
preserved by f . If e were not injective, there would be a Levy cycle for f consisting
of curves surrounding elements of A with same image under e. Now if e is injective
then the tuning constructed in Theorem 4.35 is trivial, so f is geometric.
4.8. Conjugacy and centralizer problem. We are now ready to show how conjugacy and centralizer problems may be solved in geometric bisets. We review the
algebraic interpretation of expanding maps: by [4, Theorem A], a map is expanding if and only if its minimal biset is contracting. We recall briefly that a left-free
G-G-biset B is contracting if, for every basis X Ď B, there exists a finite subset
N Ď G with the following property: for every g P G there is n P N such that
X n g P N X n Ď B bn .
The minimal such subset N is called the nucleus associated with pB, Xq. Note
that different bases yield different nuclei, but that finiteness of N is one basis implies
its finiteness for all other bases, or even for all finite sets X generating B qua left
G-set.
The biset B may be given by its structure map X ˆ G Ñ G ˆ X, written
px, gq ÞÑ pg@x, xg q; it is defined by the equality xg “ pg@xqxg in B. Then B is
contracting with nucleus N if for every g P G every sufficiently long iteration of the
maps g ÞÑ g@x eventually gives elements in N .
We may assume, without loss of generality, that XN Ď N X holds. For finitely
generated groups, there is perhaps more intuitive formulation of contraction: there
exists a proper metric on G and constants λ ă 1, C such that }g@x} ď λ}g} ` C
for all g P G, x P X.
Lemma 4.39. Let G BG be a contracting biset; choose a basis X of B, and let
N Ď G be the associated nucleus.
Then every symbolic finite orbit pbi qiPI is conjugate to one in which bi P N X for
all i P I.
Proof. Write every bi in the form gi xi with gi P G and xi P X. Conjugate the
symbolic orbit by pgi qiPI ; then each bi becomes gi´1 bi gf piq “ xi gf piq “ gi1 x1i P B for
some gi1 P G, x1i P X. Note that each gi1 is a state of some gi . Conjugate again by
pgi1 qiPE , etc.; after a finite number of steps, each bi will belong to N X.
Lemma 4.40. Let f : pS 2 , A, ordq ý be a geometric map. Then for every n P N
the number of n-periodic points of f is finite.
Proof. If f is doubly covered by a torus endomorphism z ÞÑ M z `q, then its periodn points are all solutions to z P p1 ´ M q´1 q ` p1 ´ M n q´1 Z2 ; the image of this set
under R2 Ñ R2 {xZ2 , ´zy is finite, of size at most detp1 ´ M n q.
If f is expanding, then consider its biset B, which is orbisphere contracting,
see [4, Theorem A]. By Lemma 4.39, there are finitely many conjugacy classes of
period-n finite symbolic orbits in B, and by Theorem 4.22, every symbolic orbit
shadows a unique finite poly-orbit.
r Ñ G be a forgetful morphism of groups and let
Theorem 4.41. Let G
r Ñ G BG and
r BG
r
G
r1
r BG
r
G
Ñ GB1G
r is geometbe two forgetful biset morphisms as in (8). Suppose furthermore that B
ric of degree ą 1. Denote by pGa , Ba qaPAr and pG1a , Ba1 qaPAr the portraits of bisets
r and B
r 1 in B and B 1 respectively.
induced by B
42
LAURENT BARTHOLDI AND DZMITRY DUDKO
r B
r 1 are conjugate under ModpGq
r if and only if there exists φ P ModpGq
Then B,
φ
1
φ
such that B – B and the portraits pGa , Baφ qaPAr and pG1a , Ba1 qaPA\E are conjugate.
Furthermore, the centralizer of the portrait pGa , Ba qaPAr is trivial, and the cenr of B
r is isomorphic, via the forgetful map ModpGq
r Ñ ModpGq, to
tralizer ZpBq
ˇ φ φ
(
φ P ZpBq ˇ pGa , Ba q r „ pGa , Ba q r
aPA
aPA
and is a finite-index subgroup of ZpBq.
r B
r 1 are conjugate, then certainly their images and subbisets are conProof. If B,
jugate. Conversely, let φ P ModpGq be such that B φ – B 1 and the portraits
r B
r1
pGφa , Baφ qaPAr and pG1a , Ba1 qaPA\E are conjugate, and let fr, fr1 be maps realizing B,
respectively, with f geometric. By Theorem 2.19, we have fr1 “ mfr for a knitting
mapping class m. Since m is liftable by Proposition 2.8, we have mfr “ f m1 for
some knitting mapping class m1 and therefore m1 fr1 pm1 q´1 “ m1 fr. Repeating with
m1 , we obtain mpkq ¨ ¨ ¨ m1 fr1 pmpkq ¨ ¨ ¨ m1 q´1 “ mpkq fr for all k P N, and mpkq “ 1
when k is large enough by Lemma 4.27. Therefore, fr, fr1 are conjugate, and so are
r B
r1.
B,
r follows. The centralizer of the portrait pGa , Ba q r
The description of ZpBq
aPA
r is geometric) all points shadowed by bisets Ba are
is trivial, because (since B
unmarked in G. The ZpBq-orbit of pGa , Ba qaPAr contains finitely many conjugacy
classes by Lemma 4.40, since all of them are induced from a geometric biset, and
therefore are encoded in periodic or pre-periodic points for a map realizing B.
5. Algorithmic questions
We give some algorithms that decide whether two portraits of bisets are conjugate, thereby reducing the conjugacy and centralizer problem from orbisphere maps
to their minimal orbisphere (with only the post-critical set marked).
All the algorithms we describe make use of the symbolic description of maps via
bisets, and are inherently quite fast and practical. Their precise performance, and
implementation details, be studied in the last paper [6] of this series.
There are two implementations of Algorithms 1.1 and 1.2, one for tGTor/2u
maps and one for tExpu maps. We describe them in separate subsections.
We already gave the following algorithms:
Algorithm 5.1 ([4, Algorithms 5.1 and 5.2]). Given an orbisphere biset G BG ,
Decide whether B is the biset of a map double covered by a torus endomorphism,
and if so Compute parameters M, q for a torus endomorphism z ÞÑ M z`q covering
B.
Algorithm 5.2 ([4, Algorithms 5.4 and 5.5]). Given an orbisphere biset G BG ,
Decide whether B is the biset of a tGTor/2u or an tExpu map, and in particular
whether B is geometric.
5.1. tGTor/2u maps. In the case of tGTor/2u maps, we can decide, without access
to an oracle, whether two such maps are conjugate, and we can compute their
centralizers, as follows. We shall need the following fact:
Theorem 5.3 (Corollary of [10]). There is an algorithm deciding whether two
matrices M, N P Mat`
2 pZq are conjugate by an element X P SL2 pZq, and produces
such an X if it exists.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
43
There is an algorithm computing, as a finitely generated subgroup of SL2 pZq, the
centralizer of M P Mat`
2 pZq.
Algorithm 5.4. Given G BG , G CG two minimal tGTor/2u bisets
Decide whether B and C are conjugate by an element of ModpGq, and if so construct a conjugator; and compute the centralizer ZpBq, as follows:
(1) If B˚ ‰ C˚ as maps on peripheral conjugacy classes, then return fail.
(2) Using Proposition 4.5(A), identify G with Z2 ¸ t˘1u, and with Algorithm 5.1 present B and C as BM v and BN w respectively, see (50).
(3) Using Theorem 5.3 check whether M and N are conjugate. If not, return
no; otherwise find a conjugator X and compute the centralizer ZpM q of M .
(4) Check whether there is a Y P ZpM q such that BpY Xq0 conjugates BM v to
BN w , as follows. The orbit of BM v under conjugation of
tBY 0 | Y P ZpM q, BpY Xq0 P ModpGqu
is finite and hence computable; so is its image under X 0 . Check whether
BN w belongs to it; if not, return no, and if yes, return yes and the conjugator BpY Xq0 .
(5) The centralizer of B is
tBY 0 | Y P ZpM q, BY 0 P ModpGq, and BY 0 centralizes BM v u,
naturally embeds as a subgroup of finite index in ZpM q.
Here is an algorithmic version of Proposition 4.18:
Algorithm 5.5. Given a minimal tGTor/2u orbisphere biset
xZ2 , ´z ` ry, a base point b0 P B specifying the map
(55)
Φ:
G BG
G BG
with G “
b xZ2 ,´z`ry R2 Ñ xZ2 ,´z`ry R2 : pb0 , zq ÞÑ M ´1 z
rÑA
r of the dynamics of B on its pe(see Proposition 4.18), an extension f˚ : A
ripheral classes, and a portrait of bisets pGa , Ba qaPAr ,
Compute pra qaPAr shadowed by pGa , Ba qaPAr : ra “ ΦpBa b rf˚ paq q, the local groups
Gr̄a where r̄a is the image of ra in R2 {G (see (54)), and the relative centralizer
ZD ppGa , Ba qaPAr q, which is a finite abelian group, as follows:
(1) Choose ba P Ba . Proceed through all periodic cycles E of I. Choose e P E
and solve the linear equation
´
¯
re “ Φ|E| be b bf˚ peq b ¨ ¨ ¨ b bf |E|´1 peq b re ;
˚
n
(2)
(3)
(4)
(5)
the equation takes form p1 ` θM qri “ t with θ P t˘1u (see (53)) and has
a unique solution by Lemma 4.19.
r
Inductively compute ra “ Φpba b rf˚ paq q for all a P A.
For a P A we have Gr̄a “ Ga .
r check whether ra ´ z P G. If ra ´ z R G, then Gr̄a is the trivial
For a P AzA
subgroup; otherwise Gr̄a “ xra ´ zy.
By a finite check compute ZD ppGa , Ba qaPAr q: by Lemma 4.24 it is the set
of self-conjugacies of the corresponding
homotopy pseudo-orbit, and by
ś
Lemma 4.20 it is a subgroup of aPAr Gr̄a , and is therefore an easily computable finite group.
44
LAURENT BARTHOLDI AND DZMITRY DUDKO
Algorithm 5.6. Given a minimal tGTor/2u orbisphere biset G BG , an extension
rÑ A
r of the dynamics of B on its peripheral classes, and two portraits of
f˚ : A
bisets pGa , Ba qaPAr and pG1a , Ba1 qaPAr ,
Decide whether pGa , Ba qaPAr and pG1a , Ba1 qaPAr are conjugate as follows:
(1) Normalize the portraits in such a manner that Ga “ G1a and Ba “ Ba1 for
all a P A; by Lemma 2.17 this follows from the conjugacy for subgroups:
find pℓa qaPA P GA with Gℓaa “ G1a and conjugate pGa , Ba qaPAr by pℓa qaPAr
with ℓa “ 1 if a R A.
(2) Identify G with xZ2 , ´z ` ry and choose b0 P B characterizing the map Φ,
see (55).
(3) Using Algorithm 5.5 compute the points ra and ra1 shadowed by pGa , Ba qaPAr
and pG1a , Ba1 qaPAr respectively.
r then return no.
(4) Check if r̄a “ r̄a1 in R2 {G. If r̄a ‰ r̄a1 for some a P A,
1
Otherwise find ℓa P G with ℓa ra “ ra with ℓa “ 1 for a P A and conjugate
pG1a , Ba1 q by pra qaPAr . This reduces to original problem to the case ra1 “ ra .
(5) Using Algorithm 5.5 compute the (finite
ś abelian) local groups Gr̄a and by
a finite check decide if an element in aPAr Gr̄a conjugates pGa , Ba qaPAr to
pG1a , Ba1 qaPAr .
r
and
If pr̄a qaPAr is an actual orbit, then Gr̄a are trivial groups for all a R AzA
Step (5) can be omitted. This is the case when the algorithm is called from Algorithm 5.12.
Algorithm 5.7. Given a minimal tGTor/2u orbisphere biset G BG and an extenrÑA
r of the dynamics of B on its peripheral classes,
sion f˚ : A
Produce a list of all conjugacy classes of portraits of bisets pGa , Ba qaPAr in B
with dynamics f˚ as follows:
(1) Write G “ xZ2 , ´z ` ry and B “ BM v , using Algorithm 5.1. Choose b0 P B
characterizing the map Φ, see (55).
(2) Using Algorithm 5.5 compute the orbit pr̄a qaPA of M : R2 {G ý. Produce a
list of all possible poly-orbits pr̄a qaPAr extending pr̄a qaPA .
(3) For every poly-orbit pr̄a qaPAr find a portrait pGa , Ba qaPAr that shadow pr̄a qaPAr .
(4) Using Algorithm 5.5 compute the (finite) local groups Gr̄a . By Lemma 4.20
the finite set
tpGa , Ba ha qaPAr | ha P Gr̄a , ha “ 1 if a P Au
contains a representative of every conjugacy class of portraits that shadows
pr̄a qaPAr . Using Algorithm 5.5 produce a list of all conjugacy classes of
portraits of bisets that shadow pr̄a qaPA .
5.2. tExpu maps. We now turn to expanding maps, and start by a short example
showing how the conjugacy problem may be solved algorithmically. We note, by
following the proof of Lemma 4.39, that every portrait of bisets can be algorithmically conjugated to one in which the terms belong to N X, for X a basis and N the
associated nucleus.
Example 5.8. Suppose E “ teu with f˚ peq “ e, and let pGa , Ba qaPA\E and
pGa , Ca qaPA\E be two portraits of bisets. Then Ge “ 1 and Be “ tbu and Ce “ tcu.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
45
The portraits pGa , Ba qaPA\E and pGa , Ca qaPA\E are conjugate if and only if
there exists ℓ P G such that ℓ´1 bℓ “ c.
Write Be “ tgxu in a basis X of B, with associated nucleus N ; recall that N is
symmetric, contains 1 and generates G. Then gx is conjugate to xg “ pg@xqxg “
g 1 x1 . After iterating finitely many times the process gx „ xg “ g 1 x1 , we can assume
g P N , as in Lemma 4.39. Similarly, we may replace Ce by a conjugate biset thyu
with h P N .
Find, by direct search, a t P N with N t´3 X Ě XN t ; such a t ě 4 exists because
B is contracting. Then Be and Ce are conjugate if and only if there exists ℓ P G
with bℓ “ ℓc, namely gxℓ “ ℓhy. Let u P N be such that ℓ P N u zN u´1 ; if u ě t
then gxℓ P N N u´3 X while ℓhy R N u´1 N X, a contradiction. Therefore the search
for a conjugator ℓ is constrained to ℓ P N t´1 .
As in Example 5.8 we have
Lemma 5.9. Let G BG be a contracting biset; choose a basis X of B, and let N Ď G
be the associated nucleus. Suppose that t P N is such that N t´3 X Ě XN t .
If two symbolic orbits pbi qiPI and pci qiPI with bi , ci P N X are conjugate by gi ,
then gi P N t´1 .
Proof. Suppose gi P N u zN u´1 for u ě t and some i. We can also assume that
gj P N u for all j P I. Then on one hand, gi bi P N u N X; but on the other hand
gi bi “ ci gf˚ piq P N u´3 N X. This is a contradiction.
Algorithm 5.10. Given a minimal tExpu orbisphere biset G BG , an extension
rÑ A
r of the dynamics of B on its peripheral classes, and two portraits of
f˚ : A
bisets pGa , Ba qaPAr and pG1a , Ba1 qaPAr ,
Decide whether pGa , Ba qaPAr and pG1a , Ba1 qaPAr are conjugate, and Compute the
centralizer of pGa , Ba qaPAr , which is a finite group, as follows:
r “ A\J \I with f n pJq Ď A and f˚ pIq Ď I. Normalize pGa , Ba q r
(1) Write A
˚
aPA
and pG1a , Ba1 qaPAr such that Ga “ G1a and Ba “ Ba1 ; by Lemma 2.17 this
follows from the conjugacy for subgroups: find pℓa qaPA P GA with Gℓaa “ G1a
and conjugate pGa , Ba qaPAr by pℓa qaPAr with ℓa “ 1 if a R A.
(2) Check whether there is pℓa qaPA\J with ℓa “ 1 for a P A conjugating
pGa , Ba qaPA\J and pG1a , Ba1 qaPA\J . If not, return no. This reduces the
conjugacy problem of portraits to the conjugacy problem of symbolic orbits: writing Bi “ tbi u and Bi1 “ tb1i u for i P I solve a conjugacy problem
of pbi qiPI and pb1i qiPI .
(3) Write the biset B in the form GX for a basis X, and let N be its nucleus.
Find, by direct search, a t P N with N t´3 X Ě XN t ; such a t ě 4 exists
because B is contracting.
(4) Write bi “ gi xi and replace bi with xi gi “ gi1 x1i . After iterating finitely
many times this process, we obtain bi P N X by Lemma 4.39. By a similar
iteration, we can assume b1i P N X.
(5) Answer whether pbi qiPI and pb1i qiPI are conjugate by elements in N t´1 . This
is correct by Lemma 5.9.
(6) By a direct search compute the centralizer of pGa , Ba qaPAr : the centralizer of
pGa , Ba qaPA\J is trivial, while elements centralizing pGa , Ba qaPI are within
N t´1 .
46
LAURENT BARTHOLDI AND DZMITRY DUDKO
Algorithm 5.11. Given a minimal tExpu orbisphere biset G BG and an extension
rÑA
r of the dynamics of B on its peripheral classes,
f˚ : A
Produce a list of all conjugacy classes of portraits of bisets pGa , Ba qaPAr in B
with dynamics f˚ as follows:
r “ A \ J \ I with f n pJq Ď A and f˚ pIq Ď I. Produce a list LA\J
(1) Write A
˚
of all conjugacy classes of portraits of bisets pGa , Ba qaPA\J .
(2) Write the biset B in the form GX for a basis X, and let N be its nucleus. Produce a list of all symbolic orbits pbi qiPI with bi P N X. Using
Algorithm 5.10 produce a list LI of all conjugacy classes of symbolic orbits
pbi qiPI with bi P N X.
(3) Combine LA\J and LI to produce a list LA\J ˆ LI of all conjugacy classes
of portraits of bisets pGa , Ba qaPAr by setting Bi – tbi u and Gi – t1u for
i P I.
5.3. Decidability of conjugacy and centralizer problems. The statements of
Theorem E can be turned into an algorithm:
r r, rC
r r two geometric bisets,
Algorithm 5.12. Given Gr B
G G G
And given an oracle that decides whether two minimal tExpuztGTor/2u bisets are
conjugate, and computes their centralizers,
r and C
r are conjugate by an element of ModpGq,
r and if so conDecide whether B
r
struct a conjugator; and compute the centralizer ZpBq, as follows:
r˚ ‰ C
r˚ as maps on peripheral conjugacy classes, then return no.
(1) If B
r
r r Ñ G1 CG1 be the maximal forgetful morphisms.
(2) Let Gr BGr Ñ G BG and Gr C
G
1
If G ‰ G , then return no.
(3) From now on assume G “ G1 . Using Algorithm 5.4 or the oracle, decide
whether B, C are conjugate; if not, return no. Using Algorithm 5.4 or the
oracle compute the centralizer ZpBq.
(4) From now on, assume B φ “ C for some φ P ModpGq. Compute the action
of ZpBq on the list of all conjugacy classes of portraits of bisets (this list
is finite by Corollary 4.21; Algorithms 5.7 and 5.11 compute the list) and
´1
´1
check whether pGa , Ba qaPAr and pGφa , Caφ qaPAr belong to the same orbit.
Return no if the answer is negative. Otherwise replace φ by an element in
ZpBqφ such that pGφa , Baφ qaPAr and pGa , Ca qaPAr are conjugate portraits.
r of φ, and compute k P ModpGq
r
(5) Choose an arbitrary lift φr P ModpGq
r
φ
r – C.
r It follows that k is a knitting element. Compute
such that k B
r 1 k2 ¨¨¨kpn´1q
pnq
r φk
r proceed until k pnq is identity
inductively k
with k pnq B
– C;
r 1 k 2 ¨ ¨ ¨ k pn´1q .
(this is guaranteed by Lemma 4.27). Return φk
r
(6) To compute the centralizer of B, consider first ZpBq. Compute the action
of ZpBq on the list of all conjugacy classes of portraits of bisets (this list
is finite by Corollary 4.21; Algorithms 5.7 and 5.11 compute the list) and
compute the stabilizer Z0 pBq of the ZpBq-action. Note that Z0 pBq is a
finite-index subgroup of ZpBq.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
47
(7) List a generating set S of Z0 pBq. For every φ P S compute its lift φp P
r with B
r ψp – B:
r compute first an arbitrary lift φr of φ, then inducModpGq
r 1 k2 ¨¨¨kpn´1q
pnq
r φk
r until k pnq is identity, and set
tively define k
with k pnq B
–B
1
2
pn´1q
r k ¨¨¨k
φp :“ φk
. Return tφp | φ P Su.
Proof of Corollary F. We claim that Algorithms 5.5, 5.6, 5.7, 5.10, 5.11 are efficient.
Indeed, the efficiency of Step 4 of Algorithm 5.10 follows from expansion. The
efficiency of Steps 5 and 7 of Algorithm 5.12 follows from Lemma 4.27. All the
remaining steps are obviously efficient.
We note that, in the case of p2, 2, 2, 2q bisets, the oracle itself is efficient. Indeed the oracle reduces to solving conjugacy and centralizer problems in the group
SL2 pZq, and Theorem 5.3 is efficient, because SL2 pZq has a free subgroup of index
12.
Corollary 5.13. There is an algorithm that, given two geometric bisets G BG and
decides whether B and C are conjugate, and computes the centralizer of B.
H CH ,
Proof. We briefly sketch an algorithm that justifies an existence of an oracle: given
two minimal tExpu orbisphere bisets, whether they are conjugate and computes
their centralizers; details will appear in [6].
Let B, C be two minimal tExpu orbisphere G-G-bisets. They admit a decomposition into rational maps along the canonical obstruction, which is computable
by [22]. The graph of bisets along this decomposition is computable, and rational maps may be computed for each of the small bisets in the decomposition, e.g.
by giving their coëfficients as algebraic numbers with floating-point enclosures to
distinguish them from their Galois conjugates. The bisets B, C are conjugate precisely when their respective rational maps are conjugate and the twists along the
canonical obstruction agree; the first condition amounts to finite calculations with
algebraic numbers, while the second is the topic of [5, Theorem A].
The centralizer of a rational map is trivial, and [5, Theorem A] shows that the
centralizer of B is computable.
Proof of Corollary G. If the rational map is p2, 2, 2, 2q, then the oracle is efficient,
as we noted above, so Corollary F applies. In the other case, the rational map is
hyperbolic, so it has trivial centralizer and no oracle is needed in the application of
Algorithm 5.12.
6. Examples
Finally, in this brief section, we consider some examples of portraits of bisets.
The first ones come from marking points on the Basilica map f pzq “ z 2 ´ 1, and
more generally maps with three post-critical points. The second ones from cyclic
bisets (which are particularly simple, but to which our main results apply with
restrictions because these bisets have automorphisms).
6.1. Twisted marked Basilica. Consider the Basilica polynomial
p t0, ´1, 8uq ý .
f pzq :“ z 2 ´ 1 : pC,
It has two fixed points α and β, with α P p´1, 0q and β ą 1. Let us take α to be
the basepoint and let G BG be the biset of f . Denote respectively by γ´1 and γ0
48
LAURENT BARTHOLDI AND DZMITRY DUDKO
γ´1 ´1 ‚
α
‚
γ0
´α
‚
‚ 0
‚
β
x2
γβ
t
Figure 1. The dynamical plane of z 2 ´ 1. Loops γ´1 , γ0 , γβ circle around 0, ´1, β respectively. The curve x2 connects α to its
preimage ´α. The simple closed curve t surrounds t´1, βu.
the loops circling around ´1 and 0 as in Figure 1, and let γ8 “ pγ´1 γ0 q´1 be the
loop around infinity. Then
G “ xγ´1 , γ0 , γ8 | γ8 γ´1 γ0 y.
The basepoint α has two preimages α and ´α. Let x1 be the constant path at
α and let x2 be a path slightly below 0 connecting α to ´α. The presentation of
B in the basis S :“ tx1 , x2 u is
γ´1 “!1, γ0 "p1, 2q,
γ0 “!γ´1 , 1",
´1 ´1
γ8 “!γ´1
γ0 , 1"p1, 2q.
We shall consider first the effect of marking a fixed point, and then of marking
some preimages of the post-critical set.
p t8, 0, ´1, βuq ý.
6.1.1. Marking β. Consider a marked Basilica frpzq “ z 2 ´ 1 : pC,
Let γβ be the loop around β depicted in Figure 1. The fundamental group of
p t8, 0, ´1, βuq, based at α, is
pC,
r “ xγ´1 , γ0 , γβ , γ8 | γ8 γ´1 γβ γ0 y
G
r ։ G sends γβ to 1. The presentation of the biset B
r “ Bpfrq
and the forgetful map G
in the basis S is
γ´1 “!1, γ0 "p1, 2q,
γβ “!1, γβ ",
γ0 “!γ´1 , 1",
´1 ´1
γ8 “!γ´1
γ0 , γβ´1 "p1, 2q.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
49
r “ t8, ´1, 0, βu; then B
r has a minimal portrait of
Write A “ t8, ´1, 0u and A
bisets
r ´1 “ xγ´1 y,
G
r´1 “ G
r´1 x1 ,
B
rβ “ xγβ y,
G
rβ “ G
r´1 x2 ,
B
r ´1 ,
r0 “ G
r0 x1 \ G
r0 x2 “ x1 G
B
r 0 “ xγ0 y,
G
r8 “ xγ8 y,
G
r8 “ G
r8 pγ´1 x1 q \ G
r8 pγ´1 x2 γβ q.
B
r ։ B the portrait pG
ra , Ba q r projects to
Under the forgetful intertwiner B
aPA
G´1 “ xγ´1 y,
B´1 “ G´1 x1 ,
Gβ “ 1,
Bβ “ tx2 u,
G0 “ xγ0 y,
B0 “ G0 x1 \ G0 x2 “ x1 G´1 ,
G8 “ xγ8 y,
B8 “ G8 pγ´1 x1 q \ G8 pγ´1 x2 q.
Note that Bβ “ tx2 u encodes β in the sense that x2 shadows β. This can be seen
directly as follows: set x02 “ x2 and let xi`1
be the f -lift of xi2 that starts at xi2 p1q.
2
i
Then the sequence of endpoints x2 p1q converges to β, and the infinite concatenation
x02 #x12 # ¨ ¨ ¨ is a path from α to β.
Let now T be the clockwise Dehn twist around the simple closed curve t surr is given by
rounding an interval 0 and β as in Figure 1. The action of T on G
pγ
T˚ : γ´1 Ñ γ´1´1
γβ q´1
pγ´1 γβ q´1
, γβ Ñ γβ
, γ0 Ñ γ0 , γ8 Ñ γ8 .
r of g :“ T ˝ f is B
r b BT˚ . Let us show that g is conjugate to
The biset C
2
p
z ´ 1 : pC, t8, ´1, 0, αuq ý. By Theorem 2.19,
ra , G
r a q r “ pB
ra , G
ra q r EpT q.
pC
aPA
aPA
Note that we have Ca “ Ba for all a P A, so it remains to compute Cβ .
Let t̄ P π1 pCzt0, ´1u, βq be the simple loop below 0 circling around ´1; then
T “ pushpt̄q. Let ℓβ : r0, 1s Ñ Czt´1, 0u (using notations of Lemma 2.16) be an arc
from α to β slightly below 0 so that γβ may be homotoped to a small neighborhood
´1
´1
of ℓβ . We have Cβ “ Bβ γ´1
; the claim then follows from x2 γ´1
“ x1 and the fact
that x1 shadows α.
More generally, suppose that gr “ pushps̄qfrpushpt̄q for some motions s̄, t̄ P
π1 pCzt0, ´1u, βq of β; then
´1
Cβ “ pℓβ #s̄#ℓ´1
β qBβ pℓβ #t̄#ℓβ q.
The process
´1
pℓβ #s̄#ℓ´1
β qx2 pℓβ #t̄#ℓβ q “ g1 xip1q „ xip1q g1 “ g2 xip2q „ xip2q g2 “ . . .
eventually terminates in either x1 or x2 . In the former case, gr is conjugate to z 2 ´
p t8, 0, ´1, αuq ý and in the latter case r
p t8, 0, ´1, βuq ý.
1 : pC,
g is conjugate to z 2 ´1 : pC,
6.1.2. Marking 1 and
?
2. Consider now
?
p t8, ´1, 0, 1, 2uq ý
frpzq “ z 2 ´ 1 : pC,
50
LAURENT BARTHOLDI AND DZMITRY DUDKO
´1
‚
ℓ´1
α
‚
ℓ0
0
‚
´1 1
´α ℓ0 Òf
ℓ1
1
‚?
2
ℓ´1
1 Òf
?
2
‚
ℓ?2
Figure 2. Curves ℓ´1 , ℓ0 , ℓ1 , ℓ?2 and their lifts.
?
?
r “ t1, 2u and A “ t8, ´1, 0u. The dynamics are 2 ÞÑ 1 ÞÑ 0 Ø ´1.
with AzA
r “ Bpfrq in the basis S:
As in §6.1.1 we readily compute a presentation of B
γ´1 “!1, γ0 "p1, 2q,
γ?2 “!1, 1",
γ1 “!1, γ?2 ",
γ0 “!γ´1 , γ1 ",
´1 ´1
´1
"p1, 2q,
γ8 “!γ´1
γ0 , γ1´1 γ?
2
r “ xγ8 , γ´1 , γ? , γ1 , γ0 | γ8 γ´1 γ? γ1 γ0 y.
with G
2
2
As in Lemma 2.16, let ℓ´1 and ℓ0 Ă R respectively be simple arcs from α to ´1
and to 0 so that γ´1 and γ0 may be homotoped to small neighborhoods
of γ´1 and
?
γ0 respectively. Let ℓ1 and ℓ?2 be simple arcs from α to 1 and 2 slightly below 0
as in Figure 2.
r “ Bpfrq. Then
Denote by pGa , Ba qaPAr the portrait of bisets in B induced by B
Ga and Ba are the same as in §6.1.1 for all a P A, while (by Lemma 2.18)
?
1
2
´1
?
?
B1 “ tℓ1 #ℓ´1
0 Òf u “ tx2 u and B 2 “ tℓ 2 #ℓ1 Òf u “ tx2 u.
We consider again some twists of fr. ?Consider gr “ m1 frm2 with m1 , m2 trivial
rel A. Suppose that m1 moves 1 and 2 along s1 P π1 pCzt´1, 0u, 1q and s?2 P
?
?
while m2 moves 1 and 2 along t1 P π1 pCzt´1, 0u, 1q
π1 pCzt´1, 0u, 2q respectively
?
and t?2 P π1 pCzt´1, 0u, 2q respectively. If pGa , Ca qaPAr is the portrait of bisets in
B induced by Bpr
g q, then
(
C1 “ pℓ1 #s1 #ℓ´1
1 qx2 ,
)
!
?
2
´1
?
?
C?2 “ pℓ?2 #s?2 #ℓ´1
1 qx2 pℓ 2 #t1 Òf #ℓ 2 q .
Write C1 “ th1 x2 u and C?2 “ th2 xj u. We may conjugate pC1 , C2 q to ptx2 u, txj h1 uq,
?
p t8, ´1, 0, 1, 2uq ý;
and write xj h1 “ h2 xk . If k “ 1, then r
g is conjugate to z 2 ´ 1 : pC,
?
p t8, ´1, 0, 1, ´ 2uq ý.
otherwise gr is conjugate to z 2 ´ 1 : pC,
6.1.3. Mapping class bisets. We continue the discussion from §6.1.2, and consider
the related mapping class bisets. The biset M pBq is of course reduced to tBu, with
ModpGq “ 1.
r is a left-free ModpGq-biset
r
On the other hand, the biset M pBq
of degree 2.
?
This can be seen in various ways: analytically, the point 2 may be moved to
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
51
?
?
2
r
1, 2uq, pz 2?´
´ 2, and a basis
? of M pBq may be chosen as tpz ´ 1, t8, ´1, 0, ?
1, t8, ´1, 0, 1, ´ 2uqu. More symbolically, the action of exchanging 2 with ´ 2
amounts to changing, in the wreath recursion of fr, the entry ‘γ1 “ !1, γ?2 "’ into
‘γ1 “ !γ?2 , 1"’.
r is nevertheless connected, and M pBq
r “ M ˚ pBq:
r inNote that the biset M pBq
deed the right action by the mapping class that pushes 1 once along the circle
t|z| “ 1u has the effect of exchanging the two left orbits.
?
Example 6.1. Let us now consider the sets A “ t8, ´1, 0, 1u and D “ t 2u, still
p Aq ý and fr “ f : pC,
p A \ Dq ý. Then the fibres
with the map f pzq “ z 2 ´ 1 : pC,
˚ r
of the map FD,D : M pf q “ M pfrq Ñ M pf q are not connected (by the second claim
of Proposition 2.21).
p
Indeed the left action of ModpCzA,
Dq has two orbits, while the right action
does not identify these orbits since 1 is not allowed to move around a critical value.
6.2. Belyi maps. The Basilica map f pzq “ z 2 ´ 1 is an example of dynamical
Belyi map, namely a map whose post-critical set consists of 3 points. All such
maps are realizable as holomorphic maps, and the three points may be normalized
as t0, 1, 8u.
In this subsection, we briefly state how the main results of this article simplify considerably. We concentrate on the dynamical situation, namely a map
f : pS 2 , Aq ý covered by fr: pS 2 , A \ Dq ý, with #A “ 3.
In that case, ModpS 2 , Aq “ 1, and we have a short exact sequence
1 Ñ knBraidpS 2 , A \ Dq Ñ ModpS 2 , A \ Dq Ñ π1 pS 2 zA, Dq Ñ 1.
Assume first that D consists only of fr-fixed points. Then the extension of bisets
decomposition of M pfrq from Theorem B reduces to the statement that M pfrq is
an inert extension of Bpf qD . The case of D consisting of a single fixed point was
considered in [5, §8.2].
More concretely: if we are given a wreath recursion G Ñ G ≀ S Ó. for Bpf q, with
G “ π1 pS 2 zA, ˚q, then a generating set for ModpS 2 , A\Dq can be chosen to consist
of #D copies of a generating set of G, corresponding to point pushes of D in S 2 zA,
together with some additional knitting elements. A wreath recursion for M pfrq will
then be of the form ModpS 2 , A \ Dq Ñ ModpS 2 , A \ Dq ≀ pS D qÓ., consisting of D
parallel copies of the wreath recursion of Bpf q. The wreath recursion associated
with the knitting elements is trivial.
In case D consists of periodic points of period ą 1, then the actions of G on
the left and the right should be appropriately permuted. If D contains ni cycles
of period i, then abstractly M pfrq will consist in the direct product of ni copies of
Bpf i q “ Bpf qbi .
Example 6.2. Consider the map frpzq “ f pzq “ z 3 with A “ t0, 1, 8u and D “
tω “ expp2πi{3qu. The biset M pf q is of course reduced to tf u, while M pfrq “
p
p
ModpCzA,
Dq – π1 pCzAq
with trivial right action. Indeed D is not in the image
of f , so pushing ω has no effect.
p
However, M ˚ pfrq has two left ModpCzA,
Dq-orbits by Lemma 2.20. Representa3
tives may be chosen as tpz , A\ tωuq, pz 3 , A\ tω 2 uqu, or equivalently (if the marked
p Aq ý
set is to remain A \ D) as the maps tz 3 , z 3 ˝ mu for a homeomorphism m : pC,
2
that pushes ω to ω .
52
LAURENT BARTHOLDI AND DZMITRY DUDKO
6.3. Cyclic bisets. We finally consider the easy case of a cyclic biset, namely the
p t0, 8uq ý for some d P Zzt´1, 0, 1u.
biset G BG of a monomial map f pzq “ z d : pC,
Then G – Z and B is a left-free right-principal biset. Choosing b0 P B we can
identify B “ b0 G with Z with the actions are given by
m ¨ b ¨ n “ dm ` b ` n.
Observe first that b Ñ b ` k is an automorphism of B for all k P Z. Therefore,
contrarily Proposition A we have AutpBq – Z.
Suppose that pGa , Ba qaPAr is a portrait of bisets in B. Then AutpBq acts on
portraits by
pGa , Ba qaPAr ` k :“ pGa , Ba ` kqaPAr .
We denote by pGa , Ba qaPAr { AutpBq the orbit of this action. We can now adjust
Theorem C for cyclic bisets:
r ։ G be a forgetful morphism of groups with G – Z, and let
Lemma 6.3. Let G
p t0, 8uq ý.
B
be
the
biset
of
z d : pC,
G G
There is then a bijection between, on the one hand, conjugacy classes of portraits
of bisets pBa qaPAr in B considered up to the action of AutpBq and, on the other
r G-bisets
r
r ։ G considered up to composition with
hand, Gprojecting to B under G
the biset of a knitting element. This bijection maps every minimal portrait of bisets
r to pBa q r { AutpBq.
of B
aPA
p t0, 8uq so as to remove automorphisms,
Proof. Mark an extra fixed point in pC,
and apply Theorem C.
p t0, 8uYDq ý and let pGa , Ba q r
To illustrate Lemma 6.3, consider frpzq “ z d : pC,
aPA
p t0, 8uq. Let us also assume that d ą 1. For
be the induced portrait of bisets on pC,
a P D write Ba “ tba u with ba P B – Z.
Let us consider a twisted map gr “ mfrn. For a P D let ma and na be the number
of times m and n push a around 0. If pGa , Ca qaPAr is the induced portrait of bisets,
then Ca “ tdma ` bd ` nf˚ paq u “ tca u for a P A. For a P D set
xa – ca {d ` cf˚ paq {d2 ` cf˚2 paq {d3 ` . . .
mod Z P R{Z.
Then pCa qaPD shadows pxa qaPD . The map r
g is unobstructed if and only if all points
xa are pairwise different. If gr is unobstructed, then pxa qdPD considered up to the
action
pxa qaPD Ñ pxa ` kqaPD , k P Z{pd ´ 1qZ
is a complete conjugacy invariant of gr.
References
[1] Laurent Bartholdi and Volodymyr V. Nekrashevych, Thurston equivalence of topological polynomials, Acta Math. 197 (2006), no. 1, 1–51, DOI 10.1007/s11511-006-0007-3, available at
arXiv:math.DS/0510082. MR2285317 (2008c:37072)
[2] Laurent Bartholdi and Dzmitry Dudko, Algorithmic aspects of branched coverings, Ann. Fac.
Sci. Toulouse 5 (2017), 1219–1296, DOI 10.5802/afst.1566, available at arXiv:cs/1512.05948.
[3]
, Algorithmic aspects of branched coverings I/V. Van Kampen’s theorem for bisets,
Groups Geom. Dyn. (December 2015), to appear, available at arXiv:cs/1512.08539.
, Algorithmic aspects of branched coverings IV/V. Expanding maps, Trans. Amer.
[4]
Math. Soc., posted on October 2016, to appear, DOI 10.1090/tran/7199, available at
arXiv:math/1610.02434.
ERASING MAPS, ORBISPACES, AND THE BIRMAN EXACT SEQUENCE
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
53
, Algorithmic aspects of branched coverings II/V. Sphere bisets and their decompositions (2016), submitted, available at arXiv:math/1603.04059.
, Algorithmic aspects of branched coverings V/V. Symbolic and floating-point algorithms, 2017, in preparation.
Matthieu Calvez, Fast Nielsen-Thurston classification of braids, Algebr. Geom. Topol. 14
(2014), no. 3, 1745–1758, DOI 10.2140/agt.2014.14.1745. MR3212582
James W. Cannon, William J. Floyd, Walter R. Parry, and Kevin M. Pilgrim, Nearly Euclidean Thurston maps, Conform. Geom. Dyn. 16 (2012), 209–255, DOI 10.1090/S1088-41732012-00248-2. MR2958932
Benson Farb and Dan Margalit, A primer on mapping class groups, Princeton Mathematical
Series, vol. 49, Princeton University Press, Princeton, NJ, 2012. MR2850125
Fritz J. Grunewald, Solution of the conjugacy problem in certain arithmetic groups, Word
problems, II (Conf. on Decision Problems in Algebra, Oxford, 1976), Stud. Logic Foundations Math., vol. 95, North-Holland, Amsterdam-New York, 1980, pp. 101–139. MR579942
(81h:20054)
Geoffrey Hemion, On the classification of homeomorphisms of 2-manifolds and the classification of 3-manifolds, Acta Math. 142 (1979), no. 1-2, 123–155, DOI 10.1007/BF02395059.
MR512214 (80f:57003)
John H. Hubbard and Dierk Schleicher, The spider algorithm, Complex dynamical systems
(Cincinnati, OH, 1994), Proc. Sympos. Appl. Math., vol. 49, Amer. Math. Soc., Providence,
RI, 1994, pp. 155–180. MR1315537
Yutaka Ishii and John Smillie, Homotopy shadowing, Amer. J. Math. 132 (2010), no. 4,
987–1029, DOI 10.1353/ajm.0.0126. MR2663646 (2012d:37055)
Thomas Koberda and Johanna Mangahas, An effective algebraic detection of the NielsenThurston classification of mapping classes, J. Topol. Anal. 7 (2015), no. 1, 1–21, DOI
10.1142/S1793525315500016. MR3284387
Henry W. Levinson, Decomposable braids and linkages, Trans. Amer. Math. Soc. 178 (1973),
111–126, DOI 10.2307/1996692. MR0324684
Russell Lodge, Boundary values of the Thurston pullback map, Conform. Geom. Dyn. 17
(2013), 77–118, DOI 10.1090/S1088-4173-2013-00255-5. MR3063048
H. A. Masur and Y. N. Minsky, Geometry of the complex of curves. II. Hierarchical structure,
Geom. Funct. Anal. 10 (2000), no. 4, 902–974, DOI 10.1007/PL00001643. MR1791145
Volodymyr V. Nekrashevych, Self-similar groups, Mathematical Surveys and Monographs,
vol. 117, American Mathematical Society, Providence, RI, 2005. MR2162164 (2006e:20047)
Volodymyr Nekrashevych, Combinatorial models of expanding dynamical systems, Ergodic
Theory Dynam. Systems 34 (2014), no. 3, 938–985, DOI 10.1017/etds.2012.163. MR3199801
Kevin M. Pilgrim, Combinations of complex dynamical systems, Lecture Notes in Mathematics, vol. 1827, Springer-Verlag, Berlin, 2003. MR2020454 (2004m:37087)
Alfredo Poirier, Critical portraits for postcritically finite polynomials, Fund. Math. 203
(2009), no. 2, 107–163, DOI 10.4064/fm203-2-2. MR2496235 (2010c:37095)
Nikita Selinger, Topological characterization of canonical Thurston obstructions, J. Mod.
Dyn. 7 (2013), no. 1, 99–117, DOI 10.3934/jmd.2013.7.99. MR3071467
Nikita Selinger and Michael Yampolsky, Constructive geometrization of Thurston maps
and decidability of Thurston equivalence, Arnold Math. J. 1 (2015), no. 4, 361–402, DOI
10.1007/s40598-015-0024-4, available at arXiv:1310.1492. MR3434502
Jing Tao, Linearly bounded conjugator property for mapping class groups, Geom. Funct. Anal.
23 (2013), no. 1, 415–466, DOI 10.1007/s00039-012-0206-3. MR3037904
William P. Thurston, On the geometry and dynamics of diffeomorphisms of surfaces, Bull.
Amer. Math. Soc. (N.S.) 19 (1988), no. 2, 417–431, DOI 10.1090/S0273-0979-1988-15685-6.
MR956596 (89k:57023)
Heiner Zieschang, Elmar Vogt, and Hans-Dieter Coldewey, Surfaces and planar discontinuous
groups, Lecture Notes in Mathematics, vol. 835, Springer, Berlin, 1980. Translated from the
German by John Stillwell. MR606743 (82h:57002)
E-mail address: [email protected]
54
LAURENT BARTHOLDI AND DZMITRY DUDKO
École Normale Supérieure, Paris and Mathematisches Institut, Georg-August Universität zu Göttingen
E-mail address: [email protected]
Jacobs University, Bremen
| 4math.GR
|
DeepLesion: Automated Deep Mining, Categorization
and Detection of Significant Radiology Image Findings
using Large-Scale Clinical Lesion Annotations
arXiv:1710.01766v2 [cs.CV] 10 Oct 2017
Ke Yan? , Xiaosong Wang? , Le Lu, and Ronald M. Summers
Department of Radiology and Imaging Sciences,
National Institutes of Health Clinical Center, Bethesda, MD 20892
{ke.yan, xiaosong.wang, le.lu, rms}@nih.gov
Abstract. Extracting, harvesting and building large-scale annotated radiological
image datasets is a greatly important yet challenging problem. It is also the bottleneck to designing more effective data-hungry computing paradigms (e.g., deep
learning) for medical image analysis. Yet, vast amounts of clinical annotations
(usually associated with disease image findings and marked using arrows, lines,
lesion diameters, segmentation, etc.) have been collected over several decades and
stored in hospitals’ Picture Archiving and Communication Systems. In this paper,
we mine and harvest one major type of clinical annotation data – lesion diameters annotated on bookmarked images – to learn an effective multi-class lesion
detector via unsupervised and supervised deep Convolutional Neural Networks
(CNN). Our dataset is composed of 33,688 bookmarked radiology images from
10,825 studies of 4,477 unique patients. For every bookmarked image, a bounding box is created to cover the target lesion based on its measured diameters. We
categorize the collection of lesions using an unsupervised deep mining scheme to
generate clustered pseudo lesion labels. Next, we adopt a regional-CNN method
to detect lesions of multiple categories, regardless of missing annotations (normally only one lesion is annotated, despite the presence of multiple co-existing
findings). Our integrated mining, categorization and detection framework is validated with promising empirical results, as a scalable, universal or multi-purpose
CAD paradigm built upon abundant retrospective medical data. Furthermore, we
demonstrate that detection accuracy can be significantly improved by incorporating pseudo lesion labels (e.g., Liver lesion/tumor, Lung nodule/tumor, Abdomen
lesions, Chest lymph node and others). This dataset will be made publicly available (under the open science initiative).
1
Introduction
Computer-aided detection/diagnosis (CADe/CADx) has been a highly prosperous and
successful research field in medical image processing. Many commercial software packages have been developed for clinical usage and screening. Recent advances (e.g.,
automated classification of skin lesions [3], detection of liver lesion [1], pulmonary
embolism [10]) have attracted even more attention to the application of deep learning paradigms to CADe/CADx. Deep learning, namely Convolutional Neural Network
?
These two authors contributed equally.
2
(CNN) based algorithms, perform significantly better than conventional statistical learning approaches combined with hand-crafted image features. However, these performance gains are often achieved at the cost of requiring tremendous amounts of training
data accompanied with high quality labels. Unlike general computer vision tasks, medical image analysis currently lacks a substantial, large-scale annotated image dataset
(comparable to ImageNet [2] and MS COCO [6]),for two main reasons: 1) The conventional methods for collecting image labels via Google search + crowd-sourcing from
average users cannot be applied in the medical image domain, as medical image annotation reuqires extensive clinical expertise; 2) Significant inter and intra-observer variability (among even well-trained, experienced radiologists) frequently occurs, and thus
may compromise reliable annotation of a large amount of medical images, especially
considering the great diversity of radiology diagnosis tasks.
Current CADe/CADx methods generally target one particular type of diseases or lesions, such as lung nodules, colon polyps or lymph nodes [7]. Yet, this approach differs
from the methods radiologists routinely apply to read medical image studies and compile radiological reports. Multiple findings can be observed and are often correlated. For
instance, liver metastases can spread to regional lymph nodes or other body parts. By
obtaining and maintaining a holistic picture of relevant clinical findings, a radiologist
will be able to make a more accurate diagnosis. However, it remains greatly challenging
to develop a universal or multi-purpose CAD framework, capable of detecting multiple
disease types in a seamless fashion. Such a framework is crucial to building an automatic radiological diagnosis and reasoning system.
In this paper, we attempt to address these challenges by first introducing a new
large-scale dataset of bookmarked radiology images, which accommodate lesions from
multiple categories. Our dataset, named DeepLesion, is composed of 33,688 bookmarked images from 10,825 studies of 4,477 patients (see samples in Fig. 1). For each
bookmarked image, a bounding box is generated to indicate the location of the lesions.
Furthermore, we integrate an unsupervised deep mining method to compute pseudo
image labels for database self-annotating. Categories of Liver lesion/tumor, Lung nodule/tumor, Abdomen lesions, Chest lymph node and others are identified by our computerized algorithm instead of radiologists’ annotation, which may be infeasible. After
obtaining the dataset, we develop an automatic lesion detection approach to jointly localize and classify lesion candidates using discovered multiple categories. Last, how the
unsupervisedly-learned pseudo lesion labels affect the deep CNN training strategies and
the quantitative performance of our proposed multi-class lesion detector is investigated.
2
Methods
In this section, we first describe how our DeepLesion dataset is constructed. Next, we
propose an unsupervised deep learning method to mine the latent lesion categories in
each image. This method involves an iterative process of deep image feature extraction,
image clustering and CNN model retraining. Finally, we present a multi-class object
detection approach to detect lesions of multiple categories.
3
(a)
(b)
(c)
Fig. 1. Three sample bookmarked images illustrated with annotation lesion patches (i.e., yellow
dashed boxes). The outputs from our proposed multi-category lesion detection framework are
shown in colored boxes with LiVer lesion (LV) in Red, Lung Nodule (LN) in Orange, ABdomen
lesion (AB) in Green, Chest Lymph node (CL) in magenta and other MiXed lesions (MX) in blue.
(a) A spleen metastasis is correctly detected along with several liver and abdomen metastases; (b)
Two large lymph nodes in mediastinum are all correctly detected; (c) All three lung nodules are
detected despite two small ones not being annotated in this bookmarked image.
2.1
DeepLesion Dataset
Radiologists routinely annotate hundreds of clinically meaningful findings in medical
images, using arrows, lines, diameters or segmentations to highlight and measure different disease patterns to be reported. These images, called “bookmarked images”, have
been collected over close to two decades in our institute’s Picture Archiving and Communication Systems (PACS). Without loss of generality, in this work, we study one type
of bookmark in CT images: lesion diameters. Each pair of lesion diameters consists of
two lines, one measuring the longest diameter and the second measuring its longest
perpendicular diameter in the plane of measurement. We extract the lesion diameter coordinates from the PACS server and convert into corresponding positions in the image
plane coordinates, noted as {(x11 , y11 ), (x12 , y12 )}; {(x21 , y21 ), (x22 , y22 )}. A bounding box (lef tx , topy , width, height) is computed to cover a rectangular area enclosing
the lesion measurement with 20 pixel padding in each direction, i.e., (xmin −20, ymin −
20, xmax − xmin + 40, ymax − ymin + 40) where xmin = M in(x11 , x12 , x21 , x22 )
and xmax = M ax(x11 , x12 , x21 , x22 ), and similarly for ymin and ymax . The padding
range can capture the lesion’s full spatial extent with sufficient image context. We thus
generate 33,688 bookmarked radiology images from 10,825 studies of 4,477 unique patients, and each bookmarked image is associated with a bounding box annotation of the
enclosed lesion. Sample bookmarked images and bounding boxes are shown in Fig. 1.
2.2
Unsupervised Lesion Categorization
The images in our constructed lesion dataset contain several types of lesions commonly
observed by radiologists, such as lung nodule/lesion, lymph node, and liver/kidney
lesion. However, no detailed precise category labels for each measured lesion have
been provided. Obtaining such from radiologists would be highly tedious and timeconsuming, due to the vast size and comprehensiveness of DeepLesion. To address this
4
Fig. 2. Lesion categorization framework via unsupervised and iteratively-optimized deep CNNs.
problem, we propose a looped deep optimization procedure for automated category discovery, which generates visually coherent and clinically-semantic image clusters. Our
algorithm is conceptually simple: it is based on the hypothesis that the optimization procedure will “converge” to more accurate labels, which will lead to better trained CNN
models. Such models, in turn, will generate more representative deep image features,
which will allow for creating more meaningful lesion labels via clustering.
As a pre-processing step, we crop the lesion patches from the original DICOM
slides using the dilated bounding boxes (described in Sec. 2.1) and resize them, prior
to feeding them into the CNN model. As shown in Fig. 2, our iterative deep learning
process begins by extracting deep CNN features for each lesion patch using the ImageNet [2]) pre-trained VGG-16 [9] network. Next, it applies k-means clustering to the
deep feature encoded lesion patches after k is determined via model selection [5]. Next,
it fine-tunes the current VGG-16 using the new image labels obtained from k-means.
This yields an updated CNN model for the next iteration. The optimization cycle terminates once the convergence criteria have been satisfied.
Encoding Lesion Patches using Deep CNN Features: The VGG-16 [9] CNN architecture is adopted for patch encoding and CNN model fine-tuning to facilitate the
iterative procedure. The image features extracted from the last fully-connected layer
(e.g., FC6/FC7 of VGG-16) are used, as they are able to capture both the visual appearance and the spatial layout of any lesion, with its surrounding context.
Convergence in Patch Clustering and Categorization: We hypothesize that the
newly generated clusters will converge to the “oracle” label clusters, after undergoing
several staged of cluster optimization. Two convergence measurements are employed:
Purity [5] and Normalized Mutual Information (NMI). We assess both criteria by computing empirical similarity scores between clustering results from two adjacent iterations. If the similarity score exceeds a pre-defined threshold, the optimal clustering
driven categorization of lesion patches has been attained. For each iteration, we randomly shuffle the lesion patches and divide the data into three subsets: training (75%),
validation (10%) and testing (15%). Therefore the “improving-then-saturating” trajectory of the CNN classification accuracy on the testing set can also indicate the convergence of the clustering labels (i.e., optimal image labels have been obtained).
2.3
Multi-category Lesion Detection
Using the bounding boxes (Sec. 2.1) and their corresponding newly generated pseudocategory labels (Sec. 2.2), we develop a multi-class lesion detector adapted from the
5
Fig. 3. Flow chart of the lesion detection algorithm. Bookmarked clinical annotations provide the
ground-truth bounding boxes of lesions for detector training. In detection, the dashed and solid
boxes indicate the ground-truth annotation and its predicted lesion detection, respectively.
Faster RCNN method [8]. An input image is first processed by several convolutional
and max pooling layers to produce feature maps, as shown in Fig. 3. Next, a region
proposal network (RPN) parses the feature maps and proposes candidate lesion regions.
It estimates the probability of “target/non-target” on a fixed set of anchors (candidate
regions) on each position of the feature maps. Furthermore, the location and size of each
anchor are fine-tuned via bounding box regression. Afterwards, the region proposals
and the feature maps are sent to a Region of Interest (RoI) pooling layer, which resamples the feature maps inside each proposal to a fixed size (we use 7×7 here). These
feature maps are then fed into several fully-connected layers that predict the confidence
scores for each lesion class and run another bounding box regression for further finetuning. Non-maximum suppression (NMS) is then applied to the feature maps. Finally,
the system returns up to five detection proposals with the highest confidence scores
(> 0.5), as each image only has one bookmarked clinical annotation.
The ImageNet pretrained VGG-16 [9] model is adopted as the backbone of Faster
RCNN [8]. It is useful to remove the last pooling layer (pool4) in VGG-16 to enhance
the resolution of the feature map and to increase the sampling ratio of positive samples
(candidate regions that contain lesions). In our experiments, removing pool4 improves
the accuracy by ∼ 15%. It is critical to set the anchor sizes in RPN to fit the size of
ground-truth bounding boxes in DeepLesion dataset. Hence, we use anchors of three
scales (48, 72, 96) and aspect ratios (1:1, 1:2, 2:1) to cover most of the boxes.
For image preparation, we use the ground-truth lesion bounding boxes derived in
Sec. 2.1 incorporating enlarged spatial contexts. Each full-slice image in the detection
phase is resized, so that the longest dimension is of 512 pixels. We then train the network
as demonstrated in Fig. 3 in a multi-task fashion: two classification and two regression
losses are jointly optimized. This end-to-end training strategy is more efficient than
the four-step method in the original Faster RCNN implementation [8]. During training,
each mini-batch has 4 images, and the number of region proposals per image is 32. We
use the Caffe toolbox and the Stochastic Gradient Descent (SGD) optimizer. The base
6
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 4. Six sample detection results are illustrated with the annotation lesion patches as yellow
dashed boxes. The outputs from our proposed detection framework are shown in colored boxes
with LiVer lesion (LV) in Red, Lung Nodule (LN) in Orange, ABdomen lesion (AB) in Green,
Chest Lymph node (CL) in magenta and other MiXed lesions (MX) in blue. (a) Four lung lesions
are all correctly detected; (b) Two lymph nodes in mediastinum is presented; (c) A Ground Glass
Opacity (GGO) and a mass are detected in the lung; (d) An adrenal nodule; (e) Correct detections
on both the small abdomen lymph node nearly aorta but also other metastases in liver and spleen;
(f) Two liver metastasis are correctly detected. Three lung metastases are detected but erroneously
classified as liver lesions .
learning rate is set to 0.001, and is reduced by a factor of 10 every 20K iterations. The
network generally converges within 60K iterations.
3
Results and Discussion
Our lesion categorization method in Sec. 2.2 partitions all lesion patches into five
classes k = 5. After visual inspection supervised by a board-certificated radiologist,
four common lesion categories are found, namely lung nodule/lesion, liver lesion, chest
lymph nodes and abdominal lesions (mainly kidney lesions and lymph nodes), with
high purity scores in each category (0.980 for Lung Nodule, 0.955 for Chest Lymph
Node, 0.805 for Liver Lesion and 0.995 for Abdomen Lesion). The per-cluster purity
scores are estimated through a visual assessment by an experienced radiologist using a
set of 800 randomly selected lesion patches (200 images per-category). The remaining
bookmarked annotations are treated as a “noisy” mixed lesion class. Our optimization
framework converges after six iterations, with a high purity score of 92.3% returned
when assessing the statistical similarity or stability of two last iterations. Meanwhile,
the top-1 classification accuracy reaches 91.6%, and later fluctuates by ±2% .
7
1
0.8
0.7
0.5
Precision
Detection accuracy
0.8
0.6
0.4
0.3
Single-class lesion detection
Multi-class lesion detection
0.2
0.6
Single-class lesion detection
Liver lesion detection
Lung nodule detection
Abdomen lesions detection
Chest lymph node detection
Mixed lesions detection
0.4
0.2
0.1
0
0.1
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0
0.9
0.2
0.4
0.6
0.8
1
Recall
IoU
(a)
(b)
Fig. 5. (a): Detection accuracy curves with different intersection-over-union (IoU) thresholds. (b):
Precision-Recall curves of single-class and five category lesion detection when IoU=0.5.
Cluster
1
2
3
4
5
Cluster category
Liver
lesion
Lung
nodule
abdomen
lesions
Chest
lymph node
Mixed
lesions
Overall
Cluster size
774
837
1270
860
1292
5033
Averaged Accuracy: w/o
categorization labels (%)
56.04
73.30
48.79
70.77
54.85
59.45
Averaged accuracy: with
categorization labels (%)
60.59
76.34
56.22
76.28
58.67
64.30
Table 1. Test detection accuracies of five deep-learned pseudo lesion categories. Note that in
DeepLesion dataset, available clinical annotations are only partial (i.e., missing lesion labels).
Hence the actual detection accuracies in both configurations are significantly higher than the
above reported values since many “false positive” detections are later verified to be true lesions.
For lesion detection, all bookmarked images are divided into training (70%), validation (15%), and testing (15%) sets, by random splitting the dataset at the patient level.
Although different lesion types may be present in an image, only one clinical annotation
per image is available. We adopt a straightforward evaluation criterion: 1) we take the
top one detected lesion candidate box (with the highest detection confidence score) per
testing image as the detection result; 2) if the intersection-over-union (IoU) between
this predicted box and the ground-truth clinical annotation box is larger than 0.5 (as
suggested by the PASCAL criterion [4]), the detection is regarded as correct, and vice
versa. The lesion category is not considered in this criterion. We denote this evaluation
metric as detection accuracy. The proposed multi-class lesion detector merely requires
88 ms to process a 512×512 test bookmarked image on a Titan X Maxwell GPU.
Two lesion detection setups or configurations are examined: single-class (all annotation bounding boxes are considered as one abnormality class), and multi-class detection
(with pseudo-category labels). Some illustrative results from the multi-category lesion
detector on the testing set are shown in Fig. 4. It can be found that our developed detector is able to detect all five types of lesions and simultaneously provide the corresponding lesion category labels. Furthermore, some detection boxes currently considered as
8
false alarms actually belong to true lesions because the lesions bounding boxes are only
partially labeled by clinical bookmark annotations. Detailed statistics of the five deeply
discovered lesion clusters in the test set are provided in Table 1. This outlines the types
of lesions in the clusters that have been verified by radiologists. The averaged accuracy
of the single-class detection is 59.45% (testing) and this score becomes 64.3% for multiclass detection (testing). From Table 1, the multi-category detector also demonstrates
accuracy improvements of 3∼8% per lesion cluster or category compared against the
one-class abnormality detector. Single-class abnormality detection appears to be a more
challenging task since it tackles detecting various types of lesions at once. This validates
that better lesion detection models can be trained if we can perform unsupervised lesion
categorization from a large collection of retrospective clinical data.
The default IoU threshold is set as 0.5. Fig. 5 (a) illustrates the detection accuracy
curves of both detection models under different IoU thresholds. The multi-category lesion detection achieves the better overall accuracy while being able to assign the lesion
labels at the same time. Fig. 5 (b) shows the corresponding detection precision-recall
curves. The performances of lung lesion detection and chest lymph node detection significantly outperform the one-class abnormality detection.
4
Conclusion
In this paper, we mine, categorize and detect one type of clinical annotations stored
in the hospital PACS system as a rich retrospective data source, to build a large-scale
Radiology lesion image database. We demonstrate the strong feasibility of employing
a new multi-category lesion detection paradigm via unified deep categorization and detection. Highly promising lesion categorization and detection performances, based on
the proposed dataset, are reported. To the best of our knowledge, this work is the first
attempt of building a scalable, multi-purpose CAD system using abundant retrospective
medical data. This is done almost effortlessly since no new arduous image annotation
workload is necessary. Our future work include extending bookmarked images to incorporate their successive slices for the scalable and precise lesion volume measurement;
extracting and integrating the lesion diagnosis prior from radiology text reports, and
improved multi-category detection methods.
References
1. A. Ben-Cohen, I. Diamant, E. Klang, M. Amitai, and H. Greenspan. Fully convolutional
network for liver segmentation and lesions detection. In MICCAI LABELS-DLMIA, 2016.
2. J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE CVPR, pages 248–255, 2009.
3. A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun.
Dermatologist-level classification of skin cancer with deep neural networks. Nature,
542(7639):115–118, 2017.
4. M. Everingham, A. Eslami, L. Van Gool, C. Williams, J. Winn, and A. Zisserman. The pascal
visual object classes challenge: A retrospective. Int. J. Comp. Vis., 111(1):98–136, 2015.
5. R. Gomes, A. Krause, and P. Perona. Discriminative clustering by regularized information
maximization. In NIPS, pages 775–783, 2010.
9
6. T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L.
Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740–755, 2014.
7. J. Liu, D. Wang, Z. Wei, L. Lu, L. Kim, E. Turkbey, and R. M. Summers. Colitis detection
on computed tomography using regional convolutional neural networks. In IEEE ISBI, 2016.
8. S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with
region proposal networks. In NIPS, pages 91–99, 2015.
9. K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image
recognition. In ICLR. arxiv.org/abs/1409.1556, 2015.
10. N. Tajbakhsh, M. B. Gotway, and J. Liang. Computer-aided pulmonary embolism detection using a novel vessel-aligned multi-planar image representation and convolutional neural
networks. In MICCAI, pages 62–69. Springer, 2015.
| 1cs.CV
|
Runtime Distributions and Criteria for Restarts
Jan-Hendrik Lorenz
arXiv:1709.10405v1 [cs.DS] 29 Sep 2017
Institut für Theoretische Informatik, Universität Ulm, 89069 Ulm, Germany
[email protected]
Abstract. Randomized algorithms sometimes employ a restart strategy. After a
certain number of steps, the current computation is aborted and restarted with a
new, independent random seed. In some cases, this results in an improved overall expected runtime. This work introduces properties of the underlying runtime
distribution which determine whether restarts are advantageous. The most commonly used probability distributions admit the use of a scale and a location parameter. Location parameters shift the density function to the right, while scale
parameters affect the spread of the distribution. It is shown that for all distributions scale parameters do not influence the usefulness of restarts and that location
parameters only have a limited influence. This result simplifies the analysis of the
usefulness of restarts. The most important runtime probability distributions are
the log-normal, the Weibull, and the Pareto distribution. In this work, these distributions are analyzed for the usefulness of restarts. Secondly, a condition for the
optimal restart time (if it exists) is provided. The log-normal, the Weibull, and the
generalized Pareto distribution are analyzed in this respect. Moreover, it is shown
that the optimal restart time is also not influenced by scale parameters and that
the influence of location parameters is only linear.
1
Introduction
Restart mechanisms are commonly used in day-to-day life. For example, when waiting
for an email response, it is common to send the original email again after some time.
Therefore it is not surprising that restart strategies are used in subjects as diverse as
biology (e.g. [20]), physics (e.g. [7]) and computer science (e.g. [21], [9]). There are at
least two large fields in computer science which utilize restarts. On the one hand network protocols often have a retransmission timer (e.g. TCP, see [19]), after a timeout the
loss of the package is assumed and therefore the message is resent. On the other hand,
probabilistic algorithms often restart after a certain number of steps without finding a
solution, those algorithms are especially common for constraint satisfaction problems
(CSP) and the well-known satisfiability problem (SAT). Although in practical use, our
impression is that the power of this algorithm paradigm is still underestimated. Restarts
can be used to improve the performance of an algorithm in regards to various measures.
For example, restarting the algorithm can help to improve the completion probability
when a deadline is present. This model was studied by Wu [23] while Lorenz [12] examined completion probabilities for parallel algorithms using restarts. Another measure
which can benefit from restarts is the expected runtime.
An important class of distributions for which restarts often improve the expected
runtime is the so-called class of heavy-tailed distributions. Heavy-tailed distributions
2. PRELIMINARIES
have a tail which decays slower than an exponential. Crovella et al. [6] showed that
transmission times in the world wide web follow a power-law tail which is a subclass
of heavy-tailed distributions. Gomes et al. [9] observed that some instances in CSP also
show a power-law tail, while Caniou and Codognet [5] examined that the log-normal
distribution is a good fit for some problems in constraint satisfaction. The log-normal
distribution also belongs to the class of heavy-tailed distributions.
Luby et al. [14] introduced two of the most important restart strategies. The fixed
cut-off strategy always restarts after the same (fixed) number of steps. They showed
that this strategy is optimal for a certain (possibly infinite) number of steps. However,
in general finding the right number of steps before restarting requires extensive knowledge of the distribution. The second strategy which Luby et al. introduced is called
Luby’s (universal) strategy which slowly increases the restart times. This strategy does
not require any knowledge of the underlying distribution and compared to the optimal
strategy the expected runtime of a process utilizing Luby’s strategy is only higher by a
logarithmic factor.
Since Luby’s strategy does not require any a-priori knowledge of the distribution,
it is nowadays more commonly used than the fixed cut-off strategy. One could, however, argue that better strategies are possible since, in a few cases, some information
is available prior to the experiment. For instance, for many algorithms, the distribution
for a certain class of problems has been observed empirically. Arbelaez et al. [2] use
a machine learning approach to predict the runtime distributions of several randomized algorithms. This knowledge can be used to obtain a better restart strategy for this
class of problems. Then, a speedup of, at least, a logarithmic factor can be expected
(compared to Luby’s strategy). Such a factor cannot be ignored in practice.
Choosing the wrong restart strategy can result in expected runtimes which are much
worse than not restarting at all. On the other hand, a good choice regarding the restart
strategy can result in a super-exponential speedup. We believe that not enough attention
has been paid to this field of research.
Our contribution: A condition for the usefulness of restart (Theorem 1) and a condition for optimal restart times (Theorem 2) is obtained. It is shown that scale parameters neither influence the usefulness nor the optimal restart times (Theorem 3 and 4).
The log-normal and the generalized Pareto distribution are analyzed for the usefulness
of restarts and their optimal restart times. The Weibull distribution is also studied for its
optimal restart times. Parameter settings for all these distributions for which restarts are
useful are obtained. Finally, it is shown that the influence of location parameters on the
usefulness of restarts is limited. This is, for all distributions discussed here, restarts are
still useful when a location parameter is present. The influence of a location Parameter
on the optimal restart time is just linear.
2
Preliminaries
In this section, the notation used throughout this work is introduced. Let X be a realvalued random variable. Then FX (t) = Pr(X ≤ t) is the cumulative distribution funcd
F (t) is the density function of X. In many
tion (cdf) of X and its derivative fX (t) = dt
2
3. MAIN RESULTS
cases the quantile function of the random variable X is helpful, it is defined as follows:
Definition 1 ([18]). Let X be a real-valued random variable with cumulative distribution function FX : R → [0, 1]. Then the Quantile function QX : [0, 1] → R is given
by
QX (p) = inf{x | F (x) ≥ p}.
(1)
For continuous, strictly monotonically increasing cumulative distribution functions FX
the quantile function QX is the inverse function of FX .
Luby et al. defined the fixed cut-off strategy.
Definition 2 ([14]). Let X be a random variable describing the runtime of a probabilistic algorithm A on some specific input. Given any value t ∈ R+ a new algorithm
At is obtained by restarting A after time t has passed without finding a solution. Then
Xt is a random variable describing the runtime of At .
When talking about restarts, restarts using the fixed cut-off strategy are meant. The
notion of usefulness is often used in this work. Restarts are called useful if there is a
t > 0 with E[Xt ] < E[X]. Throughout this work, only real-valued random variables
such that FX , fX and QX exist are considered. If it is clear from the context, the
subscript X is omitted for the functions defined here. For the results presented here,
we assume that the number of restarts is not limited.
3
Main Results
Before employing a restart strategy, it should first be considered under which conditions
restarts are useful at all. Moorsel and Wolter [16] obtained a condition for the usefulness
of restarts: Let T be a random variable describing the runtime of the process if there is
a t > 0 with E[T ] < E[T − t | T > t], then restarts are useful. They showed that this
condition is sufficient and necessary if the mean E[T ] exists. This is a property which is
often shown by heavy-tailed distributions. However, there are heavy-tailed distributions
which do not fulfill their conditions, and there are also light-tailed distributions which
do fulfill this condition. In this section, another condition for the usefulness of restarts
is provided and it is applied to several distributions.
3.1
Effective Restarts
There are several ways to describe a dataset. Two of the most commonly used values
are the median and the mean, both of which describe a ’typical’ value for this dataset.
If the mean lies to the right of the median one can speak of (positively) skewed data.
While outliers contribute linearly to the mean, the median is very resistant to outliers.
Therefore a big difference between the median and the mean can be explained by either
many outliers or a few, but extreme outliers. In both of these cases, restarts can be
an efficient way to reduce outliers. Thus comparing the mean to the median yields a
simple condition for the usefulness of restarts: If Q(0.5)/E[X] < 0.5, then restarts are
beneficial. This holds because the expected runtime E[XQ(p) ] is bounded from above
3
3. MAIN RESULTS
by Q(p)
p . This idea can be easily generalized. If the mean is large because of a small
number of disproportionately long runs, but there are also many short runs, then restarts
are useful. Measuring this inequality is a well-known field in economics which is known
as income inequality metrics. One of those metrics, the Lorenz curve, turns out to be
helpful in the following.
Definition 3 ([13]). Let X be a real-valued random variable. Then the Lorenz curve
L : [0, 1] → [0, 1] is given by:
Rp
Q(x)dx
.
(2)
L(p) = 0
E[X]
Q(p)
. If the mean is infinite, then it is
The derivative L0 of L is given by L0 (p) = E[X]
clear that restarts are always useful as long as some quantile exists. This is a property
which can, for example, be observed for some power-laws. The next theorem provides
a necessary and sufficient condition.
Theorem 1. Let X be a real-valued random variable, then restart are useful if and only
if there is a p ∈ [0, 1) such that
(1 − p)L0 (p) + L(p) < p.
(3)
Proof. The expected runtime with an unbounded number of restarts after Q(p) steps is
given by (see [22]):
E[XQ(p) ] =
1−p
Q(p) + E[X | X < Q(p)]
p
The conditional expectation E[X | X < Q(p)] is defined by
Rp
R Q(p)
0
(4)
xf (x)dx
p
which is
Q(u)du
.
p
This can be obtained by substitution u = F (x). Inserting these
equivalent to 0
identities into E[X] > E[XQ(p) ] and dividing by E[X] yields:
p > (1 − p)L0 (p) + L(p).
(5)
t
u
This completes the proof.
The difference with the condition in [16] is that the existence of E[X] is not required.
Also, since p is limited to [0, 1) it is algorithmically easy to find intervals for which
restarts are useful, while in the condition in [16] the variable is often unbounded. If the
condition in Equation 3 would be an equality instead of an inequality, then the condition
would describe quantiles where restarts are neither harmful nor helpful. Wolter [22]
showed that for the exponential distribution restarts are neither helpful nor harmful.
3.2
Optimal Restarts
In the previous section, a condition for the usefulness of restarts was introduced. This
section focuses on optimal restart times. Wolter [22] provided a relationship between
4
3. MAIN RESULTS
the optimal restart times and the inverse hazard rate. Here a condition for the optimal
restart time is shown by using quantile functions. This classification is used to analyze
the optimal restart times of several distributions. It is shown that a condition for the
optimal restart time can be expressed solely in terms of the quantile function.
Theorem 2. Let X be a real-valued random variable with the quantile function Q and
its antiderivative Q. Then all optimal restart times Q(p) have to fulfill:
(p − 1)Q(p) + p(1 − p)Q0 (p) − Q(p) + Q(0) = 0,
(6)
where Q0 is the derivative of the quantile function Q.
Proof. The expected runtime under restart after Q(p) steps is given by:
E[XQ(p) ] =
Q(p) − Q(0)
1−p
Q(p) +
.
p
p
(7)
By equating the derivative with zero the function can be minimized. After multiplying
the derivative with p2 the obtained condition is:
(p − 1)Q(p) + p(1 − p)Q0 (p) − Q(p) + Q(0) = 0,
(8)
where Q0 is the derivative of Q. This completes the proof.
t
u
Next, the influence of scale parameters on the condition from Theorem 2 is investigated.
3.3
Scale Parameter
In this section, it is shown that for every continuous family of distributions there is one
parameter which does not have an effect: the scale parameter.
Definition 4 ([15]). Let X be a real-valued, continuous random variable with cdf FX .
A new random variable Y with cdf FY is obtained by the identity FX (x) = FY ( βx ) for
d
β > 0. The parameter β is called a scale-parameter. This is denoted as Y = βX.
With this definition, the main results of the section can be derived.
Theorem 3. Let X be a real-valued, continuous random variable such that restarts are
d
useful. Then for Y = βX restarts are also useful for every β > 0.
Proof. Let X and Y = βX be random variables with β > 0. Then the quantile function
QY , its antiderivative QY and the mean E[Y ] are given by:
QY (p) = βQX (p),
(9)
QY (p) = βQX (p),
(10)
E[Y ] = βE[X]
(11)
Thus, Y fulfills the condition from Theorem 1 iff X fulfills it.
5
t
u
3. MAIN RESULTS
Since the derivative Q0Y is given by Q0Y = βQ0X , the result can be extended to show
that scale parameters do not change the optimal restart time. The proof is similar in its
nature and is therefore omitted.
Theorem 4. Let X be a real-valued, continuous random variable and let q ∈ (0, 1) be
such that QX (q) is the optimal restart time. Let β > 0 be a positive, real number, then
QβX (q) is the optimal restart time for the random variable βX.
These findings show that scale parameters can be ignored in the analysis for restart
times. For several commonly used distributions the properties from Theorem 1 and
Theorem 2 can be applied.
3.4
Log-normal
Due to the central limit theorem, the log-normal distribution arises by the product of
n i.id random variables. Barrero et al. [4] observed log-normally distributed times of
several evolutionary algorithms, including genetic programming, particle swarm optimization, and genetic algorithms. Muñoz et al. [17] empirically showed that the runtime
of several path planning algorithms such as A∗ and T heta∗ follow log-normal distributions. Frost et al. [8] argued that the runtime of several backtracking algorithms follow
log-normal distributions in the case of unsolvable binary CSP instances. Arbelaez [1]
studied the runtime distributions of two SAT solvers and found that for randomly generated instances the log-normal distribution is a good fit. Thus, the log-normal distribution
is commonly used to describe the runtime of local search algorithms. The log-normal
distribution is defined as follows:
Definition 5 ([18]). Let X be a real-valued random variable. If there are parameters
µ > 0, σ > 0, such that the random variable U with
U=
log (X) − µ
σ
(12)
is standard normal distributed, then X is said to be log-normally distributed. In this
case the mean E[X] and the quantile function QX are given by
E[X] = eµ+σ
2
/2
QX (p) = eµ+σ
√
,
2erf
(13)
−1
(2p−1)
,
(14)
where erf −1 is the inverse error function.
The erf −1 function is not analytically solvable, but there are numerical approaches. An
antiderivative Q of QX and the derivative Q0 can be obtained by calculations:
1 µ+σ2 /2
σ
−1
erf √ − erf (2p − 1) ,
Q(p) = − e
(15)
2
2
√
2 √
−1
−1
Q0 (p) = eµ+ 2·σ·erf (2p−1)+(erf (2p−1)) σ 2π.
(16)
Where erf is the error function which only can be computed numerically. With these
definitions the usefulness of restarts for the log-normal distribution can be estimated.
6
3. MAIN RESULTS
Fig. 1. A regionplot for the log-normal distribution. The blue area denotes parameter settings
which fulfill the condition from Theorem 1. For
σ < 0.48 the numerical approach could not
find values of p such that restarts are useful. For
high values of σ even low values of p yield an
improved expected runtime under restart, while
for low values of σ only very high values of p
improve the expected runtime.
2.0
1.5
σ 1.0
0.5
0.0
0.0
0.2
0.4
0.6
0.8
1.0
p
Theorem 5. Let X be a log-normal distributed random variable. Then there is a p ∈
(0, 1) such that
E[XQ(p) ] < E[X].
(17)
Proof. Due to Theorem 3, the scale parameter eµ can be ignored for the analysis of the
usefulness of restarts. Therefore, let µ = 0. Note that E[XQ(p) ] converges to E[X] as
p approaches one. The derivative with respect to p of the expected runtime E 0 [XQ(p) ]
is obtained similarly to Theorem 2 and is given by:
E 0 [XQ(p) ] =
(p − 1)
(1 − p) 0
1
Q(p) +
Q (p) − 2 (Q(p) + Q(0)).
p2
p
p
(18)
The limit of E 0 [XQ(p) ] as p approaches one is analyzed. In this case, it can be seen that
(p−1)
1
p2 Q(p) converges to zero and p2 (Q(p) + Q(0)) converges to E[X]. Therefore the
analysis focuses on the limit of (1 − p)Q0 (p).
√
(1 − p)
lim (1 − p)Q0 (p) = lim √
2π
(19)
2 σ
p→1
p→1 − 2·σ·erf −1 (2p−1)−(erf −1 (2p−1))
e
√ √
−1
√
σ 2e 2·σ·erf (2p−1)
2 σ 2erf −1 (2p−1)
→∞
(20)
= lim √
=
lim
σ
e
p→1
2σ + erf −1 (2p − 1) p→1
This limit is obtained by applying L’Hospital’s rule twice. Thus, since E[XQ(p) ] converges to E[X] and E 0 [XQ(p) ] approaches positive infinity, there is a p ∈ (0, 1) with
E[XQ(p) ] < E[X].
t
u
Figure 1 shows parameter combinations for p and σ for which restarts are useful.
Since neither the error function nor the inverse error function can be solved analytically,
numerical methods have been used to obtain those results.
The optimal restart times, as presented in Theorem 2, are shown in Figure 2a. It can
be seen that for high σ values the optimal restart time quickly approaches Q(0), while
for small σ values the optimal restart time converges to Q(1).
Figure 2b shows the comparison of a log-normal distribution without restarts to a
log-normal distribution with restart at the optimal time. The plot has the same shape for
all values of µ, and only the values on the y-axis differ. It can be observed for values
up to approximately σ ≈ 0.8 the difference in the expected runtime is marginal. On the
σ2
other hand, while the expected value eµ+ 2 behaves super-exponentially for increasing
values of σ, the expected runtime with restarts after an optimal number of steps starts
to decrease at about σ ≈ 1.1. Therefore, in real applications restarts should only be
employed for high values of σ.
7
3. MAIN RESULTS
σ
15
3.0
E[X opt]
E[X]
2.5
10
2.0
1.5
5
1.0
0.5
0.2
0.4
0.6
0.8
1.0
p
(a) The brown line denotes parameter settings of σ and p which
minimize the expected runtime
under restart.
σ
0.5
1.0
1.5
2.0
2.5
3.0
(b) This plot shows the expected runtime for a log-normal
distribution without restart as a dashed line. The blue line
represents the expected runtime with restarts at the optimal restart time.
Fig. 2. The figures show the optimal restart quantiles for a log-normal distribution with µ = 1 on
the left, and the expected runtime and the optimal mean with restarts on the right.
3.5
Generalized Pareto
The Pickands–Balkema–de Haan theorem (see [3]) states that the excess probability
(u)
Pr(X − u ≤ y | X > u) = F (u+y)−F
converges in distribution towards the
1−F (u)
generalized Pareto distribution for a large number of distributions as u → ∞. Due to
this, it can be used to model the tail of distributions. The generalized Pareto distribution
(GP) includes the exponential, the Pareto, and the uniform distribution. Since the Pareto
distribution is a subclass of the GP, the GP is also well suited to describe power-law
decays in the tail. Corvella et al. [6] observed that network transmission times follow a
power-law in the tail. Gomes and Selman [9] found that the runtime of the quasi-group
completion problem is also well described by a power-law. At this point, the generalized
Pareto distribution is formally defined:
Definition 6 ([18]). Let X be a real-valued random variable. If the cdf FX is given by
−1/k
kx
FX (x) = 1 − 1 +
(21)
σ
for σ > 0, k ∈ R, then X has a generalized Pareto (GP) distribution.
σ
For k < 1 the mean of the GP is given by E[X] = 1−k
, otherwise the mean is infinite.
For k = 0 the GP is equivalent to the exponential distribution, and for all k > 0 it
takes the form of a Pareto distribution. For k ≤ −0.5 the GP has finite support, and for
k = −1 it becomes a uniform distribution. Compare [18] for these results. The Quantile
function Q, an antiderivative Q and the derivative Q0 are given by:
σ
(1 − p)−k − 1 ,
(22)
Q(p) =
k
σ
(1 − p)1−k
Q(p) = − p +
,
(23)
k
1−k
Q0 (p) = σ(1 − p)−1−k .
(24)
8
3. MAIN RESULTS
These definitions can be used to analyze the usefulness of restarts.
Theorem 6. Let X be a generalized Pareto distribution, there is a p ∈ [0, 1) with
E[XQ(p) ] < E[X]
(25)
if and only if k > 0.
Proof. Due to Theorem 3, the scale parameter σ does not influence the usefulness of
restarts. Therefore, define σ = 1 for the rest of this proof. For the case k ≥ 1 the mean
is infinite. Thus, restarts are useful for all p ∈ (0, 1). Thus, only the case where k < 1
is considered. Then the Lorenz curve L and its derivative L0 are given by:
Q(p) − Q(0)
1
1 − (1 − p)1−k − (1 − k)p ,
=
E[X]
k
1−k
Q(p)
=
(1 − p)−k − 1 .
L0 (p) =
E[X]
k
L(p) =
(26)
(27)
By inserting the Lorenz function and its derivative in inequality 3, the following can be
obtained by some calculations:
p > −(1 − p)1−k + 1
⇔k log (1 − p) < 0
(28)
(29)
For p ∈ (0, 1) the left side of the equation k log (1 − p) is negative if and only if k is
positive. Therefore restarts are useful for all p ∈ [0, 1) iff k > 0.
t
u
This is consistent with other results in this field. Wolter [22] showed that Restarts are
useful for the Pareto distribution, they are not useful for the uniform distribution and are
neither helpful nor harmful for the exponential distribution. The result presented here
is stronger since all of those three distributions are subclasses of the generalized Pareto
distribution while other distributions can be obtained by using different parameters. The
next theorem analyzes the optimal restart time.
Theorem 7. Let X be a generalized Pareto distributed random variable with k > 0,
then the optimal restart time is zero and the optimal expected runtime under restart
E[XQ(0) ] is E[XQ(0) ] = σ.
Proof. Due to Theorem 4, the scale parameter σ does not have any influence on the
optimal restart time. Hence, it can be set to σ = 1 for the analysis. Then the condition
as in Theorem 2 is:
1
1
(1 − p)1−k
−k
−p−
+ (1 − p) (1 − p) − 1 − p(1 − p)−k = 0. (30)
k 1−k
1−k
By transforming the equation, this condition can be simplified to (1−p)−k (1−kp) = 1.
This is obviously true for p = 0. Actually, it can also be shown that p = 0 is the only
value which fulfills the condition. Differentiating (1 − p)−k (1 − kp) with respect to p
(k−1)kp
yields (1−p)
k (p−1) . For k ∈ (0, 1) this is strictly positive, and for k ∈ (1, ∞) this is
9
3. MAIN RESULTS
strictly negative for p 6= 0. For this reason, the condition is either strictly monotonically
increasing or strictly monotonically decreasing. Therefore, every other value of p does
not fulfill equation 30. The expected runtime is given by E[XQ(p) ] = 1−p
p Q(p) +
Q(p)−Q(0)
.
p
In case of GP the limit exists for p → 0, it can be obtained by applying
L’Hospital’s rule and is given by E[XQ(0) ] → Q0 (0) = σ.
t
u
It is noteworthy that the expected runtime does not depend on the shape parameter k
anymore. When comparing expected runtimes under restart with the expected runtimes
σ
for k ∈ (0, 1), it is easy to see that for k → 0 both
without restarts E[X] = 1−k
runtimes converge against σ. This is consistent with the fact that the generalized Pareto
distribution becomes the exponential distribution for k = 0.
3.6
Weibull
The Weibull distribution was extensively analyzed by Wolter [22]. The Weibull distribution is one of the three limiting distributions of the Fisher-Tippett-Gnedenko Theorem
in case of the minimum value (see for example [11]). Therefore it is a likely candidate when observing the minimum of n i.id random variables X1 , . . . , Xn . Frost et al.
[8] observed that the runtime distributions of several backtracking algorithms can be
reasonably well described by Weibull distributions for solvable binary CSP instances
at the 50% satisfiability point. Hoos and Stützle [10] examined the runtime of a SATsolver (GSAT). They found that for non-optimal parameter settings the runtime can be
described as a Weibull distribution. Barrero et al. [4] studied generation based models
without selective pressure, they argue that the generations-to-success can be modeled
by a Weibull distribution. The Weibull distribution is defined as follows:
Definition 7 ([18]). Let X be a real-valued random variable. The random variable X
has a Weibull distribution if and only if
x k
FX (x) = 1 − e−( a )
(31)
for some fixed k > 0. Then the quantile function QX is given by:
QX (p) = a(− log (1 − p))1/k .
(32)
Wolter [22] showed that restarts are always useful for k < 1. In case k > 1 restarts
are always harmful and in case k = 1 the Weibull distribution becomes the exponential
distribution, therefore restarts are neither useful nor harmful. The same results can be
obtained by the technique from Theorem 1. The derivative Q0 and an antiderivative Q
of Q are given as follows:
1
(− log (1 − p)) k −1
Q (p) = a
,
k(1 − p)
1
Q(p) = aγ 1 + , − log (1 − p) ,
k
0
(33)
(34)
Rx
where γ(z, x) = 0 tz−1 e−t dt is the incomplete gamma function. The optimal restart
time is analyzed in the next theorem.
10
3. MAIN RESULTS
Theorem 8. Let X be a Weibull distributed random variable with k < 1. Then the optimal restart time is zero and the optimal expected runtime under restart is E[XQ(0) ] = 0.
Proof. The quantile function, its derivative and the antiderivative can be inserted into
Theorem 2 which yields:
1
1
p
γ 1 + , − log (1 − p) +(− log (1 − p)) k (1−p− (− log (1 − p))−1 ) = 0. (35)
k
k
The scale parameter a can be ignored due to Theorem 4. Therefore, define a = 1 for
1
this proof. Since k1 − 1 is strictly positive for k < 1 the term kp (− log (1 − p)) k −1
approaches zero as p approaches zero, hence, equation 35 approaches zero. Then, the
(1−p))
expected runtime E[XQ(p) ] approaches Q0 (p) = a(− logk(1−p)
0 if k < 1
E[XQ(0) ] = a if k = 1
∞ if k > 1.
1 −1
k
.
(36)
Zero is, by definition, the lowest possible runtime, restarting at Q(0) is therefore the
optimal restart strategy for Weibull distributions with k < 1.
t
u
It is remarkable that the optimal expected runtime in the case of k < 1 is not dependent
on any parameter of the distribution.
3.7
Location Parameter
Up to now, several multiplicative variants of random variables X were discussed. Nonetheless, many commonly used distributions require an additive location parameter b ∈ R
which shifts the support of the cdf. In this section, the results are augmented with this
extension and it is shown that the influence of the location parameter b is limited.
In the following, let X be a random variable without location parameter and let
Y = X + b be a random variable with some location parameter b. The expected value
is known to be a linear function, therefore E[Y ] = E[X] + b. Similar results follow
easily for the quantile function. Since FY (x) = FX (x − b) holds QY (p) = QX (p) + b
directly follows. Then an antiderivative of QY is given by QX (p) + pb where QX is
an antiderivative of QX . Define c ∈ R such that b = cE[X]. With these identities the
usefulness of restarts can be reestimated:
1−p
1
(QX (p) + b) + b + (QX (p) − QX (0))
p
p
⇔p + (p − 1)c > (1 − p)L0X (p) + LX (p)
E[X] + b >
(37)
(38)
is the new condition for the usefulness of restarts. This inequality assumes E[X] > 0,
for E[X] < 0 the ‘greater than’ sign becomes a ‘less than’ sign. However, this only
makes sense if the location parameter shifts the support of the distribution to strictly
positive values. It can be shown that the condition for the optimal restart time only
changes by the location parameter itself. This can be shown by similar transformations;
the proof is therefore omitted.
11
4. DISCUSSION
Theorem 9. Let X be a random variable and let Y = X + b be a random variable
with location parameter b ∈ R. Then the condition for the optimal restart time for Y is:
(p − 1)QX (p) + p(1 − p)Q0X (p) − QX (p) + QX (0) = b.
(39)
If Q0X dominates (1 − p) in the neighborhood of p = 1, then the left side of equation 39
approaches infinity. This implies that restarts are useful for an arbitrarily large b since
E[XQ(1) ] = E[X]. This is the case for all distributions considered in this article.
Corollary 1. Let X be log-normal, GP, or Weibull distributed, with parameters such
that restarts are useful. Let b ∈ R+ , then there is a p ∈ (0, 1) with
E[(X + b)Q(p) ] < E[X + b].
(40)
Note, that this is not a general property which is true for all distributions which admit
useful restarts. Counterexamples are distributions with a finite support. With this relationship, it is reasonable to analyze distributions without location parameters and scale
parameters. This simplifies the analysis whether a restart strategy should be employed
and if so, which strategy should be chosen.
4
Discussion
This work discussed the relationship between the quantile function and restarts using
the fixed cut-off strategy. Theorem 1 uses the quantile function and its antiderivative
to provide a condition for the usefulness of restarts, while Theorem 2 established a
condition for the optimal restart times. It was proven in Theorem 3 and 4 that scale
parameters can be ignored in the context of restarts. In section 3.7 it was shown that
the influence of location parameters on the usefulness of restarts is limited. For a large
group of distributions the usefulness of restarts is not affected at all by the presence of
a location parameter. Secondly, the optimal restart times are just linearly influenced by
a location parameter. Therefore, it often suffices to analyze the remaining parameters.
Several commonly used distributions were observed for their usefulness under restart.
In the following, the log-normal distribution (compare section 3.4) and the generalized Pareto distribution (compare section 3.5) were discussed. It was shown that in the
case of the log-normal distribution restarts are always useful. In case of the generalized
Pareto distribution, restarts are useful iff the shape parameter k is greater than zero. The
optimal restart times and optimal expected runtimes under restart of the log-normal, the
generalized Pareto and the Weibull distribution were discussed. The expected runtime
without restart for the log-normal distribution increases super-exponentially in σ. For
the log-normal distribution, it was numerically observed that for increasing parameter
σ the optimal restart time is decreasing. And while the expected value without restarts
increases super-exponentially, the expected value with restart at the optimal time starts
to decrease at about σ ≈ 1.1. It is also interesting to see that for low values of σ the
speedup with restarts is marginal. This is especially important if the parameters of the
distribution are not completely known. Figure 3 represents an example of this behavior.
This shows that choosing a suboptimal restart time can easily result in expected runtimes under restart which are worse than not employing a restart strategy. Therefore, in
12
4. DISCUSSION
Fig. 3. This figure depicts the expected runtime of a log-normally distributed random
variable X with µ = 0 and σ = 0.7 as a
dashed line. The blue line is the expected runtime with restart after QX (p) steps.
1.34
E[X Q (p)]
E[X]
1.32
1.30
1.28
0.0
0.2
0.4
0.6
0.8
1.0
p
practice it can be better to not employ a restart strategy if the parameters are estimates
and σ is estimated to be low.
For the Weibull distribution, it was shown that the optimal restart time and the expected runtime is zero. A similar behavior was observed for the GP distribution. The
optimal restart time is also zero, its optimal expected runtime under restart, however,
approaches σ. It is noteworthy that in these two cases the expected runtime is no longer
dependent on the shape parameter and in the case of the Weibull distribution also not dependent from the scale parameter. For algorithms, an intuitive description of the restart
quantile zero in, e.g., runtime distributions with location b and scale a emerges from the
optimum restart condition (p − 1)QX (p) + p(1 − p)Q0X (p) − QX (p) + QX (0) = b/a.
If a b, the optimal restart time approaches Q(0) = b, i.e., the algorithm’s behavior
before b dominates all subsequent steps. It is also remarkable that some distributions
which show suboptimal behavior without restarts yield low runtimes when an optimal
restart strategy is applied.
References
1. Arbelaez, A., Truchet, C., Codognet, P.: Using sequential runtime distributions for the parallel speedup prediction of SAT local search. Theory and Practice of Logic Programming
13(4-5), 625–639 (2013)
2. Arbelaez, A., Truchet, C., O’Sullivan, B.: Learning sequential and parallel runtime distributions for randomized algorithms. In: ICTAI 2016: 28th International Conference on Tools
with Artificial Intelligence, San Jose, California, USA. pp. 655–662. IEEE (2016)
3. Balkema, A.A., De Haan, L.: Residual life time at great age. The Annals of probability pp.
792–804 (1974)
4. Barrero, D.F., Muñoz, P., Camacho, D., R-Moreno, M.D.: On the statistical distribution of
the expected run-time in population-based search algorithms. Soft Computing 19(10), 2717–
2734 (2015)
5. Caniou, Y., Codognet, P.: Sequential and Parallel Restart Policies for Constraint-Based Local
Search. In: Proceedings of the 2013 IEEE 27th International Symposium on Parallel and
Distributed Processing Workshops and PhD Forum. pp. 1754–1763. IEEE Computer Society
(2013)
6. Crovella, M.E., Taqqu, M.S., Bestavros, A.: Heavy-tailed probability distributions in the
World Wide Web. A practical guide to heavy tails 1, 3–26 (1998)
7. Evans, M.R., Majumdar, S.N.: Diffusion with Stochastic Resetting. Physical Review Letters
106(16), 160601 (2011)
8. Frost, D., Rish, I., Vila, L.: Summarizing CSP Hardness with Continuous Probability Distributions. In: Proceedings of the Fourteenth National Conference on Artificial Intelligence
13
4. DISCUSSION
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
and Ninth Conference on Innovative Applications of Artificial Intelligence. pp. 327–333.
AAAI’97/IAAI’97, AAAI Press (1997)
Gomes, C.P., Selman, B., Crato, N.: Heavy-tailed distributions in combinatorial search. In:
International Conference on Principles and Practice of Constraint Programming. pp. 121–
135. Springer (1997)
Hoos, H.H., Stützle, T.: Evaluating las vegas algorithms: pitfalls and remedies. In: Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence. pp. 238–245. Morgan Kaufmann Publishers Inc. (1998)
Kotz, S., Nadarajah, S.: Extreme value distributions: theory and applications. World Scientific (2000)
Lorenz, J.H.: Completion Probabilities and Parallel Restart Strategies under an Imposed
Deadline. PloS one 11(10) (2016)
Lorenz, M.O.: Methods of measuring the concentration of wealth. Publications of the American statistical association 9(70), 209–219 (1905)
Luby, M., Sinclair, A., Zuckerman, D.: Optimal speedup of Las Vegas algorithms. Information Processing Letters 47(4), 173–180 (1993)
Meyer, J.: Two-moment decision models and expected utility maximization. The American
Economic Review pp. 421–430 (1987)
Moorsel, A.P.A.v., Wolter, K.: Analysis and algorithms for restart. In: Proceedings of the
First International Conference on the Quantitative Evaluation of Systems. pp. 195–204
(2004)
Muñoz, P., Barrero, D.F., R-Moreno, M.D.: Run-time analysis of classical path-planning
algorithms. In: Research and Development in Intelligent Systems XXIX, pp. 137–148.
Springer (2012)
Norman, L., Kotz, S., Balakrishnan, N.: Continuous Univariate Distributions, Vol. 1 of wiley
series in probability and mathematical statistics: applied probability and statistics (1994)
Paxson, V., Allman, M., Chu, J., Sargent, M.: Computing TCP’s retransmission timer. Tech.
rep. (2011)
Reuveni, S., Urbakh, M., Klafter, J.: Role of substrate unbinding in Michaelis–Menten enzymatic reactions. Proceedings of the National Academy of Sciences of the United States of
America 111(12), 4391–4396 (2014)
Schöning, U.: A Probabilistic Algorithm for k-SAT and Constraint Satisfaction Problems.
In: Proceedings of the 40th Annual Symposium on Foundations of Computer Science. pp.
410–. FOCS ’99, IEEE Computer Society, Washington, DC, USA (1999)
Wolter, K.: Stochastic Models for Fault Tolerance: Restart, Rejuvenation and Checkpointing.
Springer Science & Business Media (2010)
Wu, H.: Randomization and restart strategies. Master’s thesis, University of Waterloo (2006)
14
| 8cs.DS
|
Statistical Inference for Oscillation Processes
arXiv:1412.4912v3 [math.ST] 12 Aug 2016
Rainer Dahlhaus∗, Thierry Dumont†, Sylvain Le Corff‡, Jan C. Neddermeyer§
Abstract
A new model for time series with a specific oscillation pattern is proposed. The model consists
of a hidden phase process controlling the speed of polling and a nonparametric curve characterizing
the pattern, leading together to a generalized state space model. Identifiability of the model is
proved and a method for statistical inference based on a particle smoother and a nonparametric
EM algorithm is developed. In particular, the oscillation pattern and the unobserved phase process
are estimated. The proposed algorithms are computationally efficient and their performance is
assessed through simulations and an application to human electrocardiogram recordings.
Keywords. Oscillation process; phase estimation; instantaneous frequencies; oscillation pattern;
Rao-Backwellized particle smoother; Rössler attractor; electrocardiogram.
1
Introduction
In this paper, we propose a model for the statistical analysis of oscillating time series. In its simplest
form, the model is a generalized state space model (GSSM) with nonlinear observation equation
Yt = f (φt ) + εt ,
t ∈ N+
iid
with εt ∼ N (0, σε2 ) ,
(1)
where f is an unknown 2π-periodic function (the oscillation pattern), {φt }t∈N+ is an unobserved
stochastic phase process (the internal clock of the oscillator), and the {εt }t∈N+ are independent of the
process {φt }t∈N+ . The phase process {φt }t will be an integrated process which may either slow down
or speed up the cycle, for example φt may be modeled by the state equation
∆φt = ω + ηt ,
t ∈ N+
iid
with ηt ∼ N (0, ση2 ),
ω ∈ (0, π) ,
(2)
where ∆φt := φt − φt−1 , and φ0 is assumed to be uniformly distributed on [0, 2π] independent of the
Pt
other variables. In this case the ∆φt are iid and φt = φ0 + ω t + St with St := s=1 ηs being a random
walk with variance var(St ) = ση2 t.
It is informative to compare the above model with the model φt = φ0 + ω t + Rt where Rt is not
integrated but stationary. These two models reflect two completely different situations with different
type of oscillators: The latter can be used as a model where the oscillator sticks except from small
∗ Institute
of Applied Mathematics, Heidelberg University, Germany.
Université Paris-Ouest, Nanterre, France.
‡ Laboratoire de Mathématiques d’Orsay, Univ. Paris-Sud, CNRS, Université Paris-Saclay, 91405 Orsay, France.
§ DZ BANK AG, Frankfurt, Germany.
† MODAL’X,
1
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Figure 1: 1000 Observations from an ECG
Figure 2: 1415 Observations from a Rössler Attractor
deviations to an external “pacemaker” (say where a hormone is driven by the circadian rhythm), while
the integrated phase model of this paper is e.g. a suitable model for ECG data (cf. Figure 1 - to
be analyzed in Section 4.3) and for the data of a Rössler attractor (cf. Figure 2 - to be analyzed in
Section 4.2) where one does not know in advance at which point of the cycle we will be at a specific
time point. This is reflected by a constant variance of φt in the stationary model and an increasing
variance of φt in the present model (2).
We also mention the classical “hidden frequency” model of Hannan (1973) and others (for an
overview see Quinn and Hannan (2001) and Li (2013)), which in its simplest form is Yt = A cos(ωt)+εt
with a stationary process εt . In the notation of this paper this work dealt with the case f (·) = cos(·)
and φt = ωt. There exist many papers with rather deep mathematical results on the estimation of ω
in this framework - in particular via maximization of the periodogram which originally dates back to
Schuster (1897). The case of a time varying hidden frequency ωt is more in the spirit of this paper. In
this case one may use the maximization of the periodogram on rolling data windows (c.f. Paraschakis
and Dahlhaus, 2011).
Statistical inference for the above model is challenging since (1) is a nonparametric regression model
with unknown regressors (but, due to periodicity, with asymptotically infinite many replications). In
Section 2 and Section 6, we prove identifiability of model (1), (2). Dumont and Le Corff (2014) have
proved identifiability for a different nonparametric GSSM.
In Section 3 we then show how the unknown oscillation pattern f and the unobserved phase process
2
Dahlhaus et al.
Statistical Inference for Oscillation Processes
(φt )t can be estimated / predicted. In that section we consider the more general model
Yt = at f (φt ) + bt + εt ,
(3)
which includes a time varying amplitude at and a time varying baseline bt . We also allow for the more
general state equation (13) which guarantees positivity of the phase increments and includes some
dependence of the phase increments. Maximum likelihood based estimation of φt , at and bt requires
computation of the smoothing distributions, i.e. the posterior distributions of the sequence of hidden
states given all observations Y0 , . . . , YT . These distributions cannot be computed explicitly using the
GSSM model of this paper, and we propose in Section 3 an efficient (fixed-lag) Rao-Blackwellized
particle smoother that combines the Kalman smoother and the fixed-lag particle smoother introduced
and analyzed in Olsson et al. (2008). Estimates of φt , at and bt can then be obtained as means of the
smoothing distributions. In Section 3.3 and 3.4, a nonparametric EM algorithm is developed for the
estimation of the function f and other parameters. For a recent overview of sequential Monte Carlo
methods see Douc et al. (2014).
As a by-product we also obtain a method for nonlinear phase estimation in the case where f (x) =
cos(x) is known. Such an estimator is needed in several applications - for example for the detection
of phase synchronization of chaotic oscillators (cf. Pikovsky et.al., 2001); in neuroscience for the
investigation of functional coupling of different brain regions (cf. Fell and Axmacher, 2011), and in
engineering for channel decoding (Chen et.al., 2003). Furthermore, the estimation of instantaneous
frequencies is a key step in the widely used empirical mode decomposition (cf. Huang et.al., 1998).
Traditional approaches for phase estimation are based on the Hilbert transform (cf. Pikovsky et.al.,
2001) or the Wavelet transform (c.f. Grossmann et.al., 1989). In practice, these methods often fail
not only in situations of fast varying frequencies but also in cases where the signal is corrupted by
noise. In such a situation statistical methods such as variants of Hannan’s hidden frequency approach
or the method of this paper are beneficial. Phase estimation with a GSSM has been used in Tsakonas
et.al. (2008), in the context of communication systems in Amblard (2003), and in the context of audio
processing in Dubois and Davy (2007) - in all cases in combination with a different model for the state
and only for f (x) = cos(x) (i.e. in particular under the assumption that f is known).
As mentioned above, Section 2 contains identifiability and Section 3 statistical inference in the
framework of a GSSM. Nonlinear phase estimation in the special case f (·) = cos(·) is explored in a
simulation study in Section 4.1 and for the nonlinear Rössler-attractor in Section 4.2 followed by an
application to ECG data in Section 4.3. Section 5 contains concluding remarks and the appendices in
Section 6 and 7 the proofs.
3
Dahlhaus et al.
2
Statistical Inference for Oscillation Processes
Identifiability of Oscillation Processes
Since both the function f and the phases φt are unobserved, identifiability is a nontrivial issue discussed
in this section. One comment beforehand: as usual identifiability means identifiability of f and the
parameters of the process φt and not “identifiability” of φt . The latter is not possible - the best we
can achieve is e.g. the determination of the best predictor of φt given the observations. We start with
a heuristic discussion of the identifiability problems:
1) Shifting the oscillation pattern: it is obvious that the starting point of the oscillation pattern
cannot be identified. Formally,
f (φt ) = f˜(φ̃t ) with f˜(x) = f (x − θ) and φ̃t = φt + θ .
Sometimes there exists a natural starting point known from the applied problem at hand - in other
cases one can just define the starting point arbitrarily.
2) Non-identifiability of the mean phase increment: the mean phase increment ω can only be identified from the data under additional assumptions. For example, for the model (2) where φt =
φ0 + ω t + σ St , we have
f (φt ) = f (φ0 + ω t + σ St ) = f (φ0 + (ω + 2π`) t + σ St ) ,
i.e. ω is only identifiable mod 2π. Furthermore, if f is symmetric then
f (φt ) = f (−φt ) = f (−φ0 − ω t − σ St ) = f (−φ0 + (2π − ω) t − σ St ) ,
i.e. ω and 2π − ω can only be identified under the additional assumption that f is not symmetric
or the distribution of ηt is not symmetric. Since we are mainly interested in the case where each
cycle contains several data points (corresponding to a small ω), we assume throughout this paper
that ω ∈ (0, π) guaranteeing identifiability.
Note that the above discussion is related to the Nyquist frequency and aliasing: If we regard the
original signal as continuous in time with phase process φt = φ0 + ω t + ση B(t) where B(t) denotes
a Brownian motion, and we sample the process at times ∆t, then the same arguments as above
show that we can only identify ω ∈ 0, π/∆ with aliases at 2π/∆ − ω and all frequencies shifted
by 2π`/∆, ` ∈ Z.
3) Minimal period length / the basic cycle: let
n
repl(f ) := sup r ∈ R f
·
r
o
is 2π periodic
be the number of periodic replications in f . Then f˜(·) := f
·
repl(f )
(4)
is called the basic cycle of the
oscillation. In Theorem 2.2, we prove that repl(f ) ∈ N+ and that the basic cycle is unique up to
4
Dahlhaus et al.
Statistical Inference for Oscillation Processes
the shift of the starting point. Thus, if we fix the starting point θf in the basic cycle we have
fbasic (x) := f
x − θ
f
(5)
repl(f )
with a unique fbasic . If we have two representations with different f1 and f2 then
f1
x−θ
x−θ
x−θ
1
2
),
= fbasic (x) = f2
i.e. f1 (x) = f2 (
repl(f1 )
repl(f2 )
γ
with γ = repl(f2 )/repl(f1 ) and θ = (θ2 −θ1 )/repl(f1 ) (see Theorem 2.1 and 2.2 below).
It is important to note how the phases transform when moving from the oscillation pattern f1 to
(1)
f2 . If {φt } fulfills model (2) we have
(1)
(1)
(1)
(2)
(2)
f1 (φt ) = f2 (φt ) with φt
=
φt − θ
γ
(2)
i.e. ∆φt
=
∆φt
γ
=
ω1
σ1
+
ηt .
γ
γ
(2)
Thus {φt } also fulfills model (2) with γω2 = ω1 and γσ2 = σ1 (see also Theorem 2.1). For
identifiability reasons, we usually assume that the period length is the minimal one, i.e. we use
the basic cycle as our oscillation pattern.
In practice, the discrimination between the basic cycle and multiple replications is often clear
from eye-inspection (as in Figure 1). One can incorporate this external information into the EM
algorithm from Section 3.4 in a quite elegant way: as demonstrated above the information on the
multiplicity of the cycle (say f (x) = fbasic (rx) with r ∈ N+ ) is also contained in the drift-parameter
which becomes ω = ωbasic /r. Incorporating external information can then be achieved by choosing
an appropriate starting value for ω in the algorithm (heuristically the EM algorithm then finds
that local maximum which corresponds to the basic cycle). In practice, we may count the number
of basic cycles in the data, multiply it by 2π and divide it by the number of time points leading
to a rough estimate of ω which we use as the initial value. The ECG-example in Section 4.3
shows that this works remarkably well - even with the uninformative starting value f (0) ≡ 0. The
starting value for ω then allows for a reasonable simulation of φt (the particles) leading to improved
estimates of f in the next iteration steps.
4) Time-warping:
non-identifiability due to time-warping seems to be a serious problem. Time
warping means a transformation of the observation model of the form
yt = (f ◦ g −1 ) g(φt mod 2π) + εt ,
with an increasing function g : [0, 2π] → [0, 2π] leading to the new oscillation pattern f˜ = f ◦ g −1
and the new phases φ̃t = g(φt mod 2π). The problem of time-warping has been discussed in
nonparametric regression in a large number of papers (cf. Kneip and Gasser, 1992; Wang and
Gasser, 1997).
5
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Luckily, the present model rules out time warping in a very natural way: if φ0 ∼ U [0, 2π] and
the increments ∆φt are independent of φ0 then φt mod 2π ∼ U [0, 2π]. However, this is only
true for φ̃t if g(x) = x, i.e. time warping is automatically ruled out. If we abstain from the
assumption φ0 ∼ U [0, 2π], then the assumption of stationarity of the increments implies that
D
φt mod 2π → U [0, 2π], meaning that the assumption of stationarity of the increments prevents
time-warping.
In the more complicated model yt = at f (φt ) + bt + εt with amplitude at and baseline bt (modeled by
stochastic processes) there arise two additional identifiability issues.
5) Amplitude of f : in case of a time varying amplitude at , we remove non-identifiability by assuming
Eat = 1, i.e. at measures the relative deviation of the amplitude over time; in case of a known
oscillation pattern (e.g. f (x) = cos(x)), we make no assumption on Eat .
6) Level of f : in case of a time varying baseline bt , we remove non-identifiability by assuming Ebt = 0
i.e. bt measures the deviation of the baseline over time; in case of a known oscillation pattern, we
make no assumption on Ebt .
We now prove identifiability in a strict sense of model (1), (2). In addition to the assumptions stated
in Section 1 we assume that f is a 2π periodic function with f ∈ F, where
X
F := f : R → R f (x) =
ck eikx with {ck } ∈ `2 (Z) .
k∈Z
2
2
Theorem 2.1. Let f? ∈ F be a non-constant oscillation pattern and ω? ∈ (0, π), σε?
, ση?
be the
parameter set of the state space model (1), (2). Assume that there exist f ∈ F and ω ∈ (0, π), σε2 ,
ση2 such that the finite dimensional distributions of {Yk }k∈N+ given by the two parameter sets are the
same. Under the additional assumption that ω? ∈
/ πQ, there exists a γ ∈ Q and a θ ∈ [0, 2π) such
that σε = σε? , γση = ση? , γω = ω? , and f ( x−θ
γ ) = f? (x) for all x ∈ R.
The proof is given in Section 6. We just mention here that we can identify σε , ση , ω, and the
squared Fourier-coefficients of f from the second order structure of {Yk } while for the identification
of f higher order moments are needed. The assumption ω? ∈
/ πQ is dispensable in our opinion but we
were unable to prove the result without it. In order to gain a deeper insight into the situation and to
explain the constant γ we define
ck (f ) :=
1
2π
Z
π
f (x) e−ikx dx
(6)
−π
and the sequence {κi (f )}i≥1 by
κ1 (f ) = inf {k ≥ 1; ck (f ) 6= 0}
and κi+1 (f ) = inf {k ≥ κi (f ) + 1; ck (f ) 6= 0} .
6
(7)
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Theorem 2.2. Assume that there exist an oscillation pattern f and ω ∈ (0, π), σε2 , ση2 such that the
conditions of Theorem 2.1 hold. Then, the basic cycle fbasic (·) defined in (5) is unique and every
oscillation pattern is an `-times replication of the basic cycle with ` ∈ N+ . Thus also repl(f ) ∈ N0 .
Furthermore,
repl(f ) = max ` ∈ N+ ck (f ) = 0 ∀ k 6= `N+ ,
(8)
and we have for γ from Theorem 2.1
γ=
repl(f )
repl(f? )
and
γ=
κi (f )
for all i ∈ N+ .
κi (f? )
(9)
In addition, the {κi (fbasic )}i∈N+ are setwise coprime.
Note that repl(f ) is not necessarily equal to κ1 (f ). An example is f (x) := cos(2x)+cos(3x).
3
Statistical Inference for Oscillation Processes
As our model is a GSSM statistical inference may be performed using Sequential Monte Carlo techniques. In the following, the estimation of the oscillation pattern f and of the other parameters uses
a fixed-lag particle smoother in combination with a nonparametric MCEM algorithm.
3.1
The Model
For practical purposes we modify/generalize the model (1), (2) in two aspects.
1) In order to guarantee positivity of the increments, we use for the phase differences ∆φt the
ACD(1,0) (autoregressive conditional duration) model:
∆φt = α + β∆φt−1 ηt ,
(10)
where the ηt are e.g. Beta or Gamma distributed with Eηt = 1 (if Eηt 6= 1, then ηt can be replaced
by η̃t = ηt /Eηt ). We assume that α, β > 0, β < 1, and α < π(1 − β). It then can be shown that the
(unconditional) mean of the phase increments is
ω = E[∆φt ] =
α
< π.
1−β
(11)
The ACD model was originally introduced by Engle and Russell (1998) as a model for the dependency
of the durations between consecutive transactions in financial markets.
2) For data such as the ECG data we add a time varying amplitude at and a time varying baseline
bt as hidden states. In order to keep the computations simple we use a Gaussian random walk
model. This has in particular the advantage that we can use a (fixed-lag) Rao-Blackwellized particle
smoother that combines a sequential particle smoother for φt with a Kalman smoother for at and bt
(Doucet et.al., 2000b). Estimates of φt , at and bt can then be obtained as the means of the smoothing
7
Dahlhaus et al.
Statistical Inference for Oscillation Processes
distributions. For a recent overview of sequential Monte Carlo methods see Douc et al. (2014).
Therefore, we use the observation model:
yt = at f (φt ) + bt + εt ,
with
φt
ψt
=
φt−1 + (α + βψt−1 )ηt
(α + βψt−1 )ηt .
at
µa
at−1
µ
ξ
and
−
=A
− a
+ t ,
bt
µb
bt−1
µb
ζt
(12)
(13)
where (ξt , ζt )T ∼ N (0, Q) and εt ∼ N (0, σε2 ). It is assumed that εt , ηt and (ξt , ζt )T are mutually
and serially independent. For simplicity we assume that A = diag(1, 1) (this is a typical trend
model) and Q is diagonal. We assume that µa = 1 and µb = 0 in case where the oscillation pattern
f is a nonparametric function. If (say) f (·) = cos(·) we assume that µa and µb are parameters
to be estimated. In the setting of constant (but unknown) amplitude and baseline, one may use
(at , bt )T = (µa , µb )T , which simplifies the estimation significantly.
3.2
Rao-Blackwellized Fixed-Lag Particle Smoothing
Statistical inference of the model introduced in Section 3.1 using an EM algorithm requires an approximation of the joint smoothing distributions p(x0:t |y1:T ) for 0 ≤ t ≤ T . In order to approximate
these posterior distributions we use a particle smoother (cf. Doucet et.al., 2001). The aim of particle
filters and smoothers introduced in Gordon et al. (1993) and Kitagawa (1996) is to obtain recursively
an approximation of the posterior distributions of the states given the observations using importance
sampling and importance resampling steps. These approximations are based on a set of points, the
particles, associated with importance weights.
i
Assume that the set of weighted particles {(xi0:t−1 , ωt−1
)}N
i=1 approximates p(x0:t−1 |y1:t−1 ) by
PN
i
i=1 ωt−1 δxi0:t−1 (x0:t−1 ), with δ the Dirac delta function. Then, for all 1 ≤ i ≤ N ,
- Sample xit ∼ p(xt |xit−1 ).
i
p(yt |xit ).
- Compute the importance weight ω̆ti ∝ ωt−1
i N
i N
The new particles {(xi0:t , ωti )}N
i=1 , where {ωt }i=1 are the normalized weights obtained from {ω̆t }i=1 ,
approximate the posterior distribution p(x0:t |y1:t ). To avoid weight degeneracy, a resampling step that
maps the particle system {(xi0:t , ωti )}N
i=1 onto an equally-weighted particle system is introduced, where
i N
each new particle is chosen in {xi0:t }N
i=1 according to {ωt }i=1 . Resampling is carried out whenever
PN
the effective sample size (Kong et.al., 1994) ( i=1 (ωti )2 )−1 is below some threshold. We favor the
systematic resampling with threshold 0.2N (see Douc et.al., 2005, for alternatives).
To avoid degeneracy of past samples as new observations become available, we propose to use the
fixed lag smoother of Olsson et al. (2008). For computational reasons, this method is likely to be
8
Dahlhaus et al.
Statistical Inference for Oscillation Processes
more efficient than algorithms based on forward backward decompositions as the ‘Forward Filtering
Backward Smoothing algorithm’ of Huerzeler and Künsch (1998) and Doucet et.al. (2000a) or the
‘Forward Filtering Backward Simulation algorithm’ of Godsill et.al. (2004). The basic idea of fixed-lag
smoothing is that p(x0:t |y1:T ) should be close to p(x0:t |y1:t(`) ), with t(`) := min{t + `, T }, for a well
chosen lag ` (no resampling step is performed on the past samples when observations are obtained
after time t(`)). As noted by Doucet et.al. (2000b), the posterior distribution can be decomposed as
p(x0:t(`) |y1:t(`) ) = p(a0:t(`) , b0:t(`) |y1:t(`) , φ0:t(`) ) p(φ0:t(`) , ψ0:t(`) |y1:t(`) ) .
p(a0:t(`) , b0:t(`) |y1:t(`) , φ0:t(`) ) is then computed by a Kalman smoother while p(φ0:t(`) , ψ0:t(`) |y1:t(`) ) is
i
i
approximated by samples {(φi0:t(`) , ψ0:t(`)
)T , ωt(`)
}N
i=1 from the particle smoother. Thus the above
relation implies that the marginal densities p(at , bt , φt |y1:T ) are approximated by
p(at , bt , φt |y1:T ) ≈
N
X
ω̃ti N at , bt (ãit , b̃it )T , Σ̃it δφit (φt ) ,
(14)
i=1
i
, (ãit , b̃it ) := (ait|t(`) , bit|t(`) ) = E (at , bt ) y1:t(`) , φi0:t(`) and Σ̃it := Σit|t(`) are computed
where ω̃ti = ωt(`)
with the Kalman smoother. Smoothing by marginalization has been criticized for causing sample
impoverishment (Doucet et.al., 1999). While this is true in general, it is not an issue in the setting of
this article because the lag ` is not large and the resampling frequency is low. In contrast to smoothing
algorithms which proceed backwards in time (c.f. Godsill et.al., 2004; Doucet et.al, 2000a), smoothing
by marginalization has the advantage that it can be applied on-line. When the observation at time t(`)
comes in, the estimates of time t can be updated using the fixed-lag smoothing density. In addition,
it is computationally very cheap. The RBPS has computational costs O(`N T ) for T smoothing time
steps. At each iteration, only the particles for times t(`) − 1, . . . , t are required, implying a storage
requirement of O(`N ). The following algorithm can e.g. be found in Shumway and Stoffer (2011),
Property 6.2 and 6.3.
Rao-Blackwellized Fixed-Lag Particle Smoother (RBPS)
Initialization (for t = 0)
• For i = 1, . . . , N : Sample (φi0 , ψ0i )T ∼ p(φ0 , ψ0 ), set ω0i = 1, and choose ai0 , bi0 , Σi0 according to
prior knowledge.
Filtering (for t = 1, 2, . . .)
1. Kalman Prediction Step
• For i = 1, . . . , N : Compute with µ := (µa , µb )T
(ait|t−1 , bit|t−1 )T = µ + A (ait−1 , bit−1 )T − µ ,
9
Σit|t−1 = AΣit−1 AT + Q.
Dahlhaus et al.
Statistical Inference for Oscillation Processes
2. Importance Sampling Step
i
• For i = 1, . . . , N : Sample (φit , ψti )T ∼ p(φt , ψt |φit−1 , ψt−1
), compute
i
Ft|t−1
= Cti Σit|t−1 (Cti )T + σε2 with Cti = (f (φit ), 1) and evaluate importance weights
i
i
i
ω̆ti ∝ ωt−1
p(yt |y1:t−1 , φi0:t ) = ωt−1
N yt Cti (ait|t−1 , bit|t−1 )T , Ft|t−1
.
• For i = 1, . . . , N : Normalize importance weights ωti = ω̆ti /(
PN
3. Resampling Step if ( i=1 (ωti )2 )−1 < 0.2N
PN
j=1
ω̆tj ).
4. Kalman Updating Step
• For i = 1, . . . , N : Compute
ait , bit
T
=
Σit
=
n
T
T o i
ait|t−1 , bit|t−1 + Σit|t−1 (Cti )T yt − Cti ait|t−1 , bit|t−1
(Ft|t−1 )−1 ,
n
o
i
Σit|t−1 − Σit|t−1 (Cti )T Cti Σit|t−1 (Ft|t−1
)−1 .
Smoothing
5. Kalman Smoothing Step (for k = t(`) − 1, . . . , t)
• For i = 1, . . . , N : Compute
Vki
=
(ãik , b̃ik )T
=
Σik AT (Σik+1|k )−1 ,
n
T
T o
(aik , bik )T + Vki ãik+1 , b̃ik+1 − aik+1|k , bik+1|k
,
Σ̃ik
=
Σik + Vki (Σ̃ik+1 − Σik+1|k )(Vki )T ,
Σ̃ik,k−1
=
i T
i
Σik Vk−1
)T + Vki (Σ̃ik+1,k − AΣik (Vk−1
) ,
with initial values
(ãit(`) , b̃it(`) )T = (ait(`) , bit(`) )T , Σ̃it(`) = Σit(`)
i
i
and Σ̃it(`),t(`)−1 = I − Kt(`)
Ct+`
A Σit(`)−1 ,
i
i
i
where Kt(`)
:= Σit(`)|t(`)−1 (Ct(`)
)T (Ft(`)|t(`)−1
)−1 is the Kalman gain.
T
Furthermore aik+1|k , bik+1|k = A(aik , bik )T + (I − A)µ.
6. Result
PN
PN
i
i
ãik , baseline estimate b̂k = i=1 ωt(`)
b̃ik , and
• Obtain amplitude estimate âk = i=1 ωt(`)
PN
i
phase estimate φ̂k = i=1 ωt(`)
φik for time k = t.
10
Dahlhaus et al.
3.3
Statistical Inference for Oscillation Processes
The EM Algorithm for the parametric MLE
In this section, we assume that f is known and estimate θ = (α, β, σε2 , µ, A, Q) based on a stochastic
EM algorithm (Dempster et.al., 1977). Shumway and Stoffer (1982) had introduced the EM algorithm
for linear Gaussian state space models. Wei and Tanner (1990) (see also Tanner 1993) had proposed to
replace the E-step by a Monte Carlo integration leading to the MCEM Algorithm. In the present model
the equations for the ‘Gaussian part’ take the same form (conditional on φt ) as the original equations
leading to a partial MCEM Algorithm. This reduces the computational complexity considerably.
Assume that the signal yt is received up to time T . The EM algorithm maximizes the likelihood
pθ (y1:T ) iteratively. In the E-step, the Q(θ|θ(m) ) = Eθ(m) [log pθ (x0:T , y1:T )|y1:T ] is approximated,
where θ(m) is the current estimate. We have
Q(θ|θ(m) ) = Eθ(m) [log p(φ0 , ψ0 )|y1:T ] +
T
X
Eθ(m) [log pθ (yt |xt )|y1:T ]
(15)
t=1
+
T
X
Eθ(m) [log pθ (at , bt |at−1 , bt−1 )|y1:T ] +
t=1
T
X
Eθ(m) [log pθ (φt , ψt |φt−1 , ψt−1 )|y1:T ] .
t=1
It follows that Q(θ|θ(m) ) could be approximated through smoothing particles, which are generated
with respect to the parameter value θ(m) . Due to the computational complexity we use in this paper
only the fixed lag smoother, i.e. we replace Eθ(m) [ · |y1:T ] by Eθ(m) [ · |y1:t(`) ] which can be calculated
by the RBPS. The difference should be minor for reasonable ` (see Olsson et al. (2008) for an explicit
control of the Lp -mean error of the fixed lag smoother when applied to additive functionals). With
S̃ti := E (at , bt )T (at , bt ) y1:t(`) , φi0:t(`) = Σ̃it + (ãit , b̃it )T (ãit , b̃it ) and Cti := (f (φit ), 1) ,
we obtain
Q̂(θ|θ
(m)
T
N
o
1 n 2
1 XX i
i i i T
i i
i T
2
ω̃ log 2π + log σε? + 2 yt − 2Ct (ãt , b̃t ) yt + Ct S̃t (Ct )
) = const −
2 t=1 i=1 t
σε?
T
N
1 XX i
−
ω̃t 2 log 2π + log |Q| + tr Q−1 Σ̃it + (ãit , b̃it )T − µ (ãit , b̃it ) − µT
2 t=1 i=1
− Q−1 A Σ̃it−1,t + (ãit−1 , b̃it−1 )T − µ (ãit , b̃it ) − µT
− Q−1 Σ̃it,t−1 + (ãit , b̃it )T − µ (ãit−1 , b̃it−1 ) − µT AT
+ Q−1 A Σ̃it−1 + (ãit−1 , b̃it−1 )T − µ (ãit−1 , b̃it−1 ) − µT AT
+
T X
N
X
i
ω̃ti log pα,β (φit , ψti |φit−1 , ψt−1
).
t=1 i=1
11
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Maximization with respect to σε2 , µ, A, and Q yields in the M-step the estimates
(σε2 )(m+1)
=
µ(m+1)
=
A(m+1)
=
T
N
o
1 XX i n 2
ω̃t yt − 2Cti (ãit , b̃it )T yt + Cti S̃ti (Cti )T ,
T t=1 i=1
T
N
1 XX i i i T
1
,
ω̃t (ãt , b̃t ) + Op
T t=1 i=1
T
! T N
!
T X
N
−1
X
XX
i
i
i
i
i
i
ω̃t Σ̃t,t−1 + D̃t,t−1
,
ω̃t Σ̃t−1 + D̃t−1
t=1 i=1
(m+1)
Q
=
1
T
(
T X
N
X
t=1 i=1
ω̃ti
Σ̃it
+
D̃ti
−A
(m+1)
t=1 i=1
T X
N
X
ω̃ti
Σ̃it−1,t
+
i
D̃t−1,t
)
,
t=1 i=1
with
D̃ti
=
i
D̃t,t−1
=
T
(ãit , b̃it )T − µ(m+1) (ãit , b̃it )T − µ(m+1) ,
T
(ãit , b̃it )T − µ(m+1) (ãit−1 , b̃it−1 )T − µ(m+1) .
In case where µ is assumed to be known we set µ(m+1) = µ (e.g. when f is a nonparametric function
µa and µb are usually set to 1 and 0 respectively). For α and β, numerical maximization is required
because no closed-form expression can be derived.
7. (Parametric) EM Step (see also 8. below)
• Compute (σε2 )(m+1) , µ(m+1) ,A(m+1) , and Q(m+1) .
• Obtain α(m+1) and β (m+1) by numerical maximization of Q̂t (α, β|α(m) , β (m) ) to .
3.4
Nonparametric Estimation of the Oscillation Pattern
If f is unknown we have to maximize in addition the second term in (15) with respect to f . The other
b i := (f (m+1) (φi ), 1)).
estimates remain unchanged including (σε2 )(m+1) where Cti must be replaced by C
t
t
For simplicity we ignore the other terms, i.e. we maximize
Q(f |f (m) ) ∝ const −
T
X
Ef (m) [{yt − at f (φt ) − bt }2 |y1:t(`) ]
(16)
t=1
with respect to f . As for nonparametric maximum likelihood estimation we need some regularization
in order to obtain a proper estimator. The basic idea for regularization now is to approximate the
densities p(at , bt , φt |y1:t ) in (16) instead of (14) by the kernel density
p(at , bt , φt |y1:t(`) ) ≈
N
X
ω̃ti N at , bt (ãit , b̃it )T , Σ̃it Kh φt − φit ,
(17)
i=1
where Kh (·) :=
1
·
hK h
with a bandwidth h and a kernel K(·). Without prior knowledge both are
equally good. The following proposition shows that this leads to an estimator for f which is also
12
Dahlhaus et al.
Statistical Inference for Oscillation Processes
based on kernel approximations. h might be chosen adaptively (say as in Lepski et.al., 1997) but the
situation is different here since N can be chosen arbitrarily large - leading to N T data points thus
allowing for choosing h arbitrarily small. In any case the optimal h depends on the unknown f and
the number of data meaning that h should be the same in each iteration step.
Proposition 3.1. Suppose that the density p(at , bt , φt |y1:t ) is as in (17). Then Q(f |f (m) ) is maximized by the estimate
f˜(m+1) (φ) =
PT
PN
ω̃ti Kh (φ − φit ) mod 2π yt ãit − (S̃ti )12
.
i
PT PN
i
i
t=1
i=1 ω̃t Kh (φ − φt ) mod 2π (S̃t )11
t=1
i=1
(18)
The analogue result holds when using, instead of the fixed lag smoother, the filter (where ãit and S̃ti
are replaced by ait and Sti respectively) or the complete smoother.
Proof. See the appendix.
Each step of the above nonparametric EM algorithm improves the likelihood in “nearly all cases”.
The latter restriction comes from the approximation (17) which needs to be “good enough”. More
precisely we have with Jensen’s inequality
log
T
T
X
pf (m+1) (y1:T ) X
≥
Ef (m) {yt −at f (m) (φt )−bt }2 |y1:T −
Ef (m) {yt −at f (m+1) (φt )−bt }2 |y1:T .
pf (m) (y1:T )
t=1
t=1
If f˜(m+1) would maximize (16) we had pf (m+1) (y1:T ) ≥ pf (m) (y1:T ). Since we have used the approximation (17) this is however not guaranteed in a strict sense.
Improving the speed of convergence of the MSEM - algorithm:
An example for the iteration steps of the nonparametric EM-estimate is given in Figure 6 (for details
see Section 4.3). If no prior information on the shape of f is used (when choosing fˆ(0) ) then convergence
may be slow (since a poor fˆ(0) will lead to poor particles and then again to a poor fˆ(1) ). To speed
up convergence we may invoke two kinds of additional information into the algorithm that speed up
convergence considerably. The idea is to use these corrections only during the first few steps:
1) The first correction uses that the empirical distribution of the points φt mod 2π converges to
an uniform distribution. This follows since the increments ∆φt are supposed to be stationary. By
using this property we can exclude time warping (see 4) in Section 2) and the correction consists of
transforming the phases accordingly with the edf of the φt mod 2π. Let F̂φ be a smoothed version of
this edf given by
Z
F̂φ (y) =
y
p̂φ (φ) dφ,
(19)
0
where
p̂φ (φ) =
T X
N
X
ω̃ti Khφ (φ − φit ) mod 2π
t=1 i=1
13
(20)
Dahlhaus et al.
Statistical Inference for Oscillation Processes
(note that it is computationally more efficient to use a frequency polygon instead of (20) since then
the distribution function and the inverse distribution can be easily calculated).
We then remove possible time warping by transforming f and the particles to
φ̌ mod 2π
fˇ(m+1) (φ̌) = fˆ(m+1) F̂φ−1
2π
and φ̌it = 2π F̂φ (φit mod 2π) + bφit /(2π)c .
(for simplicity we denote in step 2) the new fˇ(m+1) and φ̌it again by fˆ(m+1) and φit ).
2) The second correction uses the information that all 2π-periodic behavior of the signal is due to
the periodicity of f and not to any periodic behavior of the amplitude at and the baseline bt (this
follows from the independence of the process φt from at and bt ). Using this information means to
remove all 2π-periodic structures from the amplitude and baseline estimates and to transfer them to
the oscillation pattern f , i.e. to make the transformation
fˇ(m+1) (φ) = âper (φ) × fˆ(m+1) (φ) + b̂per (φ);
ǎit = ãit /âper (φit mod 2π);
Here
PT
where
PT
ā(φ) =
PN
ω̃ti Kha (φ − φit ) mod 2π ãit /ā(φit )
,
PT PN
i
i
t=1
i=1 ω̃t Kha (φ − φt ) mod 2π
t=1
âper (φ) =
b̌it = b̃it − b̂per (φit mod 2π).
i=1
PN
t=1
PT
t=1
i=1
ω̃ti I{−2π<φit −φ≤2π} ãit
PN
i=1
ω̃ti I{−2π<φit −φ≤2π}
,
φ ∈ (0, 2π],
φ ∈ (0, ∞),
smoothes the amplitude and the baseline by an average over 2 periods (the necessity of ā(φ) becomes
obvious when considering the example where ãit is linear). Similarly
PT PN
ω̃ti Khb (φ − φit ) mod 2π b̃it − b̄(φit )
b̂per (φ) = t=1 Pi=1
,
PN
T
i
i
t=1
i=1 ω̃t Khb (φ − φt ) mod 2π
with
PT
b̄(φ) =
PN
t=1
PT
t=1
i=1
PN
ω̃ti I{−2π<φit −φ≤2π} b̃it
i=1
ω̃ti I{−2π<φit −φ≤2π}
,
φ ∈ (0, 2π],
φ ∈ (0, ∞).
Performing first step 1 and then 2 has the disadvantage that all ãit have to be recalculated after
step 1 which is time consuming. We therefore have combined both steps by
φ̆ mod 2π
φ̆ mod 2π
φ̆ mod 2π
fˇ(m+1) (φ̆) = âper F̂φ−1
× fˆ(m+1) F̂φ−1
+ b̂per F̂φ−1
2π
2π
2π
(21)
8. Nonparametric EM Step
• Minimize the third and fourth term of (15) (with y1:T replaced by y1:t+l ) as described above
to get estimates µ(m+1) ,A(m+1) , Q(m+1) , α(m+1) and β (m+1) , and the second term of (15)
to get estimates f˜(m+1) (defined by (18)) and (σε2 )(m+1) .
14
Dahlhaus et al.
Statistical Inference for Oscillation Processes
• Perform the correction steps 1) and 2) by using (21). The resulting estimator fˆ(m+1) is used
in the next iteration. Iterate this until convergence. In the final step set fˆ(m+1) = f˜(m+1)
without using the correction.
• Keep the initial particles and weights φi0 , ãi0 , b̃i0 , and ω̃0i , i = 1, . . . , N , for the next iteration
(they have been updated due to the fixed lag smoother).
To start the iteration an initial guess fˆ(0) is required where prior information may be used. In
Section 4.3 on ECG recordings it turned out that the uninformative function fˆ(0) ≡ 0 may suffice.
However, one should choose the initial values α(0) and β (0) such that the theoretical phase increment
E[∆φt ] = α(0) /(1 − β (0) ) is close to the empirical one. This can be achieved by counting the number
of cycles in the data as in Section 4.3 (see also (11)).
4
Simulations and Data Examples
In this section we apply the proposed algorithms to simulated data including a noisy Rössler attractor
and to human electrocardiogram recordings.
4.1
Simulated Data
We generate observations yt , t = 1, . . . , 1000, from the GSSM defined through (13) and
yt = at cos(φt ) + bt + εt ,
with true at = 0.2 sin(2πt/1000) + 0.4 and bt = 0.4t/750 1t≤750 + (0.4 − 0.4(t − 750)/250) 1t>750 .
The ACD model parameters are set to α = 0.2 and β = 0.01. Two levels of observation noise are
investigated: σε2 = 0.01 and σε2 = 0.16. The parameters (α, β, σε2 , vec(Q)) are estimated with the
(parametric) EM algorithm and we set A = diag(1, 1). For both noise levels, the EM algorithm
obtained estimates (α̂, β̂, σ̂ε2 ) which were very close to the true values after a few iterations. For Q,
we obtained diag(10−4 , 5 × 10−5 ).
Figures 3 shows the true values and the estimated values for the high noise level (the low noise
level even looks better). The estimates of the amplitude, baseline, and (folded) phase are computed
by the RBPS with N = 500 particles and lag l = 100. In addition, the figures display a signal
reconstruction based on the estimates (bottom plot) consisting of the estimated denoised observations
ŷt = ât cos φ̂t + b̂t . For comparison, also the non-noisy observations yt − εt are given. In the low-noise
setting, the estimates are even more accurate.
4.2
The Noisy Rössler Attractor
We now consider the Rössler attractor with configuration
x˙1 = −x2 − x3 ,
x˙2 = x1 + .15x2 ,
15
x˙3 = .4 + x3 (x1 − 8.5).
Statistical Inference for Oscillation Processes
−1.0 0.0
1.0
Dahlhaus et al.
200
400
600
800
1000
0
200
400
600
800
1000
0
200
400
600
800
1000
0
200
400
600
800
1000
0
200
400
600
800
1000
−0.4
0.2 0.6
0.0
0.2
0.4
0.2
0.4
0.6
0
2
4
6
0
Figure 3: The estimation results of the RBPS for the simulated signal with N (0, 0.16) noise (from top to bottom):
The simulated noisy observations; the folded estimated phase (circles) and the folded true phase (solid circles); the
estimated amplitude (dotted line) and the true amplitude (solid line); the estimated baseline (dotted line) and the
true baseline (solid line); the simulated non-noisy signal (solid line) and the denoised signal obtained from the RBPS
estimates (dotted line).
16
Statistical Inference for Oscillation Processes
−30
0
20
Dahlhaus et al.
20
40
60
80
100
120
140
0
20
40
60
80
100
120
140
0
20
40
60
80
100
120
140
0
20
40
60
80
100
120
140
−15
0
10
0
2
4
6
0
2
4
6
0
Figure 4: Estimation results for the noisy Rössler attractor. The plots show (from top to bottom): x1 -component
of the Rössler attractor with additive i.i.d. N (0, 40) noise; the folded Hilbert phase (circles) compared with the true
folded phase (solid circles); the folded phase estimated with the RBPS (circles) compared with the true folded phase
(solid circles); the (non-noisy) x1 -component of the Rössler attractor compared with the reconstructed (denoised) signal
based on the amplitude and phase estimates of the RBPS.
17
Dahlhaus et al.
Statistical Inference for Oscillation Processes
The Rössler attractor and related systems are, for instance, used to model population dynamics
(Blasius et.al., 1999; Lloyd and May, 1999). We focus on the x1 component for which the (folded)
phase can be defined by means of arctan(x2,t /x1,t ) (cf. Pikovsky et al., 1997). It is assumed that x1,t
is not observed directly but through yt = x1,t + εt . A standard method for estimating the phase is to
apply the Hilbert transform (cf. Rosenblum et. al., 1996).
We now use the cosine model
yt = at cos(φt ) + εt
(22)
in combination with the RBPS and the EM algorithm for estimation (i.e bt ≡ 0).
We integrate the Rössler system with step size 0.1 using the Runge-Kutta method (Press et al.
1992, pp. 710-714) and we add i.i.d. Gaussian noise to the x1 -component. Again, two noise levels
have been considered: N (0, 4) and N (0, 40) (but we only display the high noise level - see Figure 4).
As parameter estimates we obtain (α̂, β̂)T = (0.2, 0.02)T , Q̂ = diag(0.9, 0) (the second value is set to
zero), and σ̂ε2 close to the true value. A was set to diag(1, 0). The RBPS is applied with N = 1000
particles and lag l = 200. For the computation of the Hilbert phase a running window of 100 data
points is used. The (folded) phase estimates of the Hilbert transform and our method together with
the true phase are presented in the second and third plot of Figures 4 (in the third plot the true
phases and the estimates are almost identical). It can be observed, that the phase estimate of the
RBPS is much closer to the true phase than the Hilbert phase. The bottom plot shows the (non-noisy)
x1 -component of the Rössler attractor along with the denoised signal ŷ = ât cos(φ̂t ), where ât and φ̂t
are obtained from the RBPS. Note, that even in the high noise case, the denoised signal is very close
to the true signal. The results in the low noise case are even better.
4.3
Application to Human Electrocardiogram Recordings
Human ECG recordings are characterized by a specific oscillation pattern f , amplitude changes, and
baseline shifts. The oscillation pattern depends on certain characteristics of the specific human being.
The baseline shifts are typically caused by respiration or body movements (Clifford et.al. 2006). The
estimation of the oscillation pattern f is for example of importance for the diagnosis of several heart
diseases - see Figure 3.6. and Table 3.1 in Clifford et.al. (2006). For inference we use the model
yt = at f (φt ) + bt + εt ,
i.e. the model (12)-(13) with µa = 1, µb = 0 and unknown oscillation pattern f . We use ECG
recordings obtained from the PhysioBank database1 . The data are sampled at a frequency of 0.01
seconds for a duration of 10 seconds (leading to 1000 observations) - see the top plot of Figure 5. The
RBPS and the nonparametric EM algorithm are applied to the data in order to obtain estimates for
φt , at , bt , and f . As initial oscillation pattern we use the trivial choice fˆ(0) ≡ 0. The only “prior”
1 http://www.physionet.org/physiobank/
18
Statistical Inference for Oscillation Processes
−0.5
0.5
1.5
Dahlhaus et al.
2
4
6
8
10
0
2
4
6
8
10
0
2
4
6
8
10
0
2
4
6
8
10
−0.02
0.00
0.02
0.90
1.00
1.10
0 1 2 3 4 5 6
0
Figure 5: Estimation results for the ECG recordings. The plots show (from top to bottom): the ECG data points;
the folded phase, the amplitude, and the baseline estimated by the RBPS.
information we use is that the dataset covers roughly 11 cycles leading an average increase of ∆φt of
about 2π/90. According to this we choose as initial values in the first iteration step β (0) = 0.1 and
α(0) = (1 − β (0) )2π/90 (see (11) and 3) in Section 2). The estimates for the amplitude, baseline, and
phase computed by the RBPS (applied with N = 100 particles and ` = 10) are given in Figure 5.
It can be seen that the amplitude changes significantly over time. In contrast, the baseline is almost
constant. The estimates of the oscillation pattern fˇ(m) for the iterations m = 1, . . . , 9 are shown in
Figure 6 (with h = 0.01 and the Epanechnikov kernel in (18)). Observe how rapidly the estimates
of the oscillation pattern converge. Finally, the estimated fˇ(9) is compared with one period of the
data in Figure 7 (note that this is one oscillation and not the true curve). In this application we have
chosen small N and l in order to demonstrate that this choice already leads to good results.
5
Concluding Remarks
In this paper we have proposed a general model for oscillation processes with a quasi-periodic component. The key ingredients are a nonparametric oscillation pattern f and the modeling of the
19
0.5
−0.5
−0.5
0.5
1.5
Statistical Inference for Oscillation Processes
1.5
Dahlhaus et al.
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0.5
−0.5
1.5
0.5
−0.5
1.5
0.5
−0.5
1.5
0.5
−0.5
−0.5
0.5
1.5
−0.5
0.5
1.5
−0.5
0.5
1.5
−0.5
0.5
1.5
1
1.5
0
Figure 6: The initial oscillation pattern fˆ(0) and the estimated oscillation patterns fˇ(m) for the iterations m = 1, . . . , 9
0.5 1.0 1.5
−0.5
−0.5
0.5 1.0 1.5
of the nonparametric EM algorithm.
2.6
2.8
3.0
3.2
3.4
0
1
2
3
4
5
6
Figure 7: Left plot: A fraction of the ECG recordings (note that this is one oscillation and not the true curve). Right
plot: Estimated oscillation pattern fˇ(9) obtained after nine iterations of the nonparametric EM algorithm.
20
Dahlhaus et al.
Statistical Inference for Oscillation Processes
unobserved phase process of the system by a nonlinear state space model. The situation is challenging since the model is a nonparametric regression model with unobserved regressors. We have
proven identifiability of the model for the specific phase model (2) which is the simplest possible phase
model. We have chosen this model for a basic understanding but also as a framework for first stage
mathematical investigations. The identifiability results of Section 2 and Section 6 have already been
quite challenging for this model and the asymptotic properties of the nonparametric EM-algorithm
introduced in Section 3.4 are even in this model still unknown. Proving consistency and asymptotic
normality of this nonparametric EM estimator is very challenging: using results similar to Olsson et
al. (2008) a first step would be to prove that the MCEM - estimate of this paper is a valid approximation of the usual exact nonparametric EM algorithm (following Fort and Moulines, 2003, or Le
Corff and Fort, 2013). Then, the asymptotic properties of the exact nonparametric EM algorithm had
to be established for the present GSSM - which remains an open problem in this case since the state
consists of an integrated process modulo 2π.
From an applied view we prefer the more general state equation (13) which guarantees positivity
of the phase increments and includes some dependence of consecutive phase increments. For this
more general model we have estimated the unobserved phases by a computationally efficient RaoBlackwellized Particle Smoother (RBPS) which allows for simultaneous estimation of the amplitude,
the baseline, and the phase. For the unknown oscillation pattern f we have derived a nonparametric
EM algorithm and applied it to ECG recordings. In another example the method has been used for
estimating the nonlinear phase of a noisy Rössler attractor. The examples confirm that the RBPS
provides in general a good estimate and maybe a worthwhile alternative to common methods for
estimating the instantaneous frequency such as the Hilbert transform - in particular in cases of large
observation noise. The good performance for noisy signals is due to the fact that the observation noise
is modeled explicitly.
In the ECG-application the algorithm needed about 30 minutes on a personal computer, i.e. 3
minutes per EM-step. We are optimistic that with massive parallelization over the particles and/or
reduction of the lag of the fixed lag smoother the RBPS can be performed online. The EM algorithm
may also be used pseudo-online by always switching to new parameter values after a fixed number of
time steps. A better option seems to develop a stochastic approximation EM algorithm as in Delyon
et.al. (1999) for the present model.
6
Appendix 1: Identifiability
In this appendix we discuss identifiability of the model (1),(2) in detail and prove Theorem 2.1 and
2.2 from Section 2. We first argue that we can assume without loss of generality in all proofs that the
sequence {κi (f )}i∈N+ from (7) is setwise coprime.
21
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Proposition 6.1. Suppose that {κi (f )}i∈N+ is not setwise coprime. If d is the greatest common
divisor define f¯(x) := f xd and φ̄t := dφt . Then f¯(·) is 2π-periodic with ck (f¯) = ckd (f ) and κi (f¯) =
κi (f )/d, meaning that {κi (f¯)}i∈N+ is setwise coprime. Furthermore, ω̄ = d ω and σ̄η = d ση .
Proof. Since d is a common divisor of {κi (f )}i∈N+ we have
f (x) =
X
ck (f ) eikx =
k∈Z
X
ckd (f ) eikdx
k∈Z
and therefore
X
x
¯
=
ckd (f ) eikx ,
f (x) = f
d
k∈Z
i.e. f¯(·) is 2π-periodic with ck (f¯) = ckd (f ), dκi (f¯) = κi (f ). The rest is obvious.
We now define the autocovariance function of Yk by
ω,σ 2
2
Γf,σ2η (`) := E [Y` Y0 ] − E [Y0 ] .
ε
Lemma 6.2. We have
ω,σ 2
Γf,σ2η (`)
ε
(
=
P∞
2
|ck (f )| cos(k`ω) e−`k
P+∞
σε2 + 2 k=1 |ck (f )|2 ,
2
k=1
2
ση2 /2
Proof. Let g` (x) be the probability density function of φ` − φ0 =
,
` ∈ N+
(23)
` = 0.
P`
k=1
ηk + ω` ∼ N (ω`, ση2 `). Since
f is real, we have with ck := ck (f ) for all k ≥ 1 c−k = ck and therefore for ` ∈ N+
ω,σ 2
Γf,σ2η (`)
ε
∞
X
2
= E [f (φ` )f (φ0 )] − E [f (φ0 )] =
ck1 ck2 E eik1 φ` eik2 φ0 − |c0 |2
k1 ,k2 =−∞
=
=
1
2π
∞
X
Z
eik1 x eik2 (x+y) g` (y) dx dy − |c0 |2
ck1 ck2
[0,2π]×R
k1 ,k2 =−∞
∞
X
2 2
2
|ck | eik`ω e−`k ση /2 −
|c0 |2 = 2
k=−∞
∞
X
2
|ck | cos(k`ω) e−`k
2
ση2 /2
.
k=1
ω,σ 2
by using the characteristic function of the normal distribution. The derivation for Γf,σ2η (0) is straightε
forward (use e.g. the Parseval equality).
We now prove Theorem 2.1 and at the same time the second relation from (9).
Proof of Theorem 2.1.
We set c?k := ck (f? ), κ?i := κi (f? ), and κi := κi (f ).
2
1) Identifiability of ση?
and definition of γ:
ω,σ 2
ω? ,σ 2
Γf,σ2η (`) = Γf? ,σ2η? (`) for all ` ≥ 0 implies
ε
ε?
∞
X
2
|ck (f )| cos(k`ω) e−`k
2
σ 2 /2
k=1
=
∞
X
k=1
22
2
|c?k | cos(k`ω? ) e−`k
2
2
ση?
/2
.
Dahlhaus et al.
Statistical Inference for Oscillation Processes
By definition of κ1 and κ?1 , for all ` ≥ 1
∞
X
2
|cκ1 (f )| cos(κ1 `ω) +
2
|ck (f )| cos(k`ω) e−`(k
2
−κ21 )ση2 /2
k=κ1 +1
2
2 2
−`/2(κ?2
1 ση? −κ1 ση )
=e
∞
X
|c?κ?1 |2 cos(κ?1 `ω? ) +
2
|c?k | cos(k`ω? ) e−`(k
2
2
−κ?2
1 )ση? /2
. (24)
k=κ?
1 +1
Whatever ω, lim sup`→+∞ cos(κ1 ω`) = 1 if κ1 ω ∈ πQ, the set {cos(κ1 ω`); ` ≥ 1} is finite and
cos(κ1 ω`) equals one for infinitely many `; if κ1 ω ∈
/ πQ, κ1 ωZ + 2πZ is a dense subset of R and
[−1, 1] is the set of limit points of {cos(κ1 ω`); ` ≥ 1} . Therefore, the lim sup of the left hand side
of (24) is |cκ1 (f )| while the one of the right hand side is 0 if κ?1 ση? > κ1 ση or +∞ if κ?1 ση? < κ1 ση .
2
?
Therefore, κ21 ση2 = κ?,2
| = |cκ1 (f )|. Define
1 ση? and |cκ?
1
γ :=
ση?
κ1
=
.
κ?1
ση
(25)
2) Identifiability of ω? and {|c?k |}k≥0 :
We obtain from (24) with γση = ση?
∞
X
|cκ1 (f )|2 cos(κ1 `ω) +
2
|ck (f )| cos(k`ω) e−`(k
2
−κ21 )ση2 /2
k=κ1 +1
= |c?κ?1 |2 cos(κ?1 `ω? ) +
∞
X
2
|c?k | cos(k`ω? ) e−`(k
2
2
−κ?,2
1 )ση? /2
. (26)
k=κ?
1 +1
Then, since |c?κ?1 | = |cκ1 (f )|,
2 sin
=
γω + ω
2
γω − ω
? ?
κ?1 ` sin
κ1 ` = cos(κ?1 `ω? ) − cos(κ1 `ω)
2
(27)
∞
∞
2
2
X
X
2
?2
2
|ck (f )|
|c?k |
−`(k2 −κ21 )ση2 /2
cos(k`ω)
e
cos(k`ω? ) e−`(k −κ1 )ση? /2 .
−
?
?
2
2
|cκ? |
|cκ? |
?
k=κ1 +1
If
?
γω+ω? ?
κ1
2
k=κ1 +1
1
γω−ω? ?
κ1
2
∈
/ πZ and
(28)
1
∈
/ πZ, (27) does not converge to 0 as ` → ∞ while (28) does.
Therefore there exists k1 ∈ Z such that either γω + ω? =
2π
κ?
1
k1 or γω − ω? =
2π
κ?
1
k1 . This implies
cos(κ1 `ω) = cos(κ?1 `ω? ) and, combined with |cκ1 (f )| = |c?κ?1 | and (26):
∞
X
2
−`(k2 −κ21 )ση2 /2
|ck (f )| cos(k`ω) e
=
∞
X
2
|c?k | cos(k`ω? ) e−`(k
2
2
−κ?2
1 )ση? /2
.
k=κ?
1 +1
k=κ1 +1
Note, that κ?i is the ith non-zero coefficient of c?k . Iterating these steps recursively proves for all
i ∈ N+ κi = γκ?i and
∞
X
2
|ck (f )| cos(k`ω) e−`(k
2
−κ2i )ση2 /2
=
∞
X
2
|c?k | cos(k`ω? ) e−`(k
2
2
−κ?2
i )ση? /2
.
k=κ?
i +1
k=κi +1
This implies |cκi (f )| = |c?κ?i | for all i ∈ N+ (all other coefficients are 0) and E [Yk ] = c0 (f ) = c?0 .
23
Dahlhaus et al.
Statistical Inference for Oscillation Processes
We now show that γ = 1 if each of the sequences {κi }i∈N+ and {κ?i }i∈N+ is setwise coprime (see
Proposition 6.1). Then there exist p ≥ 1 and (a1 , . . . , ap )0 ∈ Zp such that
p
X
ai κ?i = 1 .
(29)
i=1
Therefore,
γ=γ
p
X
ai κ?i =
p
X
i=1
ai κi ∈ Z .
i=1
Since for all i ≥ 1 γκ?i = κi , γ is a divisor of all the κi ’s implying γ = 1.
We also obtain from above the existence of a sequence `i
∈
Z such that either
(ω − ω? )κ?i = 2π `i (*) or (ω + ω? )κ?i = 2π `i (**). Suppose first that (*) holds for some index
`
i and (**) holds for some index j 6= i. Then ω? = π κj? − κ`i? which is in contradiction to ω? ∈
/ πQ.
j
i
This means that either (*) or (**) must hold simultaneously for all i.
Suppose now that (*) holds for all i. Then (29) implies
ω − ω? = (ω − ω? )
p
X
ai κ?i = 2π
i=1
p
X
ai `i ∈ 2πZ.
i=1
The assumption ω, ω? ∈ (0, π) then implies ω = ω? .
Finally, suppose that (**) holds for all i. This leads in the same way to ω + ω? ∈ 2πZ which is in
contradiction with ω, ω? ∈ (0, π). Thus, we finally obtain ω = ω? .
2
3) Identifiability of σε?
:
2
then is a direct consequence of (23).
The identifiability of σε?
4) Identifiability of f? :
We need to show that ck (f ) = c?k for all k ∈ N0 . We know already know from 2) that c0 (f ) = c?0
and |ck (f )| = |c?k | for all k ∈ N+ . The identification of the ck (f ) for all k ≥ 0 can be obtained
using the higher order moments: for all p ≥ 1 and all (`1 , . . . , `p ) ∈ Np+ , we define
ω? ,σ 2
Ψf? ,p η? (`1 , . . . , `p ) := E f? (φ0 )f? (φ`1 ) . . . f? (φ`1 +···+`p ) .
(30)
As the finite dimensional distributions of {Yk }k∈N+ are the same and the {εk }k∈N+ are independent
centered Gaussian random variables (implying that with σε2 also all moments Eε`t are identifiable),
2
ω? ,σ 2
η?
we know that for all p ≥ 1 and all `p ∈ Np+ , Ψω,σ
f,p (`p ) = Ψf? ,p (`p ). Let further
!
2
2
ση?
ω? ,ση?
2
ψ`
(k) := exp ik`ω? −
`k
2
(31)
2
be the characteristic function of a Gaussian random variable with mean `ω? and variance `ση?
.
Write, for all −∞ < k1 , . . . , kp < ∞,
d?k1 ,...,kp := c?−(k1 +···+kp ) c?k1 . . . c?kp
dk1 ,...,kp := c−(k1 +···+kp ) (f ) ck1 (f ) . . . ckp (f ) .
24
Dahlhaus et al.
Statistical Inference for Oscillation Processes
ω? ,σ 2
2
We know from 2) that σ 2 = ση?
and ω = ω? . Then for all p ≥ 1 and all `p ∈ Np+ , Ψf? ,p η? (`p ) =
2
ω? ,ση?
Ψf,p
(`p ) implies because of Lemma 6.3
2
ω? ,ση?
X
ψ `1
2
ω? ,ση?
(k1 + · · · + kp ) . . . ψ`p
(kp ) d?k1 ,...,kp − dk1 ,...,kp = 0 .
−∞<k1 ,...,kp <∞
Notice that, since {c?k }k≥0 and {ck }k≥0 belong to `2 (Z), the coefficients d?k1 ,...,kp −dk1 ,...,kp , k1 , . . . , kp ∈
Z are bounded. Therefore, by Lemma 6.5, for all −∞ < k1 , . . . , kp < ∞, d?k1 ,...,kp = dk1 ,...,kp . Due
to (29) we can decompose any k ∈ Z as k = ka1 κ1 + · · · + kap κp . Thus d?κ1 ,...,κp = dκ1 ,...,κp yields
c?−k (c?κ1 )a1 k . . . (c?κp )ap k = c−k (f )(cκ1 (f ))a1 k . . . (cκp (f ))ap k .
Therefore,
"
c−k (f ) =
c?−k
#k
(c?κ1 )a1 . . . (c?κp )ap
(cκ1 (f ))a1 . . . (cκp (f ))ap
.
As, for all 1 ≤ i ≤ d |cκi (f )| = |c?κi | =
6 0, there exists θ ∈ [0, 2π) such that
c−k (f ) = c?−k e−ikθ .
which completes the proof.
Proof of Theorem 2.2.
Let f? be a non-constant oscillation pattern and
ν(f? ) = max j ∈ N+ ck (f? ) = 0 ∀ k 6= jN+ .
(32)
Note that ν(f? ) < ∞ (otherwise f? were constant). We now prove that ν(f? ) = repl(f? ). Since
ck (f? ) = 0 apart from k = ` ν(f? ), with some ` ∈ Z, we obtain
X
X
x
x
¯
=
c` ν(f? ) (f? ) ei(` ν(f? )) ν(f? ) =
c` ν(f? ) (f? ) ei` x ,
f? (x) := f?
ν(f? )
`∈Z
`∈Z
implying that f¯? is 2π-periodic and repl(f? ) ≥ ν(f? ). The sequence {κi (f¯? )}i∈N+ must be setwise
coprime (if the sequence had a common factor m also m ν(f? ) would fulfill the above requirement and
ν(f? ) were not the maximum).
Suppose now that f is another oscillation pattern and ν(f ) and f¯ are defined as above. Then it
follows from 2) above that there exists a γ with κi (f¯) = γκi (f¯? ) for all i ∈ N+ . Since both sequences
{κi (f¯? )}i∈N+ and {κi (f¯)}i∈N+ are setwise coprime it follows as in 2) that γ = 1. As in 4) we therefore
obtain f¯(x − θ) = f¯? (x) with some θ ∈ (0, 2π) and f (x) = f¯? ν(f ) x + θ . In particular we have
ν(f? ) = repl(f? ) and the basic cycle is unique.
x−θ
Since f? ν(fx? ) = f¯? (x) = f¯(x − θ) = f ν(f
) we obtain from Theorem 2.1
γ=
ν(f )
repl(f )
=
.
ν(f? )
repl(f? )
25
Dahlhaus et al.
6.1
Statistical Inference for Oscillation Processes
Technical Lemmata
The proof of the following results are not difficult and are therefore omitted.
ω? ,σ 2
2
ω? ,ση?
Lemma 6.3. Let Ψf? ,p η? (`1 , . . . , `p ) and ψ`
be as in (30) and (31) respectively. Then for all
(`1 , . . . , `p ) ∈ Np+ ,
ω? ,σ 2
Ψf? ,p η? (`1 , . . . , `p ) =
2
ω? ,ση?
X
c?−(k1 +···+kp ) c?k1 . . . c?kp ψ`1
2
ω? ,ση?
(k1 + · · · + kp ) . . . ψ`p
(kp ) .
(33)
−∞<k1 ,...,kp <∞
∞
Lemma 6.4. Let {zk }∞
k=−∞ be complex numbers such that {zk }k=−∞ ∈ `∞ (Z). Then, for all ` ≥ 1,
2
ω? ,ση?
{ψ`
2
ω? ,ση?
(k)zk }k∈Z ∈ `1 (Z), where ψ`
is defined in (31). Assume that ω? ∈
/ πQ. If for all ` ≥ 1,
X
2
ω? ,ση?
ψ`
(k)zk = 0 ,
−∞<k<∞
then we have zk = 0 for all k ∈ Z.
The proof of Theorem 2.1 4) relies on the following lemma which is a corollary of Lemma 6.4.
Lemma 6.5. Assume that ω? ∈
/ πQ. Let Z ∈ `∞ (Zp ) be a bounded complex sequence indexed by
(k1 , . . . , kp ) ∈ Zp , p ≥ 1, satisfying, for any `1 , · · · , `p ≥ 1,
2
ω? ,ση?
X
ψ `1
2
ω? ,ση?
(k1 + · · · + kp ) . . . ψ`p
(kp )zk1 ,...,kp = 0 .
−∞<k1 ,...,kp <∞
Then we have zk1 ,··· ,kp = 0 for any k1 , · · · , kp .
7
Appendix 2: The nonparametric EM-estimate
Proof of Proposition 3.1. Under the assumption that f is 2π periodic, it can be seen from (16)
PT R
2
that t=1
yt − at f (φt mod 2π) − bt pf (m) (at , bt , φt |y1:t+` ) dat dbt dφt needs to be minimized with
respect to f (φ). By using the density p(at , bt , φt |y1:t+` ) from (17) the above expression becomes
T X
N
X
ω̃ti
2πZ
Z
Z
2
yt − at f (φ) − bt N at , bt (ãit , b̃it )T , Σ̃it Kh (φ − φit ) mod 2π dat dbt dφ
0
t=1 i=1
Z
=
0
T X
N
2π X
ω̃ti yt2 + (S̃ti )11 f (φ)2 + (S̃ti )22 − 2yt ãit f (φ) − 2yt b̃it + 2(S̃ti )12 f (φ)
t=1 i=1
× Kh (φ − φit ) mod 2π dφ.
For fixed φ minimization with respect to f (φ) now yields
f˜(m+1) (φ) =
PT
PN
ω̃ti Kh (φ − φit ) mod 2π yt ãit − (S̃ti )12
.
i
PT PN
i
i
t=1
i=1 ω̃t Kh (φ − φt ) mod 2π (S̃t )11
t=1
i=1
i.e. the result. For the filter and the smoother the proofs are the same.
26
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Acknowledgements: This research has been supported by the Deutsche Forschungsgemeinschaft
under DA 187/15-1, through the Research Training Group RTG 1953 and by Heidelberg University
under Frontier D.801000/08.023. It has been conducted as part of the project Labex MME-DII
(ANR11-LBX-0023-01). The authors are very grateful to two anonymous referees for their helpful
comments.
References
Amblard, P. O., Brossier, J. M., and Moisan, E. (2003). Phase tracking: what do we gain from
optimality? Particle filtering versus phase-locked loops. Signal Processing 83, 151-167.
Blasius, B., Huppert, A., and Stone, L. (1999) Complex dynamics and phase synchronization in
spatially extended ecological systems. Nature 399, 354-359.
Chen, R.-R., Koetter, R., Madhow, U., and Agrawal, D. (2003) Joint noncoherent demodulation and
decoding for the block fading channel: A practical framework for approaching Shannon capacity.
In IEEE Transactions on Communications 51, 1676-1689.
Clifford, G.D., Azuaje, F., and McSharry, P.E. (2006) Advanced Methods and Tools for ECG Data
Analysis. Norwood: Artech House.
Delyon, B., Lavielle, M., and Eric Moulines, E. (1999) Convergence of a stochastic approximation
version of the EM algorithm. Annals of Statistics 27, 94–128.
Dempster, A.P., Laird, N.M., and Rubin, D.B. (1977) Maximum Likelihood from Incomplete Data
via the EM Algorithm. Journal of the Royal Statistical Society, Series B 39, 1-38.
Douc, R., Cappé, O., and Moulines, E. (2005) Comparison of resampling schemes for particle filtering.
In Proceedings of the 4th International Symposium on Image and Signal Processing and Analysis,
64-69.
Douc, R., Moulines, E. and Stoffer, D.S. (2014) Nonlinear time series: Theory, Methods and Applications with R Examples. CRC Press.
Doucet, A., Gordon, N.J., and Krishnamurthy, V. (1999) Particle Filters for State Estimation of
Jump Markov Linear Systems. IEEE Transactions on Signal Processing 49, 613-624.
Doucet, A., Godsill, S. and Andrieu, C. (2000a) On sequential Monte Carlo sampling methods for
Bayesian filtering. Stat. Comput. 10, 197-208.
Doucet, A., De Freitas, N., Murphy, K. and Russell, S. (2000b) Rao-Blackwellised particle filtering
for dynamic Bayesian networks. In Proceedings of the Sixteenth conference on uncertainty in
artificial intelligence, 176-183.
Doucet, A., De Freitas, N., and Gordon, N. (ed.) (2001) Sequential Monte Carlo Methods in Practice,
Springer Verlag, New York.
Dubois, C. and Davy, M. (2007) Joint detection and tracking of time-varying harmonic components:
a flexible Bayesian approach. IEEE Trans. Audio, Speech, Language Process. 15(4), 1283Ű1295.
Dumont, T. and Le Corff, S. (2014) Nonparametric regression on hidden Φ-mixing variables: identifiability and consistency of a pseudo-likelihood based estimation procedure. arXiv:1209.0633.
Engle, R.F., and Russell, J.R. (1998) Autoregressive conditional duration: A new model for irregularly spaced transaction data. Econometrica 66, 1127-1162.
Fell, J., and Axmacher, N. (2011) The role of phase synchronization in memory processes. Nat. Rev.
Neurosci. 12(2), 105-118.
27
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Fort, G., and Moulines, E. (2003) Convergence of the Monte Carlo expectation maximization for
curved exponential families. Annals of Statistics 31 (4), 1220-1259.
Godsill, S.J., Doucet, A., and West, M. (2004) Monte Carlo Smoothing for Nonlinear Time Series.
Journal of American Statistical Association 99, 156-168.
Gordon, N. J., Salmond, D. J., and Smith, A. F. (1993) Novel approach to nonlinear/non-Gaussian
Bayesian state estimation. IEE Proceedings F-Radar and Signal Processing 140 (2), 107-113.
Grossmann, A. , Kronland-Martinet, R., and Morlet, J. (1989) Reading and Understanding Continuous Wavelet Transforms. Wavelets, Time-Frequency Methods and Phase Space, ed. J.M.
Combes, Springer Verlag, Berlin.
Hannan, E.J. (1973) The Estimation of Frequency. Journal of Applied Probability 10, 513-519.
Huang, N.E., Shen, Z., Long, S.R., Wu, M.L., Shih, H.H., Zheng, Q., Yen, N.C., Tung, C.C., and
Liu, H.H. (1998) The empirical mode decomposition and Hilbert spectrum for nonlinear and
nonstationary time series analysis. Proc. Roy. Soc. London A 454, 903-995.
Hürzeler, M., and Künsch, H. (2001) Approximating and maximising the Likelihood for a General
State-Space Model. Sequential Monte Carlo Methods in Practice, eds. A. Doucet, N. de Freitas
and N. Gordon, New York: Springer.
Kitagawa, G. (1996) Monte Carlo filter and smoother for non-Gaussian nonlinear state space models.
Journal of Computational and Graphical Statistics 5(1), 1-25.
Kneip, A. and Gasser, Th. (1992) Statistical Tools to Analyze Data Representing a Sample of Curves.
The Annals of Statistics 20, 1266-1305.
Kong, A., Liu, J., and Wong, W. (1994) Sequential imputation and Bayesian missing data problems.
Journal of American Statistical Association 89, 278-288.
Le Corff, S., and Fort, G. (2013) Online Expectation Maximization based algorithms for inference in
hidden Markov models. Electronic Journal of Statistics 7, 763-792.
Lepski, O.V., Mammen, E., and Spokoiny, V.G. (1997) Optimal spatial adaptation to inhomogeneous
smoothness: an approach based on kernel estimates with variable bandwidth selectors. The
Annals of Statistics 25, 929-947.
Li, T.H. (2013) Time Series with Mixed Spectra, CRC Press.
Lloyd, A.L., and May, R.M. (1999) Synchronicity, chaos and population cycles: spatial coherence in
an uncertain world. Trends in Ecology & Evolution 14, 417-418.
Myers, C.S. and L. R. Rabiner, L.R. (1981) A comparative study of several dynamic time-warping
algorithms for connected word recognition. The Bell System Technical Journal 60, 1389-1409.
Olsson, J., Cappe, O., Douc, R. and Moulines, E. (2008) Sequential Monte Carlo smoothing with
application to parameter estimation in nonlinear state space models. Bernoulli 14, 155-179.
Paraschakis, K. and Dahlhaus, R. (2012) Frequency and phase estimation in time series with quasi
periodic components. Journal of Time Series Analysis 33, 13-31 .
Pikovsky, A.S., Rosenblum, M.G., Osipov, G.V., and Kurths, J. (1997) Phase synchronization of
chaotic oscillators by external driving. Physica D 104, 219-238.
Pikovsky, A., Rosenblum, M., and Kurths, J. (2001), Synchronization. Cambridge University Press.
Press, W.H., Teukolsky, S.A., Vetterling, W.T., and Flannery, B.P. (1992), Numerical Recipes in C
(2nd ed.). Cambridge University Press.
Quinn, B.G. and Hannan, E.J. (2001) The Estimation and Tracking of Frequency. Cambridge University Press.
Rosenblum, M.G., Pikovsky, A.S., and Kurths, J. (1996) Phase Synchronization of Chaotic Oscillators. Physical Review Letters 76, 1804-1807.
28
Dahlhaus et al.
Statistical Inference for Oscillation Processes
Schuster, A. (1897) On lunar and solar periodicities of earthquakes. Proc. Roy. Soc. 61, 455-465.
Shumway, R.H. and Stoffer, D.S. (1982) An approach to time series smoothing and forecasting using
the EM algorithm. J. Time Series Anal. 3, 253–264.
Shumway, R.H. and Stoffer, D.S. (2011) Time Series Analysis and Its Applications: With R Examples
(3rd ed). Springer Verlag, New York.
Tanner, M.A. (1993) Tools for Statistical Inference: Methods for the Exploration of Posterior Distributions and Likelihood Functions (2nd ed.). Springer Series in Statistics, New York.
Tsakonas, E., Sidiropoulos, N. and Swami, A. (2008) Optimal particle filters for tracking a timevarying harmonic or chirp signal. IEEE Trans. Signal Process. 56(10), 4598-4610.
Wang, K. and Gasser, Th. (1997) Alignment of curves by dynamic time warping. Annals of Statistics,
25, 1251-1276.
Wei, G. C. G. and Tanner, M. A. (1990) A Monte Carlo implementation of the EM algorithm and
the poor man’s data augmentation algorithms. Journal of the American Statistical Association
85, 699-704.
29
| 10math.ST
|
Published as a conference paper at ICLR 2015
F IT N ETS : H INTS FOR T HIN D EEP N ETS
arXiv:1412.6550v4 [cs.LG] 27 Mar 2015
Adriana Romero1, Nicolas Ballas2 , Samira Ebrahimi Kahou3 , Antoine Chassang2 ,
Carlo Gatta4 & Yoshua Bengio2 †
1
Universitat de Barcelona, Barcelona, Spain.
2
Université de Montréal, Montréal, Québec, Canada. † CIFAR Senior Fellow.
3
École Polytechnique de Montréal, Montréal, Québec, Canada.
4
Centre de Visió per Computador, Bellaterra, Spain.
A BSTRACT
While depth tends to improve network performances, it also makes gradient-based
training more difficult since deeper networks tend to be more non-linear. The recently proposed knowledge distillation approach is aimed at obtaining small and
fast-to-execute models, and it has shown that a student network could imitate the
soft output of a larger teacher network or ensemble of networks. In this paper,
we extend this idea to allow the training of a student that is deeper and thinner
than the teacher, using not only the outputs but also the intermediate representations learned by the teacher as hints to improve the training process and final
performance of the student. Because the student intermediate hidden layer will
generally be smaller than the teacher’s intermediate hidden layer, additional parameters are introduced to map the student hidden layer to the prediction of the
teacher hidden layer. This allows one to train deeper students that can generalize
better or run faster, a trade-off that is controlled by the chosen student capacity.
For example, on CIFAR-10, a deep student network with almost 10.4 times less
parameters outperforms a larger, state-of-the-art teacher network.
1
I NTRODUCTION
Deep networks have recently exhibited state-of-the-art performance in computer vision tasks such
as image classification and object detection (Simonyan & Zisserman, 2014; Szegedy et al., 2014).
However, top-performing systems usually involve very wide and deep networks, with numerous
parameters. Once learned, a major drawback of such wide and deep models is that they result
in very time consuming systems at inference time, since they need to perform a huge number of
multiplications. Moreover, having large amounts of parameters makes the models high memory
demanding. For these reasons, wide and deep top-performing networks are not well suited for
applications with memory or time limitations.
There have been several attempts in the literature to tackle the problem of model compression to
reduce the computational burden at inference time. In Bucila et al. (2006), authors propose to train
a neural network to mimic the output of a complex and large ensemble. The method uses the ensemble to label unlabeled data and trains the neural network with the data labeled by the ensemble, thus
mimicking the function learned by the ensemble and achieving similar accuracy. The idea has been
recently adopted in Ba & Caruana (2014) to compress deep and wide networks into shallower but
even wider ones, where the compressed model mimics the function learned by the complex model, in
this case, by using data labeled by a deep (or an ensemble of deep) networks. More recently, Knowledge Distillation (KD) (Hinton & Dean, 2014) was introduced as a model compression framework,
which eases the training of deep networks by following a student-teacher paradigm, in which the
student is penalized according to a softened version of the teacher’s output. The framework compresses an ensemble of deep networks (teacher) into a student network of similar depth. To do so,
the student is trained to predict the output of the teacher, as well as the true classification labels. All
previous works related to Convolutional Neural Networks focus on compressing a teacher network
or an ensemble of networks into either networks of similar width and depth or into shallower and
wider ones; not taking advantage of depth.
1
Published as a conference paper at ICLR 2015
Depth is a fundamental aspect of representation learning, since it encourages the re-use of features,
and leads to more abstract and invariant representations at higher layers (Bengio et al., 2013). The
importance of depth has been verified (1) theoretically: deep representations are exponentially more
expressive than shallow ones for some families of functions (Montufar et al., 2014); and (2) empirically: the two top-performers of ImageNet use deep convolutional networks with 19 and 22 layers,
respectively (Simonyan & Zisserman, 2014) and (Szegedy et al., 2014).
Nevertheless, training deep architectures has proven to be challenging (Larochelle et al., 2007;
Erhan et al., 2009), since they are composed of successive non-linearities and, thus result in highly
non-convex and non-linear functions. Significant effort has been devoted to alleviate this optimization problem. On the one hand, pre-training strategies, whether unsupervised (Hinton et al., 2006;
Bengio et al., 2007) or supervised (Bengio et al., 2007) train the network parameters in a greedy layerwise fashion in order to initialize the network parameters in a potentially good basin of attraction.
The layers are trained one after the other according to an intermediate target. Similarly, semisupervised embedding (Weston et al., 2008) provides guidance to an intermediate layer to help learn
very deep networks. Along this line of reasoning, (Cho et al., 2012) ease the optimization problem
of DBM by borrowing the activations of another model every second layer in a purely unsupervised
scenario. More recently, (Chen-Yu et al., 2014; Szegedy et al., 2014; Gulcehre & Bengio, 2013)
showed that adding supervision to intermediate layers of deep architectures assists the training of
deep networks. Supervision is introduced by stacking a supervised MLP with a softmax layer on top
of intermediate hidden layers to ensure their discriminability w.r.t. labels. Alternatively, Curriculum
Learning strategies (CL) (Bengio, 2009) tackle the optimization problem by modifying the training
distribution, such that the learner network gradually receives examples of increasing and appropriate
difficulty w.r.t. the already learned concepts. As a result, curriculum learning acts like a continuation
method, speeds up the convergence of the training process and finds potentially better local minima
of highly non-convex cost functions.
In this paper, we aim to address the network compression problem by taking advantage of depth.
We propose a novel approach to train thin and deep networks, called FitNets, to compress wide and
shallower (but still deep) networks. The method is rooted in the recently proposed Knowledge Distillation (KD) (Hinton & Dean, 2014) and extends the idea to allow for thinner and deeper student
models. We introduce intermediate-level hints from the teacher hidden layers to guide the training
process of the student, i.e., we want the student network (FitNet) to learn an intermediate representation that is predictive of the intermediate representations of the teacher network. Hints allow
the training of thinner and deeper networks. Results confirm that having deeper models allow us to
generalize better, whereas making these models thin help us reduce the computational burden significantly. We validate the proposed method on MNIST, CIFAR-10, CIFAR-100, SVHN and AFLW
benchmark datasets and provide evidence that our method matches or outperforms the teacher’s
performance, while requiring notably fewer parameters and multiplications.
2
M ETHOD
In this section, we detail the proposed student-teacher framework to train FitNets from shallower
and wider nets. First, we review the recently proposed KD. Second, we highlight the proposed hints
algorithm to guide the FitNet throughout the training process. Finally, we describe how the FitNet
is trained in a stage-wise fashion.
2.1
R EVIEW
OF
K NOWLEDGE D ISTILLATION
In order to obtain a faster inference, we explore the recently proposed compression framework
(Hinton & Dean, 2014), which trains a student network, from the softened output of an ensemble of
wider networks, teacher network. The idea is to allow the student network to capture not only the
information provided by the true labels, but also the finer structure learned by the teacher network.
The framework can be summarized as follows.
Let T be a teacher network with an output softmax PT = softmax(aT ) where aT is the vector of
teacher pre-softmax activations, for some example. In the case where the teacher model is a single
network, aT represents the weighted sums of the output layer, whereas if the teacher model is the
result of an ensemble either PT or aT are obtained by averaging outputs from different networks
(respectively for arithmetic or geometric averaging). Let S be a student network with parameters
2
Published as a conference paper at ICLR 2015
WS and output probability PS = softmax(aS ), where aS is the student’s pre-softmax output. The
student network will be trained such that its output PS is similar to the teacher’s output PT , as well
as to the true labels ytrue . Since PT might be very close to the one hot code representation of the
sample’s true label, a relaxation τ > 1 is introduced to soften the signal arising from the output
of the teacher network, and thus, provide more information during training1. The same relaxation
is applied to the output of the student network (PτS ), when it is compared to the teacher’s softened
output (PτT ):
a
a
S
T
, PτS = softmax
.
(1)
PτT = softmax
τ
τ
The student network is then trained to optimize the following loss function:
LKD (WS ) = H(ytrue , PS ) + λH(PτT , PτS ),
(2)
where H refers to the cross-entropy and λ is a tunable parameter to balance both cross-entropies.
Note that the first term in Eq. (2) corresponds to the traditional cross-entropy between the output of
a (student) network and labels, whereas the second term enforces the student network to learn from
the softened output of the teacher network.
To the best of our knowledge, KD is designed such that student networks mimic teacher architectures
of similar depth. Although we found the KD framework to achieve encouraging results even when
student networks have slightly deeper architectures, as we increase the depth of the student network,
KD training still suffers from the difficulty of optimizing deep nets (see Section 4.1).
2.2
H INT- BASED T RAINING
In order to help the training of deep FitNets (deeper than their teacher), we introduce hints from the
teacher network. A hint is defined as the output of a teacher’s hidden layer responsible for guiding
the student’s learning process. Analogously, we choose a hidden layer of the FitNet, the guided
layer, to learn from the teacher’s hint layer. We want the guided layer to be able to predict the output
of the hint layer. Note that having hints is a form of regularization and thus, the pair hint/guided
layer has to be chosen such that the student network is not over-regularized. The deeper we set the
guided layer, the less flexibility we give to the network and, therefore, FitNets are more likely to
suffer from over-regularization. In our case, we choose the hint to be the middle layer of the teacher
network. Similarly, we choose the guided layer to be the middle layer of the student network.
Given that the teacher network will usually be wider than the FitNet, the selected hint layer may
have more outputs than the guided layer. For that reason, we add a regressor to the guided layer,
whose output matches the size of the hint layer. Then, we train the FitNet parameters from the first
layer up to the guided layer as well as the regressor parameters by minimizing the following loss
function:
1
LHT (WGuided , Wr ) = ||uh (x; WHint ) − r(vg (x; WGuided ); Wr )||2 ,
(3)
2
where uh and vg are the teacher/student deep nested functions up to their respective hint/guided
layers with parameters WHint and WGuided , r is the regressor function on top of the guided layer
with parameters Wr . Note that the outputs of uh and r have to be comparable, i.e., uh and r must
be the same non-linearity.
Nevertheless, using a fully-connected regressor increases the number of parameters and the memory
consumption dramatically in the case where the guided and hint layers are convolutional. Let Nh,1 ×
Nh,2 and Oh be the teacher hint’s spatial size and number of channels, respectively. Similarity, let
Ng,1 × Ng,2 and Og be the FitNet guided layer’s spatial size and number of channels. The number of
parameters in the weight matrix of a fully connected regressor is Nh,1 ×Nh,2 ×Oh ×Ng,1 ×Ng,2 ×Og .
To mitigate this limitation, we use a convolutional regressor instead. The convolutional regressor
is designed such that it considers approximately the same spatial region of the input image as the
teacher hint. Therefore, the output of the regressor has the same spatial size as the teacher hint.
Given a teacher hint of spatial size Nh,1 × Nh,2 , the regressor takes the output of the Fitnet’s guided
1
For example, as argued by Hinton & Dean (2014), with softened outputs, more information is provided
about the relative similarity of the input to classes other than the one with the highest probability.
3
Published as a conference paper at ICLR 2015
Teacher Network
FitNet
WSM
WTL
.
.
.
WT
WS
.
.
.
W*Guided =
(WGuided, Wr)
W*S =
(WS)
WSg
Wr
WT h
WHint
WGuided.
.
.
.
.
.
WHint
WS
WGuided
WT
WT 2
WT
W*Guided
WS2
1
WS1
(a) Teacher and Student Networks
(b) Hints Training
(c) Knowledge Distillation
Figure 1: Training a student network using hints.
layer of size Ng,1 × Ng,2 and adapts its kernel shape k1 × k2 such that Ng,i − ki + 1 = Nh,i ,
where i ∈ {1, 2}. The number of parameters in the weight matrix of a the convolutional regressor is
k1 × k2 × Oh × Og , where k1 × k2 is significantly lower than Nh,1 × Nh,2 × Ng,1 × Ng,2 .
2.3
F IT N ET S TAGE - WISE T RAINING
We train the FitNet in a stage-wise fashion following the student/teacher paradigm. Figure 1 summarizes the training pipeline. Starting from a trained teacher network and a randomly initialized
FitNet (Fig. 1 (a)), we add a regressor parameterized by Wr on top of the FitNet guided layer and
train the FitNet parameters WGuided up to the guided layer to minimize Eq. (3) (see Fig. 1 (b)).
Finally, from the pre-trained parameters, we train the parameters of whole FitNet WS to minimize
Eq. (2) (see Fig. 1 (c)). Algorithm 1 details the FitNet training process.
Algorithm 1 FitNet Stage-Wise Training.
The algorithm receives as input the trained parameters WT of a teacher, the randomly initialized
parameters WS of a FitNet, and two indices h and g corresponding to hint/guided layers, respectively. Let WHint be the teacher’s parameters up to the hint layer h. Let WGuided be the FitNet’s
parameters up to the guided layer g. Let Wr be the regressor’s parameters. The first stage consists in
pre-training the student network up to the guided layer, based on the prediction error of the teacher’s
hint layer (line 4). The second stage is a KD training of the whole network (line 6).
1:
2:
3:
4:
Input: WS , WT , g, h
Output: WS∗
WHint ← {WT 1 , . . . , WT h }
WGuided ← {WS 1 , . . . , WS g }
Intialize Wr to small random values
∗
WGuided
← argmin LHT (WGuided , Wr )
WGuided
5: {WS 1 , . . . , WS g } ← {WGuided ∗1 , . . . , WGuided ∗g }
6: WS∗ ← argmin LKD (WS )
WS
2.4
R ELATION
TO
C URRICULUM L EARNING
In this section, we argue that our hint-based training with KD can be seen as a particular form of
Curriculum Learning (Bengio, 2009). Curriculum learning has proven to accelerate the training
convergence as well as potentially improve the model generalization by properly choosing a sequence of training distributions seen by the learner: from simple examples to more complex ones.
A curriculum learning extension (Gulcehre & Bengio, 2013) has also shown that by using guidance
hints on an intermediate layer during the training, one could considerably ease training. However,
Bengio (2009) uses hand-defined heuristics to measure the “simplicity” of an example in a sequence
and Gulcehre & Bengio (2013)’s guidance hints require some prior knowledge of the end-task. Both
of these curriculum learning strategies tend to be problem-specific.
Our approach alleviates this issue by using a teacher model. Indeed, intermediate representations
learned by the teacher are used as hints to guide the FitNet optimization procedure. In addition, the
teacher confidence provides a measure of example “simplicity” by means of teacher cross-entropy
4
Published as a conference paper at ICLR 2015
term in Eq. (2). This term ensures that examples with a high teacher confidence have a stronger
impact than examples with low teacher confidence: the latter correspond to probabilities closer to
the uniform distribution, which exert less of a push on the student parameters. In other words, the
teacher penalizes the training examples according to its confidence. Note that parameter λ in Eq.
(2) controls the weight given to the teacher cross-entropy, and thus, the importance given to each
example. In order to promote the learning of more complex examples (examples with lower teacher
confidence), we gradually anneal λ during the training with a linear decay. The curriculum can be
seen as composed of two stages: first learn intermediate concepts via the hint/guided layer transfer,
then train the whole student network jointly, annealing λ, which allows easier examples (on which
the teacher is very confident) to initially have a stronger effect, but progressively decreasing their
importance as λ decays. Therefore, the hint-based training introduced in the paper is a generic
curriculum learning approach, where prior information about the task-at-hand is deduced purely
from the teacher model.
Algorithm
# params
Compression
FitNet
∼2.5M
Teacher
∼9M
Mimic single
∼54M
Mimic single
∼70M
Mimic ensemble
∼70M
State-of-the-art methods
Maxout
Network in Network
Deeply-Supervised Networks
Deeply-Supervised Networks (19)
Accuracy
91.61%
90.18%
84.6%
84.9%
85.8%
Algorithm
# params
Compression
FitNet
∼2.5M
Teacher
∼9M
State-of-the-art methods
Maxout
Network in Network
Deeply-Supervised Networks
90.65%
91.2%
91.78%
88.2%
Table 1: Accuracy on CIFAR-10
3
R ESULTS
ON
Accuracy
64.96%
63.54%
61.43%
64.32%
65.43%
Table 2: Accuracy on CIFAR-100
B ENCHMARK DATASETS
In this section, we show the results on several benchmark datasets2 . The architectures of all networks
as well as the training details are reported in the supplementary material.
3.1
CIFAR-10
AND
CIFAR-100
The CIFAR-10 and CIFAR-100 datasets (Krizhevsky & Hinton, 2009) are composed of 32x32 pixel
RGB images belonging to 10 and 100 different classes, respectively. They both contain 50K training
images and 10K test images. CIFAR-10 has 1000 samples per class, whereas CIFAR-100 has 100
samples per class. Like Goodfellow et al. (2013b), we normalized the datasets for contrast normalization and applied ZCA whitening.
CIFAR-10: To validate our approach, we trained a teacher network of maxout convolutional layers as reported in Goodfellow et al. (2013b) and designed a FitNet with 17 maxout convolutional
layers, followed by a maxout fully-connected layer and a top softmax layer, with roughly 1/3 of
the parameters. The 11th layer of the student network was trained to mimic the 2nd layer of the
teacher network. Like in Goodfellow et al. (2013b); Chen-Yu et al. (2014), we augmented the data
with random flipping during training. Table 1 summarizes the obtained results. Our student model
outperforms the teacher model, while requiring notably fewer parameters, suggesting that depth is
crucial to achieve better representations. When compared to network compression methods, our
algorithm achieves outstanding results; i.e., the student network achieves an accuracy of 91.61%,
which is significantly higher than the top-performer 85.8% of Ba & Caruana (2014), while requiring roughly 28 times fewer parameters. When compared to state-of-the-art methods, our algorithm
matches the best performers.
One could argue the choice of hinting the inner layers with the hidden state of a wide teacher network. A straightforward alternative would be to hint them with the desired output. This could be
addressed in a few different ways: (1) Stage-wise training, where stage 1 optimizes the 1st half of
the network w.r.t. classification targets and stage 2 optimizes the whole network w.r.t. classification
2
Code to reproduce the experiments publicly available: https://github.com/adri-romsor/FitNets
5
Published as a conference paper at ICLR 2015
targets. In this case, stage 1 set the network parameters in a good local minima but such initialization did not seem to help stage 2 sufficiently, which failed to learn. To further assist the training
of the thin and deep student network, we could add extra hints with the desired output at different
hidden layers. Nevertheless, as observed in (Bengio et al., 2007), with supervised pre-training the
guided layer may discard some factors from the input, which require more layers and non-linearity
before they can be exploited to predict the classes. (2) Stage-wise training with KD, where stage 1
optimizes the 1st half of the net w.r.t. classification targets and stage 2 optimizes the whole network
w.r.t. Eq. (2). As in the previous case, stage 1 set the network parameters in a good local minima
but such initialization did not seem to help stage 2 sufficiently, which failed to learn. (3) Jointly
optimizing both stages w.r.t. the sum of the supervised hint for the guided layer and classification
target for the output layer. We performed this experiment, tried different initializations and learning
rates with RMSprop (Tieleman & Hinton, 2012) but we could not find any combination to make the
network learn. Note that we could ease the training by adding hints to each layer and optimizing
jointly as in Deeply Supervised Networks (DSN). Therefore, we built the above-mentioned 19-layer
architecture and trained it by means of DSN, achieving a test performance of 88.2%, which is significantly lower than the performance obtained by the FitNets hint-based training (91.61%). Such
result suggests that using a very discriminative hint w.r.t. classification at intermediate layers might
be too aggressive; using a smoother hint (such as the guidance from a teacher network) offers better
generalization. (4) Jointly optimizing both stages w.r.t. the sum of supervised hint for the guided
layer and Eq. (2) for the output layer. Adding supervised hints to the middle layer of the network
did not ease the training of such a thin and deep network, which failed to learn.
CIFAR-100: To validate our approach, we trained a teacher network of maxout convolutional layers
as reported in Goodfellow et al. (2013b) and used the same FitNet architecture as in CIFAR-10.
As in Chen-Yu et al. (2014), we augmented the data with random flipping during training. Table 2
summarizes the obtained results. As in the previous case, our FitNet outperforms the teacher model,
reducing the number of parameters by a factor of 3 and, when compared to state-of-the-art methods,
the FitNet provides near state-of-the-art performance.
3.2
SVHN
The SVHN dataset (Netzer et al., 2011) is composed by 32 × 32 color images of house numbers
collected by GoogleStreet View. There are 73,257 images in the training set, 26,032 images in the
test set and 531,131 less difficult examples. We follow the evaluation procedure of Goodfellow et al.
(2013b) and use their maxout network as teacher. We trained a 13-layer FitNet composed of 11
maxout convolutional layers, a fully-connected layer and a softmax layer.
Algorithm
# params
Compression
FitNet
∼1.5M
Teacher
∼4.9M
State-of-the-art methods
Maxout
Network in Network
Deeply-Supervised Networks
Algorithm
# params
Compression
Teacher
∼361K
Standard backprop
∼30K
KD
∼30K
FitNet
∼30K
State-of-the-art methods
Maxout
Network in Network
Deeply-Supervised Networks
Misclass
2.42%
2.38%
2.47%
2.35%
1.92%
Table 3: SVHN error
Misclass
0.55%
1.9%
0.65%
0.51%
0.45%
0.47%
0.39%
Table 4: MNIST error
Table 3 shows that our FitNet achieves comparable accuracy than the teacher despite using only 32%
of teacher capacity. Our FitNet is comparable in terms of performance to other state-of-art methods,
such as Maxout and Network in Network.
3.3
MNIST
As a sanity check for the training procedure, we evaluated the proposed method on the MNIST
dataset (LeCun et al., 1998). MNIST is a dataset of handwritten digits (from 0 to 9) composed of
28x28 pixel greyscale images, with 60K training images and 10K test images. We trained a teacher
network of maxout convolutional layers as reported in Goodfellow et al. (2013b) and designed a
6
Published as a conference paper at ICLR 2015
FitNet twice as deep as the teacher network and with roughly 8% of the parameters. The 4th layer
of the student network was trained to mimic the 2nd layer of the teacher network.
Table 4 reports the obtained results. To verify the influence of using hints, we trained the FitNet
architecture using either (1) standard backprop (w.r.t. classification labels), (2) KD or (3) Hintbased Training (HT). When training the FitNet with standard backprop from the softmax layer, the
deep and thin architecture achieves 1.9% misclassification error. Using KD, the very same network
achieves 0.65%, which confirms the potential of the teacher network; and when adding hints, the
error still decreases to 0.51%. Furthermore, the student network achieves slightly better results than
the teacher network, while requiring 12 times less parameters.
3.4
AFLW
AFLW (Koestinger et al., 2011) is a real-world face database, containing 25K annotated images. In
order to evaluate the proposed framework in a face recognition setting, we extracted positive samples
by re-sizing the annotated regions of the images to fit 16x16 pixels patches. Similarly, we extracted
25K 16x16 pixels patches not containing faces from ImageNet (Russakovsky et al., 2014) dataset,
as negative samples. We used 90% of the extracted patches to train the network.
In this experiment, we aimed to evaluate the method on a different kind of architecture. Therefore,
we trained a teacher network of 3 ReLU convolutional layers and a sigmoid output layer. We designed a first FitNet (FitNet 1) with 15 times fewer multiplications than the teacher network, and a
second FitNet (FitNet 2) with 2.5 times fewer multiplications than the teacher network. Both FitNets
have 7 ReLU convolutional layers and a sigmoid output layer.
The teacher network achieved 4.21% misclassification error on the validation set. We trained both
FitNets by means of KD and HT. On the one hand, we report a misclassification error of 4.58%
when training FitNet 1 with KD and a misclassification error of when 2.55% when training it with
HT. On the other hand, we report a missclassifation error of 1.95% when training FitNet 2 with KD
and a misclassification error of 1.85% when training it with HT. These results show how the method
is extensible to different kind of architectures and highlight the benefits of using hints, especially
when dealing with thinner architectures.
4
A NALYSIS
OF EMPIRICAL RESULTS
We empirically investigate the benefits of our approach by comparing various networks trained using standard backpropagation (cross-entropy w.r.t. labels), KD or Hint-based Training (HT). Experiments are performed on CIFAR-10 dataset (Krizhevsky & Hinton, 2009).
We compare networks of increasing depth given a fixed computational budget. Each network is composed of successive convolutional layers of kernel size 3 × 3, followed by a maxout non-linearity
and a non-overlapping 2 × 2 max-pooling. The last max-pooling takes the maximum over all remaining spatial dimensions leading to a 1 × 1 vector representation. We only change the depth and
the number of channels per convolution between different networks, i.e. the number of channels per
convolutional layer decreases as a network depth increases to respect a given computational budget.
4.1
A SSISTING
THE TRAINING OF DEEP NETWORKS
In this section, we investigate the impact of HT. We consider two computational budgets of approximately 30M and 107M operations, corresponding to the multiplications needed in an image
forward propagation. For each computational budget, we train networks composed of 3, 5, 7 and
9 convolutional layers, followed by a fully-connected layer and a softmax layer. We compare their
performances when they are trained with standard backpropagation, KD and HT. Figure 2 reports
test on CIFAR-10 using early stopping on the validation set, i.e. we do not retrain our models on the
training plus validation sets.
Due to their depth and small capacity, FitNets are hard to train. As shown in Figure 2(a), we could
not train 30M multiplications networks with more than 5 layers with standard backprop. When
using KD, we succesfully trained networks up to 7 layers. Adding KD’s teacher cross-entropy to
the training objective (Eq. (2)) gives more importance to easier examples, i.e. samples for which
the teacher network is confident and, can lead to a smoother version of the training cost (Bengio,
2009). Despite some optimization benefits, it is worth noticing that KD training still suffers from
7
Published as a conference paper at ICLR 2015
88.5
91.0
88.0
90.5
87.5
90.0
89.5
Test Accuracy
Test Accuracy
87.0
86.5
86.0
85.5
85.0
88.5
88.0
87.5
84.5
87.0
Back Propagation
Knowledge Distillation
Hint Training
84.0
83.5
83.0
89.0
5
7
9
Number of Layers
Back Propagation
Knowledge Distillation
Hint Training
86.5
86.0
85.5
11
(a) 30M Multiplications
5
7
9
Number of Layers
11
(b) 107M Multiplications
Figure 2: Comparison of Standard Back-Propagation, Knowledge Distillation and Hint-based Training on CIFAR-10.
Network
Teacher
FitNet 1
FitNet 2
FitNet 3
FitNet 4
# layers
5
11
11
13
19
# params
∼9M
∼250K
∼862K
∼1.6M
∼2.5M
# mult
∼725M
∼30M
∼108M
∼392M
∼382M
Acc
90.18%
89.01%
91.06%
91.10%
91.61%
Speed-up
1
13.36
4.64
1.37
1.52
Compression rate
1
36
10.44
5.62
3.60
Table 5: Accuracy/Speed Trade-off on CIFAR-10.
the increasing depth and reaches its limits for 7-layer networks. HT tends to ease these optimization
issues and is able to train 13-layer networks of 30M multiplications. The only difference between
HT and KD is the starting point in the parameter space: either random or obtained by means of the
teacher’s hint. On the one hand, the proliferation of local minima and especially saddle points in
highly non-linear functions such as very deep networks highlights the difficulty of finding a good
starting point in the parameter space at random (Dauphin et al., 2014). On the other hand, results in
Figure 2(a) indicate that HT can guide the student to a better initial position in the parameter space,
from which we can minimize the cost through stochastic gradient descent. Therefore, HT provides
benefits from an optimization point of view. Networks trained with HT also tend to yield better
test performances than the other training methods when we fix the capacity and number of layers.
For instance, in Figure 2(b), the 7-layers network, trained with hints, obtains a +0.7% performance
gain on the test set compared to the model that does not use any hints (the accuracy increases from
89.45% to 90.1%). As pointed by Erhan et al. (2009), pre-training strategies can act as regularizers.
These results suggest that HT is a stronger regularizer than KD, since it leads to better generalization
performance on the test set. Finally, Figure 2 highlights that deep models have better performances
than shallower ones given a fixed computational budget. Indeed, considering networks that are
trained with hints, an 11-layer network outperforms a 5-layer network by an absolute improvement
of 4.11% for 107M multiplications and of 3.4% for 30M multiplications. Therefore, the experiments
validate our hypothesis that given a fixed number of computations, we leverage depth in a model to
achieve faster computation and better generalization.
In summary, this experiment shows that (1) using HT, we are able to train deeper models than with
standard back-propagation and KD; and (2) given a fixed capacity, deeper models performed better
than shallower ones.
4.2
T RADE - OFF B ETWEEN M ODEL P ERFORMANCE
AND
E FFICIENCY
To evaluate FitNets efficiency, we measure their total inference times required for processing
CIFAR-10 test examples on a GPU as well as their parameter compression. Table 5 reports both
the speed-up and compression rate obtained by various FitNets w.r.t. the teacher model along with
their number of layers, capacity and accuracies. In this experiment, we retrain our FitNets on training plus validation sets as in Goodfellow et al. (2013b), for fair comparison with the teacher.
8
Published as a conference paper at ICLR 2015
FitNet 1, our smallest network, with 36× less capacity than the teacher, is one order of magnitude
faster than the teacher and only witnesses a minor performance decrease of 1.3%. FitNet 2, slightly
increasing the capacity, outperforms the teacher by 0.9%, while still being faster by a strong 4.64
factor. By further increasing network capacity and depth in FitNets 3 and 4, we improve the performance gain, up to 1.6%, and still remain faster than the teacher. Although a trade-off between speed
and accuracy is introduced by the compression rate, FitNets tend to be significantly faster, matching
or outperforming their teacher, even when having low capacity.
A few works such as matrix factorization (Jaderberg et al., 2014; Denton et al., 2014) focus on
speeding-up deep networks’ convolutional layers at the expense of slightly deteriorating their performance. Such approaches are complementary to FitNets and could be used to further speed-up the
FitNet’s convolutional layers.
Other works related to quantization schemes (Chen et al., 2010; Jégou et al., 2011; Gong et al.,
2014) aim at reducing storage requirements. Unlike FitNets, such approaches witness a little decrease in performance when compressing the network parameters. Exploiting depth allows FitNets
to obtain performance improvements w.r.t. their teachers, even when reducing the number of parameters 10×. However, we believe that quantization approaches are also complementary to FitNets and
could be used to further reduce the storage requirements. It would be interesting to compare how
much redundancy is present in the filters of the teacher networks w.r.t. the filters of the FitNet and,
therefore, how much FitNets filters could be compressed without witnessing significant performance
drop. This analysis is out of the scope of the paper and is left as future work.
5
C ONCLUSION
We proposed a novel framework to compress wide and deep networks into thin and deeper ones, by
introducing intermediate-level hints from the teacher hidden layers to guide the training process of
the student. We are able to use these hints to train very deep student models with less parameters,
which can generalize better and/or run faster than their teachers. We provided empirical evidence
that hinting the inner layers of a thin and deep network with the hidden state of a teacher network
generalizes better than hinting them with the classification targets. Our experiments on benchmark
datasets emphasize that deep networks with low capacity are able to extract feature representations
that are comparable or even better than networks with as much as 10 times more parameters. The
hint-based training suggests that more efforts should be devoted to explore new training strategies
to leverage the power of deep networks.
ACKNOWLEDGMENTS
We thank the developers of Theano (Bastien et al., 2012) and Pylearn2 (Goodfellow et al., 2013a)
and the computational resources provided by Compute Canada and Calcul Québec. This work has
been partially supported by NSERC, CIFAR, and Canada Research Chairs, Project TIN2013-41751,
grant 2014-SGR-221 and Spanish MINECO grant TIN2012-38187-C03.
R EFERENCES
Ba, J. and Caruana, R. Do deep nets really need to be deep? In NIPS, pp. 2654–2662. 2014.
Bastien, F., Lamblin, P., Pascanu, R., Bergstra, J., Goodfellow, I., Bergeron, A., Bouchard, N., Warde-Farley,
D., and Bengio, Y. Theano: new features and speed improvements. Deep Learning & Unsupervised Feature
Learning Workshop, NIPS, 2012.
Bengio, Y. Learning deep architectures for AI. Foundations and trends in Machine Learning, 2009.
Bengio, Y., Lamblin, P., Popovici, D., and Larochelle, H. Greedy layer-wise training of deep networks. In
NIPS, pp. 153–160, 2007.
Bengio, Y., Courville, A., and Vincent, P. Representation learning: A review and new perspectives. TPAMI,
2013.
Bucila, C., Caruana, R., and Niculescu-Mizil, A. Model compression. In KDD, pp. 535–541, 2006.
Chen, Yongjian, Guan, Tao, and Wang, Cheng. Approximate nearest neighbor search by residual vector quantization. Sensors, 10(12):11259–11273, 2010.
9
Published as a conference paper at ICLR 2015
Chen-Yu, L., Saining, X., Patrick, G., Zhengyou, Z., and Zhuowen, T. Deeply-supervised nets. CoRR,
abs/1409.5185, 2014.
Cho, Kyunghyun, Raiko, Tapani, Ilin, Alexander, and Karhunen, Juha. A two-stage pretraining algorithm for
deep Boltzmann machines. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2012.
Dauphin, Y., Pascanu, R., Gulcehre, C., Cho, K., Ganguli, S., and Bengio, Y. Identifying and attacking the
saddle point problem in high-dimensional non-convex optimization. In NIPS, 2014.
Denton, Emily L, Zaremba, Wojciech, Bruna, Joan, LeCun, Yann, and Fergus, Rob. Exploiting linear structure
within convolutional networks for efficient evaluation. In NIPS, pp. 1269–1277. 2014.
Erhan, D., Manzagol, P.A., Bengio, Y., Bengio, S., and Vincent, P. The difficulty of training deep architectures
and the effect of unsupervised pre-training. In AISTATS, pp. 153–160, 2009.
Gong, Yunchao, Liu, Liu, Yang, Min, and Bourdev, Lubomir. Compressing deep convolutional networks using
vector quantization. CoRR, abs/1412.6115, 2014.
Goodfellow, I. J., Warde-Farley, D., Lamblin, P., Dumoulin, V., Mirza, M., Pascanu, R., Bergstra, J., Bastien,
F., and Bengio, Y. Pylearn2: a machine learning research library. arXiv preprint arXiv:1308.4214, 2013a.
Goodfellow, I.J., Warde-Farley, D., Mirza, M., Courville, A., and Bengio, Y. Maxout networks. In ICML,
2013b.
Gulcehre, C. and Bengio, Y. Knowledge matters: Importance of prior information for optimization. In ICLR,
2013.
Hinton, G. E., Osindero, S., and Teh, Y.-W. A fast learning algorithm for deep belief nets. Neural Computation,
18(7):1527–1554, 2006.
Hinton, G. Vinyals, O. and Dean, J. Distilling knowledge in a neural network. In Deep Learning and Representation Learning Workshop, NIPS, 2014.
Jaderberg, M., Vedaldi, A., and Zisserman, A. Speeding up convolutional neural networks with low rank
expansions. In BMVC, 2014.
Jégou, Hervé, Douze, Matthijs, and Schmid, Cordelia. Product quantization for nearest neighbor search. IEEE
TPAMI, 33(1):117–128, 2011.
Koestinger, M., Wohlhart, P., Roth, P.M., and Bischof, H. Annotated facial landmarks in the wild: A large-scale,
real-world database for facial landmark localization. In First IEEE International Workshop on Benchmarking
Facial Image Analysis Technologies, 2011.
Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. Master’s thesis, Department of Computer Science, University of Toronto, 2009.
Larochelle, H., Erhan, D., Courville, A., Bergstra, J., and Bengio, Y. An empirical evaluation of deep architectures on problems with many factors of variation. In ICML, pp. 473–480, 2007.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-based learning applied to document recognition.
Proceedings of the IEEE, 86(11):2278–2324, November 1998.
Montufar, G.F., Pascanu, R., Cho, K., and Bengio, Y. On the number of linear regions of deep neural networks.
In NIPS. 2014.
Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and Ng, A. Reading digits in natural images with
unsupervised feature learning. In Deep Learning & Unsupervised Feature Learning Workshop, NIPS, 2011.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,
Bernstein, M., Berg, A. C., and Fei-Fei, L. ImageNet Large Scale Visual Recognition Challenge, 2014.
Simonyan, K. and Zisserman, A. Very deep convolutional networks for large-scale image recognition. CoRR,
abs/1409.1556, 2014.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., D.A., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going
deeper with convolutions. CoRR, abs/1409.4842, 2014.
Tieleman, T. and Hinton, G. Lecture 6.5—RmsProp: Divide the gradient by a running average of its recent
magnitude. COURSERA: Neural Networks for Machine Learning, 2012.
Weston, J., Ratle, F., and Collobert, R. Deep learning via semi-supervised embedding. In ICML, 2008.
10
Published as a conference paper at ICLR 2015
A
S UPPLEMENTARY M ATERIAL : N ETWORK A RCHITECTURES
T RAINING P ROCEDURES
AND
In the supplementary material, we describe all network architectures and hyper-parameters used
throughout the paper.
A.1
CIFAR-10/CIFAR-100
In this section, we describe the teacher and FitNet architectures as well as hyper-parameters used in
both CIFAR-10/CIFAR-100 experiments.
A.1.1
T EACHERS
We used the CIFAR-10/CIFAR-100 maxout convolutional networks reported in Goodfellow et al.
(2013b) as teachers. Both teachers have the same architecture, composed of 3 convolutional hidden
layers of 96-192-192 units, respectively. Each convolutional layer is followed by a maxout nonlinearity (with 2 linear pieces) and a max-pooling operator with respective windows sizes of 4x4,
4x4 and 2x2 pixels. All max-pooling units have an overlap of 2x2 pixels. The third convolutional
layer is followed by a fully-connected maxout layer of 500 units (with 5 linear pieces) and a top
softmax layer. The CIFAR-10/CIFAR-100 teachers are trained using stochastic gradient descent
and momentum. Please refer to Goodfellow et al. (2013b) for more details.
A.1.2
F IT N ETS
Here, we describe the FitNet architectures used in the Section 3 and Section 4. Each FitNet is
composed of successive zero-padded convolutional layers of kernel size 3 × 3, followed by a maxout
non-linearity with two linear pieces. A non-overlapping 2 × 2 max-pooling follows some of the
convolutional layers; each network has a total of 3 max-pooling units. The last max-pooling takes
the maximum over all remaining spatial dimensions, leading to a 1 × 1 vector representation. The
last convolutional layer is followed by a fully-connected and a softmax layer, as the ones on CIFAR10/100 teachers.
Table 6 describes the architectures used for the depth experiment in Figure 2. Table 7 describes the
architectures for the efficiency-performance trade-off experiment in Table 5. The results reported in
Table 1, Table 2 and Table 3 correspond to the FitNet 4 architecture.
All FitNet parameters were initialized randomly in U(-0.005,0.005). We used stochastic gradient
descent with RMSProp (Tieleman & Hinton, 2012) to train the FitNets, with an initial learning rate
0.005 and a mini-batch size of 128. Parameter λ in Eq. (2) was initialized to 4 and decayed linearly
during 500 epochs reaching λ = 1. The relaxation term τ was set to 3.
On CIFAR-10, we divided the training set into 40K training examples and 10K validation examples.
We trained stage 1 by minimizing Eq. (3) and stopped the training after 100 epochs of no validation
error improvement, performing a maximum of 500 epochs. After that, we trained stage 2 by minimizing Eq. (2) using RMSprop, the same stopping criterion and the same hyper-parameters as stage
1. We picked the optimal number of epochs according to the above-mentioned stopping criterion
and retrained the FitNet on the whole 50K training examples (training + validation sets).
On CIFAR-100, we trained directly on the whole training set using stochastic gradient descent with
RMSprop, the same hyper-parameters as CIFAR-10 FitNets and the number of epochs determined
by CIFAR-10 stopping criterion.
A.2
MNIST
In this section, we describe the teacher and FitNet architectures as well as the hyper-parameters used
in the MNIST experiments.
We trained a teacher network of maxout convolutional layers as reported in Goodfellow et al.
(2013b). The teacher architecture has three convolutional maxout hidden layers (with 2 linear pieces
each) of 48-48-24 units, respectively, followed by a spatial max-pooling of 4x4-4x4-2x2 pixels, with
11
Published as a conference paper at ICLR 2015
5 Layer
conv 3x3x64 (3x3x128)
pool 2x2
7 Layer
conv 3x3x16 (3x3x32)
conv 3x3x32 (3x3x64)
pool 2x2
9 Layer
conv 3x3x16 (3x3x32)
conv 3x3x32 (3x3x32)
pool 2x2
conv 3x3x64 (3x3x128)
pool 2x2
conv 3x3x32 (3x3x80)
conv 3x3x64 (3x3x80)
pool 2x2
conv 3x3x64 (3x3x128)
pool 8x8
conv 3x3x64 (3x3x128)
pool 8x8
conv 3x3x32 (3x3x64)
conv 3x3x32 (3x3x80)
conv 3x3x32 (3x3x80)
pool 2x2
conv 3x3x48 (3x3x96)
conv 3x3x64 (3x3x128)
pool 8x8
fc
softmax
hint: 2←2
fc
softmax
hint: 4←2
fc
softmax
hint: 5←2
11 Layer
conv 3x3x16 (3x3x16)
conv 3x3x16 (3x3x32)
conv 3x3x16 (3x3x32)
pool 2x2
conv 3x3x32 (3x3x48)
conv 3x3x32 (3x3x64)
conv 3x3x32 (3x3x80)
pool 2x2
conv 3x3x48 (3x3x96)
conv 3x3x48 (3x3x96)
conv 3x3x64 (3x3x128)
pool 8x8
fc
softmax
hint: 7←2
Table 6: Fitnet architectures with a computational budget of 30M (or 107M) of multiplications:
conv sx × sy × c is a convolution of kernel size sx × sy with c outputs channels; pool sx × sy is
a non-overlapping pooling of size sx × sy ; fc stands for fully connected. hint: FitNet ← teacher
specifies the hint and guided layers used for hint-based training, respectively.
FitNet 1
conv 3x3x16
conv 3x3x16
conv 3x3x16
pool 2x2
FitNet 2
conv 3x3x16
conv 3x3x32
conv 3x3x32
pool 2x2
FitNet 3
conv 3x3x32
conv 3x3x48
conv 3x3x64
conv 3x3x64
pool 2x2
conv 3x3x32
conv 3x3x32
conv 3x3x32
pool 2x2
conv 3x3x48
conv 3x3x64
conv 3x3x80
pool 2x2
conv 3x3x80
conv 3x3x80
conv 3x3x80
conv 3x3x80
pool 2x2
conv 3x3x48
conv 3x3x48
conv 3x3x64
pool 8x8
conv 3x3x96
conv 3x3x96
conv 3x3x128
pool 8x8
conv 3x3x128
conv 3x3x128
conv 3x3x128
pool 8x8
fc
softmax
hint: 6←2
fc
softmax
hint: 6←2
fc
softmax
hint: 8←2
FitNet 4
conv 3x3x32
conv 3x3x32
conv 3x3x32
conv 3x3x48
conv 3x3x48
pool 2x2
conv 3x3x80
conv 3x3x80
conv 3x3x80
conv 3x3x80
conv 3x3x80
conv 3x3x80
pool 2x2
conv 3x3x128
conv 3x3x128
conv 3x3x128
conv 3x3x128
conv 3x3x128
conv 3x3x128
pool 8x8
fc
softmax
hint: 11←2
Table 7: Performance-Efficiency FitNet architectures.
12
Published as a conference paper at ICLR 2015
an overlap of 2x2 pixels. The 3rd hidden layer is followed by a fully-connected softmax layer. As is
Goodfellow et al. (2013b), we added zero padding to the second convolutional layer.
We designed a FitNet twice as deep as the teacher network and with roughly 8% of the parameters.
The student architecture has 6 maxout convolutional hidden layers (with 2 linear pieces each) of
16-16-16-16-12-12 units, respectively. Max-pooling is only applied every second layer in regions of
4x4-4x4-2x2 pixels, with an overlap of 2x2 pixels. The 6th convolutional hidden layer is followed
by a fully-connected softmax layer.
The teacher network was trained as described in Goodfellow et al. (2013b). The FitNet was trained
in a stage-wise fashion as described in Section 2. We divided the training set into a training set of
50K samples and a validation set of 10K samples.
All network parameters where initialized randomly in U(-0.005,0.005). In the first stage, the 4th
layer of the FitNet was trained to mimic the 2nd layer of the teacher network, by minimizing Eq.
(3) through stochastic gradient descent. We used a mini-batch size of 128 samples and fixed the
learning rate to 0.0005. We initialized λ to 4 and decayed it for the first 150 epochs until it reached
1. The training was stopped according to the following criterion: after 100 epochs of no validation
error improvement and performning a maximum of 500 epochs. We used the same mini-batch size,
learning rate and stopping criterion to train the second stage. The relaxation term τ was set to 3.
A.3
SVHN
In this section, we describe the teacher and FitNet architectures as well as the hyper-parameters used
in the SVHN experiments.
We used SVHN maxout convolutional network described in as Goodfellow et al. (2013b) teacher.
The network is composed of 3 convolutional hidden layers of 64-128-128 units, respectively, followed by a fully-connected maxout layer of 400 units and a top softmax layer. The teacher training
was carried out as in Goodfellow et al. (2013b).
We used the FitNet 4 architecture outlined in Table 7, initializing the network parameters randomly
in U(-0.005,0.005) and training with the same hyper-parameters as in CIFAR-10. In this case, we
used the same early stopping as in CIFAR-10, but we did not retrain the FitNet on the whole training
set (training + validation). The same hyper-parameters where used for both stages.
A.4
AFLW
In this section, we describe the teacher and FitNet architectures as well as the hyper-parameters used
in the AFLW experiments.
We trained a teacher network of 3 ReLU convolutional layers of 128-512-512 units, respectively,
followed by a sigmoid layer. Non-overlapping max-pooling of size 2 × 2 was performed after the
first convolutional layer. We used receptive fields of 3-2-5 for each layer, respectively.
We designed two FitNets of 7 ReLU convolutional layers. Fitnet 1’s layers have 16-32-32-32-3232-32-32 units, respectively, followed by a sigmoid layer. Fitnet 2’s layers have 32-64-64-64-6464-64-64 units, respectively, followed by a sigmoid layer. In both cases, we used receptive fields of
3 × 3 and, due to the really small image resolution, we did not perform any max-pooling.
All network parameters of both FitNets where initialized randomly in U(-0.05,0.05). Both FitNets
were trained in the stage-wise fashion described in Section 2. We used 90% of the data for training.
In the first stage, the 5th layer of the FitNets were trained to mimic the 3rd layer of the teacher
network, by minimizing Eq. (3) through stochastic gradient descent. We used a mini-batch size of
128 samples and initialized the learning rate to 0.001 and decayed it for the first 100 epochs until
reaching 0.01. We also used momentum. We initialized momentum to 0.1 and saturated it to 0.9 at
epoch 100. We picked the best validation value after a 500 epochs. We used the same mini-batch
size, learning rate and stopping criterion to train the second stage. The relaxation term τ was set to
3.
13
| 9cs.NE
|
Ensembles of Multiple Models and Architectures
for Robust Brain Tumour Segmentation
K. Kamnitsas, W. Bai∗ , E. Ferrante∗ , S. McDonagh∗ , M. Sinclair∗
N. Pawlowski, M. Rajchl, M. Lee, B. Kainz, D. Rueckert, B. Glocker
arXiv:1711.01468v1 [cs.CV] 4 Nov 2017
Biomedical Image Analysis Group, Imperial College London, UK
∗
Equal contribution, in alphabetical order
Abstract. Deep learning approaches such as convolutional neural nets
have consistently outperformed previous methods on challenging tasks
such as dense, semantic segmentation. However, the various proposed
networks perform differently, with behaviour largely influenced by architectural choices and training settings. This paper explores Ensembles
of Multiple Models and Architectures (EMMA) for robust performance
through aggregation of predictions from a wide range of methods. The approach reduces the influence of the meta-parameters of individual models and the risk of overfitting the configuration to a particular database.
EMMA can be seen as an unbiased, generic deep learning model which
is shown to yield excellent performance, winning the first position in the
BRATS 2017 competition among 50+ participating teams.
1
Introduction
Brain tumours are among the most fatal types of cancer [1]. Out of tumours that
originally develop in the brain, gliomas are the most frequent [2]. They arise from
glioma cells and, depending on their aggressiveness, they are broadly categorized
into high and low grade gliomas [3]. High grade gliomas (HGG) develop rapidly
and aggressively, forming abnormal vessels and often a necrotic core, accompanied by surrounding oedema and swelling [2]. They are malignant, with high
mortality and average survival rate of less than two years even after treatment
[3]. Low grade gliomas (LGG) can be benign or malignant, grow slower, but they
may recur and evolve to HGG, thus their treatment is warranted. For treatment,
patients undergo radiotherapy, chemotherapy and surgery [1].
Firstly for diagnosis and monitoring the tumour’s progression, then for treatment planning and afterwards for assessing the effect of treatment, various neuroimaging protocols are employed. Magnetic resonance imaging (MRI) is widely
used in both clinical routine and research studies. It facilitates tumour analysis
by allowing estimation of extent, location and investigation of its subcomponents
[2]. This however requires accurate delineation of the tumour, which proves challenging due to its complex structure and appearance, the 3D nature of the MR
images and the multiple MR sequences that need to be consulted in parallel for
informed judgement. These factors make manual delineation time-consuming
and subject to inter- and intra-rater variability [4].
2
Kamnitsas, Bai, Ferrante, McDonagh, Sinclair, et al.
Fig. 1: Left to right: FLAIR; manual annotation of a BRATS’17 subject, where
yellow depicts oedema surrounding tumour core; confidence of a CNN predicting
oedema, trained with cross-entropy or IoU loss. Although overall performance is
similar, training with IoU (or Dice, not shown) loss alters the CNN’s behaviour,
which tends to output only highly confident predictions, even when false.
Automatic segmentation systems aim at providing an objective and scalable
solution. Representative early works are the atlas-based outlier detection method
[5] and the joint segmentation-registration framework, often guided by a tumour
growth model [6,7,8]. The past few years saw rapid developments of machine
learning methods, with Random Forests being among the most successful [9,10].
More recently, convolutional neural networks (CNN) have gained popularity by
exhibiting very promising results for segmentation of brain tumours [11,12,13].
A variety of CNN architectures have been proposed, each presenting different
strengths and weaknesses. Additionally, networks have a vast number of meta
parameters. The multiple configuration choices for a system influence not only
performance but also its behaviour (Fig. 1). For instance, different models may
perform better with different types of pre-processing. Consequently, when investigating their behaviour on a given task, findings can be biased. Finally, a
configuration highly optimized on a given database may be an over-fit, and not
generalise to other data or tasks.
In this work we push towards constructing a more reliable and objective deep
learning model. We bring together a variety of CNN architectures, configured
and trained in diverse ways in order to introduce high variance between them. By
combining them, we construct an Ensemble of Multiple Models and Architectures
(EMMA), with the aim of averaging away the variance and with it model- and
configuration-specific behaviours. Our approach leads to: (1) a system robust to
unpredictable failures of independent components, (2) enables objective analysis
with a generic deep learning model of unbiased behaviour, (3) introduces the
new perspective of ensembling for objectiveness. This is in contrast to common
ensembles, where a single model is trained with small variations such as initial
seeds, which renders the ensemble biased by the main architectural choices. As
a first milestone in this endeavour, we evaluated EMMA in the Brain Tumour
Segmentation (BRATS) challenge 2017. Our method won the first position in
the final testing stage among 50+ competing teams. This indicates the reliability
of the approach and paves the way for its use in further analysis.
EMMA for Robust Brain Tumour Segmentation
2
3
Background: Model bias, variance and ensembling
Feedforward neural networks have been shown capable of approximating any
function [14]. They are thus models with zero bias, possible of no systematic
error. However they are not a panacea. If left unregularized they can overfit noise
in the training data, which leads to mistakes when they are called to generalise.
Coupled with the stochasticity of the optimization process and the multiple
local minima, this leads to unpredictable inconsistent errors between different
instances. This constitutes models with high variance. Regularization reduces the
variance but increases the bias, as expressed in the bias/variance dilemma [15].
Regularization can be explicit, such as weight decay that prevents networks from
learning rare noisy patterns, or implicit, such as the local connectivity of CNN
kernels, which however does not allow the model to learn patterns larger than
the its receptive field. Architectural and configuration choices thus introduce
bias, altering the behaviour of a network.
One route to address the bias/variance dilemma is ensembling. By combining
multiple models, ensembling seeks to create a higher performing model with low
variance. The most popular combination rule is averaging, which is not sensitive
to inconsistent errors of the singletons [16]. Commonly, instances of a network
trained with different initial weights or from multiple final local minima are
ensembled, with the majority correcting irregular errors. Intuitively, only inconsistent errors can be averaged out. Lack of consistent failures can be interpreted
as statistical independency. Thus methods for de-correlating the instances have
been developed. The most popular is bagging [17], commonly used for random
forests. It uses bootstrap sampling to learn less correlated instances from different subsets of the data.
The above works often discuss ensembling as a means of increasing performance. [18] approached high variance from the scope of unreliability. They
discussed ensembling as a type of N-version programming, which advocates reliability through redundancy. When producing N-versions of a program, versions
may fail independently but through majority voting they behave as a reliable
system. They formalize intuitive requirements for reliability: a) the target function to be covered by the ensemble and b) the majority to be correct. This in
turn advocates diversity, independence and overall quality of the components.
Biomedical applications are reliability-critical and high variance would deter
the use of neural networks. For this reason we set off to investigate robustness
of diverse ensembles. Diverting from the above works, we introduce another
perspective of ensembling: creating an objective, configuration-invariant model
to facilitate objective analysis.
3
Ensembles of Multiple Models and Architectures
A variety of CNN architectures has shown promising results in recent literature.
Regarding the architectures, they commonly differ in depth, number of filters and
how they process multi-scale context among others. Such architectural choices
4
Kamnitsas, Bai, Ferrante, McDonagh, Sinclair, et al.
Fig. 2: Our ensemble of diverse networks, EMMA (red), averages out the bias
infused by individual model configurations m, to approximate more reliably the
true posterior (black), while being robust to suboptimal configurations. Posteriors on the left were obtained from multiple perceptrons, trained to classify
clusters centred on 10 and -10 as a toy example, with different losses, regularizations and noise in the training labels. Their ensemble provides reliable estimates.
bias the model behaviour. For instance, models with large receptive fields may
show improved localisation capabilities but can be less sensitive to fine texture
than models emphasizing local information. Strategies to handle class imbalance is another performance relevant parameter. Common strategies are training with class-weighted sampling or class-weighted cross entropy. As analysed in
[13], these methods strongly influence the sensitivity of the model to each class.
Furthermore, the choice of the loss function impacts results. For example, we
observed that networks trained to optimize Intersection over Union (IoU), Dice
or similar losses [19] tend to give worse confidence estimations than when trained
with cross entropy (Fig. 1). Finally, the setting of hyper-parameters for the optimization can strongly affect performance. It is often observed by practitioners
that the choice of the optimizer and its configuration, for instance the learning
rate schedule, can make the difference between bad and good segmentation.
The sensitivity to such meta-parameters is a greater problem than merely a
time-consuming manual optimization of configurations:
– A configuration setting optimized on one set of training data may be overfitting them and not perform well on unseen data or another task. This can
be viewed as another source of high model variance (Sec. 2).
– By biasing the behaviour of the model, it also biases the findings of any
analysis performed with it.
We now formalize the problem and our perspective of ensembling as a solution
as follows. Given training data X with labels Y , we need to learn the generating
process P (y|x). This is commonly approximated by a model P (y|x; θm , m), which
has trainable parameters θm that are learnt via an optimization process that
EMMA for Robust Brain Tumour Segmentation
5
minimizes:
θm = min d(P (Y |X; θm , m), P (Y |X))
θm
(1)
where d is a distance (defined by the type of loss) computed at the points given
by the training data, while m represents the choice of the meta-parameters. It
is commonly neglected although it conditions (biases) the learnt estimator. To
take it into account, we instead define m as a stochastic variable over the space
of meta-parameter configurations, with a corresponding prior P (m). In order to
learn a model of P (y|x) unbiased by m, we marginalize out its effect:
X
X
P (y|x) =
P (y, m|x) =
P (y|x, m)P (m)
m
m
≈
X
∀m∈E
P (y|x; θm , m)
1
= PEM M A (y|x)
|E|
(2)
Here E is the set of models within the ensemble. The prior P (m) is considered
uniform over a subspace of m that is covered by the models in E and zero
elsewhere. Note we have arrived at the standard ensembling with averaging, by
considering that each individual model P (y|x; θm , m) approximates a conditional
P (y|x, m) on m, and the true posterior is approximated by the ensemble which
marginalizes away effects of m. Note that the case of a single model configured
by m can be derived from the above, by setting a dirac prior P (m) = δ(m).
Thus the ensemble relaxes a pre-existing neglected strong prior.
The above formulation presents averaging ensembles from a new perspective:
The marginalization over a subspace of the joint P (y|x, m) offers generalisation,
regularising the (manual) optimization process of m from falling into minima
where P (Y |X, m) overfits P (Y |X) on the given training data (Y, X) (Fig. 2).
Moreover, the process leads to a more objective approximation of P (y|x) where
the biasing effect of m has been marginalized out. The exposed limitations agree
with the requirements for ensembling mentioned in Sec. 2: we need to restrict
the subspace of m into an area of relatively high quality models and we need to
cover it with a relatively small number of models, thus diversity is key.
In the remainder of this section we describe the main properties of the models
used to construct the collection E of EMMA, which cover various contemporary
architectures, configured and trained under different settings1 .
3.1
DeepMedic
Model description: The first architecture we employ is DeepMedic, originally
presented in [20,13]. It is a fully 3D, multi-scale CNN, designed with a focus
on efficient processing of 3D images. For this, it employs parallel pathways that
take as input down-sampled context, avoiding to convolve large volumes at full
resolution to remain computationally cheap. Although originally developed for
1
Implementation and configuration details considered less important for this work
were omitted to avoid cluttering the manuscript.
6
Kamnitsas, Bai, Ferrante, McDonagh, Sinclair, et al.
Fig. 3: We used two DeepMedics [13] in our experiments. The smaller of the two
is depicted, where the number of feature maps and their dimension at every
layer are depicted in the format (N umber × Size). The second model used in
the ensemble is wider, with double the number of feature maps at every layer.
All kernels and feature maps are 3D, even though not depicted for simplicity.
segmentating brain lesions, it was found promising on diverse tasks, such as
segmentation of the placenta [21], making it a good component for a robust
ensemble. We include two deepMedic models in EMMA. The first is the residual
version previously employed in BRATS 2016 [22], depicted in Fig. 3. The second
is a wider variant, with double the number of feature maps at each layer.
Training details: The models are trained by extracting multi-scale image segments with a 50% probability centred on healthy tissue and 50% probability on
tumour as proposed in [13]. The wider variant is trained on larger inputs of width
34 and 22 for the two scales respectively. They are trained with cross-entropy
loss, with all meta-parameters adopted from the original configuration.
3.2
FCN
Model description: We integrate three 3D FCNs [23] in EMMA. A schematic of
the first architecture is depicted in Fig. 4. The second FCN is constructed larger,
replacing each convolutional layer with a residual block with two convolutions.
The third is also residual-based, but with one less down-sampling step. All layers
use batch normalisation, ReLUs and zero-padding.
Training details: We draw training patches of width 64 for the first and 80 voxels
for the residual-based FCNs, with an equal probability from each label. They
were trained using Adam. The first was trained to optimize the IoU loss [19]
while the Dice was used similarly for the other two.
3.3
U-Net
Model description: We employ two 3D versions of the U-Net architecture [24] in
our ensemble. The main elements of the first architecture are depicted in Fig. 5.
In this version we follow the strategy suggested in [25] to reduce model complexity, where skip connections are implemented via summations of the signals in the
up-sampling part of the network, instead of the concatenation originally used.
The second architecture is similar but concatenates the skip connections and uses
EMMA for Robust Brain Tumour Segmentation
7
Fig. 4: Schematic of one of the FCN architecture used in EMMA. Shown are
number of feature maps per layer. All kernels and feature maps are 3D, even
though not depicted for simplicity.
Fig. 5: Schematic of an adapted Unet used in our experiments. Depicted are
number of feature maps per layer. All kernels and feature maps are 3D, even
though not depicted for simplicity.
strided convolutions instead of max pooling. All layers use batch normalisation,
ReLUs and zero-padding.
Training Details: The U-Nets were trained with input patches of size 64×64×64.
The patches were sampled only from within the brain, with equal probability
being centred around a voxel from each of the four labels. They were trained
minimizing cross entropy via AdaDelta and Adam respectively, with different
optimization, regularization and augmentation meta-parameters.
3.4
Ensembling
The above models are all trained completely separately. At testing time, each
model segments individually an unseen image and outputs its class-confidence
maps. The models are then ensembled into EMMA, according to eq. 2. For this,
the ensemble’s confidence maps for each class are created by calculating for each
voxel the average confidence of the individual models for the voxel to belong to
this class. The final segmentation made by the EMMA is performed by assigning
to each voxel the class with the highest confidence.
8
3.5
Kamnitsas, Bai, Ferrante, McDonagh, Sinclair, et al.
Implementation details
The original implementation of DeepMedic was used for the corresponding two
models, along with the default meta-parameters, publicly available on https://
biomedia.doc.ic.ac.uk/software/deepmedic/. The FCNs were implemented
using DLTK, a deep learning library with a focus on medical imaging applications
that allowed quick implementation and experimentation (https://github.com/
DLTK/DLTK). Finally, an adaptation of the Unet will be released on https://
gitlab.com/eferrante.
4
4.1
Evaluation
Material
Our system was evaluated on the data from the Brain Tumour Segmentation
Challenge 2017 (BRATS) [4,26,27,28]. The training set consists of 210 cases with
high grade glioma (HGG) and 75 cases with low grade glioma (LGG), for which
manual segmentations are provided. The segmentations include the following
tumour tissue labels: 1) necrotic core and non enhancing tumour, 2) oedema, 4)
enhancing core. Label 3 is not used. The validation set consists of 46 cases, both
HGG and LGG but the grade is not revealed. Reference segmentations for the
validation set are hidden and evaluation is carried out via an online system that
allows multiple submissions. In the testing phase of the competition, a test set
of 146 cases is provided to the teams, and the teams have a 48 hours window
for a single submission to the system. For evaluation, the 3 predicted labels
are merged into different sets of whole tumour (all labels), the core (labels 1,4)
and the enhancing tumour (label 4). For each subject, four MRI sequences are
available, FLAIR, T1, T1 contrast enhanced (T1ce) and T2. The datasets are
pre-processed by the organisers and provided as skull-stripped, registered to a
common space and resampled to isotropic 1mm3 resolution. Dimensions of each
volume are 240 × 240 × 155.
4.2
Preprocessing: Ensembling intensity normalisation methods
We experimented with three different versions of intensity normalisation as preprocessing: 1) Z-score normalisation of each modality of each case individually,
with the mean and stdev of the brain intensities. 2) Bias field correction followed
by (1). 3) Bias field correction, followed by piece-wise linear normalisation [29],
followed by (1). Preliminary comparisons were inconclusive. We instead chose
to average away the normalisation’s effect with EMMA. Three instances of each
network were trained, each on data processed with different normalisation. They
were applied to correspondingly processed images for inference and all results
were averaged in EMMA (Fig. 6).
EMMA for Robust Brain Tumour Segmentation
9
Fig. 6: Results are affected by normalization. To make a system robust to this
factor, we introduce in EMMA models trained on differently normalized data.
4.3
Results
We provide the results that EMMA achieved on the validation and testing set
of the BRATS’17 challenge2 on Table 1. Our system won the competition by
achieving the overall best performance in the testing phase, based on Dice score
(DSC) and Haussdorf distance. We also show results achieved on the validation
set by the teams that ranked in the next two positions at the testing stage.
No testing-phase metrics are available to us for these methods. We note that
EMMA achieves similar levels of performance on validation and test sets, even
though the latter contains data from different sources, indicating the robustness
of the method. In comparison, competing methods were very good fits for the
validation set, but did not manage to retain the same levels on the testing set.
This emphasizes the importance of research towards robust and reliable systems.
Table 1: Performance of EMMA on the validation and test sets of BRATS 2017
(submission id biomedia1). Our system achieved the top segmentation performance in the testing stage of the competition. For comparison we show the
performance on validation set of the teams that ranked in the next two position.
Performance of other teams in the testing stage is not available to us.
DSC
Sensitivity
Hausdorff 95
Enh. Whole Core Enh. Whole Core Enh. Whole Core #submits
EMMA (val)
UCL-TIG (val)
MIC DKFZ (val)
EMMA (test)
2
73.8
78.6
73.2
72.9
90.1
90.5
89.6
88.6
79.7 78.3
83.8 77.1
79.7 79.0
78.5
-
89.5
91.5
89.6
-
76.2
82.2
78.1
-
4.50
3.28
4.55
36.0
4.23
3.89
6.97
5.01
6.56
6.48
9.48
23.1
2
21
2
1
Leaderboard: https://www.cbica.upenn.edu/BraTS17/lboardValidation.html
10
Kamnitsas, Bai, Ferrante, McDonagh, Sinclair, et al.
Fig. 7: FLAIR, T1ce and manual annotation of a case in the training set, along
with automatic segmentation from preliminary version of EMMA consisting of
six models. Green arrows point inconsistent mistakes by the individual model
that are corrected by the ensembling, while red arrow shows a consistent mistake.
5
Conclusion
Neural networks have been proven very potent, yet imperfect estimators, often
making unpredictable errors. Biomedical applications are reliability-critical however. For this reason we first concentrate on improving robustness. Towards this
goal we introduced EMMA, an ensemble of widely varying CNNs. By combining
a heterogeneous collection of networks we construct a model that is insensitive
to independent failures of CNN components and thus generalises well (Fig. 7).
We also introduced the new perspective of ensembling for objectiveness. By
marginalizing out via ensembling the biased behaviour introduced by configuration choices, EMMA is a model more fit for objective analysis. Even though the
individual networks have straight-forward architectures and were not optimized
for the task, EMMA won the first position in the final testing stage of BRATS
2017 competition among 50+ teams, indicating strong generalisation.
By being robust to suboptimal configurations of its components, EMMA
may offer re-usability on different tasks, which we aim to explore in the future.
EMMA could also be useful in unbiased investigation of factors such as sensitivity
of CNNs to different sources of domain shift that is strongly affecting large-scale
studies [30], or estimating amount of training data required for a task. Finally,
EMMA’s uncertainty could serve as a more objective measure of what type of
patients or tumours are most challenging to learn.
EMMA for Robust Brain Tumour Segmentation
6
11
Acknowledgements
This work is supported by the EPSRC (EP/N023668/1, EP/N024494/1 and
EP/P001009/1) and partially funded under the 7th Framework Programme by
the European Commission (CENTER-TBI: https://www.center-tbi.eu/). KK is
supported by the Presidents PhD Scholarship of Imperial College London. EF
is beneficiary of an AXA Research Fund postdoctoral grant. NP is supported by
Microsoft Research through its PhD Scholarship Programme and the EPSRC
Centre for Doctoral Training in High Performance Embedded and Distributed
Systems (HiPEDS, Grant Reference EP/L016796/1). We gratefully acknowledge
the support of NVIDIA with the donation of GPUs for our research.
References
1. DeAngelis, L.M.: Brain tumors. New Engl. Journ. of Med. 344(2) (2001) 114–123
2. Bauer, S., Wiest, R., Nolte, L.P., Reyes, M.: A survey of mri-based medical image
analysis for brain tumor studies. Physics in med. and biol. 58(13) (2013) R97
3. Louis, D.N., et al.: The 2016 world health organization classification of tumors
of the central nervous system: a summary. Acta neuropathologica 131(6) (2016)
803–820
4. Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J.,
Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor
image segmentation benchmark (brats). IEEE TMI 34(10) (2015) 1993–2024
5. Prastawa, M., Bullitt, E., Ho, S., Gerig, G.: A brain tumor segmentation framework
based on outlier detection. Med. Image Anal. 8(3) (2004) 275–283
6. Gooya, A., Pohl, K.M., Bilello, M., Biros, G., Davatzikos, C.: Joint segmentation
and deformable registration of brain scans guided by a tumor growth model. In:
MICCAI, Springer (2011) 532–540
7. Parisot, S., Duffau, H., Chemouny, S., Paragios, N.: Joint tumor segmentation and
dense deformable registration of brain mr images. In: MICCAI. (2012) 651–658
8. Bakas, S., Zeng, K., Sotiras, A., Rathore, S., Akbari, H., Gaonkar, B., Rozycki,
M., Pati, S., Davatzikos, C.: Glistrboost: combining multimodal mri segmentation,
registration, and biophysical tumor growth modeling with gradient boosting machines for glioma segmentation. In: Brainlesion: Glioma, Multiple Sclerosis, Stroke
and Traumatic Brain Injuries, Springer (2015) 144–155
9. Zikic, D., Glocker, B., Konukoglu, E., Criminisi, A., Demiralp, C., Shotton, J.,
Thomas, O.M., Das, T., Jena, R., Price, S.J.: Decision forests for tissue-specific
segmentation of high-grade gliomas in multi-channel mr. In: MICCAI, Springer
(2012) 369–376
10. Le Folgoc, L., Nori, A.V., Ancha, S., Criminisi, A.: Lifted auto-context forests
for brain tumour segmentation. In: Brainlesion: Glioma, Multiple Sclerosis, Stroke
and Traumatic Brain Injuries, Springer (2016) 171–183
11. Urban, G., Bendszus, M., Hamprecht, F., Kleesiek, J.: Multi-modal brain tumor
segmentation using deep convolutional neural networks. BRATS-MICCAI (2014)
12. Pereira, S., Pinto, A., Alves, V., Silva, C.A.: Brain tumor segmentation using
convolutional neural networks in mri images. IEEE TMI 35(5) (2016) 1240–1251
13. Kamnitsas, K., Ledig, C., Newcombe, V.F., Simpson, J.P., Kane, A.D., Menon,
D.K., Rueckert, D., Glocker, B.: Efficient multi-scale 3D CNN with fully connected
CRF for accurate brain lesion segmentation. Med. Image Anal. 36 (2017) 61–78
12
Kamnitsas, Bai, Ferrante, McDonagh, Sinclair, et al.
14. Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are
universal approximators. Neural networks 2(5) (1989) 359–366
15. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance
dilemma. Neural Networks 4(1) (2008)
16. Kittler, J., Hatef, M., Duin, R.P., Matas, J.: On combining classifiers. IEEE
transactions on pattern analysis and machine intelligence 20(3) (1998) 226–239
17. Breiman, L.: Bagging predictors. Machine learning 24(2) (1996) 123–140
18. Sharkey, A.J., Sharkey, N.E.: Combining diverse neural nets. The Knowledge
Engineering Review 12(3) (1997) 231–247
19. Nowozin, S.: Optimal decisions from probabilistic models: the intersection-overunion case. In: CVPR. (2014) 548–555
20. Kamnitsas, K., Chen, L., Ledig, C., Rueckert, D., Glocker, B.: Multi-scane 3d
convolutional neural networks for lesion segmentation in brain mri. in proc of
ISLES-MICCAI (2015)
21. Alansary, A., Kamnitsas, K., Davidson, A., Khlebnikov, R., Rajchl, M., Malamateniou, C., Rutherford, M., Hajnal, J.V., Glocker, B., Rueckert, D., et al.: Fast fully
automatic segmentation of the human placenta from motion corrupted mri. In:
MICCAI, Springer (2016) 589–597
22. Kamnitsas, K., Ferrante, E., Parisot, S., Ledig, C., Nori, A.V., Criminisi, A., Rueckert, D., Glocker, B.: Deepmedic for brain tumor segmentation. In: Brainlesion:
Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. (2016) 138–149
23. Long, J., et al.: Fully convolutional networks for semantic segmentation. In: CVPR.
(2015) 3431–3440
24. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI, Springer (2015) 234–241
25. Guerrero, R., Qin, C., Oktay, O., Bowles, C., Chen, L., Joules, R., Wolz, R., ValdesHernandez, M., Dickie, D., Wardlaw, J., et al.: White matter hyperintensity and
stroke lesion segmentation and differentiation using convolutional neural networks.
arXiv:1706.00935 (2017)
26. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann,
J., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri
collections with expert segmentation labels and radiomic features. Nature Scientific
Data (2017)
27. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann,
J., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for
the pre-operative scans of the tcga-gbm collection. The Cancer Imaging Archive
(2017)
28. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann,
J., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the
pre-operative scans of the tcga-lgg collection. The Cancer Imaging Archive (2017)
29. Nyúl, L.G., Udupa, J.K., Zhang, X.: New variants of a method of mri scale standardization. IEEE TMI 19(2) (2000) 143–150
30. Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V., Simpson, J., Kane,
A., Menon, D., Nori, A., Criminisi, A., Rueckert, D., et al.: Unsupervised domain
adaptation in brain lesion segmentation with adversarial networks. In: Information
Processing in Medical Imaging, Springer (2017) 597–609
| 2cs.AI
|
arXiv:1802.07045v1 [cs.CV] 20 Feb 2018
Latent RANSAC
Simon Korman
Weizmann Institute of Science, Rehovot, Israel
Roee Litman
General Motors, Hertzliya, Israel
[email protected]
[email protected]
Abstract
We present a method that can evaluate a RANSAC hypothesis in constant time, i.e. independent of the size of
the data. A key observation here is that correct hypotheses
are tightly clustered together in the latent parameter domain. In a manner similar to the generalized Hough transform we seek to find this cluster, only that we need as few
as two votes for a successful detection. Rapidly locating
such pairs of similar hypotheses is made possible by adapting the recent ”Random Grids” range-search technique. We
only perform the usual (costly) hypothesis verification stage
upon the discovery of a close pair of hypotheses. We show
that this event rarely happens for incorrect hypotheses, enabling a significant speedup of the RANSAC pipeline.
The suggested approach is applied and tested on three
robust estimation problems: camera localization, 3D rigid
alignment and 2D-homography estimation. We perform rigorous testing on both synthetic and real datasets, demonstrating an improvement in efficiency without a compromise in accuracy. Furthermore, we achieve state-of-the-art
3D alignment results on the challenging “Redwood” loopclosure challenge.
Figure 1. 3D alignment result of two methods on a pair of fragments from the “Redwood” dataset [11]. Our method (right) produces a correct alignment, even though the putative matches contain a mere 2% of inliers, which is very challenging. On the left,
we see a failure case of the method from [36], even though they
manage to increase inlier rate up to 7%. Runtimes for this example
are 74ms for our method, and 163ms for [36].
to facilitate matching, including ones that are learned. However, once these features are matched, for a parametric
model to be fitted, robust estimation methods like RANSAC
are used to cope with corrupted sets of putative matches.
Geometric models that are commonly amenable to such
a robust estimation process include: 2D-homography, camera localization, the essential and fundamental matrices that
describe epipolar constraints between images, rigid 3D motion and more.
1.1. Background and prior art
1. Introduction
Consensus maximization has proven a useful robust estimation approach to solving a wide variety of fitting and
alignment problems in computer vision.
Research in this field can be broadly divided into global
and local optimization methods. Global methods [27, 35,
10, 8] use different strategies to explore the entire solution space enjoy the advantage of having a deterministic
nature. Our method, however, belongs to the family of local methods which are typically extremely fast randomized
algorithms, potentially equipped with probabilistic success
guarantees.
While the proposed method is presented in the context
of RANSAC, it is closely related-to and inspired-by other
works in the field, such as Hough voting. We cover these
topics briefly.
Despite the recent success of (deep-) learning based
methods in computer vision, numerous applications still use
“old-fashioned” robust estimation methods for model fitting, such as RANSAC [19]. This is especially true for
problems of a strong geometric nature such as image alignment, camera localization and 3D reconstruction. Robust
estimation methods of these types largely follow the “hypothesize and test” paradigm which has strong roots in
statistics, and are highly attractive due to their ability to fit
a model to data that is highly corrupted with outliers. Additionally, they have been successfully applied to many problems in computer vision and robotics achieving real time
performance.
As an example, in the field of image (or shape) alignment, novel features and descriptors have been introduced
1
generate
hypothesis
start
stop
criterion
Latent
RANSAC
verify
hypothesis
"vanilla" RANSAC
parameterize
in latent space
hash using
random grid
found
collision
no
collision
Figure 2. A flow chart of RANSAC compared to the suggested method. We propose an alternative flow (in blue), in which after a
hypothesis is generated it first undergoes a hash procedure, and only verified if a (valid) collision is detected. The stop criterion has to be
modified as well, to ensure a second good hypothesis is drawn with high probability.
RANdom SAmple Consensus (RANSAC) [19] is one of
the de-facto golden standards for solving robust estimation
problems in a practical manner. Under this paradigm, the
space of solutions is explored by repeatedly selecting random minimal subsets of a set of given measurements (e.g.
putative matches), for which a model hypothesis is fitted.
These hypotheses are verified by counting the measurements that agree with them up to a predefined tolerance.
This process is repeated until a desired probability to draw
at-least one pure set of inliers is achieved.
Several techniques that extend the ‘naive’ RANSAC procedure described above, are covered in a recent comprehensive survey by Raguram et al. [28]. This survey also suggests USAC – a framework that combines some RANSAC
extensions, yielding excellent results on a variety of problems in terms of accuracy, efficiency, and stability. Two
notable extensions are PROSAC [12], which incorporates
prior knowledge on the measurements by prioritizing their
sampling accordingly, as well as LO-RANSAC [14] which
performs efficient local optimization on a model that has
been verified.
While the previous extensions can be seamlessly applied
along with the suggested method, the following extension
is similar in nature to ours in that it aims to speed up the
verification step, but it does so in a very different manner.
The Sequential Probability Ratio Test (SPRT) [13] extension, adopted by USAC, is based on Wald’s sequential test
[32]. It attempts to reject a “bad” model with high probability, after inspecting only a small fraction of the measurements. While the test is theoretically solid, it relies on two
parameters that are assumed to be known a priori, and in
practice need to be adaptively estimated at runtime. It is
reported to have achieved an improvement of 20% in evaluation time compared to the simpler bail-out test [9].
Generalized Hough transform (GHT) originated from
an algorithm for line detection in images [21], which was
later generalized to handle arbitrary shapes [18, 7]. The key
idea behind this method is that partial observations of the
model are casted as votes into a (quantized) solution space,
in which the object can be detected as a mode (the location
with the most votes). In practice, GHT has not been shown
to scale well to solution spaces of high dimensionality (i.e.
higher than 3), and typically requires numerous votes for a
mode to be accurately detected.
Between RANSAC and GHT. Some works bare resemblance to both of the mentioned approaches. Our work can
be seen as one of these: While it fits naturally into the
RANSAC pipeline, it has some similarities to GHT in the
sense that it seeks to find the mode in the parameter domain,
only that it needs as few as two votes to detect it.
A method by Den-Hollander et al. [15] also lays somewhere between RANSAC and GHT: To increase the probability of obtaining a pure set of inlier matches, a subminimal set is drawn. The remaining degrees of freedom (DoF) are resolved using a voting scheme in a lowdimensional setting. As with all Hough-like methods,
an adequate parameterization of the remaining DoF is required. The authors of [15] provide such a parameterization
for the problem of fundamental matrix estimation.
Our method bares a strong resemblance to the “Randomized Hough Transform” (RHT) [33] of Xu et al. in that
a vote is casted into a single cell in the solution domain,
generated from a randomly selected minimal set. However,
unlike [33], we deal with a hypothesis in constant time and
space, rather than logarithmic, thanks to the Random Grid
hashing mechanism that we adapt. In addition, while RHT
deals with robust curve-fitting (of up to 3 dimensions), we
successfully apply our method on a variety of problem domains of higher dimensionality (up to 8 dimensions).
1.2. Contributions
The main novelty of the presented method is its ability to
handle RANSAC hypotheses in constant time, regardless of
the number of measurements (e.g. matches). We show that
it is beneficial to handle hypotheses in the latent space, due
to an efficient parametrization and hashing scheme that we
devise, which can quickly filter candidate hypotheses until
a pair of correct ones are drawn. While this approach comes
at the expense of a small increase in the number of hypotheses to be examined, it allows for a significant speedup of the
RANSAC pipeline.
The new proposed modifications to RANSAC are accompanied by a rigorous analysis which results in an updated stopping criterion and a well understood overall probability of success. Finally, we validate our method using
challenging data in the problems of 2D-homography estimation, 2D-3D based camera localization and rigid-3D
alignment, showing state-of-the-art results.
2. Method
The ‘vanilla’ RANSAC pipeline can be divided into
three main components: hypothesis generation, hypothesis
verification and the adaptive stopping mechanism. The proposed Latent-RANSAC hypothesis handling fits naturally
into the aforementioned pipeline, as can be seen in Figure 2,
highlighted in blue. The additional modules we propose act
as a ‘filter’ that avoids the need to verify the vast majority of generated hypotheses: Instead of verifying each hypothesis by applying it on all of the matches (a costly process that takes time linear in the number of matches), we
check in constant time if a previously generated hypothesis
‘collides’ with the current one, i.e. whether they are close
enough (in a sense that will be clarified below). Only the
very few hypotheses that pass this filtering stage progress to
the verification stage for further processing. As a result of
the proposed change, the RANSAC stopping criterion needs
to modified to guarantee a probability of encountering a second good hypothesis rather than just one.
screening. Since our approach performs the majority of
the screening according to some ’similarity’ in the space of
transformations, we seek a parametrization of the transformation space in which distances between transformations
can be defined explicitly. More formally, we define such a
parametrization by an embedding hypotheses into some λdimensional space Rλ , which we call the latent space1 . We
consider the distance between transformations to be given
by the `∞ metric between the embedded (or latent) vectors
(λ-tuples). Our goal is to use an embedding in which the
distance between any pair of hypotheses f1 and f2 is tightly
related to the difference in the way these hypotheses act on
matches in the source domain, i.e. to the difference in magnitudes of their residual errors on the matches. Ideally, for
any set of matches M ,
kf1 − f2 k∞ ∝ max |err(f1 , m) − err(f2 , m)|.
m∈M
2D homography. We describe here the parameterization
we use for the space of 2D homographies, which are given
by projective matrices in R3×3 . Following previous works
(e.g. [24, 16]), we use the 4pt parameterization [6] that represents a 2D-homography H ∈ R3×3 by an 8-tuple vH ,
defined by the coordinates in the target image that are the
result of applying H on the four corners of the source image, as illustrated in Figure 3.
w
(x3 , y3 )
(h, 0)
Outline We begin by covering the 3 key components of
our method: parametrization of the solution space (Section 2.1), Random Grids hashing (Section 2.2) and the modified stopping criterion (Section 2.3). We conclude this part
of the paper in Section 2.4, with an analysis of our Random
Grids hashing process.
Preliminary definitions In our setup, the goal is to robustly fit a geometric model (transform) to a set of matches
(correspondences), w.l.o.g. in Euclidean space, where a
match m = (p, q) is an ordered pair of points p ∈ Rd and
0
0
q ∈ Rd . For a geometric transform f : Rd → Rd and
match m = (p, q) the residual error of the match m with
0
respect to f is the Euclidean distance in Rd given by:
err(f, m) = kq − f (p)k.
(1)
Given a set of matches M = {mi } and a tolerance t, the inlier rate achieved by a transform f is defined as the fraction
of matches mi ∈ M for which err(f, mi ) ≤ t. We denote
the maximal inlier rate for a match-set M by ω.
2.1. Parametrization of the solution domain
In the RANSAC pipeline, matches are used both for the
generation of hypothesis candidates, as well as for their
(2)
(h, w)
(x2 , y2 )
H
h
(0, 0)
(0, w)
source image
(x4 , y4 )
(x1 , y1 )
target image
Figure 3. Illustration of the 4pt homography parametrization
[6]. A homography H is represented in the latent space by mapping the location of the four corners of the source image onto the
target image, resulting in the 8-tuple vH = (x1 , y1 , . . . , x4 , y4 ).
As was noted in [24], this parametrization has the key
property that the difference between match errors of two
well-behaved homographies is bounded by the `∞ distance
between their 4pt representations.
The special Euclidean group SE(3), used to describe
rigid motion in R3 , will be used here to solve the problems of Perspective-n-Point (PnP) estimation and Rigid 3D
alignment. We follow a parametrization that was suggested
and used in a line of works of Li et al. [34, 8]. The group
SE(3) can be described as the product between two subgroups SE(3) = SO(3) × R3 , namely 3D translations and
1 the latent space dimension λ typically being the number of degrees of
freedom of the transformation space.
the special orthogonal group (3D rotations). Each of these
3-dimensional sub-groups is parameterized as a 3-tuple, resulting in a 6-tuple representation (r, t) defined as follows:
The axis-angle vector (3-tuple) r represents the 3D rotation
matrix given by Rr = exp([r]x ), where exp(·) is the matrix exponential and [·]x denotes the skew-symmetric matrix
representation. Such vectors r reside in the radius-π ball
that is contained in the 3D cube [−π, π]3 . The translation
3-tuple t is a vector in the cube [−ξ, ξ]3 that contains the
relevant bounded range of translations for a large enough ξ.
Similar to the case of the 2D-homography parametrization, it is proved in [34] that the difference between match
errors of two rigid motions is bounded by the `2 distance
between their parametrization.
input: (incremental) A candidate transform (matrix)f
parameters: number of tables L; tolerance t; cell
dim. c; parametrization dim. λ;
initialization:
foreach i = 1, . . . , L do
1. Initialize an empty hash table Ti .
2. Randomize offset Oi ∼ U ([0, c]λ )
end
insertion and collision check for hypothesis f :
foreach i = 1, . . . , L do
1. Let v be the embedding of f
2. The hash index for v is: τv = hash
2.2. Random Grids hashing
Given such a embedding of a generated hypotheses, the
heart of our method boils down to a nearest neighbor query
search of the current vector through all vectors representing
previously generated hypotheses. More precisely, the task
needed to be performed is a range search query for vectors
that are at a distance of up to a certain tolerance t.
A recent work of Aiger et al. [2] turns out to be extremely suitable for this task. They propose Random Grids
- a randomized hashing method based on a very simple idea
of imposing randomly shifted ‘grids’ over the vector space,
checking for vectors that ‘collide’ in a common cell. The
Random Grids algorithm is very fast, and simple to implement - even in comparison with the closely related LSHbased algorithms [4], since the grid is axis aligned and it
is uniform (consists of cells in Rd with equal side length).
Most important, and essential for the speed of our method,
is that the range search is done in constant time (i.e. it does
not depend on the number of vectors searched against), as
opposed to the RANSAC hypothesis validation that requires
applying the model and measuring errors on (typically hundreds of) point matches or even the logarithmic-time solution proposed in RHT [33].
Hashing scheme. We are given a representation of the
transform f as a vector v ∈ Rλ (for a λ-dimensional parameterization). In the Random Grids [2] setting, we hash
v into L hash tables {Ti }L
i=1 , each associated with an independent random grid, which is defined by a uniform random
shift Oi ∼ U ([0, c]λ ), where c is the cell side length and λ
is the dimension of the latent vector v. The cell index for
v in the table Ti isobtained by concatenating the integer
i
into a single scalar (where b·c means
vector zi = v+O
c
“floor” operation). The entire hashing process - initialization, insertion and collision checking, is given in detail in
Algorithm 1.
v+Oi
c
3. If the cell Ti [τv ] is occupied by a vector u, report
a collision of f if kv − uk∞ < t
4. Store v in Ti [τv ]
end
Algorithm 1: Latent-RANSAC hypothesis handling.
2.3. Latent-RANSAC stopping criterion
The classical analysis of RANSAC provides a simple
formula for the number of iterations n required to reach a
certain success probability p0 (e.g. 0.99). It is based on the
assumption that it is sufficient to have a single ‘good’ iteration in which a pure set of inliers is drawn. Note that this
assumption is made for the simplicity of the analysis and is
only theoretical, since it ignores e.g. the presence of inlier
noise and several possible degeneracies in the data.
Formally, let Gn be the random variable that counts the
number of such good iterations out of n attempts. For a
minimal set of size γ and data with inlier rate of ω, it holds
that
p0 = 1 − P [Gn = 0] = 1 − (1 − p)n
(3)
where p = ω γ . The number of iterations n required to guarantee a desired success probability p0 is therefore:
n=
log(1 − p0 )
log(1 − p)
(4)
A similar simplified analysis can be applied to the
Latent-RANSAC scheme. Ignoring the presence of inlier
noise, the existence of (at least) two ‘good’ iterations is
needed for a collision to be detected and the algorithm to
succeed. Therefore, by the binomial distribution we have
that
p0 = P [Gn ≥ 2] = 1 − (1 − p)n − n·p·(1 − p)n−1 . (5)
Based on equations (4) and (5), we plot in Figure 4 the ratio between the number of required iterations n in the case
2.5
iteration ratio
p0 = 0.999
2
p0 = 0.99
p0 = 0.95
1.5
1
-1
0
inlier rate ω
10
10
Figure 4. The ratio between stopping criterions of LatentRANSAC (5) and of RANSAC (4). Ratios are shown as a function
of inlier rate ω for several success probabilities (color coded) and
for γ (minimal sample size) values of 3 (dashed) and 4 (solid). See
text for details.
of Latent-RANSAC versus the case of RANSAC. The ratio is given as a function of the inlier rate ω, at 3 different
success rates p0 (color coded), for the two different cases
γ = 3 (e.g. in Rigid 3D motion estimation) and γ = 4 (e.g.
in homography estimation). Interestingly, the ratio attains
a small value (less than 2) for inlier rates below ω = 0.95,
and converges to small constant values as the inlier rate decreases. The very high inlier rates for which the ratio is
large are of no concern, since the absolute number n is extremely low in this range.
In the next section, as part of an analysis of the Random
Grid hashing, we derive a more realistic stopping criterion
that depends also on the success probability of the Random
Grid based collision detection, which clearly depends on the
inlier noise level.
2.4. Random Grid analysis
We cover two aspects of Random Grids. First, we extend the stopping criterion from Section 2.3 to consider the
probability that a colliding pair of good hypotheses will be
detected. Next, we discuss causes of false collision detection, which can have an affect on the algorithm runtime.
stopping criterion. Let R(i) be the event that the random
grid component succeeds (detects a collision), given i good
iterations out of a total of n. We can now update Equation (5), taking this success probability into account:
X
p0 =
P [Gn = i]·P [R(i)] ≥ P [Gn ≥ 2]·P [R(2)] (6)
i≥2
where the inequality holds due to the fact that P [R(i)]
monotonically increases with i.
A final lower bound on p0 (from which the stopping criterion is determined) can be obtained by substituting the
expression for P [Gn ≥ 2] from (5) into (6) together with
a lower bound on P [R(2)] which we provide next.
Recall that R(2) is the event that the random grid hashing
succeeds given that two successful hypotheses were gener-
ated. We will, more explicitly, denote this event by RL (2),
for a random grid that uses L hash tables.
The analysis in [2] is rather involved since it deals with
the Euclidean `2 distance. Using `∞ distances we are able
to derive the success probability of finding a true collision
in our setup, as a function of the random grid parameters,
in a simpler manner. Assuming a tolerance t in the latent
space, determined by (inlier) noise level of the data, using a
random grid with cell dimension c and a single table results
in
γ
t
(7)
P [R1 (2)] ≥ 1 −
c
since a pair of pure-inlier transformations (which differ by
at most t) must share the same independently offsetted bin
indices in each of the γ dimensions.
Finally, using L hash tables, randomly and independently generated, we obtain:
L
P [R(2)] = P [RL (2)] ≥ 1 − (1 − P [R1 (2)])
(8)
False collisions. We now discuss the expected number of
false collisions that are found by the hashing scheme. It is
important to understand why false collision might happen,
as they have an effect on the overall runtime of our pipeline.
Recall that n is the overall number of iterations of the
pipeline, and hence it is also the total number of samples
inserted into each hash table. There are two kinds of false
collisions to consider. The first kind happens due to the
fact that the random grid cell size c might be larger than
the tolerance t. Following the recommendation in [1] we
set the cell size c to be not much larger than the tolerance
t, resulting in a small number of such false collisions. In
any case, this kind of collision has a small impact on the
runtime, since it will be filtered by the tolerance test (step 3
in Algorithm 1) at constant time cost.
The second kind of false collision is one that passes the
tolerance test (step 3). Since it is not the true model, it is
associated with some inlier rate ζ. If ζ ω, the probability
of this collision appearing before we have reached the stopping criterion is negligible. Empirically, we observe very
few (typically less than 15) collisions that pass the tolerance test up to the stopping of the algorithm. These are the
only kind of collisions that incur a non-negligible penalty
(in runtime only) since they invoke the verification process
that every “vanilla” RANSAC hypothesis goes through.
3. Results
In order to evaluate our method, we performed extensive
tests on both real and synthetic data. The Latent-Ransac algorithm is applied to the problems of 2D-homography estimation (Section 3.1), Perspective-n-Point (PnP) estimation
(Section 3.2) and Rigid 3D alignment (Section 3.3). It is
compared with ordinary RANSAC, with or without the well
portion of the data
1
0.8
0.6
0.4
[11]
Redwood
Redwood
ZuBuD
[30]
ZuBuD
0.2
0
0
10
20
30
40
50
60
70
inlier rate [%]
Figure 5. Inlier rate cumulative distribution (CDF) of the two
real data sets we use. The dashed curve was taken only over Redwood pairs with provided ground truth.
known SPRT [13] extension, which is a very different technique for accelerating RANSAC’s model verification phase.
Implementation details. Our method naturally extends
the standard RANSAC pipeline, according to the changes
highlighted in Figure 2. Our implementation (for which we
use the shorthand LR) extends the excellent C++ RANSAC
implementation USAC [28], with the noted changes in
the specific modules. This enables an easy way to compare with a state-of-the-art RANSAC implementation, and
allows our method to enjoy the same extensions used
by USAC (such as its local optimization component LORANSAC [14]). In addition, their implementation includes
the SPRT [13] extension, the most commonly used acceleration of RANSAC’s model verification phase. We use the
shorthand SPRT to refer to RANSAC using this extension.
In addition we use the OpenGV library [22] for PnP
model fitting (Kneip’s P3P algorithm [23]) and for Rigid 3D
model fitting (Arun’s algorithm [5]). We make our modifications to the code available at anonymized.
Parameter were tuned by a simple coordinate descent on
a subset of the data and kept fixed throughout. Parameters
common to all settings: probability of success p0 = 0.99,
number of hash tables L = 4; Random Grid cell size
c = 1.8t, where the tolerance t (and the RANSAC threshold
used) are specified separately for each experiment; Maximal number of iterations n = 5 × 106 ; Hash table size
of n/10, resulting in addressing indices of 19 bits (see the
supplementary materials for hash table implementation details).
3.1. 2D-homography estimation
We create a large body of 2D-homography estimation instances using the Zurich buildings data-set [30]. The dataset consists of sequences of 5 snapshots taken from different street-level viewpoints for 201 buildings or houses. The
images are typically dominated by planar facade structures
and hence each pair of images in a sequence is related by
a 2D homography (or perhaps more than one in the case of
several planes).
We computed SIFT [25] features for each image and created sets of corresponding features for each of the 10 (ordered) pairs of images in a sequence using the VLFeat library [31]. Then, we ran both RANSAC and LR on each
pair 10 times and saved the highest detected inlier rate of
any of the fitted homographies as the ’optimal’ inlier rate for
the image pair. A small set of image pairs (168 out of 2010)
with very low ω was manually removed from the evaluation
since the inlier feature locations did not reside on an actual
single plane in the scene, and were very noisy.
The 1842 resulting matching instances are challenging:
many pairs have low inlier rates, that result from (i) the
planar area of interest typically covering only part of each
image; (ii) large viewpoint changes; (iii) large presence of
repetitive patterns (e.g. windows or pillars). See Figure 5
for the distribution of inlier rates for this data-set.
We ran LR with a tolerance t of 70 pixels in the latent
domain and RANSAC and SPRT with a threshold of 5 pixels. These parameters were chosen to give the best results,
which are summarized in Table 1. We arrange the image
pairs into four groups according to their ’optimal’ inlier rate
(defined above), and the size of each group is shown at the
bottom of the table. For each group we report the average,
and 95-percentile of runtimes for each method. We also report each method’s success rate (averaged over all pairs in
the group), which is the ratio between the detected inlierrate and the ’optimal’ inlier rate (i.e. the highest one from
any method).
measure
method
inlier rate range (in %)
0-10
10-20 20-40 40-100
RANSAC 1,536 (6,627) 56 (153)
SPRT
1,157 (4,994) 34 (95)
avg. (95%)
LR
1,102 (4,559) 38 (92)
runtime
success
RANSAC
SPRT
LR
# of pairs
8 (16)
6 (11)
8 (13)
8 (15)
9 (16)
12 (18)
98.01%
97.60%
97.50%
98.20%
98.39%
98.47%
98.75%
98.78%
98.90%
99.32%
99.50%
99.65%
210
390
650
592
Table 1. 2D Homography fitting on Zurich Buildings [30]. Best
results are shown in bold. See text for further details.
As can be seen, SPRT and LR (modestly) accelerate
RANSAC at the harder inlier rate ranges, where the overall
runtime is longer. LR achieves the best acceleration for the
lowest range, while SPRT does better in the 20-40 range.
In terms of accuracy - both SPRT and LR are in par with
RANSAC, achieving high success rates for the different inlier rate groups.
The detailed runtime breakdowns shown in Figure 6
(left) reveals two important points that should be made here.
First, in homography estimation, methods that accelerate
the RANSAC evaluation stage (i.e. LR and SPRT) have
LR
SPRT
RANSAC
0
0.05
match drawing
0.1
0
0.15
model fitting
2
hashing
4
6
model verification
8
10
12
local optimization
Figure 6. Runtime breakdown per pipeline module comparing LR to SPRT and RANSAC. These are average per-instance runtimes
(in seconds) for 2D-homography estimation (left) and PnP estimation (right) taken over each of the entire data-sets used for the evaluation.
a relatively small potential improvement gap since the runtime of RANSAC-based homography estimation is dominated by the model fitting stage (this is not the case for the
other problems we deal with, as will be seen later). Second,
the improved acceleration of LR in the lower ranges is significant when considering the overall time taken to fit the
entire data-set, since the majority of time is spent on these
difficult cases which are surprisingly not rare (11% and 21%
of all pairs are in the 0-10 and 10-20 ranges respectively).
3.2. Perspective-n-Point (PnP) estimation
In order to compare methods on the PnP problem (camera localization from n 2D-3D correspondences), we follow
the benchmark proposed in [22], using a large bank of synthetic instances with controlled values of inlier-rate ω and
inlier-noise σ. Specifically, a synthetic inlier set is created
by first picking a random camera position in 3D, followed
by projecting random 3D locations (within the camera view)
to 2D angular coordinates, followed by adding noise σ. The
1 − ω fraction of outlier matches are generated similarly,
except that the 2D locations are distributed uniformly in the
image. Each instance contains 1, 000 such 2D-3D correspondences, generated in this manner.
We chose to use Kneip’s algorithm [23] for minimal
sample P3P model fitting (for all methods) due to its good
accuracy-efficiency trade off compared to other alternatives.
Since only 3 matches are needed for the minimal sample fitting problem, RANSAC generally needs very few iterations
to robustly reveal the true underlying model in typical cases.
We therefore add noise at levels of 0.5◦ to 2◦ and consider
challenging inlier rates in the range between 3% and 20%.
We ran RANSAC and SPRT with a threshold of 4σ radians, LR with a tolerance t of 7σ radians (in the latent domain), and the translation-to-angle ratio of the embedding
cm
. The results, summarized in table 2, show that
was 2.1 rad
LR achieve more than an order of magnitude acceleration
compared to RANSAC, at only a minor loss in accuracy in
all configurations. SPRT achieves similar acceleration factor at the higher inlier rates (10% and 20%). It is not as
efficient at the lower range (3% and 5%) and much less accurate at the lowest inlier rate of 3%.
For the PnP case, Figure 6 (right) reveals that the huge
acceleration happens due to the vast majority (over 95%)
of time spent by RANSAC on model verification, which is
practically eliminated by LR. This is possible mainly due
to the existence of fast fitting algorithms (e.g. [23]) for the
P3P problem.
noise level method
3
inlier rate (%)
5
10
20
0.5◦
RANSAC 34,213±3,800 7,109±1,560 887±106 114±20
SPRT 60,030±1,639 874±235 71±14 10±1
LR
2,404±733
439±289 60±12 13±3
1◦
RANSAC 34,185±4,479 7,160±1,218 894±196 115±28
SPRT 61,194±20,602 954±323 75±18 10±2
LR
2,374±452
588±208 63±18 12±2
2◦
RANSAC 30,880±4,434 7302±1039 898±104 118±28
SPRT 57,461±24,507 1,002±188 81±11 12±3
LR
2,635±886
554±176 74±19 14±5
success
RANSAC
SPRT
LR
101.11%
28.94%
93.06%
99.03%
100.00%
98.17%
98.12% 97.73%
98.17% 97.62%
97.63% 96.95%
Table 2. Results for PnP estimation on synthetic data. Runtimes are given in median ± std format, taken over the 20 random instances used for each noise-level/inlier-rate combination.
Similarly, average success rates are reported (for each instance we
count the number of detected inliers as a percentage of the true
number of inliers).
3.3. Rigid 3D alignment
To evaluate the Rigid 3D alignment application of
Latent-RANSAC, we use the registration challenge of the
recent “Redwood” benchmark proposed by Choi et al. [11].
This dataset was generated from four synthetic 3D scenes,
each divided into 52 point-cloud fragments on average.
While from synthetic origin, these fragments contain highfrequency noise and low-frequency distortion that simulate
scans created by consumer depth cameras.
The challenge is to perform global 3D registration between every pair of fragments of a given scene, in order to
provide candidate pairs for trajectory loop closure. A correctly ‘detected’ pair is one for which the point clouds overlap by at least 30% and the reported transformation is sufficiently accurate (see [11] for details). The main goal in this
benchmark, as stated by [11], is to achieve high recall while
relying on a post-process to later remove false-matches.
Aside from the benchmark, Choi et al. [11] present a
simple extension (CZK) to the Point-Cloud-Library (PCL)
0.25
precision
0.2
0.15
0.1
0.05
0
0
0.1
0.2
openCV
[17] OpenCV
4pcs
[3] 4PCS
[36]FGR
FGR
[11]CZK
CZK
super4pcs
[26] super4PCS
pcl
[20] PCL
LR
LR
LR(strict)
(strict)
LR
0.3
0.4
0.5
0.6
0.7
0.8
recall
Figure 7. Performance on the “redwood” benchmark [11].
Our method achieves state-of-the-art recall in the standard setting
(marked by a red ‘x’), while using a stricter threshold (marked
by a red ‘+’) dominates all previous result in both precision and
recall. See description in the text for further details.
[20] implementation of [29]. The method of CZK showed
state-of-the-art performance, while comparing to previous
methods like OpenCV [17], 4PCS [3] and its extension super4PCS [26]. Fast Global Registration (FGR) [36] is a recent novel optimization process presented by Zhou et al.,
which achieves an order of magnitude runtime acceleration on this dataset, at a competitive recall-precision performance. They perform the costly nearest-neighbor (NN)
search only once (unlike previous methods which use them
in their inner loop), while introducing several fast and simple methods to filter false matches.
We chose to follow [11, 20, 36] and feed our framework
with putative matches based on FPFH features [29], using a
tolerance t of 24cm in the latent space, and translation-tocm
angle ratio of 4.8 rad
for the embedding. Inspired by FGR
we perform the NN search only once. We then only apply one single filter (also used in [11, 36]), an approximate
congruency validation on each minimal sample drawn.
Figure 7 shows a comparison of Latent RANSAC to
other results reported in [36]. Our method clearly achieves
the highest recall value (the main goal), at a precision
slightly below that of CZK. Furthermore, we are able to
dominate all previous results in precision and recall simultaneously by using a slightly stricter setting, when reporting
only pairs with overlap of over 40% rather than 30%.
We attribute our high performance mainly to the fact
that we perform almost no filtering to the putative matches,
such as the bidirectional search and tuple filtering done in
[36], normal-agreement in [11, 20] or the drawing of a nonminimal set of 4 matches in [11]. Using this “naive” nearest neighbor FPFH feature matching avoids filtering of true
correspondences (enabling higher recall), but this comes at
the cost of some very low inlier-rates, as can be seen in Figure 5. Our algorithm is able to deal with such inlier rates
successfully (and efficiently), as was shown in the other ex-
periments of this section and in Figure 1.
Another attractive property of our method in this benchmark is its runtime, presented in Table 3. Our runtime is
close to that of FGR, which we outperform significantly in
terms of recall. Note, however, that our method is actually
faster than FGR whenever the inlier rate is above 5%, as the
number of iterations given by (5) is very low. Verification
consumes a considerable part of the runtime of methods like
[11, 20], while we perform the costly overlap verification
only upon the detection of a collision (13.3 times per run
on average). Additionally, we perform overlap calculation
only once as done in FGR.
method
PCL [20]
avg. time (sec)
3.8
CZK [11]
7.5
FGR [36]
0.21
LR
0.40
Table 3. Average runtimes on the “redwood dataset”, excluding normals and FPFH [29] calculation time which are 24ms and
300ms on average, respectively. A breakdown of our method’s
timing includes 84ms for feature matching, 305ms for the latent
RANSAC pipeline, and 13ms for overlap calculation.
4. Future work
In this work we presented Latent-RANSAC: a novel
speed-up of the hypothesis handling stage of the RANSAC
pipeline. We have shown its advantages on challenging
matching problems, that include very low inlier rates, in
the domains of homography estimation, camera localization
and rigid 3D motion estimation.
Latent-RANSAC has the potential to be extended to additional domains. Of particular interest is finding an appropriate parametrization of the more challenging fundamental matrix domain, which is classically tackled using
RANSAC.
The good results that Latent-RANSAC achieves on the
”Redwood” benchmark come to show the advantage of being able to handle highly corrupted ”raw”s data (over 60%
of the fragment pairs have under 10% inlier rate). This is
since the alternative of filtering the data to reduce the rate
of outliers comes at the risk of loss of informative data. The
challenge, however, remains to do so efficiently, especially
for search spaces of high dimensionality.
References
[1] D. Aiger, H. Kaplan, and M. Sharir. Reporting neighbors in
high-dimensional euclidean space. SIAM Journal on Computing, 43(4):1363–1395, 2014. 5
[2] D. Aiger, E. Kokiopoulou, and E. Rivlin. Random grids:
Fast approximate nearest neighbors and range searching for
image search. In Proceedings of the IEEE International Conference on Computer Vision, pages 3471–3478, 2013. 4, 5
[3] D. Aiger, N. J. Mitra, and D. Cohen-Or. 4-points congruent
sets for robust pairwise surface registration. ACM Transactions on Graphics (TOG), 27(3):85, 2008. 8
[4] A. Andoni and P. Indyk. Near-optimal hashing algorithms for
approximate nearest neighbor in high dimensions. In Foundations of Computer Science, 2006. FOCS’06. 47th Annual
IEEE Symposium on, pages 459–468. IEEE, 2006. 4
[5] K. S. Arun, T. S. Huang, and S. D. Blostein. Least-squares
fitting of two 3-d point sets. IEEE Transactions on pattern
analysis and machine intelligence, (5):698–700, 1987. 6
[6] S. Baker, A. Datta, and T. Kanade. Parameterizing homographies. Technical Report CMU-RI-TR-06-11, 2006. 3
[7] D. H. Ballard. Generalizing the hough transform to detect
arbitrary shapes. Pattern recognition, 13(2):111–122, 1981.
2
[8] D. Campbell, L. Petersson, L. Kneip, and H. Li. Globallyoptimal inlier set maximisation for simultaneous camera
pose and feature correspondence. In The IEEE International
Conference on Computer Vision (ICCV), Oct 2017. 1, 3
[9] D. P. Capel. An effective bail-out test for ransac consensus
scoring. In BMVC, 2005. 2
[10] T.-J. Chin, P. Purkait, A. Eriksson, and D. Suter. Efficient
globally optimal consensus maximisation with tree search.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 2413–2421, 2015. 1
[11] S. Choi, Q.-Y. Zhou, and V. Koltun. Robust reconstruction
of indoor scenes. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2015. 1, 6, 7, 8
[12] O. Chum and J. Matas. Matching with prosac-progressive
sample consensus. In Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference
on, volume 1, pages 220–226. IEEE, 2005. 2
[13] O. Chum and J. Matas. Optimal randomized ransac. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
30(8):1472–1482, 2008. 2, 6
[14] O. Chum, J. Matas, and J. Kittler. Locally optimized ransac.
In Pattern Recognition, pages 236–243. Springer, 2003. 2, 6
[15] R. J. Den Holl and E. A. Hanjalic. A combined ransac-hough
transform algorithm for fundamental matrix estimation. In in
18th British Machine Vision Conference. University of. Citeseer, 2007. 2
[16] D. DeTone, T. Malisiewicz, and A. Rabinovich. Deep image
homography estimation. arXiv preprint arXiv:1606.03798,
2016. 3
[17] B. Drost, M. Ulrich, N. Navab, and S. Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In Computer Vision and Pattern Recognition (CVPR),
2010 IEEE Conference on, pages 998–1005. Ieee, 2010. 8
[18] R. O. Duda and P. E. Hart. Use of the hough transformation
to detect lines and curves in pictures. Communications of the
ACM, 15(1):11–15, 1972. 2
[19] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image
analysis and automated cartography. Communications of the
ACM, 24(6):381–395, 1981. 1
[20] D. Holz, A. E. Ichim, F. Tombari, R. B. Rusu, and S. Behnke.
Registration with the point cloud library: A modular framework for aligning in 3-d. IEEE Robotics & Automation Magazine, 22(4):110–124, 2015. 8
[21] P. V. Hough. Machine analysis of bubble chamber pictures.
In International conference on high energy accelerators and
instrumentation, volume 73, page 2, 1959. 2
[22] L. Kneip and P. Furgale. Opengv: A unified and generalized approach to real-time calibrated geometric vision. In
Robotics and Automation (ICRA), 2014 IEEE International
Conference on, pages 1–8. IEEE, 2014. 6, 7
[23] L. Kneip, D. Scaramuzza, and R. Siegwart. A novel
parametrization of the perspective-three-point problem for a
direct computation of absolute camera position and orientation. In Computer Vision and Pattern Recognition (CVPR),
2011 IEEE Conference on, pages 2969–2976. IEEE, 2011.
6, 7
[24] R. Litman, S. Korman, A. Bronstein, and S. Avidan. Inverting ransac: Global model detection via inlier rate estimation.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 5243–5251, 2015. 3
[25] D. G. Lowe.
Distinctive image features from scaleinvariant keypoints. International journal of computer vision, 60(2):91–110, 2004. 6
[26] N. Mellado, D. Aiger, and N. J. Mitra. Super 4pcs fast
global pointcloud registration via smart indexing. In Computer Graphics Forum, volume 33, pages 205–215. Wiley
Online Library, 2014. 8
[27] C. Olsson, O. Enqvist, and F. Kahl. A polynomial-time
bound for matching and registration with outliers. In Computer Vision and Pattern Recognition, 2008. CVPR 2008.
IEEE Conference on, pages 1–8. IEEE, 2008. 1
[28] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M.
Frahm. Usac: a universal framework for random sample consensus. IEEE transactions on pattern analysis and machine
intelligence, 35(8):2022–2038, 2013. 2, 6
[29] R. B. Rusu, N. Blodow, and M. Beetz. Fast point feature
histograms (fpfh) for 3d registration. In Robotics and Automation, 2009. ICRA’09. IEEE International Conference
on, pages 3212–3217. IEEE, 2009. 8
[30] H. Shao, T. Svoboda, and L. Van Gool. Zubud - zurich buildings database for image based recognition. Computer Vision Lab, Swiss Federal Institute of Technology, Switzerland,
Tech. Rep, 260:20, 2003. 6
[31] A. Vedaldi and B. Fulkerson. VLFeat: An open and portable
librar of computer vision algorithms, 2008. 6
[32] A. Wald. Sequential analysis. Courier Corporation, 1973. 2
[33] L. Xu, E. Oja, and P. Kultanen. A new curve detection
method: randomized hough transform (rht). Pattern recognition letters, 11(5):331–338, 1990. 2, 4
[34] J. Yang, H. Li, and Y. Jia. Go-icp: Solving 3d registration efficiently and globally optimally. In Proceedings of the IEEE
International Conference on Computer Vision, pages 1457–
1464, 2013. 3, 4
[35] Y. Zheng, S. Sugimoto, and M. Okutomi. Deterministically
maximizing feasible subsystem for robust model fitting with
unit norm constraint. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 1825–1832.
IEEE, 2011. 1
[36] Q.-Y. Zhou, J. Park, and V. Koltun. Fast global registration.
In European Conference on Computer Vision, pages 766–
782. Springer, 2016. 1, 8
| 1cs.CV
|
Popularity in the generalized Hospital Residents setting
Meghana Nasre and Amit Rawat
arXiv:1609.07650v3 [cs.DS] 18 Dec 2016
Indian Institute of Technology Madras, India
Abstract. We consider the problem of computing popular matchings in a bipartite graph G = (R ∪
H, E) where R and H denote a set of residents and a set of hospitals respectively. Each hospital h has
a positive capacity denoting the number of residents that can be matched to h. The residents and the
hospitals specify strict preferences over each other. This is the well-studied Hospital Residents (HR)
problem which is a generalization of the Stable Marriage (SM) problem. The goal is to assign residents to
hospitals optimally while respecting the capacities of the hospitals. Stability is a well-accepted notion
of optimality in such problems. However, motivated by the need for larger cardinality matchings,
alternative notions of optimality like popularity have been investigated in the SM setting. In this paper,
we consider a generalized HR setting – namely the Laminar Classified Stable Matchings (LCSM+ )
problem. Here, additionally, hospitals can specify classifications over residents in their preference lists
and classes have upper quotas. We show the following new results: We define a notion of popularity
and give a structural characterization of popular matchings for the LCSM+ problem. Assume n =
|R|+|H| and m = |E|. We give an O(mn) time algorithm for computing a maximum cardinality popular
matching in an LCSM+ instance. We give an O(mn2 ) time algorithm for computing a matching that
is popular amongst the maximum cardinality matchings in an LCSM+ instance.
1
Introduction
Consider an academic institution where students credit an elective course from a set of available courses.
Every student and every course rank a subset of elements from the other set in a strict order of preference.
Each course has a quota denoting the maximum number of students it can accommodate. The goal is to
allocate to every student at most one course respecting the preferences. This is the well-studied Hospital
Residents problem [7]. We consider its generalization where, in addition, a course can classify students – for
example, the students may be classified as under-graduates and post-graduates and department-wise and so
on. Depending on the classifications, a student may belong to multiple classes. Apart from the total quota,
each course now has a quota for every class. An allocation, in this setting, has to additionally respect the
class quotas. This is the Classified Stable Matching problem introduced by Huang [10].
Stability is a de-facto notion of optimality in settings where both set of participants have preferences.
Informally, an allocation of students to courses is stable if no unallocated student-course pair has incentive
to deviate from the allocation. Stability is appealing for several reasons – stable allocations are guaranteed
to exist, they are efficiently computable and all stable allocations leave the same set of students unallocated [9]. However, it is known [13] that the cardinality of a stable allocation can be half the size of the
largest sized allocation possible. Furthermore, in applications like student-course allocation, leaving a large
number of students unallocated is undesirable. Thus, it is interesting to consider notions of optimality which
respect preferences but possibly compromise stability in the favor of cardinality. Kavitha and Huang [11,13]
investigated this in the Stable Marriage (SM) setting where they considered popularity as an alternative to
stability. At a high level, an allocation of students to courses is popular if no majority wishes to deviate from
the allocation. Here, we consider popularity in the context of two-sided preferences and one-sided capacities
with classifications.
We formally define our problem now – we use the familiar hospital residents notation. Let G = (R∪H, E)
be a bipartite graph where |R ∪ H| = n and |E| = m. Here R denotes the set of residents, H denotes the set
of hospitals and every hospital h ∈ H has an upper quota q(h) denoting the maximum number of residents
h can occupy. A pair (r, h) ∈ E denotes that r and h are mutually acceptable to each other. Each resident
(resp. hospital) has a strict ordering of a subset of the hospitals (resp. residents) that are acceptable to him
or her (resp. it). This ordering is called the preference list of a vertex. An assignment (or a matching) M in
G is a subset of E such that every resident is assigned to at most one hospital and a hospital h is assigned
at most q(h) residents. Let M (r) (resp. M (h)) denote the hospital (resp. the set of residents) which are
assigned to r (resp. h) in M . A hospital h is under-subscribed if |M (h)| < q(h). A matching M is stable if no
unassigned pair (r, h) wishes to deviate from M . The goal is to compute a stable matching in G. We denote
it by HR+ throughout the paper 1 . The celebrated deferred acceptance algorithm by Gale and Shapley [7]
proves that every instance of the HR+ problem admits a stable matching.
A generalization of the HR+ problem is the Laminar Classified Stable Matching (LCSM) problem introduced by Huang [10]. An instance of the LCSM+ problem is an instance of the HR+ problem where
additionally, each hospital h is allowed to specify a classification over the set of residents in its preference
list. A class Ckh of a hospital h is a subset of residents in its preference list and has an associated upper
quota q(Ckh ) denoting the maximum number of residents that can be matched to h in Ckh . (In the LCSM
problem [10], classes can have lower quotas as well.) We assume that the classes of a hospital form a laminar
set. That is, for any two classes Cjh and Ckh , either the two classes are disjoint (Cjh ∩ Ckh = ∅), or one is contained inside the other (Cjh ⊂ Ckh or Ckh ⊂ Cjh ). Huang suitably modified the classical definition of stability
to account for the presence of these classifications. He showed that every instance of the LCSM+ problem
admits a stable matching which can be computed in O(mn) time [10]. A restriction of the LCSM+ problem,
denoted by Partition Classified Stable Matching (PCSM+ ), is where the classes of every hospital partition
the residents in its preference list.
Motivated by the need to output larger cardinality matchings, we consider computing popular matchings
in the LCSM+ problem. The notion of popularity uses votes to compare two matchings. Before we can define
voting in the LCSM+ setting, it is useful to discuss voting in the context of the SM problem.
Voting in the SM setting: Let G = (R ∪ H, E) be an instance of the SM problem and let M and M ′ be
any two matchings in G. A vertex u ∈ R ∪ H (where each hospital h has q(h) = 1) prefers M over M ′ and
therefore votes for M over M ′ if either (i) u is matched in M and unmatched in M ′ or (ii) u is matched
in both M and M ′ and prefers M (u) over M ′ (u). A matching M is more popular than M ′ if the number
of votes that M gets as compared to M ′ is greater than the number of votes that M ′ gets as compared to
M . A matching M is popular if there does not exist any matching that is more popular than M . In the
SM setting it is known that a stable matching is popular, however it was shown to be minimum cardinality
popular matching [11]. Huang and Kavitha [11,13] gave efficient algorithms for computing a max-cardinality
popular matching and a popular matching amongst max-cardinality matchings in an SM instance.
Voting in the capacitated setting: To extend voting in the capacitated setting, we assign a hospital h
as many votes as its upper quota q(h). This models the scenario in which hospitals with larger capacity
get a larger share of votes. For the HR+ problem, a hospital h compares the most preferred resident in
M (h) \ M ′ (h) to the most preferred resident in M ′ (h) \ M (h) (and votes for M or M ′ as far as those two
residents are concerned) and so on. For this voting scheme, we can obtain analogous results for computing
popular matchings in the HR+ problem via the standard technique of cloning (that is, creating q(h) copies
of a hospital h and appropriately modifying preference lists of the residents and hospitals 2 ). However,
our interest is in the LCSM+ problem, for which we are not aware of any reduction to the SM problem.
Furthermore, we show that the straightforward voting scheme as defined in the HR+ does not suffice for
the LCSM+ problem. Therefore, we define a voting scheme for a hospital which takes into consideration the
classifications as well as ensures that every stable matching in the LCSM+ instance is popular. We show the
following results:
– We define a notion of popularity for the LCSM+ problem. Since our definition ensures that stable
matchings are popular – this guarantees the existence of popular matchings in the LCSM+ problem.
– We give a characterization of popular matchings for the LCSM+ problem, which is a natural extension
of the characterization of popular matchings in SM setting [11].
1
2
We use HR+ instead of HR for consistency with other problems discussed in the paper.
For every hospital in the cloned graph, its preference list is the same as in the original instance. For every hospital h, fix an ordering
of its clones. The preference list of a resident r in the cloned instance is obtained by replacing the occurrence of h by the fixed
ordering of its clones. We refer the reader to [4,14] for details.
2
– We obtain the following algorithmic results. An O(m + n) (resp. O(mn)) time algorithm for computing a
maximum cardinality popular matching in a PCSM+ (resp. LCSM+ ) instance. An O(mn) (resp. O(mn2 ))
time algorithm for computing a popular matching amongst maximum cardinality matchings in a PCSM+
(resp. LCSM+ ) instance.
Very recently, independent of our work, two different groups [4,12] have considered popular matchings in
the one-to-many setting. Brandl and Kavitha [4] have considered computing popular matchings in the HR+
problem. In their work as well as ours, a hospital h is assigned as many votes as its capacity to compare
two matchings M and M ′ . In contrast, by the definition of popularity in [4], a hospital h chooses the most
adversarial ordering of residents in M (h) \ M ′ (h) and M ′ (h) \ M (h) for comparing M and M ′ . However, it
is interesting to note that in an HR+ instance the same matching is output by both our algorithms. On the
other hand, we remark that the model considered in our paper is a more general one than the one considered
in [4]. Kamiyama [12] has generalized our work and the results in [4] using a matroid based approach.
We finally remark that one can consider voting schemes where a hospital is given a single vote instead
of capacity many votes. In one such scheme, a hospital compares the set of residents in M (h) and M ′ (h) in
lexicographic order and votes accordingly. However, when such a voting is used, it is possible to construct
instances where a stable matching is not popular. The techniques in this paper use the fact that stable
matchings are popular, therefore it is unclear if our techniques will apply for such voting schemes.
Related Work: The notion of popularity was introduced by Gärdenfors [8] in the context of stable matchings. In [1] Abraham et al. studied popularity in the one-sided preference list model. As mentioned earlier,
our work is inspired by a series of papers where popularity is considered as an alternative to stability in the
stable marriage setting by Huang, Kavitha and Cseh [5,11,13]. Biró et al. [3] give several practical scenarios
where stability may be compromised in the favor of size. The PCSM+ problem is a special case of the Student Project Allocation (SPA) problem studied by Abraham et al. [2]. They gave a linear time algorithm
to compute a stable matching in an instance of the SPA problem. In this paper, we use the algorithms of
Abraham et al. [2] and Huang [10] for computing stable matchings in the PCSM+ and LCSM+ problems.
Both these algorithms follow the standard deferred acceptance algorithm of Gale and Shapley with problem
specific modifications. We refer the reader to [2] and [10] for details.
Organization: In Section 2 we define the notion of popularity, in Section 3 we present the structural characterization of popular matchings. In Section 4 we describe our algorithms to compute a maximum cardinality
popular matching, and a popular matching amongst maximum cardinality matchings. We conclude with a
short discussion about popular matchings in the LCSM problem.
2
Stability and popularity in the LCSM+ problem
Consider an instance G = (R ∪ H, E) of the LCSM+ problem. As done in [10], assume that for every h ∈ H
there is a class C∗h containing all the residents in the preference list of h and q(C∗h ) = q(h). For a hospital h,
let T (h) denote the tree of classes corresponding to h where C∗h is the root of T (h). The leaf classes in T (h)
denote the most refined classifications for a resident whereas as we move up in the tree from a leaf node to
the root, the classifications gets coarser.
To define stable matchings in the LCSM problem, Huang introduced the notion of a blocking group w.r.t. a
matching. Later, Fleiner and Kamiyama [6] defined a blocking pair which is equivalent to a blocking group of
Huang. We use the definition of stability from [6] which we recall below. A set S = {r1 , . . . , rl } is feasible for
a hospital h if |S| ≤ q(h) and for every class Cjh of h (including the root class C∗h ), we have |Cjh ∩ S| ≤ q(Cjh ).
A matching M in G is feasible if every resident is matched to at most one hospital, and M (h) is feasible for
every hospital h ∈ H. A pair (r, h) ∈
/ M blocks M iff both the conditions below hold:
– r is unmatched in M , or r prefers h over M (r), and
– either the set M (h) ∪ {r} is feasible for h, or there exists a resident r′ ∈ M (h), such that h prefers r over
r′ , and (M (h) \ {r′ }) ∪ {r} is feasible for h.
A feasible matching M in G is stable if M does not admit any blocking pair.
3
2.1
Popularity
To define popularity, we need to specify how a hospital compares two sets M (h) and M ′ (h) in an LCSM+
setting, where M and M ′ are two feasible matchings in the instance.
Illustrative example Consider the following LCSM+ instance where R = {r1 , . . . , r4 } and H = {h1 , . . . , h3 }
and the preference lists of the residents and hospitals are as given in Figure 1(a) and (b) respectively. The
preferences can be read as follows: resident r1 has h1 as his top choice hospital. Resident r2 has h2 as
its top choice hospital followed by h1 which is his second choice hospital and so on. For h ∈ {h2 , h3 } we
have q(h) = 1 and both these hospitals have a single class C∗h containing all the residents in the preference list of h and q(C∗h ) = q(h). For hospital h1 we have q(h1 ) = 2 and the classes provided by h1 are
C1h1 = {r1 , r2 }, C2h1 = {r3 , r4 }, C∗h1 = {r1 , r2 , r3 , r4 } with quotas as follows: q(C1h1 ) = q(C2h1 ) = 1 and
q(C∗h1 ) = 2. We remark that the example in Figure 1 is also a PCSM+ instance. Figure 1(c) shows the tree
T (h1 ).
r1
r2
r3
r4
:
:
:
:
h1
h2 , h1 , h3
h1 , h2
h1
(a)
h1 : r2 , r3 , r4 , r1
h2 : r3 , r2
h3 : r2
r1 , r2 , r3 , r4
r1 , r2
(b)
C1h1
C∗h1
r3 , r4
C2h1
(c)
Fig. 1. (a) Resident preferences, (b) Hospital preferences, (c) T (h1 ). The matchings M = {(r1 , h1 ), (r2 , h2 ), (r3 , h1 )},
M ′ = {(r2 , h1 ), (r3 , h2 ), (r4 , h1 )}, and M ′′ = {(r1 , h1 ), (r2 , h3 ), (r3 , h2 ), (r4 , h1 )} are all feasible in the instance.
Consider the two feasible matchings M and M ′ defined in Fig. 1. Note that M is stable in the instance
whereas the edge (r3 , h1 ) blocks M ′ . While comparing M and M ′ , the vote for every vertex u in the instance
except h1 is clear – u compares M (u) with M ′ (u) and votes accordingly. In order for h1 to vote between
M and M ′ , the hospital compares between M (h1 ) = {r1 , r3 } and M ′ (h1 ) = {r2 , r4 }. A straightforward way
is to compare r3 with r2 (the most preferred resident in M (h1 ) to the most preferred resident in M ′ (h1 ))
and then compare r1 with r4 (second most preferred resident in M (h1 ) to second most preferred resident
in M ′ (h1 )). Thus, both the votes of h1 are in favor of M ′ when compared with M . Such a comparison has
two issues – (i) it ignores the classifications given by h1 , and (ii) the number of votes that M ′ gets when
compared with M is more than the number of votes that M gets as compared to M ′ . Therefore M ′ is more
popular than M which implies that M (a stable matching) is not popular.
We propose a comparison scheme for hospitals which addresses both the issues. In the above example,
we note that r1 ∈ M (h) has a corresponding resident r2 ∈ M ′ (h) to be compared to in one of the most
refined classes C1h1 (see Figure 1(c)). Thus, we compare r1 with r2 . The resident r3 ∈ M (h) is compared to
r4 ∈ M (h) another leaf class C2h1 . According to this comparison, h1 is indifferent between M and M ′ and
M ′ is no longer more popular than M . Note that, although in the example, both the comparisons happen
in a leaf class, this may not be the case in a general instance. Finally, we note that the matching M ′′ is a
popular matching in the instance and is strictly larger in size than the stable matching M .
We formalize the above observations in the rest of the section. To take into account the classifications, for
a hospital h and the matchings M and M ′ , we set up a correspondence between residents in M (h) \ M ′ (h)
and the residents in M ′ (h) \ M (h). That is, we define:
corr : M (h) ⊕ M ′ (h) → M (h) ⊕ M ′ (h) ∪ {⊥}
For a resident r ∈ M (h) ⊕ M ′ (h) we denote by corr(r) the corresponding resident to which r gets compared
when the hospital h casts its votes. We let corr(r) = ⊥ if r does not have a corresponding resident to be
compared to from the other matching. The pseudo-code for the algorithm to compute the corr function is
given below.
4
Algorithm 1 Correspondence between residents of M (h) and M ′ (h)
1: procedure Find-Correspondence(h, M, M ′ )
2:
let T (h) be the classification tree associated with h
3:
set corr(r) = ⊥ for each r ∈ M (h) ⊕ M ′ (h)
4:
Y = M (h) \ M ′ (h); Y ′ = M ′ (h) \ M (h)
5:
while Y 6= ∅ and Y ′ =
6 ∅ do
6:
for each class Cjh in T (h) do
7:
Xj = Cjh ∩ Y
8:
Xj′ = Cjh ∩ Y ′
9:
10:
11:
12:
13:
14:
Let Cfh be one of the most refined classes for which Xf 6= ∅ and Xf′ 6= ∅.
for k = 1, . . . , min(|Xf |, |Xf′ |) do
let r be the k-th most preferred resident in Xf
let r ′ be the k-th most preferred resident in Xf′
set corr(r) = r ′ , and corr(r ′ ) = r
Y = Y \ {r}; Y ′ = Y ′ \ {r ′ }
The algorithm begins by setting corr for every r ∈ M (h) ⊕ M ′ (h) to ⊥. The algorithm maintains two
sets of residents Y = M (h) \ M ′ (h) and Y ′ = M ′ (h) \ M (h) for whom corr needs to be set. As long as
the sets Y and Y ′ are both non-empty, the algorithm repeatedly computes for every class Cjh (including
the root class C∗h ) the sets Xj = Cjh ∩ Y and Xj′ = Cjh ∩ Y ′ . The algorithm then chooses one of the most
refined classes, say Cfh in T (h), for whom Xf and Xf′ are both non-empty. Finally, residents in Xf and Xf′
are sorted according to the preference ordering of h and the corr of the k-th most preferred resident in Xf
is set to the k-th most preferred resident in Xf′ , where k = 1, . . . , min{|Xf |, |Xf′ |}.
For r ∈ R, and any feasible matching M in G, if r is unmatched in M then, M (r) = ⊥. A vertex prefers
any of its neighbours over ⊥. For a vertex u ∈ R ∪ H, let x, y ∈ N (u) ∪ {⊥}, where N (u) denotes the
neighbours of u in G.
voteu (x, y) = +1
= −1
= 0
if u prefers x over y
if u prefers y over x
if x = y
Using the above notation, the vote of a resident is easy to define – a resident r prefers M ′ over M iff the
term Vr > 0, where Vr = voter (M ′ (r), M (r)).
Recall that a hospital h uses q(h) votes to compare M and M ′ . Let q1 (h) = |M (h) ∩ M ′ (h)| (number of
common residents assigned to h in M and M ′ ) and q2 (h) = q(h) − max{|M (h)|, |M ′ (h)|} (number of unfilled
positions of h in both M and M ′ ). Our voting scheme ensures that q1 (h) + q2 (h) votes of h remain unused
when comparing M and M ′ . A hospital h prefers M ′ over M iff the term Vh > 0, where Vh is defined as
follows:
X
voteh (r, corr(r))
Vh = (|M ′ (h)| − |M (h)|) +
r∈M ′ (h)\M(h)
&&
corr(r)6=⊥
The first term in the definition of Vh counts the votes of h w.r.t. the residents from either M or M ′ that
did not find correspondence. The second term counts the votes of h w.r.t. the residents each of which has
a corresponding resident from the other matching. We note that in the SM setting, corr(r) will simply be
M (h). Thus, our definition of votes in the presence of capacities is a natural generalization of the voting
scheme in the SM problem.
Let us define the term ∆(M ′ , M ) as the difference between the votes that M ′ gets over M and the votes
that M gets over M ′ .
X
X
∆(M ′ , M ) =
Vr +
Vh
r∈R
5
h∈H
Definition 1. A matching M is popular in G iff for every feasible matching M ′ , we have ∆(M ′ , M ) ≤ 0.
2.2
Decomposing M ⊕ M ′
Here, we present a simple algorithm which allows us to decompose edges of components of M ⊕ M ′ in an
instance into alternating paths and cycles. Consider the graph G̃ = (R ∪ H, M ⊕ M ′ ), for any two feasible
matchings M and M ′ in G. We note that the degree of every resident in G̃ is at most 2 and the degree of
every hospital in G̃ is at most 2 · q(h). Consider any connected component C of G̃ and let e be any edge in
C. We observe that it is possible to construct a unique maximal M alternating path or cycle ρ containing e
using the following simple procedure. Initially ρ contains only the edge e.
1. Let r ∈ R be one of the end points of the path ρ, and assume that (r, M (r)) ∈ ρ. We grow ρ by adding
the edge (r, M ′ (r)). Similarly if an edge from M ′ is incident on r in ρ, we grow the path by adding the
edge (r, M (r)) if it exists.
2. Let h ∈ H be one of the end points of the path ρ, and assume that (r, h) ∈ M \ M ′ belongs to ρ. We
extend ρ by adding (corr(r), h) if corr(r) is not equal to ⊥. A similar step is performed if the last edge
on ρ is (r, h) ∈ M ′ \ M .
3. We stop the procedure when we complete a cycle (ensuring that the two adjacent residents of a hospital
are corr for each other according to the hospital), or the path can no longer be extended. Otherwise we
go to Step 1 or Step 2 as applicable and repeat.
The above procedure gives us a unique decomposition of a connected component in G̃ into alternating paths
and cycles. Note that a hospital may appear multiple times in a single path or a cycle and also can belong
to more than one alternating paths and cycles. Figure 2 gives an example of the decomposition of the two
feasible matchings in the instance in Figure 1.
r1
r1
h1
h1
r2
r2
h2
h2
r3
r3
r4
r4
h1
(a)
(b)
Fig. 2. M and M ′ are feasible matchings in the example as defined in Fig. 1. (a) G̃ = (R ∪ H, M ⊕ M ′ ); bold edges
belong to M , dashed edges belong to M ′ . (b) shows the decomposition of the edges of the component of G̃ into a
single path.
Let YM⊕M ′ denote the collection of alternating paths and alternating cycles obtained by decomposing
every component of G̃.
We now state a useful property about any alternating path or cycle in YM⊕M ′ .
Lemma 1. If ρ is an alternating path or an alternating cycle in YM⊕M ′ , then M ⊕ ρ is a feasible matching
in G.
Proof. Let hr′ , h, ri be any sub-path of ρ, where r′ = corr(r), and (r, h) ∈ M . We prove that (M (h) \ {r}) ∪
{r′ } is feasible for h. Let Cih (resp. Cjh ) be the unique leaf class of T (h) containing r (resp. r′ ). See Figure 3.
We consider the following two cases:
6
C∗h
Ckh
Cth
.., r, ..
.., r ′ , ..
Cih
Cjh
Fig. 3. The classification tree T(h) for a hospital h.
– r and r′ belong to the same leaf class in T (h), i.e. Cih = Cjh . In this case, it is easy to note that
(M (h) \ {r}) ∪ {r′ } is feasible for h.
– r and r′ belong to different leaf classes of T (h), i.e. Cih 6= Cjh . Observe that |(M (h) \ {r}) ∪ {r′ }| can
violate the upper quota only for those classes of T (h) which contain r′ but do not contain r. Let Ckh be
the least common ancestor of Cih and Cjh in T (h). It suffices to look at any class Cth which lies in the
path from Ckh to Cjh excluding the class Ckh and show that |(M (h) ∩ Cth ) ∪ {r′ }| ≤ q(Cth ). As r′ = corr(r)
and r ∈
/ Cth , we claim that |M (h) ∩ Cth | < |M ′ (h) ∩ Cth | ≤ q(Cth ). The first inequality is due to the fact
that r′ did not find a corresponding resident in the set (M (h) \ M ′ (h)) ∩ Cth . The second inequality is
because M ′ is feasible. Thus, (M (h) ∩ Cth ) ∪ {r′ } does not violate the upper quota for Cth . Therefore
(M (h) \ {r}) ∪ {r′ } is feasible for h.
We note that the hospital h may occur multiple times on ρ. Let M (h)ρ denote the set of residents matched
to h restricted to ρ. To complete the proof of the Lemma, we need to prove that (M (h) \ M (h)ρ ) ∪ M ′ (h)ρ
is feasible for h. The arguments for this follow from the arguments given above.
⊓
⊔
As was done in [13], it is convenient to label the edges of M ′ \ M and use these labels to compute ∆(M ′ , M ).
Let (r, h) ∈ M ′ \ M ; the label on (r, h) is a tuple:
(voter (h, M (r)), voteh (r, corr(r)))
Note that since we are labeling edges of M ′ \ M , both entries of the tuple come from the set {−1, 1}. With
these definitions in place, we are ready to give the structural characterization of popular matchings in an
LCSM+ instance.
3
Structural characterization of popular matchings
Let G = (R ∪ H, E) be an LCSM+ instance and let M and M ′ be two feasible matchings in G. Using
the corr function, we obtain a correspondence of residents in M (h) ⊕ M ′ (h) for every hospital h in G. Let
G̃ = (R ∪ H, M ⊕ M ′ ) and let YM⊕M ′ denote the collection of alternating paths and cycles obtained by
decomposing every component of G̃. Finally, we label the edges of M ′ \ M using appropriate votes. The goal
of these steps is to is to rewrite the term ∆(M ′ , M ) as a sum of labels on edges.
We note that the only vertices for whom their vote does not get captured on the edges of M ′ \ M are
vertices that are matched in M but not matched in M ′ . Let U denote the multi-set of vertices that are end
points of paths in YM⊕M ′ such that there is no M ′ edge incident on them. Note that the same hospital can
belong to multiple alternating paths and cycles in YM⊕M ′ , therefore we need a multi-set. All vertices in U
7
prefer M over M ′ and hence we add a −1 while capturing their vote in ∆(M ′ , M ). We can write ∆(M ′ , M )
as:
X
X
X
∆(M ′ , M ) =
{voter (h, M (r)) + voteh (r, corr(r))}
−1 +
x∈U
ρ∈YM ⊕M ′
(r,h)∈(M ′ ∩ρ)
We now delete the edges labeled (−1, −1) from all paths and cycles ρ in YM⊕M ′ . This simply breaks paths
and cycles into one or more paths. Let this new collection of paths and cycles be denoted by ỸM⊕M ′ . Let
Ũ denote the multi-set of vertices that are end points of paths in ỸM⊕M ′ such that there is no M ′ edge
incident on them. We rewrite ∆(M ′ , M ) as:
X
X
X
{voter (h, M (r)) + voteh (r, corr(r))}
−1 +
∆(M ′ , M ) =
x∈Ũ
ρ∈ỸM ⊕M ′
(r,h)∈(M ′ ∩ρ)
Theorem below characterizes a popular matching.
Theorem 1. A feasible matching M in G is popular iff for any feasible matching M ′ in G, the set ỸM⊕M ′
does not contain any of the following:
1. An alternating cycle with a (1, 1) edge,
2. An alternating path which has a (1, 1) edge and starts with an unmatched resident in M or a hospital
which is under-subscribed in M .
3. An alternating path which has both its ends matched in M and has two or more (1, 1) edges.
Proof. We show that if M is a feasible matching such that for any M ′ the set ỸM⊕M ′ does not contain (1),
(2), (3) as in Theorem 1, then M is popular in G.
Assume for the sake of contradiction that M satisfies the conditions of Theorem 1, and yet M is not
popular. Therefore there exists a feasible matching M ∗ such that ∆(M ∗ , M ) > 0. Consider the set YM ∗ ⊕M .
Recall that this set is a collection of paths and cycles and the edges of M ∗ \ M are labeled. Let ρ be any path
or cycle in YM ∗ ⊕M and let ∆(M ∗ , M )ρ denote the difference between the votes of M ∗ and M when restricted
to the residents and hospitals in ρ. Since ∆(M ∗ , M ) > 0, there exists a ρ such that ∆(M ∗ , M )ρ > 0. Note
that ρ is present in YM ∗ ⊕M ; using the presence of ρ we establish the existence of a ρ′ ∈ ỸM ∗ ⊕M of the form
(1), (2) or (3) which contradicts our assumption. We consider three cases depending on the structure of ρ.
1. ρ is an alternating cycle or ρ is an alternating path which starts and ends in an M edge:
Since ρ ∈ YM ∗ ⊕M , and ∆(M ∗ , M )ρ > 0, it implies that there are more edges in ρ labeled (1, 1) than the
number of edges labeled (−1, −1). We now delete the edges labeled (−1, −1) from ρ; this breaks ρ in to
multiple alternating paths. Note that each of these paths (say ρ′ ) start and end with an M edge and are
also present in ỸM ∗ ⊕M . Furthermore, since ρ contained more number of edges labeled (1, 1) than the
number of edges labeled (−1, −1), it is clear that there exists at least one ρ′ which has two edges labeled
(1, 1). This is a path of type (3) from the theorem statement and therefore contradicts our assumption
that M satisfied the conditions of the theorem.
2. ρ is an alternating path which starts or ends in an M ∗ edge:
The proof is similar to the previous case except that when we delete from ρ the edges labeled (−1, −1)
we get paths ρ′ ∈ ỸM ∗ ⊕M which are paths of type (2) or type (3) from the theorem statement. This
contradicts the assumption that M satisfied the conditions of the theorem.
This completes the proof of one direction of the Theorem. To prove the other direction, we prove the
contrapositive of the statement. That is, if for any feasible matching M ′ , ỸM⊕M ′ contains (1), (2) or (3),
then M is not popular in G. We first assume that ρ ∈ ỸM⊕M ′ satisfying (1), (2), or (3) is also present in
YM⊕M ′ . Under this condition, it is possible to get a more popular matching than M by the following three
cases.
8
– Let M2 = M ⊕ ρ be a matching in G; by Lemma 1 we know that M2 is feasible in G. Comparing M2 to
M yields two more votes for M2 . Hence, M2 is more popular than M .
– If ρ is an alternating path in ỸM⊕M ′ , which has both its endpoints matched in M , and contains more
than one edge labeled (1, 1). Then similar to the case above M2 = M ⊕ ρ is more popular than M .
– If ρ is an alternating path in ỸM⊕M ′ , which has exactly one of its endpoints matched in M , and contains
an edge labeled (1, 1), then again M2 = M ⊕ ρ is more popular than M .
Now let us assume that ρ ∈ ỸM⊕M ′ is not present in YM⊕M ′ . In such a case, ρ is contained in a larger path
or a cycle ρ′ ∈ YM⊕M ′ obtained by combining ρ with other paths in ỸM⊕M ′ and adding the deleted (−1, −1)
edges. Using the larger path or cycle ρ′ we can construct a matching that is more popular than M . Note
that we need to use paths or cycles in YM⊕M ′ to obtain another matching, since we have to ensure that the
matching obtained is indeed feasible in the instance and the correspondences are maintained.
⊓
⊔
We now prove that every stable matching in an LCSM+ instance is popular.
Theorem 2. Every stable matching in an LCSM+ instance G is popular.
Proof. Let M be a stable matching in G. For any feasible matching M ′ in G consider the set YM⊕M ′ . To
prove that M is stable, it suffices to show that there does not exist a path or cycle ρ ∈ YM⊕M ′ such that
an edge of ρ is labeled (1, 1). For the sake of contradiction, assume that ρ is such a path or cycle, which has
an edge (r′ , h) ∈ M ′ \ M labeled (1, 1). Let r = corr(r′ ), where (r, h) ∈ M ∩ ρ. From the proof of Lemma 1
we observe that (M (h) \ {r}) ∪ {r′ } is feasible for h, therefore the edge (r′ , h) blocks M contradicting the
stability of M .
⊓
⊔
4
Popular matchings in LCSM+ problem
In this section we present efficient algorithms for computing (i) a maximum cardinality popular matching, and
(ii) a matching that is popular amongst all the maximum cardinality matchings in a given LCSM+ instance.
Our algorithms are inspired by the reductions of Kavitha and Cseh [5] where they work with a stable marriage
instance. We describe a general reduction from an LCSM+ instance G to another LCSM+ instance Gs . Here
s = 2, . . . , |R|. The algorithms for the two problems are obtained by choosing an appropriate value of s.
The graph Gs : Let G = (R ∪ H, E) be the input LCSM+ instance. The graph Gs = (Rs ∪ Hs , Es ) is
constructed as follows: Corresponding to every resident r ∈ R, we have s copies of r, call them r0 , . . . , rs−1
in Rs . The hospitals in H and their capacities remain unchanged; however we have additional dummy
hospitals each of capacity 1. Corresponding to every resident r ∈ R, we have (s − 1) dummy hospitals
d0r , . . . , drs−2 in Hs . Thus,
Rs = { r0 , . . . , rs−1 | ∀r ∈ R}; Hs = H ∪ { d0r , . . . , drs−2 | ∀r ∈ R}
We use the term level-i resident for a resident ri ∈ Rs for 0 ≤ i ≤ s − 1. The preference lists corresponding
to s different residents of r in Gs are:
– For a level-0 resident r0 , its preference list in Gs is the preference list of r in G, followed by the dummy
hospital d0r .
followed by preference list
– For a level-i resident ri , where 1 ≤ i ≤ s − 2, its preference list in Gs is di−1
r
of r in G, followed by dir .
– For a level-(s − 1) resident rs−1 , its preference list in Gs is the dummy hospital drs−2 followed by the
preference list of r in G.
The preference lists of hospitals in Gs are as follows.
– The preference list for a dummy hospital dir is ri followed by ri+1 .
– For h ∈ H, its preference list in Gs , has level-(s − 1) residents followed by level-(s − 2) residents, so on
upto the level-0 residents in the same order as in h’s preference list in G.
9
Finally, we need to specify the classifications
of the hospitals in Gs . For every class Cih in the instance G,
S
h
we have a corresponding class C̄i = r∈C h {r0 , . . . , rs−1 } in Gs , such that q(C̄ih ) = q(Cih ). We note that
i
|C̄ih | = s · |Cih |. Let Ms be a stable matching in Gs . Then Ms satisfies the following properties:
(I1 ) Each dir ∈ Hs for 0 ≤ i ≤ s − 2, is matched to one of {ri , ri+1 } in Ms .
(I2 ) The above invariant implies that for every r ∈ R at most one of {r0 , . . . , rs−1 } is assigned to a non-dummy
hospital in Ms .
(I3 ) For a resident r ∈ R, if ri is matched to a non-dummy hospital in Ms , then for all 0 ≤ j ≤ i − 1,
. This also implies that in Ms all
Ms (rj ) = djr . Furthermore, for all i + 1 ≤ p ≤ s − 1, Ms (rp ) = dp−1
r
residents r0 , . . . , rs−2 are matched and only rs−1 can be left unmatched in Ms .
These invariants allow us to naturally map the stable matching Ms to a feasible matching M in G. We
define a function map(Ms ) as follows.
M = map(Ms ) = {(r, h) : h ∈ H and (ri , h) ∈ Ms for exactly one of 0 ≤ i ≤ s − 1}
We outline an algorithm that computes a feasible matching in an LCSM+ instance G. Given G and s,
construct the graph Gs from G. Compute a stable matching Ms in Gs . If G is an LCSM+ instance we
use the algorithm of Huang [10] to compute a stable matching in G. If G is a PCSM+ instance, it is easy
to observe that Gs is also a PCSM+ instance. In that case, we use the algorithm of Abraham et al.[2] to
compute a stable matching. (The SPA instance is different from a PCSM+ instance, however, there is a
easy reduction from the PCSM+ instance to SPA, we give the reduction (refer Appendix A.1) for the sake
of completeness.). We output M = map(Ms ) whose feasibility is guaranteed by the invariants mentioned
earlier. The complexity of our algorithm depends on s and the time required to compute a stable matching
in the problem instance.
In the rest of the paper, we denote by M the matching obtained as map(Ms ) where Ms is a stable
matching in Gs . For any resident ri ∈ R, we define
map−1 (ri , Ms ) = riji
= ris−1
where 0 ≤ ji ≤ s − 1 and Ms (riji ) is a non-dummy hospital
otherwise.
Recall by Invariant (I3 ), exactly one of the level copy of ri in Gs is matched to a non-dummy hospital in Ms .
For any feasible matching M ′ in G consider the set YM⊕M ′ – recall that this is a collection of M alternating
paths and cycles in G. For any path or cycle ρ in YM⊕M ′ , let us denote by ρs = map−1 (ρ, Ms ) the path or
cycle in Gs obtained by replacing every resident r in ρ by map−1 (r, Ms ). Recall that if a resident r is present
in the class Cjh defined by a hospital h in G, then in the graph Gs , ri ∈ C̄jh for i = 0, . . . , s − 1. The map−1
function maps a resident r in G to a unique level-i copy in Gs . Using Lemma 1 and these observations we
get the following corollary.
Corollary 1. Let ρ be an alternating path or an alternating cycle in YM⊕M ′ , then Ms ⊕ ρs is a feasible
matching in Gs , where ρs = map−1 (ρ, Ms ).
The following technical lemma is useful in proving the properties of the matchings produced by our algorithms.
Lemma 2. Let ρ be an alternating path or an alternating cycle in YM⊕M ′ , and ρs = map−1 (ρ, Ms ).
1. There cannot be any edge labeled (1, 1) in ρs .
j′
2. Let hraja , h, rbjb i be a sub-path of ρs , where h = Ms (rbjb ). Then, the edge (raa , h) ∈
/ ρs cannot be labeled
(1, 1), where ja′ < ja .
Proof. Let hraja , h, rbjb i be a sub-path of ρs , where h = Ms (rbjb ) (Figure 4). As Ms ⊕ ρs is feasible in Gs
(Corollary 1), the set (Ms (h) \ {rbjb }) ∪ {raja } is feasible for h in Gs . Now since (raja , h) is labeled (1, 1),
the edge (raja , h) blocks Ms contradicting its stability. This proves (1). To prove (2), assume that the edge
j′
j′
(raa , h) ∈
/ ρs is labeled (1, 1). The residents raja and raa belong to the same class (say C̄kh ) in Gs , hence
j′
j′
(Ms (h) \ {rbjb }) ∪ {raa } is feasible for h. Thus the edge (raa , h) blocks Ms contradicting its stability.
10
raja
h
rbjb
j′
raa
j′
Fig. 4. The edges (raja , h) and (rbjb , h) belong to ρ, while the edge (raa , h) does not belong to ρ.
4.1
Maximum cardinality popular matching
Let G = (R∪H, E) be an instance of the LCSM+ problem where we are interested in computing a maximum
cardinality popular matching. We use our generic reduction with the value of the parameter s = 2. Since G2
is linear in the size of G, and a stable matching in an LCSM+ instance can be computed in O(mn) time [10],
we obtain an O(mn) time algorithm to compute a maximum cardinality popular matching in G. In case
G is a PCSM+ instance, we use the linear time algorithm in [2] for computing a stable matching to get a
linear time algorithm for our problem. The proof of correctness involves two things – we first show that M
is popular in G. We then argue that it is the largest size popular matching in G. We state the main theorem
of this section below.
Theorem 3. Let M = map(M2 ) where M2 is a stable matching in G2 . Then M is a maximum cardinality
popular matching in G.
We break down the proof of Theorem 3 in two parts. Lemma 3 shows that the assignment M satisfies
all the conditions of Theorem 1. Lemma 5 shows that the matching output is indeed the largest size popular matching in the instance. Let M ′ be any assignment in G. Recall the definition of ỸM⊕M ′ – this set
contains M alternating paths and M alternating cycles in G and the edge labels on the M ′ edges belong to
{(−1, 1), (1, −1), (1, 1)}.
Lemma 3. Let M = map(M2 ) where M2 is a stable matching in G2 and let M ′ be any feasible assignment
in G. Consider the set of alternating paths and alternating cycles ỸM⊕M ′ . Then, the following hold:
1. An alternating cycle C in ỸM⊕M ′ , does not contain any edge labeled (1, 1).
2. An alternating path P in ỸM⊕M ′ that starts or ends with an edge in M ′ , does not contain any edge
labeled (1, 1).
3. An alternating path P in ỸM⊕M ′ which starts and ends with an edge in M , contains at most one edge
labeled (1, 1).
Proof. We first prove the parts (1) and (2). Recall that M = map(M2 ) where M2 is a stable matching in
G2 . Assume that ρ = hu0 , v1 , u1 , . . . , vk , uk i where for each i = 0, . . . , k, vi = M (ui ) (in case ui is a hospital,
vi ∈ M (ui )). In case ρ is a cycle, all subscripts follow mod k arithmetic. The existence of ρ in ỸM⊕M ′ implies
that there is an associated M2 alternating path or an M2 alternating cycle ρ2 = map−1 (ρ, M2 ) in G2 .
Now assume for the sake of contradiction that ρ contains an edge e = (ra , hb ) ∈
/ M labeled (1, 1) for some
a = 0, . . . , k, and b = 0, . . . , k. We observe the following about preferences of ra and hb in G.
(O1 ) ra prefers hb over ha = M (ra ).
(O2 ) hb prefers ra over rb ∈ M (hb ), where rb = corr(ra ).
11
Using the presence of an edge labeled (1, 1) in ρ, we will contradict the stability of M2 in G2 . Consider
/ M2 . We consider the four cases that
the edge e′ = (raja , hb ) in G2 . Since hb = M2 (rbjb ), we observe that e′ ∈
can arise depending on the values of ja and jb .
1. ja = jb = 0
2. ja = jb = 1
3. ja = 1 and jb = 0
4. ja = 0 and jb = 1
Recall observation (O1 ), and the fact that the residents do not change their preferences in G2 w.r.t.
the hospitals originally in G. This implies in all the four cases above, the resident raja prefers hb over
ha = M2 (raja ). Using (O2 ) and the fact that a hospital h in G2 prefers level-1 residents over level-0 residents,
we can conclude the following. For the cases (1), (2) and (3), hospital hb prefers raja over rbjb , which implies
that the pair (raja , hb ) is labeled (1, 1), and thus forms a blocking pair w.r.t. M2 (using Lemma 2(1)).
We now consider the three different cases for ρ depending on whether ρ is a path or a cycle. When ρ is
a path, we break down its proof in two cases – (i) ρ starts or ends with a resident unmatched in M . (ii) ρ
starts or ends with an under-subscribed hospital. In each of the different possibilities for ρ, we show that the
stability of M2 can be contradicted even in case (4), i.e. when ja = 0 and jb = 1.
– ρ = hr0 , h1 , r1 , . . . , hk−1 , rk−1 i is an alternating path that starts or ends with a resident which is
jk−1
unmatched in M . Here ρ2 = map−1 (ρ, M2 ) = hr0j0 , h1 , r1j1 , . . . , hk−1 , rk−1
i and for t = 0, . . . , k − 1,
jt ∈ {0, 1}.
Using invariants (I1 ), (I2 ), and (I3 ), we conclude that a resident r remains unmatched in M2 when its
level-0 copy is matched to the dummy hospital dr , and the level-1 copy is unmatched in M2 . Therefore,
the first resident on the path ρ2 is a level-1 resident. Furthermore, the second resident on the path r1
has to be a level-1 resident. Otherwise, as r01 is unmatched in M2 and h1 prefers a level-1 resident over a
level-0 resident, the edge (r01 , h1 ) will be labeled (1, 1), and thus forms a blocking pair w.r.t. M2 (using
Lemma 2(1)).
We consider an edge e ∈ ρ such that b = a + 1. In case (4), we observe that as j0 = j1 = 1, ja = 0, and
a < b, there exists an index x in ρ2 such that there is a transition from a level-1 resident to a level-0
1
resident. That is, (rx0 , hx ) ∈ M2 and (rx−1
, hx ) ∈
/ M2 both belong to ρ2 .
We enumerate the possible labels for the edge ex = (rx−1 , hx ) in G.
1
• If ex is labeled (1, 1) or (1, −1), then the edge (rx−1
, hx ) is labeled (1, 1), and thus blocks M2 (using
Lemma 2(1)).
0
• If ex is labeled (−1, 1), then the edge (rx−1
, hx ) is labeled (1, 1), and thus blocks M2 (using Lemma 2(2)).
– ρ = hh0 , r1 , h1 , . . . , rk−1 , hk−1 i is an alternating path that starts or ends with an under-subscribed
jk−1
hospital. Here ρ2 = map−1 (ρ, M2 ) = hh0 , r1j1 , h1 , . . . , rk−1
, hk−1 i and for t = 1, . . . , k − 1, jt ∈ {0, 1}.
0
Observe that if j1 = 1, then (r1 , h0 ) is labeled (1, 1), as h0 is unmatched in M , and r10 prefers h0
to dr1 (dr1 = M2 (r10 ) using invariants (I1 ), (I2 ), and (I3 )), contradicting the stability of Ms (using
Lemma 2(1)). Thus, it must be the case that j0 = 0. Note that the edge (r1 , h0 ) can not be labeled (1, 1)
in G, as h0 being under-subscribed prefers being matched to r1 , and residents do not change their votes,
and thus the edge (r10 , h0 ) is labeled (1, 1), contradicting the stability of M2 (using Lemma 2(1)).
We consider an edge e ∈ ρ such that a = b + 1. In case (4), we observe that as j1 = 0, jb = 1, and a > b,
there exists an index x in ρ2 such that there is a transition from a level-0 resident to a level-1 resident.
1
, hx ) ∈
/ M2 both belong to ρ2 . Using an argument similar to in the case
That is, (rx0 , hx ) ∈ M2 and (rx+1
0
1
, hx ) is labeled (1, 1), and therefore
above, we can show that either the edge (rx+1
, hx ) or the edge (rx+1
forms a blocking pair w.r.t. M2 .
– ρ = hr0 , h0 , r1 , h1 , . . . , rk , hk , r0 i is an alternating cycle. Here ρ2 = map−1 (ρ, Ms ) = hr0j0 , h0 , r1j1 , h1 , . . . , rkjk , hk , r0j0 i
and for t = 1, . . . , k − 1, jt ∈ {0, 1}.
We consider an edge e ∈ ρ such that a = b + 1. As ja = 0 and jb = 1, and b < a, this is a transition
from a level-1 resident to a level-0 resident in the cycle ρ2 . To complete the cycle ρ2 there must exist an
index x such that there is a transition from a level-0 resident to a level-1 resident. That is, (rx0 , hx ) ∈ M2
1
and (rx+1
, hx ) ∈
/ M2 both belong to ρ2 . Using an argument similar to as in the first case, we can show
1
0
that either the edge (rx+1
, hx ) or the edge (rx+1
, hx ) is labeled (1, 1), and therefore forms a blocking pair
w.r.t. M2 .
12
We now prove part (3) of the lemma. Consider P = hr0 , h0 , . . . , rk−1 , hk−1 i where for each i = 0, . . . , k−1,
M (ri ) = hi . The existence of P in ỸM⊕M ′ implies that there exists an M2 alternating path P2 in G2 . Here
jk−1
, hk−1 i, and for t = 0, . . . , k − 1, jt ∈ {0, 1}.
P2 = hr0j0 , h0 , . . . , rk−1
For the sake of contradiction assume that P contains at least two edges, e1 = (rx , hx−1 ), e2 = (ry , hy−1 )
for some x, y = 1, . . . , k − 1, w.l.o.g. x 6= y, x < y and e1 , e2 are labeled (1, 1). We observe the following about
preferences of rx , ry and hx−1 , hy−1 in G.
(O1 ) rx prefers hx−1
ry prefers hy−1
(O2 ) hx−1 prefers rx
hy−1 prefers ry
over hx = M (rx ).
over hy = M (ry ).
over rx−1 ∈ M (hx−1 ).
over ry−1 ∈ M (hy−1 ).
Using the presence of the edges e1 and e2 labeled (1, 1) in P , we will contradict the stability of M2
j
j
in G2 . Consider the edges e′1 = (rxjx , hx−1 ) and e′2 = (ryy , hy−1 ) in G2 , and since hx−1 = M2 (rxx−1 ) and
jy−1
hy−1 = M2 (ry ), note that e′1 , e′2 ∈
/ M2 .
We first consider the edge e′1 , and consider the four cases that can arise depending on the values of jx
and jx−1 .
1. jx−1 = jx = 0
2. jx−1 = jx = 1
3. jx−1 = 0 and jx = 1
4. jx−1 = 1 and jx = 0
Recall observation (O1 ), and the fact that the residents do not change their preferences in G2 w.r.t the
hospitals originally in G. This implies that in all the four cases above, the resident rxjx prefers hx−1 over
hx = M2 (rxjx ). Using (O2 ) and the fact that a hospital h in G2 prefers level-1 residents over level-0 residents,
jx−1
we can conclude the following. For the cases (1), (2) and (3), hospital hx−1 prefers rxjx over rx−1
, which
jx
implies that the pair (rx , hx−1 ) is labeled (1, 1), which contradicts the stability of Ms (using Lemma 2(1)).
With a similar analysis for the edge e′2 , we conclude that the first three cases do not arise. There is only
one case left to consider, when jx−1 = 1, jx = 0 and jy−1 = 1, jy = 0. As x 6= y, x < y, and jx = 0, jy−1 = 1,
there exists an index ℓ in P2 such that there is a transition from a level-0 resident to a level-1 resident. That
1
is, (rℓ0 , hℓ ) ∈ M2 and (rℓ+1
, hℓ ) ∈
/ M2 both belong to P2 .
We enumerate the possible labels for the edge eℓ = (rℓ+1 , hℓ ) in G.
1
– If eℓ is labeled (1, 1) or (1, −1), then the edge (rℓ+1
, hℓ ) is labeled (1, 1), which contradicts the stability
of Ms (using Lemma 2(1)).
0
– If eℓ is labeled (−1, 1), then the edge (rℓ+1
, hℓ ) is labeled (1, 1), which contradicts the stability of Ms
(using Lemma 2(2)).
⊓
⊔
This completes the proof.
Lemma 4. There is no augmenting path with respect to M in ỸM⊕M ′ .
Proof. Let P = hr0 , h1 , r1 , h2 , . . . , hk−1 , rk−1 , hk i be an augmenting path where for each i = 1, . . . , k −
1, M (ri ) = hi . The existence of P in ỸM⊕M ′ implies that there exists an M2 augmenting path P2 =
jk−1
hr0j0 , h1 , r1j1 , h2 , . . . , hk−1 , rk−1
, hk i in G2 , and for t = 0, . . . , k − 1, jt ∈ {0, 1}.
Using invariants (I1 ), (I2 ), and (I3 ), we conclude that a resident r remains unmatched in M2 when its
level-0 copy is matched to the dummy vertex dr , and the level-1 copy is unmatched in M2 . Therefore the first
resident on the path P2 is a level-1 resident. The second resident on the path r1 has to be a level-1 resident,
otherwise the edge (r01 , h1 ) will be labeled (1, 1), and thus contradict the stability of M2 (using Lemma 2(1)).
This is because r01 prefers being matched to h1 than being unmatched in M2 , and h1 prefers level-1 resident
0
0
is matched to
, hk ) is labeled (1, 1), as rk−1
over a level-0 resident. Observe that jk−1 = 0, else the pair (rk−1
dr (by invariants (I1 ) and (I2 )), which is at the end of its preference list, and hk is unmatched in M ′ .
0
Therefore the path P2 is of the form hr01 , h1 , r11 , h2 , . . . , hk−1 , rk−1
, hk i. As j0 = j1 = 1 and jk−1 = 0,
there exists an index x in P2 such that there is a transition from a level-1 resident to a level-0 resident. That
1
is, (rx0 , hx ) ∈ M2 and (rx−1
, hx ) ∈
/ M2 both belong to P2 .
We enumerate the possible labels for the edge ex = (rx−1 , hx ) in G.
13
1
– If ex is labeled (1, 1) or (1, −1), then the edge (rx−1
, hx ) is labeled (1, 1), which contradicts the stability
of Ms (using Lemma 2(1)).
0
– If ex is labeled (−1, 1), then the edge (rx−1
, hx ) is labeled (1, 1), which contradicts the stability of Ms
(using Lemma 2(2)).
This contradicts our assumption that P is augmenting with respect to M in ỸM⊕M ′ .
⊓
⊔
Lemma 5. There exists no popular matching M ∗ in G such that |M ∗ | > |M |.
Proof. For contradiction, assume that such an assignment M ∗ exists in G. Consider the set YM⊕M ∗ ; recall
that this set contains alternating paths and cycles possibly containing edges labeled (−1, −1). Since |M ∗ | >
|M | there must exist an augmenting path P in YM⊕M ∗ . We first claim that the path P must contain at least
one edge labeled (−1, −1). If not, then the path P is also contained in ỸM⊕M ∗ . However, by Lemma 4 there
is no augmenting path with respect to M in ỸM⊕M ′ for any feasible matching M ′ in G.
We now remove all edges from P which are labeled (−1, −1). This breaks the path into sub-paths say
P1 , P2 , . . . , Pt for some t ≥ 1, where P1 and Pt have one endpoint unmatched in M . Consider the path P1 ;
since P1 does not contain any (−1, −1) edge this implies that P1 ∈ ỸM⊕M ∗ . Without loss of generality,
assume that P1 starts with a resident r which is unmatched in M . Thus using Lemma 3(2), P1 does not
contain any edge labeled (1, 1). Let us denote by ∆(M ∗ , M )P1 the difference between votes of M ∗ and M
restricted to vertices of path P1 . It is clear that ∆(M ∗ , M )P1 < 0. Also, for each i = 2, . . . , t − 1, the
alternating paths Pi have both of their endpoints matched in M . Thus we have ∆(M ∗ , M )Pi ≤ 0 as there
can be at most one (1, 1) edge (by Lemma 3(3)) in these paths, but the endpoints prefer M , as they are
matched in M but not in M ′ . If Pt exists, then a argument similar as given for P1 , we have ∆(M ∗ , M )Pt < 0.
Using these observations, we conclude that M is more popular than M ∗ , a contradiction to the assumption
that M ∗ and M are both popular.
Thus, for any given matching M ∗ such that |M ∗ | > |M |, we know that M is more popular than such a
matching. This completes the proof of the lemma, and shows that the matching M = map(M2 ) is a maximum
cardinality popular matching in G.
⊓
⊔
4.2
Popular matching amongst maximum cardinality matchings
In this section we give an efficient algorithm for computing a matching which is popular amongst the set of
maximum cardinality matchings. The matching M that we output cannot be beaten in terms of votes by any
feasible maximum cardinality matching. Our algorithm uses the generic reduction with a value of s = |R|
= n1 (say). Thus, |Rn1 | = n21 , and |Hn1 | = |H| + O(n21 ). Furthermore, |En1 | = O(mn1 ) where m = |E|.
Thus the running time of the generic algorithm presented earlier with s = n1 for an LCSM+ instance is
O(mn · n1 ) = O(mn2 ) and for a PCSM+ instance is O(mn1 ) = O(mn).
To prove correctness, we show that the matching output by our algorithm is (i) maximum cardinality
and (ii) popular amongst all maximum cardinality feasible matchings. Let M = map(Mn1 ) and M ∗ be any
maximum cardinality feasible matching in G. Consider the set YM⊕M ∗ , and let ρ be an alternating path or
an alternating cycle in YM⊕M ∗ . Let ρn1 = map−1 (ρ, Mn1 ) denote the associated alternating path or cycle in
Gn1 . We observe that every hospital on the path ρn1 is a non-dummy hospital since ρn1 was obtained using
the inverse-map of ρ. We observe two useful properties about such a path or cycle ρn1 in Gn1 . We show that
if for a hospital h ∈ ρn1 , the level of the unmatched resident incident on h is greater than the level of the
matched resident incident on h, then such a level change is gradual, and the associated edge in ρ has the
label (−1, −1). Lemma 6, gives a proof of these.
Lemma 6. Let ρn1 be an alternating path or an alternating cycle in Gn1 and let h be a hospital which has
degree two in ρn1 . Let hraja , h, rbjb i be the sub-path containing h where M (rbjb ) = h. If ja > jb , we claim the
following:
1. ja = jb + 1.
2. The associated edge (ra , h) ∈ ρ is labeled (−1, −1).
14
Proof. We first prove that ja = jb + 1. For contradiction, assume that ja > jb + 1. Observe that h prefers all
the level-ja residents over any level-jb resident. We consider the edge e = (ra , h) in the graph G. We claim
that the label for the edge e cannot be (1, 1) or (1, −1), otherwise the edge (raja , h) is labeled (1, 1) in Gn1 as
the residents do not change their votes. Similarly, we claim that the label for the edge e cannot be (−1, 1)
or (−1, −1), as raja −1 is matched in Mn1 to the last dummy on its preference list, djraa−1 = Mn1 (raja −1 ) (by
invariant (I3 )), and prefers h to djraa−1 , and h prefers all the level-(ja − 1) residents over any level-jb resident.
In this case the edge (raja −1 , h) is labeled (1, 1) in Gn1 , and thus blocks Mn1 (by Lemma 2(2)).
To prove part (b), we assume ja = jb + 1. We enumerate the possible labels for the edge e = (ra , h) in G.
– If e is labeled (1, 1) or (1, −1), then the edge (raja , h) is labeled (1, 1), as raja prefers h over Mn1 (raja ), and h
prefers any level-ja resident over a level-jb resident. Thus, the edge (raja , h) blocks Mn1 (by Lemma 2(1)).
– If e is labeled (−1, 1), then the edge (rajb , h) is labeled (1, 1), as rajb prefers h over djrba = Mn1 (rajb ), and h
prefers rajb over rbjb according to its preference list. Thus, the edge (rajb , h) blocks Mn1 (by Lemma 2(2)).
Thus, the only possible label for the edge (ra , h) is (−1, −1).
⊓
⊔
We use Lemma 7 to prove that M is a maximum cardinality matching in G.
Lemma 7. Let M ∗ be any feasible maximum cardinality matching in G. Then there is no augmenting path
with respect to M in YM⊕M ∗ .
Proof. For the sake of contradiction assume that the path P = hr0 , h1 , r1 , . . . , hk−1 , rk−1 , hk i is an augmenting path where for each i = 1, . . . , (k − 1), M (ri ) = hi . Here r0 is unmatched in M , and hk is
under-subscribed in M . The existence of P in YM⊕M ∗ implies that there exists an Mn1 augmenting path
jk−1
Pn1 = map−1 (P, Mn1 ) = hr0j0 , h1 , r1j1 , . . . , hk−1 , rk−1
, hk i in Gn1 , and for t = 0, . . . , k −1, jt ∈ {0, . . . , n1 −1},
ji
where hi = Mn1 (ri ).
Since r0 is unmatched in M , by invariant (I3 ), it implies that for 0 ≤ i ≤ n1 − 2, Mn1 (r0i ) = dir0 , and
n1 −1
r0
is unmatched in Mn1 . This implies that the first resident in the path Pn1 is a level-(n1 − 1) resident
r0n1 −1 . The second resident on the path Pn1 also has to be a level-(n1 − 1) resident. If not, then the edge
(r0n1 −1 , h1 ) is labeled (1, 1) since h1 prefers r0n1 −1 to any resident at a level lower than n1 − 1 and r0n1 −1 is
jk−1
0
= rk−1
. If not, then the
unmatched in Mn1 . The last resident in the path Pn1 is a level-0 resident i.e. rk−1
0
0
0
edge (rk−1 , hk ) is labeled (1, 1), as rk−1 is matched to the last dummy hospital (drk−1 ) on its preference list
(by invariant (I3 )), and hk is under-subscribed in Mn1 .
Thus, in the path Pn1 , the first two residents are level-(n1 − 1), while the last resident is level-0. Recall
that the path Pn1 was obtained as an inverse-map of the path P in G. Since the path P contains at most n1
residents (possibly all of the residents in G), the path Pn1 also contains at most n1 residents. From Lemma 6
we observe that the difference in the levels of two residents in a sub-path of Pn1 can be at most one. Thus,
it must be the case that residents at all the levels n1 − 1 to 0 are present in Pn1 . However, since there are
two residents at level-(n1 − 1) (first two residents) and one resident at level-0 (last resident), it is clear that
residents at all levels from n1 − 1 to 0 cannot be accommodated in a path containing at most n1 residents.
This contradicts the existence of such a path Pn1 in Gn1 which implies that the assumed augmenting
path P with respect to M cannot exist. This proves that M = map(Mn1 ) is a max-cardinality matching in
G.
⊓
⊔
We can now conclude that the set YM⊕M ∗ is a set of alternating (and not augmenting) paths and
alternating cycles. It remains to show that M is popular amongst all maximum cardinality feasible matchings
in G. Let M ∗ be any feasible maximum cardinality matching in G. In Lemma 8 we show that if there is an
edge (r, h) ∈ M ∗ \ M labeled (1, 1) in ρ, then in ρn1 , for the hospital h, the level of its unmatched neighbour
(resident) is lower than the level of its matched neighbour (resident).
Lemma 8. If an edge (ra , h) ∈ ρ is labeled (1, 1), then in ρn1 for the sub-path hraja , h, rbjb i where Mn1 (rbjb ) =
h, we have ja < jb .
15
Proof. Let an edge e = (ra , h) be labeled (1, 1) in ρ. We observe the following about preferences of ra and h
in G.
(O1 ) ra prefers h over M (ra ).
(O2 ) h prefers ra over rb ∈ M (h), where rb = corr(ra ).
/ Mn1 . Consider the three cases
Consider the edge e′ = (raja , h) in Gn1 , as h = Mn1 (rbjb ) it implies e′ ∈
that can arise depending on the values of ja and jb .
1. ja = jb
2. ja > jb
3. ja < jb
Recall observation (O1 ), and the fact that the residents do not change their preferences in Gn1 w.r.t. the
hospitals originally in G. This implies in all the three cases above, the resident raja prefers h over Mn1 (raja ).
Using (O2 ) and the fact that a hospital h in Gn1 prefers level-p residents over level-q residents, when p > q,
we can conclude the following. For the cases (1) and (2), hospital h prefers raja over rbjb , which implies that
the pair (raja , h) is labeled (1, 1), which is a blocking pair for Mn1 (using Lemma 2(1)). This contradicts the
⊓
⊔
stability of Mn1 . We therefore conclude that ja < jb .
Lemma 9 shows that in an alternating path in YM⊕M ∗ with exactly one endpoint unmatched in M or an
alternating cycle, the number of edges labeled (1, 1) cannot exceed the number of edges labeled (−1, −1).
Lemma 9. Let ρ be an alternating path or an alternating cycle in YM⊕M ∗ . Then the number of edges labeled
(1, 1) in ρ is at most the number of edges labeled (−1, −1).
Proof. Depending on the nature of ρ we have three different cases.
– ρ is an alternating path which starts with an unmatched resident in M .
– ρ is an alternating path which starts with a hospital which is under-subscribed in M .
– ρ is an alternating cycle.
The proof idea is similar in all the three cases. In each of the above, we consider ρn1 = map−1 (ρ, Mn1 ). For
every edge labeled (1, 1) in ρ we show a change (increase / decrease) in the level of the residents which are
neighbours of a particular hospital. We show that each such change must be complemented with another
change (decrease / increase resp.) in the level of the residents which are neighbours to some other hospital.
Finally, we show that the second type of change translates to a (−1, −1) edge in ρ.
Let ρ be an alternating cycle. Consider a hospital hi ∈ ρ for which there is an edge (r, h) labeled (1, 1)
ji−1
, hi , riji i be a sub-path of ρ
incident on it in ρ. Consider the associated hospital hi ∈ ρn1 . W.l.o.g. let hri−1
when traversing ρn1 in counter-clock-wise direction. By Lemma 8, we know that the level of the unmatched
resident incident on hi is lower than the level of the matched resident incident on hi . Thus there is an
increase in level of residents when at hi (while traversing ρn1 in counter-clockwise direction). This is true
for any hk ∈ ρn1 where the associated hospital in ρ has a (1, 1) edge incident on it. We now recall from
Lemma 6(2) that whenever a hospital hk ∈ ρn1 has a level decrease, the associated edge in ρ is labeled
(−1, −1). Furthermore the decrease in levels at a hospital is gradual. Thus, it must be the case that the
number of (1, 1) edges in ρ is at most the number of (−1, −1) edges in ρ.
In case ρ is an alternating path starting at an unmatched resident, we show that in the path ρn1 the
first two residents are level-(n1 − 1) residents (see Claim 1 below for a proof). Furthermore, consider the
first edge (r, h) ∈ ρ that is labeled (1, 1). The associated hospital hi has a increase in the level of its two
neighbouring residents. However, since ρn1 started with two level-(n1 − 1) residents (which is the highest
level possible). Therefore, there must have been some hospital hk preceding hi in ρn1 which has a decrease
in the levels of the two neighbours. Using these facts it is easy to prove the following:
Number of (1, 1) edges in ρ ≤ Number of increases in ρn1 ≤ Number of decreases in ρn1 ≤ Number of
(−1, −1) edges in ρ.
This completes the proof in case ρ is a path starting at an unmatched resident.
Finally, we are left with the case when ρ is an alternating path which starts with a hospital hi which is
under-subscribed in M . We show that in the associated path ρn1 , the first resident is a level-0 resident (see
16
Claim 2 below for a proof). Note that in this case since the path starts at a hospital, whenever we have an
edge labeled (1, 1) in ρ, the associated hospital hi in ρn1 has an increase in the levels of the two neighbouring
residents. Now first resident in the path is at the lowest possible level, it must be the case that there is a
hospital hk preceding hi in ρ for which there is a decrease in the level of the neighbouring residents. Now
using arguments similar to those in the case of path starting at an unmatched resident, we conclude that
the number of (1, 1) edges in ρ is at most the number of (−1, 1) edges in ρ.
Claim 1: ρ = hr0 , h1 , r1 , . . .i starts with an unmatched resident. As r0 is unmatched in M , by invariant
(I3 ), it implies that for 0 ≤ i ≤ n1 − 2, Mn1 (r0i ) = dir , and r0n1 −1 is unmatched in Mn1 . Therefore the first
resident r0j0 on the path ρn1 is a level-(n1 − 1) resident, that is j0 = n1 − 1. Furthermore, the second resident
on the path r1 has to be a level-(n1 − 1) resident. If not, then as r0n1 −1 is unmatched in Mn1 and h1 prefers
a level-(n1 − 1) resident to a level-v resident (v < n1 − 1), the edge (r0n1 −1 , h1 ) will be labeled (1, 1), and
thus blocks Mn1 contradicting its stability (using Lemma 2(1)).
Claim 2: ρ = hh0 , r1 , h1 , . . .i starts with an under-subscribed hospital. The first resident r1j1 on the path ρn1
has to be a level-0 resident, that is j1 = 0. If not, i.e. if j1 = 1, then d0r0 = Mn1 (r10 ) (by invariant (I3 )) d0r0 in
Mn1 , and prefers h1 to d0r0 . The hospital h1 on the other hand is under-subscribed in Mn1 and prefers being
matched to r10 in Mn1 . Thus, the edge (r10 , h1 ) is labeled (1, 1), and blocks Mn1 contradicting its stability
(using Lemma 2(2)).
⊓
⊔
Thus, we get the following theorem:
Theorem 4. Let M = map(Mn1 ) where Mn1 is a stable matching in Gn1 . Then M is a popular matching
amongst all maximum cardinality matchings in G.
Discussion: A natural question is to consider popular matchings in LCSM instances. An LCSM instance
need not admit a stable matching. However we claim that restricted to LCSM instances which admit a stable
matching, our results hold without any modification. To obtain the result, we claim that Lemma 1 holds in
the presence of lower quotas on classes. Additionally, if the given LCSM instance G admits a stable matching,
the graph Gs for s = 1, . . . , n1 also admits a stable matching. We thank Prajakta Nimbhorkar for pointing
this to us.
Acknowledgement: We thank the anonymous reviewers whose comments have improved the presentation.
References
1. D. J. Abraham, R. W. Irving, T. Kavitha, and K. Mehlhorn. Popular Matchings. SIAM Journal on Computing,
37(4):1030–1045, 2007.
2. D. J. Abraham, R. W. Irving, and D. F. Manlove. Two Algorithms for the Student-Project Allocation Problem.
J. of Discrete Algorithms, 5(1):73–90, 2007.
3. P. Biró, D. Manlove, and S. Mittal. Size Versus Stability in the Marriage Problem. Theoretical Computer Science,
411(16-18):1828–1841, 2010.
4. F. Brandl and T. Kavitha. Popular Matchings with Multiple Partners. CoRR, abs/1609.07531, 2016.
5. Á. Cseh and T. Kavitha. Popular Edges and Dominant Matchings. In Proceedings of the Eighteenth Conference
on Integer Programming and Combinatorial Optimization, pages 138–151, 2016.
6. T. Fleiner and N. Kamiyama. A Matroid Approach to Stable Matchings with Lower Quotas. In Proceedings of
the Twenty-third Annual ACM-SIAM Symposium on Discrete Algorithms, pages 135–142, 2012.
7. D. Gale and L. Shapley. College Admissions and the Stability of Marriage. American Mathematical Monthly,
69:9–14, 1962.
8. P. Gärdenfors. Match Making: assignments based on bilateral preferences. Behavioural Sciences, 20:166–173,
1975.
9. D. Gusfield and R. W. Irving. The Stable Marriage Problem: Structure and Algorithms. MIT Press, 1989.
10. C.-C. Huang. Classified Stable Matching. In Proceedings of the Twenty-First Annual ACM-SIAM Symposium
on Discrete Algorithms, pages 1235–1253, 2010.
11. C.-C. Huang and T. Kavitha. Popular Matchings in the Stable Marriage Problem. In Proceedings of 38th
International Colloquium on Automata, Languages and Programming , pages 666–677, 2011.
17
12. N. Kamiyama. Popular Matchings with Two-Sided Preference Lists and Matroid Constraints. Technical Report
MI 2016-13, 2016.
13. T. Kavitha. A Size-Popularity Tradeoff in the Stable Marriage Problem. SIAM Journal on Computing, 43(1):52–
71, 2014.
14. Z. Király. Better and Simpler Approximation Algorithms for the Stable Marriage Problem. Algorithmica,
60(1):3–20, 2011.
A
A.1
Appendix
Reduction from a PCSM+ instance to an SPA instance
An instance of SPA [2] consists of students, projects and lecturers. Each lecturer has an upper bound on
the maximum number of students that he/she is willing to advise. Each project has an upper bound on the
number of students it can accommodate. Each project is owned by exactly one lecturer. Each student has a
preference ordering over a subset of the projects, and each lecturer has a preference over the students.
We detail on the reduction from PCSM+ instance to an SPA instance here. For a resident r in the PCSM+
instance, a corresponding student sr is introduced in the SPA instance. For each hospital h, a lecturer lh with
capacity q(h) is added in the SPA instance. For each class Cjh in the classification provided by a hospital h,
a project pj is associated with the lecturer lh , and the upper-bound of pj is equal to q(Cjh ). The preference
list of lh is obtained from its corresponding hospital h. If the resident r is the k-th most preferred resident
in the preference list of h, then the student sr is the k-th most preferred student in the preference list of lh .
Similarly, the preference list of a student sr is created from its corresponding resident r. Let Cjh be the class
that the resident r appears in the classification provided by the k-th most preferred hospital in its preference
list, then pj is the k-th most preferred project in the preference list for sr . As the classifications associated
with every hospital in the PCSM+ instance are a partition over its preference list, there is no ambiguity in
describing the preference of the students.
18
| 8cs.DS
|
It Takes (Only) Two:
Adversarial Generator-Encoder Networks
arXiv:1704.02304v3 [cs.CV] 6 Nov 2017
Dmitry Ulyanov
Skolkovo Institute of Science and Technology, Yandex
[email protected]
Andrea Vedaldi
University of Oxford
[email protected]
Victor Lempitsky
Skolkovo Institute of Science and Technology
[email protected]
Abstract
We present a new autoencoder-type architecture that is trainable in an unsupervised
mode, sustains both generation and inference, and has the quality of conditional and
unconditional samples boosted by adversarial learning. Unlike previous hybrids of
autoencoders and adversarial networks, the adversarial game in our approach is set
up directly between the encoder and the generator, and no external mappings are
trained in the process of learning. The game objective compares the divergences of
each of the real and the generated data distributions with the prior distribution in
the latent space. We show that direct generator-vs-encoder game leads to a tight
coupling of the two components, resulting in samples and reconstructions of a
comparable quality to some recently-proposed more complex architectures.
1
Introduction
Deep (Variational) Auto Encoders (AEs [2] and VAEs [14, 24]) and deep Generative Adversarial
Networks (GANs [8]) are two of the most popular approaches to generative learning. These methods
have complementary strengths and weaknesses. VAEs can learn a bidirectional mapping between
a complex data distribution and a much simpler prior distribution, allowing both generation and
inference; on the contrary, the original formulation of GAN learns a unidirectional mapping that only
allows sampling the data distribution. On the other hand, GANs use more complex loss functions
compared to the simplistic data-fitting losses in (V)AEs and can usually generate more realistic
samples.
Several recent works have looked for hybrid approaches to support, in a principled way, both sampling
and inference like AEs, while producing samples of quality comparable to GANs. Typically this is
achieved by training a AE jointly with one or more adversarial discriminators whose purpose is to
improve the alignment of distributions in the latent space [3, 19], the data space [4, 17] or in the joint
(product) latent-data space [5, 6]. Alternatively, the method of [31] starts by learning a unidirectional
GAN, and then learns a corresponding inverse mapping (the encoder) post-hoc.
While compounding autoencoding and adversarial discrimination does improve GANs and VAEs, it
does so at the cost of added complexity. In particular, each of these systems involves at least three
deep mappings: an encoder, a decoder/generator, and a discriminator. In this work, we show that
this is unnecessary and that the advantages of autoencoders and adversarial training can be combined
without increasing the complexity of the model.
The source code is available at https://github.com/DmitryUlyanov/AGE
X
e(X)
Encoder e
Z
Generator g
e(g(Z))
g(Z)
Latent space
Data space
Figure 1: Our model (AGE network) has only two components: the generator g and the encoder
e. The learning process adjusts their parameters in order to align a simple prior distribution Z in
the latent space and the data distribution X. This is done by adversarial training, as updates for
the generator aim to minimize the divergence between e(g(Z)) and Z (aligning green with gray),
while updates for the encoder aim to minimize the divergence between e(X) (aligning blue with
gray) and to maximize the divergence between e(g(Z)) and Z (shrink green “away” from gray). We
demonstrate that such adversarial learning gives rise to high-quality generators that result in the close
match between the real distribution X and the generated distribution g(Z). Our learning can also
incorporate reconstruction losses to ensure that encoder-generator acts as autoencoder (section 2.2).
In order to do so, we propose a new architecture, called an Adversarial Generator-Encoder (AGE)
Network (section 2), that contains only two feed-forward mappings, the encoder and the generator,
operating in opposite directions. As in VAEs, the generator maps a simple prior distribution in latent
space to the data space, while the encoder is used to move both the real and generated data samples
into the latent space. In this manner, the encoder induces two latent distributions, corresponding
respectively to the encoded real data and the encoded generated data. The AGE learning process
then considers the divergence of each of these two distributions to the original prior distribution.
There are two advantages of this approach. First, due to the simplicity of the prior distribution,
computing its divergence to the latent data distributions reduces to the calculation of simple statistics
over small batches of images. Second, unlike GAN-like approaches, real and generated distributions
are never compared directly, thus bypassing the need for discriminator networks as used by GANs.
Instead, the adversarial signal in AGE comes from learning the encoder to increase the divergence
between the latent distribution of the generated data and the prior, which works against the generator,
which tries to decrease the same divergence (Figure 1). Optionally, AGE training may include
reconstruction losses typical of AEs.
The AGE approach is evaluated (section 3) on a number of standard image datasets, where we show
that the quality of generated samples is comparable to that of GANs [8, 23], and the quality of
reconstructions is comparable or better to that of the more complex Adversarially-Learned Inference
(ALI) approach of [6], while training faster. We further evaluate the AGE approach in the conditional
setting, where we show that it can successfully tackle the colorization problem that is known to be
difficult for GAN-based approaches. Our findings are summarized in section 4.
Other related work. Apart from the above-mentioned approaches, AGE networks can be related
to several other recent GAN-based systems. Thus, they are related to improved GANs [26] that
proposed to use batch-level information in order to prevent mode collapse. The divergences within
AGE training are also computed as batch-level statistics.
Another avenue for improving the stability of GANs has been the replacement of the classifying
discriminator with the regression-based one as in energy-based GANs [30] and Wasserstein GANs [1].
Our statistics (the divergence from the prior distribution) can be seen as a very special form of
regression. In this way, the encoder in the AGE architecture can be (with some reservations) seen as
a discriminator computing a single number similarly to how it is done in [1, 30].
2
Adversarial Generator-Encoder Networks
This section introduces our Adversarial Generator-Encoder (AGE) networks. An AGE is composed
of two parametric mappings: the encoder eψ (x), with the learnable parameters ψ, that maps the data
space X to the latent space Z, and the generator gθ (z), with the learnable parameters θ, which runs
2
in the opposite direction. We will use the shorthand notation f (Y ) to denote the distribution of the
random variable f (y), y ∼ Y .
The reference distribution Z is chosen so that it is easy to sample from it, which in turns allow
to sample gθ (Z) unconditionally be first sampling z ∼ Z and then by feed-forward evaluation of
x = gθ (z), exactly as it is done in GANs. In our experiments, we pick the latent space Z to be an
M -dimensional sphere SM , and the latent distribution to be a uniform distribution on that sphere
Z = Uniform(SM ). We have also conducted some experiments with the unit Gaussian distribution
in the Euclidean space and have obtained results comparable in quality.
The goal of learning an AGE is to align the real data distribution X to the generated distribution gθ (Z)
while establishing a correspondence between data and latent samples x and z. The real data distribution X is empirical and represented by a large number N of data samples {x1 , x2 , ...xN }. Learning
amounts to tuning the parameter ψ and θ to optimize the AGE criterion, discussed in section 2.1.
This criterion is based on an adversarial game whose saddle points correspond to networks that align
real and generated data distribution (g(Z) = X). The criterion is augmented with additional terms
that encourage the reciprocity of the encoder e and the generator g (section 2.2). The details of the
training procedure are given in section 2.3.
2.1
Adversarial distribution alignment
The GAN approach to aligning two distributions is to define an adversarial game based on a ratio
of probabilities [8]. The ratio is estimated by repeatedly fitting a binary classifier that distinguishes
between samples obtained from the real and generated data distributions. Here, we propose an
alternative adversarial setup with some advantages with respect to GAN’s, including avoiding
generator collapse [7].
The goal of AGE is to generate a distribution g(Z) in data space that is close to the true data
distribution X. However, direct matching of the distributions in the high-dimensional data space, as
done in GAN, can be challenging. We propose instead to move this comparison to the simpler latent
space. This is done by introducing a divergence measure ∆(P kQ) between distributions defined
in the latent space Z. We only require this divergence to be non-negative and zero if, and only
if, the distributions are identical (∆(P kQ) = 0 ⇐⇒ P = Q).1 The encoder function e maps the
distributions X and g(Z) defined in data space to corresponding distributions e(X) and e(g(Z)) in
the latent space. Below, we show how to design an adversarial criterion such that minimizing the
divergence ∆(e(X), e(g(Z))) in latent space induces the distributions X and g(Z) to align in data
space as well.
In the theoretical analysis below, we assume that encoders and decoders span the class of all
measurable mappings between the corresponding spaces. This assumption, often referred to as
non-parametric limit, is justified by the universality of neural networks [10]. We further make the
assumption that there exists at least one “perfect” generator that matches the data distribution, i.e.
∃g0 : g0 (Z) = X.
We start by considering a simple game with objective defined as:
max min V1 (g, e) = ∆( e(g(Z))ke(X) ) .
e
g
(1)
As the following theorem shows, perfect generators form saddle points (Nash equilibria) of the
game (1) and all saddle points of the game (1) are based on perfect generators.
Theorem 1. A pair (g ∗ , e∗ ) forms a saddle point of the game (1) if and only if the generator g ∗
matches the data distribution, i.e. g ∗ (Z) = X.
The proofs of this and the following theorems are given in the supplementary material.
While the game (1) is sufficient for aligning distributions in the data space, finding such saddle points
is difficult due to the need of comparing two empirical (hence non-parametric) distributions e(X)
and e(g(Z)). We can avoid this issue by introducing an intermediate reference distribution Y and
comparing the distributions to that instead, resulting in the game:
max min V2 (g, e) = ∆(e(g(Z))kY ) − ∆(e(X)kY ).
e
1
g
We do not require the divergence to be a distance.
3
(2)
Importantly, (2) still induces alignment of real and generated distributions in data space:
Theorem 2. If a pair (g ∗ , e∗ ) is a saddle point of game (2) then the generator g ∗ matches the data
distribution, i.e. g ∗ (Z) = X. Conversely, if the generator g ∗ matches the data distribution, then for
some e∗ the pair (g ∗ , e∗ ) is a saddle point of (2).
The important benefit of formulation (2) is that, if Y is selected in a suitable manner, it is simple
to compute the divergence of Y to the empirical distributions e(g(Z)) and e(X). For convenience,
in particular, we choose Y to coincide with the “canonical” (prior) distribution Z. By substituting
Y = Z in objective (2), the loss can be extended to include reconstruction terms that can improve
the quality of the result. It can also be optimized by using stochastic approximations as described in
section 2.3.
Given a distribution Q in data space, the encoder e and divergence ∆(·kY ) can be interpreted as
extracting statistics F (Q) = ∆(e(Q)kY ) from Q. Hence, game (2) can be though of as comparing
certain statistics of the real and generated data distributions. Similarly to GANs, these statistics are
not fixed but evolve during learning.
We also note that, even away from the saddle point, the minimization ming V2 (g, e) for a fixed e does
not tend to collapse for many reasonable choice of divergence (e.g. KL-divergence). In fact, any
collapsed distribution would inevitably lead to a very high value of the first term in (2). Thus, unlike
GANs, our approach can optimize the generator for a fixed adversary till convergence and obtain a
non-degenerate solution. On the other hand, the maximization maxe V2 (g, e) for some fixed g can
lead to +∞ score for some divergences.
2.2
Encoder-generator reciprocity and reconstruction losses
In the previous section we have demonstrated that finding a saddle point of (2) is sufficient to align
real and generated data distributions X and g(Z) and thus generate realistically-looking data samples.
At the same time, this by itself does not necessarily imply that mappings e and g are reciprocal.
Reciprocity, however, can be desirable if one wishes to reconstruct samples x = g(z) from their
codes z = e(x).
In this section, we introduce losses that encourage encoder and generator to be reciprocal. Reciprocity
can be measured either in the latent space or in the data space, resulting in the loss functions based on
reconstruction errors, e.g.:
LX (gθ , eψ ) = Ex∼X kx − gθ eψ (x) k1 ,
(3)
2
LZ (gθ , eψ ) = Ez∼Z kz − eψ gθ (z) k2 .
(4)
Both losses (3) and (4) thus encourage the reciprocity of the two mappings. Note also that (3) is the
traditional pixelwise loss used within AEs (L1-loss was preferred, as it is known to perform better in
image synthesis tasks with deep architectures).
A natural question then is whether it is helpful to minimize both losses (3) and (4) at the same time or
whether considering only one is sufficient. The answer is given by the following statement:
W and Q be aligned by the mapping f (i.e. f (W ) = Q) and
Theorem 3. Let the two distributions
let Ew∼W kw − h f (w) k22 = 0. Then, for w ∼ W and q ∼ Q, we have w = h(f (w)) and
q = f (h(q)) almost certainly, i.e. the mappings f and h invert each other almost everywhere on the
supports of W and Q. Furthermore, Q is aligned with W by h, i.e. h(Q) = W .
Recall that Theorem 2 establishes that the solution (saddle point) of game (2) aligns distributions
in the data space. Then Theorem 3 shows that when augmented with the latent space loss (4), the
objective (2) is sufficient to ensure reciprocity.
2.3
Training AGE networks
Based on the theoretical analysis derived in the previous subsections, we now suggest the approach to
the joint training of the generator in the encoder within the AGE networks. As in the case of GAN
training, we set up the learning process for an AGE network as a game with the iterative updates over
the parameters θ and ψ that are driven by the optimization of different objectives. In general, the
optimization process combines the maximin game for the functional (2) with the optimization of the
reciprocity losses (3) and (4).
4
(a) Real images
(b) AGE samples
(c) [Real, AGE reconstr.] (d) [Real, ALI reconstr.]
Figure 2: Samples (b) and reconstructions (c) for Tiny ImageNet dataset (top) and SVHN dataset
(bottom). The results of ALI [6] on the same datasets are shown in (d). In (c,d) odd columns show
real examples and even columns show their reconstructions. Qualitatively, our method seems to
obtain more accurate reconstructions than ALI [6], especially on the Tiny ImageNet dataset, while
having samples of similar visual quality.
In particular, we use the following game objectives for the generator and the encoder:
θ̂ = arg min V2 (gθ , eψ̄ ) + λLZ (gθ , eψ̄ ) ,
θ
ψ̂ = arg max (V2 (gθ̄ , eψ ) − µLX (gθ̄ , eψ )) ,
ψ
(5)
(6)
where ψ̄ and θ̄ denote the value of the encoder and generator parameters at the moment of the
optimization and λ, µ is a user-defined parameter. Note that both objectives (5), (6) include only
one of the reconstruction losses. Specifically, the generator objective includes only the latent space
reconstruction loss. In the experiments, we found that the omission of the other reconstruction loss
(in the data space) is important to avoid possible blurring of the generator outputs that is characteristic
to autoencoders. Similarly to GANs, in (5), (6) we perform only several steps toward optimum at
each iteration, thus alternating between generator and encoder updates.
By maximizing the difference between ∆(eψ (gθ̄ (Z))kZ) and ∆(eψ (X)kZ), the optimization process
(6) focuses on the maximization of the mismatch between the real data distribution X and the
distribution of the samples from the generator gθ̄ (Z). Informally speaking, the optimization (6) forces
the encoder to find the mapping that aligns real data distribution X with the target distribution Z,
while mapping non-real (synthesized data) gθ̄ (Z) away from Z. When Z is a uniform distribution on
a sphere, the goal of the encoder would be to uniformly spread the real data over the sphere, while
cramping as
much of synthesized data as possible together assuring non-uniformity of the distribution
eψ gθ̄ (Z) .
Any differences (misalignment) between the two distributions are thus amplified by the optimization
process (6) and force the optimization process (5) to focus specifically on removing these differences.
Since the misalignment between X and g(Z) is measured after projecting the two distributions into
the latent space, the maximization of this misalignment makes the encoder to compute features that
distinguish the two distributions.
3
Experiments
We have validated AGE networks in two settings. A more traditional setting involves unconditional
generation and reconstruction, where we consider a number of standard image datasets. We have
5
also evaluated AGE networks in the conditional setting. Here, we tackle the problem of image
colorization, which is hard for GANs. In this setting, we condition both the generator and the encoder
on the gray-scale image. Taken together, our experiments demonstrate the versatility of the AGE
approach.
3.1
Unconditionally-trained AGE networks
Network architectures: In our experiments, the generator and the encoder networks have a similar
structure to the generator and the discriminator in DCGAN [23]. To turn the discriminator into the
encoder, we have modified it to output an M -dimensional vector and replaced the final sigmoid layer
with the normalization layer that projects the points onto the sphere.
Divergence measure: As we need to measure the divergence between the empirical distribution and
the prior distribution in the latent space, we have used the following measure. Given a set of samples
on the M -dimensional sphere, we fit the Gaussian Normal distribution with diagonal covariance
matrix in the embedding M -dimensional space and we compute the KL-divergence of such Gaussian
with the unit Gaussian as
KL(QkN (0, I)) = −
M
M
1 X s2j + m2j
+
− log(sj )
2
M j=1
2
(7)
where mj and sj are the means and the standard deviations of the fitted Gaussians along various
dimensions. Since the uniform distribution on the sphere will entail the lowest possible divergence
with the unit Gaussian in the embedding space among all distributions on the unit sphere, such
divergence measure is valid for our analysis above. We have also tried to measure the same divergence non-parametrically using Kozachenko-Leonenko estimator [15]. In our initial experiments,
both versions worked equally well, and we used a simpler parametric estimator in the presented
experiments.
Hyper-parameters: We use ADAM [13] optimizer with the learning rate of 0.0002. We perform two generator updates per one encoder update for all datasets. For each dataset we tried
λ ∈ {500, 1000, 2000} and picked the best one. We ended up using µ = 10 for all datasets. The
dimensionality M of the latent space was manually set according to the complexity of the dataset.
We thus used M = 64 for CelebA and SVHN datasets, and M = 128 for the more complex datasets
of Tiny ImageNet and CIFAR-10.
Results: We evaluate unconditional AGE networks on several standard datasets, while treating the
system [6] as the most natural reference for comparison (as the closest three-component counterpart
to our two-component system). The results for [6] are either reproduced with the author’s code or
copied from [6].
In Figure 2, we present the results on the challenging Tiny ImageNet dataset [25] and the SVHN
dataset [21]. We show both samples g(z) obtained for z ∼ Z as well as the reconstructions g e(x)
alongside the real data samples x. We also show the reconstructions obtained by [6] for comparison.
Inspection reveals that the fidelity of [6] is considerably lower for Tiny ImageNet dataset.
In Figure 3, we further compare the reconstructions of CelebA [18] images obtained by the AGE
network, ALI [6], and VAE [14]. Overall, the fidelity and the visual quality of AGE reconstructions
are roughly comparable or better than ALI. Furthermore, ALI takes notoriously longer time to
converge than our method, and the reconstructions of ALI after 10 epochs (which take six hours) of
training look considerably worse than AGE reconstructions after 10 epochs (which take only two
hours), thus attesting to the benefits of having a simpler two-component system.
Next we evaluate our method quantitatively. For the model trained on CIFAR-10 dataset we compute
Inception score [26]. The AGE score is 5.90 ± 0.04, which is higher than the ALI [6] score of
5.34 ± 0.05 (as reported in [28]) and than the score of 4.36 ± 0.04 from [26]. The state-of-the-art
from [28] is higher still (7.72 ± 0.13). Qualitative results of AGE for CIFAR-10 and other datasets
are shown in supplementary material.
We also computed log likelihood for AGE and ALI on the MNIST dataset using the method of [29]
with latent space of size 10 using authours source code. ALI’s score is 721 while AGE’s score is 746.
The AGE model is also superior than both VAE and GAN, which scores are 705.375 and 346.679
respectively as evaluated by [29].
6
Orig.
AGE
10 ep.
ALI
10 ep.
ALI
100 ep.
VAE
Orig.
AGE
10 ep.
ALI
10 ep.
ALI
100 ep.
VAE
Orig.
AGE
10 ep.
ALI
10 ep.
ALI
100 ep.
VAE
Figure 3: Reconstruction quality comparison of our method with ALI [6] and VAE [14]. The first
column in each set shows examples from the test set of CelebA dataset. In the other columns the
reconstructions for different methods are presented: column two for ours method, three and four for
ALI and five for VAE. We train our model for 10 epochs and compare to ALI, trained for the same
number of epochs (column three). Importantly one epoch for our method takes 3 times less time than
for ALI. For a fair comparison we also present the results of ALI, trained till convergence.
Finally, similarly to [6, 5, 23] we investigated whether the learned features are useful for discriminative
tasks. We reproduced the evaluation pipeline from [6] for SVHN dataset and obtained 23.7% error
rate in the unsupervised feature learning protocol with our model, while their result is 19.14%. At
the moment, it is unclear to us why AGE networks underperform ALI at this task.
3.2
Conditional AGE network experiments.
Recently, several GAN-based systems have achieved very impressive results in the conditional setting,
where the latent space is augmented or replaced with a second data space corresponding to different
modality [11, 32]. Arguably, it is in the conditional setting where the bi-directionality lacking in
conventional GANs is most needed. In fact, by allowing to switch back-and-forth between the data
space and the latent space, bi-directionality allows powerful neural image editing interfaces [31, 3].
Here, we demonstrate that AGE networks perform well in the conditional setting. To show that, we
have picked the image colorization problem, which is known to be hard for GANs. To the best of our
knowledge, while the idea of applying GANs to the colorization task seems very natural, the only
successful GAN-based colorization results were presented in [11], and we compare to the authors’
implementation of their pix2pix system. We are also aware of several unsuccessful efforts to use
GANs for colorization.
To use AGE for colorization, we work with images in the Lab color space, and we treat the ab color
channels of an image as a data sample x. We then use the lightness channel L of the image as an
input to both the encoder eψ (x|L) and the generator gθ (z|L), effectively conditioning the encoder
and the generator on it. Thus, different latent variables z will result in different colorizations x for
the same grayscale image L. The latent space in these experiments is taken to be three-dimensional.
The particular architecture of the generator takes the input image L, augments it with z variables
expanded to constant maps of the same spatial dimensions as L, and then applies the ResNet type
architecture [9, 12] that computes x (i.e. the ab-channels). The encoder architecture is a convolutional
network that maps the concatenation of L and x (essentially, an image in the Lab-space) to the latent
space. The divergence measure is the same as in the unconditional AGE experiments and is computed
7
(a) Colorizations – AGE network (top rows) vs. pix2pix [11] (bottom rows)
(b) Color transfer
Figure 4: (a) Each pane shows colorizations of the input grayscale image (emphasized) using
conditional AGE networks (top rows) and pix2pix [11] with added noise maps (bottom rows). AGE
networks produce diverse colorizations, which are hard to obtain using pix2pix. (b) In each row we
show the result of color transfer using the conditional AGE network. The color scheme from the first
image is transferred onto the second image.
“unconditionally” (i.e. each minibatch passed through the encoder combines multiple images with
different L).
We perform the colorization experiments on Stanford Cars dataset [16] with 16,000 training images
of 196 car models, since cars have inherently ambiguous colors and hence their colorization is
particularly prone to the regression-to-mean effect. The images were downsampled to 64×64.
We present colorization results in Figure 4. Crucially, AGE generator is often able to produce
plausible and diverse colorizations for different latent vector inputs. As we wanted to enable pix2pix
GAN-based system of [11] to produce diverse colorizations, we augmented the input to their generator
architecture with three constant-valued maps (same as in our method). We however found that their
system effectively learns to ignore such input augmentation and the diversity of colorizations was
very low (Figure 4a).
To demonstrate the meaningfulness of the latent space learned by the conditional AGE training, we
also demonstrate the color transfer examples, where the latent vector z1 = eψ (x1 |L1 ) obtained by
encoding the image [x1 , L1 ] is then used to colorize the grayscale image L2 , i.e. x2 = gθ (z1 |L2 )
(Figure 4b).
4
Conclusion
We have introduced a new approach for simultaneous learning of generation and inference networks.
We have demonstrated how to set up such learning as an adversarial game between generation and
inference, which has a different type of objective from traditional GAN approaches. In particular the
objective of the game considers divergences between distributions rather than discrimination at the
level of individual samples. As a consequence, our approach does not require training a discriminator
network and enjoys relatively quick convergence.
We demonstrate that on a range of standard datasets, the generators obtained by our approach provides
high-quality samples, and that the reconstructions of real data samples passed subsequently through
the encoder and the generator are of better fidelity than in [6]. We have also shown that our approach
is able to generate plausible and diverse colorizations, which is not possible with the GAN-based
system [11].
8
Our approach leaves a lot of room for further experiments. In particular, a more complex latent space
distribution can be chosen as in [19], and other divergence measures between distributions can be
easily tried.
References
[1] Martín Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein GAN. Proc. ICLR, 2017.
[2] Yoshua Bengio. Learning deep architectures for AI. Foundations and Trends in Machine
Learning, 2(1):1–127, 2009.
[3] Andrew Brock, Theodore Lim, James M. Ritchie, and Nick Weston. Neural photo editing with
introspective adversarial networks. Proc. ICLR, 2017.
[4] Tong Che, Yanran Li, Athul Paul Jacob, Yoshua Bengio, and Wenjie Li. Mode regularized
generative adversarial networks. Proc. ICLR, 2017.
[5] Jeff Donahue, Philipp Krähenbühl, and Trevor Darrell. Adversarial feature learning. Proc.
ICLR, 2017.
[6] Vincent Dumoulin, Ishmael Belghazi, Ben Poole, Alex Lamb, Martín Arjovsky, Olivier Mastropietro, and Aaron C. Courville. Adversarially learned inference. Proc. ICLR, 2017.
[7] Ian J. Goodfellow. NIPS 2016 tutorial: Generative adversarial networks. CoRR, abs/1701.00160,
2017.
[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil
Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. NIPS,
pages 2672–2680, 2014.
[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. In Proc. CVPR, pages 770–778, 2016.
[10] Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are
universal approximators. Neural Networks, 2(5):359 – 366, 1989.
[11] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with
conditional adversarial networks. In Proc. CVPR, 2017.
[12] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer
and super-resolution. In European Conference on Computer Vision, 2016.
[13] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proc. ICLR,
2015.
[14] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. Proc. ICLR, 2014.
[15] L. F. Kozachenko and N. N. Leonenko. Sample estimate of the entropy of a random vector.
Probl. Inf. Transm., 23(1-2):95–101, 1987.
[16] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for
fine-grained categorization. In Proc.ICCV 3DRR Workshop, pages 554–561, 2013.
[17] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, and Ole Winther. Autoencoding beyond
pixels using a learned similarity metric. CoRR, abs/1512.09300, 2015.
[18] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the
wild. In ICCV, pages 3730–3738. IEEE Computer Society, 2015.
[19] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, and Ian J. Goodfellow. Adversarial
autoencoders. Proc. ICLR, 2016.
[20] Youssef Marzouk, Tarek Moselhy, Matthew Parno, and Alessio Spantini. An introduction to
sampling via measure transport. arXiv preprint arXiv:1602.05023, 2016.
[21] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng.
Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on
Deep Learning and Unsupervised Feature Learning 2011, 2011.
[22] G. Owen. Game Theory. Academic Press, 1982.
[23] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with
deep convolutional generative adversarial networks. Proc. ICLR, 2016.
9
[24] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation
and approximate inference in deep generative models. arXiv preprint arXiv:1401.4082, 2014.
[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Fei-Fei
Li. Imagenet large scale visual recognition challenge. International Journal of Computer Vision,
115(3):211–252, 2015.
[26] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training gans. In Advances in Neural Information Processing Systems
(NIPS), pages 2226–2234, 2016.
[27] Cédric Villani. Optimal transport: old and new, volume 338. Springer Science & Business
Media, 2008.
[28] David Warde-Farley and Yoshua Bengio. Improving generative adversarial networks with
denoising feature matching. In Proc. ICLR, 2017.
[29] Yuhuai Wu, Yuri Burda, Ruslan Salakhutdinov, and Roger B. Grosse. On the quantitative
analysis of decoder-based generative models. Proc ICLR, 2017.
[30] Junbo Jake Zhao, Michaël Mathieu, and Yann LeCun. Energy-based generative adversarial
network. Proc. ICLR, 2017.
[31] Jun-Yan Zhu, Philipp Krähenbühl, Eli Shechtman, and Alexei A. Efros. Generative visual
manipulation on the natural image manifold. In Proc. ECCV, 2016.
[32] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image
translation using cycle-consistent adversarial networks. CoRR, abs/1703.10593, 2017.
10
5
Appendix
In this supplementary material, we provide proofs for the theorems of the main text (restating these
theorems for convenience of reading). We also show additional qualitative results on several datasets.
A
Proofs
Let X and Z be distributions defined in the data and the latent spaces X , Z correspondingly. We
assume X and Z are such, that there exists an invertible almost everywhere function e which
transforms the latent distribution into the data one g(Z) = X. This assumption is weak, since for
every atomless (i.e. no single point carries a positive mass) distributions X, Z such invertible function
exists. For a detailed discussion on this topic please refer to [20, 27]. Since Z is up to our choice
simply setting it to Gaussian distribution (for Z = RM ) or uniform on sphere for (Z = SM ) is good
enough.
Lemma A.1. Let X and Y to be two distributions defined in the same space. The distributions are
equal X = Y if and only if e(X) = e(Y ) holds for for any measurable function e : X → Z.
Proof. It is obvious, that if X = Y then e(X) = e(Y ) for any measurable function e.
Now let e(X) = e(Y ) for any measurable e. To show that X = Y we will assume converse:
X 6= Y . Then there exists a set B ∈ FX , such that 0 < PX (B) 6= PY (B) and a function
e, such that corresponding set C = e(B) has B as its preimage B = e−1 (C). Then we have
PX (B) = Pe(X) (C) = Pe(Y ) (C) = PY (B), which contradicts with the previous assumption.
Lemma A.2. Let (g 0 , e0 ) and (g ∗ , e∗ ) to be two different Nash equilibria in a game
ming maxe V (g, e). Then V (g, e) = V (g 0 , e0 ).
Proof. See chapter 2 of [22].
Theorem 1. For a game
max min V1 (g, e) = ∆( e(g(Z))ke(X) )
e
g
(8)
(g ∗ , e∗ ) is a saddle point of (8) if and only if g ∗ is such that g ∗ (Z) = X.
Proof. First note that V1 (g, e) ≥ 0. Consider g such that g(Z) = X, then for any e: V1 (g, e) = 0.
We conclude that (g, e) is a saddle point since V1 (g, e) = 0 is a maximum over e and minimum over
g.
Using lemma A.2 for saddle point (g ∗ , e∗ ): V1 (g ∗ , e∗ ) = 0 = maxe V1 (g ∗ , e), which is only possible
if for all e: V1 (g ∗ , e) = 0 from which immediately follows g(Z) = X by lemma A.1.
Lemma A.3. Let function e : X → Z be X-almost everywhere invertible, i.e.
∃e−1 : PX ({x 6= e−1 (e(x))}) = 0. Then if for a mapping g : Z → X holds e(g(Z)) = e(X), then
g(Z) = X.
Proof. From definition of X-almost everywhere invertibility follows PX (A) = PX (e−1 (e(A))) for
any set A ∈ FX . Then:
PX (A) = PX (e−1 (e(A))) = Pe(X) (e(A)) =
= Pe(g(Z)) (e(A)) = Pg(Z) (A).
Comparing the expressions on the sides we conclude g(Z) = X.
Theorem 2. Let Y to be any fixed distribution in the latent space. Consider a game
max min V2 (g, e) = ∆(e(g(Z))kY ) − ∆(e(X)kY ) .
e
g
(9)
If the pair (g ∗ , e∗ ) is a Nash equilibrium of game (9) then g ∗ (Z) ∼ X. Conversely, if the fake and
real distributions are aligned g ∗ (Z) ∼ X then (g ∗ , e∗ ) is a saddle point for some e∗ .
11
Proof.
• As for a generator which aligns distributions g(Z) = X: V2 (g, e) = 0 for any e we conclude
by A.2 that the optimal game value is V2 (g ∗ , e∗ ) = 0. For an optimal pair (g ∗ , e∗ ) and
arbitrary e0 from the definition of equilibrium:
0 = ∆(e∗ (g ∗ (Z))kY ) − ∆(e∗ (X)kY ) ≥
≥ ∆(e0 (g ∗ (Z))kY ) − ∆(e0 (X)kY ) .
(10)
For invertible almost everywhere encoder e0 such that ∆(e0 (X)kY ) = 0 the first term is
zero ∆(e0 (g ∗ (Z))kY ) = 0 since inequality (10) and then e0 (g ∗ (Z)) = e0 (X) = Y . Using
result of the lemma A.3 we conclude, that g ∗ (Z) = X.
• If g ∗ (Z)
=
X then ∀e
:
V2 (g ∗ , e∗ ) = V2 (g ∗ , e) = 0 = maxe0 V2 (g ∗ , e0 ).
e(g ∗ (Z))
=
e(X)
and
The corresponding optimal encoder e∗ is such that g ∗ ∈ arg ming ∆(e∗ (g(Z))kY ).
Note that not for every optimal encoder e∗ the distributions e∗ (X) and e∗ (g ∗ (Z)) are
aligned with Y . For example if e∗ collapses X into two points then for any distribution X: e∗ (X) = e∗ (g ∗ (Z)) = Bernoulli(p). For the optimal generator g ∗ the parameter p is such, that for all other generators g 0 such that e∗ (g 0 (Z)) ∼ Bernoulli(p0 ):
∆(e∗ (g ∗ (Z))kY ) ≤ ∆(e∗ (g 0 (Z))kY ).
W and Q be aligned by the mapping f (i.e. f (W ) = Q) and
Theorem 3. Let the two distributions
let Ew∼W kw − h f (w) k2 = 0. Then, for w ∼ W and q ∼ Q, we have w = h(f (w)) and
q = f (h(q)) almost certainly, i.e. the mappings f and h invert each other almost everywhere on the
supports of W and Q. More, Q is aligned with W by h: h(Q) = W .
Proof. Since Ew∼W kw − h f (w) k2 = 0, we have w = h(f (w)) almost certainly for w ∼ W .
Using this and the fact that f (w) ∼ Q for w ∼ W we derive:
Eq∼Q kq − f h(q) k2 = Ew∼W kf (w) − f h(f (w)) k2 =
= Ew∼W kf (w) − f (w)k2 = 0 .
Thus q = f (h(q)) almost certainly for q ∼ Q.
To show alignment h(Q) = W first recall the definition of alignment. Distributions are aligned
f (W ) = Q iff ∀Q̄ ∈ FQ : PQ (Q̄) = PW (f −1 (Q̄)). Then ∀W̄ ∈ FW we have
PW (W̄ ) = PW (h(f (W̄ ))) = PW (f −1 (f (W̄ ))) =
= PQ (f (W̄ )) = PQ (h−1 (W̄ )) .
Comparing the expressions on the sides we conclude h(Q) = W .
B
Additional qualitative results.
In the figures, we present additional qualitative results and comparisons for various image datasets.
See figure captions for explanations.
12
(a) Real
(b) DCGAN [23]
(c) AGE (distribution alignment)
(d) AGE (full)
Figure 5: We compare CIFAR-10 samples from DCGAN [23] (b) to the samples generated using
our ablated model trained without reconstruction terms in (c) using distribution alignment only. The
model, trained with the reconstruction terms is still able to produce diverse samples (d), but also
allows inference (Figure 6).
(a) [Real, AGE network]
(b) [Real, ALI [6]]
Figure 6: Comparison in reconstruction quality to ALI [6] for the CIFAR-10 dataset. For both figures
real examples are shown in odd columns and their reconstructions are shown in even columns. The
real examples come from test set and were never observed by the model during training.
13
(a) Real images
(b) AGE samples
Figure 7: Real examples (a) and samples (b) from our model trained on CelebA dataset.
Figure 8: Latent space interpolation between two images from CelebA dataset. The original images
are presented on the two sides.
Figure 9: In all experiments except this one, we do not use data space reconstruction loss in the
objective of the generator. These figure demonstrates the degradation occurring when this term is
added. Odd columns correspond to real images and even to reconstructions.
14
Figure 10: More color transfer results: in each triplet the color scheme is transferred from the first
image onto the second image using the conditional AGE model.
15
| 1cs.CV
|
Popular Conjectures as a Barrier for
Dynamic Planar Graph Algorithms
Amir Abboud˚
arXiv:1605.03797v1 [cs.DS] 12 May 2016
1
1
and Søren Dahlgaard:2
Stanford University, [email protected]
University of Copenhagen, [email protected]
2
Abstract
The dynamic shortest paths problem on planar graphs asks us to preprocess a planar graph
G such that we may support insertions and deletions of edges in G as well as distance queries
between any two nodes u, v subject to the constraint that the graph remains planar at all times.
This problem has been extensively studied in both the theory and experimental communities over
the past decades and gets solved millions of times every day by companies like Google, Microsoft,
and Uber. The best known algorithm performs queries and updates in Õpn2{3 q time, based
on ideas of a seminal paper by Fakcharoenphol and Rao [FOCS’01]. A p1 ` εq-approximation
?
algorithm of Abraham et al. [STOC’12] performs updates and queries in Õp nq time. An
algorithm with a more practical Oppoly log nq runtime would be a major breakthrough. However,
such runtimes are only known for a p1 ` εq-approximation in a model where only restricted weight
updates are allowed due to Abraham et al. [SODA’16], or for easier problems like connectivity.
In this paper, we follow a recent and very active line of work on showing lower bounds
for polynomial time problems based on popular conjectures, obtaining the first such results
for natural problems in planar graphs. Such results were previously out of reach due to the
highly non-planar nature of known reductions and the impossibility of “planarizing gadgets”.
We introduce a new framework which is inspired by techniques from the literatures on distance
labelling schemes and on parameterized complexity.
Using our framework, we show that no algorithm for dynamic shortest paths or maximum
weight bipartite matching in planar graphs can support both updates and queries in amortized
1
Opn 2 ´ε q time, for ε ą 0, unless the classical all-pairs-shortest-paths problem can be solved in
truly subcubic time, which is widely believed to be impossible. We extend these results to obtain
strong lower bounds for other related problems as well as for possible trade-offs between query
and update time. Interestingly, our lower bounds hold even in very restrictive models where only
weight updates are allowed.
˚
Part of the work was performed while visiting the Simons Institute for the Theory of Computing, Berkeley, CA.
Research partially supported by Virginia Vassilevska Williams’s NSF Grants CCF-1417238 and CCF-1514339, and
BSF Grant BSF:2012338.
:
Part of this work was done while the author was visiting Stanford University. Research partly supported by
Mikkel Thorup’s Advanced Grant DFF-0602-02499B from the Danish Council for Independent Research under the
Sapere Aude research career programme.
1
Introduction
The dynamic shortest paths problem on planar graphs is to preprocess a planar graph G, e.g. the
national road network, so that we are able to efficiently support the following two operations:
• At any point, we might insert or remove an edge pu, vq in G, e.g. in case a road gets congested
due to an accident. Such updates are subjected to the constraint that the planarity of the
graph is not violated. We may also consider another natural variant in which we are only
allowed to update the weights of existing edges.
• We want to be able to quickly answer queries that ask for the length of the shortest path
between two given nodes u and v, in the most current graph G.
This is a problem that gets solved millions of times every day by companies like Google, Microsoft,
and Uber on graphs such as road networks with many millions of nodes. It is thus a very important
question in both theory and practice whether there exists data structures that can perform updates
and (especially) queries on graphs with n nodes in polylogarithmic or even nop1q time.
Shortest paths problems on planar graphs provide an ideal combination of mathematical simplicity
and elegance with faithful modeling of realistic applications of major industrial interest. The literature
on the topic is too massive for us to survey in this paper: the current draft of the book “Optimization
Problems in Planar Graphs” by Klein and Mozes [42] dedicates four chapters to the algorithmic
techniques for shortest paths by the theory community. While near-optimal algorithms are known
for most variants of shortest paths on static planar graphs, the dynamic setting has proven much
more challenging.
Since an s, t-shortest path in a planar graph can be found in near-linear time (linear time for
non-negative weights) [39, 28], there is a naïve algorithm for the dynamic problem that spends Õpnq
time on queries. After progress on other related problems on dynamic planar graphs [29, 30, 24,
31, 61, 44, 39], the first sublinear bound was obtained in the seminal paper of Fakcharoenphol and
Rao [28], which introduced new techniques that led to major results for other problems like Max
Flow (even on static graphs) [17, 47]. The amortized time per operation was Opn2{3 log7{3 nq and
Opn4{5 log13{5 nq if negative edges are allowed, and follow up works of Klein [43], Italiano et al. [40],
and Kaplan et al. [41] reduced the runtime to Opn2{3 log5{3 nq (even allowing negative weights), and
5{3
n
most recently, Gawrychowski and Karczmarz [33] reduced it further to Opn2{3 loglog
q. In fact
4{3
log n
?
these algorithms give a trade-off on the update and query time of Õpn{ rq and Õprq, for all r. The
problem has also been extensively studied from an engineering viewpoint on real-world transportation
networks (see [21] for a survey). State of the art algorithms [13, 20, 22, 34, 60] are able to exploit
further structure of road networks (beyond planarity) and process updates to networks with tens of
millions of nodes in milliseconds.
In a recent SODA’16 paper, Abraham et al. [6] study worst case bounds under a restricted but
realistic model of dynamic updates in which a base graph G is given and one is allowed to perform
only weight updates subject to the following constraint: For any updated graph G1 it must hold that
dG pu, vq ď dG1 pu, vq ď M ¨ dG pu, vq for all u, v and some parameter M . (Note that this will hold if,
for example, the weight of each edge only changes to within a factor of M .) In this model, the authors
obtain a p1 ` εq-approximation algorithm that maintains updates in Oppoly log n ¨ M 4 {ε3 q time.
?
Without this restriction, the best known p1 ` εq-approximation algorithms use Õp nq updates [44, 7].
Thus, when M is small, this model allows for a major improvement over the above results which
1
require polynomial time updates. But is it enough to allow for exact algorithms with subpolynomial
updates? Such a result would explain the impressive experimental performance of state of the art
algorithms.
On the negative side, Eppstein showed that Ωplog n{ log log nq time is required in the cell probe
model [27] for planar connectivity (and therefore also shortest path). However, an unconditional
logωp1q n lower bound is far beyond the scope of current techniques (see [18]). In recent years,
much stronger lower bounds were obtained for dynamic problems under certain popular conjectures
[57, 54, 3, 46, 38, 5, 19]. For example, Roditty and Zwick [57] proved an n2´op1q lower bound for
dynamic single source shortest paths in general graphs under the following conjecture.
Conjecture 1 (APSP Conjecture). There exists no algorithm for solving the all pairs shortest paths
(APSP) problem in general weighted (static) graphs in time Opn3´ε q for any ε ą 0.
However, the reductions used in these results produce graphs that are fundamentally non-planar,
such as dense graphs on three layers, and popular approaches for making them planar, e.g. by
replacing each edge crossing with a small “planarizing gadget”, are provably impossible (this was
recently shown for matching [36] and is easier to show for problems like reachability and shortest
paths). Due to this and other challenges no (conditional) polynomial lower bounds were known for
any natural problem on (static or dynamic) planar graphs.
On a more general note, an important direction for future research on the fine-grained complexity
of polynomial time problems (a.k.a. Hardness in P) is to understand the complexity of fundamental
problems on restricted but realistic classes of inputs. A Recent result along these lines is the
observation that the n2´op1q lower bound for computing the diameter of a sparse graph [56] holds
even when the treewidth of the graph is Oplog nq [4]. In this paper, we take a substantial step in this
direction, proving the first strong (conditional) lower bounds for natural problems on planar graphs.
1.1
Our Results
We present the first conditional lower bounds for natural problems on planar graphs using a new
framework based on several ideas for conditional lower bounds on dynamic graphs combined with
ideas from parameterized complexity [50, 51] and labeling schemes [32]. We believe that this
framework is of general interest and might lead to more interesting results for planar graphs. Our
framework shows an interesting connection between dynamic problems and distance labeling and also
slightly improves the result of [32] providing a tight lower bound for distance labeling in weighted
planar graphs (this is discussed in Section 1.2).
Our first result is a conditional polynomial lower bound for dynamic shortest paths on planar
graphs. Like several recent results [64, 2, 1, 5, 58, 19], our lower bound is based on the APSP
conjecture. Perhaps the best argument for this conjecture is the fact that it has endured decades
of extensive algorithmic attacks. Moreover, due to the known subcubic equivalences [64, 1, 58], the
conjecture is false if and only if several other fundamental graph and matrix problems can be solved
substantially faster.
Theorem 1. No algorithm can solve the dynamic APSP problem in planar graphs on N nodes with
amortized query time qpN q and update time upN q such that qpN q ¨ upN q “ OpN 1´ε q for any ε ą 0
unless Conjecture 1 is false. This holds even if we only allow weight updates to G.
Thus, under the APSP conjecture, there is no hope for a very efficient dynamic shortest paths
algorithm on planar graphs with provable guarantees. We show that an algorithm achieving Opn1{2´ε q
2
time for both updates and queries is unlikely, implying that the current upper bounds achieving
Õpn2{3 q time are not too far from being conditionally optimal. Furthermore, our result implies that
any algorithm with subpolynomial query time must have linear update time (and the other way
around). Thus, the naïve algorithm of simply computing the entire shortest path every time a query
is made is (conditionally) optimal if we want nop1q update time.
An important property of Theorem 1 is that our reduction does not even violate planarity with
respect to a fixed embedding. Thus, we give lower bounds even for plane graph problems, which in
many cases allow for improved upper bounds over flexible planar graphs (e.g. for reachability[23, 11]).
Moreover, our graphs are grid graphs which are subgraphs of the infinite grid, a special and highly
structured subclass of planar graphs. Finally, as stated in Theorem 1 our lower bound holds even
for the edge weight update model of Abraham et al. [6], where each edge only ever changes its
weight to within a factor of M ą 1. While they obtain fast polylogpnq time p1 ` εq-approximation
algorithm in this model, we show that an exact answer with the same query time likely requires
linear update time and that an algorithm with Opn1{2´ε q runtime for both is highly unlikely. Thus,
further theoretical restrictions need to be added in order to explain the impressive performance on
real road networks.
We also extend Theorem 1 to the case in which we only need to maintain one s, t distance (the
s, t-shortest path problem). While this problem is equivalent to the APSP version in general (as we
may connect s and t to any two nodes u, v we wish to know the distance between) this may violate
planarity and especially a fixed embedding. We show that this problem exhibits similar trade-offs
under Conjecture 1 even if we are only allowed to update weights. Finally, we note that in the case
of directed planar graphs allowing negative edge weights our techniques can be extended to show
the same hardness result for any approximation under Conjecture 1.
Next, we seek a lower bound for the unweighted version of the problem, which arguably, is of
more fundamental interest. Typically, a conditional lower bound under the APSP conjecture for a
weighted problem can be modified into a lower bound for its unweighted version under the Boolean
Matrix Multiplication (BMM) Conjecture [25, 57, 64, 3, 1]. While for combinatorial algorithms the
complexity of BMM is conjectured to be cubic, it is known that using algebraic techniques there is
an Opnω q algorithm, where ω ă 2.373 [63, 49]. When reducing to dynamic problems, however, lower
bounds under BMM are often under a certain online version of BMM for which, Henzinger et al.
[38] conjecture that there is no truly subcubic algorithms, even using algebraic techniques. This
Online Matrix Vector Multiplication (OMv) Conjecture is stated formally in Section 5.
The OMv conjecture implies strong lower bounds for many dynamic problems on general graphs
[38], via extremely simple reductions [3, 38]. Our next result is a significantly more involved reduction
from OMv to dynamic shortest paths on planar graphs, giving unweighted versions of the theorems
above. The lower bounds are slightly weaker but they still rule out algorithms with subpolynomial
update and query times, even in grid graphs. We remark that all lower bounds under the APSP
conjecture in this paper, such as Theorem 1, also hold under OMv.
Theorem 2. No algorithm can solve the dynamic APSP problem in unit weight planar graphs on N
nodes with amortized query time qpN q and update time upN q such that maxpqpN q2 ¨ upN q, qpN q ¨
upN q2 q “ OpN 1´ε q for any ε ą 0 unless the OMv conjecture of [38] is false. This holds even if we
only allow weight updates.
1
For instance, Theorem 2 shows that no algorithm is likely to have Opn 3 ´ε q amortized time for
1
both queries and updates. It also shows that if we want to have nop1q for one we likely need n 2 ´op1q
3
time for the other.
Combined with previous results, our theorems reveal a mysterious phenomenon: there are two
contradicting separations between planar graphs and small treewidth graphs, in terms of the time
complexity of dynamic problems related to shortest paths (under popular conjectures). To illustrate
these separations, consider the dynamic s, t-shortest path problem and the dynamic approximate
diameter problem. For s, t-shortest path, planar graphs are much harder, they require n1{3´op1q
update or query time by Theorem 2 (under OMv), while on small (polylog) treewidth graphs there
is an algorithm achieving polylog updates and queries [6]. On the other hand, for approximate
diameter, planar graphs are provably easier under the Strong Exponential Time Hypothesis (SETH).
A naive algorithm that runs the known Õpnq time static algorithm for p1 ` εq approximate diameter
on planar graphs after each update [62], shatters an n2´op1q SETH-based lower bound for a p4{3 ´ δq
approximation for diameter on graphs with treewidth Oplog nq [3]1 .
We demonstrate the potential of our framework to yield further strong lower bounds for important
problems in planar graphs by proving such a result for another well-studied problem in the graph
theory literature, namely Maximum Weight Matching.
Maintaining a maximum matching in general dynamic graphs is a difficult task: the best known
algorithm by Sankowski [59] has an Opn1.495 q amortized update time, and it is better than the
simple Opmq algorithm (that looks for an augmenting path after every update) only in dense
graphs. Recent results show barriers for much faster algorithms via conjectures like OMv and 3-SUM
[3, 46, 38, 19]. To our knowledge, this Opmq update time is the best known for planar graphs and no
lower bound is known. Meanwhile, there has been tremendous progress on approximation algorithms
[53, 12, 52, 14, 16, 15, 37, 45, 55, 10], both on general and planar graphs, as well as for the natural
Maximum Weight Matching (see the references in [26] for the history of this variant). Planar graphs
have proven easier to work with in this context: the state of the art deterministic algorithm for
?
maintaining a p1 ` εq-maximum matching in general graphs has Op mq update time [35], while in
planar graphs the bound is Op1q [55].
We show a strong polynomial lower bound for Max Weight Matching on planar graphs, that
holds even for bipartite graphs with a fixed embedding into the plane and even in grid graphs. The
lower bound is similar to Theorem 1 and shows a trade-off between query and update time.
Theorem 3. No algorithm can solve the dynamic maximum weight matching problem in bipartite planar graphs on N nodes with amortized update time upN q and query time qpN q such
1
that maxpqpN q, upN qq “ OpN 2 ´ε q for any ε ą 0 unless Conjecture 1 is false. Furthermore, if
qpN q ě upN q the algorithm cannot have qpN q ¨ upN q “ OpN 1´ε q. This holds even if the planar
embedding of G never changes.
Finally, we use our framework to show lower bounds for various other problems, like dynamic
girth and diameter. We also argue that our bounds can be turned into worst-case bounds for
incremental and decremental versions of the same problems.
1.2
Techniques and relations to distance labeling
To prove the results mentioned above we introduce a new framework for reductions to optimization
problems on planar graphs. As mentioned we combine ideas from previous lower bound proofs
1
This lower bound follows from observing that the reduction from CNF-SAT to dynamic diameter [3] produces
graphs with logarithmic treewidth. For more details on an analogous observation w.r.t. the lower bound for diameter
in static graphs, see [4].
4
for dynamic graph problems with an approach inspired by the framework of Marx for hardness of
parameterized geometric problem (via the Grid Tiling problem) [50, 51] and a graph construction
from the research on labelling schemes by Gavoille et al. [32].
Gavoille et al. [32] used a family of grid-like graphs to prove an (unconditional) lower bound of
?
?
Ωp nq on the label size of distance labeling in weighted planar graphs along with a Op n log nq
upper bound. (A full discussion of distance labeling schemes is outside the scope of this paper. For
details on this we refer to [32, 9, 8]). In this paper we generalize their family of graphs to a family of
grid graphs capable of representing general matrices with weights in rpolypnqs via shortest paths
?
distances. Using our construction with the framework of [32], we obtain a tight Ωp n log nq lower
bound on the size of distance labeling in weighted planar graphs (and even grid graphs).
Our main approach works by reducing from the pmin, `q-Matrix-Multiplication problem which is
known to be equivalent to APSP (see [64]): Given two n ˆ n matrices A, B with entries in rpolypnqs,
compute a matrix C “ A ‘ B such that Cri, js “ minkPrns Ari, ks ` Brk, js. By concatenating grid
graphs from the family described above we are able to represent one of the matrices in the product
and we can then simulate the multiplication process via updates and shortest paths queries.
In a certain intuitive sense, our connection between dynamic algorithms and labeling
schemes is
?
the reverse direction of the one shown by Abraham et al. [7] to obtain their Õp N q update time
p1 ` εq-approximation algorithm for dynamic APSP. Their algorithm utilizes a clever upper bound
for the so-called forbidden set distance labeling problem, while our lower bound constructions have a
clever lower bound for labeling schemes embedded in them.
2
A grid construction
In order to reduce to problems on planar graphs we will need a planar construction, which is able
to capture the complications of problems like OMv and APSP. To do this we will employ a grid
construction based on the one used in [32] to prove lower bounds on distance labeling for planar
graphs. Our construction takes a matrix as input and produces a grid graph representing that
matrix. We first present a boolean version similar to the one from [32] and then modify it to obtain
a version taking matrices with integer entries as input. This modified matrix also immediately leads
?
to a tight Ωp n log nq lower bound for distance labeling in planar graphs with weights in rpolypnqs
when combined with the framework of [32].
Definition 1. Let M be a boolean R ˆ C matrix. We will call the following construction the grid
embedding of M :
Let GM be a rectangular grid graph with R rows and C columns. Denote the node at intersection
pi, jq by ui,j (u1,1 is top-left and uR,C is bottom-right). Add C nodes a1 , . . . , aC and edges pu1,j , aj q
above GM . Similarly add the nodes b1 , . . . , bR and edges pui,C , bi q to the right of GM . Now subdivide
each vertical edge adding the node vi,j above ui,j , and subdivide each horizontal edge adding the
node wi,j to the right of ui,j . Finally, for each entry of M such that Mi,j “ 1 add the node xi,j and
edges pvi,j , xi,j q and pwi,j , xi,j q to the graph.
The weights of GM are as follows: Each edge pui,j , vi`1,j q and paj , v1,j q has weight 2j ´ 1. Each
edge pwi,j , ui`1,j q and pwi,C , bi q has weight 2R ´ 2. The edge pui,j , wi,j q has weight 2. All remaining
edges have weight 1.
We will call the two-edge path vi,j Ñ xi,j Ñ wi,j a shortcut from vi,j to wi,j as it has length 1 less
than the path vi,j Ñ ui,j Ñ wi,j . Clearly, the grid embedding of a R ˆ C matrix has OpRCq nodes.
5
It is also easy to see that such a grid embedding is a subgraph of a 2R ` 1 ˆ 2C ` 1 rectangular grid.
The construction of Definition 1 for a 3 ˆ 3 matrix can be seen in Figure 1.
110
011
010
a1
a2
1
3
1
1
a3
5
1
2
1
4
3
1
3
v3,2
5
x3,2
5
u3,2 w3,2
b1
b2
b3
Figure 1: Illustration of the construction of Definition 1. The shortest path from a2 to b2 is
highlighted in red. Most edge weights are omitted for clarity.
Proposition 1. Let M be a boolean R ˆ C matrix and let GM be its grid embedding as defined in
Definition 1. Then for any 1 ď i ď R, 1 ď j ď C and i ă k ď R the shortest path distance from ui,j
to bk is exactly
pk ´ iq ¨ 2j ` 2R ¨ pC ´ j ` 1q
if Mk,j “ 0 and
pk ´ iq ¨ 2j ` 2R ¨ pC ´ j ` 1q ´ 1
otherwise.
Proof. Consider any shortest path from any ui,j to bk . Such a path must always go either “right” or
“down” (if i “ k the path must always go right). Essentially for every step to the left we pay at least
4R ´ 1 but can at most save 2R: paying 2R going left and 2R going right, possibly saving 1 with a
shortcut, and saving 2 for each vertical edge.
Now we will show the claim by induction on the sum i ` j. Clearly, for uR,C to bR the distance
is exactly 2R. Now consider ui,j and assume k ą i as the case of k “ i is trivial. There are three
cases to consider:
1. The path from ui,j goes through wi,j and then ui,j`1 . By the induction hypothesis this path
has length at least
2R ` pk ´ iq ¨ 2pj ` 1q ` 2R ¨ pC ´ jq ´ 1 ě 2R ¨ pC ´ j ` 1q ` pk ´ iq ¨ 2j .
2. The path from ui,j goes through vi`1,j and then xi`1,j , wi`1,j , and ui`1,j`1 . This path is only
available if Mi`1,j “ 1. If k “ i ` 1, this distance is exactly
2j ´ 1 ` 2R ¨ pC ´ j ` 1q .
6
Otherwise, by the induction hypothesis, it is at least
pk ´ i ´ 1q ¨ 2pj ` 1q ` 2R ¨ pC ´ jq ´ 1 ` 2j ` 2R ´ 1 ě pk ´ iq ¨ 2j ` 2R ¨ pC ´ j ` 1q
for k ą i ` 1
3. The path from ui,j goes through vi`1,j and then ui`1,j . By the induction hypothesis, if
Mk,j “ 1, the length of this path is
pk ´ i ´ 1q ¨ 2j ` 2R ¨ pC ´ j ` 1q ´ 1 ` 2j “ pk ´ iq ¨ 2j ` 2R ¨ pC ´ j ` 1q ´ 1
and otherwise it is
pk ´ i ´ 1q ¨ 2j ` 2R ¨ pC ´ j ` 1q ` 2j .
It is easy to verify that taking the path down and right as illustrated in Figure 1 gives exactly the
distances in the proposition, finishing the proof.
The following useful property of our grid construction follows.
Corollary 1. Let M and GM be as in Proposition 1. Then for any 1 ď k ď R, 1 ď j ď C, the
distance between aj and bk in GM is exactly determined by whether Mk,j “ 1. In this case the
distance is 2R ¨ pC ´ j ` 1q ` 2jk ´ 1 and it is 2R ¨ pC ´ j ` 1q ` 2jk otherwise.
The following generalization for matrices with integer weights will be useful when reducing from
APSP.
Definition 2. Let M be a R ˆ C matrix with integer weights in t0, . . . , Xu. We will call the
following construction the grid embedding of M .
Let GM be the grid embedding from Definition 1 for the all ones matrix of size R ˆ C and
multiply the weight of each edge by X 2 . Furthermore, for each edge pvi,j , xi,j q increase its weight by
Mi,j .
Corollary 2. Let M be a R ˆ C matrix with integer weights in t0, . . . , Xu and let GM be its grid
embedding. Then for any 1 ď k ď R, 1 ď j ď C, the distance between aj and bk in GM is exactly
X 2 ¨ p2R ¨ pC ´ j ` 1q ` 2jk ´ 1q ` Mk,j
Corollary 2 follows from Corollary 1 by observing that any path from aj to bk not using the
shortcut at intersection pk, jq has distance at least X 2 ¨ p2R ¨ pC ´ j ` 1q ` 2jkq and since Mk,j ă X 2
this distance is longer than using the shortcut. We remark that it would have been sufficient to
multiply the weights by pX ` 1q instead of X 2 , but we do so to simplify a later argument.
3
Hardness of dynamic APSP in planar graphs
We will first show the following, simpler theorem and then generalize it to show trade-offs between
query and update time.
Theorem 4. No algorithm can solve the dynamic APSP problem in planar graphs on N nodes with
1
amortized update and query time OpN 2 ´ε q for any ε ą 0 unless Conjecture 1 is false. This holds
even if only weight updates are allowed.
7
The main idea in proving Theorem 4 is to reduce from the APSP problem by first reducing to
pmin, `q-Matrix-Mult and use the grid construction from Section 2 to represent the matrices to be
multiplied. We then perform several shortest paths queries to simulate the multiplication process.
Below, we first present a naïve and faulty approach explaining the main ideas of the reduction. We
then show how to mend this approach giving the desired result.
Attempt 1. Consider the following algorithm for solving an instance, A ‘ B of the pmin, `q-MatrixMult problem, where A and B are n ˆ n matrices. We may assume that A and B have integer
weights in t0, . . . , Xu for some X “ polypnq.
We let the initial graph of the problem be the grid embedding GB of B according to Definition 2
along with a special vertex t. Also add the edges pbk , tq for each 1 ď k ď n. Now we wish to
construct C “ A ‘ B one row at a time. Such a row is a pmin, `q-product of a row in A and the
entire matrix B. Thus, for each row, i, of A we have a phase as follows:
1. For each 1 ď k ď n update the weight of the edge pbk , tq to be Ai,k .
2. For each 1 ď j ď n query the distance between aj and t.
The idea of each phase is that the distance between aj and t should correspond to the value of
Ci,j “ mink Ai,k ` Bk,j . Observe, that the distance from aj to t using the edge pbk , tq is exactly
X 2 ¨ p2n ¨ pn ´ j ` 1q ` 2jk ´ 1q ` Bk,j ` Ai,k
by Corollary 2. The dominant term in this expression increases with k and thus no matter what
Bk,j and Ai,k are (for k ą 1), the shortest path from aj to t will simply pick k “ 1 minimizing the
above expression. If we instead set the weight of each edge pbk , tq to X 2 ¨ 2jpn ´ kq ` Ai,k we get the
distance of using this edge to be
X 2 ¨ p2n ¨ pn ´ j ` 1q ` 2jk ´ 1q ` Bk,j ` Ai,k ` X 2 ¨ 2jpn ´ kq “ X 2 ¨ p2npn ` 1q ´ 1q ` Bk,j ` Ai,k .
It follows that the shortest path from aj to t is free to pick any k while only affecting the Bk,j ` Ai,k
term, which means that the shortest distance will be achieved by picking the k minimizing this
term, which would give us exactly X 2 ¨ p2npn ` 1q ´ 1q ` Ci,j . This approach therefore allows us
to correctly calculate C “ A ‘ B. However, the weight of the edge pbk , tq now depends on which
aj we are querying implying that we have to update this weight for each aj leading to a total of
Opn3 q updates. By using this approach we are thus not able to make any statement about the time
required for updates. We may try to assign edges and weights differently, but such approaches run
into similar issues.
Observe that the graph created has N “ Opn2 q nodes. Thus, if we were able to perform only
Opn2 q total queries and updates the result of Theorem 4 would follow.
♦
In order to circumvent this dependence on j when assigning weights to the edges pbk , tq we instead
replace t by another grid whose purpose is to “normalize” the distance for each aj . By doing this we
can connect the grids with edges whose weight is independent of j. This step deviates significantly
from the construction of [32] and is inspired by the grid tiling framework of Marx [50, 51].
Proof of Theorem 4. We follow the same approach as in Attempt 1, but with a few changes. Define
the initial graph G as follows: Let GB be as before and let G1B be the grid embedding of B mirrored
along the vertical axis with all shortcuts removed. Now for each 1 ď k ď n add the edge pbk , b1k q and
define G to be this graph.
Now we perform a phase for each row i of A as follows:
8
1. For each 1 ď k ď n set the weight of the edge pbk , b1k q to be X 2 ¨ p2pn ` 1qpn ´ kqq ` Ai,k .
2. For each j query the distance between aj and a1n´j`1 .
An example of this construction for n “ 3 can be seen in Figure 2.
a2
a1
a3
a'3
X2
a'2
a'1
5X2
X2·16+Ai,1
X2+B2,1
b1
X2
b'1
X2·8+Ai,2
b2
b'2
Ai,3
b3
b'3
Figure 2: Example of a phase in the graph G in the reduction of Theorem 4. The highlighted path
illustrates a shortest path between a1 and a13 as an example of a query. Most edge weights have
been omitted for clarity.
From the query between nodes aj and a1n´j`1 above during phase i we can determine the entry
Ci,j of the output matrix. To see this, consider the distance from aj to a1n´j`1 at the time of query.
This path has to go via some edge pbk , b1k q. From Corollary 2 we know that this distance is exactly
dG paj , a1n´j`1 q “ dpaj , bk q ` wpbk , b1k q ` dpb1k , a1n´j`1 q
“ X 2 ¨ p2n ¨ pn ´ j ` 1q ` 2jk ´ 1q ` Bk,j ` X 2 ¨ p2pn ` 1qpn ´ kqq ` Ai,k
` X 2 ¨ p2n ¨ j ` 2pn ´ j ` 1qkq
“ X 2 ¨ 2npn ` 1q ` X 2 ¨ 2kpn ` 1q ` X 2 ¨ 2pn ` 1qpn ´ kq ` Bk,j ` Ai,k ´ X 2
“ X 2 ¨ 4npn ` 1q ´ X 2 ` Bk,j ` Ai,k .
The crucial property that our construction achieves is that the dominant term of this expression is
independent of k. Thus, the shortest path will choose to go through the edge pbk , b1k q that minimizes
Bk,j ` Ai,k , implicitly giving us Ci,j . Subtracting X 2 ¨ p4npn ` 1q ´ 1q from the queried distance
gives exactly the value of Ci,j and the algorithm therefore correctly computes C.
Following the analysis from Attempt 1 we have that any algorithm with an amortized running
1
time of OpN 2 ´ε q for both updates and queries contradicts Conjecture 1.
3.1
Trade-offs
Theorem 4 above shows that no algorithm can perform both updates and queries in amortized time
1
OpN 2 ´ε q unless Conjecture 1 is false. We will now show how to generalize these ideas to show
Theorem 1.
9
Proof of Theorem 1. The proof follows the same structure as the proof for Theorem 4, but instead
of reducing from pmin, `q-Matrix-Mult on n ˆ n matrices we reduce from an unbalanced version.
Let A and B be n ˆ nβ and nβ ˆ nα matrices respectively for some 0 ă α, β ď 1. We define the
initial graph G from B in the same manner as in Theorem 4. We then have a phase for each row i
of A as follows:
1. For each 1 ď k ď nβ set the weight of the edge pbk , b1k q to be X 2 ¨ p2pnα ` 1qpnβ ´ kqq ` Ai,k .
2. For each 1 ď j ď nα query the distance between aj and a1nα ´j`1 .
The entry Ci,j is exactly the distance dG paj , a1nα ´j`1 q from the ith phase minus X 2 ¨p4nβ pnα `1q´1q.
The correctness of the above reduction follows directly from the proof of Theorem 4 as well as
Corollary 2.
Now observe that the graph G from the above reduction has N “ Θpnα`β q nodes and we perform
a total of Opn1`α q queries and Opn1`β q updates2 – that is, at most Opnq updates per row and Opnq
queries per column. Any algorithm solving this problem must use total time n1`α`β´op1q time unless
Conjecture 1 is false. It follows that either updates must take nα´op1q amortized time or queries
must take nβ´op1q amortized time.
Assume now that an algorithm exists such that queries take OpN γ q amortized time for any
0 ă γ ă 1. We wish to show that this algorithm cannot perform updates in amortized time
OpN 1´γ´ε q for any ε ą 0. Pick β “ γ ` ε{2 and set α “ 1 ´ β. We now use the above reduction to
create a dynamic graph G with N “ Opnα`β q “ Opnq nodes. Since queries do not take nβ´op1q time
it follows from the above discussion that updates must take nα´op1q “ n1´γ´ε{2´op1q time. Since this
is polynomially greater than OpN 1´γ´ε q the claim follows.
4
Hardness of dynamic maximum weight matching in bipartite planar graphs
In this section we will demonstrate the generality of our reduction framework by showing Theorem 3.
Proof of Theorem 3. We start by showing how to reduce from pmin, `q-Matrix-Mult to minimum
weight perfect matching, where the weight of such a matching corresponds to the shortest path
distance between aj and a1n´j`1 similar to the proof of Theorem 1. We then describe how to use
this reduction further to get a problem instance for maximum weight matching.
Let A, B be an instance to the pmin, `q-Matrix-Mult problem of sizes n ˆ nβ and nβ ˆ nα
respectively. Consider the grid embedding GB of B. We first replace each node of GB by two nodes
connected by an edge of weight 0. For aj , ui,j , xi,j , and vi,j denote the corresponding nodes with
superscript d and u (for “down” and “up”). For bi and wi,j denote the corresponding nodes with
superscript l and r (for “left” and “right”). Now, for each original edge in GB we replace it as follows
keeping its weight:
d q
• pui,j , vi,j q Ñ puui,j , vi,j
r
• pui,j , wi,j´1 q Ñ puui,j , wi,j´1
q
l q
• pui,j , wi,j q Ñ pudi,j , wi,j
d , xu q
• pvi,j , xi,j q Ñ pvi,j
i,j
u
• pui,j , vi`1,j q Ñ pudi,j , vi`1,j
q
l q
• pxi,j , wi,j q Ñ pxdi,j , wi,j
2
We also perform Opnα`β q updates to create the initial graph (depending on the model), however we will choose α
and β such that this term is dominated.
10
u q
• paj , v0,j q Ñ padj , v0,j
r , bl q
• pwi,C , bi q Ñ pwi,C
i
This construction is illustrated in Figure 3. We call this modified grid structure ḠB . Observe
that there are no edges between “up” and “left” vertices or between “down” and “right”. It follows
that the graph is bipartite and that these two sets of nodes make up the two partitions.
vu2,2
vd2,2
xu2,2
xd2,2
uu2,2
ud2,2
l
r
w2,2
w2,2
Figure 3: Grid construction for minimum weight perfect matching. Thick edges correspond to
original edges and have the same weight as in GB . Thin edges have weight 0.
We now replace the grids GB and G1B by ḠB and Ḡ1B in the initial graph G from the proof of
Theorem 1. The edges pbk , b1k q are replaced by pbrk , b1lk q. We will use the following observation.
Proposition 2. The graph resulting from joining two grids ḠB in the way of Figure 2 has a unique
perfect matching.
Proof. It is easy to see that simply matching all weight 0 edges gives a perfect matching. Thus we
need to show that this is the only perfect matching. We will show the claim by a simple “peeling”
argument.
Observe that au1 only has one incident edge, so the edge pau1 , ad1 q must be in any perfect matching
u only has one adjacent edge, so
and we may “peel” away these two nodes. It now follows that v1,1
u
d
pv1,1 , v1,1 q has to be in any perfect matching and we may peel away these nodes. Now uu1,1 only has
u , uu , etc. This peels away the entire first column. Now each
one adjacent edge and so on for v2,1
2,1
l with only one adjacent edge. Peeling
xui,1 has only one adjacent edge matching this leaves each wi,1
these nodes away leaves us with a smaller grid and we may start the argument over with a2 .
By doing this we see that the edge joining bri and b1li cannot be in a perfect matching as pbri , bli q
has to be. Thus we can repeat the same argument on the second grid.
We now add two additional nodes s and t to the initial graph and perform a phase for each row
i of A as follows:
1. For each 1 ď k ď nβ set the weight of the edge pbrk , b1lk q to be X 2 ¨ p2pnα ` 1qpnβ ´ kqq ` Ai,k .
11
2. For each 1 ď j ď nα do the following three steps: 1) add the edges ps, auj q and pt, a1u
nα ´j`1 q, 2)
query the minimum weight perfect matching, 3) delete the two edges.
d
1d
Since the edges ps, auj q and pt, a1u
nα ´j`1 q have to be in any perfect matching this leaves aj and anα ´j`1
unmatched. Any perfect matching now has to “connect” these two nodes by a path of original (weight
ą 0) edges. The weight of a perfect matching in G then corresponds to the length of a shortest
path from aj to a1nα ´j`1 in the graph from the proof of Theorem 1. It follows that we get the same
trade-offs for minimum weight perfect matching as for APSP with the exception that the trade-off
only holds when qpN q ě upN q since we perform Op1q updates for each query.
To show the same result for maximum weight matching we may simply perform the following two
changes: 1) pick a sufficiently large integer y and set the weight of each edge to y minus its weight
2
in the above reduction, and 2) when adding the edges ps, auj q and pt, a1u
nα ´j`1 q assign them weight y
such that any maximum weight matching has to include these two edges and will have weight
y2 `
N ´4
¨ y ´ dG˚ paj , a1nα ´j`1 q ,
2
where G˚ denotes the corresponding graph in the proof of Theorem 1.
5
Unweighted
The proofs of the previous sections rely heavily on the weighted grid from Section 2. We may
generalize the ideas to the unweighted case by instead using the grid of Definition 1 and subdividing
the edges giving us somewhat weaker bounds. This gives us Theorem 2.
The problem we reduce from is the online matrix-vector problem from [38]. We may define this
problem as follows: Let M be a n ˆ n matrix and let v 1 , . . . , v n be n boolean vectors arriving in an
online fashion. The task is to pre-process M such that we can output the product M v i for each i
before seeing v i`1 . It was conjectured
in [38] that this problem takes n3´op1q time, while the best
?
3
Ωp
log
nq
known upper bound is n {2
[48]. Known reductions from [64] show that this conjecture
implies a n1`α`β´op1q bound for the following problem: Let α, β ą 0 be fixed constants and let M
be a boolean nβ ˆ nα matrix (see [38] for the details). After preprocessing M , n boolean vector pairs
β
α
pu1 , v 1 q, . . . , pun , v n q P t0, 1un ˆ t0, 1un arrive one at a time and the task is to compute pui qT M v i
before being presented with the i ` 1th vector pair for every i. We will use this problem called the
OuMv problem to reduce to unit weight dynamic APSP in planar graphs below.
Proof of Theorem 2. Consider the reduction from Theorem 1 using a nβ ˆ nα grid. We will use a
similar approach to solve the OuMv problem below.
Let M be the nβ ˆ nα matrix of the OuMv problem and create GM according to Definition 1
(note that this grid embedding is different from the one used in the proof of Theorem 1). We also add
G1M similarly to the proof of Theorem 1. We then subdivide each edge into a path of the same length.
We also add to G a path of length 2pnα ` 1qpnβ ´ kq connecting bk and b1k for each 1 ď k ď nβ . We
then disconnect bk and b1k from this path.
We perform a phase as follows for each vector pair pui , v i q:
1. For each k such that uik “ 1 connect bk and b1k to their respective path.
2. For each j such that vji “ 1 query the distance from aj to a1nα ´j`1 .
12
3. Remove all the edges added in step 1.
If the answer to any of the queries during the ith phase is 4npn ` 1q ´ 1 the answer to the ith product
is 1 and otherwise the answer is 0. This follows from Corollary 1 in the same way as Theorem 1.
By subdividing the edges we get a graph with N “ Opn2β`α ` n2α`β q nodes. We perform
Opn1`α q queries and Opn1`β q updates. It follows from the OMv conjecture that the entire process
must take n1`β`α´op1q time, thus either updates take nα´op1q time or queries take nβ´op1q time.
We will assume that qpN q ě upN q and note that the other case follows symmetrically. Assume
that some algorithm can perform queries in N γ for some 13 ď γ ă 21 . We wish to show that this
algorithm cannot perform updates in time N 1´2γ´ε for any ε ą 0. To do this, pick β “ γ ` ε{3
and α “ 1 ´ 2β. Note that β ě α (corresponding to qpN q ě upN q). Thus the graph has
N “ Opn2β`α q “ Opnq nodes. It now follows by the above discussion that the algorithm cannot
perform updates faster than nα´op1q “ N 1´2γ´2ε{3`op1q which proves the claim.
Finally, observe that by using an n ˆ n matrix in the above reduction (i.e. α “ β “ 1) we see that
1
at least one of updates and queries have to take n 3 ´op1q amortized time (similar to Theorem 4).
To see that we may do the above reduction while keeping the dynamic graph G as a grid graph,
observe that we may multiply the weight of each edge before subdividing by a large enough constant
and then “zig-zag” the subdivided edges in order to fit the grid structure. This is illustrated in
Figure 4.
a1
a2
a3
b1
b2
b'1
b'2
Figure 4: Illustration of the grid created in the proof of Theorem 2. Dashed edges correspond to
possible shortcuts. Note that the lengths of the edges are not to scale!
6
Dynamic s, t-shortest path and related problems
In Section 3 we showed a lower bound for the trade-off between query and update time for dynamic
APSP in grid graphs conditioned on Conjecture 1. Here we will argue that the proof of Theorem 1
can be extended to show similar lower bounds for dynamic problems, where the algorithm only needs
to maintain a single value such as s, t-shortest path, girth, and diameter. We also note that the
above techniques for proving bounds in unweighted graphs also apply to the theorem below.
Theorem 5. No algorithm can solve the s, t-shortest path, girth (directed), or diameter problems
in planar graphs on N nodes with amortized update time upN q and query time qpN q such that
13
1
maxpqpN q, upN qq “ OpN 2 ´ε q for any ε ą 0 unless Conjecture 1 is false. Furthermore, if qpN q ě
upN q the algorithm cannot have qpN q ¨ upN q “ OpN 1´ε q. This holds even if the planar embedding
of G never changes.
Proof. We note that the proof follows the exact same structure as the proof of Theorem 1 and only
mention the changes needed to be made.
For s, t-shortest path and diameter we add two additional nodes s, t to the initial graph G and
when performing a query of the distance between aj and a1nα ´j`1 we instead insert edges ps, aj q and
pt, a1nα ´j`1 q of sufficiently high weight so that this is the longest distance in the graph and query
the s, t distance.
For girth we direct all horizontal edges of G to the right, all vertical edges of the left grid down
and all vertical edges of the right grid up. When doing a query we add the directed edge pa1nα ´j`1 , aj q
with weight 1, and the length of the shortest cycle then corresponds to the shortest path from aj to
a1nα ´j`1 plus 1.
In the above reductions, the condition qpN q ě upN q comes from the fact that we perform Op1q
updates for every query we make and the argument from Theorem 1 thus breaks down if we try to
argue for slower updates than queries. This makes sense from an upper bound perspective: clearly,
any algorithm with qpN q ď upN q could simply perform a query for every update, store the answer,
and then provide queries in Op1q time.
7
Weight updates
We mentioned in the previous sections that the results hold even if we only allow weight updates
instead of edge insertions/deletions. In [6] they considered this model in which the algorithm is
supplied with an initial graph G and a promise that for any updated graph G1 we have dG pu, vq ď
dG1 pu, vq ď M ¨ dG pu, vq for all u, v P G and some parameter M ą 1. The only operations allowed
are weight updates and queries. We note that all the above results for weighted graphs also hold in
this model.
As a proof sketch, consider the result of Theorem 1: The only edges whose weight changes
are the “in-between” edges pbk , b1k q whose weights are always between X 2 ¨ p2pnα ` 1qpnβ ´ kq and
X 2 ¨ p2pnα ` 1qpnβ ´ kqq ` X. Similarly, for s, t-shortest path and diameter: Assume that the edge
ps, aj q has weight y when added in the reduction of Theorem 5. We may instead initialize the graph
G with an edge ps, aj q of weight y for each 1 ď j ď nα , increase each edge to have weight M ¨ y and
then instead of adding the edge ps, aj q we decrease its weight back to y. We do the same for t and
the nodes a1j . By picking y sufficiently large we may ensure that an edge of weight M ¨ y cannot be
on the shortest path from s to t. Furthermore these changes can still be done while maintaining the
graphs as grids.
8
Worst-case bounds for partially dynamic problems
Our reductions above work in the fully dynamic setting, where edge insertions and deletions (or
weight increments and decrements) are allowed. We now show that, using standard techniques, we
can turn these amortized bounds into worst-case bounds for the same problem in the incremental
and decremental (only insertions/increments or deletions/decrements allowed). We will show the
result for dynamic APSP and note that the method is the same for the other problems.
14
Corollary 3. No algorithm can solve the incremental or decremental APSP problem for planar graphs
on N nodes with worst-case query time qpN q and update time upN q such that qpN q¨upN q “ OpN 1´ε q
for any ε ą 0 unless Conjecture 1 is false.
Proof. We present the argument for the problem when we are given an initial graph G and are only
allowed to increase weights on edges. The proof uses the same rollback technique employed before in
several papers (see e.g. [3]).
First we create the same initial graph, G, as in the proof of Theorem 1. We set the initial weight
of the edges pbk , b1k q to be X 2 ¨ p2pnα ` 1qpnβ ´ kqq. During the phase of each row i of A we keep
track of all memory changes made by the incremental data structure while increasing each edge
to have weight X 2 ¨ p2pnα ` 1qpnβ ´ kqq ` Ai,k . We then perform each distance query and instead
of deleting the incremented edges, we “roll back” the data structure using the memory changes we
kept track of, thus restoring G to its initial state. By doing this we solve the pmin, `q-Matrix-Mult
problem in the exact same way as in Theorem 1. However, we cannot ensure any requirement on
the amortized running time, as the rollback operations may essentially “restore all credit” to the
data structure in the sense of amortized analysis. Thus, the time bounds only apply to worst-case
running times.
Acknowledgements The authors would like to acknowledge Shay Mozes, Oren Weimann and
Virginia Vassilevska Williams for helpful comments and discussions.
References
[1] Amir Abboud, Fabrizio Grandoni, and Virginia Vassilevska Williams. Subcubic equivalences
between graph centrality problems, APSP and diameter. In Proc. 26th ACM/SIAM Symposium
on Discrete Algorithms (SODA), pages 1681–1697, 2015.
[2] Amir Abboud and Kevin Lewi. Exact weight subgraphs and the k-sum conjecture. In Proc.
40th International Colloquium on Automata, Languages and Programming (ICALP), pages 1–12,
2013.
[3] Amir Abboud and Virginia Vassilevska Williams. Popular conjectures imply strong lower bounds
for dynamic problems. In Proc. 55th IEEE Symposium on Foundations of Computer Science
(FOCS), pages 434–443, 2014.
[4] Amir Abboud, Virginia Vassilevska Williams, and Joshua R. Wang. Approximation and fixed
parameter subquadratic algorithms for radius and diameter in sparse graphs. In Proceedings
of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016,
Arlington, VA, USA, January 10-12, 2016, pages 377–391, 2016.
[5] Amir Abboud, Virginia Vassilevska Williams, and Huacheng Yu. Matching triangles and basing
hardness on an extremely popular conjecture. In Proc. 47th ACM Symposium on Theory of
Computing (STOC), pages 41–50, 2015.
[6] Ittai Abraham, Shiri Chechik, Daniel Delling, Andrew V. Goldberg, and Renato F. Werneck.
On dynamic approximate shortest paths for planar graphs with worst-case costs. In Proc. 27th
ACM/SIAM Symposium on Discrete Algorithms (SODA), pages 740–753, 2016.
15
[7] Ittai Abraham, Shiri Chechik, and Cyril Gavoille. Fully dynamic approximate distance oracles
for planar graphs via forbidden-set distance labels. In Proc. 44th ACM Symposium on Theory
of Computing (STOC), pages 1199–1218, 2012.
[8] Stephen Alstrup, Søren Dahlgaard, Mathias Bæk Tejs Knudsen, and Ely Porat. Sublinear
distance labeling for sparse graphs. CoRR, abs/1507.02618, 2015.
[9] Stephen Alstrup, Cyril Gavoille, Esben Bistrup Halvorsen, and Holger Petersen. Simpler, faster
and shorter labels for distances in graphs. In Proc. 27th ACM/SIAM Symposium on Discrete
Algorithms (SODA), pages 338–350, 2016.
[10] Abhash Anand, Surender Baswana, Manoj Gupta, and Sandeep Sen. Maintaining approximate
maximum weighted matching in fully dynamic graphs. In Conference on the Foundations of
Software Technology and Theoretical Computer Science (FSTTCS), pages 257–266, 2012.
[11] Rinat Ben Avraham, Haim Kaplan, and Micha Sharir. A faster algorithm for the discrete fréchet
distance under translation. arXiv preprint arXiv:1501.03724, 2015.
[12] Surender Baswana, Manoj Gupta, and Sandeep Sen. Fully dynamic maximal matching in o
(log n) update time. In Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual
Symposium on, pages 383–392. IEEE, 2011.
[13] Reinhard Bauer. Dynamic speed-up techniques for dijkstra?s algorithm. Master’s thesis, Institut
für Theoretische Informatik-Universität Karlsruhe (TH), page 41, 2006.
[14] Aaron Bernstein and Cliff Stein. Fully dynamic matching in bipartite graphs. In Automata,
Languages, and Programming, pages 167–179. Springer, 2015.
[15] Aaron Bernstein and Cliff Stein. Faster fully dynamic matchings with small approximation
ratios. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete
Algorithms, pages 692–711. SIAM, 2016.
[16] Sayan Bhattacharya, Monika Henzinger, and Giuseppe F Italiano. Deterministic fully dynamic
data structures for vertex cover and matching. In Proceedings of the Twenty-Sixth Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 785–804. SIAM, 2015.
[17] Glencora Borradaile, Philip N. Klein, Shay Mozes, Yahav Nussbaum, and Christian Wulff-Nilsen.
Multiple-source multiple-sink maximum flow in directed planar graphs in near-linear time. In
Proc. 52nd IEEE Symposium on Foundations of Computer Science (FOCS), pages 170–179,
2011.
[18] Raphaël Clifford, Allan Grønlund, and Kasper Green Larsen. New unconditional hardness
results for dynamic and online problems. In Proc. 56th IEEE Symposium on Foundations of
Computer Science (FOCS), pages 1089–1107, 2015.
[19] Søren Dahlgaard. On the hardness of partially dynamic graph problems and connections to
diameter. arXiv preprint arXiv:1602.06705, 2016. To appear at ICALP’16.
[20] Daniel Delling, Andrew V Goldberg, Thomas Pajor, and Renato F Werneck. Customizable
route planning. In Experimental algorithms, pages 376–387. Springer, 2011.
16
[21] Daniel Delling, Peter Sanders, Dominik Schultes, and Dorothea Wagner. Engineering route
planning algorithms. In Algorithmics of large and complex networks, pages 117–139. Springer,
2009.
[22] Daniel Delling and Dorothea Wagner. Landmark-based routing in dynamic graphs. In Experimental algorithms, pages 52–65. Springer, 2007.
[23] Krzysztof Diks and Piotr Sankowski. Dynamic plane transitive closure. In Algorithms–ESA
2007, pages 594–604. Springer, 2007.
[24] Hristo N Djidjev, Grammati E Pantziou, and Christos D Zaroliagis. Computing shortest paths
and distances in planar graphs. In Automata, Languages and Programming, pages 327–338.
Springer, 1991.
[25] Dorit Dor, Shay Halperin, and Uri Zwick. All-pairs almost shortest paths. SIAM Journal on
Computing, 29(5):1740–1759, 2000.
[26] Ran Duan and Seth Pettie. Linear-time approximation for maximum weight matching. Journal
of the ACM (JACM), 61(1):1, 2014.
[27] David Eppstein. Dynamic connectivity in digital images. Information Processing Letters,
62(3):121–126, 1997.
[28] Jittat Fakcharoenphol and Satish Rao. Planar graphs, negative weight edges, shortest paths,
and near linear time. Journal of Computer and System Sciences, 72(5):868–889, 2006. See also
FOCS’01.
[29] Greg N Frederickson. Data structures for on-line updating of minimum spanning trees, with
applications. SIAM Journal on Computing, 14(4):781–798, 1985.
[30] Zvi Galil and Giuseppe F Italiano. Maintaining biconnected components of dynamic planar
graphs. Springer, 1991.
[31] Zvi Galil, Giuseppe F Italiano, and Neil Sarnak. Fully dynamic planarity testing. In Proceedings
of the twenty-fourth annual ACM symposium on Theory of computing, pages 495–506. ACM,
1992.
[32] Cyril Gavoille, David Peleg, Stéphane Pérennes, and Ran Raz. Distance labeling in graphs.
Journal of Algorithms, 53(1):85–112, 2004. See also SODA’01.
[33] Pawel Gawrychowski and Adam Karczmarz. Improved bounds for shortest paths in dense
distance graphs. CoRR, abs/1602.07013, 2016.
[34] Robert Geisberger, Peter Sanders, Dominik Schultes, and Christian Vetter. Exact routing in
large road networks using contraction hierarchies. Transportation Science, 46(3):388–404, 2012.
[35] Madhu Gupta and Rongkun Peng. Fully dynamic (1+ e)-approximate matchings. In Foundations
of Computer Science (FOCS), 2013 IEEE 54th Annual Symposium on, pages 548–557. IEEE,
2013.
17
[36] Rohit Gurjar, Arpita Korwar, Jochen Messner, Simon Straub, and Thomas Thierauf. Planarizing
gadgets for perfect matching do not exist. In Mathematical Foundations of Computer Science
2012, pages 478–490. Springer, 2012.
[37] Meng He, Ganggui Tang, and Norbert Zeh. Orienting dynamic graphs, with applications to
maximal matchings and adjacency queries. In Algorithms and Computation, pages 128–140.
Springer, 2014.
[38] Monika Henzinger, Sebastian Krinninger, Danupon Nanongkai, and Thatchaphol Saranurak.
Unifying and strengthening hardness for dynamic problems via the online matrix-vector multiplication conjecture. In Proc. 47th ACM Symposium on Theory of Computing (STOC), pages
21–30, 2015.
[39] Monika R Henzinger, Philip Klein, Satish Rao, and Sairam Subramanian. Faster shortest-path
algorithms for planar graphs. journal of computer and system sciences, 55(1):3–23, 1997.
[40] Giuseppe F Italiano, Yahav Nussbaum, Piotr Sankowski, and Christian Wulff-Nilsen. Improved
algorithms for min cut and max flow in undirected planar graphs. In Proceedings of the forty-third
annual ACM symposium on Theory of computing, pages 313–322. ACM, 2011.
[41] Haim Kaplan, Shay Mozes, Yahav Nussbaum, and Micha Sharir. Submatrix maximum queries
in monge matrices and monge partial matrices, and their applications. In Proceedings of the
twenty-third annual ACM-SIAM symposium on Discrete Algorithms, pages 338–355. SIAM,
2012.
[42] Philip Klein and Shay Mozes. Optimization algorithms for planar graphs, 2014. Book draft
available at http://planarity.org.
[43] Philip N Klein. Multiple-source shortest paths in planar graphs. In SODA, volume 5, pages
146–155, 2005.
[44] Philip N Klein and Sairam Subramanian. A fully dynamic approximation scheme for shortest
paths in planar graphs. Algorithmica, 22(3):235–249, 1998.
[45] Tsvi Kopelowitz, Robert Krauthgamer, Ely Porat, and Shay Solomon. Orienting fully dynamic
graphs with worst-case time bounds. In Automata, Languages, and Programming, pages 532–543.
Springer, 2014.
[46] Tsvi Kopelowitz, Seth Pettie, and Ely Porat. Higher lower bounds from the 3sum conjecture.
In Proc. 27th ACM/SIAM Symposium on Discrete Algorithms (SODA), pages 1272–1287, 2016.
[47] Jakub Łącki, Yahav Nussbaum, Piotr Sankowski, and Christian Wulff-Nilsen. Single source–all
sinks max flows in planar digraphs. In Foundations of Computer Science (focs), 2012 Ieee 53rd
Annual Symposium on, pages 599–608. IEEE, 2012.
[48] Kasper Green Larsen and Ryan Williams. Faster online matrix-vector multiplication. arXiv
preprint arXiv:1605.01695, 2016.
[49] François Le Gall. Powers of tensors and fast matrix multiplication. In Proceedings of the 39th
international symposium on symbolic and algebraic computation, pages 296–303. ACM, 2014.
18
[50] Dániel Marx. On the optimality of planar and geometric approximation schemes. In Foundations
of Computer Science, 2007. FOCS’07. 48th Annual IEEE Symposium on, pages 338–348. IEEE,
2007.
[51] Dániel Marx. The square root phenomenon in planar graphs. In ICALP (2), page 28, 2013.
[52] Ofer Neiman and Shay Solomon. Simple deterministic algorithms for fully dynamic maximal
matching. ACM Transactions on Algorithms (TALG), 12(1):7, 2015. See also STOC’13.
[53] Krzysztof Onak and Ronitt Rubinfeld. Maintaining a large matching and a small vertex cover.
In Proceedings of the forty-second ACM symposium on Theory of computing, pages 457–464.
ACM, 2010.
[54] Mihai Patrascu. Towards polynomial lower bounds for dynamic problems. In Proc. 42nd ACM
Symposium on Theory of Computing (STOC), pages 603–610, 2010.
[55] David Peleg and Shay Solomon. Dynamic (1+ ε)-approximate matchings: a density-sensitive
approach. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete
Algorithms, pages 712–729. SIAM, 2016.
[56] Liam Roditty and Virginia Vassilevska Williams. Fast approximation algorithms for the diameter
and radius of sparse graphs. In Proc. 45th ACM Symposium on Theory of Computing (STOC),
pages 515–524, 2013.
[57] Liam Roditty and Uri Zwick. On dynamic shortest paths problems. Algorithmica, 61(2):389–401,
2011. See also ESA’04.
[58] Barna Saha. Language edit distance and maximum likelihood parsing of stochastic grammars:
Faster algorithms and connection to fundamental graph problems. In Foundations of Computer
Science (FOCS), 2015 IEEE 56th Annual Symposium on, pages 118–135. IEEE, 2015.
[59] Piotr Sankowski. Faster dynamic matchings and vertex connectivity. In Proc. 18th ACM/SIAM
Symposium on Discrete Algorithms (SODA), pages 118–126, 2007.
[60] Dominik Schultes and Peter Sanders. Dynamic highway-node routing. In Experimental Algorithms, pages 66–79. Springer, 2007.
[61] Sairam Subramanian. A fully dynamic data structure for reachability in planar digraphs. In
Algorithms-ESA’93, pages 372–383. Springer, 1993.
[62] Oren Weimann and Raphael Yuster. Approximating the diameter of planar graphs in near
linear time. ACM Transactions on Algorithms, 12(1):12, 2016. See also ICALP’13.
[63] Virginia Vassilevska Williams. Multiplying matrices faster than coppersmith-winograd. In
Proceedings of the forty-fourth annual ACM symposium on Theory of computing, pages 887–898.
ACM, 2012.
[64] Virginia Vassilevska Williams and Ryan Williams. Subcubic equivalences between path, matrix
and triangle problems. In Foundations of Computer Science (FOCS), 2010 51st Annual IEEE
Symposium on, pages 645–654. IEEE, 2010.
19
| 8cs.DS
|
Maximizing a Monotone Submodular Function with a Bounded
Curvature under a Knapsack Constraint
arXiv:1607.04527v1 [cs.DS] 15 Jul 2016
Yuichi Yoshida∗
National Institute of Informatics and Preferred Infrastructure, Inc.
[email protected]
February 26, 2018
Abstract
We consider the problem of maximizing a monotone submodular function under a knapsack
constraint. We show that, for any fixed ǫ > 0, there exists a polynomial-time algorithm with an
approximation ratio 1 − c/e − ǫ, where c ∈ [0, 1] is the (total) curvature of the input function.
This approximation ratio is tight up to ǫ for any c ∈ [0, 1]. To the best of our knowledge,
this is the first result for a knapsack constraint that incorporates the curvature to obtain an
approximation ratio better than 1 − 1/e, which is tight for general submodular functions.
As an application of our result, we present a polynomial-time algorithm for the budget
allocation problem with an improved approximation ratio.
∗
Supported by JSPS Grant-in-Aid for Young Scientists (B) (26730009), MEXT Grant-in-Aid for Scientific Research
on Innovative Areas (24106003), and JST, ERATO, Kawarabayashi Large Graph Project.
1
Introduction
In this paper, we consider the problem of maximizing a monotone submodular function under a
knapsack constraint. Specifically, given a monotone submodular function f : 2E → R+ and a weight
function w : E → [0, 1], we aim to solve the following optimization problem:
maximize f (S) subject to w(S) ≤ 1 and S ⊆ E,
P
where w(S) =
e∈S w(e). This problem has wide applications in machine learning tasks such
as sensor placement [8], document summarization [11, 12], maximum entropy sampling [10], and
budget allocation [14]. Although this problem is NP-hard in general, it is known that we can
achieve (1 − 1/e)-approximation in polynomial time [15], and this approximation ratio is indeed
tight [5].
Although it is useful to know that we can always obtain (1 − 1/e)-approximation in polynomial
time, it is observed that a simple greedy method outputs even better solutions in real applications
(see, e.g., [8]), and it is more desirable if we can guarantee a better approximation ratio by making
assumptions on the input function. One such assumption is the notion of curvature, introduced
by Conforti and Cornuéjols [4]. For a monotone submodular function f : 2E → R+ , the (total)
curvature of f is defined as
fE−e(e)
cf = 1 − min
.
e∈E
f (e)
Intuitively speaking, the curvature measures how close f is to a linear function. To see this, note
that cf ∈ [0, 1] and cf = 0 if and only if f is a linear function.
It was shown in [4] that, for maximizing a monotone submodular function under a cardinality
constraint, the greedy algorithm achieves an approximation ratio (1 − e−cf )/cf , and the result was
extended to a matroid constraint [17]. Recently, Sviridenko et al. [16] obtained a polynomial-time
algorithm for a matroid constraint with an approximation ratio 1 − cf /e, and showed that this
approximation ratio is indeed tight for every cf ∈ [0, 1] even under a cardinality constraint (note
that 1 − cf /e is strictly larger than (1 − e−cf )/cf except when cf = 0 or cf = 1).
In this paper, we extend these results to a knapsack constraint and present a polynomial-time
algorithm under a knapsack constraint with an approximation ratio 1 − cf /e. More specifically, we
show the following:
Theorem 1.1. There is an algorithm that, given a monotone submodular function f : 2E → R+ ,
a weight function w : E → [0, 1], and ǫ ∈ (0, 1), outputs a (random) set S ⊆ E with w(S) ≤ 1 (with
probability one) satisfying
cf
−
ǫ
f (O).
[f
(S)]
≥
1
−
E
e
Here, O ⊆ E is an optimal solution to the problem, i.e., O is a set with w(O) ≤ 1 that maximizes
poly(1/ǫ)
f . The running time is O n5 + n4 polylog(n) · 1ǫ
, where n = |E|.
Note that the approximation ratio 1 − cf /e is indeed tight for every cf ∈ [0, 1] because the lower
bound given by [16] holds even for a cardinality constraint. To the best of our knowledge, this is the
first result for a knapsack constraint that incorporates the curvature to obtain an approximation
ratio better than 1 − 1/e, which is tight for general submodular functions.
We can apply our algorithm to all the above-mentioned applications to obtain a better solution
when the input function has a small curvature. As a representative example, we consider the
1
budget allocation problem [1], which models a marketing process that allocates a given budget
among media channels, such as TV, newspapers, and the Web, in order to maximize the impact
on customers. We model the process using a bipartite graph on a vertex set A ∪ B, where A and
B correspond to the media channels and the customers, respectively, and an edge (a, b) ∈ A × B
represents the potential influence of media channel a on customer b. In a simplified setting where
we can use each channel at most once, each media channel a ∈ A can activate a customer with
a predetermined probability pa ∈ [0, 1]. Then, we have to find
Pa set S ⊆ A that maximizes the
expected number of activated customers subject to w(S) := a∈S w(a) ≤ 1, where w(a) is the
cost of using media channel a. We can formulate this problem as the maximization of a monotone
submodular function f : 2A → R+ under the knapsack constraint w(S) ≤ 1. We can show that
cf ≤ 1 − minb∈B p|Γ(b)|−1 , where Γ(b) denotes the set of neighbors of b in the bipartite graph. By
Theorem 1.1, this immediately gives the approximation ratio 1 − cf /e − ǫ for this problem. The
actual model is more general and discussed in detail in Section 5.
1.1
Proof technique
Now, we present the outline of our proof. Let f : 2E → R+ be the input function and O ⊆ E be the
optimal solution, i.e., O is the set that maximizes f among the sets with weight at most one. We
assume that cf = 1 − Ω(ǫ); otherwise, we can use a standard algorithm [15] to achieve the desired
approximation ratio. Using the argument in [16], we can decompose the input function f into a
monotone submodular function g : 2E → R+ and a linear function ℓ : 2E → R+ such that, if we
can compute a set S ⊆ E with w(S) ≤ 1 and f (S) = g(S) + ℓ(S) ≥ (1 − 1/e)g(O) + ℓ(O), then S
is a (1 − cf /e)-approximate solution. Moreover, by slightly changing the argument in [16], we can
also assume that cg = 1 − Ω(ǫ(1 − cf )) = 1 − Ω(ǫ2 ).
In order to find the desired set S ⊆ E, we use a variant of the continuous greedy algorithm [3]
that simultaneously optimizes g and ℓ. In this algorithm, we consider continuous versions of g,
ℓ, and w, denoted by G : [0, 1]E → R+ , L : [0, 1]E → R+ , and W : [0, 1]E → R+ , respectively.
We note that the function G is called the multilinear extension of g, and that L and W are linear
functions. We start with the zero vector x ∈ [0, 1]E and then iteratively update it. The algorithm
consists of 1/ǫ iterations, and roughly speaking, in each iteration, we find a vector v ∈ [0, 1]E with
the following properties: (i) G(x + ǫv) − G(x) ≥ ǫ(G(x ∨ 1O ) − G(x)), (ii) L(ǫv) = ǫL(v) ≥ ǫℓ(O),
and (iii) W (ǫv) = ǫW (v) ≤ w(O). Then, we update x by adding ǫv. Here, 1O is the characteristic
vector of the set O and ∨ is the coordinate-wise maximum. Intuitively speaking, these conditions
mean that moving along the direction v from x is no worse than moving towards x ∨ 1O . We can
find such a vector v by linear programming. Then, after 1/ǫ iterations, we get a vector x ∈ [0, 1]E
such that G(x) ≥ (1 − 1/e)G(1O ) = (1 − 1/e)g(O), L(x) ≥ ℓ(O), and W (x) ≤ w(O). Finally, we
obtain a set S ⊆ E by rounding the vector x, where each element e ∈ E is added with probability
x(e).
Unfortunately, this strategy does not work as is. Here, a crucial issue is that we cannot show
the concentration of the weight in the rounding step. To address this issue, by borrowing an idea
from [2], we split the elements into large and small ones, where an element is said to be small if
g(e) ≤ ǫ6 g(O) and ℓ(e) ≤ ǫ6 ℓ(O), and is said to be large otherwise (in our analysis, it is more
convenient to define large and small elements in terms of g and ℓ instead of w). Then, since the
curvature of g is bounded away from one, we can bound the number of large elements in O by a
2
function of ǫ.1 Let OL , OS ⊆ O be the set of large and small elements in O, respectively. Further,
we let OL = {o1 , . . . , om }. Then, in each iteration, we do the following: For each i ∈ {1, . . . , m},
we find an element ei such that (i) G(x ∨ 1ei ) − G(x) ≥ G(x ∨ 1oi ) − G(x), (ii) ℓ(ei ) ≥ ℓ(oi ), and
(iii) w(ei ) ≤ w(oi ). Then, we update x by adding ǫ1ei . Here, 1e is a characteristic vector of the
element e ∈ E. Intuitively speaking, adding ei to the current solution is no worse than adding oi .
For small items, we find a vector v as before by considering the characteristic vector 1OS ; then,
we update x by adding ǫv. In the rounding step, we handle large and small elements separately.
Note that, for each i ∈ {1, . . . , m}, we have computed 1/ǫ elements (through 1/ǫ iterations). Then,
we chose one of them uniformly at random and add it to the output set. An advantage of this
rounding procedure is that we can guarantee that the chosen element for i ∈ {1, . . . , m} has weight
at most w(oi ). For small elements, we apply the previous rounding procedure with a minor tweak
to guarantee that the output set has weight at most one.
In order to realize this idea, we need to address several additional issues. First, as we do
not know the set O, we do not know values related to O, such as g(O), ℓ(O), G(x ∨ 1O ), and
G(x ∨ 1oi ). Hence, we cannot determine whether an element is small or large, and we cannot find
the desired vector or element in each iteration. We address this issue by guessing these values.
For example, we can show a lower bound and an upper bound on g(O) that are O(n) times apart.
This means that we can find a (1 − ǫ)-approximation to g(O) in the geometric sequence of length
O(log1+ǫ n) = O((log n)/ǫ) between the lower and upper bounds. If we naively guess all the values,
poly(1/ǫ)
as we have 1/ǫ iterations, the resulting time complexity will be poly(n)· (log n)/ǫ
. However,
2
since the function g has curvature 1 − Ω(ǫ ), we can reduce the number of candidate values and
thus improve the time complexity to poly(n) · (1/ǫ)poly(1/ǫ) .
1.2
Related work
As mentioned earlier, it has been shown that the greedy method achieves (1 − e−cf )/cf approximation for a cardinality constraint [4]. The result was extended to a matroid constraint by
Vondrák [17]. He showed that the result actually holds if we replace cf with the curvature to
∗
the optimum c∗f , and the approximation ratio (1 − e−cf )/c∗f is tight. Sviridenko et al. [16] improved
the approximation ratio to 1 − cf /e − ǫ for a matroid constraint (and hence, a cardinality constraint), which is unattainable with c∗f , and showed that the approximation ratio 1 − cf /e is tight
even for a cardinality constraint.
Curvature has been used to explain the empirical performance of the greedy method. Sharma et al. [13]
considered maximum entropy sampling on Gaussian radial basis functions (RBF) kernels, which
can be modeled as the maximization of a monotone submodular function, and showed that the
curvature of this problem is close to zero.
The maximization of a submodular function under a knapsack constraint has been studied extensively. Sviridenko obtained a (1−1/e)-approximation algorithm with time complexity O(n5 ) [15].
We can also obtain (1 − 1/e − ǫ)-approximation with a constant number of knapsack constraints;
however, the time complexity blows up to npoly(1/ǫ) [9]. It has been claimed in [2] that, for any
fixed ǫ > 0, there is a (1 − 1/e − ǫ)-approximation algorithm with time complexity Õ(n2 ). However,
as mentioned in the footnote, their argument has a drawback. Several approximation guarantees
have been achieved in [7] using various parameters of the input function. However, none of them
1
Although it is claimed in [2] that the number of large elements is bounded for any submodular function, it is not
true in general.
3
has an approximation ratio better than 1 − 1/e based solely on the assumption that the curvature
is bounded.
1.3
Organization
The remainder of this paper is organized as follows. Section 2 introduces the definitions used
throughout the paper and reviews the basic properties of submodular functions. Section 3 explains
the reduction to a joint approximation of a monotone submodular function and a monotone linear
function. Section 4 presents a joint approximation algorithm. Section 5 describes an application
to the budget allocation problem.
2
Preliminaries
For an integer n ∈ N, let [n] denote the set {1, . . . , n}. In this paper, the symbol E always denotes
a (finite) domain of a function.
P
For a function w : E → R and a subset S ⊆ E,Pwe define w(S) = e∈S w(e). Similarly, for a
vector x ∈ RE and a set S ⊆ E, we define x(S) = e∈S x(e). For an elementPe ∈ E, we define 1e
as the unit vector whose e-th element is 1. For a set S ⊆ E, we define 1S as e∈S 1e .
Let f : 2E → R be a function. For an element e ∈ E, we simply write f (e) to denote f ({e}).
For a set S ⊆ E, we define a function fS : 2E → R as fS (T ) = f (S ∪ T ) − f (S). We say that f is
submodular if, for any S, T ⊆ E,
f (S) + f (T ) ≥ f (S ∩ T ) + f (S ∪ T ).
An equivalent condition is the diminishing return property, which
P requires fS (e) ≥ fT (e) for any
S ⊆ T ( E and e ∈ E \ T . We say that f is linear if f (S) = e∈S f (e) holds for every S ⊆ E.
Note that, if f is submodular (resp., linear), then fS is also submodular (resp., linear).
For a vector x ∈ [0, 1]E , let R(x) denote a random set, where each element e ∈ E is included
in the set with probability x(e). For a submodular function f : 2E → R, the multilinear extension
F : [0, 1]E → R of f is defined as
X
Y
Y
F (x) := E[f (R(x))] =
f (S)
x(e)
(1 − x(e)).
S⊆E
e∈S
e∈E\S
For an element e ∈ E and a vector x ∈ [0, 1]E , we define ∂e F (x) as the slope of F at x in the
direction of 1e . The following fact is well known (see, e.g., [6]):
∂e F (x) =
E[fR(x) (e)]
F (x ∨ 1e ) − F (x)
=
.
1 − xe
1 − xe
(1)
The following lemma bounds the marginal gain of F when adding ǫy to x:
Lemma 2.1. Let f : 2E → R+ be a monotone submodular function and x, y ∈ [0, 1]E be vectors
such that x + ǫy ∈ [0, 1]E . Then,
X
F (x + ǫy) − F (x) ≥ ǫ
y(e) E[fR(x+ǫy) (e)].
e∈E
4
Proof.
ordering of elements in E. For i ∈ {0, 1, . . . , n}, let xi =
PiLet e1 , . . . , en be an arbitrary
x + j=1 ǫy(e)1e . Note that x0 = x and xn = x + ǫy. Then, we have
F (x + ǫy) − F (x) =
X
∂ei F (xi−1 )ǫy(ei )
(by multilinearity of F )
i∈[n]
≥
X
E[fR(xi−1 ) (e)]ǫy(ei )
(by (1))
i∈[n]
≥ǫ
X
y(ei ) E[fR(xn ) (e)].
X
y(ei ) E[fR(x+ǫy) (e)].
(by submodularity of f )
i∈[n]
=ǫ
i∈[n]
We frequently use the following form of Chernoff’s bound.
Lemma 2.2 (Relative+Additive Chernoff’s bound [2]). Let
P X1 , . . . , Xn be independent random
variables such that Xi ∈ [0, 1] for every i ∈ [n]. Let X = n1 i∈[n] Xi and µ = E[X]. Then, for any
α ∈ (0, 1) and β > 0, we have
nαβ
Pr[|X − µ| > αµ + β] ≤ 2 exp −
.
3
This immediately gives the following sampling algorithm:
Corollary 2.3. Suppose that we can obtain independent samples of a random variable X bounded
in [0, d]. Let µ = E[X]. Then, there exists an algorithm, denoted by Estimateα,β,δ (X), that, given
α, β, δ ∈ (0, 1), outputs a value µ̂ such that |µ̂ − µ| ≤ αµ + βd with probability at least 1 − δ. The
number of samples used by the algorithm is O(log(1/δ)/(αβ)).
3
Reduction
In this section, we prove Theorem 1.1 using the following theorem, which gives a joint approximation
of a monotone submodular function and a monotone linear function.
Theorem 3.1. There is an algorithm that, given a monotone submodular function g : 2E → R+ ,
a monotone linear function ℓ : 2E → R+ , a weight function w : E → [0, 1], and ǫ ∈ (0, 1), outputs
a (random) set S ⊆ E with w(S) ≤ 1 satisfying
1
g(O) + ℓ(O) − ǫ g(O) + ℓ(O) .
[g(S)
+
ℓ(S)]
≥
1
−
E
e
4
polylog(n)
Here, O = arg maxT ⊆E:w(T )≤1 g(T )+ℓ(T ) is an optimal solution. The running time is O n (1−c
·
2
g)
poly(1/ǫ)/(1−cg )
1
1
, where n = |E|.
ǫ log 1−cg
The proof of Theorem 3.1 is given in Section 4. In the remainder of this section, we prove
Theorem 1.1 using Theorem 3.1. The argument is similar to that used in [16], but it is more subtle
here because the running time in Theorem 3.1 depends on the curvature of g. We use the following
lemma.
5
Lemma 3.2 (Lemma 2.1 of [16]). If f : 2E → R+ is a monotone submodular function, then
P
e∈E fE−e (e) ≥ (1 − cf )f (S) for all S ⊆ E.
Theorem 3.3. There is an algorithm that, given a monotone submodular function f : 2E → R+ ,
a weight function w : E → [0, 1], and ǫ ∈ (0, 1), outputs a (random) set S ⊆ E with w(S) ≤ 1
satisfying
cf
− ǫ f (O).
E[f (S)] ≥ 1 −
e
4
polylog(n)
·
Here, O = arg maxT ⊆E:w(T )≤1 f (T ) is an optimal solution. The running time is O n (1−c
2
f)
poly(1/ǫ)/(1−cf )
1
1
, where n = |E|.
ǫ log 1−cf
Proof. Define the functions g, ℓ : 2E → R+ such that
ǫX
ℓ(S) = 1 −
fE−e (e) and g(S) = f (S) − ℓ(S)
2
e∈S
for every S ⊆ E.
It is not hard to see that ℓ is a nonnegative monotone linear function and that g is a nonnegative
monotone submodular function. Moreover, the curvature of g is
cg = 1 − min
e∈E
ǫ(1 − cf )
gE−e (e)
ǫ
fE−e (e) − (1 − ǫ/2)fE−e (e)
fE−e (e)
= 1 − min
≤ 1 − min
=1−
.
e∈E
g(e)
f (e) − (1 − ǫ/2)fE−e (e)
2 e∈E f (e)
2
Further, Lemma 3.2 implies that for any set S ⊆ E,
ǫ
ǫ
ǫX
fE−e (e) ≥ 1 −
(1 − cf )f (S) ≥ 1 − cf −
f (S).
ℓ(S) = 1 −
2
2
2
e∈S
By applying Theorem 3.1 to g, ℓ, w, and ǫ/2, we can find a (random) set S ⊆ E with w(S) ≤ 1
satisfying
1
ǫ
g(O) + ℓ(O) − g(O) + ℓ(O)
E[f (S)] = E[g(S) + ℓ(S)] ≥ 1 −
e
2
1
ǫ
1
f (O) + ℓ(O) − f (O)
= 1−
e
e
2
1 − cf − ǫ/2
1
ǫ
≥ 1−
f (O) +
f (O) − f (O)
e
e
2
cf
≥ 1−
− ǫ f (O).
e
The running time is clearly as stated.
Now, we prove our main theorem.
Proof of Theorem 1.1. If cf < 1 − eǫ, then we run the algorithm in Theorem 3.3. The approximapoly(1/ǫ)/(1−cf )
4
polylog(n)
1
1
=
log
tion factor is 1 − cf /e − ǫ and the running time is O n (1−c
·
2
ǫ
1−cf
f)
poly(1/ǫ)
O n4 polylog(n) · 1ǫ
.
6
If cf ≥ 1 − eǫ, then we simply run the O(n5 )-time (1 − 1/e)-approximation algorithm presented
in [15]. Then, the approximation factor is
1−
cf
1 − eǫ
1
≥1−
−ǫ≥1−
− ǫ.
e
e
e
In both cases, the approximation factor is at least 1 − cf /e − ǫ whereas the running time is as
desired.
4
Proof of Theorem 3.1
In this section, we prove Theorem 3.1. Throughout this section, G : [0, 1]E → R+ denotes the
multilinear extension of g, while L : [0, 1]E → R+ and W : [0, 1]E → R+ denote the following linear
functions:
X
X
L(x) =
x(e)ℓ(e) and W (x) =
x(e)w(e).
e∈E
e∈E
Furthermore, we define dg = maxe∈E g(e), dℓ = maxe∈E ℓ(e), and dg,ℓ = max(dg , dℓ ). Note that
we have g(O) ≤ ndg ≤ ndg,ℓ , ℓ(O) ≤ ndℓ ≤ ndg,ℓ , and dg,ℓ ≤ g(O) + ℓ(O) ≤ 2ndg,ℓ . Recall that
n = |E|.
In Section 4.1, we argue that we need to deal with “small” and “large” elements separately
in order to guarantee that we get a set satisfying the knapsack constraint after rounding. Our
algorithm updates a vector x ∈ [0, 1]E in several iterations and then rounds it. In Section 4.2, we
present an algorithm that updates x by adding a vector supported on small elements. In Section 4.3,
we present the entire algorithm that computes the vector x by taking large elements into account.
Then, in Section 4.4, we describe our rounding procedure. We need to guess several parameters
when running the algorithms in Section 4.2 and 4.3. Our final algorithm with the guessing process
is presented in Section 4.5.
4.1
Small elements
Our algorithm computes a vector x ∈ [0, 1]E with W (x) ≤ 1 and then rounds it to a set. A natural
rounding method is to simply output the random set R(x). Then, we can guarantee that the
expected objective values E[g(R(x))] and E[ℓ(R(x))] are sufficiently large and the expected weight
E[w(R(x))] is at most one. However, we cannot guarantee the concentration of w(R(x)) because
some elements have large contributions to the weight. To resolve this issue, we say that elements
in E are small if
g(e) ≤ ǫ6 g(O) and ℓ(e) ≤ ǫ6 ℓ(O).
Then, we can freely remove some of small elements for decreasing the weight without decreasing
the value significantly. Further, we can prove that the number of large elements is bounded by a
polynomial in ǫ and cg .
An issue here is that we do not know O; hence, we cannot determine whether an element is
small. To resolve this issue, we guess the values of g(O) and ℓ(O). Without loss of generality, we
can assume that ǫ/n = (1 − ǫ)k for some integer k; otherwise, we slightly decrease the value of ǫ.
Then, we define a set Vǫ,n (g, ℓ) = {ndg,ℓ , (1 − ǫ)nd, . . . , ǫdg,ℓ , 0}, and we use the values in Vǫ (g, ℓ) to
7
guess g(O) and ℓ(O). Since g(O) ≤ ndg,ℓ and ℓ(O) ≤ ndg,ℓ hold, there exist some vg , vℓ ∈ Vǫ,n (g, ℓ)
such that
(1 − ǫ)vg − ǫdg,ℓ ≤ g(O) ≤ vg
and (1 − ǫ)vℓ − ǫdg,ℓ ≤ ℓ(O) ≤ vℓ .
(2)
We say that an element e ∈ E is small with respect to (vg , vℓ ) if
g(e) ≤ ǫ6 vg
and ℓ(e) ≤ ǫ6 vℓ .
Otherwise, we say that an element e ∈ E is large with respect to (vg , vℓ ). Let EL (vg , vℓ ) ⊆ E
and ES (vg , vℓ ) ⊆ E be the sets of large and small elements, respectively, with respect to (vg , vℓ ).
Further, we define OL (vg , vℓ ) = EL (vg , vℓ ) ∩ O and OS (vg , vℓ ) = ES (vg , vℓ ) ∩ O. We omit vg and vℓ
from these notations if they are clear from the context.
When vg and vℓ satisfy (2), we can upper bound the number of large elements in O:
Lemma 4.1. If vg and vℓ satisfy (2), then we have |OL | = O (1−c1g )ǫ6 .
Proof. Let mℓ be the number of elements e ∈ O with ℓ(e) > ǫ6 vℓ . Then, we have ǫ6 vℓ mℓ ≤ ℓ(O).
Since vℓ ≥ ℓ(O), we have mℓ ≤ 1/ǫ6 .
Let {o1 , . . . , omg } be the set of elements e ∈ O with g(e) > ǫ6 vg . Then, we have
X
X
g(O) ≥
g{o1 ,...,oi−1 } (oi ) ≥ (1 − cg )
g(oi ) ≥ (1 − cg )ǫ6 vg mg .
i∈[mg ]
Since vg ≥ g(O), we have mg ≤
i∈[mg ]
1
.
(1−cg )ǫ6
Then, we have |OL | ≤ mg + mℓ = O( (1−c1g )ǫ6 ).
In addition to the values of g(O) and ℓ(O), the value of |OL | is not also not known. However, we
can easily guess it because there are only O( (1−c1g )ǫ6 ) choices from Lemma 4.1. We use the symbol
m to denote the guessed value of |OL |.
For each choice of vg , vℓ , and m, we compute a (random) set that jointly maximizes g and ℓ, and
the final output is the best one among them. Since |Vǫ,n (g, ℓ)| = O(log1/(1−ǫ) (n/ǫ)) = O(log(n/ǫ)/ǫ),
2
2
(n/ǫ)
this guessing process makes the running time O log(n/ǫ)/ǫ · (1−c1g )ǫ6 = O log
times
8
(1−cg )ǫ
larger. The details will be explained in Section 4.5.
4.2
Subroutine for handling small elements
Here, we explain a subroutine that finds a vector v ∈ [0, 1]E supported on the set ES of small
in order to update the current vector
elements (with respect to the current guesses vg and vℓ ) P
E
x ∈ R . We want v to satisfy the following properties: (i) e∈ES v(e) E[gR(x) (e)] ≥ E[gR(x) (OS )],
(ii) L(v) ≥ ℓ(OS ), and (iii) W (v) ≤ w(OS ).
There are several issues in finding such a vector v: We cannot exactly calculate E[gR(x) (e)] (e ∈
ES ); hence, we need to estimate it. Further, we do not know the values E[gR(x) (OS )] and ℓ(OS ). In
the subroutine presented here, we assume that their guessed values, denoted by γ and λ, respectively,
are given as a part of the input. Once we succeed in accurately estimating E[gR(x) (e)] (e ∈ ES )
and the given guessed values γ and λ are sufficiently accurate, we can find the desired vector v by
solving a linear program. A detailed description of the subroutine is given in Algorithm 1.
Now, we analyze Algorithm 1. From Corollary 2.3 and the fact that gS (e) ≤ dg ≤ dg,ℓ for every
S ⊆ E and e ∈ E, we have the following:
8
Algorithm 1 SmallElementsǫ,δ (g, ℓ, w, ES , γ, λ, x)
Input: A monotone submodular function g : 2E → R+ , a monotone linear function ℓ : 2E → R, a
weight function w : E → [0, 1], ǫ, δ ∈ (0, 1), a set of small elements ES , guessed values γ, λ, and
a vector x ∈ [0, 1]E .
Output: A vector v ∈ [0, 1]E .
1: For each e ∈ ES , let θ(e) = Estimateǫ,ǫ/n,δ/n (gR(x) (e)).
2: Find a vector v ∈ [0, 1]E supported on ES that minimizes W (v) subject to
v · θ ≥ (1 − ǫ)γ − ǫdg,ℓ
3:
and L(v) ≥ λ.
by linear programming.
return v.
Proposition 4.2. With probability at least 1 − δ, we have
(1 − ǫ) E[gR(x) (e)] −
ǫdg,ℓ
ǫdg,ℓ
≤ θ(e) ≤ (1 + ǫ) E[gR(x) (e)] +
n
n
for every e ∈ ES .
We formalize the concept that γ and λ are sufficiently accurate, and then show that Algorithm 1
outputs a desired vector with accurate γ and λ.
Definition 4.3. We say that γ and λ are good guesses if
E[gR(x) (OS )] ≥ γ ≥ (1 − ǫ) E[gR(x) (OS )] − ǫdg,ℓ
and
ℓ(OS ) ≥ λ ≥ (1 − ǫ)ℓ(OS ) − ǫdg,ℓ
hold, respectively.
Since E[gR(x) (OS )] ≤ ndg,ℓ and ℓ(OS ) ≤ ndg,ℓ hold, we can find good guesses by trying all the
values in the set Vǫ,n (g, ℓ).
Lemma 4.4. Suppose that γ and λ are good guesses. Then, Algorithm 1 returns a vector v ∈ [0, 1]E
supported on ES such that
P
3
(i)
e∈ES v(e) E[gR(x) (e)] ≥ (1 − ǫ) E[gR(x) (OS )] − 3ǫdg,ℓ ,
(ii) L(v) ≥ (1 − ǫ)ℓ(OS ) − ǫdg,ℓ , and
(iii) W (v) ≤ w(OS ),
with probability at least 1 − δ. The time complexity of Algorithm 1 is O(n4 + n2 log(n/δ)/ǫ2 ).
Proof. With probability at least 1 − δ, the consequence of Proposition 4.2 holds. In what follows,
we assume that this occurs.
The vector 1OS satisfies
1OS · θ =
X
e∈OS
θ(e) ≥
X
ǫdg,ℓ
(1 − ǫ) E[gR(x) (e)] −
n
e∈OS
≥ (1 − ǫ) E[gR(x) (OS )] − ǫdg,ℓ ≥ (1 − ǫ)γ − ǫdg,ℓ .
9
Furthermore, we have L(1OS ) = ℓ(OS ) ≥ λ. Hence, the vector v is well defined, and in particular,
we have W (v) ≤ W (OS ).
Then, we have
X
e∈ES
X
ǫdg,ℓ
≥ (1 − ǫ)
v(e)θ(e) − ǫdg,ℓ
v(e) θ(e) −
n
e∈ES
e∈ES
≥ (1 − ǫ) (1 − ǫ)γ − ǫdg,ℓ − ǫdg,ℓ ≥ (1 − ǫ)2 γ − 2ǫdg,ℓ
≥ (1 − ǫ)2 (1 − ǫ) E[gR(x) (OS )] − ǫdg,ℓ − 2ǫdg,ℓ
v(e) E[gR(x) (e)] ≥ (1 − ǫ)
X
≥ (1 − ǫ)3 E[gR(x) (OS )] − 3ǫdg,ℓ .
It is easy to confirm (ii) and (iii). The time complexity for computing θ is O(n2 log(n/δ)/ǫ2 )
from Corollary 2.3, and the time complexity for solving the linear program is O(n4 ) by using the
ellipsoid method. The total time complexity is bounded by O(n4 + n2 log(n/δ)/ǫ2 ).
4.3
Continuous greedy algorithm with guessing
In this section, we present an algorithm whose goal is to output a vector x ∈ [0, 1]E such that (i)
G(x) ≥ (1 − 1/e)g(O), (ii) L(x) ≥ ℓ(O), and (iii) W (x) ≤ w(O).
Our algorithm is a variant of the continuous greedy algorithm [3] but differs in the following
aspects: we consider two functions g and ℓ, and we handle large and small elements separately. Let
m be an integer given as a parameter, which is a guessed value of |OL | (with respectSto the current
b=
values of vg and vℓ ). Then, we make copies E1 , . . . , Em of E and define a set E
i∈[m] Ei ∪ ES .
b
Then, we define a function b
g : 2E → R+ as gb(S1 , S2 , . . . , Sm , SS ) = g(S1 ∪ · · · ∪ Sm ∪ SS ). We note
b be the multilinear extension of b
that gb is a monotone submodular function. Let G
g.
E
We introduce a vector yi ∈ [0, 1] for each i ∈ [m] and another vector z ∈ [0, 1]E . We always
guarantee that yi (i ∈ [m]) is supported on Ei and z is supported on ES . Our algorithm runs
in 1/ǫ iterations and updates the vectors yi (i ∈ [m]) and z in each iteration. Here, we assume
that 1/ǫ is an integer; otherwise, we P
slightly decrease ǫ. The final output is the sequence of vectors
(y1 , . . . , ym , z), and their sum x := i∈[m] yi + z will satisfy the conditions stated initially in this
section. We call the first iteration the iteration at time 0, the second iteration the iteration at time
ǫ, and so on.
To explain how we update the vectors, we introduce several notations. For t = {0, ǫ, . . . , 1}, we
define yit (i ∈ [m]) and z t as the vectors yi and z immediately before the iteration at time t. We
note that yi0 = 0 (i ∈ [m]) and z 0 = 0 hold. We define yi1 (i ∈ [m]) and z 1 as yi (i ∈ [m]) and
z, respectively, after the iteration at time 1 − ǫ. P
Note that the algorithm outputs the sequence of
t
t
1 , z 1 ). Then, we define xt =
vectors (y11 , . . . , ym
i∈[m] yi + z . Further, for t ∈ {0, ǫ, . . . , 1 − ǫ}
P
P
and i ∈ {0, 1, . . . , m}, we define xti = j≤i yjt+ǫ + j>i yjt + z t , i.e., the vector obtained after the
iteration at time t − ǫ followed by updating y1 , . . . , yi . Note that xt0 = xt .
As in the argument in Section 4.1, we try all possible values for guessing |OL |. Hence, in what
follows, we assume that the guessed value m is correct, i.e., m = |OL |.
Let o1 , . . . , om be the large elements in O, i.e., OL = {o1 , . . . , om }. For i ∈ [m], let obi be the
bL = {b
b =O
bL ∪ OS ⊆ E.
b For each i ∈ [m],
copy of oi in Ei . Then, we define O
o1 , . . . , obm } and O
t+ǫ
t
t
we update the vector yi to yi by finding an element ei ∈ Ei and adding the vector ǫ1eti . Here,
we want the element eti to satisfy (i) E[b
gR(xti−1 ) (eti )] ≥ E[b
gR(xti−1 ) (b
oi )], (ii) ℓ(eti ) ≥ ℓ(oi ), and (iii)
10
Algorithm 2 GuessingContinuousGreedyǫ,δ (g, ℓ, w, EL , ES , m, {γit }, {γSt }, {λi }, λS )
Input: A monotone submodular function g : 2E → R+ , a monotone linear function ℓ : 2E → R, a
weight function w : E → [0, 1], ǫ, δ ∈ (0, 1), a set of large and small elements EL and ES , an
integer m ∈ N, guessed values {γit }i∈[m],t∈{0,ǫ,...,1−ǫ} , {γSt }t∈{0,ǫ,...,1−ǫ} , {λi }i∈[m] , and λS .
Output: A vector x ∈ [0, 1]E .
b
b
1: yi ← 0 ∈ [0, 1]E for i ∈ [m] and z ← 0 ∈ [0, 1]E .
2: for (t ← 0; t ≤ 1 − ǫ; t ← t + ǫ) do
3:
for (i ← 1; i ≤ m; i ← i + 1) do
4:
θi (e) ← Estimateǫ,ǫ/m,ǫδ/(2nm) (E[b
gR(x) ](e)) for each e ∈ Ei .
5:
Let e = arg min{w(e) | e ∈ Ei , θi (e) ≥ (1 − ǫ)γit − ǫd/m, ℓ(e) ≥ λi }.
6:
yi ← yi + ǫ1e .
P
7:
v ← SmallElementsǫ,ǫδ/2 (b
g , ℓ, w, ES , γSt , λS , i∈[m] yi + z).
8:
z ← z + ǫv.
9: return (y1 , . . . , ym , z).
w(eti ) ≤ w(oi ). As we do not know the values of E[b
gR(xti−1 ) (b
oi )] and ℓ(oi ), the algorithm requires
t
their guessed values γi and λi , respectively. We do not have to guess w(oi ) because we will choose
the element with the minimum weight satisfying (i) and (ii).
Then, we update the vectorP
z t to z t+ǫ by finding a vector v t and adding the vector ǫv t . Here, we
b
gR(x) (e)] ≥ E[b
gR(xti−1 ) (OS )] (note that OS ⊆ ES ⊆ E),
want the vector v to satisfy (i) e∈ES v(e) E[b
(ii) L(v) ≥ ℓ(OS ), and (iii) W (v) ≤ w(OS ). Such a vector can be found by calling SmallElements
with the guessed values γSt and λS for E[b
gR(xti−1 ) (OS )] and ℓ(OS ), respectively. A detailed description
of the algorithm is given in Algorithm 2.
b t+ǫ ) − G(xt ) ≥ ǫ(g(O) − G(x
b t+ǫ )), which is sufficient to show that
We will show that G(x
b 1 ) is close to the (1 − 1/e)-approximation to g(O). We will also show that L(x1 ) ≥ ℓ(O) and
G(x
W (x) ≤ w(O).
For t ∈ {0, ǫ, . . . , 1 − ǫ} and i ∈ [m], let θit be the θi used in the iteration at time t. From
Lemma 2.2 and the union bound, we immediately have the following:
Proposition 4.5. With probability at least 1 − δ/2, we have
(1 − ǫ) E[b
gR(xti−1 ) (e)] −
ǫdg,ℓ
ǫdg,ℓ
gR(xti−1 ) (e)] +
≤ θit (e) ≤ (1 + ǫ) E[b
m
m
for every t ∈ {0, ǫ, . . . , 1 − ǫ}, i ∈ [m], and e ∈ Ei .
We formalize the concept that γit and λi are sufficiently accurate.
Definition 4.6. For t ∈ {0, ǫ, . . . , 1 − ǫ} and i ∈ [m], we say that γit is a good guess if
gR(xti−1 ) (b
oi )] ≥ γit ≥ (1 − ǫ) E[b
gR(xti−1 ) (b
oi )] −
E[b
holds. For i ∈ [m], we say that λi is a good guess if
ℓ(oi ) ≥ λi ≥ (1 − ǫ)ℓ(oi ) −
holds.
11
ǫdg,ℓ
m
ǫdg,ℓ
m
Since E[b
gR(xti−1 ) (oi )] ≤ dg,ℓ and ℓ(oi ) ≤ dg,ℓ hold for every i ∈ [m], we can find good guesses by
trying all the values in the set Vǫ,m(g, ℓ)/m := {v/m | v ∈ Vǫ,m (g, ℓ)}.
Lemma 4.7. Suppose that the consequence of Proposition 4.5 holds and that {γit } and {λi } are
good guesses. Then, for every t ∈ {0, ǫ, . . . , 1 − ǫ}, we have the following:
(i) E[b
gR(xti−1 ) (eti )] ≥ (1 − ǫ)3 E[b
gR(xti−1 ) (b
oi )] − 3ǫdg,ℓ /m,
(ii) ℓ(eti ) ≥ (1 − ǫ)ℓ(oi ) − ǫdg,ℓ /m for i ∈ [m], and
(iii) w(eti ) ≤ w(oi ) for i ∈ [m].
Proof. Fix t ∈ {0, ǫ, . . . , 1 − ǫ} and i ∈ [m]. Note that we have
gR(xti−1 ) (b
oi )] −
θit (oi ) ≥ (1 − ǫ) E[b
ǫdg,ℓ
ǫdg,ℓ
≥ (1 − ǫ)γit −
m
m
and ℓ(oi ) ≥ λi . Since oi (in Ei ) is a candidate for eti , the element eti is well defined. In particular,
we have w(eti ) ≤ w(oi ) because eti is chosen as the element with the minimum element satisfying
the conditions.
We have
(1 + ǫ) E[b
gR(xti−1 ) (eti )] +
2ǫdg,ℓ
ǫdg,ℓ
ǫdg,ℓ
gR(xti ) (oi )] −
≥ θit (eti ) ≥ (1 − ǫ)γit −
≥ (1 − ǫ)2 E[b
.
m
m
m
Rearranging this inequality, we get (i). Further, (ii) is immediate from the fact that ℓ(eti ) ≥ λi .
We say that γSt for t ∈ {0, ǫ, . . . , 1 − ǫ} and λS are good guesses if they are good guesses in the
sense of Definition 4.3. Then, we have the following:
Lemma 4.8. Suppose that {γit }, {λiP
}, {γSt }, and λS are good guesses. Then, Algorithm 2 returns
vectors y1 , . . . , ym , z such that x := i∈[m] yi + z satisfies the following:
b
(i) G(x)
≥ (1 − 1/e − O(ǫ))g(O) − 6ǫdg,ℓ ,
(ii) L(x) ≥ (1 − O(ǫ))ℓ(O) − 2ǫdg,ℓ , and
(iii) W (x) ≤ w(O),
2
with probability at least 1 − δ. The running time is O( nm
log
ǫ3
nm
ǫδ
+
n4
ǫ
+
n2
ǫ3
log
1
ǫδ ).
Proof. With probability 1−δ/2, the consequence of Proposition 4.5 holds. Further, with probability
1 − δ/2, all the invocations of SmallElements succeed in outputting vectors with the guarantees
in Lemma 4.4. By the union bound, all these occur with probability at least 1 − δ. In what follows,
we assume that this occurs.
12
First, we check (i). For each t ∈ {0, ǫ, . . . , 1 − ǫ}, we have
b tm ) − G(x
b t) =
G(x
≥ǫ
m
X
b tj−1 )
b tj−1 + ǫ1et ) − G(x
G(x
j
j=1
m
X
gR(xtj−1 ) (etj )]
E[b
b
(By concavity of G)
j=1
≥ ǫ(1 − ǫ)3
m
X
3ǫdg,ℓ
m
gR(xtj−1 ) (b
oj )] −
E[b
j=1
(By (i) of Lemma 4.7)
m
X
3
≥ ǫ(1 − ǫ)
oj ) − 3ǫ2 dg,ℓ
E gbR(xtm )∪{bok |k∈[j−1]}(b
j=1
bL ) − b
= ǫ(1 − ǫ)3 E b
g(R(xtm ) ∪ O
g(R(xtm )) − 3ǫ2 dg,ℓ
bL )]) − 3ǫ2 dg,ℓ .
≥ ǫ(1 − ǫ)3 (E[b
gR(xtm ) (O
For each t ∈ {0, ǫ, . . . , 1 − ǫ}, we have
b t+ǫ ) − G(x
b tm ) = G(x
b tm + ǫv) − G(x
b tm )
G(x
X
≥ǫ
v(e) E[b
gR(xt+ǫ ) (e)]
e∈E
gR(xt+ǫ ) (OS )] − 3ǫdg,ℓ
≥ ǫ (1 − ǫ)3 E[b
2
3
(by Lemma 2.1)
(by (i) of Lemma 4.4)
gR(xt+ǫ ) (OS )] − 3ǫ dg,ℓ .
≥ ǫ(1 − ǫ) E[b
Combining these two inequalities, we get
b t+ǫ ) − G(x
b t)
G(x
bL )]) − 3ǫ2 dg,ℓ + ǫ(1 − ǫ)3 E[b
gR(xt+ǫ ) (OS )] − 3ǫ2 dg,ℓ
≥ ǫ(1 − ǫ)3 (E[b
gR(xtm ) (O
b − 6ǫ2 dg,ℓ
gR(xt+ǫ ) (O)]
≥ ǫ(1 − ǫ)3 E[b
b − G(x
b t+ǫ ) − 6ǫ2 dg,ℓ = ǫ(1 − ǫ)3 g(O) − G(x
b t+ǫ ) − 6ǫ2 dg,ℓ .
≥ ǫ(1 − ǫ)3 gb(O)
Rewriting the above inequality, we get
g(O) −
β
b t
β
b t+ǫ ) ≤ g(O) − α − G(x ) ,
− G(x
α
1+α
where α = ǫ(1 − ǫ)3 and β = 6ǫ2 dg,ℓ . Then, by induction, we can prove that
g(O) −
β
β
1
b t) ≤
g(O)
−
.
− G(x
α
α
(1 + α)t/ǫ
Substituting t = 1 and rewriting once again, we get
1
b
G(x ) ≥ 1 −
β
1
1
β
1
(g(O) − ) ≥ 1 − − O(ǫ) g(O) −
≥ 1 − − O(ǫ) g(O) − 6ǫdg,ℓ ,
α
e
α
e
(1 + α)1/ǫ
13
assuming that ǫ is sufficiently small, say, less than 1/2.
To see (ii), we have for t ∈ {0, . . . , 1 − ǫ} that
L(x
t+ǫ
t
t
) − L(x ) = L x + ǫ
=ǫ
m
X
m
X
i=1
1eti + ǫv − L(xt )
ℓ(eti ) + ǫ(1 − ǫ)ℓ(OS ) − ǫ2 dg,ℓ
(by (ii) of Lemma 4.4)
i=1
≥ ǫ((1 − ǫ)ℓ(OL ) − ǫdg,ℓ ) + ǫ(1 − ǫ)ℓ(OS ) − ǫ2 dg,ℓ
(by (ii) of Lemma 4.7)
2
≥ ǫ(1 − ǫ)ℓ(O) − 2ǫ dg,ℓ .
By induction, we get L(x1 ) ≥ (1 − ǫ)ℓ(O) − 2ǫdg,ℓ .
To see (iii), we have for t ∈ {0, . . . , 1 − ǫ} that
m
X
1eti + ǫv − W (xt )
W (xt+ǫ ) − W (xt ) = W xt + ǫ
i=1
≤ ǫ(w(OL ) + w(OS ))
(By (iii)’s of Lemmas 4.4 and 4.7)
= ǫw(O).
By induction, we get W (x1 ) ≤ w(O).
nm
m
Finally, we analyze the time complexity. For estimating the θi ’s, we need O( nm
ǫ · ǫ2 log ǫδ ) =
2
nm
1
1
n
nm2
)=
O( ǫ3 log ǫδ ) time. The time complexity of SmallElements is at most O( ǫ · n4 + ǫ2 log ǫδ
n2
1
n4
O( ǫ + ǫ3 log ǫδ ). Hence, the running time is as desired.
4.4
Rounding
In this section, we explain how to round the vectors obtained by GuessingContinuousGreedy
(Algorithm 2).
Let (y1 , . . . , ym , z) be the vectors obtained by GuessingContinuousGreedy, and let v t be
the vector supported
on ES obtained in the iteration at time t in GuessingContinuousGreedy.
P
Note that z = t∈{0,ǫ,...,1−ǫ} v t . Our algorithm is summarized in Algorithm 3.
We use the following lemma to analyze the objective value of the output set.
Lemma 4.9 (Lemma 3.7 of [3]). Let E = E1 ∪ · · · ∪ Ek , let f : 2E → R+ be a monotone submodular
function, and for all i 6= j, we have Ei ∩ Ej = ∅. Let x ∈ RE
+ such that for each Ei we have
x(Ei ) ≤ 1. If T is a random set where we sample independently from each Ei at most one random
element, i.e., element e with probability x(e), then
E[f (T )] ≥ F (x).
b
Lemma 4.10. We have E[g(SL ∪ SS )] ≥ (1 − ǫ)G(x)
− ǫ3 vg and E[ℓ(SL ∪ SS )] ≥ (1 − ǫ)L(x) − ǫ3 vℓ .
14
Algorithm 3 Roundingǫ (w, EL , ES , m, {yi }, z)
Input: A weight function w : E → [0, 1], EL , ES ⊆ E, a set of vectors {yi }i∈[m] , and a vector z.
Output: A set S ⊆ E.
1: SL ← ∅, SS ← ∅.
2: Define z ′ ∈ [0, 1]ES as z ′ (e) = (1 − ǫ)z(e) if w(e) < ǫ3 maxt W (v t ) and z ′ (e) = 0 otherwise.
3: For each e ∈ ES , add it to SS independently with probability z ′ (e).
4: for i ∈ [m] do
5:
Add exactly one element in Ei to SL (as an element of E), where an element e ∈ Ei is
chosen with probability yi (e).
6: if w(SS ∪ SL ) ≤ 1 then
7:
return SS ∪ SL .
8: else
9:
return ∅.
Proof. Let x′ =
b ′) = G
b
G(x
P
i∈[m] yi
X
i∈[m]
+ z ′ . First, let us relate the value of the vector x to that of x′ .
X
b
yi + (1 − ǫ)z −
yi + z ′ ≥ G
i∈[m]
X
e∈ES :w(e)≥ǫ3 maxt
X
b (1 − ǫ)(
≥G
yi + z) − max g(e)
e∈ES
i∈[m]
z(e)g(e)
W (vt )
X
z(e)
e∈ES :w(e)≥ǫ3 maxt W (vt )
1
b
b
− ǫ 3 vg .
≥ (1 − ǫ)G(x)
− ǫ6 vg · 3 ≥ (1 − ǫ)G(x)
ǫ
Next, we note that we get SL by selecting exactly one random element from each Ei , which is
a copy of EL , and we get SS by sampling independently from v. Hence, by applying Lemma 4.9
with sets E1 , . . . , Em and sets {{e} | e ∈ SS }, we get
b
− ǫ3 vg .
E[g(SL ∪ SS )] ≥ (1 − ǫ)G(x)
By a similar argument, we get L(x′ ) ≥ (1 − ǫ)L(x) − ǫ3 vℓ , and we have E[ℓ(SL ∪ SS )] ≥
(1 − ǫ)L(x) − ǫ3 vℓ .
Next, we show that the probability that the weight of the output set exceeds w(O) decays
exponentially.
Lemma 4.11. For any γ ≥ 1, we have w(SL ∪ SS ) ≤ γw(O) with probability 1 − exp −Ω(γ/ǫ2 ) .
Proof. Recall that, for each i ∈ [m], the vector yi is the sum of 1/ǫ elements e0i , . . . , e1−ǫ
i , and
t
we pick one of them in Algorithm 3. By the condition w(ei ) ≤ w(oi ) for every i ∈ [m] and
t ∈ {0, ǫ, . . . , 1 − ǫ}, the weight of the large elements after the rounding will be less than that of the
large elements of the optimal solution. Hence, it is sufficient to prove that w(SS ) ≤ γw(OS ) holds
with probability 1 − exp(−Ω(γ/ǫ2 )), where SS is the set obtained by rounding z ′ .
First, note that
E[w(SS )] = E[w(R(z ′ ))] ≤ (1 − ǫ) E[w(R(z))] ≤ (1 − ǫ) max W (v t ) ≤ (1 − ǫ)w(OS ).
t
15
For each e ∈ E, we set up a random variable Xe to be Xe = w(e)/(ǫ3 w(OS )) if e ∈ SS and
Xe =P0 otherwise. Note that each Xe is bounded in [0, 1] because maxt W (v t ) ≤ w(OS ). For
X = e∈ES Xe , we have µ := E[X] = E[w(SS )]/(ǫ3 w(OS )) ≤ (1 − ǫ)/ǫ3 .
Invoking Lemma 2.2 with α = ǫ/2 and β = γ/(2ǫ3 ), we have
h
h
i
γi
Pr[w(SS ) > γw(OS )] = Pr X ≥ 3 ≤ Pr X ≥ (1 + α)µ + β
ǫ
γ
αβ
= exp −Ω 2 .
≤ 2 exp −
3
ǫ
Lemma 4.12. Algorithm 3 outputs a (random) set S with w(S) ≤ 1 satisfying
b
+ L(x)) − O(ǫ) · (g(O) + ℓ(O) + vg + vℓ ).
E[g(S) + ℓ(S)] ≥ (1 − ǫ)(G(x)
Proof. It is clear that we always have w(S) ≤ 1.
Now, we analyze the objective value attained
by S. For any γ ≥ 1, the probability that
w(SL ∪ SS ) > γw(O) is at most exp −Cγ/ǫ2 for some C > 0 by Lemma 4.11. Note that, if T ⊆ E
satisfies w(T ) ≤ γw(O), then g(T ) + ℓ(T ) ≤ γ(g(O) + ℓ(O)) from the submodularity of g + ℓ. By
Lemma 4.10, we have
E[g(S) + ℓ(S)]
≥ E[g(SL ∪ SS ) + ℓ(SL ∪ SS )] −
Z
∞
γ(g(O) + ℓ(O)) exp(−Cγ/ǫ2 )dγ
1
ǫ4 + Cǫ2
exp(−C/ǫ2 )(g(O) + ℓ(O))
C2
b
= (1 − ǫ)(G(x)
+ L(x)) − O(ǫ) · (g(O) + ℓ(O) + vg + vℓ ).
b
≥ (1 − ǫ)G(x)
− ǫ3 vg + (1 − ǫ)L(x) − ǫ3 vℓ −
4.5
Putting things together
Now, we present our entire algorithm. The idea is to simply guess vg , vℓ , m, {γit }, {λi }, {γSt },
and λS , run Algorithm 2 with the guessed values, and then round the obtained vectors using
Algorithm 3.
Naively, we have O(|Vǫ,n (g, ℓ)|O(1/ǫ) ) = O((log(n/ǫ)/ǫ)O(1/ǫ) ) choices for the sequence {γSt }. We
can decrease the number of choices since g has a bounded curvature. If we have a guess γS0 such
that γS0 ≥ g(OS ) ≥ (1 − ǫ)γS0 , then we must have γS0 ≥ gS (OS ) ≥ (1 − ǫ)(1 − cg )γS0 for any set
S ⊆ E. Hence, it suffices to consider sequences whose maximum and minimum values are within
a factor of (1 − ǫ)(1 − cg ). Let Vǫ,n,γ 0 (g, ℓ) := {v ∈ Vǫ,n (g, ℓ) | v ≥ (1 − ǫ)(1 − cg )γS0 }. Then, the
S
number of such sequences is at most |Vǫ,n (g, ℓ)| · |Vǫ,n,γ 0 (g, ℓ)|O(1/ǫ) , which is much smaller than
S
O((log(n/ǫ)/ǫ)O(1/ǫ) ).
A detailed description of our algorithm is given in Algorithm 4.
Proof of Theorem 3.1. Consider the case that vg and vℓ satisfy (2), m = |OL |, and {γit }, {λi }, {γSt },
and λS are good guesses. Let S be the (random) set obtained with these guesses. By Lemma 4.12,
we have
b
+ L(x)) − O(ǫ)(g(O) + ℓ(O) + vg + vℓ ).
E[g(S) + ℓ(S)] ≥ (1 − ǫ)(G(x)
16
(3)
Algorithm 4 Knapsack
Input: A monotone submodular function g : 2E → R+ , a linear function ℓ : 2E → R, a weight
function w : E → R+ , and ǫ ∈ (0, 1).
Output: A set S ⊆ E satisfying w(S) ≤ 1.
1: for each choice of vg , vℓ ∈ Vǫ,n (g, ℓ) do
2:
EL ← the set of large elements with respect to vg and vℓ .
3:
ES ← the set of small elements with respect to vg and vℓ .
4:
S ← ∅.
5:
M := ⌊ (1−c1g )ǫ6 ⌋
6:
for each choice of m from {0, 1, . . . , M } do
7:
for each choice of {γit }, {λi } from Vǫ,m(g, ℓ)/m do
8:
for each choice of γS0 , λS from Vǫ,n (g, ℓ) do
9:
for each choice of {γSǫ , . . . , γS1−ǫ } from Vǫ,n,γ 0 (g, ℓ) do
S
10:
(y1 , . . . , ym , z) ← GuessingContinuousGreedyǫ,ǫ (g, ℓ, w, EL , ES , m, {γit }, {λi }, {γSt }, λS ).
11:
S ← Roundingǫ (w, EL , ES , m, {yi }, z).
12:
S ← S ∪ {S}.
13: return arg maxS∈S g(S) + ℓ(S).
Conditioned on the event that GuessingContinuousGreedy succeeds, by (i) and (ii) of Lemma 4.8,
we get
(3) ≥ (1 − ǫ)(1 − 1/e − O(ǫ))g(O) + (1 − ǫ)(1 − O(ǫ))ℓ(O) − O(ǫ)(g(O) + ℓ(O) + vg + vℓ + 8dg,ℓ )
≥ (1 − 1/e)g(O) + ℓ(O) − O(ǫ)(g(O) + ℓ(O)).
(4)
Since GuessingContinuousGreedy succeeds with probability at least 1 − ǫ, we get
E[g(S) + ℓ(S)] ≥ (1 − ǫ) · (4) ≥ (1 − 1/e)g(O) + ℓ(O) − O(ǫ)(g(O) + ℓ(O)).
Since Algorithm 4 outputs the set with the maximum objective, we have the desired property on
the objective value.
It is clear that the output of Algorithm 4 has weight at most 1 because Rounding always outputs
a set of weight at most 1.
For arbitrary γ ∈ Vǫ,n (g, ℓ), the time complexity of Algorithm 4 is
n4 n2
nM
1
· |Vǫ,n (g, ℓ)|O(1) · |Vǫ,n,γ (g, ℓ)|O(1/ǫ) · |Vǫ,m (g, ℓ)|O(M/ǫ)
+
+
log
ǫ3
ǫ
ǫ
ǫ3
ǫ
nM 2
n4 n2
1 O(1/ǫ) log(M/ǫ) O(M/ǫ)
nM
1 log(n/ǫ) O(1) 1
·
·
·
+
+
log
log
log
=O
ǫ3
ǫ
ǫ
ǫ3
ǫ
ǫ
ǫ
1 − cg
ǫ
4
2
log n O(1)
n
n
n
1
n
·
·
log
+
+ 3 log
=O
2
15
(1 − cg ) ǫ
(1 − cg )ǫ
ǫ
ǫ
ǫ
ǫ
1
1 O(1/ǫ) 1
1 O(1/((1−cg )ǫ7 ))
log
log
·
ǫ
1 − cg
ǫ
1 − cg
n4 polylog(n) 1
1 poly(1/ǫ)/(1−cg )
log
·
.
=O
(1 − cg )2
ǫ
1 − cg
O
nM 2
log
17
Hence, we have the desired time complexity.
By replacing ǫ with ǫ/C for a large constant C (to change O(ǫ) to ǫ), we have the desired
result.
5
The Budget Allocation Problem
In this section, we bound the curvature of the submodular function that represents the budget
allocation problem, and we confirm that our algorithm can be applied to the budget allocation
problem in order to obtain an approximation factor better than 1 − 1/e.
We formally define the budget allocation problem. The input consists of a bipartite graph
with the bipartition A ∪ B, a weight function w : A → [0, 1], a capacity function c : A → N,
and a probability function p : A → [0, 1]. Intuitively speaking, the sets A and B correspond to
media channels and customers, respectively. Each edge (a, b) in the bipartite graph represents the
potential influence of P
media channel a on customer b. Consider a budget allocation b ∈ ZA
+ to A
with b(a) ≤ c(a) and a∈A b(a)w(s) ≤ 1. If a node a is allocated a budget of b(a), it makes b(a)
independent trials to activate each adjacent node b. The probability that b is activated by a in
each trial is p(a). Thus, the probability that b becomes active is
Y
1−
p(a)b(a) ,
a∈Γ(b)
where Γ(b) denotes the set of nodes in A adjacent to b. Hence, the expected number of activated
target nodes is
X
Y
1−
p(a)b(a) .
b∈B
a∈Γ(b)
The objective of this problem is to find the budget allocation that maximizes the expected number
of activated target nodes.
We can recast the
S problem using a submodularEfunction. For each a ∈ A, let Ea = {(a, i) | i ∈
c(a)}, and let E = a∈A Ea . Then, we define f : 2 → R+ as
Y
X
1−
p(a)|S∩Ea | .
f (S) =
b∈B
a∈Γ(b)
Further, we define w′ : 2E → [0, 1] to be w′ ((a, i)) = w(a). Then, the budget allocation problem is
equivalent to maximizing f (S) subject to w′ (S) ≤ 1.
We now observe several properties of f .
Lemma 5.1. Let S ( E and (a, i) ∈ E \ S. Then,
X
Y
fS ((a, i)) =
(1 − p(a))
p(a′ )|S∩Ea′ | .
b∈B:a∈Γ(b)
a′ ∈Γ(b)
Q
Proof. For each b ∈ B, we define a function g b : 2E → R+ as g b (T ) = 1 − a∈Γ(b) p(a)|T ∩Ea | . Note
P
that fS ((a, i)) = b∈B gSb ((a, i)).
If a 6∈ Γ(b), then we clearly have gSb ((a, i)) = 0. If a ∈ Γ(b), then we have
Y
gSb ((a, i)) = (1 − p(a))
p(a′ )|S∩Ea′ | .
a′ ∈Γ(b)
18
Summing gSb ((a, i)) over all b ∈ B, we obtain the claim.
Corollary 5.2. The function f is submodular.
Proof. From Lemma 5.1, it is easy to see that fS ((a, i)) ≥ fT ((a, i)) holds for S ⊆ T ( E and
(a, i) ∈ E \ T .
Corollary 5.3. The curvature cf of f satisfies
cf ≤ 1 − min
min
a∈A b∈B:a∈Γ(b)
p(a)c(a)−1
Y
′
p(a′ )c(a ) .
a′ ∈Γ(b)\{a}
Proof. From Lemma 5.1, we have
fE\(a,i) ((a, i)) =
X
(1 − p(a))p(a)c(a)−1
f ((a, i)) =
′
p(a′ )c(a ) ,
a′ ∈Γ(b)\{a}
b∈B:a∈Γ(b)
X
Y
(1 − p(a)).
b∈B:a∈Γ(b)
Hence,
fE\(a,i) ((a, i))
f ((a, i))
(a,i)∈E
P
Q
c(a)−1
′ c(a′ )
b∈B:a∈Γ(b) (1 − p(a))p(a)
a′ ∈Γ(b)\{a} p(a )
P
= 1 − min
a∈A
b∈B:a∈Γ(b) (1 − p(a))
Y
′
p(a′ )c(a ) .
≤ 1 − min min p(a)c(a)−1
cf = 1 − min
a∈A b∈B:a∈Γ(b)
a′ ∈Γ(b)\{a}
From our main result (Theorem 1.1) and Corollaries 5.2 and 5.3, when the capacity of each node
a ∈ A is bounded by a constant and the number of vertices adjacent to each node b ∈ B is bounded
by a constant, we obtain a polynomial-time algorithm whose approximation ratio is strictly better
than 1 − 1/e.
Acknowledgments
We thank Takanori Maehara for providing us with the problem and insightful comments.
References
[1] N. Alon, I. Gamzu, and M. Tennenholtz. Optimizing budget allocation among channels and
influencers. In Proceedings of the 21st International Conference on World Wide Web (WWW),
pages 381–388, 2012.
[2] A. Badanidiyuru and J. Vondrák. Fast algorithms for maximizing submodular functions. In
Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages
1497–1514, 2013.
19
[3] G. Calinescu, C. Chekuri, M. Pál, and J. Vondrák. Maximizing a monotone submodular
function subject to a matroid constraint. SIAM Journal on Computing, 40(6):1740–1766,
2011.
[4] M. Conforti and G. Cornuéjols. Submodular set functions, matroids and the greedy algorithm:
Tight worst-case bounds and some generalizations of the rado-edmonds theorem. Discrete
Applied Mathematics, 7(3):251–274, 1984.
[5] U. Feige. A threshold of ln n for approximating set cover. Journal of the ACM, 45(4):634–652,
1998.
[6] M. Feldman. Maximization Problems with Submodular Objective Functions. PhD thesis, Technion, 2013.
[7] R. K. Iyer and J. A. Bilmes. Submodular optimization with submodular cover and submodular
knapsack constraints. In Proceedings of the 27th Annual Conference on Neural Information
Processing Systems (NIPS), pages 2436–2444, 2013.
[8] A. Krause, A. P. Singh, and C. Guestrin. Near-optimal sensor placements in gaussian processes:
Theory, efficient algorithms and empirical studies. Journal of Machine Learning Research,
9:235–284, 2008.
[9] A. Kulik, H. Shachnai, and T. Tamir. Maximizing submodular set functions subject to multiple
linear constraints. In Proceedings of the 20th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 545–554, 2013.
[10] J. Lee. Maximum Entropy Sampling. John Wiley & Sons, Ltd, 2006.
[11] H. Lin and J. Bilmes. Multi-document summarization via budgeted maximization of submodular functions. pages 912–920. Proceedings of the 2010 Annual Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), 2010.
[12] H. Lin and J. Bilmes. A class of submodular functions for document summarization. In
Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics:
Human Language Technologies (ACL-HLT), pages 510–520, 2011.
[13] D. Sharma, A. Kapoor, and A. Deshpande. On greedy maximization of entropy. In Proceedings
of the 32nd International Conference on Machine Learning (ICML), pages 1330–1338, 2015.
[14] T. Soma, N. Kakimura, K. Inaba, and K. Kawarabayashi. Optimal budget allocation: Theoretical guarantee and efficient algorithm. In Proceedings of the 31st International Conference
on Machine Learning (ICML), pages 351–359, 2014.
[15] M. Sviridenko. A note on maximizing a submodular set function subject to a knapsack constraint. Operations Research Letters, 32(1):41–43, 2004.
[16] M. Sviridenko, J. Vondrák, and J. Ward. Optimal approximation for submodular and supermodular optimization with bounded curvature. In Proceedings of the 26th Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA), pages 1134–1148, 2015.
20
[17] J. Vondrák. Submodularity and curvature: the optimal algorithm. RIMS Kokyuroku Bessatsu,
B23:253–266, 2010.
21
| 8cs.DS
|
Planning system for deliveries in Medellín
Catalina Patiño-Forero
Mateo Agudelo-Toro
Mauricio Toro
Universidad EAFIT
Medellín, Colombia
Universidad EAFIT
Medellín, Colombia
Universidad EAFIT
Medellín, Colombia
[email protected]
[email protected]
[email protected]
ABSTRACT
Here we present the implementation of an application capable of
planning the shortest delivery route in the city of Medellín,
Colombia. We discuss the different approaches to this problem
which is similar to the famous Traveling Salesman Problem
(TSP), but differs in the fact that, in our problem, we can visit
each place (or vertex) more than once. Solving this problem is
important since it would help people, especially stores with
delivering services, to save time and money spent in fuel,
because they can plan any route in an efficient way.
To solve this we need to construct a subgraph with the
delivering points, based on the city’s map, and it will be a
complete one i.e. all of its vertices are connected. Then we will
give the user different options that will determine which
algorithm will be used to solve the problem. Between these
options there is only one that will surely give the shortest route
and works only with twenty or less points. The other options are
quite fast but may or may not give the shortest route.
Depending on the chosen algorithm, the results in time, memory
and total distance will vary. For example, we have an algorithm
that does not create a subgraph to give an answer, so it takes less
memory and time, but will not give the total distance. Others can
give a better answer quite fast, even though they require to
compute a subgraph, but still the tour produced may not be the
shortest one. At last, there is an algorithm that can give the
shortest route every time, but needs to look through all possible
answers so it takes much more time.
For the problem of planning delivery routes in Medellín our
proposed solution to find the shortest route can be of huge help
for small companies if their couriers do not visit more than 20
points per trip.
Author Keywords
Planning, deliveries, routing, graph, complexity, shortest path.
ACM Classification Keywords
Theory of computation → Design and analysis of algorithms →
Approximation algorithms analysis → Routing and network
design problems
1. INTRODUCTION
Efficiently planning the deliveries is something really useful for
any company in the field. Here we talk about creating an
efficient program that gives an optimal delivering route for a
courier, in order to minimize the time spent traveling; the
courier can pass over one place more than once. Without this
last condition we would have a TSP which, though it is a
“simple” problem formulated over 200 years ago [9], does not
have any optimal solution for big graphs (thousands of
vertexes). Since it is simpler (and possible) to treat our problem
as TSP, we are going to do so.
We will see the different approaches to this problem and also
discuss the selection of the best available choice for our specific
case.
2. PROBLEM
As we just stated, we are trying to create an efficient program
that gives an optimal (shortest total distance) delivering route for
a courier, which minimizes the time spent traveling; this route
can repeat places which were already visited. In our case, we
will implement it for the city of Medellín in Colombia, but it
does not mean the algorithm cannot be used for other cities.
This efficient route planning request is quite difficult to compute
if we want to get an optimal answer. This is due the incredible
amount of possibilities we will have, since the idea is to use the
algorithm for real cities, for example Medellín, which has a
population that surpasses the 2 million people [11]. So it is to be
expected that the algorithm will take an incredible amount of
time to give an appropriate answer, time that may exceed what
we can spend on it. We can take the TSP as an example, which
requires a time proportional to (n-1)!/2 to execute (where n is
the number of places or nodes) [10], which means that for 20
destinations it would require a about 12 years to compute using
an average computer. We will treat our problem as TSP but
using a faster algorithm that requires less than 3 seconds to
compute the path for 20 points, but that would require 14 years
for 45 points on the same computer.
3. RELATED WORK
3.1 Minimum Spanning Tree (MST)
Given a weighted graph, the MST is the cheapest subset of
edges that keeps the graph in one connected component [1].
Those are very useful because they give approximate solutions
to the traveling salesman problem very efficiently.
One efficient way to compute the MST of a graph is the
Kruskal’s algorithm. It is a greedy algorithm that starts by
placing each vertex on its own connected component. It then
iterates over the edges having them sorted in non-descending
order, merging the two vertices connected by the current edge if,
and only if, they are currently on different components. The
complexity of this algorithm is O(m*log(m)) where m is the
number of edges.
3.2 Hamiltonian Path and Cycle
A Hamiltonian Path is a path between two vertices of a graph
that visits each vertex exactly once [5]. A Hamiltonian Cycle is
a closed loop through a graph that visits each node exactly once
[6]. A closed loop is a cycle in a graph in which the first vertex
is the same as the last [7]. A graph possessing a Hamiltonian
Path is said to be a Hamiltonian Graph [8].
There is a backtracking approach to find whether an undirected
graph is Hamiltonian. We start by creating an empty array and
adding vertex 0 to it. We will try to add the other vertices
starting from 1, but before that, we check whether it is adjacent
to the previously added vertex and if it is not already added. If
we find such vertex, we add it as part of the solution. If we do
not find a vertex then we return false [1]. Anyway, the
complexity of this algorithm is O(n!) where n is the number of
vertices, just like the naïve approach.
3.3 Eulerian Path and Cycle
An Eulerian Path is a path in a graph that visits each edge
exactly once [3], and an Eulerian Cycle is an Eulerian Path
which starts and ends in the same vertex [2]. It is similar to the
Hamiltonian path because in both we want to visit some part of
the graph only once. The difference is that in this case we want
to walk through each edge instead of visiting each vertex. This
difference changes everything: while the Hamiltonian path is an
NP-Complete problem for a general graph, finding whether a
given graph is Eulerian (has an Eulerian Cycle) can be done in
O(n + m), where n is the number of vertices in the graph and m
the number of edges.
To find whether a undirected graph is Eulerian it must have all
its non-zero degree vertices connected (which can be done using
a DFS traversal) and the number of vertices with odd degree
must be 1 (if it is 2 then the graph has a Eulerian Path instead)
[4].
3.4 Chinese Postman Problem (CPP)
In this problem, given a weighted graph, the postman wants to
find the shortest path that visits every edge at least once
returning to the starting point.
This problem can be solved in an optimal way by adding the
appropriate edges to the graph to make it Eulerian, because that
is basically what the problem is: finding an (especial) Eulerian
Cycle in the graph. Specifically, we find the shortest path
between each pair of odd-degree vertices in the graph. Adding a
path between two odd-degree vertices in G turns both of them to
even-degree, moving G closer to becoming an Eulerian graph.
Finding the best set of shortest paths to add to G reduces to
identifying a minimum-weight perfect matching in a special
graph G’. For undirected graphs, the vertices of G’ correspond
the odd-degree vertices of G, with the weight of edge (i, j)
defined to be the length of the shortest path from i to j in G. For
directed graphs, the vertices of G’ correspond to the degreeimbalanced vertices from G, with the bonus that all edges in G’
go from out-degree deficient vertices to in-degree deficient ones.
Thus, bipartite matching algorithms suffice when G is directed
[1]. Once the graph is Eulerian, the actual cycle can be extracted
using the algorithm described above.
4. DATA STRUCTURES
To implement our algorithm, we need two graphs: one to store
the city itself and other to store the points we need to visit.
For the graph of the city, we use adjacency list representation.
To achieve this we created four classes: vertex, edge, point and
pair. The first two, vertex and edge, are used directly on the
representation of the graph: we use HashMaps with vertex
objects as keys and edge objects as values. Point class represents
the latitude and longitude of the vertex in the world, and is used
to access vertices given its latitude and longitude. Pair objects
are used during the execution of A* and Dijkstra’s algorithm,
which require a priority queue of vertices (first value of the pair)
sorted by some value acquired during the execution of the
algorithm (second value of the pair).
The other is a complete graph that contains only the vertices we
want to visit. It is stored as an adjacency matrix, using a
primitive double’s 2D array with dynamically assigned integer
codes to vertices (used as indices), and where the edges are
completed using either Dijkstra’s or A* algorithms on the city’s
graph (depending on the amount of points). Since it is a
complete graph, using an adjacency matrix is better than an
adjacency list, because both need the same memory space, but
the adjacency matrix is faster when looking up for the distance
between any two vertices and that is a main requirement of our
algorithms.
There are other auxiliary data structures used during the
execution of different parts of the program. The Natural
Approximation algorithm uses a HashMap with vertices as keys
and integers as values to remember the original positions of the
vertices to visit. To read the vertices the user wants to visit, we
use a HashMap with points as keys and vertices as values,
because the URL only contains information about the point and
not the vertex itself. When reading the edges, we use a HashMap
with the code of the vertices (long data type) as keys and
vertices as values to access the vertices in the graph since the
edge’s specifications contain only the codes of the vertices it
connects and its cost (distance). In the case where the user gives
a point that it is not in our map, we compute a close enough one
and use a HashMap with vertex objects as keys and point objects
as values to store the original points of the approximated
vertices we found. This is done with the aim of returning to the
user the same points he or she entered. Finally, we use
ArrayLists of vertex objects to store both the points entered by
the user and the generated tours.
The reason to use HashMaps is that we do not require our
mapping to be stored in order, which allows us to use the most
efficient map available in Java.
5. ALGORITHM
First, the program creates the city’s graph by reading its
specifications which are given in two text files, one for the
vertices and other for the edges.
Then it reads the URL containing the points from the user, the
program finds the nearest vertex to the ones corresponding to the
given coordinates, which can be the exact same point. After this,
the program will compute the complete subgraph containing
only the points of interest. In order to create the subgraph the
program will choose between two different algorithms,
depending on the number of nodes given by the user: A*
algorithm is used if this number is less or equal to five,
otherwise it uses Dijkstra’s algorithm. For the heuristic of the
A* algorithm, it uses the Manhattan distance. Then it will
execute one of the following three algorithms, according to
user’s choice:
1.
Natural approximation: the algorithm is a simple sort over
the vertices using its 2D coordinates, using a custom
comparator to determine whether a point is to the left or to
the right of another. This algorithm has two versions: the
first (or the fastest) version just performs the sort described
above and completes the tour adding the initial vertex,
while the second version does the same as the first one, but
also computes the subgraph and the reverse tour, compares
both total lengths and returns the best tour. Notice that the
fastest version does not generate the subgraph. The method
for comparing the points is based on [11].
2.
3.
Nearest neighbor: This algorithm starts with the first vertex
and greedily chooses the nearest neighbor and moves to it.
It keeps doing the same thing until all the vertices are
visited. Finally, it adds the starting vertex to the end in
order to complete the tour. This algorithm will not always
give an optimal solution, but it’s execution time is much
faster than the brute force and can produces better tours
than natural approximation.
Brute force: we try every possible path starting with the
initial vertex, ending at i and passing through all the other
vertices different to i. To this path we add the distance from
i to the initial vertex to complete the tour.
We now present the complexity analysis for the algorithms
where n is the number of points to visit, E the number of edges
(335719) and V the number of vertices (179403) in the city’s
graph.
Build subgraph
Brute Force
Algorithm
Nearest Neighbor
Natural
approximation
(Normal mode)
Natural
approximation
(Fast mode)
Wost case complexity
Memory
Time
O(n(E+VlogV))
O(n2)
O(n2n)
O(n)
O(n22n)
O(n)
O(nlogn)
O(n)
O(nlogn)
O(n2)
Figure 2. Program’s menu
The first five options are to choose the way the user wants the
program to calculate the route, after choosing, the program will
show the time needed to generate the route and will take the user
to a Google Maps page where he or she can see this route. Since
the options number two, three, four and five need to internally
create a subgraph, the first time any of these options is chosen,
the program will show the time spent creating it.
1.
2.
3.
Table 1. Complexity table
6. IMPLEMENTATION
When starting the program the first thing the user will see is an
“Initializing” signal, shown to make the user wait until the
program builds the city’s graph. After this, the program will
show the time required to build the city’s graph and ask for the
Google Maps URL containing the points to be visited (Figure 1).
If the user inputs an invalid URL, the program will tell so and
then ask again for a new one (Figure 1). It will repeat this
process until the user gives a valid input. If the user wants to end
the program from here, he or she may input an ‘x’ (lower or
upper case) as URL and the program will finish.
4.
5.
Figure 1. Starting the program and invalid URL
After this, the program shows a warning telling the user that the
route given by our program may differ with Google’s because
the points we used to build the city’s graph are different. Then
offers a menu with 6 or 7 different options (Figure 2), depending
on how many points were given, to compute a tour which covers
all the given points and comes back to the first one (the user
should not add the first point at the end of the path).
6.
7.
“Natural approximation fast mode” will present a
route that tends to be circular and, compared with the
other options, it is the fastest, but it will not give any
distance and may not show the shortest route.
“Natural approximation normal mode” will also
present a route that tends to be circular, but this time
shows the total distance and may even show a route
that is shorter than the one obtained using last mode,
although it may still not be the shortest possible route.
“Nearest Neighbor” will give a route formed by
finding the closest point from the current one, and then
move to it. This process will repeat until reaching the
last point (corresponding to the first one). It may not
give the shortest possible route and its execution time
is similar to the second option. This option shows total
distance too.
“The best of both” will choose the best route between
the ones generated by last two options and show the
total distance; this will take a little more time. We
encourage its usage when there are more than 20
points in the URL and the user wants the shortest
route.
“Exact” will always show the shortest route with its
respective distance. It is potentially much slower than
the other options, reason why is not present in the
menu when the URL has more than twenty points,
since that would take a lot of time. If the user wants,
he can expand the maximum of points he or she can
input to ‘unlimited’. For this, when the program asks
for an URL (either when starting the program or
changing the current one) “extreme-mode” must be
written and press the enter key, then a warning will
appear and the program will ask for the URL to
continue as usual.
“Change URL” lets the user change the current URL,
if the new URL is not valid it will indicate so and ask
for a new one; if the given input is an ‘x’ the program
will end.
“Exit” will end the program.
Something important to consider is that, when the URL contains
at least two points that are not connected between them, the
program cannot calculate a distance, so it will tell the user so
and use the first option (“Natural approximation fast mode”) to
compute a possible route.
7. RESULTS
MEDELLÍN
WITH
THE
MAP
OF
7.1 Execution Time
The following table shows the time in seconds that each
algorithm takes to process a route for a different number of
vertices. The time required to build the subgraph for that same
amount of points is also shown.
Vertices
Build subgraph
Brute Force
Best of Both
Nearest Neighbor
Algorithm
Natural approximation
(Normal mode)
Natural approximation
(Fast mode)
5
0,4182
0
0
0
10
1,5146
0,0014
0
0
15
2,1448
0,0428
0,0001
0
20
2,7558
2,1206
0,0002
0
0
0
0
0
0
0
0
0
Considering current limits imposed by Google Maps in terms of
the available amounts of destinations for a single route (ten
vertices), it is easy to notice that our algorithm will run in
feasible time even when looking for an optimal solution after
building the subgraph, and in under two seconds if it has not
been build.
7.2 Memory space
The following table shows the memory in megabytes that each
algorithm takes to process a route for a different number of
vertices. Is important to notice that the memory used by the
subgraph can only be included in the algorithms that need it.
Algorithm
Table 2. Execution time (seconds) on a Core i7 2410u processor.
Build subgraph
Brute Force
Nearest Neighbor
Natural
approximation
(Normal mode)
Natural
approximation
(Fast mode)
5
27
0
3
10
58
1
4
Vertices
15
73
2
3
20
111
100
4
24
136
4771
4
3
3
3
4
3
2
4
4
4
4
Table 3. Memory used (in MB)
Remember that the Natural Approximation (Fast mode) does not
require the subgraph to be computed by the time it is called. For
the rest of the algorithms, the program only needs to compute
the subgraph once, because after building it, it will be saved and
reused until a new subgraph is required (new URL).
Is easy to notice that for all algorithms, except brute force, the
memory is almost constant. This is due, the amount of vertices is
not enough to show any difference.
As expected, the brute force algorithm is the slower one. For the
others, it is hard to compare their running times with such a little
amount of vertices.
In the other hand, one can see the incredible difference in
memory used by the brute force algorithm. This was
expected since the memory use increase exponentially
(O(n2n)). In graphic 3 we plot the obtained results.
Graphic 3. Memory for Brute force
Graphic 1. Brute force’s execution time on a Core i7 2410U
processor.
Graphic 2. Subgraph construction time on a Core i7 2410U
processor.
7.3 Total distance
The following graphic shows the total distance (in meters) of the
routes found as shortest by the different algorithms for random
sets of points.
Graphic 4. Total distance
http://mathworld.wolfram.com/EulerianCycle.html.
[Accessed 27 August 2016].
Graphic 4 shows how volatile Nearest Neighbor and Natural
Approximation algorithms are: in one case they got the optimal
answer whereas in other cases both obtained longer routes than
the optimal solution with a difference higher than 10 kilometers.
[3]
E. W. Weisstein, "Eulerian Path," Wolfram MathWord,
[Online].
Available:
http://mathworld.wolfram.com/EulerianPath.html.
[Accessed 28 August 2016].
[4]
GeeksForGeeks, "Eulerian path and circuit for
undirected
graph,"
GeeksForGeeks,
[Online].
Available: http://www.geeksforgeeks.org/eulerian-pathand-circuit/. [Accessed 28 August 2016].
[5]
E. W. Weisstein, "Hamiltonian Path," Wolfram
MathWorld,
[Online].
Available:
http://mathworld.wolfram.com/HamiltonianPath.html.
[Accessed 28 August 2016].
[6]
E. W. Weisstein, "Hamiltonian Cycle," Wolfram
MathWorld,
[Online].
Available:
http://mathworld.wolfram.com/HamiltonianCycle.html.
[Accessed 28 August 2016].
[7]
E. W. Weisstein, "Graph Cycle," Wolfram MathWorld,
[Online].
Available:
http://mathworld.wolfram.com/GraphCycle.html.
[Accessed 28 August 2016].
[8]
GeeksForGeeks, "Backtracking | Set 6 (Hamiltonian
Cycle),"
GeeksForGeeks,
[Online].
Available:
http://www.geeksforgeeks.org/backtracking-set-7hamiltonian-cycle/. [Accessed 28 August 2016].
[9]
University of Waterloo, "History of the TSP,"
University of Waterloo, Enero 2007. [Online].
Available:
http://www.math.uwaterloo.ca/tsp/history/index.html.
[Accessed 28 August 2016].
[10]
A. Levitin, Introduction to The Design and Analysis of
Algorithms, 3rd Edition, New Jersey: Pearson, 2012, p.
116.
[11]
C. e. dos, "Github," 18 Abril 2016. [Online]. Available:
https://github.com/mvpossum/eldiego/blob/master/geo
metria/orden.radial.cpp. [Accessed 4 September 2016].
The code can be found at
https://subversion.assembla.com/svn/planningdeliverysystemme
dellin/
8. CONCLUSIONS
The problem to calculate a delivery route given some points in a
map is not new and has been researched for years. As prove to
that, we can find the famous Traveling Salesman Problem or
TSP, in which the delivery route may not have repeated points,
except for the first one, that is the last one at the same time. This
problem was defined in the 1800s and even today, there are not
any algorithms that can give the best route efficiently: we still
have to choose between efficiency and precision. For the
problem of planning delivery routes in Medellín, that can be
modeled as the TSP problem, our proposed solution can be of
huge help for small companies when their couriers go out for a
route because it’s unlikely that, on a single trip, they will visit
more than 20 points. If the delivery route can consider repeated
points it will be harder to solve, so is better to simply solve TSP.
The biggest problem is to find an efficient way to give an
answer: algorithms that can give optimal answers require a lot of
time and memory, to the point that they cannot be used for big
graphs, and algorithms that can give an answer without
consuming too much resources, may not be able to give the
shortest route. Even if we apply all possible optimizations to the
code, it is still not enough to efficiently compute an optimal
solution.
Thanks to this work, we were able to understand the limitations
a computer has and the need of implementing efficient
algorithms with the appropriate data structures, otherwise a
program could take too much time to execute, time that the user
is not able or willing to spend waiting, since it can get to a point
where it needs years to compute an answer. And this gets worse
considering that right now we are living in a world were a lot of
data is stored, and only accessing it may take a lot of the
computer’s resources.
9. FUTURE WORK
Currently, there are two big limitations we would like to fix in
the future:
1.
2.
Our graph does not fit Google Map’s very well, which
makes the distances and the routes to be shown in a
very different way in many cases.
The only way to add more than 10 points on a route is
by working with the URL and the points’ latitude and
longitude.
10. REFERENCES
[1]
S. Skiena, The Algorithm Design Manual, New York:
Springer, 2008.
[2]
E. W. Weisstein,
MathWorld,
"Eulerian
[Online].
Cycle,"
Wolfram
Available:
| 8cs.DS
|
arXiv:1705.02299v1 [math.AG] 5 May 2017
BOUNDS ON THE TENSOR RANK
EDOARDO BALLICO, ALESSANDRA BERNARDI, LUCA CHIANTINI, ELENA GUARDO
Abstract. We give a sufficient criterion for a lower bound of the cactus rank
of a tensor. Then we refine that criterion in order to be able to give an
explicit sufficient condition for a non-redundant decomposition of a tensor to
be minimal and unique.
1. Introduction
The study of minimal decompositions of a given tensor T as a linear combination
of rank one tensors is a hot topic in many areas, ranging from pure algebraic geometry to applications to signal processing, big data analysis, quantum information...
Vectors vj,i ∈ Cnj +1 , j = 1, . . . , k, i = 1, . . . , r, such that
(1)
T =
r
X
λi v1,i ⊗ · · · ⊗ vk,i
i=1
for some λi ∈ C \ {0} determine a decomposition of T . We will say that the
decomposition is non-redundant (cf. Definition 2.1) if we cannot extract any proper
subset of {v1,i ⊗ · · · ⊗ vk,i }i=1,...,r which generates T .
Since we will use geometric arguments through the paper, we use a geometric
notation. Thus we identify (up to scalar multiplication) a tensor T with a point in
the projective space P(Cn1 +1 ⊗ · · · ⊗ Cnk +1 ) and a decomposition of T with a finite
subset S of the Segre embedding of the abstract product P(Cn1 +1 )×· · ·×P(Cnk +1 ) =
Pn1 × · · · × Pnk , such that T belongs to the linear span of S.
Cleary, having a non-redundant decomposition of a given tensor T does not imply
that such a decomposition is minimal, i.e. it has a minimal number of addenda (so
that r is the rank rk(T ) of T , cf. Definition 2.3).
In this paper we give a criterion that certifies if a non-redundant decomposition of
a general tensor T is also minimal, thus that it computes the rank of T (cf. Theorem
3.1). Moreover, under certain conditions, we can also show that a decomposition is
the unique minimal decomposition of T (cf. Theorem 4.6).
These two facts rely on our main result Theorem 3.1, where we give a criterion to
find a lower bound for the cactus rank (cf. Definition 2.3). The idea is geometrically
quite simple: Assume one has a non-redundant decomposition S of a tensor T ∈
P(Cn1 +1 ⊗ · · · ⊗ Cnk +1 ), then one can flatten the product Pn1 × · · · × Pnk in a
partition of two factors and study the geometry of the two projections of S, to get
the result.
One can easily compare our result with the Kruskal’s result on the identifiability
of tensors ([K77]), which implies a criterion for the minimality of a decomposition.
1
2
E. BALLICO, A. BERNARDI, L. CHIANTINI, E. GUARDO
It turns out that our criterion is geometrically simpler, and it applies in a wider
range of numerical cases.
Theorem 3.1 has the consequence that if a non-redundant decomposition projects
onto two linearly independents subset in the flattening, then it is also a minimal
one (cf. Example 3.6). This can be considered as a step towards the celebrated
Strassen’s Conjecture on the rank of a block tensor (cf. [S73]).
We show also that Theorem 3.1 provides new evidences towards the Comon’s
Conjecture, i.e. the equality between the rank and the symmetric rank of a symmetric tensor T . In a wide numerical range, if we know the existence of a symmetric
decomposition of T with sufficiently general geometric properties, then the conjecture holds for T (cf. Corollary 3.10). For instance, the Corollary applies for general
tensors in P(Sym6 (C3 )) and P(Sym8 (C3 )), i.e. for general forms of degree 6 and 8
in three variables.
Another consequence of Theorem 3.1 is described in Theorem 3.8. Given a
minimal decomposition with r addenda, then for any integer r′ such that r ≤ r′ <
−1 + Π(ni + 1), it is always possible to find a non-redundant decomposition with r′
addenda. We notice that this happens to be false for symmetric decompositions of
a symmetric tensor, even in the case of S d (C2 ) with d ≥ 4 (cf. Sylvester Theorem
in e.g. [CS11, BGI11]).
In Section 4 we focus on the identifiability of a minimal decomposition, meaning
the uniqueness of a decomposition of a tensor T with exactly rk(T ) addenda. The
main result of this section is Theorem 4.6 (that is again a consequence of Theorem
3.1), where we point out a sufficient condition for a non-redundant decomposition
to be minimal and unique. Again we compare our result with Kruskal’s bound
([K77]). Since Kruskal’s bound is sharp ([SB00], [D13]) and since our geometric
assumptions are weaker, and then easier to verify, than Kruskal’s ones, we cannot
hope to produce applications outside Kruskal’s numerical range. There are few cases
in which our the numerical range of application matches with Kruskal’s range. One
of them e.g. is given by tensors of type 3 × 2 × 2 × 2.
2. Notation and preliminaries
For a subscheme Z ⊂ Pm , we indicate with hZi the linear span of Z and with
deg(Z) its length (when it is finite). If Z is finite and reduced, we indicate with ♯Z
the cardinality of Z.
For any product of projective spaces Pn1 × · · · × Pnk call ν the Segre map
ν : Pn1 × · · · × Pnk → PM ,
M = −1 + Π(ni + 1).
In order to have a more compact notation we will always write
Y := Pn1 × · · · × Pnk
for the abstract product, and
X := ν(Y ) ⊂ PM
for the Segre variety.
For any i = 1, . . . , k call πi the projection of Pn1 × · · · × Pnk to the i-th factor.
We can generalize this notation by setting, for any collection of sub-indices u =
{u1 , . . . , ui } ⊂ {1, . . . , k},
πu = projection to the product of the factors u1 , . . . , ui .
BOUNDS ON THE RANK
3
In particular π{1,...,i} is the projection to the product of the first i factors.
For any subset u = {u1 , . . . , ui } ⊂ {1, . . . , k}, we will denote with O(u) the line
bundle on Pn1 × · · · × Pnk pull back of the hyperplane bundle on ν(πu (Y )).
We will write simply O(1) when u = {1, . . . , k}, i.e. O(1) is the pull back of the
hyperplane bundle in the Segre embedding of Pn1 × · · · × Pnk .
For any subset Z of Pn1 × · · · × Pnk we will write consequently
IZ (u) = IZ ⊗ O(u),
OZ (u) = OY ⊗ O(u)
and write IZ (1), OZ (1) when u = {1, . . . , k}.
Notice that the dimension h0 (IZ (u)) corresponds then to the co-dimension of
the linear span of ν(πu (Z)). Obviously, if Z is zero-dimensional and h1 (IZ (u)) = 0,
then also h1 (IZ (u′ )) = 0 for all u′ ⊇ u.
We will say that a finite subset S ⊂ Pn1 × · · · × Pnk has different coordinates if
for all i = 1, . . . , k the projection πi to the i-th factor is an embedding of S into
Pni .
We will need the process of residuation with respect to a divisor. Let X be a
variety For any zero-dimensional scheme Z ⊂ X and for any effective divisor D
on X the “ residue of Z w.r.t. D ” is the scheme ResD (Z) defined by the ideal
sheaf IZ : ID , where IZ , ID are the ideal sheaves of Z and D respectively. The
multiplication by local equations of D defines the exact sequence of sheaves:
(2)
0 → IResD (Z) (−D) → IZ → ID∩Z,D → 0
where the rightmost sheaf is the ideal sheaf of D ∩ Y in D.
We will identify elements T ∈ PM , which is the space of embedding of Y =
P × · · · × Pnk , as tensors of type (n1 + 1) × · · · × (nk + 1) (modulo scalars).
n1
Definition 2.1. A finite reduced subset S ⊂ Y = Pn1 ×· · ·×Pnk is a decomposition
of T if T ∈ hν(S)i (with an abuse of notation sometime we will say that also
ν(S) ⊂ X is a decomposition of T ). If moreover T ∈
/ hν(S ′ )i for any S ′ ( S,
the decomposition S is said to be not-redundant. Finally, if ♯S = min{♯S ′ | S ′ ⊂
Y and T ∈ hν(S ′ )i} then S is called a minimal decomposition of T .
Remark 2.2. Clearly “ not-redundant ” does not imply “ minimal ”. As we will
detail in Theorem 3.8, it is alway possible to build a non-minimal non-redundant
decomposition.
Our target is to study the identifiability of a tensor T ∈ PM (i.e. when T
has only one minimal decomposition) by means of the knowledge of the numbers
h0 (IA (u)), for all u ⊂ {1, . . . , k}, where A⊂ Y is a finite set which corresponds to
a decomposition of T .
We will use the following notions for the rank of T .
Definition 2.3. The rank, rk(T ), of T is the minimum r for which there exists a
minimal decomposition of T .
The cactus rank of T is the minimum r for which there exists a zero-dimensional
subscheme Γ ⊂ X with deg(Γ) = r and T ∈ hΓi.
Clearly:
rank of (T ) ≥ cactus rank of (T ).
4
E. BALLICO, A. BERNARDI, L. CHIANTINI, E. GUARDO
In analogy to the rank case, we will say that a zero-dimensional scheme Z ⊂ Y is a
“ minimal cactus decomposition ” of T ∈ PM if Z is of minimal degree among the
zero-dimensional schemes Z ′ ⊂ Y such that T ∈ hν(Z ′ )i.
If σr is the r-secant variety of X, then all tensors of rank r belong to σr .
For any tensor T of rank r, let S(T ) denote the set of all (reduced) finite subsets
S ∈ Y of cardinality r such that T ∈ hν(S)i. Of course for all S ∈ S(T ) the image
ν(S) is linearly independent, for otherwise T is contained in the span of a subset
of cardinality r′ < r, thus it has rank smaller than r.
A tensor T is identifiable when S(T ) is a singleton. The abstract product Y =
Pn1 × · · · × Pnk is “ generically identifiable in rank r ” if the general T ∈ σr is
identifiable.
The main tool for our analysis of the identifiability of a tensor T ∈ PM relies in
the following proposition that is an immediate consequence of [BB12, Lemma 1] if
we consider the residue exact sequence of A ∪ B cut by a linear space containing A.
Proposition 2.4. Consider linearly independent zero-dimensional schemes A, B ⊂
Y . Then the linear spans of the images ν(A), ν(B) in the Segre map satisfy
dim(hν(A)i ∩ hν(B)i) = dim(hν(A ∩ B)i + h1 (IA∪B (1)).
3. Rank
If we know a decomposition T = T1 + · · · + Tr of T in terms of tensors Ti of rank
1, in general we cannot directly conclude that r is the rank of T .
Our analysis will prove that, for small values of r, the rank of T is r provided
that the summands correspond to points in a suitably general geometric position.
We will give a criterion to find a lower bound for the cactus rank (and therefore
also for the rank).
Theorem 3.1. Fix a partition E ⊔ F = {1, . . . , k} of the k factors of the abstract
product Y = Pn1 ×· · ·×Pnk , i.e. E = {a1 , . . . , ak−h } and F = {b1 , . . . , bh } for some
Qh
fixed 0 < h < k. Let MF := i=1 (nbi + 1) be the affine dimension of the ambient
space of the Segre embedding of the F -factors YF := Pb1 × · · · × Pbh . Now fix 0 <
c < MF and let A ⊂ Y be a zero-dimensional scheme which satisfies h1 (IA (E)) = 0
and h0 (IA (F )) < MF − c. Take any T ∈ hν(A)i such that T ∈
/ hν(A′ )i for any
′
A ( A. Then there are no zero-dimensional schemes B ⊂ Y with deg(B) ≤ c such
that T ∈ hν(B)i.
Proof. Notice that the Segre embedding of the projection πF maps Y to PMF −1 .
Since h1 (IA (E)) = 0, we have h1 (IA (1)) = 0. The condition h0 (IA (F )) < MF −c
implies that deg(A) > c. Assume that the theorem fails and take B ⊂ Y with
deg(B) ≤ c and T ∈ hν(B)i. Since deg(A) > deg(B) and T ∈
/ hν(A′ )i for any
′
1
A ( A, we have B * A. Moreover h (IA∪B (1)) > 0, by Proposition 2.4. More
precisely, since T ∈ hν(B)i and T ∈
/ hν(A′ )i for any A′ ( A, we have
(3)
h1 (IA∪B (1)) > h1 (IA′ ∪B (1))
for all A′ ⊂ A with A ∪ B 6= A′ ∪ B.
Since deg(B) ≤ c < MF , we have h0 (IB (F )) > 0. Take a general divisor
D ∈ |IB (F )|. In other words, D is the inverse image in Y of a hyperplane in the
Segre embedding of YF and B ⊂ D. Since h0 (IA (F )) < MF − c ≤ h0 (IB (F )),
then A * D, so that (D ∩ A) ∪ B is strictly contained in A ∪ B. Hence by (3) we
BOUNDS ON THE RANK
5
get h1 (I(D∩A)∪B (1)) < h1 (IA∪B (1)). The residual exact sequence (2) applied to D
gives h1 (IResD (A∪B) (E)) > 0. Since ResD (A ∪ B) ⊆ A, we get a contradiction.
Observe that the condition h1 (IA (E)) = 0 can be satisfied only when deg(A) ≤
Qk−h
i=1 (nai + 1), the affine dimension of the ambient space of the Segre embedding
of the E-factors YE = Pa1 × · · · × Pak−h .
We can rephrase Theorem 3.1 to produce results on the rank of tensors.
Corollary 3.2. With the previous notation, if T sits in the linear span of a scheme
A ⊂ Y which satisfies the assumptions of Theorem 3.1, then the cactus rank of T
is a least c + 1. Hence also the rank of T cannot be smaller than c + 1.
Corollary 3.3. Let A ⊂ Y be a zero dimensional scheme of deg(A) = c + 1 (resp.
a finite set with ♯(A) = c + 1). With the Notation of Theorem 3.1, assume that
h1 (IA (E)) = 0 and h1 (IA (F )) = 0. Then any T ∈ hν(A)i such that T ∈
/ hν(A′ )i
for any A′ ( A has cactus rank (resp. cactus rank and rank) equal to c + 1.
Proof. It is straightforward from Theorem 3.1
Let we point out an application to the case of 3-way tensors.
Proposition 3.4. Consider k = 3 and let T be a tensor of type (n1 +1)×(n2 +1)×
(n3 + 1), which has a not-redundant decomposition T = T1 + · · · + Tr , where the Ti ’s
are tensors of rank 1. Identify each Ti with a point in X = ν(Pn1 ×Pn2 ×Pn3 ) and set
A = {T1 , . . . , Tr } ⊂ X. Call AE (resp. AF ) the projection of A to YE = Pn1 × Pn2
(resp. YF = Pn3 ).
Assume that A has different coordinates, AE is linearly independent and AF is
contained in no hyperplanes of Pn3 . Then the rank of T is at least n3 + 1.
Proof. In the notation of Theorem 3.1, take E = {1, 2} and F = {3}. Then our
assumptions on AE , AF imply that A satisfies h1 (IA (E)) = h0 (IA (F )) = 0. Thus
T cannot have a decomposition with MF − 1 = n3 summands.
Example 3.5. Kruskal’s Theorem for the identifiability of tensors ([K77]) provides
results similar to the previous proposition for the rank. We notice that the numerical
range of application of Proposition 3.4 is sometimes wider than Kruskal’s range.
For instance, consider a tensor T of type 3 × 4 × 6 having a decomposition with 6
summands. If the decomposition determines a subset A satisfying the assumptions
of Proposition 3.4, we can conclude that the rank of T is 6. We cannot get the
same conclusion directly with Kruskal’s Theorem because we are outside Kruskal’s
numerical range, since 6 > (3 + 4 + 6 − 2)/2.
The following example should be considered as a step towards the Strassen’s
Conjecture on the rank of a block tensor (see [S73]).
Example 3.6. Proposition 3.4 can give results on the rank of a sum of tensors,
when we have some information on the decompositions of the summands.
For instance, consider again tensors of type 3 × 4 × 6 and take a tensor T which
is the sum T = T ′ + T ′′ of two tensors of rank 3. Consider a decomposition of T ′
(resp. T ′′ ) in a sum of three tensors of rank 1 and call S ′ (resp. S ′′ ) the set of
cardinality 3 in the product P2 × P3 × P5 determined by the decomposition.
If the set S = S ′ ∪ S ′′ has cardinality 6 and satisfies the assumptions of Proposition 3.4 (i.e. both π{1,2} (ν(S)) and π{3} (ν(S)) are linearly independent), then we
can conclude that the rank of T is 6.
6
E. BALLICO, A. BERNARDI, L. CHIANTINI, E. GUARDO
We show below that, on the contrary, if we increase the cardinality of a decomposition, we can always construct new non-redundant decompositions of a tensor
T.
Example 3.7. Fix P ∈ Y and write P = (p1 , . . . , pk ) with pi ∈ Pni . Assume
ni > 0. Take two points bi , ci ∈ Pni such that pi 6= bi , ci but pi is contained in
the line of Pni spanned by bi and ci . Let Oi := (u1 , . . . , uk ), Qi := (v1 , . . . , vk )
be the points of Y with uj = vj for all j 6= i, ui = bi and vi = ci . We have
ν(P ) ∈ h{ν(Oi ), ν(Qi )}i, and of course ν(P ) ∈
/ hν(S ′ )i for any S ′ ( {Oi , Qi }.
We show that indeed the previous construction can yield a non-redundant decomposition. Moreover the following also show that having found a non-redundant
decomposition does not imply that it is a minimal one.
Theorem 3.8. Assume ni > 0 for at least one i. Take a finite set A ⊂ X of
cardinality r ≤ M , such that ν(A) ⊂ X is linearly independent. Take a general
T ∈ hν(A)i. Then there exists a non-redundant decomposition S ⊂ Y of T of
cardinality r + 1.
Proof. Fix P = (p1 , . . . , pk ) ∈ A and take Q1 , O1 as in Example 3.7, with i = 1,
and with the additional condition that O1 , Q1 ∈
/ (A \ {P }). We may take O1 to be
a general point of Pn1 × {p2 } × · · · × {pk }. Hence we may take O1 = (a1 , p2 , . . . , pk ),
with a1 general. Set A′ := A \ {P }.
(a) Assume ν(O1 ) ∈
/ hν(A)i. This is always possible unless hν(A)i contains
/ hν(A)i.
Pn1 × {p2 } × · · · × {pk }. Since P ∈ A, this is equivalent to ν(Q1 ) ∈
Set S := (A \ {P }) ∪ {O1 , Q1 }. We have ♯(S) = r + 1 and hν(S)i ⊇ hν(A)i
so that T ∈ hν(S)i. The set S does not depend on Q1 , but only on A and P .
To prove that S satisfies the claim, it is sufficient to prove that for a general
T ∈ hν(A)i there is no S ′ ( S with T ∈ hν(S ′ )i. It is sufficient to test all subsets
of S with cardinality r. Take S ′ ⊂ S with ♯(S ′ ) = r. If S ′ ⊃ A′ , i.e. if either
S ′ = A′ ∪ {O1 } or S ′ = A′ ∪ {Q1 }, then hν(S ′ )i ∩ hν(A)i = hν(A′ )i, because
{ν(O1 ), ν(Q1 )} ∩ hν(A)i = ∅. Thus T cannot stay in hν(S ′ )i because it cannot
stay in hν(A′ )i for a proper subset A′ of A. Hence S ′ + A′ . Since ♯(S ′ ) = r, we
have S ′ = A′′ ∪ {O1 , Q1 } with A′′ ⊂ A′ and ♯(A′′ ) = r − 2. If T ∈ hν(S ′ )i, then
hν(S ′ )i ∩ hν(A)i ⊇ hν(A′′ ) ∪ {T }i. Since ν(O1 ) ∈
/ hν(A)i, we get:
hν(S ′ )i ∩ hν(A)i = hν(A′′ ) ∪ {T }i.
The left hand side of this equality does not depends on the choice of T . Varying
T ∈ hν(A)i\hν(A′ )i we get hν(S ′ )i ⊇ hν(A)i. Since ♯(S ′ ) = ♯(A) and ν(A) is linearly
independent, we get hν(S ′ )i = hν(A)i and hence ν(O1 ) ∈ hν(A)i, a contradiction.
(b) If nj = 0 for all j > 1 we are done. Assume for instance that n2 > 0 and
/ hν(A)i.
that hν(A)i ⊃ Pn1 × {p2 } × · · · × {pk }, so that we cannot take ν(O1 ) ∈
Since ν(P ) ∈ h{ν(O1 ), ν(Q1 )}i, we get hν({O1 } ∪ A′ )i = hν(A)i. Replace thus A
with A′ ∪ {O1 } and P by O1 . Then take the construction of Example 3.7 with
i = 2. If the new general point O2 satisfies ν(O2 ) ∈
/ hν(A)i, then we conclude as in
step (a). Otherwise we have hν(A)i ⊃ Pn1 × Pn2 × {p3 } × · · · × {pk } Then continue
with i = 3 in the construction of Example 3.7, and so on. After at most k steps we
must conclude by step (a), because hAi cannot contain the whole product Y , since
♯(A) ≤ M .
Remark 3.9. Take P = (p1 , . . . , pk ), i ∈ {1, . . . , k} with ni > 0 and bi , ci as
in Example 3.7. Notice that in the previous proof we can choose for ci any point
BOUNDS ON THE RANK
7
(different from bi ) in the line spanned by bi , pi . Thus, in Theorem 3.8 we get a
positive-dimensional family of sets S ⊂ Y such that ♯(S) = r + 1, T ∈ hν(S)i and
T ∈
/ hν(S ′ )i for any S ′ ( S.
We end this section with a discussion on some consequence of Theorem 3.1 on
symmetric tensors.
Assume n := n1 = · · · = nk so that Y is a product of k copies of Pn . The
Segre map restricted to the diagonal ∆⊂ Y can be identified with the Veronese
D
map vk : Pn → PD where D + 1 = k+n
parameterizes symmetric
n . The space P
tensors T , which in turn can be identified with homogeneous polynomials (forms)
of degree k in n + 1 variables. The symmetric rank is the minimum r for which
there exists a finite subset A ⊂ Pn of cardinality r with T ∈ hvk (A)i.
A conjecture raised in [CGLM08] and well known as Comon’s Conjecture predicts
that the symmetric rank of a symmetric tensor T always coincides with the rank
of T as a normal tensor in the span of ν(Y ).
The following corollary of Theorem 3.1 implies that, if some assumptions on a
minimal symmetric decomposition A of a symmetric tensor T are satisfied, that the
symmetric rank of T coincides with its rank (and cactus rank).
Corollary 3.10. With the previous notation, consider a zero-dimensional scheme
A ⊂ Pn of degree deg(A) = c + 1. Call JA the ideal sheaf of A in Pn and assume
that h1 (JA (e)) = 0 for some e ≤ k/2. Take T ∈ hvk (A)i such that T ∈
/ hvk (A′ )i
′
for any A ( A. Then T has cactus rank equal to c + 1. If A is reduced (i.e. it is
a finite set of points), then T has also rank c + 1.
Proof. Consider any subset E ⊂ {1, . . . , k} of cardinality e and take F = {1, . . . , k}\
E. Notice that f := ♯F ≥ e, so that also h1 (JA (f )) = 0. This implies that,
considering A as a subset of ∆ ⊂ Y , in the notation of Theorem 3.1, we have
h1 (IA (E)) = 0 and h0 (IA (F )) < MF − c. The claim follows from Theorem 3.1.
We show that Corollary 3.10 provides new evidences for the Comon’s Conjecture,
sometimes even in a numerical range larger than the ones considered in previous
works on the topic (e.g. [F16] and [ZHQ16]).
Remark 3.11. Assume k = 2e even. Then by Corollary 3.10, the Comon’s conjecture holds for symmetric tensors T having a minimal symmetric decomposition
A with h1 (JA (e)) = 0.
The condition h1 (JA (e)) = 0 holds for general subsets A ⊂ Pn , as soon as
c + 1 = ♯A satisfies
n+e
(4)
c + 1 ≤ r0 :=
.
e
So Comon’s Conjecture holds for general tensors whose symmetric rank c + 1 is
bounded by r0 .
When k = 2e + 1 is odd, a similar conclusion holds with r0 := 1 + n+e
e .
Example 3.12. After the Alexander-Hirshowitz classification of defective Veronese
varieties ([AH95]), a general symmetric tensor in the span of vk (Pn ) is known to
have symmetric rank
rg = ⌈
except for a list of few exceptional cases.
n+k
k
n+1
⌉
8
E. BALLICO, A. BERNARDI, L. CHIANTINI, E. GUARDO
In general rg is bigger than our bound r0 of Remark 3.10, since, for fixed e, rg
grows asymptotically as (2n en )/(n + 1)! while r0 grows like en /n!.
Nevertheless, there are cases in which rg and r0 coincide. This happens for
(k, n) = (6, 2) or (k, n) = (8, 2). Since these two cases are not in the list of exceptional Veronese varieties, we can conclude that Comon’s Conjecture holds for
general forms of degree 6 and degree 8 in 3 variables.
Example 3.13. Take (k, n) = (4, 3). Then Remark 3.11 tells us that the Comon’s
conjecture holds for general symmetric tensors in the span of v4 (P3 ) with a decomposition with 10 summands. On the other hand, in this case the number rg is 9,
smaller than r0 .
Indeed (k, n) = (4, 3) is in the list of exceptional cases in the Alexander-Hirschowitz theorem, so that a general form of degree 4 in P3 has symmetric rank 10 > rg .
Thus Comon’s conjecture also holds for such general forms.
In some sense, Theorem 3.1 provides a new heuristic reason why the case (k, n) =
(4, 3) is exceptional: a general tensor with a non-redundant decomposition with 10
summands cannot have a decomposition with 9 summands. A similar remark holds
for the other exceptional cases of even degree: quadrics in any Pn and quartics in
P2 and P4 .
4. Identifiability
In order to get results on the identifiability of a tensor T , we need to refine the
previous analysis, and we are going to do that in this section.
We will need the following terminology for the Segre function of a finite subset
of the product, introduced in [CS16].
Definition 4.1. For any set of points S ⊂ Y , the Segre function SFS : {1, . . . , k} →
N is defined by:
SFS (i) = 1+ the dimension of the linear span of ν(π{1,...,i} (S)).
Remark that the knowledge of the sequence h0 (IS ({1, . . . , i})), i = 1, . . . , k, is
equivalent to the knowledge of the Segre function SFS .
More precisely, the definition of Segre function depends on the ordering of the
factors of the product. The knowledge of h0 (IS (u)), for all possible u ⊂ {1, . . . , k},
is equivalent to the knowledge of the Segre functions of S under all possible rearrangements of the factors.
Let us recall the following definition for a minimal dependent set of point (the
same can be found in [CS16, Definition 2.9], while where in [GKZ, Chap.7 Sec.1] a
minimal dependent set of points is called a circuit.).
Definition 4.2. A set of points S ⊂ Pm is minimally dependent, if S is linearly
dependent, but any proper subset of S is linearly independent.
We need now a list of results on the cohomology of IS , for finite sets S in a
product of projective spaces.
Definition 4.3. We say that a finite subset S ⊂ Y is degenerate if there exists an
index i such that πi (S) does not span Pni , i.e. there is a hyperplane H ⊂ Pni such
that
S ⊂ Pn1 × · · · × Pni−1 × H × Pni+1 × · · · × Pnk .
BOUNDS ON THE RANK
9
We say that T ∈ PM is degenerate if there exists a degenerate subset A ⊂ ν(Y )
such that T ∈ hAi.
Notice that if S ⊂ Y is non-degenerate, then necessarily ♯S > max{ni }.
The two results below are mainly an extension to the non-symmetric case of
results in [BGL13] and [BB12].
Lemma 4.4. Fix a finite set S ⊂ Y = Pn1 × · · · × Pnk such that x := ♯(S) ≤ k + 1,
h1 (IS (1)) > 0 and h1 (IS ′ (1)) = 0 for each S ′ ( S. Then there is E ⊂ {1, . . . , k}
such that ♯(E) = k + 2 − x and ♯(πi (S)) = 1 for all i ∈ E.
Proof. The lemma is trivial if x = 2, because OY (1) is very ample and so the
assumptions of the lemma are never satisfied in this case. Thus we may assume
x > 2 and use induction on the integer x.
The case x = 3 is also true, because ν(Y ) is cut out by quadrics and each line
contained in ν(Y ) is the image by ν of a line in one of the k factors of Y .
Thus we assume x > 3 (and so k ≥ 3).
Assume ♯(π1 (S)) ≥ 2, so that the Segre function of S satisfies SFS (1) ≥ 2. Notice
that the assumptions on S imply that SFS (k) = x−1. Take points P = (a1 , . . . , ak ),
Q = (b1 , . . . , bk ) ∈ S with P 6= Q. Since h1 (IS ′ (1)) = 0 for all S ′ ( S and x > 3,
we may find A ⊃ {P, Q} with ♯(A) = x − 1 and h1 (IA (1)) = 0. Since x ≤ k + 1,
there is a minimal integer i ∈ {2, . . . , k} such that SFS (i − 1) = SFS (i). By
[CS16, Proposition 2.5] for every minimally dependent S ′ ⊆ S with respect to the
line bundle O({1, . . . , i}) we have ♯(πi (S ′ )) = 1. Since i ≥ 2 and P 6= Q, we
have h1 (I{P,Q} ({1, . . . , i})) = 0. Hence we may find a minimally dependent set
containing {P, Q}. Thus ai = bi . Take any C = (c1 , . . . , ck ) ∈ S \ {P, Q}. If
c1 6= a1 we may take minimally dependent S ′′ ⊆ S with respect to the line bundle
O({1, . . . , i − 1}) containing {P, C} and hence ci = ai . If c1 = a1 we may take
minimally dependent S ′′ ⊆ S with respect to the line bundle O({1, . . . , i − 1})
containing {Q, C} and hence ci = bi = ai . Thus πi (S) = {ai }. The lemma is now
proved when x = k + 1, by setting E = {i} .
If x < k + 1, we apply the same proof to the projection πu (X) where u =
{1, . . . , k} \ {i} and conclude by descending induction on the integer k + 1 − x.
Lemma 4.5. Fix a finite set S ⊂ Y such that x := ♯(S) ≤ k + n1 , hπ1 (S)i = Pn1 ,
h1 (IS (1)) > 0 and h1 (IS ′ (1)) = 0 for each S ′ ( S. Then there is E ⊂ {2, . . . , k}
such that ♯(E) = k + 2 − x and ♯(πi (S)) = 1 for all i ∈ E.
Proof. Take n1 + 1 points A1 , . . . , Ani +1 , say Aj = (aj,1 , . . . , aj,k ), 1 ≤ j ≤ n1 + 1,
with the property that the set {a1,1 , . . . , an1 +1,1 } spans Pn1 . Then just repeat the
proof of Lemma 4.4.
Now, with the aid of Proposition 2.4, we are ready to prove the following theorem
where we explicit a sufficient condition for a non-redundant decomposition to be
minimal and unique.
Theorem 4.6. Fix an element T ∈ PM and a non-redundant decomposition S ⊂ Y
of T and let ♯(S) = r. If m := max{n1 , . . . , nk }, then
(a) If 2r ≤ k + m, then the rank of T is r.
(b) If moreover 2r < k + m, then S(T ) = {S}, i.e. T is identifiable.
Proof. By permuting the factors of Y we may assume that n1 = m. The assumptions on T imply in particular that ν(S) is linearly independent, i.e. h1 (IS (1)) = 0.
10
E. BALLICO, A. BERNARDI, L. CHIANTINI, E. GUARDO
Let S ′ ⊂ Y be a minimal decomposition of T such that ♯S ′ ≤ r and S ′ 6= S. Then,
by Proposition 2.4, h1 (IS∪S ′ (1)) > 0.
Let S̃ ⊆ S∪S ′ be a minimal subset of S∪S ′ containing S and with h1 (IS̃ (1)) > 0.
Since S̃ ⊇ S, and X is non-degenerate, the set π1 (S̃) spans Pn1 . By Lemma 4.5 we
have ♯(S̃) ≥ k + ni + 1 and hence ♯(S ′ ) + r ≥ k + m + 1.
If 2r = k + m we get ♯(S ′ ) ≥ r. This proves that T has rank r. If 2r < k + m
we get a contradiction.
Corollary 4.7. Set m := max{n1 , . . . , nk }. Take a non-degenerate T ∈ PM . If
2rk(T ) < k + m, then T is identifiable.
Proof. Take S ∈ S(T ). Since S ⊂ Y and T ∈ hν(S)i, then X chi e’ X? is the
minimal multiprojective space containing S. Then apply part (b) of Theorem 4.6.
Remark 4.8. One cannot give a result on the identifiability of tensors without
comparing it with the celebrated Kruskal’s bound ([K77]), which is known to be
sharp ([SB00], [D13]).
Our condition on the decomposition S of the tensor T is weaker than the condition imposed by Kruskal, which requires to compute the span of any subset of πi (S)
up to cardinality ni + 1 (in order to determine the Kruskal’s rank), while we only
need to check that πi (S) generates Pni . Since the Kruskal’s rank of the projections
of S in principle can be even 1 (when πi is not injective), for low values of the rank
our result determines the identifiability of T under weaker assumptions.
Of course, as our assumptions are weaker than Kruskal’s ones, we cannot give
applications outside Kruskal’s numerical range. There are few cases in which the
numerical range of application of our result matches with Kruskal’s range. One of
them e.g. is given by tensors of type 3 × 2 × 2 × 2.
Next, we show that under some condition on the decomposition S of a tensor T ,
we can prove that any other decomposition S ′ of cardinality smaller or equal than
♯S must have projections in special position.
Definition 4.9. A zero-dimensional scheme Z ⊂ Y is said to be curvilinear if each
connected component of Z has embedding dimension 1, i.e. there exists a smooth
curve in Y containing Z.
We will indeed prove the result even when S ′ is non-reduced, provided that S ′
is curvilinear.
Notice that, of course, any reduced finite subset of Y is curvilinear.
Theorem 4.10. Set m′ := min{n1 , . . . , nk }. Fix integers s > 0, 0 < x < k such
that (m′ + 1)k−x ≥ r. Let B ⊂ Y be zero-dimensional curvilinear scheme and
S ⊂ Y be a finite set with different coordinates such that ♯(S) = r, deg(B) = x and
h1 (IS (u)) = 0 for any subset u ⊂ {1, . . . , k} of cardinality k − x.
Assume that each projection πi is an isomorphism when restricted to B (when
B is reduced this is equivalent to say that also B has different coordinates).
Then, h1 (IS∪B (1)) = 0.
Proof. Set Z := S ∪ B and assume h1 (IZ (1)) > 0. Taking S \ (B ∩ S) instead of S
we reduce to the case S ∩ B = ∅. Fix u = {2, . . . , k} so that πu is the projection to
the last k − 1 factors and write Yu = Pn2 × · · · × Pnk
BOUNDS ON THE RANK
11
(a) First assume k = 2 and hence x = 1, u = {2} and ♯(S) ≤ m + 1. Write
B = {O} with O = (O1 , O2 ). Take a general hyperplane H ⊂ Pn1 such that
O1 ∈ H1 and set W1 := H1 × Pn2 . Since H1 is a general hyperplane containing O1
and S has different coordinates, either W1 ∩ S = ∅ or W1 ∩ S is the unique point
whose image by π1 is O1 . Hence ♯(W1 ∩ Z) ≤ 2.
Since O(1) is very ample, then h1 (IW1 ∩Z (1)) = 0, so h1 (W1 , IW1 ∩Z,W1 (1)) =
0. The residual exact sequence of W1 in Y gives h1 (IZ\(Z∩W1 ) (u)) > 0. Since
Z \ (Z ∩ W1 ) ⊆ S, we get h1 (IS (u)) > 0, a contradiction.
(b) Now assume k ≥ 3 and assume that the claim holds for multiprojective
spaces with k − 1 factors. Write {P (1), . . . , P (y)} for the points of the reduced
set Bred , where 1 ≤ y ≤ x. For any i ∈ {1, . . . , k} let Ai be the set of all pairs
(P (u), Q(u)) ∈ Bred × S such that all the coordinates of P (u) and Q(u) are the
same, except the i-th one (which is different, because S ∩ B = ∅). Assume the
existence of (P (u), Q(v)) ∈ Ai and (P (u), Q(v)) ∈ Aj . Since S ∩ B = ∅, Q(u)
(resp. Q(v)) has all coordinates equal to the one of P (u), except the i-th (resp.
j-th) one, which is different. Since k ≥ 3 and S has different coordinates, we get
Q(u) = Q(v). Hence i = j. Since y ≤ x < k, there is h ∈ {1, . . . , k} with Ah = ∅.
Permuting the factors of Y we may assume h = 1, i.e. A1 = ∅. This is equivalent
to the injectivity of the map π1|S∪Bred . Since π1|B is an embedding, we get that
π1|Z is an embedding.
Fix P ∈ Bred and call P1 , . . . , Pk its components. Take a general hyperplane
H1 of Pn1 containing P1 . Since π1 (B) is curvilinear and H1 is general, we have
π1 (B) ∩ H1 = {P1 } (scheme-theoretic intersection). Set W1 := H1 × Yu . W1 is an
element of |O({1})|. Since π1|B is an embedding and π1 (B) ∩ H1 = {P1 } (schemetheoretic intersection), we have B ∩ W1 = {P } (scheme-theoretic intersection).
Since H1 is general and S has different coordinates, either W1 ∩ S = ∅ or W1 ∩ S
is the unique point of S with P1 as its first coordinate. Hence Z1 := Z ∩ W1
is always reduced, it contains P and at most another point, which is in S. Set
Zu := ResW1 (Z), B1 := ResW1 (B) and S1 := ResW1 (S) = S \ (S ∩ W1 ). We
have Zu = B1 ∪ S1 , deg(B1 ) = x − 1, B1 ⊂ B, S1 ⊆ S. Since ♯(Z1 ) ≤ 2 and
O(1) is very ample, we have h1 (IZ1 (1)) = 0. Hence h1 (W1 , IZ1 (1)) = 0. The
residual exact sequence of W1 in X gives h1 (IZu (u)) > 0. Since πu 1|Z, is an
embedding, then πu|Zu is an embedding. Hence h1 (IZu (u)) = h1 (Yu , IZu (1)). The
inductive assumption on the number of factors of the multiprojective space gives
h1 (Yu , IZu (1)) = 0, a contradiction.
Mixing the previous theorem with Proposition 2.4, we get the following.
Corollary 4.11. Set m′ := min{n1 , . . . , nk } and fix integers s > 0, 0 < x < k such
that (m′ + 1)k−x ≥ r. Take a tensor T ∈ PM with a non-redundant decomposition
S ⊂ Y such that ♯S = r and S has different coordinates. Assume h1 (IS (u)) = 0
for any subset u ⊂ {1, . . . , k} of cardinality k − x.
Then any other non-redundant decomposition S ′ ⊂ X of T of cardinality ≤ x
cannot have different coordinates.
The statement of Theorem 3.1 cannot produce corollaries on the identifiability,
because the assumptions do not include the case ♯A = ♯B.
We can however modify the proof of Theorem 3.1, adding some assumptions on
the decomposition S, which produces results which bound different the decompositions of a tensor T .
12
E. BALLICO, A. BERNARDI, L. CHIANTINI, E. GUARDO
Definition 4.12. Let W ⊂ Pr be an integral and non-degenerate projective variety.
A finite set (resp. zero-dimensional scheme) S ⊂ W is said to be set-theoretically
quasi-general (resp. scheme-theoretically quasi-general) if the set-theoretic (resp.
scheme-theoretic) intersection W ∩hSi is set-theoretically (resp. scheme-theoretical)
equal to S.
Lemma 4.13. Let W ⊂ PM be an integral and non-degenerate variety of dimension
n. Fix a general reduced subset S ⊂ W with ♯S ≤ M − n− 1. Then S is the schemetheoretic base locus of the linear system on Y chi e’ Y? induced by H 0 (IS (1)), i.e
S is scheme-theoretically quasi-general.
Proof. Let L ⊂ PM be a general linear space of dimension M − r − 1. By Bertini’s
theorem the scheme L ∩ W is a finite set of deg(Y ) points, moreover the set L ∩ W
is in linearly general position in L. Hence for any S ⊂ L ∩ W with ♯S ≤ M − r − 1,
the restriction of U := H 0 (IS (1)) to L ∩ W has S as its set-theoretic base locus.
Since L is a linear space, the restriction of U to W has base locus contained in
L ∩ W . Thus S is the base locus of the restriction of U to W . Since W is integral
and non-degenerate, a general subset A ⊂ W with cardinality at most M − n spans
a general subspace of PM with dimension ♯(A) − 1. The claim follows.
Proposition 4.14. Fix a partition E ⊔ F = {1, . . . , k} with E = {a1 , . . . , an−h }
and F = {b1 , . . . , bh } for some 0 < h < k, and a positive integer c < MF =
Qh
i=1 (nbi + 1). Let Z ⊂ Y be a zero-dimensional scheme such that ♯(Z) = c,
h1 (IZ (E)) = 0 and h1 (IZ (F )) = 0. Assume that πF (Z) is set-theoretically (resp.
scheme-theoretically) quasi-general. Take any T ∈ hν(Z)i such that T ∈
/ hν(Z ′ )i
′
for any Z ( Z.
If S is a finite set (resp. zero-dimensional scheme) such that S 6= Z, deg(S) ≤ b
and T ∈ hν(S)i, then deg(S) = b and πF (S) = πF (Z).
Proof. By Theorem 3.1 it is sufficient to do the case deg(S) = c and h0 (IS (F )) =
h0 (OY (F )) − b. In particular πF induces an embedding of S into YF == Pb1 × · · · ×
Pbk .
The proof of Theorem 3.1 works verbatim, if there is H ∈ |OY (F )| containing S,
but not containing Z. Since h0 (IS (F )) = h0 (IZ (F )), this is equivalent to require
that H 0 (IS (F )) 6= H 0 (IZ (F )), i.e. that πF (S) is not contained in the base locus
of πF (Z). Since πF (Z) is the set-theoretic (resp. scheme-theoretic) quasi-general,
the base locus of |OYF (1)| is πF (Z). Thus we get πF (S) = πF (Z).
Corollary 4.15. For each i ∈ {1, . . . , k} fix a set Fi ⊂ {1, . . . , k} such that i ∈ Fi
and set Ei := {1, . . . , k} \ Fi . Let S ⊂ Y be a finite set such that ♯(S) = r, where:
Y
Y
r<
(nh + 1).
(nj + 1) and r ≤
j∈Fi
h∈Ei
Take any T ∈ hν(S)i, such that T ∈
/ hν(S ′ )i for any S ′ ( S. Assume that:
h0 (IS (Fi )) = h0 (OY (Fi )) − r
and
h0 (IS (Ei )) = h0 (OY (Ei )) − r
for all i = {1, . . . , k}. Assume moreover that each πFi (S) is set-theoretically (resp.
scheme-theoretical) quasi-general.
If S ′ 6= S is a zero-dimensional subscheme of degree ≤ r such that hν(S ′ )i
contains T , then S ′ is a finite set, ♯(S ′ ) = r and πi (S ′ ) = πi (S) for all i.
BOUNDS ON THE RANK
13
Notice that (unfortunately) we are not able to conclude that S = S ′ : they can
be different even if πi (S ′ ) = πi (S) for all i. Namely the points can differ by a
rearrangement of the coordinates.
Proof. Since h0 (IS (Fi )) = h0 (OY (Fi )) − ♯(S), each πFi |S is injective. Assume that
S ′ exists. By Proposition 4.14 applied to Fi , S ′ is a finite set with ♯(S ′ ) = r, πFi |S ′
injective and πFi (S ′ ) = πFi (S). Thus πi (S ′ ) = πi (S) for all i.
Remark 4.16. Corollary 4.15 does not provide the identifiability of a tensor T : it
simply bounds strictly the locus where different decompositions of the same tensor
T could lie. We observe that, on the other hand, the numerical range of application
of Corollary 4.15 is wider than the range of Kruskal’s criterion of identifiability.
Just to give an example, consider tensors of type 3 × 3 × 6, corresponding (mod
scalars) to points in the space P53 which contains the Segre embedding of P2 ×P2 ×P5 .
Kruskal’s criterion for the identifiability applies only when the rank r is bounded
by r ≤ (3 + 3 + 6 − 2)/2 = 5. Our Corollary 4.15 applies, taking F1 = F2 = {1, 2},
F3 = {3} and checking the geometric assumptions, even for r = 6.
References
[AH95]
Alexander J. and A. Hirschowitz A., Polynomial interpolation in several variables, J.
Algebraic Geom., 4 (1995), 201222.
[BB12]
Ballico E. and Bernardi A., Decomposition of homogeneous polynomials with low rank,
Math. Z. 271 (2012), 1141–1149.
[BGI11]
Bernardi A., Gimigliano A. and Idà M., Computing symmetric rank for symmetric
tensors, J. Symbolic. Comput. 46 (2011), 34–55.
[BGL13]
Buczyński J., Ginenski A. and Landsberg J. M., Determinantal equations for secant
varieties and the Eisenbud-Koh-Stillman conjecture, J. London Math. Soc. 88 (2013),
1–24.
[COV17] Chiantini L., Ottaviani G. and Vannieuwenhoven N., Effective criteria for specific identifiability of tensors and forms, SIAM J. Matrix Anal. Applic. to appear,
(arXiv:1609.00123).
[CS16]
Chiantini L. and Sacchi D., Segre functions in Multiprojective Spaces and Tensor
Analysis, in: G. Casnati et al., From Classical to Modern Algebraic Geometry, Corrado Segre’s Mastership and Legacy, Springer - Trends in History and Science (2016),
361–374.
[CS11]
Comas G, M. Seiguer M., On the rank of a binary form, Found Comput Math 11
(2011), 65–78.
[CGLM08] Comon P., Golub G., Lim L. H., and Mourrain B., Symmetric tensors and symmetric
tensor rank, SIAM J. Matrix Anal. Applic., 30 (2008), 1254–1279.
[D13]
Dersken H., Kruskal’s uniqueness inequality is sharp, Lin. Alg. Applic., 438 (2013),
708–712.
[F16]
Friedland S., Remarks on the symmetric rank of symmetric tensors, SIAM J. Matrix
Anal. Applic., 37 (2016), 320–337.
[GKZ]
Gelfand I. M., M. Kapranov M. and Zelevinsky A, Discriminants, Resultants, and
Multidimensional Determinants. Birkäuser, Zurich (1994).
[K77]
Kruskal J. B., Three-way arrays: rank and uniqueness of trilinear decompositions,
with applications to arithmetic complexity and statistics, Lin. Alg. Applic. 18 (1977),
95–138.
[SB00]
Sidiropoulos N. D. and Bro R., On the uniqueness of multilinear decomposition of
N-way arrays, J. Chemometrics 14 (2000), 229–239.
[S73]
Strassen V., Vermeidung von Divisionen, J. Reine Angew. Math. 264 (1973), 184-202.
[ZHQ16] Zhang X., Huang Z. H., and Qi L., Comon’s conjecture, rank decomposition, and
symmetric rank decomposition of symmetric tensors, SIAM J. Matrix Anal. Applic.,
37 (2016), 1719–1728.
14
E. BALLICO, A. BERNARDI, L. CHIANTINI, E. GUARDO
(E. Ballico, A. Bernardi) Dipartimento di Matematica, Università di Trento
E-mail address: [email protected], [email protected]
(L. Chiantini) Dipartimento di Ingegneria dell’Informazione e Scienze Matematiche,
Università di Siena
E-mail address: [email protected]
(E. Guardo) Dipartimento di Matematica e Informatica, Università di Catania
E-mail address: [email protected]
| 0math.AC
|
TORIC IDEALS AND THEIR CIRCUITS
arXiv:1210.3204v1 [math.AC] 11 Oct 2012
HIDEFUMI OHSUGI AND TAKAYUKI HIBI
Abstract. In this paper, we study toric ideals generated by circuits. For toric
ideals which have squarefree quadratic initial ideals, a sufficient condition to be
generated by circuits is given. In particular, squarefree Veronese subrings, the
second Veronese subrings and configurations arising from root systems satisfy the
condition. In addition, we study toric ideals of finite graphs and characterize the
graphs whose toric ideals are generated by circuits u − v such that either u or v
is squarefree. There exists several classes of graphs whose toric ideals satisfy this
condition and whose toric rings are nonnormal.
Introduction
Let Zd×n be the set of all d × n integer matrices. A configuration of Rd is a
matrix A ∈ Zd×n , for which there exists a hyperplane H ⊂ Rd not passing the
origin of Rd such that each column vector of A lies on H. Throughout this paper,
we assume that the columns of A are pairwise distinct. Let K be a field and
−1
K[T, T −1 ] = K[t1 , t−1
1 , . . . , td , td ] the Laurent polynomial ring in d variables over
K. Each column vector a = (a1 , . . . , ad )⊤ ∈ Zd (= Zd×1 ), where (a1 , . . . , ad )⊤ is
the transpose of (a1 , . . . , ad ), yields the Laurent monomial T a = ta11 · · · tadd . Let
A ∈ Zd×n be a configuration of Rd with a1 , . . . , an its column vectors. The toric
ring of A is the subalgebra K[A] of K[T, T −1 ] which is generated by the Laurent
monomials T a1 , . . . , T an over K. Let K[X] = K[x1 , . . . , xn ] be the polynomial ring in
n variables over K and define the surjective ring homomorphism π : K[X] → K[A]
by setting π(xi ) = T ai for i = 1, . . . , n. We say that the kernel IA ⊂ K[X] of π is the
toric ideal of A. It is known that, if IA 6= {0}, then IA is generated by homogeneous
binomials of degree ≥ 2. More precisely,
D +
E
−
IA = X u − X u ∈ K[X] u ∈ KerZ (A)
where KerZ (A) = {u ∈ Zn | Au = 0}. Here u+ ∈ Zn≥0 (resp. u− ∈ Zn≥0 ) is the
positive part (resp. negative part) of u ∈ Zn . In particular, we have u = u+ − u− .
See [13] for details.
The support of a monomial u of K[X] is supp(u) = {xi | xi divides u} and the
support of a binomial f = u − v is supp(f ) = supp(u) ∪ supp(v). We say that an
irreducible binomial f ∈ IA is a circuit of IA if there is no binomial g ∈ IA such that
supp(g) ⊂ supp(f ) and supp(g) 6= supp(f ). Note that a binomial f ∈ IA is a circuit
of IA if and only if IA ∩ K[{xi | xi ∈ supp(f )}] is generated by f . Let CA be the set
This research is supported by JST CREST.
1
of circuits of IA and define its subsets CAsf and CAsfsf by
CAsf = {X u − X v ∈ CA | either X u or X v is squarefree },
CAsfsf = {X u − X v ∈ CA | both X u and X v are squarefree }.
It is known [13, Proposition 4.11] that CA ⊂ UA where UA is the union of all
reduced Gröbner bases of IA . Since any Gröbner basis is a set of generators, we have
IA = hUA i. Bogart–Jensen–Thomas [1] characterized the configuration A such that
IA = hCA i in terms of polytopes. On the other hand, Martinez-Bernal–Villarreal [5]
introduced the notion of “unbalanced circuits” and characterized the configuration
A such that IA = hCA i in terms of unbalanced circuits when K[A] is normal. Note
that, if K[A] is normal, then any binomial belonging to a minimal set of binomial
generators of IA has a squarefree monomial. (This fact appeared in many papers.
See, e.g., [11, Lemma 6.1].)
One of the most important classes of toric ideals whose circuits are well-studied
is toric ideals arising from finite graphs. Let G be a finite connected graph on the
vertex set [d] = {1, 2, . . . , d} with the edge set E(G) = {e1 , . . . , en }. Let e1 , . . . , ed
stand for the canonical unit coordinate vector of Rd . If e = {i, j} is an edge of G,
then the column vector ρ(e) ∈ Rd is defined by ρ(e) = ei + ej . Let AG ∈ Zd×n
denote the matrix with column vectors ρ(e1 ), . . . , ρ(en ). Then AG is a configuration
of Rd which is the vertex-edge incidence matrix of G. Circuits of IAG are completely
characterized in terms of graphs (Proposition 2.1). It is known that K[AG ] is normal
if and only if G satisfies “the odd cycle condition” (Proposition 2.3). In [10, Section
3], generators of IAG are studied when K[AG ] is normal. It is essentially shown
in [10, Proof of Lemma 3.2] that, if K[AG ] is normal, then we have IAG = hCAsfG i.
Martinez-Bernal–Villarreal [5, Theorem 3.2] also proved this fact and claimed that
the converse is true. However, as they stated in [5, Note added in proof], the converse
is false in general. Several classes of counterexamples are given in Section 2.
The content of this paper is as follows. In Section 1, we study toric ideals having
squarefree quadratic initial ideals. For such configurations, a sufficient condition to
be generated by circuits is given. In particular, squarefree Veronese subrings, the
second Veronese subrings and configurations arising from root systems satisfy the
condition. In Section 2, we study toric ideals of finite graphs. We characterize the
graphs G whose toric ideals are generated by CAsfG . A similar result is given for
CAsfsf
. By this characterization, we construct classes of graphs G such that K[AG ]
G
is nonnormal and that IAG = hCAsfG i = hCAsfsf
i.
G
1. Configurations with squarefree quadratic initial ideals
In this section, we study several classes of toric ideals with squarefree quadratic
initial ideals. It is known [13, Proposition 13.15] that, if a toric ideal IA has a
squarefree initial ideal, then K[A] is normal. First, we show a fundamental fact on
quadratic binomials in toric ideals. (Since we assume the columns of A are pairwise
distinct, IA has no binomials of degree 1.)
Proposition 1.1. Let A = (aij ) ∈ Zd×n be a configuration. Suppose that, for each
1 ≤ i ≤ d, there exists zi ∈ Z such that zi − 1 ≤ aij ≤ zi + 1 for all 1 ≤ j ≤ n. Then,
2
any quadratic binomial in IA belongs to CAsf . Moreover, if, for each 1 ≤ i ≤ d, there
exists zi ∈ Z such that zi ≤ aij ≤ zi + 1 for all 1 ≤ j ≤ n, then any quadratic
binomial in IA belongs to CAsfsf .
Proof. Suppose that, for each 1 ≤ i ≤ d, there exists zi ∈ Z such that zi − 1 ≤ aij ≤
zi + 1 for all 1 ≤ j ≤ n. It is known [13, Lemma 4.14] that there exists a vector
w ∈ Rd such that w · A = (1, 1, . . . , 1). Hence, by elementary row operations, we
may assume that A is a (0, ±1)-configuration. Let A = (a1 , . . . , an ) ∈ Zd×n and let
f ∈ IA be a quadratic binomial. Since the columns of A are pairwise distinct, f is of
the form either x1 x2 − x3 x4 or x1 x2 − x23 . Note that ♯|supp(h)| ≥ 3 for any binomial
h ∈ IA . Hence f is a circuit if f = x1 x2 − x23 .
Let f = x1 x2 −x3 x4 and suppose that f ∈
/ CA . By [13, Lemma 4.10], there exists a
circuit g = X u −X v ∈ CA such that supp(X u ) ⊂ {x1 , x2 } and supp(X v ) ⊂ {x3 , x4 }.
Since f is not a circuit, ♯|supp(g)| < 4. Hence we have ♯|supp(g)| = 3. Thus we
may assume that g = xa1 xb2 − xc3 where 1 ≤ a, b, c ∈ Z. Then, a · a1 + b · a2 = c · a3
(k) (k)
(k)
and a + b = c. Let ak = (a1 , a2 , . . . , ad )⊤ for k = 1, 2, 3. Since a1 , a2 , a3 are
(0, ±1)-vectors, we have the following for each 1 ≤ j ≤ d:
(3)
(1)
(2)
• If aj = 1, then aj = aj = 1.
(3)
(1)
(2)
• If aj = −1, then aj = aj = −1.
(3)
(1)
Since a1 and a3 are distinct, there exists 1 ≤ k ≤ d such that ak = 0 and ak 6= 0.
(1)
(2)
(2)
(1)
Then a · ak + b · ak = 0 and hence a = b and ak = −ak . Note that g = xa1 xa2 − x2a
3
should be irreducible. It then follows that a = 1 and g = x1 x2 − x23 . Thus f − g =
x23 − x3 x4 belongs to IA and hence a3 = a4 , a contradiction. Therefore f ∈ CA .
Suppose that, for each 1 ≤ i ≤ d, there exists zi ∈ Z such that zi ≤ aij ≤ zi + 1
for all 1 ≤ j ≤ n. By elementary row operations, we may assume that A is a (0, 1)configuration. Let f = x1 x2 − x23 ∈ IA . Then, a1 + a2 = 2 · a3 . Since a1 , a2 , a3 are
(0,1)-vectors, it follows that a1 = a2 = a3 , a contradiction.
By Proposition 1.1, we can prove that several classes of toric ideals are generated
by circuits.
1.1. Veronese and squarefree Veronese configurations. Let 2 ≤ d, r ∈ Z and
(r)
Vd = (a1 , . . . , an ) ∈ Zd×n be the matrix where
)
(
d
X
αi = r .
{a1 , . . . , an } = (α1 , . . . , αd )⊤ ∈ Zd αi ≥ 0,
i=1
(r)
Then, K[Vd ] is called the r-th Veronese subring of K[t1 , . . . , td ]. On the other
(r)
hand, let SVd = (a1 , . . . , an ) ∈ Zd×n be the matrix where
)
(
d
X
αi = r .
{a1 , . . . , an } = (α1 , . . . , αd )⊤ ∈ {0, 1}d
i=1
(r)
K[SVd ]
Then,
is called the r-th squarefree Veronese subring of K[t1 , . . . , td ]. It is
known (e.g., [13, Chapter 14]) that
3
Proposition 1.2. Toric ideals IV (r) and ISV (r) have squarefree quadratic initial
d
(r)
d
(r)
ideals and hence K[Vd ] and K[SVd ] are normal.
We characterize such toric ideals that are generated by circuits.
Theorem 1.3. Let 2 ≤ d, r ∈ Z. Then, we have the following:
(r)
(i) For A = SVd , the toric ideal IA is generated by CAsfsf .
(r)
(ii) For A = Vd , the toric ideal IA is generated by CAsf if and only if r = 2.
Proof. First, by Propositions 1.1 and 1.2, (i) and the “if” part of (ii) holds.
(r)
Let r ≥ 3. Since K[V2 ] is a combinatorial pure subring (see [6] for details) of
(r)
K[Vd ] for all d > 2, it is sufficient to show that IV (r) is not generated by circuits.
2
r r −1 r −2 r −3 ··· 0
(r)
Recall that the configuration V2 is
. Then the
0
1
2
3
··· r
binomial x1 x4 −x2 x3 ∈ IV (r) is not a circuit since x22 −x1 x3 belongs to IV (r) . Suppose
2
2
that 0 6= x1 x4 −xi xj belongs to IV (r) . Then ai +aj = a1 +a4 = (2r −3, 3)⊤ . Since the
2
last coordinate of ai + aj is 3, it follows that {i, j} is either {1, 4} or {2, 3}. Hence
x1 x4 − xi xj = x1 x4 − x2 x3 . Thus x1 x4 − x2 x3 is not generated by other binomials in
IV (r) as desired.
2
1.2. Configurations arising from root systems. For an integer d ≥ 2, let Φ ⊂
Zd be one of the classical irreducible root systems Ad−1 , Bd , Cd and Dd ([4, pp.
64 – 65]) and write Φ(+) for the set consisting of the origin of Rd together with all
positive roots of Φ. More explicitly,
(+)
Ad−1 = {0} ∪ {ei − ej | 1 ≤ i < j ≤ d}
(+)
= Ad−1 ∪ {e1 , . . . , ed } ∪ {ei + ej | 1 ≤ i < j ≤ d}
(+)
= Ad−1 ∪ {ei + ej | 1 ≤ i ≤ j ≤ d}
(+)
= Ad−1 ∪ {ei + ej | 1 ≤ i < j ≤ d}
Bd
Cd
Dd
(+)
(+)
(+)
where ei is the ith unit coordinate vector of Rd and 0 is the origin of Rd . For each
(+)
(+)
(+)
(+)
Φ(+) ∈ {Ad−1 , Bd , Cd , Dd }, we identify Φ(+) with the matrix whose columns
are Φ(+) and associate the configuration
e (+) =
Φ
Φ(+)
···
1
1
.
Proposition 1.4 ([2], [9]). Working with the same notation as above. Then the
e (+) ] is normal.
toric ideal IΦe (+) has a squarefree quadratic initial ideal and hence K[Φ
By Proposition 1.1, we have the following.
o
n
(+) e (+) e (+) e (+)
e
Corollary 1.5. If A ∈ Ad−1 , Bd , Cd , Dd , then IA is generated by quadratic
binomials in CAsf .
4
e (+) , B
e (+) and D
e (+) are (0, ±1) configurations, by Propositions 1.1
Proof. Since A
d−1
d
d
o
n
e (+) , B
e (+) , D
e (+) .
and 1.4, IA is generated by quadratic binomials in CAsf if A ∈ A
d−1
d
d
(+)
e . By elementary row operations, one can transform the matrix A as
Let A = C
d
follows:
A −→
(+)
Ad−1
1
···
P
1 0 ··· 0
−→
(+)
Ad−1 + 1
···
1
P
1 0 ··· 0
=Q
where 1 is the matrix with all entries equal to one and P is the matrix whose
columns are {ei + ej | 1 ≤ i ≤ j ≤ d}. Since Q is a (0, 1, 2)-configuration, IQ = IA
is generated by quadratic binomials in CAsf by Propositions 1.1 and 1.4.
2. Configurations arising from graphs
In this section, we study toric ideals arising from graphs. First, we introduce
some graph terminology. A walk of G of length q is a sequence Γ = (ei1 , ei2 , . . . , eiq )
of edges of G, where eik = {uk , vk } for k = 1, . . . , q, such that vk = uk+1 for
k = 1, . . . , q − 1. Then,
• A walk Γ is called a path if ♯|{u1, . . . , uq , vq }| = q + 1.
• A walk Γ is called a closed walk if vq = u1 .
• A walk Γ is called a cycle if vq = u1 , q ≥ 3 and ♯|{u1 , . . . , uq }| = q.
For a cycle Γ = (ei1 , ei2 , . . . , eiq ), an edge e = {s, t} of G is called a chord of Γ if s
and t are vertices of Γ and if e ∈
/ {ei1 , ei2 , . . . , eiq }. A cycle Γ is called minimal if Γ
has no chord. Let Γ = (ei1 , ei2 , . . . , ei2q ) where eik = {uk , vk } for k = 1, . . . , 2q be an
even closed walk of G. Then it is easy to see that the binomial
q
q
Y
Y
xu2ℓ v2ℓ
xu2ℓ−1 v2ℓ−1 −
fΓ =
ℓ=1
ℓ=1
belongs to IAG . Circuits of IAG are characterized in terms of graphs (see, e.g., [13,
Lemma 9.8]).
Proposition 2.1. Let G be a finite connected graph. Then f ∈ CAG if and only if
f = fΓ for some even closed walk Γ which is one of the following even closed walk:
(i) Γ is an even cycle of G;
(ii) Γ = (C1 , C2), where C1 and C2 are odd cycles of G having exactly one common vertex;
(iii) Γ = (C1 , e1 , . . . , er , C2 , er , . . . , e1 ), where C1 and C2 are odd cycles of G having no common vertex and where (e1 , . . . , er ) is a path of G which combines
a vertex of C1 and a vertex of C2 .
In particular, f ∈
/ CAsfG if and only if Γ satisfies (iii) and r > 1.
Moreover, it is known [8, Lemma 3.2] that
Proposition 2.2. Let G be a finite connected graph. Then IAG is generated by all
fΓ where Γ is one of the following even closed walk:
5
(i) Γ is an even cycle of G;
(ii) Γ = (C1 , C2), where C1 and C2 are odd cycles of G having exactly one common vertex;
(iii) Γ = (C1 , Γ1 , C2 , Γ2 ), where C1 and C2 are odd cycles of G having no common
vertex and where Γ1 and Γ2 are walks of G both of which combine a vertex
v1 of C1 and a vertex v2 of C2 .
See also [12] for characterization of generators of IAG . The normality of K[AG ] is
characterized in terms of graphs.
Proposition 2.3 ([7]). Let G be a finite connected graph. Then K[AG ] is normal if
and only if G satisfies the odd cycle condition, i.e., for an arbitrary two odd cycles
C1 and C2 in G without common vertex, there exists an edge of G joining a vertex
of C1 with a vertex of C2 .
Let A = (a1 , . . . , an ) ∈ Zd×n be a configuration. Given binomial f = u − v ∈ IA ,
we write Tf for the set of those variables ti such that ti divides π(u)(= π(v)). Let
K[Tf ] = K[{ti | ti ∈ Tf }] and let Af be the matrix whose columns are {ai | T ai ∈
K[Tf ]}. The toric ideal IAf coincides with IA ∩ K[{xi | π(xi ) ∈ K[Tf ]}]. A binomial
f ∈ IA is called fundamental if IAf is generated by f . A binomial f ∈ IA is called
indispensable if, for any system of binomial generators F of IA , either f or −f
belongs to F . A binomial f ∈ IA is called not redundant if f belongs to a minimal
system of binomial generators of IA . Given binomial f ∈ IA , it is known [11] that
• f is fundamental =⇒ f is a circuit
• f is fundamental =⇒ f is indispensable =⇒ f is not redundant
hold in general.
We give a characterization of toric ideals of graphs generated by CAsfG .
Theorem 2.4. Let G be a finite connected graph. Then the following conditions are
equivalent:
(i) IAG = hCAsfG i;
(ii) Any circuit in CAG \ CAsfG is redundant;
(iii) Any circuit in CAG \ CAsfG is not indispensable;
(iv) Any circuit in CAG \ CAsfG is not fundamental;
(v) There exists no induced subgraph of G consisting of two odd cycles C1 , C2
having no common vertex and a path of length ≥ 2 which connects a vertex
of C1 and a vertex of C2 .
In particular, if G satisfies the odd cycle condition, then G satisfies (v).
In order to prove Theorem 2.4, we need the following lemma:
Lemma 2.5. Let G be a finite connected graph which satisfies the condition (v) in
Theorem 2.4. Let C and C ′ be odd cycles of G having no common vertex and let Γ
a path of G which combines a vertex v of C and a vertex v ′ of C ′ . Then, at least
one of the following holds:
(a) There exists an edge of G joining a vertex of C with a vertex of C ′ .
6
(b) There exists an edge of G joining a vertex p of C with a vertex q of Γ where
q 6= v and {p, q} ∈
/ Γ.
(c) There exists an edge of G joining a vertex p of C ′ with a vertex q of Γ where
q 6= v ′ and {p, q} ∈
/ Γ.
Proof. The proof is by induction on the sum of the length of C and C ′ .
(Step 1.) Suppose that C and C ′ are cycle of length 3. Then, C and C ′ are
minimal. If C, C ′ and Γ satisfy none of (a), (b) and (c), then, by the condition
(v), it follows that Γ is not an induced subgraph of G. Then there exists a path Γ′
which combines v and v ′ whose vertex set is a proper subset of the vertex set of Γ.
By repeating the same argument, we may assume that the path Γ′ is an induced
subgraph of G. This contradicts to the condition (v).
(Step 2.) Let C and C ′ be odd cycles of G having no common vertex and let Γ a
path of G which combines a vertex v of C and a vertex v ′ of C ′ . If both C and C ′
are minimal, then one of (a), (b) and (c) follows from the same argument in Step
1. Suppose that C is not minimal, i.e., there exists a chord e of C. It is easy to see
that there exists a unique odd cycle Ce such that e ∈ E(Ce ) ⊂ E(C) ∪ {e}. Note
that the length of Ce is less than that of C.
If v is a vertex of Ce , then Ce , C ′ and Γ satisfy one of (a), (b) and (c) by the
hypothesis of induction. Thus C, C ′ and Γ satisfy the same condition.
Suppose that v is not a vertex of Ce for any chord e of C. Then Ce , C ′ and a path
Γ′ = (ei1 , . . . , eis , Γ) where (ei1 , . . . , eis ) is a part of C satisfy one of (a), (b) and (c)
by the hypothesis of induction. We may assume that s (≥ 1) is minimal. If C, C ′
and Γ satisfy none of (a), (b) and (c), then Ce , C and Γ′ satisfy the condition (b)
where q is not a vertex of Γ. This contradicts the minimality of s.
Proof of Theorem 2.4. In general, (i) =⇒ (iii) and (ii) =⇒ (iii) =⇒ (iv) hold. Moreover, by Proposition 2.1, (iv) =⇒ (v) holds.
(v) =⇒ (i) Suppose that G satisfies the condition (v). Let f = fΓ ∈
/ CAsfG where
Γ is an even closed walk satisfying the condition (iii) in Proposition 2.2, i.e., Γ =
(C1 , Γ1 , C2 , Γ2 ), where C1 and C2 are odd cycles of G having no common vertex,
and Γ1 and Γ2 are walks of G both of which combine a vertex v1 of C1 and a vertex
v2 of C2 . By Propositions 2.1 and 2.2, it is sufficient to show that f is redundant.
Since f does not belong to CAsfG , at least one of Γi is of length > 1. We may assume
that, except for starting and ending vertices, each Γi does not contain the vertices
of two odd cycles. (Otherwise, Γ separates into two even closed walk and hence f
is redundant.)
If there exists an edge of G joining a vertex of C1 with a vertex of C2 , then f is
redundant by [10, Proof of Lemma 3.2]. Suppose that there exists no such an edge.
(Then, in particular, the length of Γi is greater than 1 for i = 1, 2.) By Lemma 2.5,
there exists an edge of G joining a vertex p of C1 with a vertex q (6= v1 ) of Γ1 and
{p, q} does not belong to Γ. Let C1 = (V1 , V2 ) and Γ1 = (W1 , W2 ) where
• V1 and V2 are paths joining v1 and p;
• Wi is a walk joining vi and q for i = 1, 2.
7
Since the length of C1 is odd, we may assume that the length of the walk (V1 , W1 ) is
odd. Note that both Γ3 = (V1 , W1 , {q, p}) and Γ4 = (V2 , Γ2 , C2 , W2 , {q, p}) are even
closed walks. It then follows that f ∈ hfΓ3 , fΓ4 i and deg(fΓ3 ), deg(fΓ4 ) < deg(f ).
Hence f is redundant.
Thus, f is redundant and hence G satisfies the condition (i).
(v) =⇒ (ii) Suppose that G satisfies the condition (v). Let f = fΓ ∈ CAG \ CAsfG
where Γ = (C1 , e1 , . . . , er , C2 , er , . . . , e1 ) (r > 1) is an even closed walk satisfying
the condition (iii) in Proposition 2.1. Then f is redundant by the same argument
above. Thus G satisfies the condition (ii).
A similar theorem holds for CAsfsf
.
G
Theorem 2.6. Let G be a finite connected graph. Then the following conditions are
equivalent:
(i) IAG = hCAsfsf
i;
G
(ii) Any circuit in CAG \ CAsfsf
is redundant;
G
(iii) Any circuit in CAG \ CAsfsf
is not indispensable;
G
(iv) Any circuit in CAG \ CAsfsf
is not fundamental;
G
(v) There exists no induced subgraph of G consisting of two odd cycles C1 , C2
having no common vertex and a path of length ≥ 1 which connects a vertex
of C1 and a vertex of C2 .
Proof. As stated in Proof of Theorem 2.4, it is sufficient to show “(v) =⇒ (i)” and
“(v) =⇒ (ii).” Suppose that G satisfies (v). By Theorem 2.4, IAG = hCAsfG i and
any circuit in CAG \ CAsfG is redundant. Thus, in order to prove (i) and (ii), it is
sufficient to show that any circuit in CAsfG \ CAsfsf
is redundant. Let f be a binomial
G
sf
sfsf
in CAG \ CAG . By Proposition 2.1, f = fΓ where Γ is an even closed walk which
consists of two odd cycles C1 and C2 having no common vertex and an edge e0 of G
which combines a vertex of C1 and a vertex of C2 . Since G satisfies the condition
(v), Γ is not an induced subgraph of G. If there exists an edge e′ (6= e0 ) of G joining
a vertex of C1 with a vertex of C2 , then f is redundant by [10, Proof of Lemma 3.2].
Suppose that there exists no such an edge. Since Γ is not an induced subgraph of
G, we may assume that C1 is not minimal. Then there exists a chord e of C1 and
an odd cycle Ce such that e ∈ E(Ce ) ⊂ E(C) ∪ {e}. If v is a vertex of Ce , then f
is redundant by [10, Proof of Lemma 3.2]. Suppose that v is not a vertex of Ce for
any chord e of C. Note that Ce , C2 and Γ = (ei1 , . . . , eis , e0 ) where (ei1 , . . . , eis ) is a
part of C satisfy one of (a), (b) and (c) in Lemma 2.5. Suppose that s is minimal.
Since there exists no edge e′ (6= e0 ) of G joining a vertex of C1 with a vertex of C2 ,
Ce , C2 and Γ satisfy the condition (b). This contradicts the minimality of s.
Using Theorems 2.4 and 2.6, we give several classes of graphs G such that IAG =
hCAsfG i and K[AG ] is nonnormal.
8
Example 2.7. Let G be the graph whose
1 0 0 1
1 1 0 0
0 1 1 0
AG = 0 0 1 1
0 0 0 0
0 0 0 0
0 0 0 0
vertex-edge incidence matrix is
0 0 0 1 0
0 0 0 0 0
0 0 0 1 0
1 0 0 0 0
1 1 0 0 1
0 1 1 0 0
0 0 1 0 1
Then, IAG is generated by the circuits x1 x3 −x2 x4 , x3 x4 x6 x9 −x25 x7 x8 ([10, Example
3.5]). Since G does not satisfy the odd cycle condition, K[AG ] is not normal.
Example 2.8. Let G be the graph whose vertex-edge incidence matrix is
1 0 0 1 0 0 0 0 1 0
1 1 0 0 0 0 0 0 0 0
0 1 1 0 0 0 0 0 1 0
AG = 0 0 1 1 1 0 0 1 0 0 .
0 0 0 0 1 1 0 0 0 1
0 0 0 0 0 1 1 0 0 0
0 0 0 0 0 0 1 1 0 1
Then, IAG is generated by the circuits x5 x7 − x6 x8 , x1 x3 − x2 x4 , x3 x4 x10 − x5 x8 x9 .
Since G does not satisfy the odd cycle condition, K[AG ] is not normal.
Example 2.8 is the most simple nonnormal example whose toric ideal is generated
by circuits u − v such that the two monomials u and v are squarefree. In fact,
+
+
Proposition 2.9. If IA is generated by binomials f1 = X u −X u , f2 = X v −X v
−
+
−
+
such that X u , X u , X v and X v are squarefree, then there exists a monomial
order such that {f1 , f2 } is a Gröbner basis of IA and hence K[A] is normal.
+
−
−
+
Proof. Suppose that xi ∈ supp(X u )∩supp(X v ) and xj ∈ supp(X u )∩supp(X v ).
−
+
Let w = u + v ∈ KerZ (A) and g = X w − X w . Then g belongs to IA . Since xi
+
−
belongs to supp(X u ) ∩ supp(X v ), supp(g) does not contain xi . Similarly, since
−
+
xj belongs to supp(X u ) ∩ supp(X v ), supp(g) does not contain xj . Hence g is
not generated by f1 and f2 . This contradicts that g ∈ IA . Thus, we may assume
+
−
+
+
that supp(X u ) ∩ supp(X v ) = ∅ and supp(X u ) ∩ supp(X v ) = ∅. Let < be a
lexicographic order induced by the ordering
−
+
−
+
supp(X u ) > supp(X v ) > other variables.
+
+
Then in< (f1 ) = X u and in< (f2 ) = X v are relatively prime. Hence {f1 , f2 } is
a Gröbner basis of IA . Since both in< (f1 ) and in< (f2 ) are squarefree, K[A] is
normal.
Let G1 = (V1 , E1 ) and G2 = (V2 , E2 ) be graphs such that V1 ∩V2 is a clique of both
graphs. The new graph G = G1 ♯G2 with the vertex set V = V1 ∪ V2 and edge set
E = E1 ∪ E2 is called the clique sum of G1 and G2 along V1 ∩ V2 . If the cardinality
of V1 ∩ V2 is k + 1, this operation is called a k-sum of the graphs.
9
Example 2.10. Let G be the 0-sum of two complete graphs having at least 4 vertices. Then, G satisfies the condition (v) in Theorem 2.6 and hence IAG is generated
by CAsfsf
. Since G does not satisfy the odd cycle condition, K[AG ] is not normal.
G
On the other hand, by the criterion [3, Theorem 2.1], it follows that K[AG ] does not
satisfy Serre’s condition (R1 ).
Example 2.11. Let G be the 1-sum of two complete graphs having at least 5 vertices. Then, G satisfies the condition (v) in Theorem 2.6 and hence IAG is generated
by CAsfsf
. Since G does not satisfy the odd cycle condition, K[AG ] is not normal.
G
On the other hand, by the criterion [3, Theorem 2.1], it follows that K[AG ] satisfies
Serre’s condition (R1 ).
References
[1] T. Bogart, A. N. Jensen and R. R. Thomas, The circuit ideal of a vector configuration, J.
Algebra 309 (2007), 518 – 542.
[2] I. M. Gelfand, M. I. Graev and A. Postnikov, Combinatorics of hypergeometric functions
associated with positive roots, in “Arnold–Gelfand Mathematics Seminars, Geometry and
Singularity Theory” (V. I. Arnold, I. M. Gelfand, M. Smirnov and V. S. Retakh, Eds.),
Birkhäuser, Boston, 1997, pp. 205 – 221.
[3] T. Hibi and L. Katthän, Edge rings satisfying Serre’s condition (R1 ), preprint, 2012.
arXiv:1202.4889v2 [math.CO]
[4] J. E. Humphreys, “Introduction to Lie Algebras and Representation Theory,” Second Printing,
Revised, Springer–Verlag, Berlin, Heidelberg, New York, 1972.
[5] J. Martinez-Bernal and R. H. Villarreal, Toric ideals generated by circuits, Algebra Colloq. 19
(2012), 665 – 672.
[6] H. Ohsugi, J. Herzog and T. Hibi, Combinatorial pure subrings, Osaka J. Math. 37 (2000),
745 – 757.
[7] H. Ohsugi and T. Hibi, Normal polytopes arising from finite graphs, J. Algebra 207 (1998),
409 – 426.
[8] H. Ohsugi and T. Hibi, Toric ideals generated by quadratic binomials, J. Algebra 218 (1999),
509 – 527.
[9] H. Ohsugi and T. Hibi, Quadratic initial ideals of root systems, Proceedings of the AMS, 130
(2002), 1913 – 1922.
[10] H. Ohsugi and T. Hibi, Indispensable binomials of finite graphs, J. Algebra and Its Applications
4 (2005), 421 – 434.
[11] H. Ohsugi and T. Hibi, Toric ideals arising from contingency tables, in “Commutative Algebra
and Combinatorics” (W. Bruns Ed.), Ramanujan Mathematical Society Lecture Notes Series,
Number 4, Ramanujan Math. Soc., Mysore, 2007, pp. 91 – 115.
[12] E. Reyes, C. Tatakis and A. Thoma, Minimal generators of toric ideals of graphs, Advances
in Applied Math. 48 (2012), 64 – 78.
[13] B. Sturmfels, “Gröbner bases and convex polytopes,” Amer. Math. Soc., Providence, RI, 1996.
Hidefumi Ohsugi, Department of Mathematics, College of Science, Rikkyo University, Toshima-ku, Tokyo 171-8501, Japan
E-mail address: [email protected]
Takayuki Hibi, Department of Pure and Applied Mathematics, Graduate School
of Information Science and Technology, Osaka University, Toyonaka, Osaka 5600043, Japan
E-mail address: [email protected]
10
| 0math.AC
|
Reviving the Two-state Markov Chain Approach
(Technical Report)
Andrzej Mizera1 , Jun Pang1,2 , and Qixia Yuan1?
1
arXiv:1501.01779v4 [cs.CE] 25 Oct 2016
2
University of Luxembourg, FSTC, Luxembourg
University of Luxembourg, SnT, Luxembourg
[email protected]
Abstract. Probabilistic Boolean networks (PBNs) is a well-established computational framework for modelling biological systems. The steady-state dynamics
of PBNs is of crucial importance in the study of such systems. However, for
large PBNs, which often arise in systems biology, obtaining the steady-state distribution poses a significant challenge. In fact, statistical methods for steady-state
approximation are the only viable means when dealing with large networks. In
this paper, we revive the two-state Markov chain approach presented in the literature. We first identify a problem of generating biased results, due to the size of
the initial sample with which the approach needs to start and we propose a few
heuristics to avoid such a pitfall. Second, we conduct an extensive experimental
comparison of the two-state Markov chain approach and another approach based
on the Skart method and we show that statistically the two-state Markov chain
has a better performance. Finally, we apply this approach to a large PBN model
of apoptosis in hepatocytes.
1
Introduction
Systems biology aims to study biological systems from a holistic perspective, with
the goal to provide a comprehensive, system-level understanding of cellular behaviour.
Proper functioning of a living cell requires a finely-tuned and orchestrated interplay
of many complex processes. Complex interactions within a biological system lead to
emergent properties which are crucial for sustaining life. Therefore, understanding the
machinery of life requires the use of holistic approaches which enable the study of a system as a whole, in contrast to the reductionist approach. Computational modelling plays
a prominent role in the field of systems biology. Construction and analysis of a computational model for some biological process enables the systematisation of available
biological knowledge, identification of missing biological information, provides formal
means for understanding and reasoning about the concerted interplay between different
parts of the model, finally reveals directions for future experimental work which could
provide data for better understanding of the process under study.
Unfortunately, computational modelling of biological processes that take place in
a living cell poses significant challenges with respect to the size of the state-space that
?
Supported by the National Research Fund, Luxembourg (grant 7814267) and partially by
Google Summer of Code 2014.
2
Mizera, Pang and Yuan
needs to be considered. Modelling of certain parts of cellular machinery such as gene
regulatory networks (GRNs) or signal transduction pathways often leads to dynamical models characterised by huge state-spaces of sizes that surpass the sizes of any
human-designed systems by orders of magnitude. Therefore, profound understanding
of biological processes asks for the development of new methods and approaches that
would provide means for formal analysis and reasoning about such huge systems.
In this study we concentrate on the analysis of the steady-state dynamics of biological processes, in particular GRNs, modelled as discrete-time Markov chains (DTMCs).
This is the case, for example, when the biological system under study is cast into the
mathematical/computational framework of probabilistic Boolean networks (PBNs) [1,2].
In these or other discrete-time models, e.g., dynamic Bayesian networks, the real (considered as continuous) time is not modelled. Instead, the evolution of the system is
abstracted as a sequence of consecutive events. These coarse-grained models have been
successfully applied in many systems biology studies and proved their predictive power [3].
In fact, for the study of large regulatory systems they remain the only reasonable solution. Extrapolating the ordinary differential equations model of a single elementary
building block of the network (e.g., a gene) to the whole large system would result
in a prohibitively complex model. However, moving towards a higher-level description by ignoring the molecular details allows to grasp the system-level behaviour of the
network [4]. In consequence, these coarse-grained formalisms are broadly applied in
systems biology studies of systems where the predictions of exact reaction times are
not of main interest. For example, this is the case in the study of dynamical attractors
of a regulatory network, which seem to depend on the circuit wiring rather than kinetic
constants (such as production rates or interaction rates) [5]. In this sense modelling
biological systems with more abstract, high-level view formalisms has certain unquestionable advantages.
One of the key aspects in the analysis of such dynamic systems is the comprehension of their steady-state (long-run) behaviour. For example, attractors of such systems
were hypothesised to characterise cellular phenotypes [6]. Another complementary conjecture is that attractors correspond to functional cellular states such as proliferation,
apoptosis, or differentiation [7]. These interpretations may cast new light on the understanding of cellular homeostasis and cancer progression [1]. In this work, we focus
on the computation of steady-state probabilities which are crucial for the determination
of long-run influences and sensitivities. These are measures that quantify the impact
of genes on other genes, considered collectively or individually, and that enable the
identification of elements with highest impact. In this way they provide insight into the
control mechanisms of the network.
So far the huge-state space, which often characterises dynamical models of biological systems, tempers the application of the above mentioned techniques in the analysis
of realistic biological systems. In fact, approximations with the use of Markov chain
Monte Carlo (MCMC) techniques are the only viable solution to this problem [8]. However, due to the difficulties with the assessment of the convergence rate to the steadystate distribution (see, e.g., [9]), certain care is required when applying these methods
in practice. A number of statistical methods exists, which allow to empirically determine when to stop the simulation and output estimates. We employ in our study: (1) the
Reviving the Two-state Markov Chain Approach
3
two-state Markov chain approach [10] and (2) the Skart batch-means method of [11].
The two-state Markov chain approach was introduced in 1992 by Raftery and Lewis;
and Shmulevich et al. [8] proposed its application to the analysis of PBNs in 2003.
However, to the best of our knowledge, since then it has not been widely applied for
the analysis of large PBNs. In this paper, we aim to revive its usage for approximating
steady-state probabilities of large PBNs, which often arise in computational systems
biology as models of full-size genetic networks. We identify a problem concerned with
the choice of the initial sample size in the two-state Markov chain approach. As we
show, an unconscious choice may lead to biased results. We propose a few heuristics
to avoid this problem. By extensive experiments, we show that the two-state Markov
chain approach outperforms the Skart method in most cases, where the batch-means
Skart method is considered the current state-of-the-art approach. In this way we show
that the two-state Markov chain is often the optimal choice for the analysis of large
PBN models of biological systems.
Structure of the paper. After presenting some preliminaries in Section 2, we describe
the two-state Markov chain approach in Section 3 and identify a problem of generating
biased results, due to the size of the initial sample with which the approach needs to
start. We then propose some heuristics for the approach to avoid unfortunate initialisations. We perform an extensive evaluation and comparison of the two-state Markov
chain approach and the Skart method in Section 4 on a large number of randomly generated PBNs. In most cases, the two-state Markov chain approach seems to have a better
performance in terms of computational cost. Finally, we apply the two-state Markov
chain approach to study a large PBN model of apoptosis in hepatocytes consisting of
91 nodes in Section 5. We compute the steady-state influences and long-run sensitivities and confirm previously formulated hypothesis. We conclude our paper with some
discussions in Section 6.
2
2.1
Preliminaries
Finite discrete-time Markov chains (DTMCs)
Let S be a finite set of states. A (first-order) discrete-time Markov chain is an S-valued
stochastic process {Xt }t∈N with the property that the next state is independent of the past
states given the present state. Formally, P(Xt+1 = st+1 | Xt = st , Xt−1 = st−1 , . . . , X0 =
s0 ) = P(Xt+1 = st+1 | Xt = st ) for all st+1 , st , . . . , s0 ∈ S. Here, we consider time-homogenous
Markov chains, i.e., chains where P(Xt+1 = s0 | Xt = s), denoted Ps,s0 , is independent of
t for any states s, s0 ∈ S. The transition matrix P = (Ps,s0 )s,s0 ∈S satisfies Ps,s0 > 0 and
∑s0 ∈S Ps,s0 = 1 for all s ∈ S. We denote by π a probability distribution on S. If π = π P,
then π is a stationary distribution of the DTMC (also referred to as the invariant distribution). A path of length n is a sequence s1 → s2 → · · · → sn such that Psi ,si+1 > 0 and
si ∈ S for i ∈ {1, 2, . . . , n}. State q ∈ S is reachable from state p ∈ S if there exists a path
such that s1 = p and sn = q. A DTMC is irreducible if any two states are reachable
from each other. The period of a state is defined as the greatest common divisor of the
lengths of all paths that start and end in the state. A DTMC is aperiodic if all states in
4
Mizera, Pang and Yuan
S are of period 1. A finite state DTMC is called ergodic if it is irreducible and aperiodic. By the famous ergodic theorem for DTMCs [12] an ergodic chain has a unique
stationary distribution being its limiting distribution (also referred to as the steady-state
distribution given by limn→∞ π0 Pn , where π0 is any initial probability distribution on
S). In consequence, the limiting distribution for an ergodic chain is independent of the
choice of π0 . The steady-state distribution can be estimated from any initial distribution
by iteratively multiplying it by P.
The evolution of a first-order DTMC can be described by a stochastic recurrence
sequence Xt+1 = φ (Xt ,Ut+1 ), where {Ut }t∈N is an independent sequence of uniformly
distributed real random variables over [0, 1] and the transition function φ : S × [0, 1] → S
satisfies the property that P(φ (s,U) = s0 ) = Ps,s0 for any states s, s0 ∈ S and for any U,
a real random variable uniformly distributed over [0, 1]. When S is partially ordered and
when the transition function φ (·, u) is monotonic for all u, then the Markov chain is said
to be monotone ([13,14]).
2.2
Probabilistic Boolean networks (PBNs)
A PBN G(V, F ) consists of a set of binary-valued nodes (also known as genes) V =
{v1 , v2 , . . . , vn } and a list of sets F = (F1 , F2 , . . . , Fn ). For each i ∈ {1, 2, . . . , n} the set
(i) (i)
(i)
Fi = { f1 , f2 , . . . , fl(i) } is a collection of predictor functions for node vi , where l(i) is
(i)
the number of predictor functions for vi . Each f j ∈ Fi is a Boolean function defined
with respect to a subset of nodes referred to as parent nodes of vi . There is a probability
(i)
(i)
distribution on each Fi ∈ F: c j is the probability of selecting f j ∈ Fi as the next
l(i)
(i)
predictor for vi and it holds that ∑ j=1 c j = 1. We denote by vi (t) the value of node vi at
time point t ∈ N. The state space of the PBN is S = {0, 1}n and it is of size 2n . The state
of the PBN at time t is given by s (t) = (v1 (t), v2 (t), . . . , vn (t)). The dynamics of the PBN
∞ . We consider here independent PBNs where predictor
is given by the sequence (ss(t))t=0
functions for different nodes are selected independently of each other. The transition
from s (t) to s (t + 1) is conducted by randomly selecting a predictor function for each
node vi from Fi and by synchronously updating the node values in accordance with the
selected functions. There are N = ∏ni=1 l(i) different ways in which the predictors can
be selected for all n nodes. These combinations are referred to as realisations of the
(1) (2)
(n)
PBN and are represented as n-dimensional function vectors f k = ( fk1 , fk2 , . . . , fkn ) ∈
F1 ×F2 ×. . .×Fn , where k ∈ {1, 2, . . . , N} and ki ∈ {1, 2, . . . , l(i)}. A realization selected
(i)
at time t is referred to as f (t). Due to independence, P( f k ) = P( f (t) = f k ) = ∏ni=1 cki .
In PBNs with perturbations, a perturbation parameter p ∈ (0, 1) is introduced to
sample the perturbation vector γ (t) = (γ1 (t), γ2 (t), . . . , γn (t)), where γi (t) ∈ {0, 1} and
P(γi (t) = 1) = p for all t and i ∈ {1, 2, . . . , n}. Perturbations provide an alternative way
to regulate the dynamics of a PBN: the next state is determined as s(t +1) = f (t)(ss(t)) if
γ (t) = 0 and as s(t + 1) = s(t) ⊕ γ (t) otherwise, where ⊕ is the exclusive or operator for
vectors. The perturbations, by the latter update formula, allow the system to move from
any state to any other state in one single transition, hence render the underlying Markov
chain irreducible and aperiodic. Therefore, the dynamics of a PBN with perturbations
can be viewed as an ergodic DTMC [15]. The transition matrix is given by Ps,s0 = (1 −
Reviving the Two-state Markov Chain Approach
0
5
0
p)n ∑Nk=1 1 [ f k (s)=s0 ] P( f k ) + (1 − (1 − p)n )pη(s,s ) (1 − p)n−η(s,s ) , where 1 is the indicator
function and η(s, s0 ) is the hamming distance between states s, s0 ∈ S. According to
the ergodic theory, adding perturbations to any PBN assures the long-run dynamics of
the resulting PBN is governed by a unique limiting distribution, convergence to which
is independent of the choice of the initial state. However, the perturbation probability
value should be chosen carefully, not to dilute the behaviour of the original PBN. In this
way the ‘mathematical trick’, although introduces some noise to the original system,
allows to significantly simplify the analysis of the steady-state behaviour, which is often
of interest for biological systems.
The density of a PBN is measured with its function number and parent nodes numF
ω(i), where n is the number
ber. For a PBN G, its density is defined as D(G) = 1n ∑Ni=1
of nodes in G, NF is the total number of predictor functions in G, and ω(i) is the number
of parent nodes for the ith predictor function.
Within the framework of PBNs the concept of influences is defined; it formalizes
the impact of parents nodes on a target node and enables its quantification ([23]). The
concept is based on the notion of a partial derivative of a Boolean function f with
respect to variable x j (1 ≤ j ≤ n):
∂ f (x)
= f (x( j,0) ) ⊕ f (x( j,1) ),
∂xj
where ⊕ is addition modulo 2 (exclusive OR) and for l ∈ {0, 1}
x( j,l) = (x1 , x2 , . . . , x j−1 , l, x j+1 , . . . , xn ).
The influence of node x j on function f is the expected value of the partial derivative
with respect to the probability distribution D(x):
I j ( f ) = ED
h ∂ f (x) i
∂xj
n ∂ f (x)
o
=P
= 1 = P{ f (x( j,0) ) 6= f (x( j,1) )}.
∂xj
(i)
Let now Fi be the set of predictors for xi with corresponding probabilities c j for j =
(i)
(i)
1, . . . , l(i) and let Ik ( f j ) be the influence of node xk on the predictor function f j . Then,
the influence of node xk on node xi is defined as:
l(i)
Ik (xi ) =
(i)
∑ Ik ( f j
j=1
(i)
)·cj .
The long-term influences are the influences computed when the distribution D(x) is the
stead-state distribution of the PBN.
We define and consider in this study two types of long-run sensitivities.
Definition 1. The long-run sensitivity with respect to selection probability perturbation
is defined as
(i)
(i)
sc [c j = p] = kπ̃[c j = p] − πkl ,
6
Mizera, Pang and Yuan
where k · kl denotes the l-norm, π is the steady-state distribution of the original PBN,
(i)
(i)
p ∈ [0, 1] is the new value for c j , and π̃[c j = p] is the steady-state probability distribution of the PBN perturbed as follows. The jth selection probability for node xi is
(i)
(i)
replaced with c̃ j = p and all ck selection probabilities for k ∈ I− j = {1, 2, . . . , j −
1, j + 1, . . . , l(i)} are replaced with
(i)
(i)
(i)
c̃k = ck + (c j − p) ·
(i)
ck
(i)
,
∑l∈I− j cl
The remaining selection probabilities of the original PBN are unchanged.
Definition 2. The long-run sensitivity with respect to permanent on/off perturbations
of a node xi as
sg [xi ] = max{kπ̃[xi ≡ 0] − πkl , kπ̃[xi ≡ 1] − πkl },
where π, π̃[xi ≡ 0], and π̃[xi ≡ 1] are the steady-state probability distributions of the
original PBN, of the original PBN with all f (i) ∈ Fi replaced by f˜(i) ≡ 0, and all f (i) ∈ Fi
replaced by f˜(i) ≡ 1, respectively.
Notice that the definition of long-run sensitivity with respect to permanent on/off
perturbations is similar but not equivalent to the definition of long-run sensitivity with
respect to 1-gene function perturbation of [23].
3
3.1
The Two-state Markov Chain Approach
Description
We recall the two-state Markov chain approach originally introduced in [10]. The twostate Markov chain approach is a method for estimating the steady-state probability of
a subset of states of a DTMC. In this approach the state space of an arbitrary DTMC is
split into two disjoint sets, referred to as meta states. One of the meta states, numbered
1, is the subset of interest and the other, numbered 0, is its complement. The steadystate probability of meta state 1, denoted q, can be estimated by performing simulations
of the original Markov chain. For this purpose a two-state Markov chain abstraction of
the original DTMC is considered. Let {Zt }t>0 be a family of binary random variables,
where Zt is the number of the meta state the original Markov chain is in at time t.
{Zt }t>0 is a binary (0-1) stochastic process, but in general it is not a Markov chain.
However, as argued in [10], a reasonable assumption is that the dependency in {Zt }t>0
(k)
(k)
falls off rapidly with lag. Therefore, a new process {Zt }t>0 , where Zt = Z1+(t−1)k ,
will be approximately a first-order Markov chain for k large enough. A procedure for
determining appropriate k is given in [10]. The first-order Markov chain consists of the
two meta states with transition probabilities α and β between them. See Figure 1 for
a conceptual illustration of the construction of this abstraction.
The steady-state probability estimate q̂ is computed from a simulated trajectory of
the original DTMC. The key point is to determine the optimal length of the trajectory.
Reviving the Two-state Markov Chain Approach
0
7
1
A
D
E
1 ↵
B
C
(a) Original DTMC
0
1
1
↵
(b) Two-state DTMC
Fig. 1: Conceptual illustration of the idea of the two-state Markov chain construction.
(a) The state space of the original discrete-time Markov chain is split into two meta
states: states A and B form meta state 0, while states D, C, and E form meta state 1. The
split of the state space into meta states is marked with dashed ellipses. (b) Projecting the
behaviour of the original chain on the two meta states results in a binary (0-1) stochastic
process. After potential subsampling, it can be approximated as a first-order, two-state
Markov chain with the transition probabilities α and β set appropriately.
Two requirements are imposed. First, the abstraction of the DTMC, i.e., the two-state
Markov chain, should converge close to its steady-state distribution π = [π0 π1 ]. For(k)
(k)
mally, t satisfying |P[Zt = i | Z0 = j] − πi | < ε for a given ε > 0 and all i, j ∈ {0, 1}
needs to be determined. t is the so-called ‘burn-in’ period and determines the part of
the trajectory of the two-state Markov chain that needs to be discarded. Second, the
estimate q̂ is required to satisfy P[q − r 6 q̂ 6 q + r] > s, where r is the required precision and s is a specified confidence level. This condition is used to determine the length
of the second part of the trajectory used to compute q̂, i.e., the sample size. Now, the
total required trajectory length of the original DTMC is then given by M + N, where
M = 1 + (t − 1)k and N = 1 + (dn(α, β )e − 1)k, where t = dm(α, β )e. The functions m
and n depend on the transitions probabilities α and β and are given by
2
ε(α+β )
log max(α,β
αβ (2 − α − β ) Φ−1 ( 21 (1 + s))
)
and n(α, β ) =
,
m(α, β ) =
log (|1 − α − β |)
(α + β )3
r2
where Φ−1 is the inverse of the standard normal cumulative distribution function. The
expressions for m and n were originally presented in [10]. Derivations however were not
provided and the expressions contain two oversights: in the formula for m the absolute
value is missing in the denominator and in the formula for n the inverse of Φ should be
used instead of Φ. We provide detailed derivations of the expressions for m and n in the
Appendices A and B, respectively.
Since α and β are unknown, they need to be estimated. This is achieved iteratively
in the two-state Markov chain approach of [10]. It starts with sampling an arbitrary
initial length trajectory, which is then used for estimating the values of α and β . M and
N are calculated based on these estimates. Next, the trajectory is extended to reach the
required length, and α and β values are re-estimated. The new estimates are used to
8
Mizera, Pang and Yuan
re-calculate M and N. This process is iterated until M + N is smaller than the current
trajectory length. Finally, the resulting trajectory is used to estimate the steady-state
probability of meta state 1. For more details, see [10].
The two-state Markov chain approach can be viewed as an aggregation method for
reducing the state space. Generically, aggregation methods aggregate groups of nodes
in the original Markov chain in accordance with a given partition function, which leads
to a smaller transition graph. In principle, the aggregation can be any Markov chain over
this smaller transition system. What remains is the choice of the specific aggregation
which is determined by the choice of the transition probabilities. The partition function
is also used to obtain the so-called projection of the original process, i.e., the realisation of the original Markov chain is projected through the partition function. Ideally,
the aggregated chain and the projected realisation should coincide. However, since the
projection is in general not Markovian, the aggregation which is “closest” to the projection is considered instead and the “closeness” has to be appropriately defined. The
problem of finding the optimal partition function in the case where the distance between
the projection and the aggregation is quantified by the Kullback-Leibler divergence rate
(KLDR) has been recently studied, see, e.g., [16,17,18]. The focus in these works is on
finding the optimal partition function, for which the KLDR distance between the projection and aggregation is minimised. In our case the partition function which classifies
the original states into the two meta states is specified by the biological question under
study. As shown in [17] and [18], for a given partition function, the aggregation closest to the projection can be analytically obtained provided the steady-state distribution
of the original chain is available. The steady-state probabilities are however our goal,
thus these techniques cannot be exploited for the determination of α and β transition
probabilities. The iterative, statistical estimation of the transition probabilities for the
aggregation remains the only viable solution.
3.2
The choice of the initial sample size
Given good estimates of α and β , the theory of the two-state Markov chain presented
above guarantees that the obtained value satisfies the imposed precision requirements.
However, the two-state Markov chain approach starts with generating a trajectory of the
original DTMC of an arbitrarily chosen initial length, i.e., M0 + N0 = 1 + (m0 − 1)k +
1 + (n0 − 1)k, where m0 it the ‘burn-in’ period and n0 is the sample size of the two-state
Markov chain abstraction. An unfortunate choice may lead to first estimates of α and
β that are biased and result in the new values of M and N such that M + N is either
smaller or not much larger than the initial M0 + N0 . In the former case the algorithm
stops immediately with the biased values for α, β and, more importantly, with an estimate for the steady-state probability that does not satisfy the precision requirements.
The second case may lead to the same problem. As an example we considered a two24
(0.0020214) and β = 24
state Markov chain with α = 11873
25 (0.96). The steady-state
probability distribution was [0.997899 0.002101]. With k = 1, ε = 10−6 , r = 10−3 ,
1
s = 0.95, m0 = 5, and n0 = 1, 920 the first estimated values for α and β were 1918
(0.0005214) and 1, respectively. This subsequently led to M = 2 and N = 1, 999, resulting in a request for the extension of the trajectory by 76. After the extension, the
1
new estimates for α and β were 1997
and 1, respectively. These estimates gave M = 2,
Reviving the Two-state Markov Chain Approach
9
N = 1, 920, and the algorithm stopped. The estimated steady-state probability distribution was [0.99950 0.00050], which was outside the pre-specified precision interval
given by r. Independently repeating the estimation for 104 times resulted in estimates
of the steady-state probabilities that were outside the pre-specified precision interval
90% of times. Given the rather large number of repetitions, it can be concluded that the
specified 95% confidence interval was not reached in this case.
The reason for the biased result is the unfortunate initial value for n0 and the fact that
the real value of α is small. In the initialisation phase the value of α is underestimated
and dn(α, β )e calculated based on the estimated values of α and β is almost the same
as n0 . Hence, subsequent extension of the trajectory does not provide any improvement
to the underestimated value of α since the elongation is short and the algorithm halts
after the next iteration.
To identify and avoid some of such pitfalls, we consider a number of cases and formulate some of the conditions in which the algorithm may fail to achieve the specified
precision. Let n0 be the initial size of the sample used for initial estimation of α and
β . We assume that neither α nor β is zero. Then, the smallest possible estimates for
both α and β are greater than n10 . Let us set an upper bound value for n0 to be 104 .
For most cases this boundary value is reasonable and we expect n0 to be much smaller.
Notice however that this is the case only if the real values of α and β are larger than
10−4 . We just mention here that in general the selection of a proper value for n0 heavily
depends on the real values of α and β , which are unknown a priori. From what was
stated above, it follows that both first estimates for α and β are greater than 10−4 . The
following cases are possible.
– If both α and β are small, e.g., less than 0.1, then we have that 10−4 < α, β < 0.1
and n(α, β ) > 72, 765 as can be seen by investigating the n(·, ·) function. In this
case the sample size is increased more than 7-fold which is reasonable since the
two-state Markov chain seems to be bad-mixing by the first estimates of the values
for α and β and the algorithm asks for a significant increase of the sample size. We
therefore conclude that the bad-mixing Markov chain case can be properly handled
by the algorithm.
– Both first estimates of α and β are close to 1. If α, β ∈ [0.7, 0.98], the value of
n(α, β ) is larger than 19, 000. If both α, β > 0.98 than the size of the sample drops,
but in this case the Markov chain is highly well-mixing and short trajectories are
expected to provide good estimates.
– The situation is somewhat different if one of the parameters is estimated to be
small and the other is close to 1 as in the example described above. The extension
to the trajectory is too small to significantly change the estimated value of the small
parameter and the algorithm halts.
Considering the above cases leads us to the observation that the following situation
needs to be treated with care: The estimated value for one of the parameters is close to
1
n0 , the value of the second parameter is close to 1, and n(α, β ) is either smaller or not
significantly larger than n0 .
First approach: pitfall avoidance. In order to avoid this situation, we determine n0
which in principle could lead to initial inaccurate estimates of α or β and such that the
10
Mizera, Pang and Yuan
0.01
0.001
0.0001
0.9 0.95 0.975 0.9
0.95
0.975
0.9
0.95
0.975
n0 ∈ 0/ [2, 136] 0/ [2, 1161] [2, 1383] [2, 1582] [2, 11628] [2, 13857] [2, 15847]
r
s
Table 1: Ranges of integer values for n0 that do not satisfy the ‘critical’ condition
n(α, β ) < 2 n0 for the given values of r and s.
next sample size given by dn(α, β )e would practically not allow for an improvement of
the estimates. We reason as follows. As stated above, the ‘critical’ situation may take
place when one of the parameters is estimated to be very small, i.e., close to n10 , and
the increase in the sample size is not significant enough to improve the estimate. If the
initial estimate is very small, the real value is most probably also small, but the estimate
is not accurate. If the value is underestimated to the lowest possible value, i.e., n10 , on
average the improvement can take place only if the sample size is increased at least by
n0 . Therefore, with the trade-off between the accuracy and efficiency of the method in
mind, we propose the sample size to be increased at least by n0 . Therefore, the ‘critical’
situation condition is n(α, β ) < 2 n0 . By analysing the function n(·, ·) as described in
details in Appendix D, we can determine the values of n0 that are ‘safe’, i.e., which do
not satisfy the ‘critical’ condition. We present them in Table 1 for a number of values
for r and s.
Second approach: controlled initial estimation of α and β . The formula for n is
asymptotically valid for a two-state Markov chain provided that the values for α and β
are known. However, these values are not known a priori and they need to be estimated.
Unfortunately, the original approach does not provide any control over the quality of the
initial estimate of the values of these parameters. In certain situation, e.g., as in the case
discussed above, the lack of such a control mechanism may lead to results with worse
statistical confidence level than the specified one given by s. In the discussed example
s = 95%, but this value was not reached in the performed experiment. In order to address
this problem, we propose to extend the initial phase of the two-state approach algorithm
in the following way. The algorithm samples a trajectory of the given Markov chain and
estimates the values of α and β . It might be the case that an arbitrarily chosen initial
sample size is not big enough to provide non-zero estimates for the two parameters.
If this is the case, the initial sample size is doubled and the trajectory is elongated to
collect a sample of required size. This is repeated iteratively until non-zero estimates
for both α and β are obtained. We introduce the following notation: α̂ and β̂ are the
non-zero estimates of the values of α and β , respectively. Furthermore, let n0 be the
sample size used to obtain first non-zero estimates α̂ and β̂ , i.e., n0 is either the initial
sample size or it is two to power some multiple of the initial sample size.
Once the non-zero estimates are available, the algorithm computes the sample size
required to reach the s confidence level that the true value of min(α, β ) is within a certain interval. For definiteness, let us assume from now on that α̂ < β̂ , which suggests
that min(α, β ) = α. During the execution of the procedure outlined in the following
the inequality may be inverted. If this is the case, the algorithm makes corresponding
change in the consideration of α and β .
Reviving the Two-state Markov Chain Approach
11
The aim is to have a good estimate for α. Notice that the smallest possible initial
value of α̂ is greater than n10 . We refer to n10 as the resolution of estimation. Given the
resolution, one cannot distinguish between values of α in the interval (α̂ − n10 , α̂ + n10 ).
In consequence, if α ∈ (α̂ − n10 , α̂ + n10 ), then the estimated value α̂ should be considered as optimal. Hence, one could use this interval as the one which should contain
the real value with specified confidence level. Nevertheless, although the choice of this
interval usually leads to very good results, as experimentally verified, the results are
obtained at the cost of large samples which make the algorithm stop immediately after
the initialisation phase. Consequently, the computational burden is larger than would
be required by the original algorithm to reach the desired precision specified by r and
s parameters in most cases. In order to reduce this unnecessary overhead, we consider
the interval (α̂ − α̂2 , α̂ + α̂2 ), which is wider than the previous one whenever α̂ > n10 and
leads to smaller sample sizes.
The two-state Markov chain consists of two states 0 and 1, i.e., the two meta states
of the original DTMC. We set α as the probability of making the transition from state 0
to state 1 (denoted as 0 → 1). The estimate α̂ is computed as the ratio of the number of
transitions from state 0 to state 1 to the number of transition from state 0. Let n0,α be
the number of transitions in the sample starting from state 0. Let Xi , i = 1, 2, . . . , n0,α , be
a random variable defined as follows: Xi is 1 if ith transition from meta-state 0 is 0 → 1
and 0 otherwise.
Notice that state 0 is an accessible atom in the terminology of the theory of Markov
chains, i.e., the Markov chain regenerates after entering state 0, and hence the random variables Xi , i = 1, 2, . . . , n0,α , are independent. They are Bernoulli distributed
with parameter α. The unbiased estimate of the population variance from the samn
ple, denoted σ̂ 2 , is given by σ̂ 2 = α̂ · (1 − α̂) · n0,α0,α−1 . Due to independence, σ̂ 2 is also
the asymptotic variance and, in consequence, the sample size that provides the specified confidence level for the estimate of the value of α is given by nα,s (α̂, n0,α ) =
2
n0,α
Φ−1 ( 21 (1+s))
α̂ · (1 − α̂) · n0,α −1 ·
.
α̂/2
β
. Then, given that
The Markov chain is in state 0 with steady-state probability α+β
the chain reached the steady-state distribution, the expected number of regenerations,
n·β
i.e., returns to meta state 0, in a sample of size n is given by α+β
. Therefore, the sample
size used to estimate the value of α with the specified confidence level s is given by
nα = α+β
β · nα,s (α̂, n0,α ). As the real values of α and β are unknown, the estimated
values α̂ and β̂ can be used in the above formula. If the computed nα is bigger than
the current number of transitions n0,α , we extend the trajectory to reach nα transitions
from 0 to 1 and re-estimate the values for α and β using the extended trajectory. We
repeat this process until the computed nα value is smaller than the number of transitions
used to estimate α. In this way, good initial estimates for α and β are obtained and the
original two-state Markov chain approach using the formula for n(α, β ) is run.
Third approach: simple heuristics. When performing the initial estimation of α and
β , we require both the count of transitions from meta state 0 to meta state 1 and the
count of transitions from meta-state 1 to meta state 0 be at least 3. If this condition is not
12
Mizera, Pang and Yuan
satisfied, we proceed by doubling the length of the trajectory. In this way the problem of
reaching the resolution boundary is avoided. Our experiments showed that this simple
approach in many cases led to good initial estimates of the α and β probabilities.
4
Evaluation
We implemented the two-state Markov chain approach with the simple heuristics presented in Section 3 and the Skart method of [11] in the tool ASSA-PBN, which was
specially designed for steady-state analysis of large PBNs [19] (see Section 2.2 for
the theoretical background of PBNs). We verified with experiments that with use of
the simple heuristics, the two-state Markov chain approach could meet the predefined
precision requirement even in the case of an unlucky initial sample size. For the steadystate analysis of large PBNs, applications of these two methods necessitate generation
of trajectories of significant length. To achieve this in an efficient way, we applied the
alias method [20] to sample the consecutive trajectory state. This enables ASSA-PBN,
for example, to simulate 4, 800 steps within 1s for a 2, 000 nodes PBN (state-space of
size 22,000 ).
We choose the Skart method [11] as a reference for the evaluation of the performance of the two-state Markov chain approach. The Skart method is a successor
of ASAP3, WASSP, and SBatch methods, which are all based on the idea of batch
means [11]. It is a procedure for on-the-fly statistical analysis of the simulation output, asymptotically generated in accordance with a steady-state distribution. Usually
it requires an initial sample of size smaller than other established simulation analysis
procedures [11]. In a brief, high-level summary, the algorithm partitions a long simulation trajectory into batches, for each batch computes a mean and constructs an interval
estimate using the batch means. Further, the interval estimate is used by Skart to decide
whether a steady state distribution is reached or more samples are required. For a more
detailed description of this method, see [11].
The Skart method differs in two key points with the two-state Markov chain approach. First, it specifies the initial trajectory length to be at least 1, 280, while for the
two-state Markov chain approach this information is not provided. Second, the Skart
method applies the student distribution for skewness adjustment while the two-state
Markov chain approach makes use of the normal distribution for confidence interval
calculations.
To compare the performance of the two methods, we randomly generated 882 different PBNs using ASSA-PBN. ASSA-PBN can randomly generate a PBN which satisfies
structure requirements given in the form of five input parameters: the node number, the
minimum and the maximum number of predictor functions per node, finally the minimum and maximum number of parent nodes for a predictor function. We generated
PBNs with node numbers from the set {15, 30, 80, 100, 150, 200, 300, 400, 500, 1000,
2000}. We assigned the obtained PBNs into three different classes with respect to the
density measure D: dense models with density 150–300, sparse models with density
around 10, and in-between models with density 50–100. The two-state Markov chain
approach and the Skart method were tested on these PBNs with precision r set to the
Reviving the Two-state Markov Chain Approach
0
k
tTS ≤ tSkart
tSkart ≤ tTS
5
10
15
20
69.83% 55.10% 41.02% 31.03% 25.86%
30.38% 18.68% 11.10% 7.39% 5.49%
25
13
30
22.76% 20.05%
4.43% 3.74%
Table 2: Performance comparison of the Skart and the two-state MC methods. Explanations in the text.
node
number
method
precision
trajectory size
1,000
time cost (s)
trajectory size
2,000
time cost (s)
the two-state Markov chain
0.01
0.005
0.001
35,066
6.19
64,057
20.42
133,803 3,402,637
23.53
616.26
240,662 5,978,309
67.60 1722.86
0.01
37,999
7.02
63,674
20.65
Skart
0.005
0.001
139,672 3,272,940
24.39
590.26
273,942 5,936,060
78.53 1761.05
Table 3: Approximate steady-state analysis of two large PBNs.
values in {10−2 , 5 × 10−3 , 10−3 , 5 × 10−4 , 10−4 , 8 × 10−5 , 5 × 10−5 }. We set ε to 10−10
for the two-state Markov chain approach and s to 0.95 for both methods.
The experiments were performed on a HPC cluster, with CPU speed ranging between 2.2GHz and 3.07GHz. ASSA-PBN is implemented in Java and the initial and
maximum Java virtual machine heap size were set to 503MB and 7.86GB, respectively.
We collected 5263 results with the information on the PBN node number, its density
class, the precision value, the estimated steady-state probabilities computed by the two
methods, and their CPU time costs. The steady-state probabilities computed by the two
methods are comparable in all the cases (data not shown in the paper). For each experimental result i, we compare the time costs of the two methods. Let tTS (i) and tSkart (i)
be the time cost for the two-state Markov chain approach and the Skart method, respectively. We say that the two-state Markov chain approach is by k per cent faster than the
k
. The definition for the Skart method to be faster
Skart method if (tSkartt (i)−t(i)TS (i)) > 100
Skart
than the two-state Markov chain approach is symmetric. In Table 2 we show the percentage of cases in which the two-state Markov chain approach was by k per cent faster
than Skart and vice versa for different k. In general, in about 70% of the results, the
two-state Markov chain was faster than Skart. It is also clear that the number of cases
the two-state Markov chain approach was faster than Skart is larger than in the opposite
case.
We show in Table 3 the trajectory sizes and the time costs for computing steadystate probabilities of two large PBNs using the two-state Markov chain approach and
the Skart method for different precision requirements. The two analysed PBNs consist
of 1, 000 and 2, 000 nodes, which give rise to state spaces of sizes exceeding 10300 and
10600 , respectively.
5
A Biological Case study
A multicellular organism consists of cells that form a highly organised community. The
number of cells in this system is tightly controlled by mechanisms that regulate the cell
14
Mizera, Pang and Yuan
division and the cell death. One of these mechanisms is the programmed cell death, also
referred to as apoptosis: if cells are damaged, infected, or no longer needed, the intracellular death program is activated, which leads to fragmentation of the DNA, shrinkage
of the cytoplasm, membrane changes and cell death without lysis or damage to neighbouring cells. This process is regulated by a number of signaling pathways which are
extensively linked by cross-talk interactions. In [21], a large-scale Boolean network
of apoptosis in hepatocytes was introduced, where the assigned Boolean interactions
for each molecule were derived from literature study. In [22], the original multi-value
Boolean model was cast into the PBN framework: a binary PBN model, so-called ‘extended apoptosis model’ which comprised 91 nodes (state-space of size 291 ) and 102
interactions was constructed, see Figure 2 in Appendix E for the wiring of the PBN
model. With respect to the original multi-value Boolean model of [21], the PBN model
was extended as described in [22]. For example, the possibility of activation of NF-κB
through Caspase 8 (C8*) was included. The model was fitted to steady-state experimental data obtained in response to six different stimulations of the input nodes, see [22]
for details.
As can be seen from the wiring of the network, the activation of complex2 (co2) by
RIP-deubi can take place in two ways: 1) by a positive feedback loop from activated
C8* and P → tBid → Bax → smac → RIP-deubi → co2 → C8*-co2 → C8*, and 2) by
the positive signal from UV-B irradiation (input nodes UV(1) or UV(2)) → Bax →
smac → RIP-deubi → co2. The former to be active requires the stimulation of the type
2 receptor (T2R). The latter way requires complex1 (co1) to be active, which cannot
happen without the stimulation of the TNF receptor-1. Therefore, RIP-deubi can activate co2 only in the condition of co-stimulation by TNF and either UV(1) or UV(2). In
consequence, it was suggested in [22] that the interaction of activation of co2 via RIPdeubi is not relevant and could be removed from the model in the context of modelling
primary hepatocyte. However, due to the problem with efficient generation of very long
trajectories in optPBN toolbox, quantitative analysis was hindered and this hypothesis
could not be verified ([22]).
In this work, we take up this challenge and we quantitatively investigate the relevancy of the interaction of activation of co2 via RIP-deubi. We perform an extensive
analysis in the context of co-stimulation by TNF and either UV(1) or UV(2): we compute long-term influences of parent nodes on the co2 node and the long-run sensitivities
with respect to various perturbations related to specific predictor functions and their
selection probabilities. For this purpose we apply the two-state Markov chain approach
as implemented in our ASSA-PBN tool [19] to compute the relevant steady-state probabilities for the best-fit models described in [22]. Due to the efficient implementation,
the ASSA-PBN tool can easily deal with trajectories of length exceeding 2 × 109 for
this case study.
We consider 20 distinct parameter sets of [22] that resulted in the best fit of the
‘extended apoptosis model’ to the steady-state experimental data in six different stimulation conditions. In [22], parameter estimation was performed with steady-state measurements for the nodes apoptosis, C3ap17 or C3ap17 2 depending on the stimulation
condition considered, and NF-κB. The optimisation procedure used was Particle Swarm
and fit score function considered was the sum of squared errors of prediction (SSE) and
Reviving the Two-state Markov Chain Approach
TNF and UV(1)
IRIP-deubi
Ico1
IFADD
Best fit
Min
Max
Mean
Std
0.2614
0.0000
0.3145
0.2087
0.0735
0.9981
0.9979
0.9988
0.9982
0.0002
0.9935
0.9935
0.9944
0.9937
0.0002
15
TNF and UV(2)
IRIP-deubi
Ico1
IFADD
0.2615
0.0000
0.3146
0.2088
0.0735
0.9980
0.9979
0.9990
0.9982
0.0002
0.9936
0.9936
0.9947
0.9938
0.0003
Table 4: Long-term influences of RIP-duebi, co1, and FADD on co2 in the ‘extended
apoptosis model’ in [22] under the co-stimulation of both TNF and UV(1) or UV(2).
the sum was taken over the three nodes in the six stimulation conditions. We took all
the optimisation results from the three independent parameter estimation runs of [22],
each containing 7500 parameter sets. We sorted them increasingly with respect to the
cost function value obtained during optimisation, removed duplicates, and finally took
the first 20 best-fit parameter sets.
As mentioned above, we fix the experimental context to co-stimulation of TNF and
either UV(1) or UV(2). We note that originally in [21] UV-B irradiation conditions
were imposed via a multi-value input node UV which could take on three values, i.e.,
0 (no irradiation), 1 (300 J/m2 UV-B irradiation), and 2 (600 J/m2 UV-B irradiation).
In the model of [22], UV input node was refined as UV(1) and UV(2) in order to cast
the original model into the binary PBN framework. Therefore, we consider in our study
two cases: 1) co-stimulation of TNF and UV(1) and 2) co-stimulation of TNF and
UV(2). Node co2 has two independent predictor functions: co2 = co1 ∧ FADD or co2 =
(co2)
(co2)
co1 ∧ FADD ∧ RIP-deubi. The selection probabilities are denoted as c1
and c2 ,
respectively. Their values have been optimised in [22].
We start with computing the influences with respect to the steady-state distribution, i.e., the long-term influences on co2 of each of its parent nodes: RIP-deubi, co1,
and FADD, in accordance with the definition in Section 2.2. Notice that the computation of the three influences requires several joint steady-state probabilities to be estimated with the two-state Markov chain approach, e.g., (co1=1,FADD=1,RIP-deubi=0)
or (co1=1,FADD=0). Each probability determines a specific split of the original Markov
chain. For example, in the case of the estimation of the joint steady-state probability for (co1=1,FADD=0), the states of the underlying Markov chain of the apoptosis
PBN model in which co1=1 and FADD=0 constitute meta state 1 and all the remaining states form meta state 0. Therefore, the estimation of influences is computationally
demanding. The summarised results for the 20 parameter sets are presented for the costimulation of TNF and UV(1) or TNF and UV(2) in Table 4. They are consistent across
the different parameter sets and clearly indicate that the influence of RIP-deubi on co2
is small compared to the influence of co1 or FADD on co2. However, the influence of
RIP-deubi is not negligible.
We take the analysis of the importance of the interaction between RIP-deubi and co2
further and we compute various long-run sensitivities with respect to selection probabil(co2)
ity perturbation. In particular, we perturb the selection probability c2
by ±5%, i.e.,
16
Mizera, Pang and Yuan
TNF and UV(1)
(co2)
c2
+5%
Best fit
Min
Max
Mean
Std
0.0003
0.0002
0.0008
0.0005
0.0001
−5%
0.0002
0.0002
0.0008
0.0005
0.0001
TNF and UV(2)
=0
+5%
0.0011
0.0003
0.0014
0.0009
0.0003
0.0002
0.0002
0.0012
0.0004
0.0002
−5%
0.0004
0.0002
0.0007
0.0004
0.0001
=0
0.0011
0.0002
0.0013
0.0009
0.0003
Table 5: Long-run sensitivities w.r.t selection probability perturbations for the ‘extended
apoptosis model’ of [22] under the co-stimulation of TNF and UV(1) or TNF and
UV(2).
we set the new value by multiplying the original value by (1 ± 0.05), and compute in
line with Definition 1 how the joint steady-state distribution for (apoptosis,C3ap17,NFκB) differs from the non-perturbed one with respect to the l1 norm, i.e., ||·||1 . We notice
that the computation of the full steady-state distribution for the considered PBN model
of apoptosis is practically intractable, i.e., it would require the estimation of 291 values. Therefore, we restrict the computations to the estimation of eight joint steady-state
probabilities for all possible combinations of values for (apoptosis,C3ap17,NFκB), i.e.,
the experimentally measured nodes. Each estimation is obtained by a separate run of
the two-state Markov chain approach with the split into meta states determined by the
considered probability as explained above in the case of the computation of long-term
influences. To compare the estimated distributions we choose the l1 norm after [24],
where it is used in the computations of similar types of sensitivities for PBNs to these
defined in Section 2.2. Notice that the l1 norm of the difference of two probability distributions on a finite sample space is twice the total variation distance. The latter is a wellestablished metric for measuring the distance between probability distributions defined
as the maximum difference between the probabilities assigned to a single event by the
(co2)
two distributions (see, e.g., [25]). Additionally, we check the difference when c2
is
(co2)
set to 0 (and, in consequence, c1
is set to 1). The obtained results for the 20 parameter
sets in the conditions of co-stimulation of TNF and UV(1) and co-stimulation of TNF
and UV(2) are summarised in Table 5. In all these cases, the sensitivities are very small.
(co2)
Therefore, the system turns to be insensitive to small perturbations of the value of c2 .
Also the complete removal of the second predictor function for co2 does not cause any
drastic changes in the joint steady-state distribution for (apoptosis,C3ap17,NF-κB).
Finally, we compute the long-run sensitivity with respect to permanent on/off perturbations of the node RIP-deubi in accordance with Definition 2. As before, we consider
the joint steady-state distributions for (apoptosis,C3ap17,NF-κB) and we choose the
l1 -norm. The results, given in Table 6, show that in both variants of UV-B irradiation
the sensitivities are not negligible and the permanent on/off perturbations of RIP-deubi
have impact on the steady-state distribution.
To conclude, all the obtained results indicate that in the context of co-stimulation
of TNF and either UV(1) or UV(2) the interaction between RIP-deubi and co2 plays
a certain role. Although the elimination of the interaction does not invoke significant
Reviving the Two-state Markov Chain Approach
RIP-deubi f. pert.
Best fit
Min
Max
Mean
Std
TNF & UV(1)
TNF & UV(2)
0.3075
0.3097
0.0130
0.0105
0.3595
0.3612
0.2089
0.2105
0.0823
0.0827
17
Table 6: Long-run sensitivities w.r.t permanent on/off perturbations of RIP-deubi for
the ‘extended apoptosis model’ of [22].
changes to the considered joint steady-state distribution, the long-term influence of RIPdeubi on co2 is not negligible and may be important for other nodes in the network other
than apoptosis, nodeC3ap17, or NF-κB.
6
Discussion and Conclusion
In this paper, we focused on two statistical methods for estimating steady-state probabilities of large PBNs: the two-state Markov chain approach and the Skart method.
The Skart method follows a continuous development [11], while the two-state Markov
chain approach was originally introduced by Raftery and Lewis in 1992, and only recently it was explored for the analysis of a relatively large PBN model in [22]. To revive
the application of the two-state Markov chain approach, we propose a few heuristics to
avoid a problem with the size of the initial sample which can lead to biased results. By
extensive experiments, we show that the two-state Markov chain approach outperforms
the Skart method in most cases. In the end, we illustrated the usability of the two-state
Markov chain approach on a realistic biological system.
Our work in the current paper is closely related to statistical model checking [26,27],
a simulation-based approach using hypothesis testing to infer whether a stochastic system satisfies a property. Most current tools for statistical model checking are restricted
for bounded properties which can be checked on finite executions of the system. In
recent year, both the Skart method and the perfect simulation algorithm have been
explored for statistical model checking of steady state and unbounded until properties [28,29], which was considered as a future step of statistical model checking [30].
The perfect simulation algorithm for sampling the steady-state of an ergodic DTMC is
based on the indigenous idea of the backward coupling scheme originally proposed by
Propp and Wilson in [13]. It allows to draw independent samples which are distributed
exactly in accordance with the steady-state distribution of a DTMC. However, due to the
nature of this method, each state in the state space needs to be considered at each step of
the coupling scheme. Of course, a special, more efficient variant of this method exists. If
a DTMC is monotone, then it is possible to sample from the steady-state distribution by
considering the maximal and minimal states only [13,14]. For example, this approach
was exploited in [28] for model checking large queuing networks. Unfortunately, it is
not applicable in the case of PBNs with perturbations. In consequence, the perfect simulation algorithm is only suited for at most medium-size PBNs and large-size PBNs
are out of its scope. Thus, in this paper we have only compared the performance of the
two-state Markov chain approach with the Skart method.
Moreover, in this study we have identified a problem of generating biased results
by the original two-state Markov chain approach and have proposed three heuristics
18
Mizera, Pang and Yuan
to avoid wrong initialisation. Finally, we demonstrated the potential of the two-state
Markov chain approach on a study of a large, 91-node PBN model of apoptosis in hepatocytes. The two-state Markov chain approach facilitated the quantitative analysis of
the large network and the investigation of a previously formulated hypothesis regarding the relevance of the interaction of activation of co2 via RIP-deubi. In the future,
we aim to investigate the usage of the discussed statistical methods for approximate
steady-state analysis in a wide project on systems biology. For instance, we will further apply them to develop new techniques for minimal structural interventions to alter
steady-state probabilities, which will enable the synthesis of optimal control strategies
for large regulatory networks.
Acknowledgment. Experiments presented in this paper were carried out using the HPC
facilities of the University of Luxembourg [31] (http://hpc.uni.lu).
References
1. Shmulevich, I., Dougherty, E.R., Zhang, W.: From boolean to probabilistic boolean networks
as models of genetic regulatory networks. Proceedings of the IEEE 90(11) (2002) 1778–1792
2. Trairatphisan, P., Mizera, A., Pang, J., Tantar, A.A., Schneider, J., Sauter, T.: Recent development and biomedical applications of probabilistic Boolean networks. Cell Communication
and Signaling 11 (2013) 46
3. Albert, R., Othmer, H.G.: The topology of the regulatory interactions predicts the expression
pattern of the segment polarity genes in Drosophila melanogaster. Journal of Theoretical
Biology 223(1) (2003) 1–18
4. Bornholdt, S.: Less is more in modeling large genetic networks. Science 310(5747) (2005)
449–451
5. Wagner, A.: Circuit topology and the evolution of robustness in two-gene circadian oscillators. PNAS 102(33) (2005) 11775–11780
6. Kauffman, S.A.: Homeostasis and differentiation in random genetic control networks. Nature
224 (1969) 177–178
7. Huang, S.: Genomics, complexity and drug discovery: insights from boolean network models
of cellular regulation. Pharmacogenomics 2(3) (2001) 203–222
8. Shmulevich, I., Gluhovsky, I., Hashimoto, R.F., Dougherty, E.R., Zhang, W.: Steady-state
analysis of genetic regulatory networks modelled by probabilistic Boolean networks. Comparative and Functional Genomics 4(6) (2003) 601–608
9. Cowles, M.K., Carlin, B.P.: Markov chain Monte Carlo convergence diagnostics: A comparative review. J. American Statistical Association 91(434) (1996) 883–904
10. Raftery, A.E., Lewis, S.: How many iterations in the Gibbs sampler? Bayesian Statistics 4
(1992) 763–773
11. Tafazzoli, A., Wilson, J.R., Lada, E.K., Steiger, N.M.:
Skart: A skewness- and
autoregression-adjusted batch-means procedure for simulation analysis. In: Proc. 2008 Winter Simulation Conference. (2008) 387–395
12. Norris, J.: Markov Chains. Cambridge UK: Cambridge University Press (1998)
13. Propp, J., Wilson, D.: Exact sampling with coupled markov chains and applications to statistical mechanics. Random Structures & Algorithms 9(1) (1996) 223–252
14. Bušić, A., Gaujal, B., Vincent, J.M.: Perfect simulation and non-monotone markovian
systems. In: Proceedings of the 3rd International Conference on Performance Evaluation
Methodologies and Tools. ValueTools ’08, Brussels, ICST (Institute for Computer Sciences,
Social-Informatics and Telecommunications Engineering) (2008) 27:1–27:10
Reviving the Two-state Markov Chain Approach
19
15. Shmulevich, I., Dougherty, E.R.: Probabilistic Boolean Networks: The Modeling and Control of Gene Regulatory Networks. Philadelphia PA: SIAM Press (2010)
16. Deng, K., Mehta, P.G., Meyn, S.P.: A simulation-based method for aggregating Markov
chains. In: Proceedings of the 48th IEEE Conference on Decision and Control, Shanghai,
IEEE (December 2009) 4710–4716
17. Deng, K., Mehta, P.G., Meyn, S.P.: Optimal Kullback-Leibler aggregation via spectral theory
of Markov chains. IEEE Transactions on Automatic Control 56(12) (2011) 2793–2808
18. Geiger, B.C., Petrov, T., Kubin, G., Koeppl, H.: Optimal KullbackLeibler aggregation via
information bottleneck. IEEE Transactions on Automatic Control 60(4) (2015) 1010–1022
19. Mizera, A., Pang, J., Yuan, Q.: ASSA-PBN: a tool for approximate steady-state analysis of
large probabilistic Boolean networks. In: Proc. 13th International Symposium on Automated
Technology for Verification and Analysis. LNCS, Springer (2015) Available at http://
satoss.uni.lu/software/ASSA-PBN/.
20. Walker, A.: An efficient method for generating discrete random variables with general distributions. ACM Trans. on Mathematical Software 3(3) (1977) 253–256
21. Schlatter, R., Schmich, K., Vizcarra, I.A., Scheurich, P., Sauter, T., Borner, C., Ederer, M.,
Merfort, I., Sawodny, O.: ON/OFF and beyond - a Boolean model of apoptosis. PLOS
Computational Biology 5(12) (2009) e1000595
22. Trairatphisan, P., Mizera, A., Pang, J., Tantar, A.A., Sauter, T.: optPBN: An optimisation
toolbox for probabilistic boolean networks. PLOS ONE 9(7) (2014)
23. Shmulevich, I., Dougherty, E.R., Kim, S., Zhang, W.: Probabilistic Boolean networks: a
rule-based uncertainty model for gene regulatory networks. Bioinformatics 18(2) (2002)
261–274
24. Qian, X., Dougherty, E.R.: On the long-run sensitivity of probabilistic Boolean networks.
Journal of Theoretical Biology 257(4) (2009) 560–577
25. Levin, D.A., Peres, Y., Wilmer, E.L.: Markov Chains and Mixing Times. American Mathematical Society (2009)
26. Younes, H.S., Simmons, R.: Probabilistic verification of discrete event systems using acceptance sampling. In: Proc. 14th Conference on Computer Aided Verification. Volume 2404 of
LNCS., Springer (2002) 223–235
27. Sen, K., Viswanathan, M., Agha, G.: On statistical model checking of stochastic systems. In:
Proc. 17th Conference on Computer Aided Verification. Volume 3576 of LNCS., Springer
(2005) 266–280
28. El Rabih, D., Pekergin, N.: Statistical model checking using perfect simulation. In: Proc.
7th Symposium on Automated Technology for Verification and Analysis. Volume 5799 of
LNCS. (2009) 120–134
29. Rohr, C.: Simulative model checking of steady state and time-unbounded temporal operators.
Transactions on Petri Nets and Other Models of Concurrency 8 (2013) 142–158
30. Legay, A., Delahaye, B., Bensalem, S.: Statistical model checking: An overview. In: Proc.
1st Conference on Runtime Verification. Volume 6418 of LNCS., Springer (2010) 122–135
31. Varrette, S., Bouvry, P., Cartiaux, H., Georgatos, F.: Management of an academic HPC
cluster: The UL experience. In: Proc. Conference on High Performance Computing & Simulation, IEEE CS (2014) 959–967
20
A
Mizera, Pang and Yuan
Derivation of the number of “burn-in” iterations
Let {Zt }t≥0 be a discrete-time two-state Markov chain as given in Figure 1b. Zt has
the value 0 or 1 if the system is in state 0 or state 1 at time n, respectively. The transition probabilities satisfy 0 < α, β < 1 and the transition matrix for this chain has the
following form
1−α α
P=
.
β 1−β
Matrix P has two distinct eigenvalues: 1 and λ = (1 − α − β ). Notice that |λ | < 1.
The chain is ergodic and the unique steady-state distribution is π = [π0 π1 ] =
β
α
]. Let Eπ (Zt ) denote the expected value of Zt for any fixed t ≥ 0, with re[ α+β α+β
α
spect to the steady-state distribution π. We have that Eπ (Zt ) = α+β
.
The m-step transition matrix can be written, as can be checked by induction, in the
form
λm
α −α
π π
·
Pm = 0 1 +
,
(1)
π0 π1
α + β −β β
where λ is the second eigenvalue of P.
Suppose we require m to be such that the following condition is satisfied
P[Zm = 0 | Z0 = j] P[Zm = 1 | Z0 = j] − π0 π1 < ε ε
(2)
for some ε > 0. For any vector v = [v1 v2 . . . vn ]T ∈ Rn we use |v| to denote [|v1 | |v2 |
. . . |vn |]T , where T is the transposition operator. If e0 = [1 0] and e1 = [0 1], then for
j ∈ {0, 1} we have that
P[Zm = 0 | Z0 = j] P[Zm = 1 | Z0 = j] = e j Pm .
(3)
With (1) and (3), condition (2) can be rewritten as
ej
λm
α −α
π0 π1
·
− π0 π1 < ε ε .
+
π0 π1
α + β −β β
For j = 0 and j = 1 the above simplifies to
λm
· α −α < ε ε
α +β
and
λm
· −β β < ε ε ,
α +β
respectively. Therefore, it is enough to consider the following two inequalities
λ mα
<ε
α +β
and
λ mβ
< ε,
α +β
which, since α, β > 0, can be rewritten as
|λ m | <
ε(α + β )
α
and
|λ m | <
ε(α + β )
.
β
Reviving the Two-state Markov Chain Approach
21
Equivalently, m has to satisfy
|λ m | <
ε(α + β )
.
max(α, β )
By the fact that |λ m | = |λ |m this can be expressed as
|λ |m <
ε(α + β )
.
max(α, β )
Then, by taking the logarithm to base 10 on both sides3 , we have that
ε(α + β )
m · log (|λ |) < log
max(α, β )
and in consequence, since |λ | < 1 and log |λ | < 0,
ε(α+β )
log max(α,β
)
.
m>
log (|λ |)
B
Derivation of the sample size
By the Law of Large Numbers for irreducible positive recurrent Markov chains Z n →
n
π1 a. s. with n → ∞, where Z n = 1n ∑t=1
Zt . Now, by a variant of the Central Limit
Theorem for non-independent random variables4 , for n large, Z n is approximately norα
2 = 1 αβ (2−α−β ) , see
mally distributed with mean π1 = α+β
and asymptotic variance σas
n (α+β )3
Section C for the derivation of the asymptotic variance. Let X be the standardised Z n ,
i.e.,
Z n − π1
√ .
X=
σas / n
If follows that X is normally distributed with mean 0 and variance 1, i.e., X ∼ N(0, 1).
Now, we require n to be such that the condition P[π1 − r ≤ Z n ≤ π1 + r] = s is
satisfied for some specified r and s. This condition can be rewritten as
P[−r ≤ Z n − π1 ≤ r] = s,
and further as
3
4
√
√
n Z n − π1
n
√
P[−r ·
≤
≤ r·
] = s,
σas
σas / n
σas
In fact, by the formula for change of base for logarithms, the natural logarithm (ln), the logarithm to base 2 (log2 ), or a logarithm to any other base could be used to calculate m instead of
log. Notice that m does not depend on the choice of the base of the logarithm!
Notice that the random variables Zt , Zt+1 which values are consecutive states of a trajectory
are correlated and are not independent.
22
Mizera, Pang and Yuan
which is
√
√
n
n
P[−r ·
≤ X ≤ r·
] = s.
σas
σas
Since X ∼ N(0, 1) and N(0, 1) is symmetric around 0, it follows that
P[0 ≤ X ≤ r ·
and
√
n
s
]=
σas
2
√
n
1 s
1
P[X ≤ r ·
] = + = (1 + s).
σas
2 2 2
Let Φ(·) be the standard normal cumulative distribution function. Then the above can
be rewritten as
√
n
1
) = (1 + s).
Φ(r ·
σas
2
Therefore, if we denote the inverse of the standard normal cumulative distribution function with Φ −1 (·), we have that
r·
√
n
1
= Φ −1 ( (1 + s)).
σas
2
In consequence,
n=
C
2
σas
r
Φ −1 ( 12 (1+s))
2 =
αβ (2−α−β )
(α+β )3
r
Φ −1 ( 12 (1+s))
2 .
Derivation of the asymptotic variance
√
d
By the Central Limit Theorem for stationary stochastic processes5 n(Z n − π1 ) →
−
2
2
N(0, σas ) as n → ∞, where σas is the so-called asymptotic variance given by
∞
2
σas
= Varπ (Z j ) + 2 ∑ Covπ (Z j , Z j+k )
(4)
k=1
and Varπ (·) and Covπ (·) denote the variance and covariance with respect to the steady2 . First, observe that E (Z ·
state distribution π, respectively. We proceed to calculate σas
π n
α
Zn+1 ) = α+β (1 − β ): Zn · Zn+1 6= 0 if and only if the chain is state 1 at time n and
remains in 1 at time n + 1, i.e., Zn = Zn+1 = 1. The probability of this event at steady
α
state is α+β
(1 − β ). Then, by the definition of covariance, we have that the steady-state
5
After discarding the ‘burn-in’ part of the trajectory, we can assume that the Markov chain in
a stationary stochastic process.
Reviving the Two-state Markov Chain Approach
23
covariance between consecutive random variables of the two-state Markov chain, i.e.,
Covπ (Zn , Zn+1 ) is
Covπ (Zn , Zn+1 ) = Eπ [(Zn − Eπ (Zn ))(Zn+1 − Eπ (Zn+1 ))]
α
α
)(Zn+1 −
)
= Eπ (Zn −
α +β
α +β
α
α2
= Eπ Zn Zn+1 −
(Zn + Zn+1 ) +
α +β
(α + β )2
= Eπ (Zn Zn+1 ) −
α
α2
(Eπ (Zn ) + Eπ (Zn+1 )) +
α +β
(α + β )2
α(1 − β )
α2
α2
−2
+
α +β
(α + β )2 (α + β )2
αβ (1 − α − β )
.
=
(α + β )2
=
Further, we have that Varπ (Zn ) = π0 · π1 =
αβ
(α+β )2
(variance of the Bernoulli distribu-
tion) and it can be shown that Covπ (Zn , Zn+k ) = (1 − α − β )k · Varπ (Zn ) for any k ≥ 1.
Now, according to Equation (4), we have
∞
2
σas
= Varπ (X j ) + 2 ∑ Covπ (X j , X j+k )
k=1
∞
αβ
αβ
=
+
2
(1 − α − β )k ·
∑
(α + β )2
(α
+ β )2
k=1
=
∞
αβ
2αβ
+
·
∑ (1 − α − β )k
(α + β )2 (α + β )2 k=1
αβ
2αβ
1−α −β
+
·
(α + β )2 (α + β )2
α +β
αβ (2 − α − β )
=
.
(α + β )3
=
In consequence, Z n is approximately normally distributed with mean
α
α+β
and variance
1 αβ (2−α−β )
n (α+β )3 .
D
Derivations for the pitfall avoidance heuristics
We start with analysing the minimum values n(·, ·) can attain. The function is considered on the domain D = (0, 1] × (0, 1] and, as mentioned before, the estimated values of
α and β are within the range [ n10 , 1]. Computing the partial derivatives, equating them to
zero, and solving for α and β yields α = −β , which has no solution in the considered
domain. Hence, the function has neither local minimum nor maximum on D. Let us fix
24
Mizera, Pang and Yuan
β for a moment and consider n(α, β ) as a function of α. We denote it as nβ (α). By
differentiating with respect to α, we obtain
∂
1 β α2 − β 2 − 4 α + 2 β
r2
nβ (α) =
, where cr,s =
2 .
4
∂α
cr,s
(α + β )
Φ−1 1 (1 + s)
2
p
By equating topzero and solving for α we get two solutions: α1 = 2 − β 2 − 2β + 4
and α2 = 2 + β 2 − 2β + 4. Since the second solution is always greater than 1 on the
(0, 1] interval, only the first solution is valid. The sign of the second derivative of nβ (α)
with respect to α at α1 is negative. This shows that for any fixed β , nβ (α) grows on the
interval [ n10 , α1 ], attains its maximum at α1 and decreases on the interval [α1 , 1]. Notice
that n is symmetric, i.e., n(α, β ) = n(β , α). Thus the minimumvalue
for
n could
attain
1 1
1
α and β estimated from a sample of size n0 is given by min n n0 , n0 , n n0 , 1 .
After evaluating n we get
1 1
(n0 − 1) · n0
n0 − 1
1
n
,
,1 =
.
=
and
n
n0 n0
4 cr,s
n0
cr,s · (1 + n0 )3
Now, to avoid the situation where the initial estimates of α and β lead to n(α, β ) <
2 n0 , it is enough to make sure that given r and s the following condition is satisfied:
min(n( n10 , n10 ), n( n10 , 1)) > 2 n0 . This can be rewritten as
(8 cr,s − 1) n0 + 1 ≤ 0
2 cr,s n30 + 6 cr,s n20 + (6 cr,s − 1) n0 + 2 cr,s + 1 ≤ 0
Both inequalities can be solved analytically. Given that n0 > 0, the solution of the first
inequality is
(
n0 ∈ [− 8·cr,s1 −1 , ∞)
cr,s < 18
(5)
n0 ∈ 0/
cr,s > 18 .
The solution of the second inequality is more complicated, but can be easily obtained
with computer algebra system software (e.g., MapleTM ). In Table 1 we present some
solutions for a number of values for r and s.
Reviving the Two-state Markov Chain Approach
E
25
The Boolean model of apoptosis
Fig. 2: The wiring of the probabilistic Boolean model of apoptosis originally introduced
in [22].
| 5cs.CE
|
A randomized polynomial kernel for Subset Feedback Vertex Set
Eva-Maria C. Hols and Stefan Kratsch
University of Bonn, Germany, {hols,kratsch}@cs.uni-bonn.de
arXiv:1512.02510v1 [cs.DS] 8 Dec 2015
9th December 2015
Abstract
The subset feedback vertex set problem generalizes the classical feedback vertex set problem and asks, for a given undirected graph G = (V, E), a set S ⊆ V , and an
integer k, whether there exists a set X of at most k vertices such that no cycle in G − X
contains a vertex of S. It was independently shown by Cygan et al. (ICALP ’11, SIDMA
’13) and Kawarabayashi and Kobayashi (JCTB ’12) that subset feedback vertex set
is fixed-parameter tractable for parameter k. Cygan et al. asked whether the problem also
admits a polynomial kernelization.
We answer the question of Cygan et al. positively by giving a randomized polynomial
kernelization for the equivalent version where S is a set of edges. . In a first step we show that
edge subset feedback vertex set has a randomized polynomial kernel parameterized
by |S| + k with O(|S|2 k) vertices. For this we use the matroid-based tools of Kratsch and
Wahlström (FOCS ’12) that for example were used to obtain a polynomial kernel for smultiway cut. Next we present a preprocessing that reduces the given instance (G, S, k)
to an equivalent instance (G′ , S ′ , k ′ ) where the size of S ′ is bounded by O(k 4 ). These two
results lead to a polynomial kernel for subset feedback vertex set with O(k 9 ) vertices.
1
Introduction
In the subset feedback vertex set (subset fvs) problem we are given an undirected graph
G = (V, E), a set of vertices S ⊆ V , and an integer k, and have to determine whether there
is a set X of at most k vertices that intersects all cycles that contain at least one vertex of S.
Clearly, because we can choose S = V , this is a generalization of the well-studied feedback
vertex set (fvs) problem where, given G and k, we have to determine whether some set X
of at most k vertices intersects all cycles in G. feedback vertex set has been extensively
studied in parameterized complexity: It is known to be fixed-parameter tractable (FPT) with
parameter k, i.e., solvable in time f (k) · |V |c , and after a series of improvements the fastest
known algorithms take deterministic time O∗ (3.619k ) [10] and randomized time O∗ (3k ) [2]. It
is also known to admit a polynomial kernelization [1], i.e., there is an efficient algorithm that
reduces any instance (G, k) of fvs to an equivalent instance of size polynomial in k; the best
known kernelization creates an equivalent instance with O(k2 ) vertices [19].
In 2011, Cygan et al. [3] and Kawarabayashi and Kobayashi [9] independently showed that
subset fvs is FPT. The algorithm of Cygan et al. runs in time 2O(k log k) nO(1) , while the one of
Kawarabayashi and Kobayashi runs in time O(f (k) · n2 m). Wahlström [21] then gave the first
single-exponential algorithm with running time 4k · nO(1) ; an algorithm with subexponential
dependence on k is ruled out under the Exponential-Time Hypothesis (e.g., because subset
fvs generalizes vertex cover). More recently, Lokshtanov et al. [12] gave algorithms with
deterministic time 2O(k log k) · (n + m) and randomized time O(25.6k · (n + m)).
1
Cygan et al. [3] ask whether the subset fvs problem also admits a polynomial kernelization
and suggest that the matroid-based tools of Kratsch and Wahlström [11] could be applicable.
The latter work uses representative sets of independent sets in matroids to obtain, amongst
others, polynomial kernels for s-multiway cut and deletable terminal multiway cut
(dtmwc) with O(ks+1 ) and O(k3 ) vertices, respectively. In multiway cut we are given a
graph G = (V, E), a set T ⊆ V of terminals, and an integer k and have to determine whether
deletion of at most k non-terminal vertices separates all terminals. In s-multiway cut the
terminal set has size at most s, and in dtmwc we are also allowed to delete terminals (which
is essentially the same as restricting terminals to be degree one).
Interestingly, Cygan et al. [3] also provide a polynomial-time reduction from multiway cut
to subset fvs that does not change the parameter value and, hence, is known to imply that
subset fvs is at least as hard as multiway cut regarding existence of polynomial kernels.
Accordingly, multiway cut would be the natural next target problem for attempting to find a
polynomial kernelization (after s-multiway cut and deletable terminal multiway cut).
It appears, however, that the reduction of Cygan et al. is from deletable terminal multiway
cut rather than from the more general multiway cut, and it is not obvious whether similar
ideas could yield a reduction from multiway cut to subset fvs.
Our work. We apply the matroid-based tools of Kratsch and Wahlström [11] and develop
a randomized polynomial kernelization that reduces instances (G, S, k) of subset fvs to equivalent instances with at most O(k9 ) vertices; this is our main result. Similarly to Cygan et al. [3]
we also work on edge subset fvs where S is a set of edges of G and X needs to intersect all
cycles that contain at least one edge of S; edge subset fvs and subset fvs are equivalent [3].
The result is obtained in two parts.
In the first part (Section 3) we establish a randomized polynomial kernelization for edge
subset fvs parameterized by |S|+ k that reduces to equivalent instances with at most O(|S|2 k)
vertices. Note that nontrivial instances have k < |S| since one could otherwise remove S by
deleting one endpoint of each edge in S. Thus, parameterization by |S| suffices, but O(|S|2 k)
gives a tighter overall bound than O(|S|3 ).
At high level, this part is similar to the polynomial kernelization for deletable terminal
multiway cut. We show that certain solutions X, later called dominant solutions, allow
particular path packings in the underlying graph G. For deletable terminal multiway
cut this is achieved by a fairly simple replacement argument for solutions X that are not
sufficiently well connected to connected components of G − X. For edge subset fvs the
endpoints T = V (S) of edges in S can be regarded as terminals, but this gives a different
separation property: Solutions X need not generate many connected components in G − X
since only S-cycles need to be prevented, and components may contain many vertices of T .
Rather, in G − X there must be a tree-like (or forest-like) structure with components without
S-edges playing the role of nodes and with edges given by S. Nevertheless, using the tree-like
structure, a replacement argument can be found, implying that dominant solutions must create
many components in (G − X) − S containing vertices of T and be well connected to them.
This allows to set up a gammoid on G − S with sources T and apply, as in [11], a result of
Lovász [13] (made algorithmic by Marx [15]) on representative sets in (linear) matroids that
is then guaranteed to generate a superset of X. Randomization is only needed to generate a
matrix representation for the gammoid.
In the second part (Section 4) we give a (deterministic) polynomial-time preprocessing that,
given an instance (G, S, k) of edge subset fvs, returns an equivalent instance (G′ , S ′ , k′ ) with
k′ ≤ k and |S ′ | ∈ O(k4 ). Together with the randomized kernelization from the first part this
implies the claimed randomized kernelization to O(k9 ) vertices.
2
A reduction of the number of S-edges is also a crucial ingredient in the FPT algorithm
for edge subset fvs by Cygan et al. [3]. They achieve |S| ∈ O(k3 ), but it is in a slightly
more favorable setting: Using iterative compression, it suffices to solve the task of finding a
solution X ′ of size k when given a solution X of size k + 1. (This is well known in parameterized
complexity, and we prefer not to repeat it here.) Considering some unknown solution X ′ of size
k, one can guess the intersection D of X ′ with X, by trying all O(2k+1 ) possibilities. For the
correct guess D = X ′ ∩ X, the remaining problem is to find for (G − D, S \ D, k − |D|) a solution
Z ′ of size at most k − |D| that is disjoint from Z = X \ D, since Z ′ = X ′ \ D would be such a
solution; here S \ D denotes the set of edges in S with no endpoint in D. Cygan et al. make
the nice observation that the guessing also allows to assume that there is no other solution X ′
with an even larger intersection with X.
In contrast, we cannot afford to run iterative compression for a kernelization to get a starting
solution of size k + 1 and, as is common, we have to start with an approximate solution Z, which
can be assumed to be of size at most 8k using an 8-approximation algorithm of Even et al. [6].
The idea of guessing the intersection of an optimal solution with Z is infeasible regarding
both time and the number of created instances. Thus, while several structures like z-flowers
or disjoint x,y-paths containing S-edges appear in both approaches, many things have to be
handled differently. For example, having k + 2 disjoint x,y-paths containing S-edges for x, y ∈ Z
implies that one of x and y must be in every solution of size k; Cygan et al. can stop here because
the solution would not be disjoint from Z; we need to instead store the information about x
and y to later detect S-edges that can be safely removed. Like Cygan et al., we also use Gallai’s
A-path Theorem but we avoid the 2-expansion lemma by using the properties of a blocking set
of size at most 2k differently. (Such a blocking set can be found if certain flowers of order k + 1
do not exist, using Gallai’s A-path Theorem.) Cygan et al. compute a blocking set B of size
at most 3k to find an F -flower of order |X| (with F ⊆ V outer-abundant; see [3, Definition
3.4]) under the assumption that certain F -flowers of order k + 1 do not exist and show that
there exists a solution that contains X (under the assumption that there exists a solution that
is disjoint from F ). We cannot assume that our solution is disjoint from F and have to take
another approach. Moreover, we observe that z-flowers can be found via matroid parity on an
appropriate gammoid.1
2
Preliminaries
We use standard graph notation, mostly following Diestel [5]. All graphs are undirected and
may contain multi-edges and loops; accordingly, they may contain cycles of length one and
two (formed by loops and multi-edges, respectively.) An edge e ∈ E is called a bridge if
(V, E \ {e}) has more connected components than G. For a set X ⊆ V , let G[X] denote
the subgraph of G induced by X and let NG (X) denote the neighborhood of X in G, i.e.,
NG (X) = {v ∈ V \ X | ∃u ∈ X : {u, v} ∈ E}. Given two sets X, Y ⊆ V , by E(X, Y ) we denote
the set of edges that have one endpoint in X and one endpoint in Y . For a set E ′ ⊆ E of edges
let V (E ′ ) be the set of vertices that are incident with at least one edge in E ′ . For X ⊆ V and
F ⊆ E we shorthand G − X for G[V \ X] and G − F for (V (G), E(G) \ F ); if X = {x} then we
may also write G − x instead of G − {x}. Note that the graph (G − X) − F is the same graph
as the graph (G − F ) − X and we will drop the parentheses.
For A ⊆ V a path with endpoints in A and internal vertices not in A is called an A-path.
The following theorem about A-paths was already used by Cygan et al. [3] for subset fvs and
in the quadratic kernelization for feedback vertex set by Thomassé [19].
1
The latter is deterministic by applying a specialized matroid parity algorithm due to Tong et al. [20].
3
Theorem 1 (Gallai [8]). Let A ⊆ V and k ∈ N. If the maximum number of vertex-disjoint
A-paths is strictly less than k + 1, then there exists a set B ⊆ V of at most 2k vertices that
intersect every A-path.
In particular it is possible to find either (k + 1)-disjoint A-paths or a set B that intersects
all A-paths in polynomial time. This follows from Schrijver’s proof of Gallai’s theorem [18].
Let (G, S, k) be an instance of the edge subset fvs problem. We call a cycle C an S-cycle,
if at least one edge of S is contained in C. Let x be a vertex of V . A set {C1 , C2 , . . . , Ct }
of S-cycles that contain x is called an x-flower of order t, if the sets of vertices Ci \ {x} are
pairwise disjoint. Note that if there exists a x-flower of order at least k + 1, then the vertex x
must be in every solution for (G, S, k), if one exists. A set B ⊆ V \ {x} of size t is called an
x-blocker of size t, if each S-cycle through x also contains at least one vertex of B.
Parameterized complexity. A parameterized problem is a language Q ⊆ Σ∗ × N, where
Σ is any finite set. The second component of an instance (x, k) is called the parameter. We say
that a parameterized problem Q is fixed-parameter tractable (FPT) if there exists a computable
function f : N → N and an algorithm A that on input of (x, k) ∈ Q × Σ∗ takes time at most
f (k) · |x|O(1) and correctly decides whether (x, k) ∈ Q. A kernelization of a parameterized
problem Q is an algorithm K that on input of (x, k) ∈ Σ∗ × N takes time polynomial in
|x| + k and returns an equivalent instance (x′ , k′ ) ∈ Σ∗ × k with |x′ | + k′ ≤ h(k), where h is
a computable function. The function h is called the size of the kernel. We say that K is a
polynomial kernelization if h(k) ∈ O(kc ) for some constant c. The polynomial kernelization
obtained in this paper is randomized, which means that there is a small chance for the reduced
instance to not be equivalent to the input. The error probability can be made exponentially
small in the input size without increasing the size of the kernelization. Similarly to previous
work [11], the only source for error is the need to compute a matrix representation for a particular
matroid (preliminaries on matroids follow below).
Matroids, gammoids, and representative sets. A matroid M = (U, I) consists of
a finite set U and a family I of subsets of U , called independent sets, fulfilling the following
properties: (i) ∅ ∈ I; (ii) if X ⊆ Y and Y ∈ I then also X ∈ I; and (iii) if X, Y ∈ I with
|X| < |Y | then there exists y ∈ Y \ X such that X ∪ {y} ∈ I. The rank of of a matroid M ,
denoted by r(M ), is the size of the largest independent set of the matroid M .
Let A be a matrix over an arbitrary field F . Let U be the set of columns of A and let
I be the family of all sets X ⊆ U of columns that are linearly independent over F . Then
M = (U, I) is a matroid, called the linear matroid or vector matroid of A, and we say that A
represents M . If M = (U, I) is representable over some field, then it is also representable by an
r(M ) × |U | matrix; by Gaussian elimination we can always reduce a representing matrix for M
to one with r(M ) many rows (cf. [15]). Let M1 = (U1 , I1 ) and M2 = (U2 , I2 ) be two matroids
with U1 ∩ U2 = ∅. The direct sum M1 ⊕ M2 is a matroid over U = U1 ∪ U2 with independent
sets I = {X ⊆ U | X ∩ U1 ∈ I1 , X ∩ U2 ∈ I2 }. If A1 and A2 represent the two matroids over
the same field F , then matrix A = diag(A1 , A2 ) represents M1 ⊕ M2 .
Let G = (V, E) be a graph that may have both directed and undirected edges and let S ⊆ V .
A set T ⊆ V is linked to S if there exist |T | vertex-disjoint paths from S to T . Thus every
vertex in T is endpoint of a different path from S. It holds that M = (U, I), where U ⊆ V and
I contains all sets T ⊆ U that are linked to S in G, is a matroid [17]. The matroid M is also
called the gammoid on G with sources S and ground set U ; if U = V then M is also called a
strict gammoid. Marx [15] gave a randomized polynomial-time procedure for finding a matrix
representation of a strict gammoid. The error probability can be made exponentially small in
4
the size of the graph. (This is the only source of randomness and error in our kernelization.) A
matrix representation for a gammoid for graph G = (V, E) with ground set U ( V and sources
S can be obtained from one for the strict gammoid for G and S by simply deleting columns
corresponding to elements of V \ U .
Let A, B be independent sets in a matroid. We say that A extends B if A ∩ B = ∅ and A ∪ B
is again an independent set. Note that from the independence of A∪B follows the independence
of A and B due to the second matroid property.
Definition 1. Let M = (U, I) be a matroid, let A ⊆ I, and let q ∈ N. A set A′ ⊆ A is
q-representative for A if for every independent set B of size at most q there is a set A ∈ A that
extends B if and only if there is also a set A′ ∈ A′ that extends B.
Observe that if A′ is q-representative for A and there exists a set A ∈ A that uniquely
extends some given independent set I of size at most q, then this implies that A ∈ A′ .
The following theorem of Lovász [13] proves that for any linear matroid there exist small
representative sets. It was made algorithmic by Marx [15] and, thus, permits to find representative sets in polynomial time when given a matrix representation of the matroid. A faster
algorithm for this task was developed recently by Fomin et al. [7].
Lemma 1 (Lovász [13], Marx [15]). Let M be a linear matroid of rank q + p,
and let T =
q+p
{I1 , I2 , . . . , It } be a collection of independent sets, each of size p. If |T | > p , then there is
a set I ∈ T such that T \ {I} is q-representative for T . Furthermore, given a representation A
of M , we can find such a set I in f (q, p) · (kAkt)O(1) time.
Given a gammoid M we can compute in randomized polynomial-time a representation of
the gammoid. Together with Theorem 1 it follows that given a gammoid M and a collection
T = {I1 , . . . , It } of independent sets,
each of size p, we can find in randomized polynomial time
that is q-representative for T .
a set T ′ ⊆ T of size at most q+p
p
3
Randomized polynomial kernelization for parameter |S| + k
In this section we present a randomized polynomial kernelization for edge subset fvs parameterized by |S| + k. Because deletion of one endpoint of each edge in S always constitutes
a feasible solution, nontrivial instances have |S| > k. Thus, our kernelization also works for
parameter |S| alone. However, to achieve a better bound for edge subset fvs parameterized
by k only it is beneficial to give the kernel size in terms of |S| and k rather than |S| alone.
We use representative sets of independent sets of matroids to obtain a kernel of size O(|S|2 k).
Our approach is similar to the kernelization of deletable terminal multiway cut(k) [11].
As in that paper we construct path packings such that certain vertices can be shown to be in
a representative set. Note that, unlike for multiway cut-type problems, a solution X ⊆ V will
not necessarily create many connected components. Rather, as used also in the FPT algorithm
of Cygan et al. [3], it creates a particular tree-like structure in G − X. Nevertheless, endpoints
of edges in S, denoted T := V (S), will play the role of terminals that need to be separated in a
certain way; hence a vertex x in T is called a terminal. We will focus on the graph G − S, i.e.,
with edges of S deleted, in which a solution X creates a grouping of (not deleted) terminals into
connected components. The structure of these components will be crucial for a replacement
argument (Lemma 3) that leads to the required path packing; this constitutes one of the key
arguments for our result.
The kernelization consists of four steps. In the first step we show that if an instance is
YES then there exists a solution X with a certain path packing from T to X. Then we define
5
an appropriate gammoid to find in a next step a representative set of size O(|S|2 k) which is
(essentially) a superset of X using Lemma 1. Finally we explain how to reduce the graph G,
using the superset of the last step, to obtain an equivalent instance of edge subset fvs.
Analyzing solutions. Let (G, S, k) be a yes-instance of edge subset fvs (k + |S|). We
say that a solution X for (G, S, k) is dominant, if it has minimum size and contains a maximal
number of vertices from T among solutions of minimum size. The vertices in X ∩ T correspond
to endpoints of edges in S that we delete and the vertices in X0 = X \ T block all x-y paths
with {x, y} ∈ S0 = {e ∈ S | e ∩ X = ∅}, except the one that consists of the edge {x, y}. We
show that X is linked to T in a strong sense, with vertices of X0 playing a special role.
Lemma 2. Let X be a dominant solution for (G, S, k) and x any vertex in the set X0 = X \ T .
There exist |X| + 2 paths from T to X in G − S that are vertex-disjoint except for three paths
ending in vertex x. Moreover, the paths can be chosen in such a way that each connected
component of G − X − S is intersected by at most one path.
We use Hall’s Theorem and the lemma below to prove this. For this purpose we use the two
graphs G − X and G − X − S which simplify the analysis of a dominant solution. We call a
connected component K of G − X − S interesting if it contains a terminal, i.e., if T ∩ V (K) =
(T \ X) ∩ V (K) 6= ∅, and we say that x ∈ X0 sees an interesting component K if x is adjacent
to a vertex of K in G. We extend this definition by saying that Y ⊆ X0 sees an interesting
component K if at least one vertex y ∈ Y sees K.
Lemma 3. If X is a dominant solution then every nonempty set Y ⊆ X0 sees at least |Y | + 2
interesting components of G − X − S.
Proof. Assume for contradiction that there exists a nonempty set Y ⊆ X0 that sees at most
|Y | + 1 interesting components of G − X − S. Let Ci denote the set of interesting components
of G − X − S seen by Y , and let Co denote the other components seen by Y . We will show that
there is an alternative solution X ′ = (X \ Y ) ∪ Y ′ that is smaller than X or that contains more
vertices of T , contradicting the choice of X as a dominant solution. To this end, let us consider
the graphs G − X and G − (X \ Y ) (in part repeating things that have been said earlier to get
a self-contained proof).
In G − X the components of G − X − S may be connected by edges of S and form a
tree-structure with components playing the role of vertices and edges of S whose endpoints are
not deleted being the edges of the tree: (We say tree-structure, but a forest of components,
connected by S-edges, is also fine.) There can be no cycles in this tree-structure because they
would give rise to S-cycles in G − X. Moreover, any other set X ′ of size at most k such that
G − X ′ consists of components without S-edges that are connected in a tree-like manner by
S-edges is also a valid solution. Note that non-interesting components of G − X − S are isolated
in G − X because they do not contain vertices of T , i.e., no endpoints of S-edges, so they cannot
be incident with S-edges in G − X.
In G − (X \ Y ) − S the components in Ci and Co may form larger combined components
because we do not delete the vertices in Y ; let C ′ denote the set of these components. Crucially,
because Y ⊆ X0 = X \ T , there are no additional vertices of T , i.e., T \ X = T \ (X \ Y ). Thus,
in G − (X \ Y ) the set of S-edges incident with components in C ′ is the same as the S-edges
incident with Ci in G − X; recall that no S-edges are incident with components in Co in G − X.
Note that, in general, G − (X \ Y ) will not have the tree-structure: In comparison to G − X
we are not deleting vertices of Y , which corresponds to merging some components in Ci ∪ Co .
This may lead to components in C ′ that are incident with both endpoints of some S-edges (the
equivalent of loops) and it may also create other (longer) cycles. We will see that deleting at
6
most |Y | edges of S, i.e., deleting a set Y ′ of at most |Y | endpoints of S-edges, will suffice to
get the tree-structure, making (X \ Y ) ∪ Y ′ a valid solution.
Consider a component C ′ ∈ C ′ of G − (X \ Y ) − S that fully contains all vertices of some
components Ci1 , . . . , Cia ∈ Ci and Co1 , . . . , Cob ∈ Co ; additionally it may contain vertices of Y .
(The fact that we must have full containment follows directly by comparing deletion of X from
G − S with deletion of X \ Y from G − S.) In G − (X \ Y ) the component C ′ may be incident
with S-edges and thus be part of a larger component C + ; we want to see that deleting (one
endpoint each of) at most a − 1 S-edges from C + suffices to get the tree-structure.
In G − X instead of component C + we may have several separate components because we
additionally delete the vertices of Y . Since Y sees only components in Ci ∪ Co there are at most
a + b separate components “created” from C + by deleting Y since these are all components
contained in C + that are seen by Y . Recall that components in Co are isolated in G − X and
contain no vertices of T and, thus, they do not contribute any S-edges to C + . It remains to
consider the components Ci1 , . . . , Cia that are contained in C + .
Assume first that all components Ci1 , . . . , Cia are part of a single connected component in
G − X. (Recall that they are connected components of G − X − S but may be connected by Sedges in G − X.) Thus, they are part of a single tree of components (connected by S-edges) and
not deleting Y corresponds to merging a vertices in this tree into a single one. If the tree had c
components and, thus, c − 1 S-edges then we obtain c − a + 1 components that are connected
by c − 1 S-edges in G − (X \ Y ). (Recall that Y contains no endpoints of S-edges.) It therefore
suffices to delete (c − 1) − ((c − a + 1) − 1) = a − 1 S-edges, i.e., to delete one endpoint of each
of a − 1 S-edges, to obtain the tree-structure. (Not any a − 1 edges are ok but we can keep
any c − a S-edges spanning the c − a + 1 components and delete the (c − 1) − (c − a) = a − 1
remaining S-edges.)
In general, the components Ci1 , . . . , Cia may be part of several different connected components in G − X. Nevertheless, this still means that we have a cycle-free structure of components
(seen as vertices) connected by S-edges. If overall the cycle-free structure has c components
then, being cycle-free, it has at most c − 1 S-edges. Thus, merging yields c − a + 1 components
connected by at most c − 1 S-edges and removing at most a − 1 S-edges suffices.
Overall, we get that a component C ′ ∈ C ′ that fully contains a interesting components from
Ci requires at most a − 1 vertex deletions of endpoints of S-edges to obtain the tree-structure.
Since Y sees at most |Y | + 1 such components, the worst case is achieved by a single component
C ′ containing all |Y | + 1 interesting components in Ci ; this still costs at most (|Y | + 1) − 1 = |Y |
vertex deletions, as claimed.
Let Y ′ contain all the endpoints of S-edges that we delete to get the tree-structure. We
know that |Y ′ | ≤ |Y | and thus |(X \ Y ) ∪ Y ′ | ≤ |X|. Moreover, by the initial considerations, we
know that X ′ = (X \ Y ) ∪ Y ′ is a feasible solution as G − X ′ has the required tree-structure. If
|Y ′ | < |Y |, including the case that Y ′ = ∅, then |X ′ | < |X| as Y 6= ∅; this contradicts optimality
of X (required for being a dominant solution). If |Y ′ | = |Y | then Y ′ 6= ∅ and X ′ is an optimal
solution that contains more vertices of T ⊇ Y ′ , contradicting the choice of X as a dominant
solution. Thus, every nonempty set Y must see at least |Y | + 2 connected components, as
claimed.
Now we are ready to give the proof of Lemma 2. The argument relies on Hall’s Theorem
and is similar to the one for deletable terminal multiway cut [11].
Proof of Lemma 2. We know that every nonempty set Y ⊆ X0 sees at least |Y | + 2 interesting
components. To prove existence of the required path packing we construct a bipartite graph
where one side consists of the interesting components and the other side consists of the set X0
and two copies x′ , x′′ of the vertex x ∈ X0 . We connect v ∈ X0 with an interesting component
7
K if v sees K and we connect x′ and x′′ with the same interesting components as x. For this
bipartite graph it holds that for all sets Y ⊆ X0 ∪ {x′ , x′′ }, the size of N (Y ) is at least |Y |:
This holds trivially for Y = ∅; assume there exists a nonempty set Y ⊆ X0 ∪ {x′ , x′′ } such that
|N (Y )| < |Y |. But then we have |N (Y \ {x′ , x′′ })| ≤ |N (Y )| < |Y | ≤ |Y \ {x′ , x′′ }| + 2, which is
a contradiction to Lemma 3.
Since Hall’s condition is satisfied there exists a matching M that covers X0 ∪ {x′ , x′′ }. This
matching gives rise to a path packing from T to X where exactly three paths end in x and no
other vertices occur in more than one path: For each v ∈ X ∩ T pick the path of length zero that
consists only of v. For each edge {K, v} in the matching M , where v ∈ X0 ∪ {x′ , x′′ }, pick an
arbitrary path from a terminal t ∈ V (K) ∩ (T \ X) to v that uses only vertices from V (K) ∪ {v}.
(For v ∈ {x′ , x′′ } let the path end in x and use only vertices in V (K) ∪ {x}.) Because K is an
interesting component a terminal t ∈ V (K) ∩ (T \ X) must exist, and because K is a component
of G − X − S the path contains no other vertices of X. Similarly, the path cannot contain
S-edges between vertices of K, and its final edge to v cannot be in S because v ∈ X0 = X \ T ,
i.e., because v is not endpoint of any S-edge. Moreover, since each interesting component is
matched to a single vertex v ∈ X0 ∩ {x′ , x′′ }, all the paths are vertex-disjoint except for the
three paths that share their endpoint x. This path packing, including the trivial paths from
X ∩ T to X ∩ T , contains |X| + 2 paths from T to X in G − S that are vertex-disjoint except for
the three paths sharing endpoint x. By construction, there is at most one path to any vertex
of X0 starting in any interesting component K of G − X − S, because the components are used
according to the matching M . All further paths are of length zero, consisting of only a vertex
in X ∩ T and are, thus, not contained in components of G − X − S.
Setting up the gammoid. The gammoid M that we use is the direct sum of two gammoids M1 and M2 . To construct gammoid M1 we define a graph G1 = (V1 , E1 ) that is obtained
from G − S by adding two so called sink-only copies v ′ and v ′′ for every vertex v ∈ V . A sinkonly copy of a vertex v is a vertex v ′ (or v ′′ ) that has a directed edge (u, v ′ ) for each edge {u, v};
these were already used in previous work [11]. Note that adding sink-only copies of vertices does
not affect the possible path packings to other vertices since they can only be endpoints of paths;
however, they are convenient to capture multiple vertex-disjoint paths that, intuitively, end in
the same vertex. The matroid M1 is defined to the gammoid on G1 with sources T = V (S)
and ground set V1 = {v, v ′ , v ′′ | v ∈ V }; note that the sink-only copies of vertices in T are not
sources of M1 . The rank of matroid M1 is |T |, because the set of all trivial paths is independent
and at most |T | vertices can be linked to T .
Matroid M2 is the gammoid on the directed graph G2 = Kk,n = (S2 ∪˙ V̂ , E2 ) with sources
S2 and ground set V̂ = {v̂ | v ∈ V }; the edges in E2 are directed from S2 to V̂ . In other words,
gammoid M2 is simply a uniform matroid and a (deterministic) matrix representation could
also be obtained by using a Vandermonde matrix. The rank of M2 is k = |S2 | because no more
than |S2 | vertices can be linked to S2 and every set of at most k vertices of V̂ is linked to S2 .
For the application of Lemma 1 we will use the matroid M = M1 ⊕M2 , which has rank |T |+k.
˙ 2 with appropriate sources and
(Matroid M can also be seen as a gammoid on the graph G1 ∪G
ground set but we prefer the explicit direct sum and the implied block-diagonal representation
obtained below.) Representations A1 and A2 for both M1 and M2 can be computed by a
randomized polynomial-time algorithm with exponentially small error chance [15]; hence we
get a representation for M by diag(A1 , A2 ), i.e., the block-diagonal matrix with blocks A1 and
A2 . We may assume that A1 has |T | rows and A2 has k rows since this could be achieved by
Gaussian elimination (cf. [15]).
8
Applying the representative set lemma. Let T := {{v ′ , v ′′ , v̂} | v ∈ V }. For clarity,
by the above notation, this means that v ′ , v ′′ ∈ V1 and v̂ ∈ V̂ for each v ∈ V . Using Lemma 1
we will prove that we can compute in randomized polynomial time a (|T | + k − 3)-representative
subset T ′ of T that contains for all x ∈ X0 = X \ T the set {x′ , x′′ , x̂}, where X is any dominant
solution for (G, S, k). Lemma 1 guarantees |T ′ | ∈ O((|T |+k)3 ) = O(|S|3 ), since we can compute
a matrix representation of M in randomized polynomial-time as described above. We will see
later that we can find a (|T | + k − 3)-representative set of size O(|S|2 k) by a careful look at the
proof of Lemma 1, using the fact that M is the direct sum of two gammoids and that all sets
{v ′ , v ′′ , v̂} in T have two elements from the first and one element from the second gammoid;
a similar argument for getting a smaller representative set was already used by Kratsch and
Wahlström [11].
To ensure that all sets {x′ , x′′ , x̂} with x ∈ X0 are in T ′ we have to show that for each such
set {x′ , x′′ , x̂} there exists an independent set I of size at most |T | + k − 3 such that {x′ , x′′ , x̂}
uniquely extends I among triplets in T . This directly implies that {x′ , x′′ , x̂} must be in every
(|T | + k − 3)-representative set T ′ of T .
Lemma 4. Let X be a dominant solution for (G, S, k) and let T = V (S). For all x ∈ X0 = X \T
there exists an independent set I of size at most |T | + k − 3 in M such that {x′ , x′′ , x̂} uniquely
extends I.
Proof. Let x be an arbitrary vertex of X0 . In a first step we define an independent set I
and show in a second step that {x′ , x′′ , x̂} uniquely extends I. Applying Lemma 2 implies the
existence of a path packing P of |X| + 2 paths from T to X in G − S that are vertex-disjoint
except for three paths ending in x and such that each connected component of G − X − S is
intersected by at most one path of P. This directly implies a path packing P1 in G1 from T
to X ∪ {x′ , x′′ } that is (fully) vertex-disjoint. We retain the property that at most one path
intersects the vertex set of any component of G − X − S, but note that we do not get exactly
the same property for G1 − X because of the still present sink-only copies of vertices in X. (The
latter point will be no problem and should mainly explain why we need to talk about G − X − S
and not only G1 . Note that G − S and G1 by construction share the vertex set V to be able to
refer to connected components of G − X − S and the graph G1 underlying the gammoid M1 .)
While we do not know the paths in P1 entirely, we know for sure that no vertex of X ∪{x′ , x′′ }
can be an internal vertex of any path in P1 because there is a path ending in each of those
vertices. Similarly, we may assume that no vertex of T is internal to any path of P1 : If not
then any path P ∈ P1 with internal vertex from T can be shortened to start in that vertex; this
argument cannot be repeated indefinitely (as the paths get shorter each time). There is still at
most one path intersecting the vertex set of any component of G − X − S.
Now, define T ′ ⊆ T as those vertices of T in which no path of P1 starts; there must be
exactly |T | − |P| = |T | − (|X| + 2) of them since no vertex of T is internal. Moreover, for each
component K of G − X − S, the set T ′ contains all but at most one vertex of T ∩ V (K): At most
one path of P1 can start in T ∩ V (K) and no vertex can be internal. This will be important for
proving the claim below.
Clearly, the set T ′ ∪ X ∪ {x′ , x′′ } is independent in M1 because an appropriate path packing
′
P can be obtained from P1 by adding length zero paths for each v ∈ T ′ . The set X̂ =
{x̂ | x ∈ X} ⊆ V̂ is clearly independent in M2 since it has size at most k. Thus, the set
I ′ = T ′ ∪ X ∪ {x′ , x′′ } ∪ X̂ is independent in M = M1 ⊕ M2 . Define I as I ′ \ {x′ , x′′ , x̂}, i.e.,
I = T ′ ∪ X ∪ (X̂ \ {x̂}). The size of I is at most
|T ′ | + |X| + (|X̂| − 1) = |T | − (|X| + 2) + |X| + |X| − 1 = |T | + |X| − 3 ≤ |T | + k − 3.
Clearly, {x′ , x′′ , x̂} extends I, as I ′ = {x′ , x′′ , x̂} ∪ I is independent and both are disjoint by
choice of I. We now show that no other {v ′ , v ′′ , v̂} ∈ T extends I.
9
Claim 1. If {v ′ , v ′′ , v̂} ∈ T extends I then v = x.
Proof. Suppose that {v ′ , v ′′ , v̂} extends I. Clearly, this implies that v ∈
/ X \ {x} because
otherwise {v ′ , v ′′ , v̂} would not be disjoint from X̂ \ {x̂} ⊆ I. Thus, v ∈ V \ (X \ {x}).
Assume, for contradiction, that v ∈ V \ X, i.e., that v 6= x. We know that {v ′ , v ′′ , v̂} ∪ I is
independent in M , so I1 := I ∩ V1 must be independent in M1 . Thus, there exists a collection
P ′′ of |I1 | vertex-disjoint paths from T to I1 in G1 . Because X ⊆ I1 , the paths, say Pv′ and
Pv′′ , from T to {v ′ , v ′′ } cannot have internal vertices from the set X. Furthermore, they cannot
have other sink-only copies as internal vertices. Since v ∈ V \ X, this implies that Pv′ and Pv′′
are entirely contained in some component K1 of G1 − (X ∪ {x′ , x′′ | x ∈ X}). (Component K1
corresponds to a component K of G−X −S but also has sink-only copies of each vertex.) Recall
now that in T ′ we have all but at most one vertex of T ∩ V (K) for each connected component
of G − X − S and this is also true for T ∩ V (K1 ) as V (K1 ) ∩ V = V (K). Thus, in P ′′ there
is a path v of length zero for each vertex T ′ ∩ V (K1 ), leaving at most one vertex of T to start
paths to {v ′ , v ′′ }. This is a contradiction because Pv′ and Pv′′ are entirely contained in K1 and
fully vertex-disjoint.
Thus, if v ∈ V \ X then {v ′ , v ′′ } ∪ I1 is not independent in M1 and, hence, {v ′ , v ′′ , v̂} does
not extend I in M . Together with the first paragraph this implies that v = x, as claimed.
The set I fulfills the required properties which completes the proof.
We know now that for every vertex x ∈ V \ T that is a vertex in a dominant solution the set
{x′ , x′′ , x̂} is in every (|T |+k−3)-representative set T ′ . If we define V (T ′ ) = {v | {v ′ , v ′′ , v̂} ∈ T ′ }
then this implies that X0 ⊆ V (T ′ ) for each dominant solution X. Thus, every dominant solution
X is contained in V (T ′ ) ∪ T .
Shrinking the input graph to O(|V (T ′ )∪T |) vertices. In the previous parts we have
shown that if there exists a solution for (G, S, k), then there exists a solution that is completely
contained in W := V (T ′ ) ∪ T . Using this we can make all vertices in V \ W undeletable. We
achieve this by applying the so-called torso operation to vertex set W in G; let G′ = torso(G, W ).
By definition of torso(G, W ), the resulting graph G′ has vertex set W and is derived from G[W ]
by making each pair {u, v} ⊆ W adjacent if there is a u,v-path in G with internal vertices from
V \ W . Note that we do not create double edges or loops in G′ and that all edges of S are
preserved in G′ because T ⊆ W . (The same can be achieved by iteratively selecting a vertex
v ∈ V \ W , making its neighbors a clique, and deleting v from the graph.)
Lemma 5. (G′ , S, k) has a solution if and only if (G, S, k) has a solution.
It follows from Lemma 5 that (G′ , S, k) is an equivalent instance and the graph of this
instance contains at most |W | vertices. This completes the kernelization. The correctness
of Lemma 5 follows from the fact that the torso operation preserves the separators that are
contained in W (cf. [16]). For completeness we give a short proof of the lemma.
Proof of Lemma 5. Let X be a solution for (G′ , S, k). We prove that X is also a solution for
(G, S, k) by contradiction. Assume that X is not a solution for (G, S, k). Then there exists an
S-cycle C = v1 v2 . . . vl in G − X. Note that S ⊆ E(G′ ), because T = V (S) ⊆ W and therefore
at least two vertices of C are contained in W . Now we modify C to obtain an S-cycle C ′ in G′ .
Let vi , vj ∈ W ∩ C two vertices of the cycle with i < j such that {vi+1 , . . . , vj−1 } ⊆ V \ W . By
definition there exists an edge {vi , vj } in torso(G, W ) and using these edges we obtain cycle C ′ .
Note that C ′ contains no vertex of X and contains the same edges from S that C contains. Thus
C is an S-cycle in G′ − X which contradicts the assumption that X is a solution of (G′ , S, k).
10
For the other direction we assume that (G, S, k) has a solution. Then there also exists a
dominant solution X for (G, S, k) and we know that X ⊆ W . Again we prove that X is also
a solution for (G′ , S, k) by contradiction. Assume that X is not a solution for (G′ , S, k). Then
there exists a path P between the endpoints of an edge e = {x, y} ∈ S in G′ − X that does not
use the edge e. We modify P ′ to obtain a path P in G that does not contain the edge e. If P ′
uses an edge {u, v} that is not contained in G, then there exists a u-v path in V \ W connecting
u and v. Crucially, V \ W is disjoint from X so this replacement still yields a walk that avoids
X. Overall we get a walk from x to y in G that does not contain e as an edge and that avoids X.
This walk contains a path P from x to y and this path together with the edge e is an S-cycle
in G − X which is a contradiction to the assumption that X is a solution for (G, S, k).
So far we have a kernelization that creates an equivalent instance (G′ , S, k) such that G′ has
|W | vertices. As mentioned above, Lemma 1 guarantees that |W | ∈ O(|S|3 ) and this implies
a polynomial kernel for edge subset fvs parameterized by |S|. If we use the fact that the
gammoid M is the direct sum of two gammoids M1 and M2 , and that all sets {v ′ , v ′′ , v̂} ∈ T
contain exactly two elements of M1 and one element of M2 , then we can guarantee that |W | ∈
O(|S|2 k), which is an improvement for all nontrivial instances with k < |S|.
Lemma 6. Let M = M1 ⊕ M2 be the gammoid of rank |T | + k as defined above and T =
{I1 , I2 , . . . , It } be the set of independent sets of M that we use for the kernelization. Let A be
represented by diag(A1 , A2 ) as above. If |T | > |T2 | · k1 , then there exists a set I ∈ T such that
T \ {I} is (|T | + k − 3)-representative for T .
The proof of Lemma 6 is similar to Marx [15, Lemma 4.2]. We additionally use the fact that
M is the direct sum of two gammoids to obtain that the vectors in the exterior algebra which
represent the sets in T span a space of smaller dimension.
Proof of Lemma 6. Let U be the ground set of the matroid M which equals the set of columns
of A. For each
V e ∈ U , let xe be the corresponding (|T | + k)-dimensional column vector of A
and let wi = e∈Ai xe be a vector in the exterior algebra of the linear space F |T |+k . Every wi is
the wedge product of three vectors where exactly two are from A01 and one from A02 . The
two vectors corresponding to A01 can only span a space of dimension |T2 | and the vectors
corresponding to A02 can only span a space of dimension k1 . Thus, the wi ’s span a space of
dimension at most |T2 | · k1 . If |T | > |T2 | · k1 , then the wi ’s are not independent and there
exists some vector wl that can be expressed as a linear combination of the other vectors.
One can show analogously to Marx [15, Lemma 4.2] that T \{Il } is (|T |+k−3)-representative
for T . We replicate this proof for convenience of the reader. Assume that there exists a set Y
of size
V at most |T | + k − 3 such that Il extends Y and no other set Ii , i 6= l extends Y . Let
y = e∈Y xe . One property of the wedge product is that the product of some vectors in F |T |+k
is zero if and only if they are not independent. Therefore it holds that wl ∧ y 6= 0 and wi ∧ y = 0
for every i 6= l. But wl is a linear combination of other wi ’s and by the multi-linearity of the
wedge product we get that wl ∧ y 6= 0 is a linear combination of the values wi ∧ y = 0 for i 6= l,
which is a contradiction.
As mentioned above, Marx showed in [15] that one can find in randomized polynomialtime a matrix with r(M ) rows that represents a given gammoid M . We can make this proof
algorithmic in the same way Marx did [15, Lemma 4.2]. Combined with Lemma 6 it follows
directly
that we can find a (|T | + k − 3)-representative subset T ′ of |T | whose size is at most
|T | k
2
2
2 · 1 ∈ O(|S| k). This leads to a polynomial kernel with O(|S| k) vertices for edge subset
fvs parameterized by |S| and k.
11
4
Reducing the size of S
We have seen that edge subset fvs parameterized by |S| and k has a polynomial kernel. Now
the goal is to reduce the size of the set S until |S| is polynomially bounded in k. This will lead
to a polynomial kernel of edge subset fvs parameterized by k.
To begin, we do some initial modifications to ensure that we can always find a solution of
size at most k that contains no vertex of the set V (S), if one exists. For this we first delete all
vertices v ∈ V with the property that e = {v, v} ∈ S is a loop in G; since the vertex v must be
in any solution, we decrease the value k by one. Next we delete all remaining loops, because
these loops are not in S and cannot be contained in any S-cycle. We also reduce the number of
edges between two vertices v, w ∈ V (G). If no edge that is incident to v and w is contained in
the set S, then we delete all except one edge. On the other hand, if at least one edge between
v and w is contained in S, then we delete all except two edges. One of these edges is contained
in S and the other not. In the next step we add for every edge e = {v, w} ∈ S two new vertices
ve , ue to the graph, subdivide the edge e into three edges {v, ve }, {ve , we }, {we , w}, and edit S
by replacing edge e by the edge {ve , we } in S. If a solution X of edge subset fvs contains a
vertex xe ∈ V (S), then we can instead add the vertex x to X and delete xe from X, because
every cycle that contains vertex xe also contains vertex x; hence we can always find an optimal
solution that is disjoint from V (S).
Let (G, S, k) be an instance of edge subset fvs, such that G is a graph with the above
properties. Analogous to the paper of Cygan et al. [3] we consider a solution Z of the edge
subset fvs, with the difference that our solution is an 8-approximation of the problem, to
reduce the size of S. Even et al. [6] show that there exists an 8-approximation algorithm
for subset fvs. Since subset fvs and edge subset fvs are equivalent (cf. [3]), we can
compute in polynomial time an 8-approximation for edge subset fvs and we can assume that
Z ∩ V (S) = ∅. If |Z| > 8k, then we can stop immediately because no solution of size at most k
can exist. On the other hand, if |Z| ≤ k, then Z is a solution for the problem and we are done.
The set Z is a feasible solution to edge subset fvs on (G, S, |Z|). This implies that every
edge e ∈ S is a bridge in G − Z. In a next step we also remove all edges in S from G − Z.
Every connected component in G − Z − S contains no edge from S and, following Cygan et
al. [3], we call such a component a bubble. We denote the set of bubbles by DZ and define a
graph HZ = (DZ , EDZ ) whose vertices are bubbles and with bubbles I and J being adjacent,
i.e., {I, J} ∈ EDZ , if and only if the components I and J are connected by an edge from S.
The graph HZ is a forest, because Z is a solution for (G, S, |Z|) and a cycle in HZ would give
rise to an S-cycle in G − Z. Similarly, no two bubbles can be connected by more than one edge
of S. By VI we denote the vertices that are contained in bubble I. Since |E(VI , VJ ) ∩ S| ≤ 1
for all I, J ∈ DZ and equality holds if and only if {I, J} ∈ EDZ , we can associate an edge
e = {I, J} ∈ EDZ with the one edge eS = {vI , vJ } in E(VI , VJ ) ∩ S. If we add the vertex set Z
and all edges {z, I} with the property that z ∈ Z, I ∈ DZ and E(z, VI ) 6= ∅ to the graph HZ we
obtain a graph HZ+ that contains S-cycles. Note that every S-cycle must contain a vertex of the
set Z. We partition the set of bubbles according to the number of bubbles they are connected
with.
Definition 2. A bubble I ∈ DZ is called(i) solitary, if degHZ (I) = 0; (ii) leaf, if degHZ (I) = 1;
and (iii) inner, if degHZ (I) ≥ 2. By DZs , DZl , DZi we denote the corresponding sets of of bubbles.
+
as well as the sets
Let X ⊆ V \ V (S) be a superset of Z. We define the graphs HX , HX
+
DX , EDX analogously to the graphs HZ , HZ and the sets DZ , EDZ . Observe that the number
of edges in S is at most |DZ \ DZs |, because HZ is a forest, any two bubbles are connected by at
most one S-edge, and V (S) ∩ Z = ∅.
12
So far our setup is essentially the same as the one used by Cygan et al. [3]. However, instead
of an 8-approximate solution they use the framework of iterative compression, which provides
a solution Z of size k + 1 and leaves them with the task of reducing the number of S-edges for
the problem of finding a solution Z ∗ that is disjoint from Z. Moreover, it suffices for them to
consider the case that every feasible solution (if one exists) is disjoint from Z. In this setting
they are able to reduce to an equivalent instance (or find that some assumption was violated)
with only O(k3 ) edges in S.
Thus, while many relevant structures like z-flowers or parallel x-y paths containing S-edges
are the same, many things have to be handled differently. In particular, if we find that at
least one out of two vertices x, y ∈ Z must be in the solution then we cannot stop (using the
maximality condition) but need to continue and use this information in a more direct way.
During the reduction we detect certain pairs {x, y} of different vertices with the property
that each solution of size at most k must contain at least one of the vertices (if one exists). We
store this fact as a pair-constraint. We keep and enforce this information in the final instance,
unless we decide earlier to delete x or y. By P we denote the set of pair-constraints that we have
found so far. We can interpret this set as a set of edges and by V (P) we denote all vertices that
are contained in a pair-constraint. Note that vertices from the set V (S) are never contained
in a pair-constraint from P, because there always exists a solution that is disjoint from V (S).
We need the set P to detect edges in S that may be safely deleted. To this end, we generalize
the edge subset fvs problem by adding a set of pair-constraints P to the input; we call this
problem pair-constrained edge subset fvs.
pair-constrained edge subset feedback vertex set
Parameter: k
Input: An undirected graph G, a set S ⊆ E of edges, a set P of pair-constraints and an
integer k.
Question: Does there exist a set X ⊆ V of size at most k such that G − X contains no
S-cycle and such that for each pair-constraint {x, y} ∈ P we have x ∈ X or y ∈ X?
Clearly, instances (G, S, k) of edge subset fvs and (G, S, ∅, k) of pair-constrained edge
subset fvs are equivalent. Our goal is to reduce the size of S by detecting S-edges that we
can delete from S without changing the outcome. This leads to the following definition:
Definition 3. Let (G, S, P, k) be an instance of pair-constrained edge subset fvs. We
call an edge e ∈ S irrelevant, if X ⊆ V (G) is a solution for (G, S, P, k) if and only if X is a
solution for (G, S \ {e}, P, k).
Note that if two different S-edges e and e′ are irrelevant in (G, S, P, k), then e′ is not
necessarily irrelevant in (G, S \ {e}, P, k). In addition we do not expect to find all irrelevant
edges or pair-constraints.
The reduction rules. We now present our reduction rules. Throughout we assume that
always the lowest numbered applicable rule is applied first. Correctness and efficiency of the
overall reduction process will be proved later.
Let (G, S, P = ∅, k) be an instance for pair-constrained edge subset fvs and let Z
be an 8-approximation of this problem with k < |Z| ≤ 8k that is disjoint from V (S). In the
following the graphs G − Z, G − Z − S, HZ , and HZ+ are always defined with respect to the
current instance (G, S, P, k) of pair-constrained edge subset fvs. Note that Z ⊆ V and
we delete vertices from Z if we delete the corresponding vertex in V .
Rule 1: If k < 0, or if k = 0 and there exists an S-cycle, then reduce (G, S, P, k) to some
trivial false instance, i.e. G′ := ({x}, {e = {x, x}}), S ′ := {e}, P ′ = ∅ and k′ := 0.
13
Rule 2: Delete all bridges and all connected components not containing any edge from S.
Rule 3: If there exists an edge e ∈ S such that e is a bridge in (V, E \ (S \ {e})), then reduce
to S ′ = S \ {e}.
Rules 2 and 3 ensure that each bubble I ∈ DZ is adjacent to a vertex in Z in the graph
HZ+ , i.e. for all I ∈ DZ we have EH + (VI , Z) 6= ∅: Since Rule 2 is not applicable every bubble
Z
I ∈ DZ must be adjacent to a bubble J ∈ DZ \ I, or a vertex in Z; otherwise G[VI ] would be a
connected component of G that does not contain any edge from S (VI was deleted in Rule 2).
From Rule 3 follows that a bubble I ∈ DZ must be adjacent to a vertex in Z; otherwise the
edge e ∈ N (VI ) ∩ S would be a bridge in (V, E \ (S \ {e})).
Rule 4: If there exists a vertex v in the set V (P) that is contained in at least k + 1 pairconstraints of P, then we reduce to G′ = G − v and k′ = k − 1.
Rule 5: If |P| > k 2 (and Rule 4 is not applicable), then reduce (G, S, P, k) to some trivial false
instance.
Rule 6: If there exists a z-flower of order k + 1 in G for a vertex z ∈ Z, then we reduce to
G′ := G − z and k ′ := k − 1.
For the next rules we need a maximal matching M in HZ that covers all inner bubbles DZi in
HZ . Note that two adjacent leaf bubbles I1 , I2 are not adjacent to an inner bubble and form a
K2 in HZ , hence the edge {I1 , I2 } ∈ EDZ is contained in every maximal matching in HZ . We use
this matching to detect pair-constraints in Z. To this end we introduce the following definition:
Let e = {I, J} be an edge in the matching M . We say e sees the pair {x, y} of different vertices
x, y ∈ Z respectively the vertex x ∈ Z, if {I, x}, {J, y} ∈ E(HZ+ ) or {I, y}, {J, x} ∈ E(HZ+ )
respectively {I, x}, {J, x} ∈ E(HZ+ ).
Rule 7: If at least (k + 2) edges in M see a pair {x, y} of different vertices in Z, then we add
{x, y} to the set of pair-constraints P.
Rule 8: If there exists an edge e ∈ M such that e sees no single vertex z ∈ Z and for every pair
{x, y} seen by e the pair {x, y} is a pair-constraint in P, then remove eS from S and e from
M . (Recall: If e = {I, J} ∈ M ⊆ E(HZ ), then eS is the unique edge in E(VI , VJ ) ∩ S.)
The matching M is always recomputed if, through application of rules, it does no longer
cover every inner bubble or is maximal when testing whether Rules 7 or 8 apply (i.e., if the
preceding rules do not apply). If M does cover all inner bubbles but neither Rule 7 nor 8
apply then, as we will prove later, this implies |M | ∈ O(k3 ) and, hence, that there are at most
2|M | ∈ O(k3 ) inner bubbles.
Let L = DZl \ V (M ) be the set of leaf bubbles that are not covered by M . Because the
matching covers at least all inner bubbles, we know that |S| ≤ 2|M | + |L|. Therefore we have
to find a reduction rule that reduces the number of leaf bubbles in L. Every leaf bubble in L is
adjacent to an inner bubble in HZ , because M covers all leaf bubbles that are not adjacent to
an inner bubble. To bound the number of leaf bubbles in L we define for each z ∈ Z a graph
Gz with the help of the following two sets. The first one, Lz = NH + (z) ∩ L, is the set of all
Z
leaf bubbles I that are adjacent to z in HZ+ . The other Vzi = {v ∈ V | ∃J ∈ NH + (Lz ) : v ∈ VJ }
Z
consists of all vertices that are contained in an inner bubble that is adjacent to a leaf bubble in
Lz .
14
Vzi
Lz
z
Figure 1: Graph Gz ,
S-edges in Gz
V (Gz ) = {z} ∪ Lz ∪ Vzi
E(Gz ) = EH + (z, Lz ) ∪ {{I, w} | ∃I ∈ Lz , v ∈ VI , w ∈ Vzi : {v, w} ∈ S} ∪ (E(G[Vzi ]) \ S)
Z
In the graph Gz each leaf bubble I ∈ Lz is a single vertex. We are not interested in the
internal structure of leaf bubbles in Lz , whereas we are interested in the structure of the inner
bubbles that are adjacent to the leaf bubbles in Lz . Thus we add the connected component
that corresponds to an inner bubble which is adjacent to a bubble in Lz to Gz . In order to
apply the concept of flowers and blocking sets in Gz , an edge e ∈ E(Gz ) is an S-edge in Gz
if e = {I, w} with I ∈ Lz and w ∈ Vzi . Note that e is an edge in Gz , because there exists an
S-edge e′ = {v, w} in G with v ∈ VI .
Lemma 7. If there exists no z-flower of order k + 1 in Gz for a vertex z ∈ Z, then we can find
a z-blocker Bz ⊆ Vzi \ V (S) of size at most 2k in Gz .
The lemma follows from Theorem 1 and the preprocessing as well as the construction of Gz .
Proof of Lemma 7. The number of vertex-disjoint Lz -paths in Gz − z is at most k, otherwise
the Lz -paths together with vertex z would correspond to a z-flower of order k + 1 in Gz ; this
contradicts the assumption. From Theorem 1 it follows that there exists a set Bz ⊆ V (Gz −z) =
Lz ∪ Vzi of size at most 2k intersecting every Lz -path. Since every S-cycle through z in Gz must
contain an Lz -path, Bz is a z-blocker of size at most 2k in Gz .
It remains to show that there exists a z-blocker Bz ⊆ Vzi \ V (S). First we assume that
there exists a vertex I ∈ Bz ∩ Lz . From the construction of Gz it follows that every leaf bubble
I ∈ Lz has degree one in Gz − z. Thus instead of I we can choose the vertex in NGz (I) for the
z-blocker Bz to obtain that Bz ⊆ Vzi .
In the next step we take care that Bz is also disjoint from V (S). Assume that Bz contains
a vertex ve ∈ V (S) ∩ Vzi . From the preprocessing it follows that we can add v ⊆ Vzi \ V (S) to
Bz and delete ve from Bz , because every cycle that contains ve also contains v.
Note that we delete at least as many vertices from Bz as we add to Bz , hence Bz is still of
size at most 2k.
Since no previous rule is applicable and a z-flower of order k + 1 in Gz gives rise to a zflowerSof order k + 1 in G, we find a z-blocker of size at most 2k for every vertex z ∈ Z. Let
B = z∈Z Bz be the union of all z-blockers Bz of size at most 2k. Note that the set L is the
union of all sets LzSwith z ∈ Z, because every leaf bubble is adjacent to a vertex in Z due to
Rule 2, hence L = z∈Z Lz .
15
The following lemma provides three nice properties of the graph HZ∪B = (DZ∪B , EDZ∪B )
which helps us to bound the number of leaf bubbles in L ⊆ DZl . To memorize: The set DZ∪B
is the set of bubbles in G − (Z ∪ B) − S and two bubbles I, J are adjacent in HZ∪B if and only
if E(VI , VJ ) ∩ S 6= ∅.
Lemma 8. The graph HZ∪B has the following properties:
1. For each bubble I ∈ DZ∪B there exists a bubble J ∈ DZ , such that VI ⊆ VJ .
2. For each leaf bubble J ∈ DZ there exists a leaf bubble I ∈ DZ∪B , such that VI = VJ .
i
3. Let I, J ∈ L and K ∈ DZ∪B
, such that {I, K}, {J, K} ∈ EDZ∪D . For all z ∈ Z it holds
that z ∈
/ NG (VI ) or z ∈
/ NG (VJ ).
Proof. Property 1 holds because the set B only splits bubbles of G − Z − S further (because we
are now looking at deleting Z ∪ B from G − S) and does not merge any two bubbles. Property
2 follows from the fact that the set B is disjoint from the set of leaf bubbles. Next we show
Property 3 by contradiction. We assume that some z ∈ Z is in NG (VI ) and in NG (VJ ). Then
I and J are both vertices of the graph Gz and hence both are contained in the set Lz . The
consequence is that there exists an Lz path from bubble I over bubble K to bubble J in HZ∪B
which can be extended to a Lz path in Gz not containing any vertex in B; this contradicts the
fact that Bz ⊆ B blocks all Lz -paths in Gz .
l
From Lemma 8 it follows that L ⊆ DZ∪B
; thus we can use HZ∪B to bound the number of
i
leaf bubbles in L. Let I = {J ∈ DZ∪B | E(L, J) 6= ∅} be the set of inner bubbles in HZ∪B
that are adjacent to a leaf bubble in L. Clearly the number of edges between I and L in HZ∪B
equals the number |L|. Instead of again using a matching to reduce this number we consider
more carefully the properties of these edges. For this we define the property of seeing a pair
in a slightly different way. Let e = {I, J} be an edge with I ∈ I and J ∈ L. We say that
e = {I, J} with I ∈ I and J ∈ L sees the pair {x, y} of different vertices x ∈ Z ∪ B and y ∈ Z,
+
). Observe
that a bubble in L is never adjacent to a vertex in B in
if {I, x}, {J, y} ∈ E(HZ∪B
S
the graph HZ∪B , because B ⊆ z∈Z Vzi \ V (S).
Rule 9: If at least (k+2) edges {I1 , J1 }, {I2 , J2 }, . . . {Il , Jl } with l ≥ k+2, Ii ∈ I and Ji ∈ L for
1 ≤ i ≤ l see a pair {x, y} of different vertices, such that x ∈ Z ∪ B is adjacent to Ii , y ∈ Z
is adjacent to Ji for all i ∈ {1, 2, . . . , l}, then we add {x, y} to the set of pair-constraints
P.
At first sight Rule 7 and 9 may seem somewhat similar, but on closer inspection on can
observe a decisive difference. In Rule 9 we consider only edges between two disjoint sets of
bubbles, whereas the edges in M can be between two inner bubbles, an inner bubble and a leaf
bubble, or between two leaf bubbles. For this reason we can require in Rule 9 that all bubbles
in I are adjacent to x and all bubbles in L are adjacent to y; this is not possible in Rule 7. We
will see later that we need the definite assignment of the bubbles to the vertices in Z ∪ B by
applying Rule 9.
Rule 10: If there exists an edge e = {I, J} with I ∈ I and J ∈ L such that e sees no single
vertex z ∈ B ∪ Z and for every pair {x, y} seen by e the pair {x, y} is a pair-constraint in
P, then remove eS from S, delete J from L and replace I by I ∪ J in I.
If we delete an edge e = {I, J} from S by applying Rule 10, then the consequence is that
bubbles I and J are now merged into a single bubble. Anyhow, it is sufficient to continue with
16
Rule 9, because M is still a matching that covers all inner bubbles in the current graph HZ and
B still has the properties of Lemma 8 with respect to the current graph HZ∪B . That the edge
set M is still a matching in HZ holds because we never delete an edge in M or an endpoint of
an edge in M ; we only merge an endpoint of an edge in M with an unmatched leaf bubble in
L. The first two properties of Lemma 8 obviously hold with respect to the current graph HZ .
That Property 3 also holds follows from the fact that the leaf bubbles that are still in L are the
same as before and adjacent to the same inner bubbles as before.
The reduction rules are safe. First we show that our reduction rules are safe, i.e. that
there exists a solution for (G, S, P, k) if and only if there exists a solution for (G′ , S ′ , P ′ , k′ ).
Note that Rules 1, 2, and 6 are obviously safe and Rule 3 is safe because for every S-cycle
through an edge e ∈ S that is a bridge in (V, E \ (S \ {e})) there is another S-edge e′ on the
cycle. Let us consider the set P of pair-constraints to see that Rules 4 and 5 are safe. The
set P naturally leads to the graph P = (V (P), P) and has the property that we have to pick
at least on vertex of each pair-constraint for a solution for (G, S, P, k). Hence any solution
for (G, S, P, k) must contain a vertex cover of P . Thus, Rules 4 and 5 are direct analogues
of classical reduction rules for the vertex cover problem, and hence safe. To show that the
other rules are safe, we first show a technical Lemma about a property of edges in HZ∪B .
Lemma 9. If two different edges {I1 , J1 } and {I2 , J2 } in HZ∪B with I1 , I2 ∈ I, J1 , J2 ∈
L see a vertex z ∈ Z, respectively a pair {x, y} with x ∈ Z ∪ B and y ∈ Z such that
{x, I1 }, {x, I2 }, {y, J1 }, {y, J2 } ∈ E(HZ∪B ), then it holds that they are disjoint, i.e. that I1 6= I2
and J1 6= J2 .
Proof. We first assume that I1 = I2 . This implies that J1 and J2 are leaf bubbles in L which
are adjacent to the same inner bubble I = I1 = I2 in HZ∪B . For J1 and J2 it must hold that
z ∈ NG (VIi ) respectively y ∈ NG (VIi ) for i = 1, 2. But this is a contradiction to Property 3 of
Lemma 8. On the other hand if J1 = J2 , then I1 = I2 because every leaf bubble in L sees only
one other bubble.
To show that Rules 7 and 9 are safe, we have to prove that we only add a pair {x, y} of
vertices to the set P of pair-constraints if either x or y must be in each solution of size at most
k. The (k + 2) edges that see a pair {x, y} are pairwise disjoint, because M is a matching and
+
Lemma 9 holds. Hence we have at least (k + 2) disjoint x-y paths in HZ+ respectively HZ∪B
which we can extend to at least (k + 2) disjoint x-y paths in G. This is the reason why at least
one of x and y must be in any solution and it is safe to add {x, y} to P as a pair-constraint.
It remains to show that Rules 8 and 10 are safe. For this we prove that the edges that we
delete in these rules are irrelevant. First we prove the following lemma.
Lemma 10. Let Y ⊆ V \ V (S) be a superset of Z, hence G − Y contains no S-cycle. If
e = {I, J} ∈ HY sees no single vertex y ∈ Y and for every pair {x, y} with x, y ∈ Y seen
by e the pair {x, y} is a pair-constraint in P, then eS = {vI , vJ } is irrelevant for the instance
(G, S, P, k).
Proof. Let e = {I, J} ∈ HY be an edge with the properties of the lemma and eS the single edge
in E(VI , VJ )∩S. To show that eS = {vI , vJ } is irrelevant for instance (G, S, P, k) we have to show
that X ⊆ V (G) is a solution for (G, S, P, k) if and only if X is a solution for (G, S \ {eS }, P, k).
Since every solution X for (G, S, P, k) is also a solution for (G, S \ {eS }, P, k), we only have to
show the other direction.
Let X be a solution for (G, S \ {eS }, P, k). We assume that there exists an S-cycle C in
G−X. This S-cycle C can only contain the S-edge eS ; otherwise would C be an (S \{eS })-cycle
which contradicts the fact that X is a solution for (G, S \ {eS }, P, k).
17
Claim 2. If an S-cycle C in G only contains the S-edge eS , then there exists either a vertex
y ∈ Y such that e sees the single vertex y and y is contained in cycle C or two different vertices
x, y ∈ Y such that e sees the pair {x, y} and cycle C contains x and y.
Proof. Let C be an S-cycle with the properties of the claim. Thus C must exit bubble I and
bubble J by edges that end in Y , because this is the only way to obtain a path from vI to vJ
that uses no edge from S. If these two edges share their endpoint y in Y , then e sees the single
vertex y and y is contained in C. On the other hand if these two edges have different endpoints
x, y in Y , then e sees the pair {x, y} and the vertices x, y are contained in C.
Based on Claim 2, it follows that edge e = {I, J} must see a single vertex y ∈ Y that is
contained in C or a pair {x, y} with x, y ∈ Y such that x, y are contained in C. From the
properties of edge e follows that e sees no single vertex and every pair {x, y} that is seen by
e must be contained in a pair-constraint. Let {x, y} be the pair that is seen by e such that
x, y are vertices of cycle C (Claim 2). But at least one vertex of the pair {x, y} must be in
the solution X for (G, S \ {eS }, P, k), since e sees only pairs that are contained in the set P of
pair-constraints; hence C is no cycle in G − X.
From Lemma 10 follows that we only delete an edge eS in Rule 8 and 10 when eS is irrelevant
for instance (G, S, P, k); this holds because Y = Z respectively Y = Z ∪ B is a superset of Z.
Applying the Rules. First we show that if none of the rules can be applied, then the
size of S is bounded by O(k4 ). For this we prove two lemmas. One bounds the size of M which
helps us to bound the number of inner bubbles and the other bounds the number of leaf bubbles
in L.
Lemma 11. If the matching M covers all inner bubbles in HZ and we cannot apply Rules 1
through 8, then the size of M is at most O(k3 ).
Proof. Each edge in M sees either a pair of vertices in Z or a single vertex in Z. The number
2
of pairs in Z is at most |Z|
2 ≤ |Z| . Therefore the number of pairs in Z that are not in the set
P of pair-constraints is at most |Z|2 . Because we cannot apply Rule 7, at most (k + 1) edges
in M see any pair that is not in the set of pair-constraints. Thus at most (k + 1)|Z|2 edges of
M can see a pair of vertices in Z that is not in P. The number of edges in M that see a single
vertex in Z is at most k|Z|; otherwise we can apply Rule 6, because at least one single vertex
z in Z is seen by at least k + 1 edges from M and these edges together with z are a z-flower
of order k + 1 in HZ+ which we can expand to a z-flower of order k + 1 in G. Since we cannot
apply Rules 6, 7 or 8, this leads to at most (k + 1)|Z|2 + k|Z| ∈ O(k3 ) edges in M , because
|Z| ≤ 8k.
From the lemma it follows that the number of inner bubbles in HZ is at most 2|M | ∈ O(k3 ).
Lemma 12. If we cannot apply Rules 1 through 10 then the size of L is bounded by O(k4 ).
Proof. We claim that the number of edges between bubbles in I and bubbles in L is at most
(k + 1)|Z|(|B| + |Z|) + k|Z|, if no rule is applicable. This implies that there are at most O(k4 )
leaf bubbles in L.
Each edge between bubbles in I and bubbles in L sees a pair {x, y}, such that {x, I}, {y, J} ∈
E(HZ∪B ) with x ∈ Z ∪ B is adjacent to I, y ∈ Z is adjacent to J or a vertex z in Z; hence
the number of pairs is at most |Z|(|Z| + |B|). Rule 9 adds {x, y} to P if at least (k + 2) edges
{I1 , J1 }, {I2 , J2 }, . . . , {Il , Jl } with l ≥ k + 1, Ii ∈ I and Ji ∈ L for 1 ≤ i ≤ l see the pair {x, y}
such that x ∈ Z ∪ B is adjacent to Ii and y ∈ Z is adjacent to Ji for 1 ≤ i ≤ l. This bounds the
18
number of edges between vertices in I and L which see a pair, whose vertices are not a pair in
the set P of pair-constraints, by (k + 1)|Z|(|Z| + |B|). The number of edges between vertices in
I and L that see a certain vertex z is at most k, otherwise the at least k + 1 edges between I
+
because
and L that see vertex z together with vertex z form a z-flower of order k + 1 in HZ∪B
Lemma 9 ensures that the edges are disjoint. But then we can apply Rule 6 and delete vertex
z. Hence at most k|Z| edges between vertices I and L can see a vertex in Z. This leads to at
most (k + 1)|Z|(|B| + |Z|) + k|Z| edges between vertices in I and L, because we cannot apply
Rules 6, 9 or 10; this implies that |L| ∈ O(k4 ), because |Z| ≤ 8k and |B| ≤ 2k|Z| ≤ 16k2 .
If we combine these two results, we know that |DZi | + |DZl | ∈ O(k4 ). As mentioned above
this is an upper bound for the number of edges in S, because HZ is a forest, because there is
at most one edge of S between any two bubbles, and because V (S) ∩ Z = ∅.
Finally we have to prove that we can perform the reduction in polynomial time. First we
prove that each rule is applied a polynomial number of times and second that every single rule
application can be performed in polynomial time.
Lemma 13. Each reduction rule is applied at most a number of times that is polynomially
bounded in the input size.
Proof. Note that we reduce in each rule, except Rules 7 and 9, the size of at least one of the
sets V , E, S, the value k or decide that no solution of size at most k exists. In Rules 7 and 9
we add pair constraints to P, but if P contains more than k 2 pair constraints, we either find
a vertex z ∈ V (P) that we delete in Rule 4 and reduce k by one or we decide in Rule 5 that
no solution of size at most k exists. This bounds the number of pair constraints that we add
to P during the reduction by k3 because we can decrease k at most k times. Thus, each rule is
applied at most a number of times that is polynomial in the input size.
Next we show that each single rule application can be performed in polynomial time. It is
obvious that we can apply Rules 1 through 5 in polynomial time. The following lemma addresses
Rule 6 by solving a matroid parity problem on an appropriate gammoid.
Lemma 14. Let G = (V, E), z ∈ V , and S ⊆ E. A z-flower of maximum order, i.e., a
maximum number of S-cycles that intersect only in z, can be found in (deterministic) polynomial
time.
Proof. For simplicity, we assume that there are no edges of S incident with z. If this is not
the case, then it can be checked that for each neighbor v ∈ N (z) with {v, z} ∈ S removing
{v, z} from S and adding instead all other edges incident with v to S gives the desired result.
Furthermore, we assume that no two edges of S are incident with the same vertex of G; this
can be achieved by appropriate subdivision operations, without changing the maximum order
of z-flowers.
Let {C1 , . . . , Ct } be a z-flower of order t. Each Ci gives rise to a path Pi between two different
neighbors u and v of z; all these paths are fully vertex-disjoint. By our above assumption, there
are no S-edges incident with z, hence, each Pi must contain two consecutive vertices, say si and
ti , with {si , ti } ∈ S. In this way, each path Pi can be split into two paths, Pi,s and Pi,t , from
N (v) to {si , ti }; all these 2t paths are pairwise vertex-disjoint and do not contain the vertex
z. Thus, from any z-flower of order t we get 2t vertex-disjoint paths in G − z from N (z) to
T ⊆ V (S), i.e., endpoints of S-edges, such that T can be partitioned into t two-sets of vertices
that are also edges in S. In the language of gammoids this means that T is an independent set
in the gammoid on graph G − z, with sources N (z), and ground set V (S).
Conversely, any independent set T in the mentioned gammoid implies the existence of |T |
vertex-disjoint paths in G − z from N (z) to T . If, as above, T can be partitioned into edges of
19
S then this gives rise to a z-flower of order t = |T |/2: Clearly, |T | must be even to allow for
the partition into sets of size two. Moreover, the paths are vertex-disjoint and, thus, two paths
from N (z) ending in {si , ti } ∈ S can be combined, using that {si , ti } must be an edge of G
into a single path, say Pi , from N (z) to N (z) that contains at least one edge of S. Note that,
because si and ti are ends of two paths in the packing they cannot occur in any other paths,
so this combination still yields vertex-disjoint paths in G − z. Finally, adding the vertex z, the
paths P1 , . . . , Pt can be combined into t S-cycles that intersect only in z.
Thus, the task of finding a z-flower of maximum order reduces to that of solving a matroid
parity problem on a gammoid: The underlying graph is G − z, the source set is NG (z), the
ground set is V (S), and the pairs are given by S. Recall that pairs in S are vertex-disjoint. Using
the algorithm due to Lovász [14], one may find a maximum independent set composed of pairs
in S in polynomial time, when provided with a matrix representation for the gammoid. A small
caveat would be that one would need a randomized algorithm for finding said representation.
Conveniently, specialized deterministic algorithms exist for subclasses of linear matroids; we
can use a deterministic algorithm due to Tong et al. [20] that solves the problem by reduction
to weighted matching on graphs. (Note that given a maximum independent set T composed of
pairs, the cycles of the z-flower can be found by simple disjoint paths computation for N (z) to
T in G − z.)
It remains to show that we can apply Rules 7 through 10 in polynomial time.
Lemma 15. We can apply Rule 7 and 8 in polynomial time.
Proof. First of all we store for each edge e = {I, J} ∈ M all vertices z ∈ Z seen by edge e
and all pairs {x, y} with x, y ∈ Z seen by edge e. For each edge we need at most O(|Z|2 ) time;
we only have to test for each vertex z ∈ Z respectively each pair {x, y} with x, y ∈ Z whether
{I, z}, {J, z} ∈ E(HZ ) respectively {I, x}, {J, y} ∈ E(HZ ) or {I, y}, {J, x} ∈ E(HZ ). Next we
count how many edges see a pair {x, y} with x, y ∈ Z and denote this value by c{x,y} . It takes
at most O(|E||Z|2 ) time to compute all values; we only have to count for how many edges we
store a certain pair. If a counter c{x,y} has value at least k + 2, then we add the pair {x, y} to
the set P of pair-constraints. We can check this for all counters in O(|Z|2 ) time. The above
computation corresponds to the computation we need for Rule 7. To apply Rule 8 we only have
to look at all vertices and pairs that we stored for an edge e ∈ M . If we have stored no single
vertex and only pairs that are pair-constraints in P, then e fulfills the conditions of an edge
that we delete in Rule 8. To check this for one edge takes at most O(|Z|2 ) time.
We prove that we can apply Rule 9 and 10 in polynomial time similar to how we prove that
we can apply Rule 7 and 8 in polynomial time. We only have to remember which endpoint is
adjacent to which vertex in a pair.
Lemma 16. We can apply Rule 9 and 10 in polynomial time.
Proof. First of all we store for each edge e = {I, J} with I ∈ I, J ∈ L all vertices z ∈ Z seen
by edge e and all pairs (x, y) with x ∈ Z ∪ B adjacent to I, y ∈ Z adjacent to J such that e
sees the pair {x, y}. For each edge e = {I, J} with I ∈ I, J ∈ L we need at most O(|Z ∪ B||Z|)
time; we only have to test for each vertex z ∈ Z respectively each pair (x, y) with x ∈ Z ∪ B,
y ∈ Z whether {I, z}, {J, z} ∈ E(HZ∪B ) respectively {I, x}, {J, y} ∈ E(HZ∪B ). Next we count
for how many edges we stored the pair (x, y) with x ∈ Z ∪ B, y ∈ Z and denote this value
by c(x,y) . It takes at most O(|E||Z ∪ B||Z|) time to compute all values; we only have to count
for how many edges we store a certain pair. If a counter c(x,y) has value at least k + 2, then
we add the pair {x, y} to the set P of pair-constraints. We can check this for all counters in
20
O(|Z ∪ B||Z|) time. The above computation corresponds to the computation we need for Rule
9, because we only store the pair (x, y) for an edge if the edge sees the pair {x, y}. To apply
Rule 10 we only have to look at all vertices and pairs that we stored for an edge e between
bubbles in I and bubbles in L. If we have stored no single vertex and only pairs (x, y) such
that {x, y} is a pair-constraints in P, then e fulfills the conditions of an edge that we delete in
Rule 10. To check this for one edge takes at most O(|Z ∪ B||Z|) time.
Finally, we show that we can compute the matching M and the set B in polynomial time.
Lemma 17. We can compute a maximal matching M in HZ that covers all inner bubbles in
polynomial time.
Proof. We prove the lemma by giving a simple greedy algorithm for this problem. Let T be
a connected component in HZ . Since HZ is a forest, T is a tree; take T to be rooted in an
arbitrary vertex r.
Algorithm 1 Matching
Input: A tree T with root r.
Output: A maximal matching M that covers all inner bubbles in T .
1: let v be a child of r in T
2: M ← {{r, v}}
3: for all children w of v and r with w 6= v do
4:
if w is no leaf in T then
5:
let Tw be the subtree of T rooted at w
6:
M ← M ∪ Matching(Tw , w)
7:
end if
8: end for
9: return M
It can be easily shown that M is an maximal matching in T that covers all inner bubbles,
because T is a tree. Since we can apply this algorithm to every connected component in HZ in
polynomial time, the union of all matchings leads to the required maximal matching.
It remains to show that we can find B in polynomial time. From Schrijver’s proof of Theorem
1 and the proof of Lemma 7 it follows that we can find in polynomial time either a z-flower of
order (k + 1) or a z-blocker of size at most 2k in Gz . Since there exists no z-flower in Gz when
we compute B, we compute for every z exactly once the set Bz and since B is simply the union
of all z-blockers we can compute B in polynomial time.
Finding an equivalent instance for Edge Subset Feedback Vertex Set. Up to
now we can only bound the number of edges in S for the pair-constrained edge subset
fvs problem. As mentioned above the instance (G, S, P = ∅, k) for pair-constrained edge
subset fvs is equivalent to the instance (G, S, k) of edge subset fvs. Therefore we only have
to show that we can find in polynomial time an instance of edge subset fvs that is equivalent
to the instance (G, S, P, k) of pair-constrained edge subset fvs and has at most O(k4 )
S-edges. Let {x, y} ∈ P be a pair-constraint. If there are two edges between x and y of which
at least one is contained in S, then x or y must be in any solution, because xy is an S-cycle. For
this reason, the instance (G′ , S ′ = S ∪ P, k) of edge subset fvs is equivalent to the instance
(G, S, P, k) of pair-constrained edge subset fvs, where G′ is created from G by adding
one edge {x, y} between every two vertices x and y with {x, y} ∈ P when {x, y} ∈
/ E and by
21
adding an edge {x, y} between x and y that is also contained in S ′ ; hence there are two edges
between x and y with {x, y} ∈ P in graph G′ and we add exactly one edge between x and y to
S ′ . Because we cannot apply Rule 4 or 5 to (G, S, P, k), we know that |P| ≤ k2 . This leads to
a bound of |S| + |P| ∈ O(k4 ) edges in S ′ for the edge subset fvs problem after the reduction.
Finally, we combine the results of Section 3 and Section 4 to obtain a polynomial kernel for
edge subset fvs parameterized by k. Let us first make some comments about the reduction
of the size of S and the kernelization: For the reduction of the size of S we use the fact that
we can always find a solution that is disjoint from T . This only holds because we modified
the graph accordingly. But since this is a correct reduction it holds that an input instance
(G, S, k) of edge subset fvs has a solution if and only if the output instance (G′ , S ′ , k′ ) of the
reduction in Section 4 has a solution. Thus it is no problem that we consider dominant solutions
for the kernelization in Section 3 and that the kernelization only guarantees the preservation of
dominant solutions. Every instance (G′ , S ′ , k′ ) has a dominant solution of size at most k′ when
a solution of size at most k′ exists; remember that X is a dominant solution for (G′ , S ′ , k′ ) if
it has minimum size and contains a maximal number of vertices from T ′ among solutions of
minimum size. Hence if (G′ , S ′ , k′ ) has a solution then it has a dominant solution X and X is
a dominant solution for (G′ , S ′ , k′ ) if and only if X is a dominant solution for (G′′ , S ′ , k′ ) the
output instance of the kernelization in Section 3.
Summarized, the reduction of the number of edges in S to O(k4 ) edges together with the
kernelization to O(|S|2 k) vertices for edge subset fvs parameterized by |S| and k, results in
a kernelized instance with O(k9 ) vertices for edge subset fvs parameterized by k.
5
Conclusions
We have shown that the subset fvs problem has a randomized polynomial kernelization using
the matroid-based tools of Kratsch and Wahlström [11], positively answering the question of
Cygan et al. [3]. As in previous work [11] the error-probability can be made exponentially
small without increasing the kernel size. Nevertheless, it would of course be very interesting
whether the use of randomization and/or matroids can be avoided. Furthermore, there is quite
a gap between O(k9 ) vertices and a lower bound of size O(k2−ε ) that is inherited from vertex
cover [4], conditioned on non-collapse of the polynomial hierarchy.
Other open problems regarding existence of polynomial kernels, possibly amenable to the
matroid tools, are multiway cut and directed feedback vertex set (dfvs). There is
also a directed version of subset fvs, called directed subset feedback vertex set, but
it generalizes dfvs, whose kernel status has remained open for quite some time now.
References
[1] Kevin Burrage, Vladimir Estivill-Castro, Michael R. Fellows, Michael A. Langston, Shev
Mac, and Frances A. Rosamond. The undirected feedback vertex set problem has a poly(k )
kernel. In IWPEC 2006, volume 4169 of LNCS, pages 192–202. Springer, 2006.
[2] Marek Cygan, Jesper Nederlof, Marcin Pilipczuk, Michal Pilipczuk, Johan M. M. van Rooij,
and Jakub Onufry Wojtaszczyk. Solving connectivity problems parameterized by treewidth
in single exponential time. In FOCS 2011, pages 150–159. IEEE Computer Society, 2011.
[3] Marek Cygan, Marcin Pilipczuk, Michal Pilipczuk, and Jakub Onufry Wojtaszczyk. Subset
feedback vertex set is fixed-parameter tractable. SIAM J. Discrete Math., 27(1):290–309,
2013.
22
[4] Holger Dell and Dieter van Melkebeek. Satisfiability allows no nontrivial sparsification
unless the polynomial-time hierarchy collapses. J. ACM, 61(4):23:1–23:27, 2014.
[5] Reinhard Diestel. Graph theory (graduate texts in mathematics). 2005.
[6] Guy Even, Joseph Naor, and Leonid Zosin. An 8-approximation algorithm for the subset
feedback vertex set problem. SIAM J. Comput., 30(4):1231–1252, 2000.
[7] Fedor V. Fomin, Daniel Lokshtanov, and Saket Saurabh. Efficient computation of representative sets with applications in parameterized and exact algorithms. In SODA 2014,
pages 142–151. SIAM, 2014.
[8] Tibor Gallai. Maximum-minimum sätze und verallgemeinerte faktoren von graphen. Acta
Mathematica Hungarica, 12(1-2):131–173, 1961.
[9] Ken-ichi Kawarabayashi and Yusuke Kobayashi. Fixed-parameter tractability for the subset feedback set problem and the s-cycle packing problem. J. Comb. Theory, Ser. B,
102(4):1020–1034, 2012.
[10] Tomasz Kociumaka and Marcin Pilipczuk. Faster deterministic feedback vertex set. Inf.
Process. Lett., 114(10):556–560, 2014.
[11] Stefan Kratsch and Magnus Wahlström. Representative sets and irrelevant vertices: New
tools for kernelization. In FOCS 2012, pages 450–459. IEEE Computer Society, 2012.
[12] Daniel Lokshtanov, M. S. Ramanujan, and Saket Saurabh. Linear time parameterized
algorithms for subset feedback vertex set. In ICALP 2015, volume 9134 of LNCS, pages
935–946. Springer, 2015.
[13] László Lovász. Flats in matroids and geometric graphs. Combinatorial surveys, pages
45–86, 1977.
[14] László Lovász. Matroid matching and some applications.
28(2):208–236, 1980.
J. Comb. Theory, Ser. B,
[15] Dániel Marx. A parameterized view on matroid optimization problems. Theor. Comput.
Sci., 410(44):4471–4479, 2009.
[16] Dániel Marx, Barry O’Sullivan, and Igor Razgon. Finding small separators in linear time
via treewidth reduction. ACM Transactions on Algorithms, 9(4):30, 2013.
[17] Hazel Perfect. Applications of Menger’s graph theorem. J. Math. Anal. Appl., 22:96–111,
1968.
[18] Alexander Schrijver. A short proof of Mader’s sigma-paths theorem. J. Comb. Theory, Ser.
B, 82(2):319–321, 2001.
[19] Stéphan Thomassé. A 4k 2 kernel for feedback vertex set. ACM Transactions on Algorithms,
6(2), 2010.
[20] Po Tong, Eugene L. Lawler, and Vijay V. Vazirani. Solving the weighted parity problem
for gammoids by reduction to graphic matching. Computer Science Division, University of
California, 1982.
[21] Magnus Wahlström. Half-integrality, lp-branching and FPT algorithms. In SODA 2014,
pages 1762–1781. SIAM, 2014.
23
| 8cs.DS
|
arXiv:1410.5401v2 [cs.NE] 10 Dec 2014
Neural Turing Machines
Alex Graves
Greg Wayne
Ivo Danihelka
[email protected]
[email protected]
[email protected]
Google DeepMind, London, UK
Abstract
We extend the capabilities of neural networks by coupling them to external memory resources, which they can interact with by attentional processes. The combined system is
analogous to a Turing Machine or Von Neumann architecture but is differentiable end-toend, allowing it to be efficiently trained with gradient descent. Preliminary results demonstrate that Neural Turing Machines can infer simple algorithms such as copying, sorting,
and associative recall from input and output examples.
1
Introduction
Computer programs make use of three fundamental mechanisms: elementary operations
(e.g., arithmetic operations), logical flow control (branching), and external memory, which
can be written to and read from in the course of computation (Von Neumann, 1945). Despite its wide-ranging success in modelling complicated data, modern machine learning
has largely neglected the use of logical flow control and external memory.
Recurrent neural networks (RNNs) stand out from other machine learning methods
for their ability to learn and carry out complicated transformations of data over extended
periods of time. Moreover, it is known that RNNs are Turing-Complete (Siegelmann and
Sontag, 1995), and therefore have the capacity to simulate arbitrary procedures, if properly
wired. Yet what is possible in principle is not always what is simple in practice. We
therefore enrich the capabilities of standard recurrent networks to simplify the solution of
algorithmic tasks. This enrichment is primarily via a large, addressable memory, so, by
analogy to Turing’s enrichment of finite-state machines by an infinite memory tape, we
1
dub our device a “Neural Turing Machine” (NTM). Unlike a Turing machine, an NTM
is a differentiable computer that can be trained by gradient descent, yielding a practical
mechanism for learning programs.
In human cognition, the process that shares the most similarity to algorithmic operation
is known as “working memory.” While the mechanisms of working memory remain somewhat obscure at the level of neurophysiology, the verbal definition is understood to mean
a capacity for short-term storage of information and its rule-based manipulation (Baddeley et al., 2009). In computational terms, these rules are simple programs, and the stored
information constitutes the arguments of these programs. Therefore, an NTM resembles
a working memory system, as it is designed to solve tasks that require the application of
approximate rules to “rapidly-created variables.” Rapidly-created variables (Hadley, 2009)
are data that are quickly bound to memory slots, in the same way that the number 3 and the
number 4 are put inside registers in a conventional computer and added to make 7 (Minsky,
1967). An NTM bears another close resemblance to models of working memory since the
NTM architecture uses an attentional process to read from and write to memory selectively.
In contrast to most models of working memory, our architecture can learn to use its working
memory instead of deploying a fixed set of procedures over symbolic data.
The organisation of this report begins with a brief review of germane research on working memory in psychology, linguistics, and neuroscience, along with related research in
artificial intelligence and neural networks. We then describe our basic contribution, a memory architecture and attentional controller that we believe is well-suited to the performance
of tasks that require the induction and execution of simple programs. To test this architecture, we have constructed a battery of problems, and we present their precise descriptions
along with our results. We conclude by summarising the strengths of the architecture.
2
Foundational Research
2.1
Psychology and Neuroscience
The concept of working memory has been most heavily developed in psychology to explain
the performance of tasks involving the short-term manipulation of information. The broad
picture is that a “central executive” focuses attention and performs operations on data in a
memory buffer (Baddeley et al., 2009). Psychologists have extensively studied the capacity
limitations of working memory, which is often quantified by the number of “chunks” of
information that can be readily recalled (Miller, 1956).1 These capacity limitations lead
toward an understanding of structural constraints in the human working memory system,
but in our own work we are happy to exceed them.
In neuroscience, the working memory process has been ascribed to the functioning of a
system composed of the prefrontal cortex and basal ganglia (Goldman-Rakic, 1995). Typ1
There remains vigorous debate about how best to characterise capacity limitations (Barrouillet et al.,
2004).
2
ical experiments involve recording from a single neuron or group of neurons in prefrontal
cortex while a monkey is performing a task that involves observing a transient cue, waiting
through a “delay period,” then responding in a manner dependent on the cue. Certain tasks
elicit persistent firing from individual neurons during the delay period or more complicated
neural dynamics. A recent study quantified delay period activity in prefrontal cortex for a
complex, context-dependent task based on measures of “dimensionality” of the population
code and showed that it predicted memory performance (Rigotti et al., 2013).
Modeling studies of working memory range from those that consider how biophysical
circuits could implement persistent neuronal firing (Wang, 1999) to those that try to solve
explicit tasks (Hazy et al., 2006) (Dayan, 2008) (Eliasmith, 2013). Of these, Hazy et al.’s
model is the most relevant to our work, as it is itself analogous to the Long Short-Term
Memory architecture, which we have modified ourselves. As in our architecture, Hazy
et al.’s has mechanisms to gate information into memory slots, which they use to solve a
memory task constructed of nested rules. In contrast to our work, the authors include no
sophisticated notion of memory addressing, which limits the system to storage and recall
of relatively simple, atomic data. Addressing, fundamental to our work, is usually left
out from computational models in neuroscience, though it deserves to be mentioned that
Gallistel and King (Gallistel and King, 2009) and Marcus (Marcus, 2003) have argued that
addressing must be implicated in the operation of the brain.
2.2
Cognitive Science and Linguistics
Historically, cognitive science and linguistics emerged as fields at roughly the same time
as artificial intelligence, all deeply influenced by the advent of the computer (Chomsky,
1956) (Miller, 2003). Their intentions were to explain human mental behaviour based on
information or symbol-processing metaphors. In the early 1980s, both fields considered
recursive or procedural (rule-based) symbol-processing to be the highest mark of cognition. The Parallel Distributed Processing (PDP) or connectionist revolution cast aside the
symbol-processing metaphor in favour of a so-called “sub-symbolic” description of thought
processes (Rumelhart et al., 1986).
Fodor and Pylyshyn (Fodor and Pylyshyn, 1988) famously made two barbed claims
about the limitations of neural networks for cognitive modeling. They first objected that
connectionist theories were incapable of variable-binding, or the assignment of a particular
datum to a particular slot in a data structure. In language, variable-binding is ubiquitous;
for example, when one produces or interprets a sentence of the form, “Mary spoke to John,”
one has assigned “Mary” the role of subject, “John” the role of object, and “spoke to” the
role of the transitive verb. Fodor and Pylyshyn also argued that neural networks with fixedlength input domains could not reproduce human capabilities in tasks that involve processing variable-length structures. In response to this criticism, neural network researchers
including Hinton (Hinton, 1986), Smolensky (Smolensky, 1990), Touretzky (Touretzky,
1990), Pollack (Pollack, 1990), Plate (Plate, 2003), and Kanerva (Kanerva, 2009) investigated specific mechanisms that could support both variable-binding and variable-length
3
structure within a connectionist framework. Our architecture draws on and potentiates this
work.
Recursive processing of variable-length structures continues to be regarded as a hallmark of human cognition. In the last decade, a firefight in the linguistics community staked
several leaders of the field against one another. At issue was whether recursive processing
is the “uniquely human” evolutionary innovation that enables language and is specialized to
language, a view supported by Fitch, Hauser, and Chomsky (Fitch et al., 2005), or whether
multiple new adaptations are responsible for human language evolution and recursive processing predates language (Jackendoff and Pinker, 2005). Regardless of recursive processing’s evolutionary origins, all agreed that it is essential to human cognitive flexibility.
2.3
Recurrent Neural Networks
Recurrent neural networks constitute a broad class of machines with dynamic state; that
is, they have state whose evolution depends both on the input to the system and on the
current state. In comparison to hidden Markov models, which also contain dynamic state,
RNNs have a distributed state and therefore have significantly larger and richer memory
and computational capacity. Dynamic state is crucial because it affords the possibility of
context-dependent computation; a signal entering at a given moment can alter the behaviour
of the network at a much later moment.
A crucial innovation to recurrent networks was the Long Short-Term Memory (LSTM)
(Hochreiter and Schmidhuber, 1997). This very general architecture was developed for a
specific purpose, to address the “vanishing and exploding gradient” problem (Hochreiter
et al., 2001a), which we might relabel the problem of “vanishing and exploding sensitivity.”
LSTM ameliorates the problem by embedding perfect integrators (Seung, 1998) for memory storage in the network. The simplest example of a perfect integrator is the equation
x(t + 1) = x(t) + i(t), where i(t) is an input to the system. The implicit identity matrix
Ix(t) means that signals do not dynamically vanish or explode. If we attach a mechanism
to this integrator that allows an enclosing network to choose when the integrator listens to
inputs, namely, a programmable gate depending on context, we have an equation of the
form x(t + 1) = x(t) + g(context)i(t). We can now selectively store information for an
indefinite length of time.
Recurrent networks readily process variable-length structures without modification. In
sequential problems, inputs to the network arrive at different times, allowing variablelength or composite structures to be processed over multiple steps. Because they natively
handle variable-length structures, they have recently been used in a variety of cognitive
problems, including speech recognition (Graves et al., 2013; Graves and Jaitly, 2014), text
generation (Sutskever et al., 2011), handwriting generation (Graves, 2013) and machine
translation (Sutskever et al., 2014). Considering this property, we do not feel that it is urgent or even necessarily valuable to build explicit parse trees to merge composite structures
greedily (Pollack, 1990) (Socher et al., 2012) (Frasconi et al., 1998).
Other important precursors to our work include differentiable models of attention (Graves,
4
Figure 1: Neural Turing Machine Architecture. During each update cycle, the controller
network receives inputs from an external environment and emits outputs in response. It also
reads to and writes from a memory matrix via a set of parallel read and write heads. The dashed
line indicates the division between the NTM circuit and the outside world.
2013) (Bahdanau et al., 2014) and program search (Hochreiter et al., 2001b) (Das et al.,
1992), constructed with recurrent neural networks.
3
Neural Turing Machines
A Neural Turing Machine (NTM) architecture contains two basic components: a neural
network controller and a memory bank. Figure 1 presents a high-level diagram of the NTM
architecture. Like most neural networks, the controller interacts with the external world via
input and output vectors. Unlike a standard network, it also interacts with a memory matrix
using selective read and write operations. By analogy to the Turing machine we refer to the
network outputs that parametrise these operations as “heads.”
Crucially, every component of the architecture is differentiable, making it straightforward to train with gradient descent. We achieved this by defining ‘blurry’ read and write
operations that interact to a greater or lesser degree with all the elements in memory (rather
than addressing a single element, as in a normal Turing machine or digital computer). The
degree of blurriness is determined by an attentional “focus” mechanism that constrains each
read and write operation to interact with a small portion of the memory, while ignoring the
rest. Because interaction with the memory is highly sparse, the NTM is biased towards
storing data without interference. The memory location brought into attentional focus is
determined by specialised outputs emitted by the heads. These outputs define a normalised
weighting over the rows in the memory matrix (referred to as memory “locations”). Each
weighting, one per read or write head, defines the degree to which the head reads or writes
5
at each location. A head can thereby attend sharply to the memory at a single location or
weakly to the memory at many locations.
3.1
Reading
Let Mt be the contents of the N × M memory matrix at time t, where N is the number
of memory locations, and M is the vector size at each location. Let wt be a vector of
weightings over the N locations emitted by a read head at time t. Since all weightings are
normalised, the N elements wt (i) of wt obey the following constraints:
X
wt (i) = 1,
0 ≤ wt (i) ≤ 1, ∀i.
(1)
i
The length M read vector rt returned by the head is defined as a convex combination of
the row-vectors Mt (i) in memory:
X
rt ←−
wt (i)Mt (i),
(2)
i
which is clearly differentiable with respect to both the memory and the weighting.
3.2
Writing
Taking inspiration from the input and forget gates in LSTM, we decompose each write into
two parts: an erase followed by an add.
Given a weighting wt emitted by a write head at time t, along with an erase vector
et whose M elements all lie in the range (0, 1), the memory vectors Mt−1 (i) from the
previous time-step are modified as follows:
M̃t (i) ←− Mt−1 (i) [1 − wt (i)et ] ,
(3)
where 1 is a row-vector of all 1-s, and the multiplication against the memory location acts
point-wise. Therefore, the elements of a memory location are reset to zero only if both the
weighting at the location and the erase element are one; if either the weighting or the erase
is zero, the memory is left unchanged. When multiple write heads are present, the erasures
can be performed in any order, as multiplication is commutative.
Each write head also produces a length M add vector at , which is added to the memory
after the erase step has been performed:
Mt (i) ←− M̃t (i) + wt (i) at .
(4)
Once again, the order in which the adds are performed by multiple heads is irrelevant. The
combined erase and add operations of all the write heads produces the final content of the
memory at time t. Since both erase and add are differentiable, the composite write operation is differentiable too. Note that both the erase and add vectors have M independent
components, allowing fine-grained control over which elements in each memory location
are modified.
6
Figure 2: Flow Diagram of the Addressing Mechanism. The key vector, kt , and key
strength, βt , are used to perform content-based addressing of the memory matrix, Mt . The
resulting content-based weighting is interpolated with the weighting from the previous time step
based on the value of the interpolation gate, gt . The shift weighting, st , determines whether
and by how much the weighting is rotated. Finally, depending on γt , the weighting is sharpened
and used for memory access.
3.3
Addressing Mechanisms
Although we have now shown the equations of reading and writing, we have not described
how the weightings are produced. These weightings arise by combining two addressing
mechanisms with complementary facilities. The first mechanism, “content-based addressing,” focuses attention on locations based on the similarity between their current values
and values emitted by the controller. This is related to the content-addressing of Hopfield
networks (Hopfield, 1982). The advantage of content-based addressing is that retrieval is
simple, merely requiring the controller to produce an approximation to a part of the stored
data, which is then compared to memory to yield the exact stored value.
However, not all problems are well-suited to content-based addressing. In certain tasks
the content of a variable is arbitrary, but the variable still needs a recognisable name or address. Arithmetic problems fall into this category: the variable x and the variable y can take
on any two values, but the procedure f (x, y) = x × y should still be defined. A controller
for this task could take the values of the variables x and y, store them in different addresses,
then retrieve them and perform a multiplication algorithm. In this case, the variables are
addressed by location, not by content. We call this form of addressing “location-based addressing.” Content-based addressing is strictly more general than location-based addressing
as the content of a memory location could include location information inside it. In our experiments however, providing location-based addressing as a primitive operation proved
essential for some forms of generalisation, so we employ both mechanisms concurrently.
Figure 2 presents a flow diagram of the entire addressing system that shows the order
of operations for constructing a weighting vector when reading or writing.
7
3.3.1
Focusing by Content
For content-addressing, each head (whether employed for reading or writing) first produces
a length
M key vector kt that is compared to each vector Mt (i) by a csimilarity measure
K ·, · . The content-based system produces a normalised weighting wt based on the similarity and a positive key strength, βt , which can amplify or attenuate the precision of the
focus:
exp βt K kt , Mt (i)
c
.
(5)
wt (i) ←−
P
j exp βt K kt , Mt (j)
In our current implementation, the similarity measure is cosine similarity:
K u, v =
3.3.2
u·v
.
||u|| · ||v||
(6)
Focusing by Location
The location-based addressing mechanism is designed to facilitate both simple iteration
across the locations of the memory and random-access jumps. It does so by implementing
a rotational shift of a weighting. For example, if the current weighting focuses entirely on
a single location, a rotation of 1 would shift the focus to the next location. A negative shift
would move the weighting in the opposite direction.
Prior to rotation, each head emits a scalar interpolation gate gt in the range (0, 1). The
value of g is used to blend between the weighting wt−1 produced by the head at the previous
time-step and the weighting wtc produced by the content system at the current time-step,
yielding the gated weighting wtg :
wtg ←− gt wtc + (1 − gt )wt−1 .
(7)
If the gate is zero, then the content weighting is entirely ignored, and the weighting from the
previous time step is used. Conversely, if the gate is one, the weighting from the previous
iteration is ignored, and the system applies content-based addressing.
After interpolation, each head emits a shift weighting st that defines a normalised distribution over the allowed integer shifts. For example, if shifts between -1 and 1 are allowed,
st has three elements corresponding to the degree to which shifts of -1, 0 and 1 are performed. The simplest way to define the shift weightings is to use a softmax layer of the
appropriate size attached to the controller. We also experimented with another technique,
where the controller emits a single scalar that is interpreted as the lower bound of a width
one uniform distribution over shifts. For example, if the shift scalar is 6.7, then st (6) = 0.3,
st (7) = 0.7, and the rest of st is zero.
8
If we index the N memory locations from 0 to N − 1, the rotation applied to wtg by st
can be expressed as the following circular convolution:
w̃t (i) ←−
N
−1
X
wtg (j) st (i − j)
(8)
j=0
where all index arithmetic is computed modulo N . The convolution operation in Equation (8) can cause leakage or dispersion of weightings over time if the shift weighting is
not sharp. For example, if shifts of -1, 0 and 1 are given weights of 0.1, 0.8 and 0.1, the
rotation will transform a weighting focused at a single point into one slightly blurred over
three points. To combat this, each head emits one further scalar γt ≥ 1 whose effect is to
sharpen the final weighting as follows:
w̃t (i)γt
wt (i) ←− P
γt
j w̃t (j)
(9)
The combined addressing system of weighting interpolation and content and locationbased addressing can operate in three complementary modes. One, a weighting can be
chosen by the content system without any modification by the location system. Two, a
weighting produced by the content addressing system can be chosen and then shifted. This
allows the focus to jump to a location next to, but not on, an address accessed by content;
in computational terms this allows a head to find a contiguous block of data, then access a
particular element within that block. Three, a weighting from the previous time step can
be rotated without any input from the content-based addressing system. This allows the
weighting to iterate through a sequence of addresses by advancing the same distance at
each time-step.
3.4
Controller Network
The NTM architecture architecture described above has several free parameters, including
the size of the memory, the number of read and write heads, and the range of allowed location shifts. But perhaps the most significant architectural choice is the type of neural
network used as the controller. In particular, one has to decide whether to use a recurrent
or feedforward network. A recurrent controller such as LSTM has its own internal memory
that can complement the larger memory in the matrix. If one compares the controller to
the central processing unit in a digital computer (albeit with adaptive rather than predefined
instructions) and the memory matrix to RAM, then the hidden activations of the recurrent
controller are akin to the registers in the processor. They allow the controller to mix information across multiple time steps of operation. On the other hand a feedforward controller
can mimic a recurrent network by reading and writing at the same location in memory at
every step. Furthermore, feedforward controllers often confer greater transparency to the
network’s operation because the pattern of reading from and writing to the memory matrix
is usually easier to interpret than the internal state of an RNN. However, one limitation of
9
a feedforward controller is that the number of concurrent read and write heads imposes a
bottleneck on the type of computation the NTM can perform. With a single read head, it
can perform only a unary transform on a single memory vector at each time-step, with two
read heads it can perform binary vector transforms, and so on. Recurrent controllers can
internally store read vectors from previous time-steps, so do not suffer from this limitation.
4
Experiments
This section presents preliminary experiments on a set of simple algorithmic tasks such
as copying and sorting data sequences. The goal was not only to establish that NTM is
able to solve the problems, but also that it is able to do so by learning compact internal
programs. The hallmark of such solutions is that they generalise well beyond the range of
the training data. For example, we were curious to see if a network that had been trained
to copy sequences of length up to 20 could copy a sequence of length 100 with no further
training.
For all the experiments we compared three architectures: NTM with a feedforward
controller, NTM with an LSTM controller, and a standard LSTM network. Because all
the tasks were episodic, we reset the dynamic state of the networks at the start of each
input sequence. For the LSTM networks, this meant setting the previous hidden state equal
to a learned bias vector. For NTM the previous state of the controller, the value of the
previous read vectors, and the contents of the memory were all reset to bias values. All
the tasks were supervised learning problems with binary targets; all networks had logistic
sigmoid output layers and were trained with the cross-entropy objective function. Sequence
prediction errors are reported in bits-per-sequence. For more details about the experimental
parameters see Section 4.6.
4.1
Copy
The copy task tests whether NTM can store and recall a long sequence of arbitrary information. The network is presented with an input sequence of random binary vectors
followed by a delimiter flag. Storage and access of information over long time periods has
always been problematic for RNNs and other dynamic architectures. We were particularly
interested to see if an NTM is able to bridge longer time delays than LSTM.
The networks were trained to copy sequences of eight bit random vectors, where the
sequence lengths were randomised between 1 and 20. The target sequence was simply a
copy of the input sequence (without the delimiter flag). Note that no inputs were presented
to the network while it receives the targets, to ensure that it recalls the entire sequence with
no intermediate assistance.
As can be seen from Figure 3, NTM (with either a feedforward or LSTM controller)
learned much faster than LSTM alone, and converged to a lower cost. The disparity between the NTM and LSTM learning curves is dramatic enough to suggest a qualitative,
10
cost per sequence (bits)
10
LSTM
NTM with LSTM Controller
NTM with Feedforward Controller
8
6
4
2
0
0
200
400
600
800
1000
sequence number (thousands)
Figure 3: Copy Learning Curves.
rather than quantitative, difference in the way the two models solve the problem.
We also studied the ability of the networks to generalise to longer sequences than seen
during training (that they can generalise to novel vectors is clear from the training error).
Figures 4 and 5 demonstrate that the behaviour of LSTM and NTM in this regime is radically different. NTM continues to copy as the length increases2 , while LSTM rapidly
degrades beyond length 20.
The preceding analysis suggests that NTM, unlike LSTM, has learned some form of
copy algorithm. To determine what this algorithm is, we examined the interaction between
the controller and the memory (Figure 6). We believe that the sequence of operations performed by the network can be summarised by the following pseudocode:
initialise: move head to start location
while input delimiter not seen do
receive input vector
write input to head location
increment head location by 1
end while
return head to start location
while true do
read output vector from head location
emit output
increment head location by 1
end while
This is essentially how a human programmer would perform the same task in a low2
The limiting factor was the size of the memory (128 locations), after which the cyclical shifts wrapped
around and previous writes were overwritten.
11
Figure 4: NTM Generalisation on the Copy Task. The four pairs of plots in the top row
depict network outputs and corresponding copy targets for test sequences of length 10, 20, 30,
and 50, respectively. The plots in the bottom row are for a length 120 sequence. The network
was only trained on sequences of up to length 20. The first four sequences are reproduced with
high confidence and very few mistakes. The longest one has a few more local errors and one
global error: at the point indicated by the red arrow at the bottom, a single vector is duplicated,
pushing all subsequent vectors one step back. Despite being subjectively close to a correct copy,
this leads to a high loss.
level programming language. In terms of data structures, we could say that NTM has
learned how to create and iterate through arrays. Note that the algorithm combines both
content-based addressing (to jump to start of the sequence) and location-based addressing (to move along the sequence). Also note that the iteration would not generalise to
long sequences without the ability to use relative shifts from the previous read and write
weightings (Equation 7), and that without the focus-sharpening mechanism (Equation 9)
the weightings would probably lose precision over time.
4.2
Repeat Copy
The repeat copy task extends copy by requiring the network to output the copied sequence a
specified number of times and then emit an end-of-sequence marker. The main motivation
was to see if the NTM could learn a simple nested function. Ideally, we would like it to be
able to execute a “for loop” containing any subroutine it has already learned.
The network receives random-length sequences of random binary vectors, followed by
a scalar value indicating the desired number of copies, which appears on a separate input
channel. To emit the end marker at the correct time the network must be both able to
interpret the extra input and keep count of the number of copies it has performed so far.
As with the copy task, no inputs are provided to the network after the initial sequence and
repeat number. The networks were trained to reproduce sequences of size eight random
binary vectors, where both the sequence length and the number of repetitions were chosen
randomly from one to ten. The input representing the repeat number was normalised to
have mean zero and variance one.
12
Figure 5: LSTM Generalisation on the Copy Task. The plots show inputs and outputs
for the same sequence lengths as Figure 4. Like NTM, LSTM learns to reproduce sequences
of up to length 20 almost perfectly. However it clearly fails to generalise to longer sequences.
Also note that the length of the accurate prefix decreases as the sequence length increases,
suggesting that the network has trouble retaining information for long periods.
Figure 6: NTM Memory Use During the Copy Task. The plots in the left column depict
the inputs to the network (top), the vectors added to memory (middle) and the corresponding
write weightings (bottom) during a single test sequence for the copy task. The plots on the right
show the outputs from the network (top), the vectors read from memory (middle) and the read
weightings (bottom). Only a subset of memory locations are shown. Notice the sharp focus of
all the weightings on a single location in memory (black is weight zero, white is weight one).
Also note the translation of the focal point over time, reflects the network’s use of iterative
shifts for location-based addressing, as described in Section 3.3.2. Lastly, observe that the read
locations exactly match the write locations, and the read vectors match the add vectors. This
suggests that the network writes each input vector in turn to a specific memory location during
the input phase, then reads from the same location sequence during the output phase.
13
cost per sequence (bits)
200
180
160
140
120
100
80
60
40
20
0
LSTM
NTM with LSTM Controller
NTM with Feedforward Controller
0
100
200
300
400
500
sequence number (thousands)
Figure 7: Repeat Copy Learning Curves.
Figure 7 shows that NTM learns the task much faster than LSTM, but both were able to
solve it perfectly.3 The difference between the two architectures only becomes clear when
they are asked to generalise beyond the training data. In this case we were interested in
generalisation along two dimensions: sequence length and number of repetitions. Figure 8
illustrates the effect of doubling first one, then the other, for both LSTM and NTM. Whereas
LSTM fails both tests, NTM succeeds with longer sequences and is able to perform more
than ten repetitions; however it is unable to keep count of of how many repeats it has
completed, and does not predict the end marker correctly. This is probably a consequence
of representing the number of repetitions numerically, which does not easily generalise
beyond a fixed range.
Figure 9 suggests that NTM learns a simple extension of the copy algorithm in the
previous section, where the sequential read is repeated as many times as necessary.
4.3
Associative Recall
The previous tasks show that the NTM can apply algorithms to relatively simple, linear data
structures. The next order of complexity in organising data arises from “indirection”—that
is, when one data item points to another. We test the NTM’s capability for learning an
instance of this more interesting class by constructing a list of items so that querying with
one of the items demands that the network return the subsequent item. More specifically,
we define an item as a sequence of binary vectors that is bounded on the left and right
by delimiter symbols. After several items have been propagated to the network, we query
by showing a random item, and we ask the network to produce the next item. In our
experiments, each item consisted of three six-bit binary vectors (giving a total of 18 bits
3
It surprised us that LSTM performed better here than on the copy problem. The likely reasons are that the
sequences were shorter (up to length 10 instead of up to 20), and the LSTM network was larger and therefore
had more memory capacity.
14
Figure 8: NTM and LSTM Generalisation for the Repeat Copy Task. NTM generalises
almost perfectly to longer sequences than seen during training. When the number of repeats is
increased it is able to continue duplicating the input sequence fairly accurately; but it is unable
to predict when the sequence will end, emitting the end marker after the end of every repetition
beyond the eleventh. LSTM struggles with both increased length and number, rapidly diverging
from the input sequence in both cases.
per item). During training, we used a minimum of 2 items and a maximum of 6 items in a
single episode.
Figure 10 shows that NTM learns this task significantly faster than LSTM, terminating
at near zero cost within approximately 30, 000 episodes, whereas LSTM does not reach
zero cost after a million episodes. Additionally, NTM with a feedforward controller learns
faster than NTM with an LSTM controller. These two results suggest that NTM’s external
memory is a more effective way of maintaining the data structure than LSTM’s internal
state. NTM also generalises much better to longer sequences than LSTM, as can be seen
in Figure 11. NTM with a feedforward controller is nearly perfect for sequences of up to
12 items (twice the maximum length used in training), and still has an average cost below
1 bit per sequence for sequences of 15 items.
In Figure 12, we show the operation of the NTM memory, controlled by an LSTM
with one head, on a single test episode. In “Inputs,” we see that the input denotes item
delimiters as single bits in row 7. After the sequence of items has been propagated, a
15
cost per sequence (bits)
Figure 9: NTM Memory Use During the Repeat Copy Task. As with the copy task the
network first writes the input vectors to memory using iterative shifts. It then reads through
the sequence to replicate the input as many times as necessary (six in this case). The white dot
at the bottom of the read weightings seems to correspond to an intermediate location used to
redirect the head to the start of the sequence (The NTM equivalent of a goto statement).
20
18
16
14
12
10
8
6
4
2
0
LSTM
NTM with LSTM Controller
NTM with Feedforward Controller
0
200
400
600
800
1000
sequence number (thousands)
Figure 10: Associative Recall Learning Curves for NTM and LSTM.
16
co st p er seq u en ce (bits)
40
35
30
25
LS T M
N T M w ith LS T M C o ntroller
N T M w ith F eed fo rw a rd C o n troller
20
15
10
5
0
6
8
10
12
14
16
18
20
n um be r of item s per se que nce
Figure 11: Generalisation Performance on Associative Recall for Longer Item Sequences.
The NTM with either a feedforward or LSTM controller generalises to much longer sequences
of items than the LSTM alone. In particular, the NTM with a feedforward controller is nearly
perfect for item sequences of twice the length of sequences in its training set.
delimiter in row 8 prepares the network to receive a query item. In this case, the query
item corresponds to the second item in the sequence (contained in the green box). In
“Outputs,” we see that the network crisply outputs item 3 in the sequence (from the red
box). In “Read Weightings,” on the last three time steps, we see that the controller reads
from contiguous locations that each store the time slices of item 3. This is curious because it
appears that the network has jumped directly to the correct location storing item 3. However
we can explain this behaviour by looking at “Write Weightings.” Here we see that the
memory is written to even when the input presents a delimiter symbol between items.
One can confirm in “Adds” that data are indeed written to memory when the delimiters
are presented (e.g., the data within the black box); furthermore, each time a delimiter is
presented, the vector added to memory is different. Further analysis of the memory reveals
that the network accesses the location it reads after the query by using a content-based
lookup that produces a weighting that is shifted by one. Additionally, the key used for
content-lookup corresponds to the vector that was added in the black box. This implies the
following memory-access algorithm: when each item delimiter is presented, the controller
writes a compressed representation of the previous three time slices of the item. After the
query arrives, the controller recomputes the same compressed representation of the query
item, uses a content-based lookup to find the location where it wrote the first representation,
and then shifts by one to produce the subsequent item in the sequence (thereby combining
content-based lookup with location-based offsetting).
4.4
Dynamic N-Grams
The goal of the dynamic N-Grams task was to test whether NTM could rapidly adapt to
new predictive distributions. In particular we were interested to see if it were able to use its
17
Figure 12: NTM Memory Use During the Associative Recall Task. In “Inputs,” a sequence of items, each composed of three consecutive binary random vectors is propagated to the
controller. The distinction between items is designated by delimiter symbols (row 7 in “Inputs”).
After several items have been presented, a delimiter that designates a query is presented (row 8
in “Inputs”). A single query item is presented (green box), and the network target corresponds
to the subsequent item in the sequence (red box). In “Outputs,” we see that the network correctly produces the target item. The red boxes in the read and write weightings highlight the
three locations where the target item was written and then read. The solution the network finds
is to form a compressed representation (black box in “Adds”) of each item that it can store in
a single location. For further analysis, see the main text.
memory as a re-writable table that it could use to keep count of transition statistics, thereby
emulating a conventional N-Gram model.
We considered the set of all possible 6-Gram distributions over binary sequences. Each
6-Gram distribution can be expressed as a table of 25 = 32 numbers, specifying the probability that the next bit will be one, given all possible length five binary histories. For
each training example, we first generated random 6-Gram probabilities by independently
drawing all 32 probabilities from the Beta( 21 , 12 ) distribution.
We then generated a particular training sequence by drawing 200 successive bits using
the current lookup table.4 The network observes the sequence one bit at a time and is then
asked to predict the next bit. The optimal estimator for the problem can be determined by
4
The first 5 bits, for which insufficient context exists to sample from the table, are drawn i.i.d. from a
Bernoulli distribution with p = 0.5.
18
cost per sequence (bits)
160
LSTM
NTM with LSTM Controller
NTM with Feedforward Controller
Optimal Estimator
155
150
145
140
135
130
0
200
400
600
800
1000
sequence number (thousands)
Figure 13: Dynamic N-Gram Learning Curves.
Bayesian analysis (Murphy, 2012):
P (B = 1|N1 , N0 , c) =
N1 + 21
N1 + N0 + 1
(10)
where c is the five bit previous context, B is the value of the next bit and N0 and N1 are
respectively the number of zeros and ones observed after c so far in the sequence. We can
therefore compare NTM to the optimal predictor as well as LSTM. To assess performance
we used a validation set of 1000 length 200 sequences sampled from the same distribution as the training data. As shown in Figure 13, NTM achieves a small, but significant
performance advantage over LSTM, but never quite reaches the optimum cost.
The evolution of the two architecture’s predictions as they observe new inputs is shown
in Figure 14, along with the optimal predictions. Close analysis of NTM’s memory usage
(Figure 15) suggests that the controller uses the memory to count how many ones and zeros
it has observed in different contexts, allowing it to implement an algorithm similar to the
optimal estimator.
4.5
Priority Sort
This task tests whether the NTM can sort data—an important elementary algorithm. A
sequence of random binary vectors is input to the network along with a scalar priority
rating for each vector. The priority is drawn uniformly from the range [-1, 1]. The target
sequence contains the binary vectors sorted according to their priorities, as depicted in
Figure 16.
Each input sequence contained 20 binary vectors with corresponding priorities, and
each target sequence was the 16 highest-priority vectors in the input.5 Inspection of NTM’s
5
We limited the sort to size 16 because we were interested to see if NTM would solve the task using a
binary heap sort of depth 4.
19
Figure 14: Dynamic N-Gram Inference. The top row shows a test sequence from the N-Gram
task, and the rows below show the corresponding predictive distributions emitted by the optimal
estimator, NTM, and LSTM. In most places the NTM predictions are almost indistinguishable
from the optimal ones. However at the points indicated by the two arrows it makes clear
mistakes, one of which is explained in Figure 15. LSTM follows the optimal predictions closely
in some places but appears to diverge further as the sequence progresses; we speculate that this
is due to LSTM “forgetting” the observations at the start of the sequence.
Figure 15: NTM Memory Use During the Dynamic N-Gram Task. The red and green
arrows indicate point where the same context is repeatedly observed during the test sequence
(“00010” for the green arrows, “01111” for the red arrows). At each such point the same
location is accessed by the read head, and then, on the next time-step, accessed by the write
head. We postulate that the network uses the writes to keep count of the fraction of ones and
zeros following each context in the sequence so far. This is supported by the add vectors, which
are clearly anti-correlated at places where the input is one or zero, suggesting a distributed
“counter.” Note that the write weightings grow fainter as the same context is repeatedly seen;
this may be because the memory records a ratio of ones to zeros, rather than absolute counts.
The red box in the prediction sequence corresponds to the mistake at the first red arrow in
Figure 14; the controller appears to have accessed the wrong memory location, as the previous
context was “01101” and not “01111.”
20
Figure 16: Example Input and Target Sequence for the Priority Sort Task. The input
sequence contains random binary vectors and random scalar priorities. The target sequence is a
subset of the input vectors sorted by the priorities.
Write Weightings
Read Weightings
Locatiobn
Hypothesised Locations
Time
Time
Time
Figure 17: NTM Memory Use During the Priority Sort Task. Left: Write locations
returned by fitting a linear function of the priorities to the observed write locations. Middle:
Observed write locations. Right: Read locations.
memory use led us to hypothesise that it uses the priorities to determine the relative location
of each write. To test this hypothesis we fitted a linear function of the priority to the
observed write locations. Figure 17 shows that the locations returned by the linear function
closely match the observed write locations. It also shows that the network reads from the
memory locations in increasing order, thereby traversing the sorted sequence.
The learning curves in Figure 18 demonstrate that NTM with both feedforward and
LSTM controllers substantially outperform LSTM on this task. Note that eight parallel
read and write heads were needed for best performance with a feedforward controller on
this task; this may reflect the difficulty of sorting vectors using only unary vector operations
(see Section 3.4).
4.6
Experimental Details
For all experiments, the RMSProp algorithm was used for training in the form described
in (Graves, 2013) with momentum of 0.9. Tables 1 to 3 give details about the network
configurations and learning rates used in the experiments. All LSTM networks had three
stacked hidden layers. Note that the number of LSTM parameters grows quadratically with
21
cost per sequence (bits)
140
LSTM
NTM with LSTM Controller
NTM with Feedforward Controller
120
100
80
60
40
20
0
0
200
400
600
800
1000
sequence number (thousands)
Figure 18: Priority Sort Learning Curves.
Task
Copy
Repeat Copy
Associative
N-Grams
Priority Sort
#Heads
Controller Size
Memory Size
Learning Rate
1
1
4
1
8
100
100
256
100
512
128 × 20
128 × 20
128 × 20
128 × 20
128 × 20
10−4
10−4
10−4
3 × 10−5
3 × 10−5
#Parameters
17, 162
16, 712
146, 845
14, 656
508, 305
Table 1: NTM with Feedforward Controller Experimental Settings
the number of hidden units (due to the recurrent connections in the hidden layers). This
contrasts with NTM, where the number of parameters does not increase with the number of
memory locations. During the training backward pass, all gradient components are clipped
elementwise to the range (-10, 10).
5
Conclusion
We have introduced the Neural Turing Machine, a neural network architecture that takes
inspiration from both models of biological working memory and the design of digital computers. Like conventional neural networks, the architecture is differentiable end-to-end and
can be trained with gradient descent. Our experiments demonstrate that it is capable of
learning simple algorithms from example data and of using these algorithms to generalise
well outside its training regime.
22
Task
#Heads
Controller Size
Memory Size
Learning Rate
1
1
1
1
5
100
100
100
100
2 × 100
128 × 20
128 × 20
128 × 20
128 × 20
128 × 20
10−4
10−4
10−4
3 × 10−5
3 × 10−5
Copy
Repeat Copy
Associative
N-Grams
Priority Sort
#Parameters
67, 561
66, 111
70, 330
61, 749
269, 038
Table 2: NTM with LSTM Controller Experimental Settings
Task
Network Size
Learning Rate
#Parameters
3 × 256
3 × 512
3 × 256
3 × 128
3 × 128
3 × 10−5
3 × 10−5
10−4
10−4
3 × 10−5
1, 352, 969
5, 312, 007
1, 344, 518
331, 905
384, 424
Copy
Repeat Copy
Associative
N-Grams
Priority Sort
Table 3: LSTM Network Experimental Settings
6
Acknowledgments
Many have offered thoughtful insights, but we would especially like to thank Daan Wierstra, Peter Dayan, Ilya Sutskever, Charles Blundell, Joel Veness, Koray Kavukcuoglu,
Dharshan Kumaran, Georg Ostrovski, Chris Summerfield, Jeff Dean, Geoffrey Hinton, and
Demis Hassabis.
23
References
Baddeley, A., Eysenck, M., and Anderson, M. (2009). Memory. Psychology Press.
Bahdanau, D., Cho, K., and Bengio, Y. (2014). Neural machine translation by jointly
learning to align and translate. abs/1409.0473.
Barrouillet, P., Bernardin, S., and Camos, V. (2004). Time constraints and resource sharing in adults’ working memory spans. Journal of Experimental Psychology: General,
133(1):83.
Chomsky, N. (1956). Three models for the description of language. Information Theory,
IEEE Transactions on, 2(3):113–124.
Das, S., Giles, C. L., and Sun, G.-Z. (1992). Learning context-free grammars: Capabilities and limitations of a recurrent neural network with an external stack memory. In
Proceedings of The Fourteenth Annual Conference of Cognitive Science Society. Indiana
University.
Dayan, P. (2008). Simple substrates for complex cognition. Frontiers in neuroscience,
2(2):255.
Eliasmith, C. (2013). How to build a brain: A neural architecture for biological cognition.
Oxford University Press.
Fitch, W., Hauser, M. D., and Chomsky, N. (2005). The evolution of the language faculty:
clarifications and implications. Cognition, 97(2):179–210.
Fodor, J. A. and Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A
critical analysis. Cognition, 28(1):3–71.
Frasconi, P., Gori, M., and Sperduti, A. (1998). A general framework for adaptive processing of data structures. Neural Networks, IEEE Transactions on, 9(5):768–786.
Gallistel, C. R. and King, A. P. (2009). Memory and the computational brain: Why cognitive science will transform neuroscience, volume 3. John Wiley & Sons.
Goldman-Rakic, P. S. (1995). Cellular basis of working memory. Neuron, 14(3):477–485.
Graves, A. (2013). Generating sequences with recurrent neural networks. arXiv preprint
arXiv:1308.0850.
Graves, A. and Jaitly, N. (2014). Towards end-to-end speech recognition with recurrent
neural networks. In Proceedings of the 31st International Conference on Machine Learning (ICML-14), pages 1764–1772.
24
Graves, A., Mohamed, A., and Hinton, G. (2013). Speech recognition with deep recurrent
neural networks. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE
International Conference on, pages 6645–6649. IEEE.
Hadley, R. F. (2009). The problem of rapid variable creation. Neural computation,
21(2):510–532.
Hazy, T. E., Frank, M. J., and O’Reilly, R. C. (2006). Banishing the homunculus: making
working memory work. Neuroscience, 139(1):105–118.
Hinton, G. E. (1986). Learning distributed representations of concepts. In Proceedings
of the eighth annual conference of the cognitive science society, volume 1, page 12.
Amherst, MA.
Hochreiter, S., Bengio, Y., Frasconi, P., and Schmidhuber, J. (2001a). Gradient flow in
recurrent nets: the difficulty of learning long-term dependencies.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory. Neural computation,
9(8):1735–1780.
Hochreiter, S., Younger, A. S., and Conwell, P. R. (2001b). Learning to learn using gradient
descent. In Artificial Neural Networks?ICANN 2001, pages 87–94. Springer.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective
computational abilities. Proceedings of the national academy of sciences, 79(8):2554–
2558.
Jackendoff, R. and Pinker, S. (2005). The nature of the language faculty and its implications
for evolution of language (reply to fitch, hauser, and chomsky). Cognition, 97(2):211–
225.
Kanerva, P. (2009). Hyperdimensional computing: An introduction to computing in distributed representation with high-dimensional random vectors. Cognitive Computation,
1(2):139–159.
Marcus, G. F. (2003). The algebraic mind: Integrating connectionism and cognitive science. MIT press.
Miller, G. A. (1956). The magical number seven, plus or minus two: some limits on our
capacity for processing information. Psychological review, 63(2):81.
Miller, G. A. (2003). The cognitive revolution: a historical perspective. Trends in cognitive
sciences, 7(3):141–144.
Minsky, M. L. (1967). Computation: finite and infinite machines. Prentice-Hall, Inc.
Murphy, K. P. (2012). Machine learning: a probabilistic perspective. MIT press.
25
Plate, T. A. (2003). Holographic Reduced Representation: Distributed representation for
cognitive structures. CSLI.
Pollack, J. B. (1990).
46(1):77–105.
Recursive distributed representations.
Artificial Intelligence,
Rigotti, M., Barak, O., Warden, M. R., Wang, X.-J., Daw, N. D., Miller, E. K., and Fusi,
S. (2013). The importance of mixed selectivity in complex cognitive tasks. Nature,
497(7451):585–590.
Rumelhart, D. E., McClelland, J. L., Group, P. R., et al. (1986). Parallel distributed processing, volume 1. MIT press.
Seung, H. S. (1998). Continuous attractors and oculomotor control. Neural Networks,
11(7):1253–1258.
Siegelmann, H. T. and Sontag, E. D. (1995). On the computational power of neural nets.
Journal of computer and system sciences, 50(1):132–150.
Smolensky, P. (1990). Tensor product variable binding and the representation of symbolic
structures in connectionist systems. Artificial intelligence, 46(1):159–216.
Socher, R., Huval, B., Manning, C. D., and Ng, A. Y. (2012). Semantic compositionality
through recursive matrix-vector spaces. In Proceedings of the 2012 Joint Conference on
Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1201–1211. Association for Computational Linguistics.
Sutskever, I., Martens, J., and Hinton, G. E. (2011). Generating text with recurrent neural
networks. In Proceedings of the 28th International Conference on Machine Learning
(ICML-11), pages 1017–1024.
Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Sequence to sequence learning with neural
networks. arXiv preprint arXiv:1409.3215.
Touretzky, D. S. (1990). Boltzcons: Dynamic symbol structures in a connectionist network.
Artificial Intelligence, 46(1):5–46.
Von Neumann, J. (1945). First draft of a report on the edvac.
Wang, X.-J. (1999). Synaptic basis of cortical persistent activity: the importance of nmda
receptors to working memory. The Journal of Neuroscience, 19(21):9587–9603.
26
| 9cs.NE
|
Calibration for the (Computationally-Identifiable) Masses
arXiv:1711.08513v2 [cs.LG] 16 Mar 2018
Úrsula Hébert-Johnson∗
Stanford University
Michael P. Kim†
Stanford University
Omer Reingold‡
Stanford University
Guy N. Rothblum§
Weizmann Institute
Abstract
As algorithms increasingly inform and influence decisions made about individuals, it becomes
increasingly important to address concerns that these algorithms might be discriminatory. The
output of an algorithm can be discriminatory for many reasons, most notably: (1) the data
used to train the algorithm might be biased (in various ways) to favor certain populations over
others; (2) the analysis of this training data might inadvertently or maliciously introduce biases
that are not borne out in the data. This work focuses on the latter concern.
We develop and study multicalbration – a new measure of algorithmic fairness that aims to
mitigate concerns about discrimination that is introduced in the process of learning a predictor
from data. Multicalibration guarantees accurate (calibrated) predictions for every subpopulation
that can be identified within a specified class of computations. We think of the class as being
quite rich; in particular, it can contain many overlapping subgroups of a protected group.
We show that in many settings this strong notion of protection from discrimination is both
attainable and aligned with the goal of obtaining accurate predictions. Along the way, we present
new algorithms for learning a multicalibrated predictor, study the computational complexity of
this task, and draw new connections to computational learning models such as agnostic learning.
∗
[email protected]
[email protected]. Supported in part by NSF grant CNS-122864. Part of this work was completed while the
author was visiting VMWare Research Group.
‡
[email protected]. Supported in part by NSF grant CCF-1749750.
§
[email protected] Research supported by the ISRAEL SCIENCE FOUNDATION (grant No. 5219/17).
†
1
Introduction
Fueled by rapidly growing data sets and by breakthroughs in machine learning, algorithms are
informing decisions that affect all aspects of life. From news article recommendations to criminal
sentencing decisions to healthcare diagnostics, increasingly algorithms are used to make predictions
about individuals. Often, the predictions of an algorithm form the basis for deciding how to treat
these individuals (suggesting a conservative Op-Ed, approving early parole, or initiating chemotherapy). A potential risk is that these algorithms – even given access to unbiased ground truth data –
might discriminate against groups of individuals that are protected by law or by ethics. This paper
aims to mitigate such risks of algorithmic discrimination.
We consider algorithms that predict the probabilities of events occurring for individuals. For
example, a financial institution may be interested in predicting the probability that an individual
will repay a mortgage. The institution may have at its disposal a large array of information for each
individual (as well as historic data and global information such as financial and political trends).
But as thorough as the company may be, a significant level of uncertainty will remain. Just as we
wouldn’t expect to be able to predict with absolute certainty whether it will rain on a particular
day a year from now, the financial institution wouldn’t expect to predict with absolute certainty
whether an individual will repay a loan. Thus, we consider algorithms that output, for every
individual i, a prediction xi of the probability that the event will occur for i; we call the mapping
from individuals to probabilities a predictor.1
Our focus in this paper is mitigating biases that may arise as an algorithm analyzes given data –
specifically, as the algorithm learns a predictor from data. Continuing the above example, suppose
that in a particular protected community S, on average, individuals are financially disadvantaged
and are unlikely to repay a loan. A machine-learning algorithm that aims to optimize the institution’s returns might devote resources to learning outside of S – where there is more opportunity
for gains in utility – and assign a fixed, low probability to all i ∈ S. Such an algorithm would
discriminate against the qualified members of S. If S is an underrepresented subpopulation, this
form of discrimination has the potential to amplify S’s underrepresentation by refusing to approve
members that are capable of repaying the loan.
1.1
Overview of our contributions
Focusing on such concerns, we investigate a notion we call multicalibration. Multicalibration aims
to mitigate discrimination that arises in the process of learning a predictor from given data. In a
nutshell, multicalibration guarantees highly-accurate predictions for every subpopulation of individuals identified by a specified collection C of subsets of individuals. While our results hold for
any set system C, one natural way to think about C is as a collection of subsets defined by a family
of boolean functions. That is, for each S ∈ C, we consider a function cS : X → {0, 1}, where
cS (i) = 1 if and only if i ∈ S. In particular, we will be most interested in considering families
of boolean functions that can be represented by some bounded computational circuit class – for
instance, conjunctions of a small number of boolean features or small decision trees. In the mortgage repayment example above, if the class of qualified members of S can be identified by a circuit
1
A predictor can be used for regression or, when paired with randomized rounding, for binary classification.
1
c ∈ C, then the predictions made on qualified members of S must be accurate, and the prediction
algorithm cannot ignore / discriminate against these individuals. We emphasize that the class C can
be quite rich and, in particular, can contain many overlapping subgroups of a protected group S.
In this sense, multicalibration captures the notion of calibration for all computationally-identifiable
subsets, where the notion of computational-identifiability is parameterized by the expressiveness of
C.
We present a simple, general-purpose algorithm for learning a predictor from a small set of labeled
examples that is multicalibrated with respect to any given class C. The algorithm is an iterative
method, similar to gradient descent, and can be viewed as a form of no-regret online optimization.
While the powerful online learning framework is exceptional in its broad applicability, a number of
subtleties arise when learning a multicalibrated predictor that we must deal with. In particular, the
definition of calibration refers to sets of individuals defined by the values output by the predictor.
This leads to two key challenges. First, we must deal with an appropriate notion of discretization
of the range of real-valued outputs; this challenge is mostly technical. Secondly, the fact that we
want to update sets of predictions defined by the current predictions means that any reduction
to the online learning framework is inherently an adaptive analysis procedure; as such we need to
bring in machinery from the literature on differential privacy and adpative data analysis in order
to guarantee good generalization from a small number of examples.
While we place no restrictions on the model class of the learned predictor, we show that implicitly
the algorithm learns a model that provably generalizes well to unseen data, which may be of independent interest. We demonstrate this implicit generalization by viewing the learning algorithm
as an adaptive data analysis algorithm and showing that the predictions we learn are highly compressible. In the language of circuit complexity, we show that we can build a circuit, only slightly
larger than the circuits from C, that implements the learned predictor. As a corollary, the learned
predictor is efficient in both space to represent and time to evaluate.
We also study the computational complexity of learning multicalibrated predictors for more structured classes C. We show a strong connection between the complexity of learning a multicalibrated
predictor and agnostic learning [Hau92, KSS94]. In the positive direction, if there is an efficient
(weak) agnostic learner [KMV08, Fel10] for a class C, then we can achieve similarly efficient multicalibration with respect to sets defined by C. In the other direction, we show that learning a
multicalibrated predictor on sets defined by C is as hard as weak agnostic learning on C. In this
sense, the complexity of learning a multicalibrated predictor with respect to a class C is equivalent
to the complexity of weak agnostic learning on C.
Finally, we demonstrate that the goal of multicalbration is aligned with the goal of achieving highutility predictions. In particular, given any predictor h, we can use post-processing to obtain a
multicalibrated predictor x whose accuracy is no worse than that of h (accuracy is measured in
ℓ22 distance from the benchmark p∗ ). The complexity of evaluating the predictor x is only slightly
larger than that of h. In this sense, unlike many fairness notions, multicalibration is not at odds
with predictive power and can be paired with any predictive model at essentially no cost to its
accuracy.
Organization. We begin by elaborating on our setting. The remainder of the introduction is
structured as follows. In Section 1.2 we elaborate on the notion of multicalibration and on its
2
relationship to other notions in the larger context of fairness. We outline our main results on
learning multicalibrated predictors in Section 1.3. We further elaborate on related work and on
future directions in Section 1.4. Finally, we provide a brief overview of techniques in Section 1.5.
1.2
Fairness, calibration, and multicalibration
High-level setting. For an individual i from the population X , we denote i’s outcome by oi ∈
{0, 1}. We take p∗i ∈ [0, 1] to be the probability of outcome oi = 1, conditioned on all the information
which is available to the algorithm. We denote by p∗ the vector of these probabilities for all
individuals in X . Our goal is to make a prediction xi for the value of p∗i for every individual i. As
discussed above, we would like to avoid additional malicious or inadvertent discrimination (beyond
the biases contained in the data). Thus, we refer to p∗ as the benchmark predictor for measuring
discrimination.
Calibration If we do not want a predictor x to downplay the fitness of a group
S, ∗we
can require
that it be (approximately) accurate in expectation over S; namely, that Ei∼S xi − pi ≤ α, where
α ≥ 0 is small. This means that the expectation of x and p∗ over S are almost identical. Calibration,
introduced as a fairness concept in [KMR17], strengthens this requirement by essentially asking
that for any particular value v, if we let Sv = {i ∈ S : xi = v} be the subset of S of individuals
with predicted probability v, then Ei∼Sv xi − p∗i = |v − Ei∼Sv [p∗i ]| ≤ α.2
While calibration already precludes some forms of discrimination, as a group notion of fairness, it
still allows for others (even if we assume that p∗ is perfectly fair). Indeed, weaknesses of group
notions of fairness were discussed in [DHP+ 12] (for a somewhat related notion called statistical
parity), as a motivation for introducing an individual notion of fairness (see further discussion and
comparisons below). A specific way to discriminate while satisfying calibration is to assign every
member of S the value Ei∼S [p∗i ]. While being perfectly calibrated over S, the qualified members of
S with large values p∗i will be hurt.
Balance Other notions of fairness have been studied, looking at the rate of false positives and
false negatives of predictions. Several variants of such properties have been recently studied on
their own and in connection to calibration [KMR17, Cho17, PRW+ 17, HPS16, CDPF+ 17]. Let us
briefly consider the notions referred to as balance in [KMR17]: balance for the positive class — the
expected prediction xi for yes instances (oi = 1) in group S equals the expected prediction for yes
instances outside of S; and balance for the negative class — the expected prediction for no instances
(oi = 0) in group S equals the expected prediction for no instances outside of S. In [HPS16] it is
shown how to obtain equalized odds, a definition related to error-rate balance, as a post-processing
step of “correcting” any other predictor.
While both calibration and balance (as well as other related variants) intuitively seem like good
2
Calibration is often defined in the literature with respect to the instantiations of the events rather than their
probabilities. Namely, that for every v, the fraction of i in Sv with oi = 1 is roughly v. As long as Sv is sufficiently
large and the instantiations are sufficiently independent, the two definitions are equivalent (up to small sampling
errors) by concentration bounds. Necessarily, the formal definition given in Supplementary Materials A will allow
for a small fraction of the elements in S to be misclassified due to being in a small set Sv .
3
properties to expect in a fair predictor (even if they are a bit weak), it has been shown that
calibration and balance are impossible to obtain together (in non-degenerate cases) [KMR17,Cho17,
PRW+ 17]. This inherent conflict between balance and calibration, combined with our observation
that calibration is always aligned with the goal of accurate high-utility predictions, implies that
at times, balance must be at odds with obtaining predictive utility. Our approach in this paper
towards mitigating the conflict between calibration and balance is to strengthen the protections
implied by calibration, rather than enforcing balance. While there are certainly contexts in which
“equalizing the odds” across groups is a good idea, there are also contexts where calibration is a
more appropriate notion of fairness. In particular, the benchmark predictor p∗ itself is unlikely
to be balanced. Indeed, in our setting, the fact that balance is not satisfied might simply be an
artifact of the inherent randomness in the process of sampling the outcome oi , and this motivates
our definition of multicalibration.
To illustrate this point, consider the following (intentionally-artificial) example: an algorithm is
tasked with predicting the probability of rain during 10 days of winter, in two cities, a year from
now. In city A the algorithm predicts rain on each day with probability 0.8; in city B it predicts
rain with probability 0.2 (note that the certainty of the predictions is identical for both cities). A
year passes and indeed in city A it rains on 8 of the days, whereas in city B it rains on only 2 days.
What surprising accuracy! Nevertheless, the predictions violate balance; that is, after observing
which days rained, the predictions treat the positive (respectively, negative) examples in city A and
B differently. Indeed, the mayor of city A may complain that the predictor hurt tourism to her
city: “our sunny days are just as sunny as the sunny days of city B, so why were you so much more
pessimistic about ours?” The point is that a priori the sunny and rainy days in A (respectively, B)
were indistinguishable from one another; further, there was a noticeable difference in the likelihood
of rain between A and B. Given the inherent uncertainty, it is unreasonable to expect accuracy on
a subgroup that is only identifiable a posteriori. While this example is overly-simplified, it points
to the fact that in any prediction context where there is inherent unpredictability, different false
negative (or false positive) rates between groups are not necessarily a sign of discrimination.
Multicalibration As illustrated, a principle weakness of calibration as a notion of fairness is
that the guarantees are too coarse. As far as we are aware, in the existing literature, calibration
has been applied to large, often disjoint, sets of protected groups; that is, the guarantees are only
required to hold on average over a population defined by a small number of sensitive attributes,
like race or gender. A stronger definition of fairness would ensure that the predictions on every
subpopulation would be protected, including, for instance, the qualified members of S from the
example above. The problem with such a notion is that it is information-theoretically unattainable
from a small sample of labeled examples, as it would essentially require perfect predictions. As
such, we need an intermediary definition that balances the desire to protect important subgroups
and the information bottleneck that arises when learning from a small sample.
To motivate our notion, consider an algorithm that produces a predictor x. The outcomes oi are
determined, and then an auditor comes up with a set S ′ that over-performed compared with the
predictions of x. Perhaps the learning algorithm was lazy and neglected to identify the higher
potential in S ′ ? Perhaps the individuals of S ′ were simply lucky? How can we tell? To answer
these questions, we take the following perspective: on the one hand, we can only expect a learner
to produce a predictor that is calibrated on sets that could have been identified efficiently from the
4
data at hand; on the other hand, we expect the learner to produce a predictor that is calibrated on
every efficiently-identifiable subset. This motivates our definition of multicalibration, which loosely
says: “A predictor x is α-multicalibrated with respect to a family of sets C if it is α-calibrated with
respect to every S ∈ C.”
In the spirit of the discussion above, we take C to be a family of (sufficiently large) sets of individuals, such that for every S ∈ C, the predicate i ∈ S can be evaluated from the individual’s
data within a particular family of computations (conjunctions of four attributes, bounded-depth
decision trees, or any other bounded complexity class). The more powerful the family, the stronger
the guarantee becomes; no subpopulation that can be identified by the family will be overlooked.
At the extreme, consider multicalibration with respect to the family of polynomial-size circuits; in
this case, every efficiently-identifiable subpopulation is protected! Note that the subpopulations in
C can be overlapping, with complex relationships. In particular, they may well have no explicit
dependence on sensitive attributes. In this sense, multicalibration goes far beyond calibration for
several sensitive groups.
1.3
Our Results
Our study of multicalibration follows two major themes:
• We investigate the feasibility of obtaining multicalibration; specifically, we study the learnability of multicalibrated predictors, showing both positive and negative results.
• We investigate the properties of multicalibrated predictors. While multicalibration provides
strong guarantees against forms of discrimination, we show that this protection can come at
little cost in terms of complexity and utility.
We begin with a high-level overview of our setup. For a formal description of our model and
assumptions, see Section 2. Suppose that for some universe of individuals X , we wish to predict
whether some event (ad click, loan repayment, cancer diagnosis, etc.) will occur for each individual
i ∈ X . We assume that for each individual, there is some true underlying probability p∗i that the
event will occur. We call any mapping from the universe to probabilities a predictor ; formally, a
predictor is a function3 x : X → [0, 1] that maps individuals from the universe to estimates of their
true probabilities. We denote by p∗ the benchmark predictor that gives the true probabilities.
The benchmark predictor is, itself, multicalibrated with respect to every collection of subsets C.
Thus, if we can efficiently learn a predictor from the data at hand with sufficient accuracy across the
entire population (specifically, with small ℓ1 distance from p∗ ), then the learned predictor will be
multicalibrated. That said, in most interesting situations, p∗ will be too complex to learn efficiency
with such uniform accuracy (especially given that the values of p∗ themselves will usually not be
observable). Focusing on such settings, we aim for multicalibration as a notion of protection from
discrimination.
The first question to address is whether multicalibration is feasible. For instance, it could be the
case that the requirements of multicalibration are so strong that they would require learning and
3
We will interchange between function and vector notation; generally we will denote the prediction that x assigns
to an individual i ∈ X as xi .
5
representing an arbitrarily complex function p∗ exactly, which we’ve established can be infeasible.
Our first result characterizes the complexity of representing a multicalbrated predictor. We demonstrate that multicalibration, indeed, can be achieved efficiently: for any p∗ and any collection of
large subsets C, there exists a predictor that is α-multicalibrated on C, whose complexity is only
slightly larger than the complexity required to describe the sets of C. For concreteness, we use
circuit size as our measure of complexity in the following theorem.
Theorem 1. Suppose C ⊆ 2X is collection of sets where for S ∈ C, there is a circuit of size s that
computes membership in S and |S| ≥ γ |X |. For any p∗ : X → [0, 1], there is a predictor that is
α-multicalibrated with respect to C implemented by a circuit of size O(s/α4 γ).
As stated, this result claims the existence of multicalibrated predictors whose predictions are efficient to evaluate. The existence of such predictors, while interesting from a complexity-theoretic
perspective, begs the more practical question of whether we can get our hands on such a predictor.
In fact, we prove Theorem 1 algorithmically by learning an α-multicalibrated predictor from labeled
samples. While our model assumes the existence of some underlying true probabilities p∗ , in most
applications, these probabilities will not be directly observable. As such, we design algorithms that
learn predictors from samples of individuals labeled with their outcomes; specifically, we assume
access to labeled samples (i, oi ) of individual-outcome pairs, where i is sampled according to some
distribution D on the universe and oi is the realized value of a Bernoulli trial with probability p∗i .
Naturally, in this model, our goal is to give algorithms that are efficient in terms of running time
and sample complexity.
In Section 3, we give an algorithm for learning a multicalibrated predictor from labeled samples,
whose running time scales linearly with |C| and polynomially with α and γ. A consequence of our
analysis is that naively, the sample complexity can be upper-bounded by log(|C|)/α6 γ 6 . We show
how to improve the sample complexity over the naive approach by polynomial factors in both α
and γ.
Theorem 2. Suppose C ⊆ 2X is collection of sets such that for all S ∈ C, |S| ≥ γ |X|, and
suppose set membership can be evaluated in time t. Then there is an algorithm that learns a
predictor of p∗ : X → [0, 1] that is α-multicalibrated on C from O(log(|C|)/α11/2 γ 3/2 ) samples in
time O(|C| · t · poly(1/α, 1/γ)).
Observing the linear dependence in the running time on |C|, it is natural to try and develop a
learning procedure with subpolynomial, or even polylogarithmic, dependence on |C|. Our next
results aim to characterize when this optimistic goal is possible – and when it is not. We emphasize
that the algorithm of Theorem 2 learns a multicalibrated predictor for arbitrary p∗ : X → [0, 1]
and C. In the setting where we cannot exploit structure in p∗ to learn efficiently, we might hope to
exploit structure, if it exists, in the collection of subsets C. Indeed, we demonstrate a connection
between our goal of learning a multicalibrated predictor and weak agnostic learning, introduced in
the literature on agnostic boosting [BDLM01, KMV08, KK09, Fel10]. Our next result shows that
efficient weak agnostic learning over C implies efficient learning of α-multicalibrated predictors on
C.
Theorem 3 (Informal). If there is a weak agnostic learner for C that runs in time T , then there
is an algorithm for learning an α-multicalibrated predictor on C ′ = {S ∈ C : |S| γ |X|} that runs in
time O(T · poly(1/α, 1/γ)).
6
Slightly more formally, we require a (ρ, τ )-weak agnostic learner in the sense first introduced by
[KMV08] and generalized by [Fel10]. For the specifics of the requirements and parameters, see the
formal statement in Section 4.
These results show that under the right structural assumptions on p∗ or on C, a multicalibrated
predictor may be learned more efficiently than our upper bound for the general case. Returning
to the general case, we may wonder if these structural assumptions are necessary; we answer this
question in the positive. We show that for worst-case p∗ learning a multicalibrated predictor on C
is as hard as weak agnostic learning for the class C.
Theorem 4 (Informal). If there is an algorithm for learning an α-multicalibrated predictor on a
collection of sets C ′ = {S ∈ C : |S| ≥ γN } that runs in time T , then there is an algorithm that
implements a (ρ, τ )-weak agnostic learner in time O(T · poly(1/τ )) for any ρ > 0 where τ =
poly(ρ, γ, α).
In general, agnostic learning is considered a notoriously hard computational problem. In particular,
under cryptographic assumptions [Val84, GGM84, BR17], this result implies that there is some
constant t > 0, such that any algorithm that learns an α-multicalibrated predictor requires Ω(|C|t )
time.
Finally, we return our attention to investigating the utility of multicalibrated predictors. Above,
we have argued that multicalibration provides a strong protection of groups against discrimination.
We show that this protection comes at (next to) no cost in the utility of the predictor. This result
adds to the growing literature on fairness-accuracy trade-offs [FKL16, BHJ+ 17, CG17].
Theorem 5. Suppose C ⊆ 2X is a collection of subsets of X and H is a set of predictors. There is
a predictor x that is α-multicalibrated on C such that
E [(xi − p∗i )2 ] − E [(h∗i − p∗i )2 ] < 6α,
i∼X
i∼X
where h∗ = argminh∈H Ei∼X [(h − p∗ )2 ]. Further, suppose that for all S ∈ C, |S| ≥ γN , and suppose
that set membership for S ∈ C and h ∈ H are computable by circuits of size at most s; then x is
computable by a circuit of size at most O(s/α4 γ).
We can interpret Theorem 5 in different ways based on the choice of H. Suppose there is some
sophisticated learning algorithm (say, a neural network) that produces some predictor h that obtains
exceptional performance, but may violate calibration arbitrarily. If we take H = {h}, then this
result says: enforcing calibration on h after learning does not hurt the accuracy by much. Further,
our proof will demonstrate that if calibration changes the predictions of h significantly, then this
change amounts to an improvement in accuracy. See Lemma 5.2 for the exact statement. Taking
a different perspective, we can also think of H as a set of predictors that, say, are implemented
by a circuit class of bounded complexity (e.g. conjunctions of k variables, halfspaces, circuits of
size s). This theorem shows that for any such class of predictors H of bounded complexity, there
exists a multicalibrated predictor with similar complexity that performs as well as the best h∗ ∈ H.
In this sense, with just a slight overhead in complexity, multicalibrated predictors can achieve
“best-in-class” predictions.
7
1.4
Related work
Calibration As mentioned earlier, calibration is a well-studied concept in the literature on
statistics and econometrics, particularly forecasting. For a background on calibration in this
context, see [SSV03, FH15] and the references therein. Calibration has also been studied in the
context of structured predictions where the supported set of predictions is large [KL15]. Our
algorithmic result for multicalibration bears similarity to works from the online learning literature [BM07, KP08, TTV09]. In Section 3.1, we describe a procedure for achieving a weaker notion
than multicalibration, which we refer to as multi-“accurate in expectation” (multi-AE); our result
can be viewed as a restatement of the boosting-based proof due to [TTV09] that every high-entropy
distribution is indistinguishable by small circuits from an efficiently-samplable distribution of the
same entropy. While the algorithms for learning a multi-AE and multicalibrated predictor bear
similarity, when we turn to achieving multicalibration, we have to deal with a number of nontrivial
challenges due to learning from a small sample (while guaranteeing generalization) and using an
apprropriate notion of discretization that are not necessary in [TTV09]. We are also not aware of
any prior works connecting the literature on calibration with the literature on differential privacy
and adaptive data analysis.
Between populations and individuals In [DHP+ 12], an individual notion of fairness was defined, referred to as “fairness through awareness.” This notion relied on a task-specific metric of
similarity between individuals and formalized the idea that similar individuals (under this metric)
are treated similarly. It is natural, in an array of applications, to view p∗ as defining a metric –
two individuals i and j will be assigned the distance |p∗i − p∗j |. In this work we consider cases in
which figuring out p∗ in its entirety is difficult. Thus, one can view our approach as a meaningful compromise between group fairness (satisfying calibration) and individual-calibration (closely
matching p∗i ). The multicalibration framework presented in this work served as the inspiration
for subsequent work of a subset of the authors [KRR18] investigating how to interpolate between
statistical and individual notions of “metric fairness” for general similarity metrics.
Subgroup Fairness Contemporary independent work of [KNRW17] also investigates strengthening the guarantees of notions of group fairness by requiring that these properties hold for a much
richer collection of sets. Unlike our work, their definitions require balance or statistical parity on
these collection of sets. Their motivation is similar to ours, namely to bridge the gap between
notions of individual fairness (powerful but hard to obtain) and population-level fairness (easy to
work with but weak).
Despite similar motivations, the two approaches to subgroup fairness differ in substantial ways.
As a concrete example, Theorem 5 demonstrates that achieving multicalibration is aligned with
the incentives of achieving high-utility predictors; this is not necessarily the case with balancebased notions of fairness. Indeed, in the setting considered in this work, one of the motivations
for multicalibration is a critique of balance that may only be heightened when considering “multibalance”. Consider the example in [DHP+ 12] where in a population S the strongest students apply
to Engineering whereas in the general population T they apply to Business. Even if predictor p∗
(for the probability of success in school) is balanced between S and T , forcing balance between
the two populations within Business applicants and within Engineering applicants would be unfair
8
to qualified applicants in both groups. (For more discussion, see the section below on “Corrective
discrimination”.)
On a technical level, both works draw connections between agnostic learning and the task of
finding a group on which the fairness condition is violated ( [KNRW17] refer to this as auditing).
Leveraging this connection, we show how to use an agnostic learner to learn a multicalibrated
predictor efficiently. In a similar vein, [KNRW17] show how to use an agnostic learner to efficiently
learn a classifier that satisfies balance or parity on a collection of sets, while achieving competitive
accuracy within a given class. They also implement a more-practical variant of their algorithm and
use it to conduct empirical experiments.
Preventing discrimination by algorithms is subtle; different scenarios will call for different notions
of protection. As such, it is hard to imagine a universal definition of fairness. Nevertheless,
these two independent works validate the need to investigate attainable approaches to mitigating
discrimination beyond large protected groups. It will be interesting to further understand how these
notions of subgroup fairness relate to one another, and when each approach is most appropriate.
Calibration for multiple protected sets vs. multicalibration The literature on fairness
commonly considers more than a single protected set (cf. [CSV17] which studies fairness in ranking
algorithms, with protected groups being defined by each attribute of interest). The major difference
in our work is that we think of protected groups as any group that can be efficiently identified,
rather than those defined by sensitive properties. One side benefit of the generality of our approach
is that it does not single out groups based on sensitive values for special treatment (which can be
illegal in some contexts).
Calibration, Bandits, and Regret There is a growing literature on designing fair selection
policies in the multi-arm bandits setting [JKMR16, JKM+ 17, LRD+ 17]. Recently [LRD+ 17] initiated the study of calibration in this context. Motivated by the aforementioned work of [DHP+ 12],
their notion of calibrated fairness aims to “treat similar individuals similarly” by designing sampling
policies that have low fairness regret.
Causality and discrimination in the data Discrimination can occur at any stage throughout
the algorithmic pipeline and in many forms. The most important aspect of developing a theory
of algorithmic fairness that multicalibration does not address is “unfair” data (see more below).
Kilbertus et al. [KRCP+ 17] voice an important criticism of any notion of fairness based solely on
observational criteria, as discrimination can occur through unresolved causal relationships. This
criticism applies to the notions of balance, calibration, and multicalibration, which all depend only
on the joint distribution of predictions, outcomes, and features.
Rather than attempting to provide a general-purpose definition of fairness, our work tackles a
particular concern about discrimination that can occur as part of the process of learning a predictor
from given data. In this context, we believe that multicalibration provides an important antidiscrimination guarantee.
9
Corrective discrimination Let us look deeper into the biases that may be present in the gathered data, by considering the mortgage example again: perhaps the number of members of S that
received loans in the past is small (and thus there are too few examples for fine-grained learning
within S); perhaps the attributes are too limited to identify the qualified members of S (taking
this point to the extreme, perhaps the only available attribute is membership in S). In these cases,
the data may be insufficient for multicalibration to provide meaningful guarantees. Further, even
if the algorithm was given access to unlimited rich data such that refined values of p∗ could be
recovered, there are situations where preferential treatment may be in order: after all, the salaries
of members of S may be lower due to historical discrimination.
For these reasons, our concern that balance is inconsistent with p∗ could be answered with: “yes,
and purposely so!” Indeed, [HPS16] promotes enforcing a balance-related property, called equalized odds, as a form of “corrective discrimination.” While this type of advocacy is important in
many settings, multicalibration represents a different addition to the quiver of anti-discrimination
measures, which we also believe is natural and desirable in many settings.
Consider a concrete example where multicalibration is appropriate, but equalizing error rates might
not be: Suppose a genomics company offers individuals a prediction of their likelihood of developing
certain genetic disorders. These disorders have different rates across different populations; for
instance, Tay-Sachs disease is rare in the general population, but occurs much more frequently in
the Ashkenazi Jewish population. We certainly do not want to enforce corrective discrimination
on the Ashkenazi population by down-weighting the prediction that individuals would have TaySachs (as they are endogenously more likely to have the disease). However, we also don’t want the
company to base its prediction solely on the Ashkenazi feature (either positively or negatively).
Instead, enforcing multicalibration would require that the learning algorithm investigate both the
Ashkenazi and non-Ashkenazi population to predict accurately in each group (even if this means
a higher false positive rate in the Ashkenazi population). In this case, relying on p∗ seems to be
well-aligned with promoting fairness.
Finally, we consider the interplay between multicalibration and “corrective discrimination” to be
an important direction for further research. For example, one can imagine applying corrective
measures, such as the transformation of [HPS16], to a multi-calibrated predictor.
1.5
Our Techniques
Here, we give a high-level technical overview of our results. Our techniques draw from the literature on computational learning theory, online optimization, differential privacy, and adaptive data
analysis.
Learning a multicalibrated predictor In Section 3.2, we describe an algorithm for learning α-multicalibrated predictors as stated in Theorem 2. Our algorithm is an iterative procedure. In particular, we will maintain a candidate predictor x, and at each iteration, the algorithm corrects the candidate values of some subset that violates calibration until the candidate
predictor is α-multicalibrated. Recall that calibration over a set S requires that on the subsets
Sv = {i ∈ S : xi = v} (which we will refer to throughout as categories), the expected value of the
true probabilities Ei∼Sv [p∗i ] on this set is close to v. As such, the algorithm is easiest to describe
10
in the statistical query model, where we query for noisy estimates of the true statistics on subsets
of the population and update the predictor based on these estimates. In particular, given a statistical query oracle that guarantees tolerance ω = O(αγ), the estimates will be accurate enough to
guarantee α-calibration on sets S with |S| ≥ γ |X|.
When we turn to adapting the algorithm to learn from random samples, the algorithm answers
these statistical queries using the empirical estimates on some random sample from the population.
Standard generalization arguments [KV94] show that if the set of queries we might ask is fixed in
advance, then we could bound the sample complexity needed to answer these non-adaptive queries
as Õ(log |C|/ω 2 ). Note, however, that the categories Sv whose expectations we query are selected
adaptively (i.e. with dependence on the results of prior queries). In particular, the definition of
the categories Sv depends on the current values of the predictor x; thus, when we update x based
on the result of a statistical query, the set of categories on which we might ask a statistical query
changes. In this case, we cannot simply apply concentration inequalities and take a union bound
to guarantee good generalization without resampling every time we update the predictor.
To avoid this blow-up in sample complexity, we appeal to recently-uncovered connections between differential privacy and adaptive data analysis developed in [DFH+ 15c, DFH+ 15c, BNS+ 16,
DFH+ 15b]. In Section 3.3, we show how to answer the statistical queries in a way that guarantees
the learning algorithm is differentially private. This, in turn, allows us to argue that the predictor
our algorithm learns from a small sample will be multicalibrated, not just for the observed individuals but also the unseen individuals. In particular, using a privacy-based approach with an analysis
tailored to the goal of calibration, we obtain sample complexity that depends on 1/α11/2 γ 3/2 as
opposed to the naive approach which results in 1/α6 γ 6 .
As stated earlier, Theorem 1 can be seen as a corollary of Theorem 2. Bounding the number of
iterations needed by the algorithm to converge not only upper-bounds the algorithm’s running time
and sample complexity, but also implies that the circuit complexity of the learned predictor is not
much larger than the complexity of evaluating membership for S ∈ C. We explain this implication
in Section 3.5.
The complexity of multicalibration In Section 3.4, we discuss the complexity of learning
a multicalibrated predictor and draw connections to agnostic learning [Hau92, KSS94]. We show
that the algorithm learns a multicalibrated predictor in a bounded number of iterations; however,
without additional assumptions about p∗ or C, each iteration could take Ω(|C|) time. In the cases
where |C| is large, we might hope to improve the dependence on |C| to polylogarithmic or perhaps
subpolynomial. If p∗ can be learned directly, then we can eliminate the dependence on |C| and
instead, only depend on the minimum γ such that for all S ∈ C, |S| ≤ γN .
In the case where p∗ is arbitrary, we show that improving the dependence on |C| is possible if C
is structured in a certain way, drawing a connection to the literature on agnostic learning [Hau92,
KSS94,KMV08,Fel10]. Recall, in our algorithm for learning multicalibrated predictors, we maintain
a candidate predictor x, and iteratively search for some set S ∈ C on which x is not calibrated.
To solve this search problem more quickly, we frame the search as weak agnostic learning over a
concept class derived from C and over the hypothesis class of H = {h : X → [−1, 1]}.
Specifically, consider the concept class defined by the collection of subsets C, where for each S ∈ C,
we include the concept cS : X → {−1, 1} where cS (i) = 1 if and only if i ∈ S. We show how to
11
design a “labeling” ℓ : X → [−1, 1] for individuals such that if x violates the calibration constraint
on any S ∈ C, then the concept cS correlates nontrivially with the labels over the distribution of
individuals, i.e. hcS , ℓi ≥ ρ for some ρ > 0.
Thus, if x is not yet multicalibrated on C, then we are promised that there is some concept cS with
nontrivial correlation with the labels; we observe that this promise is exactly the requirement for
a weak agnostic learner, as defined in [KMV08, Fel10]. In particular, given labeled samples (i, ℓ(i))
sampled according to D, if there is a concept cS with correlation at least ρ with ℓ, then the weak
agnostic learner returns a hypothesis h that is τ correlated with ℓ for some τ < ρ. The catch is
that this hypothesis may not be in our concept class C, so we cannot directly “correct” any S ∈ C.
Nevertheless, the labeling on individuals ℓ is designed such that given the hypothesis h, we can still
extract an update to x that will make global progress towards the goal of attaining calibration. As
long as τ is nontrvially lower bounded, we can upper bound the number of calls we need to make
to the weak learner. The details of our choice of labels and how we track progress are given in
Section 4.
Additionally, we show that our reduction to weak agnostic learning is unavoidable. In particular,
we show that if it is possible to learn a multicalibrated predictor with respect to C, then it is
possible to weak agnostic learn on C (if we view C as a concept class). Specifically, we will show
how to implement a weak agnostic learner for C, given an algorithm to learn an α-multicalibrated
predictor x with respect to C (in fact, we only need the predictor to be multicalibrated on C ′ =
{S ∈ C : |S| ≥ γ |X|}). The key lemma for this reduction says that if there is some c ∈ C that is
nontrivially correlated with the labels, then x is also nontrivially correlated with c. As there are
many natural classes C for which agnostic learning is conjectured to be hard, this gives a strong
negative result to the question of whether we can obtain speedups in the general case.
In combination, these results show that the complexity of learning a multicalibrated predictor with
respect to a class C is equivalent to the complexity of weak agnostic learning C. Section 4 contains
the formal statements and proofs that imply this equivalence.
“Best-in-class” prediction In Section 5, we turn to understanding how requiring a predictor
to be multicalibrated affects the accuracy of the predictor. We show – in contrast to many other
notions of fairness – multicalibration does not limit the utility of a predictor. In particular, given
any collection of predictors H, and any collection of subsets C, we design a procedure for obtaining
a predictor x that is α-multicalibrated on C and achieves expected squared prediction error less
than or equal to the the best predictor in H (plus a small additive error on α). Further, leveraging
Theorem 1 and Theorem 2, we show that the predictor x can be learned from samples and implemented by a circuit of comparable size to the predictors h ∈ H. To prove that multicalibration
does not negatively impact the utility, we in fact, show a much stronger statement: if applying
multicalibration to some h ∈ H changes the predictions of h significantly (i.e. if kx − hk is large),
then this change represents an improvement in accuracy (i.e. kx − p∗ k < kh − p∗ k). In this sense,
requiring multicalibration is aligned with the goals of learning a high-utility predictor.
Organization of the paper In Section 2, we provide a description of our model and the formal definitions related to multicalibration. In Section 3, we describe and analyze our algorithm
for learning multicalibrated predictors. In Section 4, we investigate the complexity of obtaining
12
multicalibration, showing a tight connection to the complexity of weak agnostic learning. Finally,
in Section 5, we demonstrate how multicalibration achieves “best-in-class” prediction.
2
Preliminaries
Predictors Let X denote the universe of N individuals over which we wish to make predictions
about some outcome o ∈ {0, 1}N . We assume that outcomes are the result of some underlying
random process, where for each i, oi is sampled independently4 as a Bernoulli random variable with
success probability p∗i . We aim to predict the underlying parameters of the process, rather than
the realized outcome of the process. A predictor x : X → [0, 1] of outcome o is some mapping from
individuals i ∈ X to [0, 1], where xi is the prediction of p∗i . We refer to p∗ as the baseline predictor.
When it is notationally convenient, we will sometimes treat predictors as vectors x ∈ [0, 1]N rather
than as functions x : X → [0, 1], where we assume a bijection from X to [N ].
Remark. Often in learning theory, we think of learning functions h : F → [0, 1] over the space of
possible features F. We find it preferable to reason about predictions about individuals; nevertheless,
in our model, we can think of xi as given by the composition of two separate functions, where
xi = x′ (φ(i)) for x′ : F → [0, 1] and φ : X → F. As φ will be fixed and assumed to be some
simple function (say, mapping individuals to their h age, height, ZIP code, . . .i) we drop the explicit
reference to F and φ.
Sampling from X Throughout, we will assume that our learning algorithms have the ability to
efficiently obtain randomly sampled individuals from X (and by rejection sampling, large subsets
of X ). Specifically, we will use i ∼ S to denote sampling i uniformly at random from S ⊆ X .
Remark. Our focus on the uniform distribution is mostly syntactic; as per our distinction between
the identities of individuals i ∈ X and their features φ(i), the uniform distribution over individuals
gives rise to a rich class of distributions over the features of individuals.
We will further assume that sampling i ∼ X is inexpensive compared to obtaining a corresponding
outcome oi ∈ {0, 1} for i ∼ X that is Bernoulli distributed with parameter p∗i . As such, when
measuring the sample complexity, we will only count the latter type of labeled samples. The
learner has a good sense of the distribution over features, but not of the outcomes that arise from
these features.
When measuring the accuracy of a predictor x, we will use the squared prediction error kp∗ − xk2 ,
as this measure of divergence penalizes large individual deviations. The ℓ22 distance will also be
useful for measuring the similarity of predictors. While we elect to use ℓ22 , much of our analysis
could be performed using any Bregman divergence; in particular, if we elected to work over arbitrary
multinomial distributions over individuals in X , we could measure accuracy in terms of the squared
Mahalanobis distance from the optimal predictor, i.e. (p∗ − x)T D(p∗ − x), where D is the diagonal
matrix in which Dii is the probability of sampling individual i ∈ X .
4
In many cases, complete independence may not hold; individuals’ outcomes may be correlated in nontrivial ways.
The only place we will use independence is to argue that the reliable statistics can be estimated from the observable
data; our arguments can be applied to any model for which one can prove the appropriate tail inequalities. For
further consideration, continue to our discussion on “observable” calibration after Claim 2.3.
13
2.1
Calibration
Next, we give formal definitions of the criteria we use to measure algorithmic discrimination. The
first notion captures the idea that, on average, we would like an algorithm’s predictions to be
unbiased.
Definition 2.1 (Accurate in expectation). For any α > 0 and S ⊆ X , a predictor x is α-accurate
in expectation (α-AE) with respect to S if
E [xi − p∗i ] ≤ α.
i∼S
That is, the predictions, averaged over the set S, are accurate up to some additive α slack. This basic
condition is necessary to ensure unbiased predictions, but it is too weak to guarantee discrimination
hasn’t occurred. In particular, suppose the underlying probabilities are such that p∗i = 1/2 for
all i ∈ S. A predictor that predicts 1 on half of the individuals in S and 0 on the other half
is accurate in expectation, but it is arguably discriminatory; there is no difference between the
individuals in S, but the predictor has artificially created two categories within this population. This
example motivates the definition of calibration. Calibration mitigates this form of discrimination
by considering the expected values over categories Sv = {i : xi = v} defined by the predictor x.
Specifically, α-calibration with respect to S requires that for all but an α-fraction of a set S, the
average of the true probabilities of the individuals receiving prediction v is α-close to v.
Definition 2.2 (Calibration). For any v ∈ [0, 1], S ⊆ X , and predictor x, let Sv = {i : xi = v}.
For α ∈ [0, 1], x is α-calibrated with respect to S if there exists some S ′ ⊆ S with |S ′ | ≥ (1 − α) |S|
such that for all v ∈ [0, 1],
E
i∼Sv ∩S ′
[xi − p∗i ] ≤ α.
Note that α-calibration with respect to S implies 2α-AE with respect to S. To see this, observe
that calibration implies that on a (1 − α)-fraction of S, the average of the values are α-close to the
expectation on this fraction; even if the other α-fraction is arbitrary, it can only introduce another
additive α error. This definition only requires the notion of calibration to hold on a (1 − α)-fraction
of each S; this is for technical reasons. See Definition 2.9 to understand better the motivations for
this choice.
Departures from prior definitions This notion of calibration departs from prior definitions in
a few ways. Earlier definitions required exact calibration with respect to X ; we find it meaningful
to consider approximate calibration. Introducing approximation into the definition of calibration
has practical motivation. In particular, we don’t expect to know the exact probabilities, nor can we
observe the entire population. Given a desired accuracy and access to samples from the population,
we can quantify how many samples are needed to guarantee calibration with good probability.
Note that in our definition of α-calibration, we require relative additive error that scales with
the size of Sv . That P
is, for some S and category Sv = {i : xi = v} ∩ S, the magnitude of the
∗
sum of errors on Sv ,
i∈Sv (v − pi ) , scales with |Sv |, not N . This relative approximation model
prevents certain “attacks” against approximate calibration. Specifically, one natural way a predictor
14
might attempt to sidestep the anti-discrimination properties of approximate calibration would be
to support many distinct values of v ∈ [0, 1], each with a very small number of individuals. If our
notion of approximation allowed for absolute errors (i.e. errors whose sum over Sv scale with N
instead of |Sv |), then this attack would be viable, leading to essentially no guarantees
Another subtle distinction is that we evaluate calibration with respect to the underlying probabilities p∗ ∈ [0, 1]N , as opposed to the realized outcomes o ∈ {0, 1}N . We refer to the earlier notion as
observable calibration. Formally, a predictor is observably α-calibrated with respect to S ⊆ X and
outcome o ∈ {0, 1}N if for all v ∈ [0, 1],
E
i∼Sv ∩S ′
[xi − oi ] ≤ α,
where Sv and S ′ are as in Definition 2.2. In our terminology, earlier references to calibration
require that a predictor be observably 0-calibrated with respect to each protected group S. Our
introduction of a non-zero α in these definitions allows us to relate our notion of calibration and
the previous notion of observable calibration.
q
log(1/ξ)
N
Claim 2.3. Let S ⊆ X , ξ > 0, and α >
is drawn
2|S| . Suppose the outcome o ∈ {0, 1}
∗
according to the true probabilities p . Then, with probability at least 1 − ξ over the draw of o, any
predictor that is α-calibrated with respect to S will be observably 2α-calibrated with respect to S.
Claim 2.3 follows directly from an application of Hoeffding’s inequality over the random outcome
o. The claim implies that for sufficiently large α, to guarantee observable 2α-calibration with high
probability, it suffices to consider our notion of α-calibration.
Independence and observability We note that to prove Claim 2.3, we use the independence
of oi ’s in our application of Hoeffding’s inequality. Throughout, the only times we invoke the
independence of the realization of the oi ’s are to prove that the empirical statistics over a sufficiently
large sample of observations will be concentrated around the corresponding statistics of the true
parameters. Thus, the independence assumption isn’t strictly necessary; our results should hold
for any model where the outcomes admit similar tail inequalities. We note, however, that if strong
dependencies exist in the realization of the outcomes, our techniques will still achieve observable
calibration from samples. Recall that in observable calibration, we compare the value v output
by
P
the predictor on a category Sv to the empirical average of outcomes over the category i∈Sv oi .
If we only need to guarantee closeness with respect to these observable outcomes, we do not need
to ensure that sums over outcomes will be concentrated around their expectation (i.e. sums of the
underlying true probabilities).
2.2
Learning model
In Section 3, we present algorithms to learn predictors satisfying accuracy-in-expectation and calibration. The algorithms can be viewed as statistical query algorithms [Kea98]. Specifically, the
algorithms only require access to approximate statistical queries of the following form.
15
Definition 2.4 (Statistical Query [Kea98]). For a subset of the universe S ⊆ X , let p∗S =
For τ ∈ [0, 1], a statistical query with tolerance τ returns some p̃(S) satisfying
P
i∈S
p∗i .
p∗S − τ N ≤ p̃(S) ≤ p∗S + τ N.
Note that this query model guarantees absolute additive error τ N . As discussed above, our notion
of calibration with respect to S asks for relative additive error; however, if we know a lower bound
on |S| ≥ γN , then, asking a statistical query with tolerance τ = αγ will guarantee relative error α
on S.
In addition to giving algorithms that work in this statistical learning framework, we also address
the question of learning a calibrated predictor from a set of samples of outcomes. Formally, we
define the access to sampled outcomes as follows.
Definition 2.5 (Random sample). For p∗ ∈ [0, 1]N , a random sample returns an individualoutcome pair (i, oi ) ∈ X × {0, 1}, where i ∼ X is drawn uniformly at random and oi is sampled
according to the Bernoulli distribution with parameter p∗i .
We say an algorithm learns a predictor from samples if its only access to the true parameters p∗ is
through random samples of this form. It’s easy to see that we can always implement a statistical
query algorithm with access to enough random samples – for every query, we could sample a fresh
set of outcomes to estimate the statistic accurately. Our goal will be to avoid resampling in order
to prevent a blow-up in the sample complexity.
2.3
Multicalibration
We introduce the notion of multicalibration, which requires calibration to hold simultaneously
on subsets of the population. We will show that multicalibration not only guarantees fairness
across protected populations, but also helps us uncover more accurate predictions. To motivate
multicalibration further, consider the following toy example: suppose p∗ is such that there is some
population S (possibly, a traditionally protected group) and a subpopulation S ′ ⊆ S with |S ′ | =
|S| /2, where for every i ∈ S ′ , p∗i = 1, and for every i ∈ S \ S ′ , p∗i = 0. The predictor x that
predicts xi = 1/2 for all i ∈ S is calibrated on the population S, but clearly is suboptimal. Further,
if S ′ was identifiable in advance, then this predictor is arguably discriminatory – there are two
clearly identifiable groups within S, but we are treating them the same way. If, however, we insist
on calibration with respect to S ′ in addition to S, then the predictor will be required to output
accurate predictions for each group. Earlier approaches to using calibration to achieve fairness, as
introduced in [KMR17], would prevent this form of discrimination for subsets S that are identified
as a protected group (defined, for example, by race), but not for subpopulations of these groups –
even if the subpopulations could be easily distinguished as outstanding.
For a collection of subsets C, we say that a predictor is α-multicalibrated on C if it is α-calibrated
simultaneously on all S ∈ C.
Definition 2.6 (α-multicalibration). Let C ⊆ 2X be a collection of subsets of X and α ∈ [0, 1]. A
predictor x is α-multicalibrated on C if for all S ∈ C, x is α-calibrated with respect to S.
We also define the corresponding definition for the weaker notion of α-AE.
16
Definition 2.7 (α-multi-AE). Let C ⊆ 2X be a collection of subsets of X and α ∈ [0, 1]. A predictor
x is α-multi-AE on C if for all S ∈ C, x is α-AE with respect to S.
Discretization Even though α-calibration is a meaningful definition if we allow for arbitrary
predictions xi ∈ [0, 1], when designing algorithms to learn calibrated predictors, it will be useful to
maintain some discretization on the values v ∈ [0, 1]. Formally, we will use the following definition.
Definition
2.8 (λ-discretization). Let λ > 0. The λ-discretization of [0, 1], denoted by Λ[0, 1] =
λ 3λ
,
,
.
.
.
,
1
− λ2 , is the set of 1/λ evenly spaced real values over [0, 1]. For v ∈ Λ[0, 1], let
2 2
λ(v) = [v − λ/2, v + λ/2)
be the λ-interval centered around v (except for the final interval, which will be [1 − λ, 1]).
At times, it will be convenient to work with a more technical variant of multicalibration, which
implies α-multicalibration. In particular, this definition will allow us to work with an explicit
discretization of the values v ∈ [0, 1]. Throughout, for a predictor x, we refer to the “categories”
Sv (x) = {i : xi ∈ λ(v)} ∩ S for all S ∈ C and v ∈ Λ[0, 1].
Definition 2.9 ((α, λ)-multicalibration). Let C ⊆ 2X be a collection of subsets of X . For any
α, λ > 0, a predictor x is (α, λ)-multicalibrated on C if for all S ∈ C, v ∈ Λ[0, 1], and all categories
Sv (x) such that |Sv (x)| ≥ αλ |S|, we have
X
xi − p∗i ≤ α |Sv (x)| .
i∈Sv (x)
We claim that if learn a predictor that satisfies (α, λ)-multicalibration, we can easily transform this
predictor into one that satisfies our earlier notion of α-multicalibration. In particular, let xλ be the
λ-discretization of a predictor x if for all i ∈ Sv (x), xλi = Ei∼Sv (x) [xi ].
Claim 2.10. For α, λ > 0, suppose C ⊆ 2X is a collection of subsets of X . If x is (α, λ)multicalibrated on C, then xλ is (α + λ)-multicalibrated on C.
Proof. Consider the categories Sv (x) where |Sv (x)| < αλ |S|. By the λ-discretization, there are at
most 1/λ such categories, so the cardinality of their union is at most (1/λ)αλ |S| = α |S|. Thus,
for each S ∈ C, there is a subset S ′ ⊆ S with |S ′ | ≥ (1 − α) |S| where for all v ∈ Λ[0, 1],
E
i∼Sv (x)∩S ′
[xi − p∗i ] ≤ α.
Further, λ-discretization will “move” the values of xi by at most λ, so overall, xλ will be (α + λ)calibrated.
Typically, we will imagine λ = Θ(α), but our results hold for any λ ∈ (0, 1]. Choosing a smaller λ
will allow the predictor to be more expressive, but will also increase the running time and sample
complexity. Choosing a larger λ leads to a decay in the calibration guarantees.
17
Representing subsets of individuals When representing collections of subsets, we will assume
that the subsets are represented implicitly. In particular, we will assume that S ∈ C is given as a
circuit cS : X → {0, 1}, where S = c−1
S (1); that is, for i ∈ X , cS (i) = 1 if and only if i ∈ S. Using
this implicit representation serves two purposes. First, in many cases, we may want to calibrate
on a collection of subsets over a large universe; in these cases, assuming an explicit representation
of each set is unreasonable. Second, associating a set S with a circuit that computes membership
in S allows us to quantify the complexity of the sets in C. In particular, it seems natural to
apply multicalibration to guarantee calibration with respect to a collection of efficiently-identifiable
subsets (say, subsets defined by conjunctions of four attributes, or any other simple circuit class).
It seems comparatively unreasonable to require calibration on, say, a random subsets, each of which
would require Ω(|X |) bits to describe.
3
Learning α-multicalibrated predictors
In this section, we prove Theorem 2; that is, we provide an algorithm for efficiently learning
α-multicalibrated predictors. The algorithms we describe are iterative and fit into the powerful
online optimization framework [SS12, Haz16] as well as the statistical query framework [Kea98].
For completeness, in Section 3.1, we describe an algorithm that solves the simpler task of learning
an α-multi-AE predictor, as a warm-up to introduce the main ideas; this algorithm was originally
discovered by [TTV09]. In Section 3.2, we describe the algorithm for learning α-multicalibrated
predictors in full. Then, in Section 3.3, we will give a nontrivial implementation of the statistical
query oracle that will imply nontrivial upper bounds on the running time and sample complexity
of the learning algorithm. This implementation borrows ideas from the literature on differentially
private query release and optimization [HR10, Ull15, HRRU14]. We conclude in Section 3.5 with
the observation that our algorithm also has implications for the circuit complexity of calibrated
predictors: for any collection of sets C, there is an α-calibrated predictor whose circuit complexity is
a small factor larger than the circuit complexity required to describe the sets in C. This establishes
Theorem 1.
3.1
α-multi-AE predictors
We begin our discussion with a simpler statistical query algorithm for learning an α-multi-AE
predictor. This algorithm serves as a warm-up for the subsequent algorithm for learning an αmulticalibrated predictor.
Algorithm 3.1 describes an iterative statistical query procedure for learning a α-multi-AE predictor
on C. Note that the problem of finding an α-multi-AE predictor for some collection of sets C can
be written as a linear program; the algorithm presented can be viewed as an instance of projected
subgradient descent (see [Haz16]). The algorithm iteratively updates a predictor x until it cannot
find a set S ∈ C where the current estimate deviates significantly from the value reported by the
statistical query. We claim that if no set violates this condition, then x is α-multi-AE on C.
Claim 3.1. If Algorithm 3.1 outputs a predictor x, then x is α-multi-AE on C.
Proof. Let pS =
P
i∈S
p∗i and xS =
P
i∈S
xi . By the assumed tolerance of the statistical queries
18
Algorithm 3.1 – Learning an α-multi-AE predictor on C
Let α, γ > 0 and let C ⊆ 2X be such that for all S ∈ C, |S| ≥ γN .
For S ⊆ X , let p̃(S) be the output of a statistical query with tolerance τ < αγ/4.
• Initialize:
◦ Let x = (1/2, . . . , 1/2) ∈ [0, 1]N
• Repeat:
◦ For each S ∈ C:
– Let ∆S = p̃(S) −
P
i∈S
xi
– If |∆S | > α |S| − τ N :
S
update xi ← xi + ∆
|S| for all i ∈ S (projecting xi onto [0, 1] if necessary)
◦ If no S ∈ C updated: exit and output x
p̃(S), we know that the queries are close to the pS . Specifically, we know |p̃(S) − p∗S | ≤ τ N for
some τ < αγ. By the termination condition and the triangle inequality, for all S ∈ C we get the
estimate |p∗S − xS | ≤ |p̃(S) − xS | + τ N ≤ α |S|; thus x is α-AE on C.
Thus, to show the correctness of the algorithm, it remains to show that the algorithm will, in fact,
terminate; we show the algorithm can make at most O(1/α2 γ) updates.
Lemma 3.2. Suppose α, γ > 0 and C ⊆ 2X such that for all S ∈ C, |S| ≥ γN . Let τ = αγ/4.
Then Algorithm 3.1 makes O(1/α2 γ) updates to x before terminating.
Proof. We use a potential argument, tracking the progress the algorithm makes on each update in
terms of the ℓ22 distance between our learned predictor x and the true predictions p∗ . Let x′ be
the predictor after updating x on set S and let π : R → [0, 1] denote projection onto [0, 1]. We
use the fact that the ℓ22 can only decrease under this projection. For notational convenience, let
P
1
∆S
=
(p̃(S) − i∈S xi ). We have
δS =
|S|
|S|
X
X
2
kp∗ − xk2 − p∗ − x′ =
(p∗i − xi )2 −
(p∗i − π(xi + δS ))2
i∈S
i∈S
X
≥
((p∗i − xi )2 − (p∗i − (xi + δS ))2 )
i∈S
X
=
(2(p∗i − xi )δS − δS2 )
i∈S
=
!
X
∗
(pi − xi ) − δS2 |S|
2δS
i∈S
≥ 2δS (δS |S| − sgn(δS )τ N ) − δS2 |S|
≥ δS2 |S| − 2 |δS | τ N.
19
By setting τ = αγ/4 and by the bound |∆S | ≥ α |S| − τ N ≥ 3α |S| /4, the final quantity is at least
Ω(α2 |S|). We also have
2
3α
αγ
3α
2
N
|S| − 2
δS |S| − 2 |δS | τ N ≥
4
4
4
3α2
=
|S| .
16
The ℓ22 distance between p∗ and any other predictor (in particular, our initial choice for x) is upperbounded by N . Thus, given that all S ∈ C have |S| ≥ γN , we make at least Ω(α2 γN ) progress in
potential at each update, so the lemma follows.
In combination, these statements show the correctness of Algorithm 3.1 and imply an upper bound
on the number of statistical queries necessary.
Theorem 3.3. For α, γ > 0 and for any C ⊆ 2X satisfying |S| ≥ γN for all S ∈ C, there is
a statistical query algorithm with tolerance τ = αγ/4 that learns a α-multi-AE predictor on C in
O(|C| /α2 γ) queries.
Recall that, trivially, we could implement this statistical query algorithm from random samples by
resampling for every query; however, in this case, we can easily improve the sample complexity
exponentially over the trivial solution. Specifically, the queries we make are non-adaptive because,
up front, we know a fixed collection of subsets whose expectation we might query. To guarantee
accurate expectations on this fixed collection, we only need enough samples to guarantee that the
sample is inaccurate on a fixed subset with very small probability, and then union bound over all
|C| subsets. Appealing to a standard generalization argument [KV94], we can show the following
theorem.
Corollary 3.4. Suppose α, γ, ξ > 0 and C ⊆ 2X is such that for all S ∈ C, |S| ≥ γN . Then
there isan algorithm
that learns an α-multi-AE predictor on C with probability at least 1 − ξ from
log(|C| /ξ)
samples.
n = Õ
α2 γ
Note that the γ dependence in the sample complexity is only 1/γ. Naively, applying the guarantees
of the statistical query oracle, we would obtain a 1/γ 2 dependence. To achieve this bound, we note
that because calibration requires relative error, we can be more judicious with our use of samples.
We will exploit this observation subsequently to obtain improvements to the sample complexity for
learning α-multicalibrated predictors.
Proof. To obtain the claimed sample complexity bound, we observe that in Algorithm 3.1, we
only use the statistical query oracle to guarantee bounds on the relative
error of each query – not
1 P
absolute error. In particular, for S ⊆ X with |S| ≥ γN , let p̄S = |S| i∈S p∗i . To run Algorithm 3.1,
we need only implement an oracle p̂(S) satisfying
p̄S − τ ≤ p̂(S) ≤ p̄S + τ.
By Chernoff bounds, for a fixed set S of cardinality at least γN , if we take Õ(t/γα2 ) independent
samples, the probability that the estimate of p̄S differs by more than α |S| is at most e−Ω(t) log(1/γα) .
20
Taking t = c(log |C| + log(1/ξ)) for an appropriate constant c, a union bound implies that for each
of poly(1/γ, 1/α) iterations, the probability that the estimate of every set S ∈ C is α-accurate is at
least 1 − ξ.
3.2
α-Multicalibrated predictors
Next, we present the full algorithm for learning α-multicalibrated predictors on C. The algorithm
is based on Algorithm 3.1 but differs in a few key ways. First, instead of updating the predictions
on entire sets S ∈ C whose overall expectation is wrong, we update the predictions on uncalibrated
categories Sv = S ∩ {i : xi = v}. This is a simple change to the algorithm in the statistical query
model; however, when we wish to implement this statistical query oracle from a finite sample, we
need to be more careful.
In particular, the categories of the predictor we learn are not fixed a priori, so our queries will be
selected adaptively based on the results of earlier statistical queries. Stated another way, we cannot
simply union bound against the collection of sets on which we wish to be calibrated. The most
naive approach to bounding the sampling complexity would be as follows: at each iteration take
a fresh sample large enough to guarantee ωN absolute error, for ω = αγ. Following our analysis
below for this naive strategy, it’s easy to see that this approach would result in Ω(1/α4 γ 4 ) iterations
and sample complexity n ≥ Ω(1/α6 γ 6 ). We will improve on this approach by polynomial factors
achieving a bound of at most O(1/α4 γ) iterations with O(1/α5/2 γ 3/2 ) samples.
To achieve these improvements, we combine two ideas. As before, we will leverage the observation that calibration only requires relative error (as in Corollary 3.4), and thus, in principle should
require fewer samples. Additionally, to avoid naively resampling but still guarantee good generalization from a small sample, we interact with the sample through a mechanism which we call a guessand-check statistical query (similar in spirit to mechanisms proposed in [HR10, BH15, GRU12]).
We show how to implement this mechanism in a manner that guarantees generalization on the
unseen data even after asking many adaptively chosen statistical queries. We defer our discussion
of privacy to Section 3.3.
Details of the algorithm Next, we give an iterative procedure to learn a (α, λ)-multicalibrated
predictor on C as described in Algorithm 3.2. The procedure is similar to Algorithm 3.1, but
deliberately interacts with its statistical queries through a so-called guess-and-check oracle. In
particular, each time the algorithm needs to know the value of a statistical query on a set S,
rather than
asking the query directly, we require that the algorithm submit its current guess
1 P
xS = |S|
i∈S xi to the oracle, as well as an acceptable “window” ω ∈ [0, 1]. Intuitively, if the
algorithm’s guess is far from the window centered around the true expectation, then the oracle will
respond with the answer to a statistical query with tolerance ω. If, however, the guess is sufficiently
close to the true value, then the oracle responds with X to indicate that the current guess is close
to the expectation, without revealing another answer.
Definition 3.5 (Guess-and-check oracle). Let q̃ : 2X × [0,P
1] × [0, 1] → [0, 1] ∪ {X}. q̃ is a guessand-check oracle with window ω0 if for S ⊆ X with pS = i∈S p∗i , v ∈ [0, 1], and any ω ≥ ω0 , the
response to q̃(S, v, ω) satisfies the following conditions:
21
• if pS − |S| v < 2ωN , then q̃(S, v, ω) = X
• if pS − |S| v > 4ωN , then q̃(S, v, ω) ∈ [0, 1]
• if q̃(S, v) 6= X, then
pS − ωN ≤ q̃(S, v, ω) |S| ≤ pS + ωN.
Note that if the guess is such that pS − |S| v ∈ [2ωN, 4ωN ], the the oracle may respond with
some ω-accurate r ∈ [0, 1] or with X. Of course, if we have a lower bound ω0 on the window over
a sequence of guess-and-check queries, we can implement the queries given access to a statistical
query oracle with tolerance τ ≤ ω0 ; it is also clear that a statistical oracle with tolerance τ can be
implemented with access to a guess-and-check oracle with window τ /4. The advantage of using this
guess-and-check framework is that it can be implemented in a differentially private manner. This
will in turn allow us to give an algorithm for learning α-multicalibrated predictors from a small
number of samples that generalizes well.
Algorithm 3.2 – Learning a (α, λ)-calibrated predictor on C
Let α, λ > 0 and let C ⊆ 2X be such that for all S ∈ C, |S| ≥ γN .
For S ⊆ X and v ∈ [0, 1], let q̃(·, ·, ·) be a guess-and-check oracle.
• Initialize:
◦ Let x = (1/2, . . . , 1/2) ∈ [0, 1]N
• Repeat:
◦ For each S ∈ C, v ∈ Λ[0, 1], for each Sv = S ∩ {i : xi = λ(v)} such that |Sv | = βN ≥ αλ |S|
P
– Let v̄ = |S1v | i∈Sv xi
– Let r = q̃(Sv , v̄, αβ/4)
– If r 6= X:
update xi ← xi + (r − v̄) for all i ∈ Sv (projecting xi onto [0, 1] if necessary)
◦ If no Sv updated, exit
• For v ∈ Λ[0,
P1]:
◦ Let v̄ = i∈λ(v) xi
◦ For i ∈ λ(v): xi ← v̄
• Output x
Algorithm 3.2 runs through each possible category Sv and if Sv is large enough, queries the oracle.
The algorithm continues searching for uncalibrated categoires until x’s guesses on all sufficiently
large categories receive X. By the definition of the guess-and-check oracle, if for some category Sv
where |SvP
| = βN the query returns X, then v̄ is at most 4 · (αβN/4) = α |Sv | far from the true
1
value Sv i∈Sv p∗i . Thus, by the stopping condition of the loop, the predictor where all i ∈ λ(v)
receive xi = v̄ will be α-calibrated on every large category. Finally, the algorithm updates x to
be λ-discretized, so by Claim 2.10, x will be (α + λ)-calibrated. Further, the number of updates
necessary to terminate is bounded.
22
Lemma 3.6. Suppose α, λ > 0 and C ⊆ 2X where for all S ∈ C, |S| ≥ γN . Algorithm 3.2
returns x after receiving at most O(1/α3 λγ) guess-and-check responses where r ∈ [0, 1] and at most
O(|C| /α4 λγ) responses r = X.
Proof. For some non-X response on Sv = {i : xi ∈ λ(v)} ∩ S, by the properties of the guess-andcheck oracle, we can lower bound the update step size. Recall, we only query on sets wherer
|Sv | = βN ≥ α |S| with a window of ω = αβ/4.
X
X
p∗i − xi =
p∗i − v̄
i∈Sv
i∈Sv
≥ 2 (αβ/4) N
= α |Sv | /2.
Letting δv = r − v̄. We can measure progress in the same way as in Lemma 3.2.
X
X
2
(p∗i − π(xi + δv ))2
(p∗i − xi )2 −
kp∗ − xk2 − p∗ − x′ =
i∈Sv
i∈Sv
≥
X
((p∗i
2
− xi ) − (p∗i − (xi + δv ))2 )
i∈Sv
=
X
(2(p∗i − xi )δv − δv2 )
i∈Sv
=
2δv
X
i∈Sv
(p∗i
!
− xi )
− δv2 |Sv |
P
Let ν = |S1v | i∈Sv (p∗i − xi ). By the properties of the guess-and-check oracle, we can rewrite δv as
ν − η for some η ∈ [−ω/β, ω/β]. This gives us a lower bound on the progress as follows.
2(ν − η)ν − (ν − η)2 |Sv | = ν 2 + νη − (η)2 |Sv |
This concave function in η is minimized at an extreme value for η (depending on the sign of ν).
Noting that |ν| ≥ α/2 and |η| ≤ ω/β = α/4, we can lower bound our progress by (α/4)2 |Sv | =
α2 βN/16 = α3 λγN/16. As kp∗ k2 ≤ N , we make at most O(1/α3 λγ) updates upper bounding the
number of non-X responses. By working with a λ-discretization, there are at most |C| /λ categories
to consider in every phase, so we receive at most O(|C| /α3 λ2 γ) X responses.
Thus, we conclude the following theorem.
Theorem 3.7. For α, λ > 0 and C ⊆ 2X where for all S ∈ C, |S| ≥ γN , there is a statistical query
algorithm that learns a (α, λ)-multicalibrated predictor with respect to C in O(|C| /α3 λ2 γ) queries.
Again, note that our output is, in fact, (α + λ)-multicalibrated, so taking λ = α, we obtain a
(2α)-multicalibrated predictor in O(|C| /α5 γ) queries.
23
3.3
Answering guess-and-check queries from a random sample
Next, we argue that we can implement a guess-and-check oracle from a set of random samples in
a manner that guarantees good generalization. This, in turn, allows us to translate our statistical
query algorithm for learning an (α, λ)-multicalibrated predictor with respect to C into an algorithm
that learns from samples. As mentioned in the beginning of Section 3, naively, we could resample
for every update the algorithm makes to the predictor. Suppose that C is such that for all S ∈
C, |S| ≥ γN ; let β = αλγ. Using our tighter analysis of Algorithm 3.2, we could take n =
Õ(log(|C|)/α2 β) samples per update to guarantee generalization, resulting in an overall sample
complexity of Õ(log(|C|)/α4 β 2 ). We show how to improve upon this approach further. In particular,
we argue that there is a differentially private algorithm that can answer the entire sequence of
guess-and-check queries accurately. Appealing to known connections between differential privacy
and adaptive data analysis [DFH+ 15c, DFH+ 15a, BNS+ 16, DFH+ 15b], this will guarantee that our
calibration algorithm generalizes given a set of Õ(log(|C|)/α5/2 β 3/2 ) random samples.
Algorithm 3.2 only interacts with the sample through the guess-and-check oracle. Thus, to give
a differentially private implementation of the algorithm, it suffices to give a differentially private
implementation of the guess-and-check oracle [DR14].
Consider the sequence of queries that Algorithm 3.2 makes to the guess-and-check oracle. We say
the sequence h(S1 , v1 , ω1 ), . . . , (Sk , vk , ωk )i is a (k, m)-sequence of guess-and-check queries if, over
the course of the k queries, the response to at most m of the queries is some r ∈ [0, 1], and the
responses to the remaining queries are all X. We will assume that we know a lower bound on the
minimum window ω = minj∈[k] ωj over all of the queries. We say that some algorithm A responds
to a guess-and-check query (S, v,P
ω) according to a random sample X if its response satisfies the
guess-and-check properties with i∈S p∗i replaced by its empirical estimate on X,
p̂S (X) =
X
|S|
oi .
|S ∩ X|
i∈S∩X
Responding to such a sequence in a differentially private manner can be achieved using techniques
from the private multiplicative weights mechanism.
Theorem 3.8 ( [HR10] ). Suppose ε, δ, ω, ξ > 0 and suppose X ∼ (X × {0, 1})n is a set of n
random samples. Then there exists an (ε, δ)-differentially private algorithm A that responds to any
(k, m)-sequence of guess-and-check queries with minimum window ω according to X provided
!
r
log(k/ξ) · m · log(1/δ)
n=Ω
ε · ω2
with probability at least 1 − ξ over the randomness of A.
Using this differentially private algorithm, we can apply generalization bounds based on privacy
developed in [DFH+ 15c, BNS+ 16, DFH+ 15a, DFH+ 15b] to show that, with a modest increase in
sample complexity, we can respond to all k guess-and-check queries.
Theorem 3.9. Let sk = h(S1 , v1 , ω̂1 ), . . . , (Sk , vk , ω̂k )i be a (k, m)-sequence of guess-and-check
queries such that for all j ∈ [k], |Sj | = βj N ≥ βN and ω̂j = Ω(αβj ). Then there is an algorithm
24
A that, given n random samples X ∼ (X × {0, 1})n , responds to sk such that for all j ∈ [k], the
response A(Sj , vj , ω̂j ; X) satisfies the guess-and-check properties with window ωj = αβj provided
log(|C| /αβξ)
n=Ω
α5/2 · β 3/2
with probability at least 1 − ξ over the randomness of A and the draw of X.
This theorem impliespthat, asymptotically, we can answer the k adaptively chosen guess-and-check
queries with only a 1/αβ factor increase in the sample complexity compared to if we knew the
queries in advance. Theorem 3.9 follows from tailoring the proof of the main “transfer” theorem
of [BNS+ 16] (Theorem 3.4) specifically to the requirements of our guess-and-check oracle and
applying the differentially private mechanism described in Theorem 3.8. Combining these theorems
and Algorithm 3.2 and the fact that β = αλγ, we obtain an algorithm for learning α-multicalibrated
predictors from random samples.
Theorem 3.10. Suppose α, λ, γ, ξ > 0, and C ⊆ 2X where for all S ∈ C, |S| ≥ γN . Then there
is an algorithm that learns
predictor with respect to C with probability at
an (α, λ)-multicalibrated
log(|C| /αλγξ)
samples.
least 1 − ξ from n = O
α4 · λ3/2 · γ 3/2
3.4
Runtime analysis of Algorithm 3.2
Here, we present a high-level runtime analysis of Algorithm 3.2 for learning an (α, λ)-calibrated
predictor on C. In Lemma 3.6, we claim an upper bound of O(|C| /α3 λ2 γ) on the number of guessand-check queries needed before Algorithm 3.2 converges. Here, we formally argue that each of
these queries can be implemented in the random sample model without much overhead, which
upper-bounds the running time of the algorithm overall. This upper bound is not immediate from
our earlier analysis, as the sets and our predictor are represented implicitly as circuits.
Claim 3.11. Algorithm 3.2 runs in time O(|C| · t · poly(1/α, 1/λ, 1/γ)), where t is an upper bound
on the time it takes to evaluate set membership for S ∈ C.
Proof. As before, let β = αλγ. First, for each S ∈ C, we need to evaluate |Sv | for Sv =
{i : xi ∈ λ(v)} ∩ S for each of the O(1/λ) values v ∈ Λ[0, 1]. We do this by sampling i ∼ X
and evaluating whether i ∈ S, and if so, checking the current value of xi . Each of the membership
queries takes at most t time and each evaluation of xi takes at most O(t/α2 β) time by the same
argument as our upper bound on the circuit size from Theorem 3.12. After Õ(1/λβ 2 ) samples,
we will be able to detect with constant probability which of the Sv have cardinality |Sv | ≥ βN .
Further, if |Sv | is large, we can estimate v̄ by evaluating the current predictor on samples from Sv ,
by rejection sampling. Similarly, to answer the guess-and-check queries, we will estimate the true
empirical estimate of the query based
P on samples from Sv and respond based on a noisy comparison
between the v̄ and the estimate of i∈Sv oi . These estimates can all be computed in poly(1/α, 1/β).
Then, as discussed in the proof of Theorem 3.12, each update to the predictor can be implemented
in time proportional to the bit complexity of the arithmetic computations, which is upper bounded
by t. Repeating this process for each S ∈ C gives the upper bound of O(|C| · t · poly(1/α, 1/λ, 1/γ)).
Finally, applying the upper bound on the number of guess-and-check queries from Lemma 3.6, the
claim follows.
25
3.5
The circuit complexity of multicalibrated predictors
As discussed in Section 1.3, an interesting corollary of our algorithm is a theorem about the complexity of representing a multicalibrated predictor. Indeed, from the definition of multicalibration
alone, it is not immediately clear that there should be succinct descriptions of multicalibrated
predictors; after all, C could contain many sets. We argue that the cardinality of C is not the
operative parameter in determining the circuit complexity of a predictor x that is multicalibrated
on C; instead it is the circuit complexity necessary to describe sets S ∈ C, as well as the cardinality
of the subsets in C, and the degree of approximation.
Leveraging Lemma 3.6, we can see that Algorithm 3.2 actually gives us a way to build up a circuit
that computes the mapping from individuals to the probabilities of our learned multicalibrated
predictor x. Suppose that for all sets S ∈ C, set membership can be determined by a circuit
family of bounded complexity; that is, for all S ∈ C, there is some cS with size at most s, such
that cS (i) = 1 if and only if i ∈ S. Then we can use this family of circuits to build a circuit
that implements x. We assume that we maintain real-valued numbers up to b ≥ log(1/α) bits of
precision.
Theorem 3.12. Suppose C ⊆ 2X is a collection of sets where each S ∈ C can be implemented by
a boolean circuit cS and for all S ∈ C, the size of cS is O(s). Then there is a predictor that is
α-multicalibrated on C implemented by a circuit of size O((s + b)/α4 γ). Further, Algorithm 3.2 can
be used to learn such a circuit.
Proof. We describe how to construct a circuit fx that, on input i, will output the prediction xi
according to the predictor learned by our algorithm. We initialize fx to be the constant function
fx (i) = 1/2 for all i ∈ X . Throughout, we will update fx based on the current outputs of fx .
Consider an iteration of Algorithm 3.2 where for some S described by cS ∈ C, we update x based
on a category Sv = S ∩ {i : xi ∈ λ(v)}. This occurs when the guess-and-check query returns some
r = q̃(Sv , v̄, ω) ∈ [0, 1]. Our goal is to implement the update to x (i.e. update fx ), such that for all
i ∈ Sv , the new value xi = r and all other values are unchanged.
We achieve this update by testing membership i ∈ S and separately testing if the current value
fx (i) = v; if both tests pass, then we update the value output by fx (i) to be r. Specifically, we
include a copy of cS and hard-code v and δv = r − v̄ into the circuit; if cS (i) = 1 and the current
value of fx (i) is in λ(v), then we update fx (i) to add the hardcoded δv to its current estimate of
xi ; if either test fails, then fx (i) remains unchanged. This logic can be implemented with addition
and subtraction circuits to a precision of λ with boolean circuits of size O(b). We string these
update circuits together, one for each iteration. Learning an (α/2, α/2)-multicalibrated predictor
with Algoirthm 3.2 only requires O(α4 γ) updates. By this upper bound, we obtain an O(α4 γ)
upper bound on the resulting circuit size.
4
Multicalibration and weak agnostic learning
Note that in the algorithm and analysis in Section 3, we’ve assumed nothing about the structure of
the underlying p∗ or C; the true probabilities could be adversarially chosen and yet, our algorithm
26
guarantees α-multicalibration on C. That said, the running time of the algorithm depends linearly
on |C|. As we imagine C to be a large, rich class of subsets of X , in many cases linear depedence
on |C| will be expensive. Thus, we turn our attention to when we can exploit structure within the
collection of subsets C to speed up the learning process.
The main running time bottleneck in the algorithms arises from searching for some S ∈ C where
calibration is violated. Without any assumptions about C, we need to loop over the collection;
however, if we can find such a set without looping over the entire collection of sets, then we would
improve the running time of the algorithm. At a high level, we will show a connection between
the agnostic learnability of C and the ability to speed up learning a multicalibrated predictor
on C. Imagining the sets S ∈ C as boolean concepts, we show that if it is possible to perform
weak agnostic learning over a class C efficiently (in the sense of [KMV08, Fel10]), then there is an
efficient search algorithm to find an update to the current predictor that will make progress towards
multicalibration.
While there are some classes for which we have weak agnostic learners, in general, agnostic learning
is considered a notoriously challenging problem. A natural question to ask is whether there is a
way to speed up learning a multicalibrated predictor that does not involve agnostic learning. We
answer this question in the negative. Roughly, we show that for a concept class C, we can use any
predictor that is multicalibrated on the large sets of C to a query for distribution-specific weak
agnostic learning on C. In this sense, the reduction to weak agnostic learning is inherent; any
efficient algorithm for multicalibration gives rise to an algorithm for weak agnostic learning.
In all, these results show that weak agnostic learning on a class C is equivalent to learning an
α-multicalibrated predictor with respect to C = {S ∈ C : |S| ≥ γN }, the large sets defined by C, up
to polynomial factors in 1/α, 1/γ where ρ and τ will be a function of α and γ.
4.1
Weak agnostic learning
For this discussion, we think of boolean concepts c ∈ C as c : X → {−1, 1}. We will overload
the notions of a concept class C of boolean functions c : X → {−1, 1} and our collection of
subsets C ⊆ 2X ; in particular, there is a natural bijection between concepts and sets: a concept
c : X → {−1, 1} defines a set S ⊆ 2X where i ∈ S if c(i) = 1 and i 6∈ S if c(i) = −1. We will connect
the problem of finding a set S ∈ C on which a predictor x violates calibration to the problem of
learning over the concept class C on a distribution D.
P
For some distribution D supported on X and x, y ∈ [−1, 1]N , let hx, yiD =
i∈X Di xi yi . This
inner product measures the correlation between x and y in [−1, 1]N over the discrete distribution
D. Throughout our discussion, we will focus on learning over the uniform distribution on X and
drop explicit reference to D. As per Remark 2, this may be a rich distribution over the features of
individuals.
In our results, we will work with the distribution-specific weak agnostic learners of [Fel10]5 .
Definition 4.1 (Weak agnostic learner). Let ρ ≥ τ > 0. Let D be a distribution supported on X .
5
Often, such learners are defined in terms of their error rates rather than correlations; the definitions are equivalent
up to factors of 2 in ρ and τ . Also, we will always work with a hypothesis class H = [−1, 1]X the set of functions
from X to [−1, 1], so we fix this class in the definition.
27
A (ρ, τ )-weak agnostic learner L for D solves the following promise problem: given a collection of
labeled samples {(i, yi )} where i ∼ D and yi ∈ [−1, 1], if there is some c ∈ C such that hc, yiD > ρ,
then L returns some h : X → [−1, 1] such that hh, yiD > τ .
Intuitively, if there is a concept c ∈ C that correlates nontrivially with the observed labels, then
the weak agnostic learner returns a hypothesis h (not necessarily from C), that is also nontrivially
correlated with the observed labels. In particular, ρ and τ are typically taken to be ρ = 1/p(d) and
τ = 1/q(d) for polynomials p(d) ≤ q(d), where d = log(|X |).
4.2
Multicalibration from weak agnostic learning
In this section, we show how we can use a weak agnostic learner to solve the search problem that
arises at each iteration of Algorithm 3.2: namely, to find an update that will make progress towards
mulitcalibration. Formally, we show the following theorem.
Theorem 4.2 (Formal statement of Theorem 3). Let ρ, τ > 0 and C ⊆ 2X be some concept
class. If C admits a (ρ, τ )-weak agnostic learner that runs in time T (|C| , ρ, τ ), then there is an
algorithm that learns a predictor that is (α, λ)-multicalibrated on C ′ = {S ∈ C : |S| ≥ γN } in time
O(T (|C| , ρ, τ ) · poly(1/α, 1/λ, 1/γ)) as long as ρ ≤ α2 λγ/2 and τ = poly(α, λ, γ).
That is, if there is an algorithm for learning the concept class C over the hypothesis class of realvalued functions H = {h : X → [−1, 1]} on the distribution of individuals in polynomial time in
log(|C|), 1/ρ, and 1/τ , then there is an algorithm for learning an α-multicalibrated predictor on the
large sets in C that runs in time polynomial in log(|C|), 1/α, 1/λ, 1/γ. For clarity of presentation in
the reduction, we make no attempts to optimize the sample complexity or running time. Indeed,
the exact sample complexity and running time will largely depend on how strong the weak learning
guarantee is for the specific class C.
We prove Theorem 4.2 by using the weak learner for C to learn a (α, λ)-multicalibrated predictor.
Recall Algorithm 3.2: we maintain a predictor x and iteratively look for a set S ∈ C where x
violates the calibration constraint on Sv = {i : xi ∈ λ(v)} ∩ S for some value v. In fact, the proof
of Lemma 3.6 reveals that we are not restricted to updates on Sv for S ∈ C. As long as there is
some uncalibrated category Sv , we can find an update that makes nontrivial progress in ℓ22 distance
from p∗ – even if this update is not on any S ∈ C – then we can bound the number of iterations it
will take before there are no more uncalibrated categories. We show that a weak agnostic learner
allows us to find such an update.
Proof. Throughout the proof, let β = αλγ, ρ = αβ/2, and τ = ρd for some constant d ≥ 1. Let x
be a predictor initialized to be the constant function xi = 1/2 for all i ∈ X .
Consider the search problem that arises during Algorithm 3.2 immediately after updating the
predictor x. Let Xv = {i : xi ∈ λ(v)} be the set of individuals in the λ-interval surrounding v. Our
goal is to determine if there is some v ∈ Λ[0, 1] and S ∈ C such that |Sv | ≥ βN , where
X
xi − p∗i ≥ α |Sv | .
i∈Sv
28
(1)
We reduce this search problem to the problem of weak agnostic learning over C on the distribution
DX . For any v ∈ Λ[0, 1], if |Xv | < βN , then clearly there is no uncalibrated category Sv with
|Sv | ≥ βN ; for each v ∈ Λ[0, 1], we will test if Xv is large enough by taking O(log(1/βξ)/β) random
draws from X .
Supposing Xv is large enough, we take a fresh sample of size n ≥ Õ(log(|C/ξ|)/β 2 τ 2 ). We take n
large enough that over all categories Sv of |Sv | ≥ βN , the observable statistics deviate from their
expectation by at most τ /4:
1 X ∗
1 X
(2)
pi ≤ τ /4
oj −
n
|Sv |
i∈Sv
j∈[n]
Additionally, assume that x is overall observably τ /4-calibrated with respect to X (recall, this
means calibrated on the set of observations). Note that observable calibration on X implies that
for each v ∈ Λ[0, 1],
1 X
(oj − xj ) ≤ τ /4.
(3)
n
j∈[n]
(If x is not τ /4-calibrated then P
for the Xv that violates calibration, offset all the values of xi from
their current values such that
i∈X xi − oi ≤ τ N/4 and resample; as in Algorithm 3.1, this will
make at least Ω(τ 2 ) progress towards p∗ ).
i
For each v ∈ Λ[0, 1], we consider the following learning problem. For i ∈ Xv , let ∆i = xi −o
2 . For
i ∈ X \ Xv , let ∆i = 0. We claim that if there is some Sv satisfying (1), then for i ∼ DX , the
labeled samples of either (i, ∆i ) or (i, −∆i ) satisfy the weak learning promise for ρ = αβ/2.
Claim 4.3. Let cS : X → {−1, 1} be the boolean
function associated with some S ∈ C. For
P
v ∈ Λ[0, 1], if Sv = {i : xi ∈ λ(v)} ∩ S satisfies i∈Sv (xi − p∗i ) ≥ α |Sv |, then
hcS , ∆iDX ≥ ρ.
Note that the supposition of the claim is satisfied when (1) holds without the absolute value. In
the case where (1) holds in the other direction, the claim will hold for −∆. The argument will be
29
identical.
hcS , ∆iDX
1 X x i − oi
=
· cS (i)
N
2
i∈X
X
1 X x i − oi
0 · cS (i)
· cS (i) +
=
N
2
i∈Xv
i∈X \Xv
X
X
1
x i − oi
x i − oi
=
−
N
2
2
i∈Sv
i∈Xv \Sv
!!
X x i − oi X x i − oi
−
−
2
2
i∈Sv
i∈Xv
i∈Sv
2 X x i − oi
− τ |Xv | /4
≥
N
2
1
=
N
X
x i − oi
2
(4)
i∈Sv
2
≥ (αβN − τ N/4)
N
≥ 2ρ − τ /2
(5)
(6)
where the inequality
from (3), (5) follows from the assumption that |Sv | ≥ βN and our
P (4) follows
∗
assumption on i∈Sv (xi − pi ), and (6) follows from our assumption that the sample size is large
enough to guarantee at most τ /8 error. Noting that τ /2 ≤ ρ gives the claim.
Thus, because the (ρ, τ )-weak agnostic learning promise is satisfied, the learner will return to us
some h : X → [−1, 1] satisfying the following inequality.
τ ≤ h∆i , hi iDv
1 X
(xi − oi ) · hi
=
2 |X |
i∈Xv
1 X
≤
(v − p∗i ) · hi + τ /8.
2 |X |
i∈Xv
where the final inequality follows by the assumed statistical accuracy. This shows that the h
returned to us by the weak agnostic learner is nontrivally correlated with x−p∗ on Xv . In particular,
if we use this h as a gradient step, updating xi → v − ηhi (projecting onto [0, 1] if necessary) for
η = Ω(τ /β), then we can guarantee that each such update will achieve τ 2 N progress in kx − p∗ k2 .
The analysis follows in the same way as the analysis of Algorithm 3.2.
4.3
Weak agnostic learning from multicalibration
In this section, we show the converse reduction. In particular, we will show that for a concept
class C, an efficient algorithm for obtaining an α-multicalibrated predictor with respect to C ′ =
{S ∈ C : |S| ≥ γN }, gives an efficient algorithm for responding to weak agnostic learning queries
on C.
30
Theorem 4.4 (Formal statement of Theorem 4). Let α, γ > 0 and suppose C ⊆ 2X is a concept
class. If there is an algorithm for learning an α-multicalibrated predictor on C ′ = {S ∈ C : |S| ≥ γN }
in time T (|C| , α, γ) then we can implement a (ρ, τ )-weak agnostic learner for C in time O(T (|C| , α, γ)·
poly(1/τ )) for any ρ, τ > 0 such that τ ≤ min {ρ − 2γ, ρ/4 − 4α}.
Proof. There are two cases to handle. First, suppose the support of cS is small; that is, for the
corresponding S ∈ C, |S| < γ. Then, consider the correlation between y and the the constant
hypothesis h(i) = −1 for all i ∈ X .
X
hy, −1i = −
yi
(7)
i∈X
= −
X
yi +
i∈S
i∈X \S
= hcS , yi − 2
X
X
yi
yi
(8)
(9)
i∈S
≥ ρ − 2γ
(10)
P
P
where (9) follows by writing hcS , yi as i∈S yi + i∈X \S yi and rearranging. Thus, for τ < ρ − 2γ,
in the case when the support of cS is small, then we can return the hypothesis −1. We can test
if the constant hypothesis is sufficiently correlated with y in poly(1/τ ) log(1/ξ) time by random
sampling to succeed with probability at least 1 − ξ.
Suppose we want to weak agnostically learn over C on sampled observations from y ∈ [−1, 1]N .
We assume P
there is some cS ∈ C such that hcS , yi > ρ. Consider some ω = ρ/4. First, we will
check if N1
i∈X yi > ω. Again, this does not require multicalibration, just sampling from y and
averaging. In this case, a constant function h(i) = 1 or h(i) = −1 is sufficiently correlated with y
to satisfy the weak agnostic learning guarantee.
P
Next, we will proceed assuming N1
i∈X yi < 2ω. We can expand the inner product between cS
and y as follows.
X
1 X
hcS , yi =
yi −
yi
(11)
N
i∈S
i∈X \S
!
X
X
1
(12)
2
yi −
yi
=
N
i∈S
i∈X
2 X
≥
yi − 2ω
(13)
N
i∈S
This means
1
N
P
i∈S
yi >
ρ
2
− ω.
Suppose we learn an x that is α-multicalibrated with respect to C ′ = C ∪ {X } on the labels y. This
implies that P
there is some X ′ ⊆ X such that |X ′ | ≥ (1 − α) |X | and for all v ∈ [−1, 1], we have
1
v − α ≤ |X ′ | i∈Xv′ yi ≤ v + α. In turn, this implies the following inequality.
v
X
sgn(v) ·
yi ≥ Xv′ · (|v| − α)
(14)
i∈Xv′
31
(x)
Then, let h(x) be the hypothesis defined as hi = sgn(xi ). Consider the inner product with y.
1 X (x)
hh(x) , yi =
hi · yi
(15)
N
i∈X
1 X X (x)
=
hi · yi
(16)
N
v∈[−1,1] i∈Xv
≥
1
N
X
sgn(v) ·
yi − α
1
Xv′ · |v| − 2α
N
v∈[−1,1]
X
1
1 X
Xv′ ·
yi − 3α
≥
N
|Sv′ |
′
i∈Sv
v∈[−1,1]
!
1 X
yi − 4α
≥
N
i∈S
ρ
≥ − ω − 4α
2
≥
(17)
i∈Xv′
v∈[−1,1]
X
X
(18)
(19)
(20)
(21)
where the first equalities follow by the definition of h(x) ; (17) follows by the choice of X ′ and αmulticalibration; (18) follows by applying (14) for each v ∈ [−1, 1]; (19) follows by substituting v for
the empirical average of y over Sv′ invoking α-multicalibration for the appropriate choice of S ′ ⊆ S;
and (20) follows by noting that we can restrict our attention to S ′ ⊆ X ′ such that |Xv′ | ≥ |Sv′ | and
the triangle inequality.
Thus, h(x) satisfies the (ρ, τ )-weak agnostic learning guarantee for any τ ≤ ρ/4 − 4α by our choice
of ω = ρ/4.
5
Multicalibration achieves “best-in-class” prediction
While our notion of multicalibration provides a protection against discrimination for groups, we
argue that this protection comes at virtually no cost in the utility of the predictor. In fact, we
argue that Algorithm 3.2 can be used as an effective post-processing step to turn any predictor,
or family of predictors, into a multicalibrated predictor that achieves comparable (or improved)
prediction error.
Suppose we are given a collection C of sets of individuals on which we wish to be multicalibrated.
Additionally, suppose we have a collection of candidate predictors H, which achieves low prediction
error but might violate calibration arbitrarily. From these collections, we would like to produce
a predictor x that is α-multicalibrated on C but achieves prediction error commensurate with the
best predictor in H; in particular, kx − p∗ k2 should be not much larger than kh∗ − p∗ k2 (and ideally
would be smaller). In this sense, the calibrated x would not only be fair, but would also achieve
(approximately) best-in-class prediction error over H.
32
Consider some h ∈ H and consider the partition of X into sets according to the predictions of h
– in particular, we will first apply a λ-discretization to the range of each h to partition X into
categories. S
That is, let Sv (h) = {i : hi ∈ λ(v)}, and note that Sv (h) is disjoint from Sv′ (h) for
v 6= v ′ , and v∈Λ[0,1] Sv (h) = X . In addition to calibrating with respect to S ∈ C, we can also ask
for calibration on Sv (h) for all h ∈ H and v ∈ Λ[0, 1]. Specifically, let S(H) = {Sv (h)}h∈H,v∈Λ[0,1] ;
we consider imposing calibration on C ∪ S(H). Calibrating in this manner protects the groups
defined by C but additionally gives a strong utility guarantee.
Theorem 5.1 (Best-in-class prediction). Suppose C ⊆ 2X is a collection of subsets of X and H is
a set of predictors. Then there is a predictor x that is α-multicalibrated on C such that
kx − p∗ k2 − kh∗ − p∗ k2 < 6αN,
where h∗ = argminh∈H kh − p∗ k2 . Further, suppose that for all S ∈ C, |S| ≥ γN , and suppose
that set membership for S ∈ C and h ∈ H are computable by circuits of size at most s; then x is
computable by a circuit of size at most O(s/α4 γ).
The proof of Theorem 5.1 actually reveals something stronger: if x is calibrated on the set S(H),
then P
for every category Sv (h) ∈ S(H), if x is significantly different from h on this category – that
is, if i∈Sv (h) (hi − xi )2 is large – then x actually achieves significantly improved prediction error
on this category compared to h. This is stated formally in Lemma 5.2.
Lemma 5.2. Suppose y is an arbitrary predictor and let S(y) = {Sv (y)}v∈Λ[0,1] . Suppose x is an
arbitrary α-multicalibrated predictor on S(y). Then for all v ∈ Λ[0, 1],
X
X
(yi − p∗i )2 − (xi − p∗i )2 ≥
(v − xi )2 − (4α + λ) |Sv (y)| .
i∈Sv (y)
Consequently,
i∈Sv (y)
ky − p∗ k2 − kx − p∗ k2 ≥ kx − yk2 − (4α + λ)N.
This lemma shows that calibrating on the categories of a predictor not only prevents the squared
prediction error from degrading beyond a small additive approximation, but it also guarantees that
if calibrating changes the predictor significantly on any category, this change represents significant progress towards the true underlying probabilities on this category. Assuming Lemma 5.2,
Theorem 5.1 follows.
Proof of Theorem 5.1. Note that if x is α-multicalibrated on C, then x is α-multicalibrated on
any C ′ ⊆ C. Consider enforcing calibration on the collection C ∪ S(H) as defined above. If x is αcalibrated on C ∪ S(H) then it is α-multicalibrated on {Sv (h)}v∈Λ[0,1] for all h ∈ H and specifically
for h∗ . By Lemma 5.2, and the fact that the squared difference is nonnegative, we obtain the
following inequality:
kh∗ − p∗ k2 − kx − p∗ k2 ≥ kx − h∗ k2 − (4α + λ)N ≥ −(4α + λ)N.
This inequality suffices to prove the accuracy guarantee; however, to also guarantee the predictor
x can be implemented by a small circuit, we have to be a bit more careful. In particular, when
calibrating, we will ignore any Sv (h) such that |Sv (h)| < λαN . Note that because we have λdiscretized, there are at most 1/λ categories; thus, excluding the sets Sv (h) where |Sv (h)| < αλN
33
introduces at most an additional αN error. Taking λ = α, in turn, this implies that the difference
in squared prediction error can be bounded as kx − p∗ k2 − kh∗ − p∗ k2 ≤ 6αN . Finally, because the
sets we want to calibrate on are at least α2 γN in cardinality, the circuit complexity bound follows
by applying Algorithm 3.2 and Theorem 3.12.
Thus, given any method for learning an accurate predictor h, we can turn it into a method for
learning a fair and accurate predictor h′ by running Algorithm 3.2 on the set of categories of h.
Combined with Theorem 3.12, this theorem shows that for any such class of predictors H of bounded
complexity, there exists a calibrated predictor with similar circuit complexity that performs nearly
as well as the best h ∈ H in terms of accuracy. Further, by Lemma 5.2, this (nearly) best-in-class
property will hold not just over the entire domain X , but on every sufficiently large category Sv (h)
identified by some h ∈ H. That is, if x is
calibrated
on S(H), then for every category Sv (h), the
average squared prediction error Ei∈Sv (h) (xi − p∗i )2 will be at most 6α worse than prediction given
by h on this set. If we view H as defining a set S(H) of “computationally-identifiable” categories,
then we can view any predictor that is calibrated on S(H) as at least as fair and at least as accurate
on this set of computationally-identifiable categories as the predictor that identified the group (up
to some small additive approximation).
We turn to proving Lemma 5.2. The lemma follows by expanding the difference in squared prediction errors and invoking the definition of α-calibration.
Proof of Lemma 5.2. Let Svu represent the set of individuals i where y ∈ λ(v) and x assigns
value u. By the assumption that x is α-calibrated on S(y), we know for every Sv (y) ∈ S(y), there
is some subset Sv′ (y) ⊆ Sv (y) such that |Sv′ (y)| ≥ (1 − α) |Sv (y)| for which x’s predictions are
′ = S ′ (y) ∩ S (x); if x is α-calibrated with respect to
approximately correct. In particular, let Svu
u
v
Sv (y), this guarantees that for all values u ∈ [0, 1], we have
X
′
p∗i − u ≤ α Svu
.
(22)
′
i∈Svu
Using this fact, and the fact that the remaining α-fraction of Sv (y) can contribute at most α |Sv (y)|
to the squared error, we can express the difference in the squared errors of y and x on Sv (y):
X
X
X
X
(yi − p∗i )2 −
(xi − p∗i )2 =
(v − p∗i + (yi − v))2 −
(xi − p∗i )2
i∈Sv (y)
i∈Sv (y)
i∈Sv (y)
=
X
i∈Sv (y)
(v − p∗i )2 −
i∈Sv (y)
≥
X
X
i∈Sv (y)
(xi − p∗i )2 + 2
X
(v − p∗i )(yi − v)
i∈Sv (y)
(2(p∗i − v)(xi − v) − (xi − v)2 ) − λ |Sv (y)| .
(23)
i∈Sv (y)
where (23) follows by the observation that if yi ∈ λ(v), then |yi − v| ≤ λ/2 and |v − p∗i | is trivially
34
bounded by 1. We bound the sum over i ∈ X of the first term:
X
X X
(p∗i − v)(u − v)
(p∗i − v)(xi − v) =
u∈[0,1] i∈Svu
i∈Sv (y)
=
=
X
(u − v)
u∈[0,1]
i∈Svu
X
X
(u − v)
X
(p∗i − v)
(u − v + p∗i − u)
i∈Svu
u∈[0,1]
=
X
|Svu | (u − v)2 + (u − v)
X
!
(p∗i − u) .
i∈Svu
u∈[0,1]
At this point, we note that |u − v| ≤ 1. Thus, we can bound the contribution of the sum over Svu
by its negative absolute value:
!
X
X
∗
2
(pi − u)
≥
|Svu | (u − v) − |u − v|
i∈Svu
u∈[0,1]
≥
X
u∈[0,1]
≥
X
u∈[0,1]
≥
X
|Svu | (u − v)2 −
|Svu | (u − v)2 −
X
(p∗i − u) +
X
(p∗i − u) − α |Sv (y)|
′
i∈Svu
′
i∈Svu
X
′
i∈Svu \Svu
(p∗i − u)
|Svu | (u − v)2 − 2α |Sv (y)|
u∈[0,1]
=
X
(v − xi )2 − 2α |Sv (y)| ,
i∈Sv (y)
where we bound the sums over Svu by invoking α-calibration and applying (22). Plugging this
bound into (23), we see that
X
X
X
(v − xi )2
(v − x∗i )2 − 2α |Sv (y)| − λ |Sv (y)| −
(yi − p∗i )2 − (xi − p∗i )2 ≥ 2
i∈Sv (y)
i∈Sv (y)
i∈Sv (y)
=
X
(v
− x∗i )2
− (4α − λ) |Sv (y)| .
i∈Sv (y)
Summing over v ∈ [0, 1], we can conclude
ky − p∗ k2 − kx − p∗ k2 ≥ kx − yk2 − (4α − λ)N
showing the lemma.
Acknowlegments The authors would like to thank Cynthia Dwork, Roy Frostig, Parikshit
Gopalan, Moritz Hardt, Aditi Raghunathan, Jacob Steinhardt, and Greg Valiant for helpful discussions related to this work.
35
References
[BDLM01]
Shai Ben-David, Philip Long, and Yishay Mansour. Agnostic boosting. In Computational Learning Theory, pages 507–516. Springer, 2001.
[BH15]
Avrim Blum and Moritz Hardt. The ladder: A reliable leaderboard for machine
learning competitions. In International Conference on Machine Learning, pages 1006–
1014, 2015.
[BHJ+ 17]
Richard Berk, Hoda Heidari, Shahin Jabbari, Matthew Joseph, Michael Kearns, Jamie
Morgenstern, Seth Neel, and Aaron Roth. A convex framework for fair regression.
arXiv preprint arXiv:1706.02409, 2017.
[BM07]
Avrim Blum and Yishay Mansour. From external to internal regret. Journal of Machine Learning Research, 8(Jun):1307–1324, 2007.
[BNS+ 16]
Raef Bassily, Kobbi Nissim, Adam Smith, Thomas Steinke, Uri Stemmer, and
Jonathan Ullman. Algorithmic stability for adaptive data analysis. In Proceedings
of the 48th Annual ACM SIGACT Symposium on Theory of Computing, pages 1046–
1059. ACM, 2016.
[BR17]
Andrej Bogdanov and Alon Rosen. Pseudorandom functions: Three decades later. In
Tutorials on the Foundations of Cryptography, pages 79–158. Springer, 2017.
[CDPF+ 17] Sam Corbett-Davies, Emma Pierson, Avi Feller, Sharad Goel, and Aziz Huq. Algorithmic decision making and the cost of fairness. KDD, 2017.
[CG17]
Alexandra Chouldechova and Max G’Sell. Fairer and more accurate, but for whom?
arXiv preprint arXiv:1707.00046, 2017.
[Cho17]
Alexandra Chouldechova. Fair prediction with disparate impact: A study of bias in
recidivism prediction instruments. Big Data, 2017.
[CSV17]
L. Elisa Celis, Damian Straszak, and Nisheeth K. Vishnoi. Ranking with fairness
constraints. arXiv preprint arXiv:1704.06840, 2017.
[DFH+ 15a] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toni Pitassi, Omer Reingold, and
Aaron Roth. Generalization in adaptive data analysis and holdout reuse. In Advances
in Neural Information Processing Systems, pages 2350–2358, 2015.
[DFH+ 15b] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold, and
Aaron Roth. The reusable holdout: Preserving validity in adaptive data analysis.
Science, 349(6248):636–638, 2015.
[DFH+ 15c] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold, and
Aaron Leon Roth. Preserving statistical validity in adaptive data analysis. In Proceedings of the forty-seventh annual ACM symposium on Theory of computing, pages
117–126. ACM, 2015.
36
[DHP+ 12]
Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard S. Zemel.
Fairness through awareness. In Innovations in Theoretical Computer Science (ITCS),
pages 214–226, 2012.
[DR14]
Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy.
Foundations and Trends R in Theoretical Computer Science, 9(3–4):211–407, 2014.
[Fel10]
Vitaly Feldman. Distribution-specific agnostic boosting. In Proceedings of the First
Symposium on Innovations in Computer Science10, 2010.
[FH15]
Dean P Foster and Sergiu Hart. Smooth calibration, leaky forecasts, finite recall, and
nash dynamics. Technical report, Working paper, 2015.
[FKL16]
Benjamin Fish, Jeremy Kun, and Ádám D. Lelkes. A confidence-based approach
for balancing fairness and accuracy. In Proceedings of the 2016 SIAM International
Conference on Data Mining, pages 144–152. SIAM, 2016.
[GGM84]
Oded Goldreich, Shafi Goldwasser, and Silvio Micali. How to construct random functions. In Foundations of Computer Science, 1984. 25th Annual Symposium on, pages
464–479. IEEE, 1984.
[GRU12]
Anupam Gupta, Aaron Roth, and Jonathan Ullman. Iterative constructions and private data release. Theory of Cryptography, pages 339–356, 2012.
[Hau92]
David Haussler. Decision theoretic generalizations of the pac model for neural net and
other learning applications. Information and computation, 100(1):78–150, 1992.
[Haz16]
Elad Hazan. Introduction to online convex optimization. Foundations and Trends R
in Optimization, 2(3-4):157–325, 2016.
[HPS16]
Moritz Hardt, Eric Price, and Nathan Srebro. Equality of opportunity in supervised
learning. In Advances in Neural Information Processing Systems, pages 3315–3323,
2016.
[HR10]
Moritz Hardt and Guy N. Rothblum. A multiplicative weights mechanism for privacypreserving data analysis. In Foundations of Computer Science (FOCS), 2010 51st
Annual IEEE Symposium on, pages 61–70. IEEE, 2010.
[HRRU14]
Justin Hsu, Aaron Roth, Tim Roughgarden, and Jonathan Ullman. Privately solving
linear programs. In International Colloquium on Automata, Languages, and Programming, pages 612–624. Springer, 2014.
[JKM+ 17]
Matthew Joseph, Michael Kearns, Jamie H. Morgenstern, Seth Neel, and Aaron Roth.
Fair algorithms for infinite and contextual bandits. arXiv preprint arXiv:1610.09559,
2017.
[JKMR16]
Matthew Joseph, Michael Kearns, Jamie H. Morgenstern, and Aaron Roth. Fairness in learning: Classic and contextual bandits. In Advances in Neural Information
Processing Systems, pages 325–333, 2016.
37
[Kea98]
Michael Kearns. Efficient noise-tolerant learning from statistical queries. Journal of
the ACM (JACM), 45(6):983–1006, 1998.
[KK09]
Varun Kanade and Adam Kalai. Potential-based agnostic boosting. In Y. Bengio,
D. Schuurmans, J. D. Lafferty, C. K. I. Williams, and A. Culotta, editors, Advances
in Neural Information Processing Systems 22, pages 880–888. Curran Associates, Inc.,
2009.
[KL15]
Volodymyr Kuleshov and Percy S Liang. Calibrated structured prediction. In Advances
in Neural Information Processing Systems, pages 3474–3482, 2015.
[KMR17]
Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. Inherent trade-offs in
the fair determination of risk scores. ITCS, 2017.
[KMV08]
Adam Tauman Kalai, Yishay Mansour, and Elad Verbin. On agnostic boosting and
parity learning. In Proceedings of the fortieth annual ACM symposium on Theory of
computing, pages 629–638. ACM, 2008.
[KNRW17] Michael Kearns, Seth Neel, Aaron Roth, and Zhiwei Steven Wu. Preventing fairness gerrymandering: Auditing and learning for subgroup fairness. arXiv preprint
arXiv:1711.05144v3, 2017.
[KP08]
Subhash Khot and Ashok Kumar Ponnuswami. Minimizing wide range regret with
time selection functions. In COLT, pages 81–86, 2008.
[KRCP+ 17] Niki Kilbertus, Mateo Rojas-Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, and Bernhard Schölkopf. Avoiding discrimination through causal reasoning. arXiv preprint arXiv:1706.02744, 2017.
[KRR18]
Michael P. Kim, Omer Reingold, and Guy N. Rothblum.
Fairness through
computationally-bounded awareness. arXiv preprint arXiv:1803.03239, 2018.
[KSS94]
Michael J. Kearns, Robert E. Schapire, and Linda M. Sellie. Toward efficient agnostic
learning. Machine Learning, 17(2-3):115–141, 1994.
[KV94]
Michael J. Kearns and Umesh Virkumar Vazirani. An introduction to computational
learning theory. MIT press, 1994.
[LRD+ 17]
Yang Liu, Goran Radanovic, Christos Dimitrakakis, Debmalya Mandal, and David C.
Parkes. Calibrated fairness in bandits. arXiv preprint arXiv:1707.01875, 2017.
[PRW+ 17]
Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q. Weinberger.
On fairness and calibration. NIPS, 2017.
[SS12]
Shai Shalev-Shwartz. Online learning and online convex optimization. Foundations
and Trends R in Machine Learning, 4(2):107–194, 2012.
[SSV03]
Alvaro Sandroni, Rann Smorodinsky, and Rakesh V Vohra. Calibration with many
checking rules. Mathematics of operations Research, 28(1):141–153, 2003.
38
[TTV09]
Luca Trevisan, Madhur Tulsiani, and Salil Vadhan. Regularity, boosting, and efficiently simulating every high-entropy distribution. In Computational Complexity,
2009. CCC’09. 24th Annual IEEE Conference on, pages 126–136. IEEE, 2009.
[Ull15]
Jonathan Ullman. Private multiplicative weights beyond linear queries. In Proceedings
of the 34th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database
Systems, pages 303–312. ACM, 2015.
[Val84]
Leslie G. Valiant. A theory of the learnable.
27(11):1134–1142, 1984.
39
Communications of the ACM,
| 8cs.DS
|
arXiv:1601.04665v1 [math.AC] 18 Jan 2016
GORENSTEIN INJECTIVE ENVELOPES AND COVERS
OVER TWO SIDED NOETHERIAN RINGS
ALINA IACOB
Abstract. We prove that the class of Gorenstein injective modules is both enveloping and covering over a two sided noetherian
ring such that the character modules of Gorenstein injective modules are Gorenstein flat. In the second part of the paper we consider the connection between the Gorenstein injective modules and
the strongly cotorsion modules. We prove that when the ring R
is commutative noetherian of finite Krull dimension, the class of
Gorenstein injective modules coincides with that of strongly cotorsion modules if and only if the ring R is in fact Gorenstein.
1. Introduction
The class of Gorenstein injective modules was introduced by Enochs
and Jenda in 1995 ([9]). Together with the Gorenstein projective and
with the Gorenstein flat modules, they are the key elements of Gorenstein homological algebra. So it is natural to consider the question of
the existence of the Gorenstein injective envelopes and covers. Their
existence is known over Gorenstein rings ([7]). It is also known ([6])
that when R is a commutative noetherian ring such that the character
modules of Gorenstein injective modules are Gorenstein flat, then the
class of Gorenstein injective modules is both enveloping and covering.
We extend these results to two sided noetherian rings (not necessarily
commutative). More precisely, we show that when R is a two sided
noetherian ring and the character module of any Gorenstein injective
left R-module is Gorenstein flat, the class of Gorenstein injective left
R-modules is both enveloping and covering.
Then we consider the relation between the Gorenstein injective modules and the strongly cotorsion modules. It is known that over a commutative noetherian ring R of finite Krull dimension, any Gorenstein
injective module is strongly cotorsion ([16]). We prove that over such
2010 Mathematics Subject Classification. 18G10, 18G25, 18G35.
key words: Gorenstein injective module, Gorenstein injective envelope, Gorenstein injective cover, strongly cotorsion module.
1
2
ALINA IACOB
a ring the class of Gorenstein injective modules coincides with that of
strongly cotorsion modules if and only if the ring R is in fact Gorenstein.
2. Preliminaries
Throughout this section R denotes an associative ring with unity. By
R-module we mean a left R-module.
Definition 1. ([7], definition 10.1.1) A module G is Gorenstein injective if there is an exact complex . . . → E1 → E0 → E−1 → . . . of
injective modules which remains exact under application of the functors Hom(Inj, −) , where Inj stands for all injective modules, and
such that G = Ker(E0 → E−1 ).
We will use the notation GI for the class of Gorenstein injective modules.
A stronger notion is that of strongly Gorenstein injective module:
Definition 2. ([2], definition 2.1) An R-module M is strongly Gorenstein injective if there exists an exact and Hom(Inj, −) exact sequence
f
f
... → E −
→E−
→ E → ...
with E injective and with M = Ker(f ).
It is known ([2], Theorem on page 3) that a module is Gorenstein
injective if and only if it is a direct summand of a strongly Gorenstein
injective one.
The Gorenstein flat modules are defined in terms of the tensor product.
Definition 3. ([7], Definition 10.3.1) An R-module G is Gorenstein
flat if there exists an exact and Inj ⊗ − exact sequence of flat modules
. . . → F1 → F0 → F−1 → . . .
such that G = Ker(F0 → F−1 ).
We will use the notation GF for the class of Gorenstein flat modules.
We recall that the character module of a left R-module M is the right
R-module M + = HomZ (M, Q/Z).
Gorenstein injective envelopes and covers over two sided noetherian rings
3
It is known ([12], Theorem 3.6) that if the ring R is right coherent then
a module G is Gorenstein flat if and only if its character module G+ is
Gorenstein injective.
Given a class of R-modules F , we will denote as usual by F ⊥ the class
of all R-modules M such that Ext1 (F, M) = 0 for every F ∈ F .
The left orthogonal class of F , denoted ⊥ F , is the class of all R N such
that Ext1 (N, F ) = 0 for every F ∈ F .
We recall that a pair (L, C) is a cotorsion pair if L⊥ = C and ⊥ C = L.
A cotorsion pair (L, C) is complete if for every R M there exists exact
sequences 0 → C → L → M → 0 and 0 → M → C ′ → L′ → 0 with C,
C ′ in C and L, L′ in L.
Definition 4. ([5], Definition 1.2.10) A cotorsion pair (L, C) is called
hereditary if one of the following equivalent statements hold:
(1) L is resolving, that is, L is closed under taking kernels of epimorphisms.
(2) C is coresolving, that is, C is closed under taking cokernels of
monomorphisms.
(3) Exti (F, C) = 0 for any F ∈ F and C ∈ C and i ≥ 1.
Our main results in section 3 are about the existence of covers and envelopes with respect to the class of Gorenstein injective modules. We
recall the following definitions:
Definition 5. ([5], Definition 1.2.3) Let C be a class of left R-modules.
A homomorphism φ : M → C is a C-preenvelope of M if C ∈ C and
if for any homomorphism φ′ : M → C ′ with C ′ ∈ C, there exists
u ∈ HomR (C, C ′ ) such that φ′ = uφ.
A C-preenvelope φ : M → C is a C-envelope if any endomorphism
u ∈ HomR (C, C) such that φ = uφ is an automorphism of C.
(Pre)covers and covers are defined dually.
In the case when C is the class of Gorenstein injective modules, a
C-(pre)cover (preenvelope respectively) is called a Gorenstein injective
(pre)cover (preenvelope respectively).
Duality pairs were introduced by Holm and Jørgensen in [11]. We recall their definition (the opposite ring is denoted Rop ):
4
ALINA IACOB
Definition 6. ([11], Definition 2.1) A duality pair over a ring R is
a pair (M, C) where M is a class of R-modules and C is a class of
Rop -modules, subject to the following conditions:
(1) For an R-module M, one has M ∈ M if and only if M + ∈ C.
(2) C is closed under direct summands and finite direct sums.
A duality pair (M, C) is called (co)product closed if the class M is
closed under (co)products in the category of all R-modules.
A duality pair (M, C) is called perfect if it is coproduct-closed, if M
is closed under extensions, and if R belongs to M.
We also recall that a cotorsion pair (L, C) is perfect if C is an enveloping
class and L is a covering class.
The following result is [11], Theorem 3.1:
Theorem. Let (M, C) be a duality pair. Then M is closed under pure
submodules, pure quotients, and pure extensions. Furthermore, the
following hold:
(a) If (M, C) is product-closed then M is preenveloping.
(b) If (M, C) is coproduct-closed then M is covering.
(c) If (M, C) is perfect then (M, M⊥ ) is a perfect cotorsion pair.
3. Gorenstein injective envelopes and covers
We recall the following results from [6]:
Lemma 1. ([6], Corollary 1) If R is left noetherian then (⊥ GI, GI) is
a complete hereditary cotorsion pair.
Theorem 1. ([6], Proposition 2) Let R be a left noetherian ring. The
class of Gorenstein injective modules is enveloping if and only if the
class ⊥ GI is covering.
We prove that if the ring R is two sided noetherian and has the property
that the character modules of Gorenstein injective left R-modules are
Gorenstein flat right R-modules, then the class of Gorenstein injective
modules is enveloping. Since the class ⊥ GI is closed under arbitrary
direct sums, it suffices to prove that ⊥ GI is the left half of a duality
pair. Then by [11], Theorem 3.1 (b), the class ⊥ GI is covering. By
Theorem 1, GI is enveloping in this case.
We start with the following
Gorenstein injective envelopes and covers over two sided noetherian rings
5
Lemma 2. Let R be a two sided noetharin ring such that for any (left)
Gorenstein injective module M, its character module, M + is a (right)
Gorenstein flat module. Then a left R-module K is in the class ⊥ GI if
and only if its character module K + is in GF ⊥ .
Proof. ⇒ Let K ∈⊥ GI. For any Gorenstein flat right R-module B we
have B + ∈ GI ([12], Theorem 3.6). It follows that Ext1 (K, B + ) = 0.
Then Ext1 (B, K + ) ≃ Ext1 (K, B + ) = 0. So K + ∈ GF ⊥ .
⇐ Assume that K is a left R-module such that K + ∈ GF ⊥ .
Since R is noetherian there exists an exact sequence 0 → K → G →
V → 0 with G Gorenstein injective and with V ∈⊥ GI (by Lemma
1). This gives an exact sequence 0 → V + → G+ → K + → 0 with G+
Gorenstein flat. Since V ∈⊥ GI we have that V + ∈ GF ⊥ . Since K +
is also in GF ⊥ it follows that G+ ∈ GF ∩ GF ⊥ . Thus G+ is flat and
therefore G is injective (by [7], Theorem 3.2.16).
So we have an exact sequence 0 → K → G → V → 0 with G ∈ Inj ⊆⊥
GI and with V ∈⊥ GI. Since (⊥ GI, GI) is a hereditary cotorsion pair,
it follows that K ∈⊥ GI.
Now we can prove:
Theorem 2. Let R be a two sided noetherian ring such that the character modules of Gorenstein injective left R-modules are Gorenstein flat
right R-modules. Then (⊥ GI, GF ⊥ ) is a duality pair. In particular the
class ⊥ GI is covering.
Proof. By Lemma 2 we have that K ∈⊥ GI if and only if K + ∈ GF ⊥ .
Any right orthogonal class (in particular GF ⊥ ) is closed under direct
products, and so it is closed under finite direct sums. Also, any right
orthogonal class (so GF ⊥ in particular) is closed under direct summands.
Thus (⊥ GI, GF ⊥ ) is a duality pair. Since the class ⊥ GI is closed under
direct sums, it follows (by [11], Theorem 3.1) that ⊥ GI is covering.
Theorem 3. Let R be a two sided noetherian ring such that the character modules of Gorenstein injective left R-modules are Gorenstein
flat right R-modules. Then the class of Gorenstein injective modules is
enveloping.
Proof. This follows from Theorem 1 and Theorem 2.
6
ALINA IACOB
We prove that over the same kind of rings the class of Gorenstein injective modules is also covering.
Theorem 4. Let R be a two sided noetherian ring such that the character modules of Gorenstein injective modules are Gorenstein flat. Then
(GI, GF ) is a duality pair.
Proof. - We prove first that K ∈ GI if and only if K + ∈ GF.
One implication is a hypothesis we made on the ring.
Assume that K + ∈ GF . Since the ring R is left noetherian there exists
an exact sequence 0 → K → G → L → 0 with G Gorenstein injective and L ∈⊥ GI. Therefore we have an exact sequence 0 → L+ →
G+ → K + → 0 with G+ Gorenstein flat, and L+ ∈ GF ⊥ (by Lemma
2). Then Ext1 (K + , L+ ) = 0, so G+ ≃ L+ ⊕ K + , and therefore
L+
T
is a Gorenstein flat right R-module. It follows that L+ ∈ GF GF ⊥ ,
so L+ is a flat right R-module, and therefore L is injective (by [7],
Theorem 3.2.16). The exact sequence 0 → K → G → L → 0 with
G Gorenstein injective and with L injective, gives that K has finite
Gorenstein injective dimension. By [3] Lemma 2.18, there exists an
exact sequence 0 → B → H → K → 0 with B Gorenstein injective and with idR H = GidR K < ∞. This gives an exact sequence
0 → K + → H + → B + → 0. Both B + and K + are Gorenstein flat
modules, so H + is also Gorenstein flat. Since idR H is finite it follows
that H + has finite flat dimension ([7], Theorem 3.2.19). But a Gorenstein flat module of finite flat dimension is flat ([7], Corollary 10.3.4).
So H + is flat and therefore H is injective. Thus GidR K = 0.
- We can prove now that (GI, GF) is a duality pair.
The first condition from the definition holds by the above.
It is known that the class of Gorenstein flat modules is closed under
(finite) direct sums, and under direct summands ([4], Theorem 4.14).
In the proof of the result about the existence of the Gorenstein injective envelopes we will use the following lemma:
Lemma 3. A direct sum of modules is a pure submodule of the direct
product of the modules.
Proof. Let (Xi )i∈I be a family of R-modules. The direct sum ⊕i∈I Xi
of the family is the direct limit of the finite direct sums Yj = ⊕i≤j Xi
Gorenstein injective envelopes and covers over two sided noetherian rings
7
of these modules. Any
Q finite direct sum Yj is a direct summand of the
direct product Y = Xi , so it is a pure submodule of Y . Therefore the
sequence 0 → A⊗Yj → A⊗Y is exact, for any R-module A. Then 0 →
lim
−→(A⊗Yj ) → lim
−→(A⊗Y ) is exact. Since the tensor product commutes
with direct limits we have an exact sequence 0 → A⊗(lim
−→Yj ) → A⊗Y .
Q
So 0 → A ⊗ (⊕i∈I Xi ) → A ⊗ Xi is exact for any R-module A.
Theorem 5. Let R be a two sided noetherian ring such that the character modules of Gorenstein injective left R-modules are Gorenstein
flat right R-modules. Then the class of Gorenstein injective modules is
covering in R − Mod.
Proof. By Theorem 4, (GI, GF) is a duality pair. Then by [11] Theorem 3.1, the class of Gorenstein injective modules is closed under pure
submodules. Since GI is closed under direct products and every direct
sum is a pure submodule of a direct product (Lemma 3), it follows that
GI is closed under arbitrary direct sums. Another application of [11],
Theorem 3.1 gives that GI is covering.
The following is an equivalent description of the condition that (GI, GF)
is a duality pair.
Proposition 1. Let R be a two sided noetherian ring such the character modules of Gorenstein injective modules are Gorenstein flat. The
following are equivalent:
1. (GI, GF) is a duality pair;
2. The class of Gorenstein injective modules, GI, is closed under pure
submodules.
Proof. 1. ⇒ 2. follows from [11], Theorem 3.1.
2. ⇒ 1. Assume that K + is a right Gorenstein flat module.
The ring R is noetherian so there is an exact sequence 0 → K → G →
L → 0 with G Gorenstein injective and with L ∈⊥ GI. By the proof of
Theorem 4, G+ ≃ L+ ⊕ K + , so the sequence 0 → K → G → L → 0 is
pure exact. Since GI is closed under pure submodules it follows that
K is Gorenstein injective. Thus K is Gorenstein injective if and only
if its character module, K + , is Gorenstein flat.
Since the class of Gorenstein flat modules is closed under direct sums
and under direct summands, it follows that (GI, GF ) is a duality pair.
8
ALINA IACOB
Our results hold over two sided noetherian rings with the property that
the character modules of left Gorenstein injective modules are Gorenstein flat right R-modules. It is known that this class of rings includes
the Gorenstein rings ([7]), as well as the commutative noetherian rings
with dualizing complexes ([11]). But these are not the only rings with
this property. We proved ([13]) that every two sided noetherian ring R
with i.d.Rop R < ∞ has the desired property.
Theorem 6. ([13], Theorem 2) Let R be a two sided noetherian ring
such that i.d.Rop R ≤ n for some positive integer n. Then the character
modules of Gorenstein injective left R-modules are Gorenstein flat right
R-modules.
Then by Theorem 3, Theorem 5 and Theorem 6 we obtain:
Theorem 7. Let R be a two sided noetherian ring such that i.d.Rop R <
∞. Then the class of left Gorenstein injective R-modules is both enveloping and covering in R − Mod.
4. When is every strongly cotorsion module Gorenstein
injective?
We recall first the following:
Definition 7. ([8], Definition 2.6) A module R M is called strongly cotorsion if Ext1 (F, M) = 0 for every module F of finite flat dimension.
It is known (see for example [19], Lemma 2.2) that M is a strongly
cotorsion module if and only if Exti (F, M) = 0 for all i ≥ 1, for any
module F of finite flat dimension.
We will use the notation SC for the class of strongly cotorsion modules,
and F for the class of modules of finite flat dimension.
We recall ([10]) that a ring R is said to be left n-perfect if every left
R-module of finite flat dimension has projective dimension less than or
equal to n.
Left perfect rings, commutative noetherian rings of finite Krull dimension, the universal enveloping algebra U(g) of a Lie algebra of dimension
n, and n-Gorenstein rings are all examples of left n-perfect rings (by
[10]).
Proposition 2. If R is left n-perfect then every Gorenstein injective
left R-module is strongly cotorsion.
Gorenstein injective envelopes and covers over two sided noetherian rings
9
Proof. Let M be a strongly Gorenstein injective left R-module. Then
there is an exact sequence 0 → M → E → M → 0 with R E an injective module. It follows that Exti (−, M) ≃ Ext1 (−, M) for all i ≥ 1.
Let R X be a module of finite flat dimension. Since R is left n-perfect,
the projective dimension of X is less than or equal to n. So Exti (X, M) =
0 for all i > n. Then by the above we have that Exti (X, M) = 0 for
all i ≥ 1. Thus M ∈ SC.
If M ′ is a Gorenstein injective R-module then there exists a strongly
Gorenstein injective module M such that M ≃ M ′ ⊕ M” (by [2]).
It follows that Ext1 (X, M ′ ) = 0 for any X ∈ F . So M ′ is strongly
cotorsion.
Proposition 3. Let R be a two sided noetherian ring such that R is
left n-perfect. If the class of strongly cotorsion modules coincides with
that of Gorenstein injective modules then i.d.Rop R < ∞.
Proof. Since R is left noetherian, (⊥ GI, GI) is a complete hereditary
cotorsion pair. By [19], Proposition 2.14, (F , SC) is also a cotorsion
pair.
By our assumption, SC is the class of Gorenstein injective modules.
It follows that ⊥ SC =⊥ GI, that is, F =⊥ GI. Since Inj ⊆⊥ GI, it
follows that every injective module has finite flat dimension.
Since R is right noetherian we have that i.d.Rop R = sup{f.d.I|R I ∈
Inj} (by [14]). It follows that i.d.Rop R ≤ n.
In particular, any commutative noetherian ring of Krull dimension
n < ∞ is left n-perfect. Over such a ring the condition i.d.Rop R < ∞
is equivalent with R being a Gorenstein ring. So using Proposition 3
we obtain the following result:
Theorem 8. Let R be a commutative noetherian ring of finite Krull
dimension. The following are equivalent:
1) The class of Gorenstein injective modules coincides with that of
strongly cotorsion modules.
2) R is a Gorenstein ring.
Proof. 1) ⇒ 2) By Proposition 3, i.d.Rop R ≤ n. Since R is commutative
noetherian and such that i.d.Rop R ≤ n it follows that R is a Gorenstein
ring.
2)⇒ 1) Since R is a Gorenstein ring the class ⊥ GI consists of all modules of finite projective dimension ([7]). By [7] again, this is the class
10
ALINA IACOB
F of modules of finite flat dimension. Then GI = (⊥ GI)⊥ = F ⊥ = SC.
References
[1] D. Bennis and N. Mahdou. Strongly Gorenstein projective, injective and flat
modules. J. Pure Appl. Algebra, (210):1709–1718, 2007.
[2] D. Bennis and N. Mahdou. Strongly Gorenstein projective, injective, and flat
modules,. J. Pure Appl. Algebra, 210:437–445, 2007.
[3] L. Christensen and A. Frankild and H. Holm. On Gorenstein projective, injective, and flat modules. A functorial description with applications,. J. Algebra,
302(1):231–279, 2006.
[4] L.W. Christensen, H.B. Foxby, and H. Holm Beyond totally reflexive modules
and back. A survey on Gorenstein dimensions. book chapter in ”Commutative Algebra: Noetherian and non-Noetherian perspectives”, 101-143. SpringerVerlag, 2011.
[5] J.R. Garcı́a Rozas. Covers and evelopes in the category of complexes of modules.
CRC Press LLC, 1999.
[6] E.E. Enochs and A. Iacob. Gorenstein injective covers and envelopes over noetherian rings. Proc. American Math. Soc., accepted.
[7] E.E. Enochs and O.M.G. Jenda. Relative Homological Algebra. Walter de
Gruyter, 2000. De Gruyter Exposition in Math.
[8] E.E. Enochs, and Z. Huang Injective envelopes and (Gorenstein) flat covers.
Algebras and Representation Theory, 15:1131–1145, 2012.
[9] E.E. Enochs and O.M.G. Jenda. Gorenstein injective and projective modules.
Mathematische Zeitschrift, (220):611–633, 1995.
[10] E.E. Enochs, and O.M.G. Jenda, and J.A. López-Ramos Dualizing modules
and n-perfect rings. Proc. Edinburgh Math. Society, 48:79–90, 2005.
[11] H. Holm, and P. J ørgensen Cotorsion pairs induced by duality pairs. J. Commut.Algebra, 1(4):621–633, 2009.
[12] H. Holm. Gorenstein homological dimensions. J. Pure and Appl. Alg., 189:167–
193, 2004.
[13] A. Iacob. Gorenstein flat preenvelopes. submitted.
[14] Y. Iwanaga On rings with finite self injective dimension II. Tsukuba J. Math.,
4:107–113, 1980.
[15] P. J ørgensen Finite flat and projective dimension. Comm. Algebra, 33(7):22752279, 2005.
[16] K. Khashyarmanesh and Sh. Salarian. On the rings whose injective hulls are
flat. Proc. of the American Math. Soc., 131(8):2239-2335, 2003.
[17] Z. Huang Gorenstein injective and strongly cotorsion modules. Israel J. Math,
to appear.
[18] C.U. Jensen On the vanishing of lim
←−. J. Algebra, 15: 151 – 166, 1970.
[19] H. Yan Strongly cotorsion (torsion-free) modules and cotorsion pairs. Bull.
Korean Math. Soc., 47(5):1041-1052, 2010.
[20] N. Ding and J. Chen The flat dimensions of injective modules. Manuscripta
Math., 78(1):165-177, 1993.
Gorenstein injective envelopes and covers over two sided noetherian rings
Alina Iacob
Georgia Southern University, Statesboro, GA 30460-8093
Email: [email protected]
11
| 0math.AC
|
Groupwise Maximin Fair Allocation of Indivisible Goods
arXiv:1711.07621v1 [cs.GT] 21 Nov 2017
Siddharth Barman∗ Arpita Biswas† Sanath Kumar Krishnamurthy‡ Y. Narahari§
Abstract
We study the problem of allocating indivisible goods among n agents in a fair manner. For
this problem, maximin share (MMS) is a well-studied solution concept which provides a fairness threshold. Specifically, maximin share is defined as the minimum utility that an agent
can guarantee for herself when asked to partition the set of goods into n bundles such that the
remaining (n − 1) agents pick their bundles adversarially. An allocation is deemed to be fair if
every agent gets a bundle whose valuation is at least her maximin share.
Even though maximin shares provide a natural benchmark for fairness, it has its own drawbacks and, in particular, it is not sufficient to rule out unsatisfactory allocations. Motivated by
these considerations, in this work we define a stronger notion of fairness, called groupwise
maximin share guarantee (GMMS). In GMMS, we require that the maximin share guarantee is
achieved not just with respect to the grand bundle, but also among all the subgroups of agents.
Hence, this solution concept strengthens MMS and provides an ex-post fairness guarantee. We
show that in specific settings, GMMS allocations always exist. We also establish the existence
of approximate GMMS allocations under additive valuations, and develop a polynomial-time
algorithm to find such allocations. Moreover, we establish a scale of fairness wherein we show
that GMMS implies approximate envy freeness.
Finally, we empirically demonstrate the existence of GMMS allocations in a large set of
randomly generated instances. For the same set of instances, we additionally show that our
algorithm achieves an approximation factor better than the established, worst-case bound.
1
Introduction
In recent years, the topic of fair division of indivisible goods has received significant attention in
the computer science, mathematics, and economics communities, see, for instance, Chapter 12
in [Brandt et al., 2016]. A central motivation behind this thread of research is the fact that classical
notions of fairness—such as envy freeness (EF)1 and proportionality2 —which were developed for
∗
Indian Institute of Science. [email protected]
Indian Institute of Science. [email protected]
‡
Chennai Mathematical Institute. [email protected]
§
Indian Institute of Science. [email protected]
1
An allocation is called envy free if every agent values her bundle at least as much she values any other agent’s
bundle [Foley, 1967, Varian, 1974, Stromquist, 1980].
2
An allocation, among n agents, is called proportionally fair if every agent’s value for her share is at least 1/n times
the total value of all the goods [Steinhaus, 1948].
†
1
divisible goods (that can be fractionally allocated), do not translate well to the indivisible case.
For instance, if there is a single indivisible good and two agents, then no allocation can guarantee
EF or proportionality. But, given that a number of real-world settings (such as budgeted course
allocations [Budish, 2011], division of inheritance, and partitioning resources in a cloud computing
environment) entail allocation of discrete/indivisible goods, it is essential to define and study
solution concepts which are applicable for a fair division of indivisible goods.
Classically, the applicability of solution concepts is studied via existence results. Understanding if and when a solution concept is guaranteed to exist is of fundamental importance in microeconomics and other related fields. Such existence results have been notably complemented
by research—in algorithmic game theory and artificial intelligence—that has focused on computational issues surrounding the underlying solution concepts. Broadly, our results contribute to
these key themes by establishing both existential and algorithmic results for a new notion of fairness.
In the context of fair division, the focus on developing efficient algorithms is motivated, in
part, by websites such as Spliddit3 [Goldman and Procaccia, 2015] and Adjusted Winner4 [Brams
and King, 2005], which offer fair solutions for dividing goods. Spliddit has attracted more than
fifty thousand users and, among other services, it computes allocations which are fair with respect
to the standard notions of fairness. One of the solution concepts considered by Spliddit is the
maximin share guarantee (MMS).
The MMS solution concept was defined in the notable work of [Budish, 2011], and it deems
an allocation to be fair if each agent gets a bundle of value greater than or equal to an agentspecific threshold, called the maximin share of the agent. Specifically, the maximin share of an agent
corresponds to the maximum value that the agent can attain for herself if she is hypothetically
asked to partition the set of goods into n bundles and, then, the remaining (n − 1) agents pick their
bundles adversarially. Hence, a risk-averse agent i would partition the goods (into n bundles)
such that value of the least desirable bundle (according to her) in the partition is maximized. The
value of the least desirable bundle in such a partition is called the maximin share of agent i. This
definition can be interpreted as a natural generalization of the classical cut-and-choose protocol.
Although maximin shares provide a natural benchmark to define fairness, this solution concept has its own drawbacks. In particular, MMS is not sufficient to rule out unsatisfactory allocations; see Section 3 for an example. Moreover, different MMS allocations can be drastically
different in terms of, say, the social welfare of the agents.
Motivated by these considerations, we define a strictly stronger notion of fairness, called groupwise maximin share guarantee (GMMS). Intuitively, GMMS provides an ex-post fairness guarantee:
it ensures that, even after the allocation has been made, the maximin share guarantee is achieved
not just with respect to the grand bundle of goods, but also among all the subgroups of agents
J ⊆ [n]. Specifically, we say that an allocation is GMMS if, for all groups J ⊆ [n] and agents i ∈ J,
the value of i’s bundle in the allocation is no less than the maximin share of i restricted to J. That
is, if the agent i were to repeat the thought experiment (of dividing all the goods allocated to the
3
4
www.spliddit.org
http://www.nyu.edu/projects/adjustedwinner/
2
agents in group J, so that the other |J| − 1 agents pick their bundles adversarially) to calculate
her maximin share restricted to J, then the value of her bundle is at least this threshold. This definition directly ensures that groupwise maximin share guarantee is a stronger solution concept:
GMMS implies MMS. In Section 3, we show that GMMS can, in fact, be arbitrarily better than an
allocation that just satisfies MMS.
GMMS also strictly generalizes pairwise maximin share guarantee (PMMS), a notion defined
by [Caragiannis et al., 2016]. In PMMS, the maximin share guarantee is required only for pairs
of agents, but not necessarily for the grand bundle. Section 2.1, provides an example which establishes that GMMS is a strict generalization of PMMS and MMS.
The relevance of GMMS is also substantiated by the fact that it implies other complementary
notions of fairness, which do not follow from MMS alone. In the context of indivisible goods,
relaxations of envy freeness, such as EF15 [Budish, 2011] and EFX6 [Caragiannis et al., 2016]
have also been studied. In Section 4 we show that (unlike MMS) GMMS fits into this scale of
fairness and, in particular, a GMMS allocation is guaranteed to be EFX (and hence EF1). These
implications essentially follow from the observation that, by definition, a GMMS allocation is
PMMS as well. [Caragiannis et al., 2016] have shown that PMMS implies EFX, and hence we
obtain the desired implications.
Throughout the paper, we focus on additive valuations, and a high-level contribution of our
work is to show that under additive valuations, a number of useful (existence, algorithmic, and
approximation) results which have been established for MMS continue to hold for GMMS as well.
1.1
Our Contributions
In addition to proving a scale of fairness for GMMS, we establish the following results:
1. Approximate groupwise maximin share allocations always exist under additive valuations.
Prior work has shown that there are instances wherein no allocation is MMS [Procaccia and
Wang, 2014, Kurokawa et al., 2016]. These non-existence results have motivated a detailed
study of approximate MMS allocations, i.e., allocations in which each agent gets a bundle
of value (multiplicatively) close to her maximin share. Along these lines, we consider approximate GMMS (see Definition 3), and show that, under additive valuations, a 1/2-GMMS
allocation always exists. In addition, such an allocation can be found in polynomial time.
2. GMMS allocations are guaranteed to exist when the valuations of the agents are either binary
or identical.
3. Analogous to the experimental results for MMS [Bouveret and Lemaı̂tre, 2014], GMMS allocations exist empirically. These simulation results indicate that we do not fall short on such
generic existence results by strengthening the maximin solution concept.
5
An allocation is said to be envy-free up to one good (EF1) if no agent envies any other after removing at most one
good from the other agent’s bundle; see Definition 4.
6
An allocation is said to be envy-free up to the least valued good (EFX) if no agent envies any other agent after
removing any positively valued good from the other agent’s bundle; see Definition 5.
3
1.2
Related Work
As mentioned earlier, [Budish, 2011] introduced the notion of maximin share guarantee (MMS),
and it has been extensively studied since then. In particular, [Bouveret and Lemaı̂tre, 2014] showed
that if the agents’ valuations are additive, then an envy free (or proportional) allocation will be
MMS as well. They also established that MMS exists under binary, additive valuations. Their
experiments, using different distributions over the valuations, did not yield a single example
wherein an MMS allocation did not exist. [Kurokawa et al., 2016] showed that MMS allocations
exist with high probability when valuations are drawn randomly.
However, [Procaccia and Wang, 2014] provided intricate counterexamples to refute the universal existence of MMS allocations, even under additive valuations. This motivated the study
of approximate maximin share allocations, α-MMS, where each agent obtains a bundle of value
at least α ∈ (0, 1) times her maximin share. [Procaccia and Wang, 2014] established the existence
of 2/3-MMS, and developed a polynomial-time algorithm to obtain such an allocation when the
number of agents is a fixed constant. Later, [Amanatidis et al., 2015] showed that a 2/3-MMS can
be computed in polynomial (in the number of players) time.
Approximate maximin share allocations have also been studied for general valuations. [Barman and Krishnamurthy, 2017] have developed an efficient algorithm which finds a 1/10-MMS
allocation under submodular valuations. More recently, [Ghodsi et al., 2017] have improved the
approximation guarantee for additive valuations to 3/4. They have also developed constant-factor
approximation guarantees for submodular and XOS valuations, along with a logarithmic approximation for subadditive valuations.
[Aziz et al., 2017] studied the fair division of indivisible chores (negatively valued goods) and
have developed an efficient algorithm which finds a 2-MMS allocation.
[Caragiannis et al., 2016] defined another important fairness notion called pairwise maximin
share guarantee (PMMS). As mentioned previously, under PMMS, the maximin share guarantee is
required only for pairs of agents, and not even for the grand bundle. They also established that
PMMS and MMS are incomparable: neither one of these solution concepts implies the other.
2
Preliminaries and Notation
We consider the problem of finding a fair allocation of a set of indivisible goods [m] = {1, . . . , m},
among a set of agents [n] = {1, . . . , n}. For a subset of goods S ⊆ [m] and integer t, let Πt (S)
denote the set of all t partitions of S. An allocation is defined as an n-partition (A1 , A2 , . . . , An ) ∈
Πn ([m]), where Ai is the set of goods allocated to agent i.
The preference of the agents over the goods is specified via valuations. Specifically, we denote the valuation of an agent i ∈ [n] for a subset of goods S ⊆ [m] by vi (S). In this work,
we assume the valuations to be non-negative and additive, i.e., vi ({g}) ≥ 0 for all g ∈ [m] and
P
vi (S) = g∈S vi ({g}). For ease of presentation, we will use vi (g) for agent i’s valuation of good
g, i.e., for vi ({g}).
As mentioned previously, the fairness notions considered in this work are defined using thresh-
4
olds called maximin shares. Formally, given an agent i, parameter k ∈ Z+ , and subset of goods S ⊆
[m], k-maximin share of i restricted to S is defined as µki (S) := max(P1 ,...,Pk )∈Πk (S) minj∈[k] vi (Pj ).
Throughout, MMSi will be used to denote the maximin share of an agent i with respect to the
grand bundle, MMSi := µni ([m]). We will now formally define maximin share allocation (MMS
allocation).
Definition 1 (Maximin Share Allocation). An allocation (A1 , . . . , An ) is said to be a maximin share
allocation (MMS) iff for all agents i ∈ [n] we have vi (Ai ) ≥ MMSi .
A different solution concept defined by [Caragiannis et al., 2016] requires the maximin share
guarantee to hold only for every pair of agents, i.e., an allocation A = (A1 , . . . , An ) is said to
achieve pairwise maximin share guarantee (PMMS) iff for all i, j ∈ [n], we have vi (Ai ) ≥ µ2i (Ai ∪ Aj )
and vj (Aj ) ≥ µ2j (Ai ∪ Aj ).
In this paper, we strengthen MMS and PMMS, and define a stronger threshold for each agent
i ∈ [n], namely groupwise maximin share (GMMSi ). Formally,
[
|J|
GMMSi := max
µi
Aj .
J⊆[n],J3i
j∈J
Now we define groupwise maximin share allocation (GMMS allocation).
Definition 2 (Groupwise Maximin Share Allocation). An allocation (A1 , . . . , An ) is said to be a groupwise maximin share allocation (GMMS) iff for all agents i ∈ [n] we have vi (Ai ) ≥ GMMSi .
|J| S
Note that an allocation (A1 , . . . , An ) is GMMS iff vi (Ai ) ≥ µi
A
for all J ⊆ [n] such
j
j∈J
that i ∈ J. Also, the threshold GMMSi is a function of the underlying allocation and, in contrast,
MMSi depends only the valuation of agent i for the m goods and the number of agents n.
The fact that there are fair division instances which do not admit an MMS allocation directly
implies that GMMS allocation are not guaranteed to exist either. Therefore, we consider approximate GMMS allocations.
Definition 3 (Approximate Groupwise Maximin Share Allocation). An allocation (A1 , . . . , An ) is
said to be an α-approximate groupwise maximin share allocation (α-GMMS) iff for all agents i ∈ [n] we
have vi (Ai ) ≥ αGMMSi .
A 1-approximate groupwise maximin share allocation is a GMMS allocation.
2.1 GMMS Strictly Generalizes MMS and PMMS
This section shows that GMMS is a distinct solution concept which strictly generalizes both MMS
and PMMS. In fact, the instance constructed in this section shows that there exists allocations
which are MMS and, furthermore, satisfy the maximin share guarantee for all subgroups of size
at most k (say), but do not satisfy the GMMS criteria.
We now formally define the notion of maximin share guarantee for a subgroup of size k. For
ease of presentation, we call such allocations k-wise fair. An allocation A = (A1 , . . . , An ) is said to
5
be k-wise fair iff forall agentsi ∈ [n] and for all size-k subsets J ⊆ [n] such that i ∈ J, the following
S
holds: vi (Ai ) ≥ µki
j∈J Aj .
In the following example we identify an allocation which is t-wise fair—for each t < k—but
not k-wise fair; here, k ≥ 4. Let us consider 3k − 4 goods and n agents, with n > 3k − 4. Since the
number of agents is greater than the number of goods, MMSi = 0 for all i ∈ [n] and, hence, all the
allocations are MMS. Consider an agent, say i = 1, and let her valuation for the 3k − 4 goods be
• 3k − 7 for the first k − 1 goods l1 , l2 , . . . , lk−1 (“large” goods),
• 3 for the next k − 2 goods d2 , . . . , dk−1 (“medium-sized” goods), and
• 1 for the remaining k − 1 goods s1 , ..., sk−1 (“small” goods).
The valuation of other agents can be set to ensure fairness for them. Let allocation A =
(A1 , . . . , An ) be such that A1 = {l1 }, A2 = {l2 , d2 }, . . ., Ak−1 = {lk−1 , dk−1 }, Ak = {s1 , ..., sk−1 }
and Ar = ∅ for all r ≥ k, i.e., no goods are allocated to the last n − k agents. We first show that
the bundle allocated to agent 1 is at least her pairwise maximin share with any other agent j ∈ [n].
Agent 1 does not envy agent k since v1 (l1 ) = 3k − 7 > k − 1 = v1 (Ak ) (since k ≥ 4). Moreover, for
the other agents j ∈ {2, . . . , k − 1}, we have µ21 ({l1 , lj , dj }) = 3k − 7 = v1 ({l1 }) = v1 (A1 ). Thus,
A ensures PMMS. Below, we present a case analysis which shows that A is also t-wise fair, for all
t ≤ k − 1.
1. Consider any subgroup J of the first k − 1 agents, containing agent 1. Let |J| = t. In this
case, exactly 2t − 1 goods are allocated to the group of agents J. In particular, this set of
goods consists of t “large” goods and t − 1 “medium-sized” goods. Hence, any t-partition
of these 2t − 1 goods will have at least one bundle with at most one good. Therefore, any
partition which maximizes the least valued (with respect to the valuation of agent 1) bundle
will have one bundle containing a “large” good, and each of the remaining t − 1 bundles
containing one “large” and one “medium-sized” good. Therefore, the allocation A1 = {l1 }
ensures maximin share guarantee for agent 1 for all subgroups J ⊆ {1, . . . , k − 1}, which
contain agent 1.
2. Next, consider any subgroup J of size t which includes agent k as well as agent 1. Here, the
set of goods allocated to agents in J consists of t − 1 “large” goods, t − 2 “medium-sized”
goods and k − 1 “small” goods. To simplify the analysis, we assume that t = k/2 and use an
averaging argument to obtain
(3k − 7)(t − 1) + 3(t − 2) + 1(k − 1)
t
2k
= 3k − 4 −
t
≤ 3k − 4 − 4
(since t ≤ k/2)
µt1 (∪j∈J Aj ) ≤
= 3k − 8
< 3k − 7 = v1 (A1 ).
6
Hence, for all such subgroups of agents of size t ≤ k/2, the bundle A1 = {l1 } satisfies the
maximin share guarantee of agent 1. A similar, but slightly involved argument establishes
the result for subgroups of size t ≤ k − 1.
3. Finally, consider a size-t subgroup J ∈ [n] such that 1 ∈ J and J contains at least one agent
h > k, i.e., J includes an agent h with an empty bundle Ah = ∅. Write J 0 := J \ {h : h > k}
and |J 0 | = t0 . The arguments used in the first two cases show that the maximin guarantee
0
for agent 1 holds for J 0 , i.e., v1 (A1 ) ≥ µt1 (∪j∈J 0 Aj ). Using the fact that t0 < t and ∪j∈J 0 Aj =
0
∪j∈J Aj , we get µt1 (∪j∈J 0 Aj ) ≥ µt1 (∪j∈J Aj ). These observations establish the required
maximin share guarantee, v1 (A1 ) ≥ µt1 (∪j∈J Aj ), for this case as well.
Overall, we get that the allocation A satisfies t-wise fairness, for any t < k. We will complete the
analysis by showing that A is not k-wise fair.
In particular, let J = [k]. Note that, here, the set ∪j∈J Aj consists of k − 1 “large” goods, k − 2
“medium-sized” goods and k−1 “small” goods. In addition, the maximin share value with respect
to this subgroup is 3k − 6: partition the goods to obtain one bundle containing all k − 2 “mediumsized” goods (each of value 3) and, additionally, k − 1 bundles each containing one “large” good
and one “small” good. Therefore, µk1 (∪j∈J Aj ) = 3k − 6 > 3k − 7 = v1 (A1 ).
This example illustrates that, in particular, GMMS is a stronger solution concept than PMMS
and MMS.
3 GMMS can be arbitrarily better than MMS
In this section, we provide a class of examples where an MMS allocation is not necessarily satisfactory in terms of agents’ valuations. In particular, we show that imposing GMMS leads to
allocations which Pareto dominate an allocation which only satisfies MMS.
Consider n agents and a set of n + 3 goods (where [m] = {g1 , . . . , gn+3 }), along with parameter
V and a sufficiently small ε; 0 < ε V . The valuations are assumed to be additive and are as
follows:
For agents i
∈ {1, . . . , n − 1},
V,
z ∈ {1, . . . , n − 3}
V /2,
z ∈ {n − 2, n − 1}
vi ({gz }) =
V /2 − ε,
z ∈ {n, n + 2}
V /2 + ε,
z ∈ {n + 1, n + 3}
For the agent
n,
V,
vn ({gz }) =
0,
ε,
z ∈ {1, . . . , n − 1}
z ∈ {n, n + 1}
z ∈ {n + 2, n + 3}
Note that in this fair division instance the maximin share of the first (n − 1) agents, MMSi = V
for all i ∈ {1, . . . , n − 1}. In addition, the maximin share of the last agent, MMSn , is 2ε.
7
Now, consider the allocation A = (A1 , . . . , An ) wherein Ai = {gi }, i ∈ {1, . . . , n − 3}, An−2 =
{gn−2 , gn−1 }, An−1 = {gn , gn+1 }, and An = {gn+2 , gn+3 }. Allocation A is MMS since vi (Ai ) =
V for each i ∈ {1, . . . , n − 1} and vn (An ) = 2ε. In this allocation the valuation of agent n is
unsatisfactorily low. Another relevant observation is that this allocation is not GMMS. Below,
we show that in this instance any GMMS allocation allocates a bundle of value V to every agent,
including agent n.
The fact that A is not GMMS follows by considering the goods allocated to agents n − 2 and
n, i.e., let S := An−2 ∪ An = {gn−2 , gn−1 , gn+1 , gn+3 }. Now, µ2n (S) = V + ε, but vn (An ) = 2ε V .
Furthermore, it can be observed that it is necessary to allocate the agent n at least one of her high
valued goods {1, . . . , n − 1} to satisfy GMMS. Thus, GMMS allocation would always ensure that
a bundle of at least V is allocated to all the agents. That is, in this instance, unsatisfactorily low
valuations can be avoided by imposing the groupwise maximin share guarantee.
4
Scale of Fairness
As mentioned earlier, envy freeness (EF) is a well-studied solution concept in the context of fair
division of divisible items. However, for indivisible goods, a simple example with one positively
valued good and two agents shows that envy-free allocations do not always exist. Hence, for
indivisible goods, natural relaxations of envy freeness—in particular, EF1 and EFX—have been
considered in the literature. We now provide formal definitions of these relaxations.
Definition 4 (Envy-free up to one good [Budish, 2011]). An allocation A = (A1 , A2 , . . . , An ) is said
to be envy-free up to one good (EF1) iff for every pair of agents i, j ∈ [n] there exists a good g ∈ Aj such
that vi (Ai ) ≥ vi (Aj \ {g}).
Definition 5 (Envy-free up to the least positively valued good [Caragiannis et al., 2016]). An allocation A = (A1 , A2 , . . . , An ) is said to be envy-free up to the least valued good (EFX) iff for every pair of
agents i, j ∈ [n] and for all goods g ∈ Aj ∩ {g 0 ∈ [m] | vi (g 0 ) > 0} (i.e., for all goods g in Aj which are
positively valued by agent i) we have vi (Ai ) ≥ vi (Aj \ {g}).
Note that the above mentioned definitions imply that, for additive valuations, an EFX allocation is necessarily EF1 as well. Next we show that, interestingly, these relaxed versions of envy
freeness are implied by GMMS, but not by MMS.
Proposition 1 (Scale of Fairness). In any fair division instance with additive valuations
1. If an allocation A is envy free (EF) then it achieves groupwise maximin share guarantee (GMMS) as
well.
2. If an allocation B is GMMS then it is EFX (and, hence, EF1) as well.
Proof. First we will show that EF implies GMMS: Assume that the allocation A = (A1 , . . . , An ) is
envy free, that is, for all i, j ∈ [n] we have vi (Ai ) ≥ vi (Aj ). Therefore, for any agent i and any group
P
of agents J ⊆ [n] we have |J|vi (Ai ) ≥ j∈J vi (Aj ). Since the valuation vi is additive, this inequal1
ity leads to the following bound vi (Ai ) ≥ |J|
vi (S); here S = ∪j∈J Aj . Now, an averaging argument
8
establishes the following inequality: for any |J|-partition of S if (P1 , P2 , . . . , P|J| ) ∈ Π|J| (S) then
|J|
vi (Ai ) ≥ min1≤k≤|J| vi (Pk ). Hence, vi (Ai ) ≥ max(P1 ,...,P|J| )∈Π|J| (S) mink vi (Pk ) = µi (S), and we
get that A is GMMS.
Next, we argue that GMMS implies EFX: [Caragiannis et al., 2016] have shown that that PMMS
implies EFX. By definition, GMMS implies PMMS. Hence, a GMMS allocation is guaranteed to
be EFX.
Note that [Caragiannis et al., 2016] also provided an example to show that maximin share
guarantee by itself does not imply EF1. Hence, an MMS allocation is not necessarily EFX. Consider a fair division instance with three agents, five goods, and each agent values each good at
1. The maximin share of all the agents is 1. Thus, the allocation A1 = {g1 , g2 , g3 }, A2 = {g4 } and
A3 = {g5 } satisfies MMS, but not EF1. This is because agents 2 and 3 continue to envy agent 1
even if a single good is removed from A1 .
This, overall, shows that while MMS is not enough to guarantee weaker notions of envy freeness, GMMS ensures fairness in terms of such notions, and secures a place in the scale of fairness.
Next we will consider the complementary direction of going from bounded envy to groupwise
maximin fairness. In particular, we will establish existence and algorithmic results for approximate GMMS by considering a solution concept which is stronger than EF1, but weaker than EFX.
Specifically, we will define allocations which are envy-free up to one less-preferred good (EFL)—see
Definition 6 below—and show that such allocations are guaranteed to exist, when the valuations
are additive. Note that, in contrast, the generic existence of EFX allocations remains an interesting
open question. Furthermore, we will prove that EFL allocations can be computed in polynomial
time and, under additive valuations, such allocations imply 1/2-GMMS.
Definition 6. An allocation A = (A1 , A2 , . . . , An ) is said to be envy-free up to one less-preferred
good (EFL) if for every pair of agents i, j ∈ [n] at least one of the following conditions hold:
• Aj contains at most one good which is positively valued by i; |Aj ∩ {g 0 | vi (g 0 ) > 0}| ≤ 1
• There exists a good g ∈ Aj such that vi (Ai ) ≥ vi (Aj \ {g}) and vi (Ai ) ≥ vi ({g}).
The fact that an EFL allocation is EF1 follows directly from the definitions of these solution
concepts. Also, note that if an allocation (A1 , . . . , An ) is EFX then for any pair of agents i, j ∈ [n]
with |Aj ∩ {g 0 | vi (g 0 ) > 0}| > 1 the second condition in the definition of EFL holds. In particular,
write Aij := Aj ∩ {g 0 | vi (g 0 ) > 0} and consider two distinct goods ĝ, g̃ ∈ Aij . Since the allocation is
EFX, we have vi (Ai ) ≥ vi (Aj \ {ĝ}) and vi (Ai ) ≥ vi (Aj \ {g̃}). Note that ĝ ∈ Aj \ {g̃}, and, hence,
vi (Ai ) ≥ vi (ĝ). This implies that good ĝ satisfies the second condition in Definition 6. Hence, any
EFX allocation is EFL as well.
With this new fairness notion, we have the following chain of implications: EF ⇒ GMMS ⇒
EFX ⇒ EFL ⇒ EF1.
9
5
An Approximation Algorithm for GMMS
Our main result in this section shows that 1/2-GMMS allocations always exist under additive
valuation, and such allocations can be found efficiently.
Theorem 1. Every fair division instance with additive valuations admits a 1/2-approximate groupwise
maximin share allocation. Furthermore, such an allocation can be found in polynomial time.
Proof-Sketch The proof proceeds in two steps. First, we provide a constructive proof for the existence of EFL allocations, under additive valuations (Section 5.1). Next, we complete the proof by
showing that EFL implies 1/2-GMMS (in Section 5.2)
5.1
Existence of EFL Allocations
This section shows that EFL allocations are guaranteed to exist when the valuations are additive.
Specifically, we develop an algorithm that always finds such an allocation.
Lemma 1. Given any fair division instance with additive valuations, Algorithm 1 finds an EFL allocation
in polynomial time.
Algorithm 1 iteratively allocates the goods and maintains a partial allocation, A = (A1 , . . . , An ),
of the goods assigned so far. In each iteration, the algorithm selects an agent i who is not currently
envied by any other agent, and allocates i an unassigned good of highest value (under vi ).
Algorithm 1 Finding an EFL Allocation
Input : n agents, m items, and valuations vi {g} for each agent i ∈ [n] and for each good g ∈ [m].
Output: An EFL allocation.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
Initialize allocation A = (A1 , A2 , . . . , An ) with Ai = ∅ for each agent i ∈ [n], and M = [m].
Initialize directed envy graph G(A) = ([n], E) where E = ∅.
while M 6= ∅ do
Pick a source vertex i of G(A). {such a vertex always exists, since G(A) is acyclic.}
Pick g ∈ arg maxj∈M vi {g}.
Update Ai ← Ai ∪ {g} and M ← M \ {g}.
while the current envy graph G(A) contains a cycle do
Update A (using Lemma 2) to remove the cycle.
end while
end while
Return A.
Throughout the execution of the algorithm, the existence of an unenvied agent is ensured by
maintaining a directed graph, G(A), that captures the envy between agents. The nodes in this
envy graph represent the agents and it contains a directed edge from i to j iff i envies j, i.e.,
vi (Ai ) < vi (Aj ). Lemma 2, established in [Lipton et al., 2004], shows that if any iteration leads
to a cycle in the envy graph G(A), then we can always resolve it to obtain an acyclic envy graph
10
without decreasing the valuation of any agents; for completeness, we provide a proof of Lemma 2.
It is relevant to that, since G(A) is acyclic for a partial allocation A, it necessarily contains a source
node, i.e., an agent i who is not envied by other agents.
Although, Algorithm 1 is similar to the algorithm developed in [Lipton et al., 2004]—which
efficiently finds EF1 allocations— here, instead of assigning goods in an arbitrary order, we always
allocate to an unenvied agent the available good she values the most. This is crucial for obtaining
an EFL allocation.
Lemma 2. [Lipton et al., 2004] Given a partial allocation A = (A1 , . . . , An ) of a subset of goods S ⊆ [m],
we can find another partial allocation B = (B1 , . . . , Bn ) of S in polynomial time such that
(i) The valuations of agents for their bundles do not decrease, that is, vi (Bi ) ≥ vi (Ai ) for all i ∈ [n].
(ii) The envy graph G(B) is acyclic.
Proof. If the envy graph of A is acyclic then the claim holds trivially. Otherwise, let us denote a
cycle in the graph G(A) by C = i1 → i2 → . . . → ik → i1 . Now, we can reallocate the bundles as
follows: for all agents not in C, i.e., k ∈
/ {i1 , i2 , . . . , ik } set A0k = Ak , and for all the agents in the
cycle set A0i to be the bundle of their successor in C, i.e., set A0ia = Ai(a+1) for 1 ≤ a < k along with
A0ik = Ai1 .
Note that after this reallocation we have vi (A0i ) ≥ vi (Ai ) for all i ∈ [n]. Furthermore, the
number of edges in G(A) strictly decreases: The edges in C do not appear in the envy graph of
(A01 , . . . , A0n ) and if an agent k starts envying an agent in the cycle, say agent ia , then k must have
been envious of ia+1 in A. Edges between agents k and k 0 which are not in C remain unchanged,
and edges going out of an agent i in the cycle C can only get removed, since i valuation for the
bundle assigned to her bundle increases. Therefore, we can repeatedly remove cycles and keep
reducing the number of edges in the envy graph to eventually a find a partial allocation B that
satisfies the stated claim.
Proof of Lemma 1. Write A = (A1 , . . . , An ) to denote the allocation returned by Algorithm 1. First,
we note that an inductive argument proves that A is EF1. In fact, we will show that the EF1
condition holds for A with respect to the last (in terms of the algorithm’s allocation order) good
g assigned to each bundle Aj . Write gt to denote the good allocated in the tth iteration of Algorithm 1; hence, the goods are allocated in the following order g1 , g2 , . . . , gm .
The initial partial allocation (∅, ∅, . . . , ∅) is EF1 (in fact, it is envy free). Now, say that in the
jth iteration the algorithm allocates good gj to agent i. Write Aj = (Aj1 , . . . , Ajn ) and Aj+1 =
j+1
(Aj+1
1 , . . . , An ), respectively, to denote the partial allocations before and after the jth iteration.
The induction hypothesis implies that Aj is EF1 with respect to the last assigned good. Therefore,
for every pair of agents r, s ∈ [n], we have vr (Ajr ) ≥ vr (Ajs \ {ga }), where ga is the last good
assigned to the bundle Ajs (i.e., for any other good gb ∈ Ajs , we have b < a).
Since the good gj is allocated to agent i, it must be the case that i is a source vertex in G(Aj ),
i.e., no agent envies i under Aj . This implies that, vr (Ajr ) ≥ vr (Aji ) = vr ((Aji ∪ {gj }) \ {gj })
for all r ∈ [n]. Note that at this point of time, gj is the last good assigned to the bundle Aji .
In addition, from the proof of Lemma 2, we know that Aj+1 is a permutation of the allocation
j
(Aj1 , Aj2 , . . . , Aji ∪ {gj }, . . . , Ajn ), and vr (Aj+1
r ) ≥ vr (Ar ) for all r ∈ [n]. Hence, for every pair of
11
j+1
agents r, s ∈ [n] there exists a good ga ∈ Aj+1
such that vr (Aj+1
\ {ga }). In addition, ga
s
r ) ≥ vr (As
j+1
is the last good assigned to the bundle As . That is, the stated property holds for Aj+1 as well.
Now, we will use this observation to prove that A is EFL. Specifically, we show the EFL
conditions hold for, say, agent 1 (analogous arguments establish the claim for the other agents).
Suppose that, during the execution of the algorithm, agent 1 receives its first good, gt , in the tth
iteration. Note that the partial allocation before the tth iteration, say At = (At1 , . . . , Atn ), satisfies
|Ati ∩ {g 0 ∈ [m] : v1 (g 0 ) > 0}| ≤ 1 for all i ∈ [n]. This bound follows from the observation that
during any previous iteration s < t, the selected source vertex i (i.e., the agent that gets a new
good during the sth iteration) does not contain any good which is positively valued by agent 1;
otherwise, i would have been envied by 1, contradicting the fact that it is a source vertex. Hence,
each bundle in At contains at most one good from the set {g 0 ∈ [m] : v1 (g 0 ) > 0}.
Let us now consider the final allocation A = (A1 , . . . , An ) and any agent j ∈ [n]. If |Aj ∩ {g 0 ∈
[m] : v1 (g 0 ) > 0}| ≤ 1, then the first condition in the definition of EFL holds and we are done,
else if |Aj ∩ {g 0 ∈ [m] : v1 (g 0 ) > 0} | > 1, then bundle Aj must have received a good after the tth
iteration. This is consequence of the above-mentioned property of the partial allocation At . Write
g` to denote the last good allocated to the bundle Aj . We have ` > t, since g` was assigned to Aj
after the tth iteration. Also, in the tth iteration good gt was selected instead of g` , hence it must be
the case that v1 (gt ) ≥ v1 (g` ).
Note that, as mentioned before, the EF1 condition holds for Aj with respect to g` , i.e., v1 (A1 ) ≥
v1 (Aj \ {g` }). In addition, Lemma 2 implies that v1 (A1 ) ≥ v1 ({gt }). Therefore, if |Aj ∩ {g 0 ∈
[m] : v1 (g 0 )}| > 1 for any j ∈ [n], then there exists a good g` such v1 (A1 ) ≥ v1 (Aj \ {g` }) and
v1 (A1 ) ≥ v1 (g` ). Hence, A is an EFL allocation.
5.2 EFL implies Approximate GMMS
Lemma 3. In any fair division instance with additive valuations, if an allocation A = (A1 , . . . , An ) is
EFL, then it is 1/2-GMMS allocation as well.
Proof. Fix an agent i and a set of k agents, J ⊆ [n] which contains i, i.e., |J| = k and J 3 i. Also, let
S = ∪j∈J Aj denote the set of all the goods allocated to the agents in J. We will show that agent
if (A1 , . . . , An ) is EFL, then vi (Ai ) is at least 12 times the maximin share of i restricted to S, i.e.,
vi (Ai ) ≥ 12 µki (S). This establishes the stated claim.
Write Si ⊆ [m] to denote the set of goods which are positively valued by i, Si := {g 0 ∈ [m] |
vi (g 0 ) > 0}. Now, among the set of agents J \ {i} consider the ones who are allocated at most one
good from Si ; specifically, let T := {j ∈ J \ {i} | |Aj ∩ Si | ≤ 1}. Write J 0 := J \ T , t0 := |J 0 |, and
S 0 = ∪j∈J 0 Aj . Note that the agent i belongs to the group J 0 , and for all j ∈ J 0 we have |Aj ∩Si | > 1.
Therefore, the fact that A is EFL implies that, for all j ∈ J 0 , there exists a good g(j) ∈ Aj such that
vi (Ai ) ≥ vi (Aj \ {g(j) }) and vi (Ai ) ≥ vi (g(j) ). In other words, for the additive valuation vi , we have
2vi (Ai ) ≥ vi (Aj ).
0
We will now establish the multiplicative bound with respect to µti (S 0 ) (the maximin share of i
0
restricted to S 0 ) and prove that µti (S 0 ) ≥ µti (S). This will complete the proof. Since vi is additive,
an averaging argument gives us
12
0
µti (S 0 ) ≤
1
1 X
1
0
v
(S
)
=
vi (Aj ) ≤ 0 2t0 vi (Ai )
i
0
0
t
t
t
0
j∈J
Here, the last inequality uses the bound 2vi (Ai ) ≥ vi (Aj ) for all j ∈ J 0 . Therefore, vi (Ai ) ≥
1 t0
0
t0
0
t
2 µi (S ). To complete the proof, we need to show that µi (S ) ≥ µi (S). Note that—while considering the maximin shares of agent i—we can restrict our attention to Si (the set of goods which are
0
0
positively valued by i). In particular, the equalities µti (S 0 ) = µti (S 0 ∩Si ) and µti (S) = µti (S ∩Si ) imply that, without loss of generality, we can work under the assumptions that S ⊆ Si and S 0 ⊆ Si .
Effectively, for all j ∈ J, we can slightly abuse notation and denote Aj ∩ Si by Aj . Now, consider
the allocation M = (M1 , . . . , Mt ) ∈
arg max
min vi (Bj ).
(B1 ,...,Bt )∈Πt (S) j∈[t]
t
µi (S). Also, note that |Aj | ≤ 1
= t − t0 bundles in M with items
We have, minj∈J vi (Mj ) =
for each agent j ∈ K = J \ J 0 .
Therefore, there are at most |K|
from S \ S 0 . We choose t0 bundles
from M = (M1 , . . . , Mt ) which do not contain any item from S \ S 0 . Let us call these new bundles
M0 = (M10 , . . . , Mt00 ). By the definition of M, the value of each such bundle for agent i is greater
than or equal to µti (S), that is, vi (Mj0 ) ≥ µti (S) for all j ∈ [t0 ]. Since we have assumed that all items
have nonnegative values, adding more items from the remaining (t − t0 ) bundles to any of the
0
bundles in M0 can only increase the value of the partitions. Thus, µti (S 0 ) ≥ µti (S), which implies
that vi (Ai ) ≥ 12 µti (S).
Lemmas 1 and 3 directly establish Theorem 1.
We note that (both in terms of the algorithm and EFL’s implication) the approximation guarantee established in Theorem 1 is essentially tight. Appendix A provides an example to establish
this observation.
6
Guaranteed Existence of GMMS Allocations in Restricted Settings
In this section, we prove the existence of GMMS for relevant valuation classes.
Theorem 2. Groupwise maximin share allocations always exist under additive, binary valuations, i.e.,
such allocations exist when the additive valuations satisfy vi (g) ∈ {0, 1} for all agents i ∈ [n] and goods
g ∈ [m].
Proof. To prove the theorem, it suffices to show that under additive, binary valuations, EF1 implies
GMMS. Recall that EF1 allocations are guaranteed to exist. Let A = (A1 , . . . , An ) be an EF1
S
allocation. Fix an agent i and a group of agents J 3 i. Also, write S := j∈J Aj . Next we complete
|J|
the proof by showing that vi (Ai ) ≥ µi (S).
Since the valuations are binary, for any subset of goods B ⊆ [m], the total value vi (B) is a non|J|
negative integer. Therefore, µi (S) must be a non-negative integer. Moreover, since the valuations
|J|
1
are additive, we have |J|
vi (S) ≥ µi (S). In addition, the facts that (a) A is an EF1 allocation and
(b) the maximum value of any good is 1 imply vi (Ai ) ≥ vi (Aj ) − 1 for all j ∈ J. Therefore,
13
vi (Ai ) =
1
1
|J|vi (Ai ) ≥
vi (Ai ) +
|J|
|J|
X
(vi (Aj ) − 1)
j∈J\{i}
1
1
|J| − 1
(vi (S) − |J| + 1) =
vi (S) −
|J|
|J|
|J|
|J| − 1
|J|
|J|
≥ µi (S) −
> µi (S) − 1.
|J|
≥
|J|
|J|
|J|
The last inequality vi (Ai ) − µi (S) > −1 implies vi (Ai ) ≥ µi (S) since vi (Ai ) and µi (S) are
|J|
integers. Overall, the inequality vi (Ai ) ≥ µi (S) shows that the EF1 allocation A is GMMS as
well. This completes the proof.
Note that the following theorem holds for general combinatorial valuations and is not limited
to the additive case.
Theorem 3. If all the n agents in a fair division instance have the same valuation (i.e., vi = vj for all
i, j ∈ [n]), then a groupwise fair allocation is guaranteed to exist.
Proof. Given a fair division instance with m goods and n agents with the same valuation function
v, we will define an order on Πn ([m]), the set of n-partitions of [m]. For vectors u, u0 ∈ Rn , we say
that u lexicographically dominates u0 if u = u0 , or there exists an index a ∈ [n] such that u(a) > u0(a)
and for all b < a we have u(b) = u0(b) ; here, u(k) and u0(k) denotes the kth smallest component
of u and u0 , respectively. Extending this definition, an allocation (A1 , . . . , An ) ∈ Πn ([m]) is said
to lexicographically dominate another allocation (B1 , . . . , Bn ) ∈ Πn ([m]) iff the vector of valuations (v(A1 ), . . . , v(An )) lexicographically dominates (v(B1 ), . . . , v(Bn )). Note that lexicographic
domination defines a total order over the equivalence classes of n-partitions. In addition, up to
permutations of the valuation vector, there exists a unique maximal allocation with respect to this
order, i.e., there exists an allocation A∗ = (A∗1 , . . . , A∗n ) which lexicographically dominates all other
allocations.
We will show that A∗ is GMMS. For contradiction, say that this is not the case, and the GMMS
|J|
condition is violated for subset J ⊆ [n] and agent i ∈ J, i.e., vi (A∗i ) < µi (S); here S := ∪j∈J A∗j .
Since the valuations are identical, the maximin share of each agent in J restricted to S is the same,
|J|
|J|
i.e., µi (S) = µk (S), for all k ∈ J. Hence, GMMS condition must also be violated for agent
arg minj∈J v(A∗j ). This observation and the definition of maximin shares imply that there exists a
|J|-partition of S, say (M1 , . . . , M|J| ), such that min1≤a≤|J| v(Ma ) > minj∈J v(A∗j ).
Now, consider an allocation B = (B1 , . . . , Bn ) obtained from A∗ by (i) replacing the |J| bundles
A∗j s (for j ∈ J) with Ma s (for 1 ≤ a ≤ |J|), and (ii) setting Bk = A∗k for all k ∈
/ J. Note that,
∗
even under any permutation of the bundles, the allocation B 6= A . Also, the construction of B
ensures that it lexicographically dominates A∗ . This contradicts the maximality (under the defined
lexicographic order) of A∗ and, hence, the stated claim follows.
14
7
Some Empirical Results
For an experimental investigation, we generated 1000 random instances, for several combinations
of n agents and m goods (n ranging from 3 to 5, and m from 3 to 11), by randomly drawing
the valuations from two different distributions (a) gaussian distribution with mean 0.5, standard
deviation 0.1 (truncated at 0 to ensure nonnegative valuations) and (b) uniform distribution [0, 1].
These are the same set of experiments that [Bouveret and Lemaı̂tre, 2014] carried out for studying
the existence of various fairness notions under additive valuations.
In addition, we considered the modification of these instances wherein all the agents agree on
the same ranking of the goods, but may have different valuations for the same item. Such instances
are said to have same-order preferences (SOP), and were studied by [Bouveret and Lemaı̂tre,
2014]. They showed that, when it comes to (empirically) finding an MMS allocation, SOP instances
are the hardest. We observe similar results for GMMS: finding a GMMS allocation, done using
brute-force search, requires noticeably more time in SOP instances, than in non-SOP instances.
Our empirical results and observations are summarized below:
1. GMMS allocations exist empirically in all randomly generated instances (which is similar to
the experimental results for MMS [Bouveret and Lemaı̂tre, 2014]). These simulation results
indicate that we may not fall short on such generic existence results by strengthening the
maximin solution concept.
2. Allocations generated by the EFL algorithm (Algorithm 1) on the random instances achieve
an approximation factor of 0.98 (on an average) which is better than our theoretically obtained worst-case bound of 0.5. Figure (1) tabulates the approximation factor for GMMS
using the EFL algorithm, averaged over 1000 instances. The approximation factor (on average) is lower for SOP than for non-SOP instances. Moreover, we observe that, for randomly
generated instances, the approximation factor is better when the number of items is a multiple of the number of agents. This may be because of the round robin nature of the algorithm
which gives an agent her most desirable good at each round.
3. As expected, our proposed EFL algorithm ran much faster than the (exponential-time) algorithm for finding exact GMMS. Our algorithm was about 107 times faster than the exact
GMMS computation on a machine with a quad Intel Core i7 processor and 32 GB RAM.
8
Conclusion and Future Work
In this paper, we introduced the concept of GMMS to address fair allocation of indivisible goods,
thereby strengthening the well-studied notions of MMS and PMMS. We established the existence
of 1/2-GMMS under additive valuations, and developed an efficient algorithm for finding such
allocations. We also proved that under specific settings exact GMMS allocations always exist. In
addition, GMMS allocations were always found in our experiments (over a large number of ran-
15
Figure 1: EFL algorithm results on approximate GMMS.
domly generated instances). This indicates why it seems nontrivial to obtain simpler7 examples
which refute the guaranteed existence of GMMS allocations. Finding an instance which admits an
MMS allocation but not a GMMS allocation remains an interesting, open problem.
Our work addresses key questions about groupwise fairness when the valuations are additive, and it motivates the study of this notion in more general settings, e.g., under submodular
valuations, or under additional constraints, such as the ones considered by [Bouveret et al., 2017].
Another interesting direction of future work is to understand groupwise fair division of indivisible
goods among strategic agents.
7
The intricate examples showing the nonexistence of MMS under additive valuations [Procaccia and Wang, 2014,
Kurokawa et al., 2016] also serve as counterexamples for GMMS.
16
Acknowledgments.
Siddharth Barman was supported by a Ramanujan Fellowship (SERB - SB/S2/RJN-128/2015).
Arpita Biswas gratefully acknowledges the support of a Google PhD Fellowship Award.
References
[Amanatidis et al., 2015] Amanatidis, G., Markakis, E., Nikzad, A., and Saberi, A. (2015). Approximation algorithms for computing maximin share allocations. In International Colloquium
on Automata, Languages, and Programming, pages 39–51. Springer.
[Aziz et al., 2017] Aziz, H., Rauchecker, G., Schryen, G., and Walsh, T. (2017). Approximation
algorithms for max-min share allocations of indivisible chores and goods. In Proceedings of the
Thirty-First AAAI Conference on Artificial Intelligence, pages 335–341.
[Barman and Krishnamurthy, 2017] Barman, S. and Krishnamurthy, S. K. (2017). Approximation
algorithms for maximin fair division. In Proceedings of the 2017 ACM Conference on Economics
and Computation, pages 647–664.
[Bouveret et al., 2017] Bouveret, S., Cechlaova, K., Elkind, E., Igarashi, A., and Peters, D. (2017).
Fair division of a graph. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, pages 135–141.
[Bouveret and Lemaı̂tre, 2014] Bouveret, S. and Lemaı̂tre, M. (2014). Characterizing conflicts in
fair division of indivisible goods using a scale of criteria. In Proceedings of International Conference
on Autonomous Agents and Multi-Agent Systems, pages 1321–1328.
[Brams and King, 2005] Brams, S. J. and King, D. L. (2005). Efficient fair division: Help the worst
off or avoid envy? Rationality and Society, 17(4):387–421.
[Brandt et al., 2016] Brandt, F., Conitzer, V., Endriss, U., Procaccia, A. D., and Lang, J. (2016). Handbook of Computational Social Choice. Cambridge University Press.
[Budish, 2011] Budish, E. (2011). The combinatorial assignment problem: Approximate competitive equilibrium from equal incomes. Journal of Political Economy, 119(6):1061–1103.
[Caragiannis et al., 2016] Caragiannis, I., Kurokawa, D., Moulin, H., Procaccia, A. D., Shah, N.,
and Wang, J. (2016). The unreasonable fairness of maximum nash welfare. In Proceedings of the
2016 ACM Conference on Economics and Computation, pages 305–322. ACM.
[Foley, 1967] Foley, D. (1967). Resource allocation in the public sector. Yale Economic Essays, 7:73–
76.
[Ghodsi et al., 2017] Ghodsi, M., HajiAghayi, M., Seddighin, M., Seddighin, S., and Yami, H.
(2017). Fair allocation of indivisible goods: Improvement and generalization. arXiv preprint
arXiv:1704.00222.
17
[Goldman and Procaccia, 2015] Goldman, J. and Procaccia, A. D. (2015). Spliddit: Unleashing fair
division algorithms. ACM SIGecom Exchanges, 13(2):41–46.
[Kurokawa et al., 2016] Kurokawa, D., Procaccia, A. D., and Wang, J. (2016). When can the maximin share guarantee be guaranteed? In Proceedings of the Thirtieth AAAI Conference on Artificial
Intelligence, volume 16, pages 523–529.
[Lipton et al., 2004] Lipton, R. J., Markakis, E., Mossel, E., and Saberi, A. (2004). On approximately fair allocations of indivisible goods. In Proceedings of the 5th ACM conference on Electronic
commerce, pages 125–131. ACM.
[Procaccia and Wang, 2014] Procaccia, A. D. and Wang, J. (2014). Fair enough: Guaranteeing approximate maximin shares. In Proceedings of the Fifteenth ACM Conference on Economics and Computation, pages 675–692. ACM.
[Steinhaus, 1948] Steinhaus, H. (1948). The problem of fair division. Econometrica, 16:101–104.
[Stromquist, 1980] Stromquist, W. (1980). How to cut a cake fairly. The American Mathematical
Monthly, 87(8):640–644.
[Varian, 1974] Varian, H. R. (1974). Equity, envy, and efficiency. Journal of economic theory, 9(1):63–
91.
Appendix
A
Tight Example for Algorithm 1
This section shows that the approximation guarantee established in Theorem 1 is essentially tight.
Specifically, consider the following fair division instance with n agents and (3n − 2) goods. The
additive valuation of agent 1 is as follows:
• For each of the first n goods—say, {l1 , l2 , . . . , ln }—the valuation of agent 1 is equal to one.
• The next (n − 1) goods, {d2 , d3 , . . . , dn }, are valued at (n − 1)/n by agent 1.
• Finally, the remaining (n − 1) goods, {s2 , s3 , . . . , sn }, are valued at 1/n.
The remaining (n − 1) agents value l1 at 2, and they value {s2 , . . . , sn } at 1. In addition, goods
{d2 , d3 , . . . , dn } along with {l2 , . . . , ln } have a value of 1/2 − 1/n for these (n − 1) agents.
Consider allocation A = (A1 , ..., An ) wherein the first agent gets bundle A1 = {l1 }, and the
remaining agents obtain Ak = {lk , dk , sk } for all k > 1. This allocation is EFL with respect to all
the agents. But, the maximin share of agent 1 is close to 2; specifically, in the following n-partition
the value of each bundle for agent 1 is equal to 2 − 1/n: {l1 , s2 , . . . , sn }, {l2 , d2 }, . . . , {ln , dn }.
Hence, as n increases, the approximation guarantee of this EFL allocation for agent 1 approaches 1/2.
18
Furthermore, we can ensure that the proposed algorithm finds this EFL allocation. Specifically,
during the algorithm’s execution, say, agent 1 picks l1 at the beginning. The other agents now
pick in the remaining iterations. Since the other agents value the goods sj s the most among the
unallocated goods, agents 2 to agent n will select s2 to sn during the initial iterations. At this point
of time, the partial allocation is {l1 }, {s2 }, . . . , {sn }.
Agent 1 is still envied the most. Now, agents 2 to n pick d2 to dn one after the other (the ties are
broken in favor of the target allocation). The intermediate allocation obtained by the algorithm
is {l1 }, {s2 , d2 }, . . . , {sn , dn }. Still, Agent 1 is the most envied one and she does not envy others.
Goods l2 to ln remain unallocated. Agents 2 to n finally select them, one after the other, to give us
the required EFL allocation.
19
| 2cs.AI
|
Development of the SP machine
Vasile Palade∗
J Gerard Wolff†
arXiv:1707.00614v1 [cs.AI] 28 Jun 2017
July 4, 2017
Abstract
This paper describes the main things that need to be done to develop the SP machine, based on the SP theory of intelligence and
its realisation in the SP computer model. The SP machine may be
developed initially as a software virtual machine with high levels of
parallel processing, hosted on a high-performance computer, or driven
by high-parallel search processes in any of the leading search engines.
An easy-to-use user interface is needed with facilities for visualisation of knowledge structures and processing. The system needs to be
generalised to work with patterns in two dimensions, including the
display of multiple alignments with 2D patterns. Research is needed
into how the system may discover low-level features in speech and
visual images. Existing strengths of the SP system in the processing of natural language may be developed towards the understanding
of natural language and its production from meanings. This may be
done most effectively in conjunction with the development of existing strengths of the SP system in unsupervised learning. Existing
strengths of the SP system in pattern recognition may be developed
for computer vision. Further work is needed on the representation
of numbers and the performance of arithmetic processes in the SP
system. A computer model is needed of SP-neural drawing on the existing conceptual model. When the SP machine is relatively mature,
new hardware may be developed to exploit opportunities to increase
the efficiency of computations. There is potential for the SP machine
∗
Dr Vasile Palade, MEng (Tech. Univ. of Bucharest), PhD (University of Galaţi), IEEE
Senior Member; Reader in Pervasive Computing, Department of Computing, Coventry
University, UK; [email protected]; +44 (0) 2477 659190; +44 (0) 7912 043982;
Skype: v palade; Web: bit.ly/1xNM5xr, bit.ly/1Buyk9E.
†
Dr Gerry Wolff, BA (Cantab), PhD (Wales), CEng, MIEEE, MBCS; CognitionResearch.org, Menai Bridge, UK; [email protected]; +44 (0) 1248 712962; +44 (0)
7746 290775; Skype: gerry.wolff; Web: www.cognitionresearch.org.
1
to be applied on relatively short timescales in such areas as information storage and retrieval, with intelligence, software engineering with
or without automation or semi-automation, and information compression.
Keywords: information compression, artificial intelligence, natural language processing, pattern recognition, computer vision, neuroscience.
1
Introduction
The SP theory of intelligence and its realisation in the SP computer model is a
unique attempt to simplify and integrate observations and concepts across artificial intelligence, mainstream computing, mathematics, and human learning, perception and cognition, with information compression as a unifying
theme. It is described in outline in Appendix A with pointers to where fuller
information may be found.
There are many potential benefits and applications of the SP system,
described in several peer-reviewed papers, referenced in Appendix A.6.
The SP system has several distinctive features and advantages compared
with AI-related alternatives [27]. Probably the most distinctive feature is a
concept of multiple alignment, borrowed and adapted from bioinformatics. A
key advantage of the multiple alignment framework as it has been developed
in the SP system is its versatility in the representation of knowledge and in
aspects of intelligence, and its potential for the seamless integration of diverse
kinds of knowledge and diverse aspects of intelligence, in any combination
(Appendix A.5).
To distinguish multiple alignment in bioinformatics from multiple alignment as it has been developed in the SP system, the latter will be referred
to as SP-multiple-alignment.
1.1
The SP machine
To realise the potential benefits and applications of the SP system, it is envisaged that the SP computer model will provide the basis for the development
of an industrial-strength SP machine, as shown schematically in Figure 1.
The main purpose of this paper is to describe what appear to be the main
things to be done to develop the full potential of the SP machine.
Initially, it will be developed as a high-parallel software virtual machine,
hosted on an existing high-performance computer, and with an easy-to-use
user interface. This may have a dual function: 1) to serve directly in practical
2
SP theory and SP computer model
High parallel
In the cloud
SP MACHINE
Open source
Good user interface
Natural language processing
Representation of knowledge
Several kinds of reasoning
Planning & problem solving
Unsupervised learning
Information compression
Pattern recognition
Information retrieval
MANY APPLICATIONS
Figure 1: A schematic view of how the SP machine may develop from the SP
theory and the SP computer model. Adapted, with permission, from Figure
6 in [23].
applications; and 2) to serve as a vehicle for further research, to explore what
can be done with the SP machine and to create new versions of it.
At some stage, there is likely to be a case for the development of new
hardware, dedicated to structures and processes in the SP system, and with
potential for substantial gains in efficiency and performance ([23, Section IX],
[22, Section III]).
1.2
Presentation
After a summary of our research strategy (Section 2), several sections describe things that need to be done in the development of the SP machine,
including a description of problems to be solved, and potential solutions to
those problems. Some areas of application may yield benefits on relatively
short timescales (Section 14).
Four of the sections are about pieces of “unfinished business” in the development of the SP machine, outlined in [20, Section 3.3]: processing information in two or more dimensions (Section 6); recognition of perceptual features
in speech and visual images (Section 7); unsupervised learning (Section 9);
and processing numbers (Section 11).
3
1.3
An invitation
In accordance with the saying that “A camel is a horse designed by a committee”, the development of a scientific theory, like writing a novel or composing
a symphony, is a task that is not easily shared. Now that the SP theory is relatively mature, it should become easier progressively for other researchers to
work on the several avenues that now need to be explored, without treading
on each others’ toes.
We would very much welcome the involvement of other researchers and
would be happy to discuss possibilities, and the necessary facilities, including
software.
2
Research strategy
Because of its importance, we say again that this programme of research is a
unique attempt to simplify and integrate observations and concepts across artificial intelligence, mainstream computing, mathematics, and human learning, perception and cognition, with information compression as a unifying
theme.
This long-term objective is now showing potential to produce dividends
in spades:
• The strengths of SP-multiple-alignment in the simplification and integration of observations and concepts across a broad canvass is a major
discovery with clear potential in the long-sought-after quest for general,
human-level artificial intelligence, and for developing an understanding
of human intelligence.
• As described in Appendix A.6, discovering the strengths of the SPmultiple-alignment concept has been a bit like hitting the jackpot on a
fruit machine, releasing a cascade of potential benefits and applications.
2.1
Maintain a strong focus on simplicity and power
The long view of this research contrasts with, for example, most research on
deep learning in artificial neural networks where undoubted successes obscure
several weaknesses in the framework (Appendix A.5), suggesting that deep
learning is likely to be a blind alley in any search for general human-level
artificial intelligence.
For these kinds of reasons, we urge researchers working with the SP concepts to maintain a strong focus on maintaining a favourable combination
4
of conceptual simplicity with descriptive or explanatory power, resisting the
temptation to take short cuts or to introduce ideas that do not integrate
well with the SP framework. That said, we believe the following kinds of
development, which are not mutually exclusive, would be reasonable:
• There may be opportunities to develop SP capabilities with industrial
strength for practical applications, without corrupting the core concepts (Section 14). An example is the potential of the SP machine to
serve as a database management system with versatility and intelligence [19].
• There may be a case on occasion to develop a hybrid system as a stopgap to facilitate practical applications, pending the development of a
solution that is theoretically clean. An example in computer vision
would be the use of existing procedures for the detection of low-level
features in images (such as lines and corners) pending an analysis in
terms of SP concepts (Section 7).
2.2
Scheduling
As we shall see, some things are probably best done before others. These
‘ground-clearing’ tasks are described early in what follows. When those tasks
have been completed, it should become easier to develop two or more aspects
of the SP machine in parallel, in tasks described in later subsections.
When the SP system has been brought close to the point where it may
prosper in a commercial environment, perhaps via applications that may be
developed on relatively short timescales (Section 14), companies may wish
to refine the system for one or more markets.
3
The creation of a high-parallel SP machine
The proposed developments will begin with SP71, the latest version of the
SP computer model, described in outline in Appendix A.3. Instructions for
obtaining the source code are in Appendix B.
To get things going, we believe that the first stage, as outlined in this
main section, will be to develop the SP machine with parallel processing,
and then, in the two following sections, to create a user interface that is
easy to use and to facilitate the visualisation of knowledge structures and
the processing of knowledge.
5
3.1
Is parallel processing really needed?
Before we apply parallel processing in the SP machine, we need to establish
that parallel processing is really necessary.
A substantial advantage of the SP system, compared with, for example,
deep learning in artificial neural networks [11], is that, on the strength of
evidence to date, the SP machine is likely to be very much less demanding of
computational resources than deep learning systems [27, Section V-E]. And
there appears to be potential to speed up processing by the judicious use of
indexing, as outlined in Section 3.4.
More generally, there appears to be potential in the SP system for very
substantial gains in computational efficiency compared with conventional
computers: by the use of a technique called “model-based coding” for the
efficient transmission of data [23, Section VIII]; and by exploiting the statistical information which is gathered as a by-product of how the system works
[23, Section IX].
So it is reasonable to ask whether the existing computer model (SP71)
might be speedy enough as it is, without the need for parallel processing?
Here are some answers to that question which are not mutually exclusive:
• Although, as mentioned above, the SP machine has the potential to be
relatively frugal in its computational demands, its full potential in that
respect will probably not be fully realised until new hardware has been
developed to take full advantage of opportunities in the structure and
workings of the SP system.
• The afore-mentioned gains in efficiency in the transmission of information via model-based coding, will not be fully realised until unsupervised learning in the SP system is more fully developed.
• Although, with AI problems, the SP system is likely to be much more
efficient than deep learning systems, AI problems are normally quite
computationally demanding compared with, for example, school-level
mathematics. For those kinds of reasons, parallel processing is likely
to be needed to achieve acceptable run times when the SP machine
is applied to what is normally the relatively demanding task of learning in unsupervised mode, and perhaps also to such things as pattern
recognition and the processing of natural languages.
• When the SP machine has been generalised to work with two-dimensional
patterns, it is likely that parallel processing will be needed to speed up
6
processing in what is likely to be the computationally-intensive process of finding good full or partial matches between two 2D patterns
(Section 6.1).
• Parallel processing will almost certainly be needed if the SP machine
is to become an industrial-strength database management system with
intelligence, as described in [19].
3.2
How to apply parallel processing
The effective application of parallel processing will require careful analysis of
the SP71 computer model (Appendix A.3) to identify parts of the program
where parallel processing may be applied.
At this stage, it seems likely that parallel processing may be applied most
effectively via a MapReduce technique,1 as, for example, in these two parts
of the program:
• With the building of SP-multiple-alignments (Appendix A.3.2), a recurring operation is searching for good full or partial matches between
one “driving” pattern and each of a set of “target” patterns (Appendix
A.3.1), where the set of target patterns may become very large.
Here, the “map” phase would be applying the search in parallel to each
pairing of a driving pattern with a target pattern, while the “reduce”
phase would be identifying the most successful matches and selecting
a subset for further processing.
• With unsupervised learning, each pattern in a relatively large set of
New patterns is processed via a search for SP-multiple-alignments that
allow the given New pattern to be encoded economically.
Here, the “map” phase would be the application of the search process to all the New patterns in parallel, and the “reduce” phase would
be a process of sifting and sorting to create one or more alternative
grammars for the set of New patterns.
No doubt, there will be other opportunities for applying the MapReduce
technique in the SP machine. There may also be opportunities for the use of
other techniques such as pipelining.2
1
2
See “MapReduce”, Wikipedia, bit.ly/2oG42kY, retrieved 2017-06-15.
See “Pipeline (computing)”, Wikipedia, bit.ly/2sq4fuf, retrieved 2017-06-21.
7
3.3
Computing environments for the application of parallel processing
In broad terms in this development, there seem to be two main kinds of
computing environment that may be used for the application of parallel processing:
• A stand-alone computer cluster. A server with one or more multicore chips and appropriate software would probably work well. Compared with most supercomputing facilities, this kind of computer cluster would give researchers unfettered access to develop programs and
to test them frequently, without restrictions.
• High-parallel search mechanisms in a search engine. An interesting
alternative would be to replace the low-level search mechanisms in the
SP machine with high-parallel search mechanisms in an existing search
engine. This would supercharge the SP machine with the power of the
given search engine, adding intelligence to it. Initially, the intelligence
would comprise the capabilities of the SP71 computer model, but there
would be clear potential for many enhancements across most aspects
of intelligence.
With regard to the second of these options, the SP machine, hosted on any
given search engine, may provide human-like intelligence with access to all the
knowledge that has been indexed in the underlying search engine. Looking
further ahead, all search engines may be adapted in this way, and some of
the processing may be offloaded to PCs and other computers around the
world, in the manner of volunteer computing projects such as SETI@home
and Folding@home.
With developments like that, the internet would, in effect, become a giant
brain with potential for human-like intelligence.
Many benefits may flow from artificial intelligence like that, in the internet
and beyond. But there are also potential dangers and other problems which
will need careful scrutiny and much debate.
3.4
Indexing as a means of speeding up processing
A prominent feature of the SP system, both in the way it does unsupervised
learning and the way it builds SP-multiple-alignments, is a process of searching for symbols that match each other, which is normally part of a process
of searching for full or partial matches between patterns.
8
Since these things are often done repeatedly with the same sets of New
and Old patterns, there is potential for substantial gains in computational
efficiency by recording matches as they are found and using that information to by-pass searching in later processing. The record of matches that
have been found is, if effect, and index into the data. This may be done
with matches between individual symbols, and may perhaps also be done for
matches between patterns.
A similar idea may be applied in SP-neural, as outlined in Section 12.2.
4
Development of an easy-to-use user interface
The user interface for the SP71 computer model is the product of ad hoc developments in response to evolving needs, without any attempt at a coherent
and streamlined design.
This has not mattered for the one user (JGW) who originated each feature
and has learned the workings of the user interface as it evolved. However, as
the SP system is opened up to new users, a more intuitive and easy-to-use
user interface is needed, taking advantage of advanced graphical techniques
that are available now.
Since all users, both local and remote, may access the SP machine via a
web browser, and to eliminate the need for users to download any software,
it is envisaged that the new user interface will be developed using HTML5
or any of its successors.
5
Development of visualisation of structures
and processes
An important concern with artificial intelligence is that it should be possible
to inspect and understand how an AI system stores its knowledge and how
it reaches its conclusions. This is important where an AI system may have
an impact on the safety of people or where its performance and conclusions
may have a bearing on court cases.
A major strength of the SP system is that all its knowledge is open to
inspection and it provides an audit trail for all its processing [23, Section XI].
Although this information is provided by the system, there is probably
scope for improving its presentation, so that structures and processes can be
seen easily and so that users can find easily what they are looking for.
9
6
Processing information in two or more dimensions
It has been recognised for some time that the SP machine should be generalised to work with information in two dimensions and probably more [18,
Section 13.2.1]. How this may be done is discussed in the following subsections.
6.1
Introduction of two-dimensional patterns
The main motive for introducing 2D patterns into the SP system is to facilitate the representation and processing of such things as photographs, paintings, drawings, and diagrams. They may also serve in the representation
and processing of procedures that operate in parallel [22, Sections V-G, V-H,
and V-I, and Appendix C]. And, as we shall see in the next subsection, 2D
patterns can play a useful part in the representation and processing of 3D
structures.
It is a simple enough matter to add 2D patterns to the SP system.
What is more challenging is how to generalise such things as the building
of SP-multiple-alignments with 2D patterns, how to represent SP-multiplealignments of 2D patterns on a 2D computer screen or a printed page, and
unsupervised learning with 2D patterns.
We can get some clues from how digital photographs that overlap each
other may be stitched together to create a panorama. However, it seems unlikely that software that is designed for the stitching together of overlapping
pictures will have the flexibility needed to achieve the 2D equivalent of the
processes in SP71 that can find good full and partial matches between two
patterns, including situations where there are substantial differences between
the patterns (Appendix A.3.1).
Generalising the pattern-matching processes in SP71 to work well with
2D patterns is likely to be a significant challenge. It is likely also that any
reasonably good solution will absorb quite a lot of computing power. To
achieve useful speeds in processing, it is likely that high levels of parallel
processing will be required (Section 3).
6.2
Modelling structures in three dimensions
At first sight, the obvious way to model 3D structures in the SP system is to
add 3D patterns to the system. But that is probably a step too far, partly
because of the complexity of matching patterns in three-dimensions and the
10
difficulty of representing SP-multiple-alignments of 3D patterns, and partly
for reasons described here:
“Patterns in one or two dimensions may serve well enough to
represent concepts with three or more dimensions. There is a long
tradition of representing buildings and other engineering artifacts
with two-dimensional plans and elevations.” [18, Section 13.2.2].
and
“Although the human brain is clearly a three-dimensional structure, there is a distinct tendency for it to be organised into layers of cells such as those found in the cortex and in the lateral
geniculate body. Although these are often folded into three dimensions, they are topologically equivalent to two-dimensional
sheets. ... these sheets of neurons would provide a natural home
for SP patterns.” ibid..
The case for using 2D patterns as a means of building and representing 3D structures is also strengthened by the existence of applications
that work like that, such as “Big Object Base” bit.ly/1gwuIfa, “Camera
3D” bit.ly/1iSEqZu, and “PhotoModeler” bit.ly/MDj70X, and also Google’s
“Streetview”,3 which builds 3D models of street plans from 2D photographs.
There is relevant discussion in [21, Sections 6.1 and 6.2].
6.3
Modelling structures in four or more dimensions
Although there are four dimensions in our everyday experience—three dimensions of space and one of time—it seems likely that, normally, we don’t
attempt to integrate them in the manner of Einstein’s space-time but concentrate on a subset of the four dimensions, such as the succession 2D images on
our retinas as we walk through a building or drive around a town. It seems
likely that, for most applications, that kind of partial integration would work
well enough.
If for any reason there is a need to represent structures in five or more
dimensions, it is probably best to borrow techniques used by mathematicians
for dealing with such structures.
3
See “Google Street View”, Wikipedia, bit.ly/1TYQmLR, retrieved 2017-06-15.
11
7
Recognition of low-level perceptual features in speech and images
As noted in Appendix A.2, the SP system is designed to represent and process stored knowledge and sensory data expressed as atomic symbols in SP
patterns. So at first sight, it should be able work with any digital information, meaning information that is founded on the atomic symbols ‘0’ and ‘1’.
But there are two main complications:
• Much digital information drawn relatively directly from the environment is encoded as, for example, hexadecimal numbers representing
such things as levels of brightness in an image. Until the SP system
can deal effectively with numbers (Section 11), it will fail to interpret
such information in an appropriate manner.
• In principle, all such encoded information may be translated into a more
“primitive” kind of representation in which, for example, the brightness
in any given portion of an image would be represented by the density of
‘0’s or ‘1’s, much as in old-style black and white newspaper photograph
[18, Section 2.2.3]. But it is not clear at present how the SP system
would work in processing information of that kind.
In general, the SP system works best with symbols representing structures
or entities that are significant in human perception. Thus:
“For the SP system to be effective in the processing of speech or
visual images, it seems likely that some kind of preliminary processing will be required to identify low-level perceptual features,
such as, in the case of speech, phonemes, formant ratios or formant transitions, or, in the case of visual images, edges, angles,
colours, luminances or textures.” [20, Section 3.3].
In broad terms, there are two ways in which the developing SP machine
may bridge the gap which is likely to exist between information that is drawn
relatively directly from the environment and the kinds of structures just
mentioned:
• As a stop-gap measure, preliminary processing of sensory data may
done by conventional methods to yield the kinds of low-level perceptual
entities with which the SP system can operate.
12
• Given the power and generality of the principles that lie a the heart of
the SP theory, it seems likely that, pending some further research, the
SP system, without stop-gap measures, will be able to process all kinds
of information in a meaningful way, including “raw” sensory data and
data that has been encoded using numbers.
Perhaps we should worry less about maintaining conceptual purity. For
example:
“The issue of representation lies at the heart of the debate between the logic-inspired and the neural-network-inspired
paradigms for cognition. In the logic-inspired paradigm, an instance of a symbol is something for which the only property is
that it is either identical or non-identical to other symbol instances. It has no internal structure that is relevant to its use;
and to reason with symbols, they must be bound to the variables
in judiciously chosen rules of inference. By contrast, neural networks just use big activity vectors, big weight matrices and scalar
non-linearities to perform the type of fast intuitive inference that
underpins effortless commonsense reasoning.” [5, p. 441]
Although the SP system is very different from most logic-inspired systems
(except for all-or-nothing matching amongst SP symbols),4 it is, at its current stage of development, somewhat choosy about its inputs compared with
many artificial neural networks. Although this is, superficially, an advantage
for artificial neural networks, we believe that the SP system will be more
satisfactory in the long run, for reasons given in Appendix A.5.
Given the generality of information compression and the alreadydemonstrated versatility of SP-multiple-alignment as a means of achieving information compression, we believe it is likely that the SP-multiple-alignment
framework will accommodate low-level features in speech and images. Although there may be a case for adopting stop-gap solutions in some applications (Section 2), there is also a need to explore how the SP system may be
applied in these areas.
4
Contrary to what is said in the quotation, it is often the case that a symbol in
logic-inspired systems, such as a variable in a system like Prolog, is not “something for
which the only property is that it is either identical or non-identical to other symbol
instances”.
13
8
The processing of natural language
A major strength of the SP-multiple-alignment framework is that it lends
itself well to the representation and processing of natural language:
• As can be seen in Figure 7 in Appendix A.2, and in many other examples in [18, 20], the building of SP-multiple-alignments can achieve the
effect of parsing a sentence into its parts and sub-parts.
• The SP-multiple-alignment concept exhibits versatility in the representation and processing of several kinds of non-syntactic knowledge
(Appendix A.5). Any of these may represent the semantics or meanings of a sentence.
• As noted in Appendix A.5, the fact that one relatively simple framework may serve in the representation and processing of diverse kinds of
knowledge means that there is potential for their seamless integration
in any combination. More specifically in the present context, there is
clear potential for the seamless integration of the syntax and semantics
of natural language.
• Preliminary examples described in [18, Section 5.7] show how this may
work in both the semantic interpretation of a simple sentence ([18,
Figure 5.18]) and in the production of a sentence from its meanings
([18, Figure 5.19]).
• The potential of the SP system for the seamless integration of diverse
kinds of knowledge and diverse aspects of intelligence (Appendix A.5)
means that there is clear potential for the integration of natural language with other aspects of intelligence such as pattern recognition,
reasoning, planning, and problem solving.
In order to demonstrate parsing via SP-multiple-alignment, the semantic
interpretation of a sentence, or the production of a sentence from meanings,
it has been necessary to construct appropriate grammars by assembling collections of SP patterns manually. This has been acceptable in the process of
developing the SP-multiple-alignment framework and demonstrating what it
can do, but it is likely to be far too slow and prone to error in the further
development of natural language processing in the SP system.
To make progress with the processing of natural language in the SP system, it will be necessary to develop this area in conjunction with the development of unsupervised learning in the SP system, as described in the next
section.
14
9
Unsupervised learning
In the light of observations in the previous section, the discussion of unsupervised learning in this section concentrates mainly on unsupervised learning
of the syntax and semantics of natural language. Success in this area is likely
to generalise to other areas and, in any case, “semantics” is a broad concept
that embraces most kinds of knowledge.
It is pertinent to mention that thinking in this area draws on an earlier
programme of research (summarised in [16]) developing computer models of
how a child learns his or her first language or languages. The emphasis on
hierarchical structures in those models proved entirely unsuitable for the SP
programme of research, so that it was necessary to develop an entirely new
conceptual framework—SP-multiple-alignment—to meet the goals of the SP
research.
Our emphasis on unsupervised learning in the SP machine is because
of the belief that most human learning is achieved without the benefit of
any kind of “teacher” or anything equivalent (cf. [2]), and because of our
belief that other kinds of learning, such as supervised learning, reinforcement
learning, learning by being told, and learning by imitation, may be seen as
special cases of unsupervised learning [22, Sections V-A.1 and V-A.2].
Probably the best place to start in developing unsupervised learning in
the SP computer model is with two weaknesses in the SP71 model, outlined
here:
“A limitation of the SP computer model as it is now is that it
cannot learn intermediate levels of abstraction in grammars (eg,
phrases and clauses), and it cannot learn the kinds of discontinuous dependencies in natural language syntax that are described
in [[20, Sections 8.1 to 8.3] and [18, Section 5.4]]. I believe these
problems are soluble and that solving them will greatly enhance
the capabilities of the system for the unsupervised learning of
structure in data ....” [20, Section 3.3].
Pending further analysis, it looks as if the problems with SP71 in discovering structures such as phrases and clauses in the syntax of natural language,
and in discovering discontinuous dependencies in that syntax, may both be
solved in some such manner as what follows.
Since the building of SP-multiple-alignments is an integral part of unsupervised learning in the SP system (Appendix A.3.3), and since every stage
in the process of building an SP-multiple-alignment yields an encoding of
15
one or more New patterns in terms of the Old patterns in the SP-multiplealignment, better results may be obtained by applying the learning processes
to those encodings as well as to the original patterns.
Probably, the best way to build up an understanding of what needs to be
done is to work with unsegmented samples of relatively simple English-like
languages generated by artificial grammars. If or when good insights and
good results can be obtained with that kind of material, the system may be
tested and refined with more challenging material, as described in the next
three subsections.
9.1
Learning the syntax of natural language
Following the steps outlined above, it seems possible now to bring the SP
system to a stage of development where it can derive plausible grammars for
the syntax of natural language via the application of unsupervised learning
to unsegmented textual samples of natural language. This would be a major
achievement since, to our knowledge, no other system can perform at that
level. Incidentally, for reasons given in [20, Section 6.2], there are likely to
be subtle but important differences between grammars derived from natural
language texts without meanings and the kinds of grammars that we learn
as children, where syntax and semantics are developed in tandem.
It seems likely that success with this task would smooth the path for
success with the tasks described next.
9.2
The unsupervised learning of non-syntactic kinds
of knowledge
A key idea in the SP theory as it is now, which is a working hypothesis in the
SP programme of research, is that one set of principles applies across the representation and processing of diverse kinds of knowledge, where “processing”
includes unsupervised learning [23, Section III].
Hence, we anticipate that if we can find good solutions for the unsupervised learning of syntax, it is likely that, with or without some further
development, they would generalise to the learning of non-syntactic, semantic
structures. With the learning of visual structures, it would be necessary for
there to be a robust solution to the previously-mentioned problem of finding
good full and partial matches between 2D patterns (Section 6.1).
16
9.3
Learning syntactic/semantic structures
A prominent feature of the environment in which most young children learn
their first language or languages is that, very often, they can hear the speech
of adults or older children at the same time as they can see or hear what
people are talking about: “Lunch is on the table”, “Here’s our bus coming”,
and so on.
It appears that this kind of environment is needed to enable young children to work out the meanings of words and grammatical forms, and something similar would be needed for experiments with the SP system. But there
appears to be a problem:
“The logician W. V. O. Quine asks us to imagine a linguist studying a newly discovered tribe. A rabbit scurries by, and a native
shouts, ‘Gavagai!’ What does gavagai mean? Logically speaking, it needn’t be ‘rabbit.’ It could refer to that particular rabbit (Flopsy, for example). It could mean any furry thing, any
mammal, or any member of that species of rabbit (say, Oryctolagus cuniculus), or any member of that variety of that species
(say, chinchilla rabbit). It could mean scurrying rabbit, scurrying
thing, rabbit plus the ground it scurries upon, or scurrying in general. It could mean footprint-maker, or habitat for rabbit-fleas.
It could mean the top half of a rabbit, or rabbit-meat-on-thehoof, or possessor of at least one rabbit’s foot. It could mean
anything that is either a rabbit or a Buick. It could mean collection of undetached rabbit parts, or ‘Lo! Rabbithood again!,’ or
‘It rabbiteth,’ analogous to ‘It raineth.’ ... Somehow a baby must
intuit the correct meaning of a word and avoid the mind-boggling
number of logically impeccable alternatives. It is an example of
a more general problem that Quine calls ‘the scandal of induction,’ which applies to scientists and children alike: how can they
be so successful at observing a finite set of events and making
some correct generalization about all future events of that sort,
rejecting an infinite number of false generalizations that are also
consistent with the original observations?” [8, pp. 147–148].
Without wishing to trivialise this problem, it appears that the SP system
has potential to provide a good solution:
• By searching for recurrent patterns and attaching more weight to patterns that occur frequently and less weight to patterns that are rare,
the SP system naturally homes in on the stronger correlations—such
17
as the correlation between the word “lunch” and what people eat at
lunchtime—and discards weaker correlations—such as the correlation
between the word “lunch” and whether or not it is raining outside.
• The problem of how to generalise from what one has learned, without
either under-generalisation or over-generalisation, is discussed in [20,
Section 5.3]5 and also in [27, Section V-H].
In brief, learning with the SP system from a body of data, I, means compressing I as much as possible and then splitting the resulting body of
data into two parts: a grammar which contains a single copy of each recurrent pattern in I; and an encoding of I which contains non-recurrent
parts of I and brief codes or references recurrent patterns in I.6 Evidence
to date shows that with learning of this kind, the grammar generalises
in a way that we judge to be “correct” and avoids what intuitively we
would judge to be under-generalisation or over-generalisation.7
Assuming that progress can be made along the lines outlined here, another
avenue that may be explored later is how children may learn the meanings of
things by making inferences from their existing knowledge. For example, if a
child that does not know the word “plantain” hears someone say something
like: “I had plantain for lunch today. It was delicious—more tasty than
potato but not overpowering”, he or she may infer quite reasonably that
plantain is something that one can eat and that it is probably something
starchy, like potato.”
In this area of research—how the syntax and semantics of natural language may be learned together—it will be necessary to provide input to the
SP system that is comparable with that described above. Also, it will probably be necessary for the SP system to have been generalised to work with
two-dimensional patterns (Section 6.1), so that syntactic knowledge may be
shown alongside semantic knowledge, with associations between them (see
also discussions in [22, Sections V-G, V-H, and V-I, and Appendix C]).
9.4
Learning language from speech
Although children learn their first language or languages by hearing people
talk, we have left that kind of learning to last because it is more challenging
5
Under-generalisation was overlooked in this account but the avoidance of
under-generalisations is achieved in exactly the same way as the avoidance of
over-generalisations.
6
Notice that recurrent patterns in I may be abstract patterns as well as concrete
patterns.
7
See also the discussion of generalisation in [16]
18
than, for example, learning grammars from text.
The main reason for the difficulty is uncertainties about the discovery of
low-level features in speech, as discussed in Section 7. Solutions in that area
will facilitate the learning of language from speech.
10
Computer vision
The generalisation of the SP machine to work with 2D patterns (Section 6.1)
is a likely prerequisite for the development of the SP system for computer
vision and as a model of human vision, as described in [21].
An important part of this development will be the unsupervised learning
or discovery of objects and classes, as discussed in [21, Section 5] and Section
9.2, below. This will include work on the discovery of low-level features in
images, as discussed in Section 7.
11
Processing numbers
“The SP model works with atomic symbols, such as ASCII characters or strings of characters with no intrinsic meaning. In itself,
the SP system does not recognise the arithmetic meaning of numbers such as “37” or “652” and will not process them correctly.
However, the system has the potential to handle mathematical
concepts if it is supplied with patterns representing Peanos axioms or similar information [18, Chapter 10]. As a stop-gap solution in the SP machine, existing technologies may provide whatever arithmetic processing may be required.” [20, Section 3.3].
In view of evidence that mathematics may be seen as a set of techniques
for compression of information, and their application [29], and bearing in
mind that the SP system works entirely via the compression of information,
and in view of the potential of the SP system as a vehicle for ‘computing’
([18, Chapter 4], [24, Section 6.6], [30]), and as a vehicle for mathematics [18,
Chapter 10], there is clear potential in the SP machine for the performance
of arithmetic and related computations.
Pending the development of capabilities for arithmetic in the SP machine,
it seems reasonable to use an existing “maths coprocessor”,8 or similar facility, as an adjunct to the SP machine, much as is done in many database
management systems today.
8
See “Coprocessor”, Wikipedia, bit.ly/2rYawia, retrieved 2017-06-23.
19
12
SP-neural
At present, the SP-neural part of the SP theory (Appendix A.4) exists only
as a conceptual model, described in some detail in [26].
To flesh out this conceptual model, it is really necessary to create a computational version of it. As with the SP71 computer model, this will help
to guard against vagueness in the theory, it will serve as a means of testing
ideas to see whether or not they work as anticipated, and it will be a means
of demonstrating what the model can do.
Probably, the best place to start would be to create a computer model of
the building of an SP-multiple-alignment in SP-neural, as discussed in [26,
Section 4] and illustrated schematically in Figure 2.
The development of unsupervised learning in the computer model of SPneural is probably best left until later, if or when there is a satisfactory
SP-neural model of the building of the neural equivalent of SP-multiplealignments. This is because the building of SP-multiple-alignments is an
important part of unsupervised learning in the SP system. We make some
remarks about unsupervised learning in Section 12.3.
In the figure, each alphanumeric symbol surrounded by a solid-line envelope represents a neural symbol (Appendix A.4), and each group of neural
symbols surrounded by a broken-line envelope represents a pattern assembly
(Appendix A.4, [26, Section 4.2]).
In the figure, nerve fibre connections between neural symbols are represented with lines which may be solid or broken. The solid lines represent
neural connections in part of the SP-multiple-alignment shown in [26, Figure
2], while the broken lines represent other neural connections that are not
part of that SP-multiple-alignment.
12.1
Inhibition
This section considers the potential role of inhibitory processes in the new
computer model, starting with some background.
12.1.1
Background
In [26, Section 9], there is a brief review of what appears to be a substantial
role for inhibitory signals between neurons in the workings of brains and
nervous systems. A tentative general rule from this research is that When,
in neural processing, two or more signals [that interact] are the same, they
tend to inhibit each other, and when they are different, they don’t [26, p. 19],
with the implication that “The overall effect should be to detect redundancy
20
NP 1 D #D N #N #NP
D 9 o n e #D
N 5 b r a v e #N
D 1 t h e #D
N 9 t a b l e #N
… t .... h .... e .... b .... r .... a .... v .... e …
|
|
|
|
|
|
|
|
t
h
e
b
r
a
v
e
receptor array
sensory data
Figure 2: A schematic representation of part of the process of building an SPmultiple-alignment in SP-neural, as discussed in [26, Section 4]. Reproduced
with permission from Figure 3 in [26].
21
in information and to reduce it, whilst retaining non-redundant information,
in accordance with the central principle in the SP theory—that much of
computing and cognition may, to a large extent, be understood as information
compression.” (ibid.).
A related idea is that the rate of firing of neurons correlates only indirectly
with the strength of stimulation, and that when stimulation is steady, neurons
typically fire at a constant rate which is the same regardless of the strength
of the stimulation. This can be seen in Figure 3 which shows variations in
the rate of firing of a sensory cell (ommatidium) in the eye of the horseshoe
crab (Limulus) in response to variations in the intensity of the light that is
falling on it.
Figure 3: Variation in the rate of firing of a single ommatidium of the eye of
a horseshoe crab in response to changing levels of illumination. Reproduced
from [9, p. 118.], with permission from the Optical Society of America.
On the left of the figure, where the light level is low, the ommatidium fires
at a steady rate of about 20 impulses per second. When a light is switched
on, there is sharp upswing in the rate of firing, but this soon settles back to
the same 20 impulses per second rate as before. When the light is switched
off, there is a sharp downswing in the rate of firing followed by a rapid return
to the previous steady rate.
This pattern of responses can be explained via inhibitory connections
22
which dampen down responses when stimulation is steady [9, 14].9 Similar
effects occur in the spatial dimension and it appears that similar explanations
apply.
12.1.2
Developments
In the building of an SP-multiple-alignment, modelling such things as the
parsing of a sentence or the recognition of a pattern, things are a bit different
from what has been described in the previous subsection. Instead of trying to
detect and remove redundancy between neighbouring elements in a sequential
or spatial pattern, the aim is to detect redundancies between two independent
patterns, such as a New pattern (as, for example, ‘I N F O R M A T I O
N’) and an Old pattern (such as ‘I N F O R M A T I O N’), or between two
partial SP-multiple-alignments at later stages of processing.
Bearing in mind the putative idea that the role of inhibition is to detect
and remove redundancies in information, a possible scheme is shown in Figure
4, with a key to the meanings of symbols in the caption to the figure.
This is how the scheme may work:
• In accordance with what was said about the results of neural processing
shown in Figure 3, it is envisaged that, without incoming signals, the
four neural symbols shown at the top of the figure will fire at a steady
intermediate rate.
• If a signal comes in from the neural symbol marked ‘a’ at the bottom
of the figure, this will have an inhibitory effect on the neural symbol
marked ‘a’ at the top of the figure. Likewise for the other two neural
pairings, ‘b’ with ‘b’, and ‘c’ with ‘c’.
• If all three of ‘a’, ‘b’, and ‘c’, in the pattern assembly have been inhibited in this way, there will be reduced inhibitory signals transmitted to
the “identifier” (ID) neural symbol marked ‘X’ in the figure.
• Since the ID neural symbol is receiving reduced inhibitory signals, its
rate of firing will increase. This may be interpreted as recognition of the
pattern ‘a b c’. That increase in the rate of firing may be transmitted
to higher-level pattern assemblies that contain ‘X’ as a “reference” to
the pattern ‘a b c’. This may lead to similar processes of recognition
at higher levels.
9
Although the two sources that have been referenced have authority in this area, they
do not explain the principles very well. A clearer account may be found in [7, pp. 65–75].
23
X
a
b
c
a
b
c
Figure 4: A schematic representation of one possible neural scheme for detecting redundancies between an incoming pattern (‘a b c’), shown at the
bottom of the figure and a stored pattern that matches it, shown above. Key:
a letter inside an unbroken envelope represents a neural symbol; a group of
neural symbols surrounded by a broken-line envelope represents a pattern
assembly; lines between neural symbols ending in a blob (a filled circle)
represent nerve fibres carrying signals towards the blob, with an inhibitory
impact on the neural symbol that is receiving the signals;
24
A possible weakness in the account just given is that, in line with what
was said about Figure 3, we may suppose that the rate of firing of the ‘X’
would quickly settle back to the average rate of firing, but there is nothing
to ensure that that would happen. Issues like this may be clarified with a
working model.
What about the ordering of symbols? With the scheme as just described,
the pattern assembly ‘X a b c’ would respond as strongly to an input pattern ‘b c a’—assuming that there were corresponding neural connections in
place—as it would to the input pattern ‘a b c’.
It looks as if the weakness in that idea is the assumption contained within
it, and it appears that the concept of a typographic pattern, provides an
answer. As was quoted in [26, Section 5.4]:
“Receptors in the retina and body surface are organized as
two-dimensional sheets, and those of the cochlea form a onedimensional line along the basilar membrane. Receptors in these
organs communicate with ganglion cells and those ganglion cells
with central neurons in a strictly ordered fashion, such that relationships with neighbors are maintained throughout. This type
of pattern, in which neurons positioned side by side in one region
communicate with neurons positioned side-by-side in the next region, is called a typographic pattern.” [13, p. 504], emphasis in
the original.
12.2
Implementation
model
of
the
SP-neural
computer
The SP-neural computer model may be developed with these main features:
• The C++ computer language. For reasons given in Appendix C, the
model is probably best implemented using the C++ computer language.
• Neural symbols as a class. Neural symbols would be implemented as
a C++ class. For each instance of the class, this would provide the
following features:
– Inhibitory connections between neurons. The means of sending an
inhibitory message to zero or more other neural symbols. The
paths for such messages may be regarded as nerve-fibre connections between neural symbols like the one between neural symbol
25
‘a’ near the bottom of Figure 4 and neural symbol ‘a’ near the
top, and likewise for the two other pairs of neural symbols in the
figure.
– Indexing of symbols that match each other. If the nerve-fibre connections just mentioned are only ever between a pair of symbols
that match each other, such as, for example, between ‘textttN’ as
the ID neural symbol in a pattern assembly like ‘textttN d o g’
(the word ‘dog’ belongs in the class ‘noun’), and ‘textttN’ as part
of the contents of a pattern assembly like ‘textttS N V’ (a sentence in its simplest form is a noun followed by a verb), then the
list of connections in each neural symbol is, in effect, an index of
matching symbols, much like the index proposed for the abstract
part of the SP theory, described in Section 3.4.
This line of thinking suggests that many neural connections in
the brain, perhaps all of them, are, in effect, direct connections
between neural symbols with a role that is very much like that
of an index in any ordinary database management system, or the
very large indices that are maintained in search engines for the
internet.
– Inhibition of ID neural symbols. Within each pattern assembly,
each neural symbol would have the means of sending an inhibitory
message to ID neural symbol at the far left of the pattern assembly,
as shown in Figure 4. Notice that any one neuron can only ever
be in one pattern assembly, or none. Hence, any one neuron only
ever needs to be able to send inhibitory messages to an ID neural
symbol in the pattern assembly of which it is a part.
– Subclasses of the neural symbol class. There may be a case for
providing two subclasses of the neural symbol class: one for C
neural symbols (each of which may be part of the contents of a
pattern assembly) and one for ID neural symbols (each of which
may serve as an identifier for a pattern assembly). These features
of SP-neural are described in [18, Section 11.3.2].10
• Pattern assemblies as a class. Pattern assemblies would also be implemented as a C++ class. For each instance of the class, it would provide
the means of storing zero or more ID-neural-symbols and zero or more
10
The term C-neuron (meaning “C neural symbol”), introduced in [18, Section
11.3.2], and the term ID-neuron (meaning “ID neural symbol”), introduced in the same
place, should probably be dropped. This is because neural symbols are not necessarily
single neurons—they may be small clusters of neurons.
26
C-neural-symbols, perhaps using linked lists. It may also store other
information such as the number of times the pattern assembly has been
recognised in a given set of New patterns.
12.3
Unsupervised learning in SP-neural
How unsupervised learning may be achieved in SP-neural is considered
mainly in [26, Section 10] with some related observations in Section 11 of
that paper. Here, we have no new insights to add to that account except to
underscore the point that:
“... in any or all of short-term memory, working memory, and
long-term memory, SP-neural may achieve the necessary speed
in the creation of new structures, combined with versatility in
the representation and processing of diverse kinds of knowledge,
by the switching on and off of synapses in pre-established neural
structures and their inter-connections ...” and that this would be
“somewhat like the way in which an ‘uncommitted logic array’
(ULA) may, via small modifications, be made to function like
any one of a wide variety of ‘application-specific integrated circuits’ (ASICs), or how a ‘field-programmable gate array’ (FPGA)
may be programmed to function like any one of a wide variety of
integrated circuits.” [26, Section 11].
13
New hardware
As was mentioned in the Introduction, there may, at some stage, be a case
for the development of new hardware, dedicated to structures and processes
in the SP system, and with potential for gains in efficiency and performance
([23, Section IX], [22, Section III]).
Since searching for matching patterns is an important part of how the SP
system works, gains in efficiency may be achieved by concentrating search
where good results are most likely to be found: “If we want to find some
strawberry jam, our search is more likely to be successful in a supermarket
than it would be in a hardware shop or a car-sales showroom.” [23, Section
IX-A.2], and the statistical knowledge in the system flows directly from the
central role of information compression in the workings of the SP system, and
from the intimate relationship that is known to exist between information
compression and concepts of prediction and probability [12, 12, 6].
27
Gains in efficiency from the use of indices (Section 3.4) may be at least
partly via the encoding of statistical aspects of the data that are being processed.
New hardware may be developed for either or both of the two versions of
the SP system: the abstract version embodied in the SP71 computer model,
and SP-neural as outlined in Appendix A.4, and discussed in Section 12.
14
Applications that may be realised on relatively short timescales
Sources of information about potential applications of the SP machine are
detailed in Appendix A.6.
Here, in brief, are some of the applications that may be developed on
relatively short timescales:
• The SP machine as an intelligent database system. With the application of parallel processing and indexing (Section 3) and the creation
of a “friendly” user interface (Section 4), the SP machine should begin
to be useful as an intelligent database system, with several advantages
over ordinary database systems, as described in [19].
• Software engineering. As outlined in [24, Section 6.6], and described in
more detail in [30], the SP system may be applied with advantage in
software engineering in two main ways:
– In conjunction with an existing programming environment, replacing an ordinary database management system with an SP intelligent database system as indicated in the previous bullet point.
– As a programming environment in its own right. As before, the
SP system may serve as an intelligent database system but, here,
there would be seamless integration with the SP system as a programming environment.
In both of these options, the SP system would provide a more streamlined way of representing such things as entity-relationship models,
compared with the relative inefficiencies of using a relational database
for that purpose. And the SP system lends itself well to the use of
object-oriented concepts for the structuring of knowledge.
With the second option, there is also potential for the application of
unsupervised learning for the automatic or semi-automatic creation of
software.
28
• Information compression. Since information compression is central in
how the SP system works, the SP machine is likely to prove useful in
situations where compression of information is the main requirement. It
may, for example, be a useful means of reducing the size of “big data”,
with corresponding benefits in storage and transmission of those data
[23, Section VII].
15
Conclusion
This paper is about the main things that need to be done to develop a mature
version of the SP machine, starting with the SP71 computer model. After
a summary of our research strategy, we describe the main tasks in what we
think is roughly the order they should be tackled:
1. Starting with the SP71 computer model, create a high-parallel version
of the SP machine as a software virtual machine hosted on a highperformance computer. An interesting alternative would be to create
the SP machine as a software virtual machine driven by high-parallel
search processes in any of the leading search engines. If this were
extended to all search engines, there would be potential to convert the
internet into something that would, in effect, be a giant brain!
2. Create a user interface for the SP machine that is easy to use.
3. Closely related to the previous task is the development of the SP system’s existing strengths in the visualisation of the structure of stored
knowledge and in the provision of an audit trail for all processing.
4. As the basis for such things as computer vision, the SP machine will
need to be generalised to work with patterns in two dimensions as well
as the 1D patterns of the SP71 model.
5. The generalisation just mentioned would include the development of
processes for the discovery of good full and partial matches between
2D patterns (and between a 2D pattern and a 1D pattern) and for the
display of SP-multiple-alignments comprising two or more 2D patterns
(with or without 1D patterns).
6. An investigation is needed into whether or how the SP machine, with
or without some modification, may discover low-level features in speech
and visual images, of the kind that appear to be significant in human
perception.
29
7. Existing strengths of the SP system in the processing of natural language may be developed towards the goal of creating an SP machine
that can understand natural language and produce natural language
from meanings. This is likely to be a major project or programme of
research, with several stages, in conjunction with the development of
unsupervised learning (next). Likely stages include: experiments with
the unsupervised learning of artificial languages: unsupervised learning
of plausible grammars from unsegmented samples of natural language
text; unsupervised learning of plausible structures for various kinds of
non-syntactic knowledge; unsupervised learning of syntactic/semantic
structures; and unsupervised learning from speech.
8. Residual problems with the existing processes for unsupervised learning need to be solved. These weaknesses are failure of the system to
discover structures with intermediate levels of abstraction, and failure to discover discontinuous dependencies in structure. A solution to
these problems will clear the path for the development of unsupervised
learning in conjunction withe the processing of natural language, as
described in the previous bullet point.
9. Existing strengths of the SP system in pattern recognition may be
developed for computer vision, guided by the insights in [21].
10. On the strength of evidence to date, it seems likely that the representation of numbers and the performance of arithmetic processes may
be accommodated within the SP framework. Pending success in those
areas, stop-gap solution may be employed where needed.
11. Somewhat independent of other developments described in this paper,
would be the development of a computer model of SP-neural, drawing
on the information and insights described in [26] and perhaps also in
[18, Chapter 11]. An important part of this development would be to
see whether or how inhibitory processes have the important role that
seems likely from evidence that is available now, and, if not, to provide
some reasons why.
12. At some stage when the SP machine is relatively mature, it is envisaged
that new hardware would be developed, mainly to exploit opportunities to increase the efficiency of computations, most notably by taking
advantage of statistical information that the SP system gathers as a
by-product of how it works.
30
There is potential for the SP machine to be applied on relatively short
timescales in such areas of application as information storage and retrieval,
with intelligence, software engineering, and information compression.
A
Outline of the SP system
To help ensure that this paper is free standing, the SP system is described
in outline here with enough detail to make the rest of the paper intelligible.
As stated in the Introduction and in Section 2, the SP programme of
research is a unique attempt to simplify and integrate observations and concepts across artificial intelligence, mainstream computing, mathematics, and
human learning, perception and cognition, with information compression as
a unifying theme.
The SP system is described most fully in [18] and more briefly in [20].
The SP theory is conceived as a brain-like system as shown schematically
in Figure 5. The system receives New information via its senses and stores
some or all of it in compressed form as Old information.
Old
(compressed)
New
(uncompressed)
Figure 5: Schematic representation of the SP system from an ‘input’ perspective.
Information compression is central in the SP system because of substantial evidence for the importance of information compression in the workings
of brains and nervous systems described in [28]. In this connection, the name
“SP” is short for Simplicity and Power. This is because information compression may be seen to be a process of maximising the simplicity of a body
of information, I, by extracting redundancy from I, whilst retaining as much
as possible of its non-redundant descriptive power.
31
A.1
SP-multiple-alignments
In the SP system, information compression is achieved largely via the powerful concept of SP-multiple-alignment, a concept that has been adapted from
bioinformatics. As illustrated in Figure 6, a multiple alignment in bioinformatics is an arrangement of two or more DNA sequences or amino-acid
sequences so that, by judicious “stretching” of sequences in a computer,
symbols that match each other are brought into line. A “good” multiple
alignment is one in which a relatively large number of matching symbols
have been aligned.
G
|
G
|
A |
|
G
|
G
G
|
G
|
G
|
G
|
G
A
|
| G
|
A C T
|
A A
|
C A
G
C
|
|
G C C C
| | | |
G C C C
|
|
|
C
A
|
A
|
A
|
A
|
A
G
|
G
|
G
|
G
|
G
G
|
G
|
G
|
G
|
G
G
|
G
|
G
|
G
|
G
A
|
A
|
|
|
A
|
A
G
|
G
|
G
|
G
|
G
G
|
G
|
G
|
G
|
G
A
T
|
|
A
|
|
|
| G C T
|
|
A
| A
|
C
G
G
G G A
|
|
| | |
G G C G
G G A
|
|
| | |
G
G A | G A
|
|
| | |
G
G
G G A
|
|
| | |
G
G
G G A
Figure 6: A ‘good’ multiple alignment amongst five DNA sequences. Reproduced with permission from Figure 3.1 in [18].
Because there is normally an astronomically large number of alternative
ways in which multiple alignments may be arranged, heuristic methods are
needed to build multiple alignments in stages, discarding all but the best few
partial multiple alignments at the end of each stage. With such methods,
it is not normally possible to guarantee that a theoretically ideal multiple
alignment has been found, only multiple alignments that are “reasonably
good”.
The following subsections describe how the multiple alignment concept
has been adapted to become the concept of SP-multiple-alignment.
A.2
Patterns, symbols, and multiple alignments in the
SP system
In the SP system, all kinds of knowledge are represented with arrays of atomic
symbols in one or two dimensions, termed patterns. Of course, a 1D pattern is
the same as a sequence in bioinformatics. At present, the SP computer model
works only with 1D patterns but it is envisaged that it will be generalised to
work with 2D patterns (Section 3).
32
In themselves, SP patterns are not very expressive. But within the SPmultiple-alignment framework they become very versatile as a means of representing diverse kinds of knowledge [31, Section 4.2].
The main change to the multiple alignment concept in the SP system
is that, in each SP-multiple-alignment, one or more of the sequences (often
only one) is a New pattern and the rest are Old, and a “good” SP-multiplealignment is one in which the New information may be encoded economically
in terms of the Old information, as described in [18, Section 2.5] and in [20,
Section 4.1].
Examples of SP-multiple-alignments created by the SP computer model
are shown in Figures 7 and 8.
The first of these shows how the parsing of a sentence may be modelled
within the SP-multiple-alignment framework. In the figure, row 0 shows a
New SP pattern representing the sentence to be parsed while rows 1 to 8 show
Old SP patterns, one per row, representing grammatical structures including
words. Row 8 shows the grammatical dependency between the plural subject
of the sentence (marked with ‘textttNp’) and the plural main verb (marked
with ‘textttVp’).
0
t
|
1
|
|
2
|
|
3
D Dp 4 t
|
4
NP D
|
5
|
|
6
|
|
7 S Num
; NP
|
|
8
Num PL ;
w
|
|
|
|
|
w
o
|
|
|
|
|
o #D
|
#D
N
|
|
|
N
k i t t e n
| | | | | |
Nr 5 k i t t e n #Nr
|
|
Np Nr
#Nr
|
|
|
|
|
|
|
|
|
|
|
Np
s
|
|
|
s #N
|
|
|
#N #NP
|
|
Vr 1
|
|
| V Vp Vr
| | |
#NP V |
|
Vp
p
|
|
|
|
|
|
|
|
|
p
l
|
|
|
|
|
|
|
|
|
l
a
|
|
|
|
|
|
|
|
|
a
y
|
|
|
|
|
|
|
|
|
y #Vr
|
#Vr #V
|
#V #S
0
1
2
3
4
5
6
7
8
Figure 7: (a) The best SP-multiple-alignment created by the SP computer
model with a store of Old patterns like those in rows 1 to 8 (representing
grammatical structures, including words) and a New pattern (representing
a sentence to be parsed) shown in row 0. Adapted with permission from
Figures 1 in [19].
In the SP-multiple-alignment shown in Figure 8, column 0 shows a New
SP pattern representing something to be recognised while columns 1 to 4
show Old SP patterns, one per row, representing categories of entity. This
way of representing SP-multiple-alignments is entirely equivalent to the way
in which the SP-multiple-alignment in Figure 7 has been represented. The
choice between them depends entirely on what fits best on the page.
33
0
1
2
3
4
T
Tibs
C ---------------- C
cat
M ------------ M
mammal
A ---------- A
animal
head ---------------------- head
carnassial-teeth
#head --------------------- #head
body ----------------------------------------- body
white-bib ------------------------------------------------ white-bib
#body ---------------------------------------- #body
legs ---------------------- legs
retractile-claws
#legs --------------------- #legs
eats ------ eats
breathes
has-senses
...
#A --------- #A
furry ------------------ furry
warm-blooded
...
#M ----------- #M
purrs --------------------------------- purrs
...
#C --------------- #C
tabby
...
#T
0
1
2
3
4
Figure 8: The best SP-multiple-alignment found by the SP model, with the
New pattern ‘white-bib eats furry purrs’ shown in column 1, and a set
of Old patterns representing different categories of animal and their attributes
shown in columns 1 to 4. Reproduced with permission from Figure 15 in [20].
34
By convention, the New pattern is always shown in column (or row) 0
and Old patterns are shown in other columns (or rows). In some applications
there is more than 1 New pattern in column (row) 0.
A.3
The SP71 computer model
The latest version of the SP computer model is SP71. Instructions for obtaining the source code are in Appendix B.
There is no comprehensive description of SP71, but SP70, a very similar
and slightly earlier version of the program is described quite fully in [18].
More detailed citations are given in the three subsections that follow.
At the level of detail we shall be considering, the description of SP70 in
[18] is an accurate description of SP71. In what follows, it will be regarded
as a description of SP71.
A.3.1
Finding good full and partial matches between patterns in
SP71
In the foundations of the SP computer model is a process for finding good
full and partial matches between two SP patterns (sequences of symbols)
that works well even when there are substantial differences between the two
patterns. This is described quite fully in [18, Appendix A].
This technique is similar in its effects to standard “dynamic programming” methods for comparing sequences [15, 10] but with these main advantages [18, Section 3.10.3.1]:
• It can match arbitrarily long sequences without excessive demands on
memory.
• For any two sequences, it can find a set of alternative matches (each
with a measure of how good it is) instead of a single ‘best’ match.
• The ‘depth’ or thoroughness of the searching can be controlled by parameters.
A.3.2
Building SP-multiple-alignments in SP71
The way in which SP-multiple-alignments are built in SP71 is described in
[18, Sections 3.9 and 3.10]. Slightly confusingly, the main subject of these
sections is SP61, but this is merely a part of the SP computer model that is
concerned with the building of SP-multiple-alignments.
35
As with multiple alignments in bioinformatics, heuristic techniques are
used in the building of SP-multiple-alignments in the SP system: searching
for good SP-multiple-alignments in stages with a weeding out of the lowerscoring SP-multiple-alignments at the end of each stage.
A.3.3
Unsupervised learning in SP71
Unsupervised learning in SP71 is described in [18, Sections 3.9 and 9.2], with
relevant pseudocode in [18, Figures 9.1 and 9.6]. The aim here is, for a given
set of New patterns, to create one or two grammars—meaning collections of
Old SP patterns—that are effective at encoding the given set of New patterns
in an economical manner.
The building of SP-multiple-alignments is an integral part of unsupervised
learning in the SP system. It provides a means of creating Old patterns
via the splitting or amalgamation of pre-existing Old patterns, and via the
direct incorporation of New patterns. And it provides a means of evaluating
candidate grammars in terms of their effectiveness at encoding the given set
of New patterns in an economical manner.
As with the building of SP-multiple-alignments, the creation of good
grammars requires heuristic search through the space of alternative grammars: creating grammars in stages and discarding low-scoring grammars at
the end of each stage.
A.3.4
Varying the thoroughness of heuristic search
A useful feature of SP71 is that, with all three of the main components
(Sections A.3.1, A.3.2, and A.3.3), the thoroughness of the searches may be
varied by varying the amount of memory space that is available for storing
intermediate results. In effect, this controls the amount of backtracking that
can be done and thereby controlling the chances of escaping from local peaks
in the search space.
A.4
SP-neural
Abstract concepts in the SP theory map quite neatly into groupings of neurons and their interconnections in a version of the theory called SP-neural
[26]: SP symbols are realised as neural symbols (single neurons or small
clusters of neurons) and SP patterns are realised as pattern assemblies. Although pattern assemblies in SP-neural are quite similar to cell assemblies as
described by Donald Hebb [3], unsupervised learning in the SP system (including what is envisaged in the development of SP-neural) is quite different
36
from “deep learning in artificial neural networks” [11]. The key differences
are described in [26, Section 10.5].
A.5
Distinctive features and advantages of the SP system
Distinctive features and advantages of the SP system are described in [27]. In
particular, Section V of that paper describes 13 problems with deep learning
in artificial neural networks and how, with the SP system, those problems
may be overcome.
The SP system also provides a comprehensive solution to a fourteenth
problem with deep learning—“catastrophic forgetting”—meaning the way in
which new learning in a deep learning system wipes out old memories.11
Key strengths of the SP system, which owe much to the central role in
the system of the SP-multiple-alignment concept, are in three inter-related
features of the system:
• Versatility in the representation of knowledge. The SP system has
proved to be a versatile framework for the representation of diverse
kinds of knowledge, and there are reasons to think that it may in principle accommodate any kind of knowledge [31, Section 4.2].
• Versatility in aspects of intelligence. The SP system has proven
strengths in several aspects of intelligence, including unsupervised
learning, pattern recognition, planning, problem solving, and several
kinds of reasoning, and there are reasons to think that SP-multiplealignment may prove to be the key to general, human-like artificial
intelligence. [31, Section 4.3].
• Seamless integration of diverse kinds of knowledge and diverse aspects
of intelligence, in any combination. Since the SP system’s strengths in
the representation of diverse kinds of knowledge and in diverse aspects
of intelligence all flow from a single coherent source—the SP-multiplealignment framework—there is clear potential for their seamless integration in any combination [31, Section 4.4]. This kind of integration
appears to be essential in any system that aspires to human-like fluidity,
versatility, and adaptability in intelligence.
11
A solution has been proposed in [4] but it appears to be partial, and it is unlikely to
be satisfactory in the long run.
37
A.6
Potential benefits and applications of the SP system
The discovery that SP-multiple-alignments could simplify and integrate observations and concepts across a broad canvass has been a bit like hitting
the jackpot on a fruit machine, releasing a cascade of potential benefits and
applications, described in a book and in several peer-reviewed papers:
• The book, Unifying Computing and Cognition [18] describes strengths
of the SP system in several areas, mainly in AI, with corresponding
potential in applications that require human-like intelligence.
• The paper “The SP theory of intelligence: benefits and applications”
[24] describes several potential benefits and applications of the SP system, including potential: for an overall simplification of computing
systems, including software; for applications in the processing of natural language; in software engineering; in bioinformatics; and several
more.
• The paper “Big data and the SP theory of intelligence” [23] describes
how the SP system may help solve nine significant problems with big
data.
• The paper “Autonomous robots and the SP theory of intelligence” [22]
describes how the SP system may help in the development of humanlike intelligence in autonomous robots.
• The paper “Application of the SP theory of intelligence to the understanding of natural vision and the development of computer vision”
[21] describes, as its title suggests, how the SP system may help in the
understanding of human vision and in the development of computer
vision.
• The paper “Towards an intelligent database system founded on the
SP theory of computing and cognition” [19] describes how the SP system may function as a database system that, in a relatively streamline
manner, can accommodate a wider variety of kinds of knowledge than
a conventional DBMS, with potential for aspects of intelligence as summarised in [31, Section 4.3].
• The paper “Medical diagnosis as pattern recognition in a framework of
information compression by multiple alignment, unification and search”
[17] describes how the SP system may serve as a repository for medical
38
knowledge and as a means of assisting medical practitioners in medical
diagnosis.
• The paper “Commonsense reasoning, commonsense knowledge, and the
SP theory of intelligence” [25] describes how the SP system can throw
light on most of the problems for AI posed by Ernest Davis and Gary
Marcus in modelling commonsense reasoning [1].
B
Source code for the SP71 computer model
The latest version of the SP computer model is SP71. At the time of writing,
the source code, written in C++ with many comments, may be obtained
via instructions under the headings “Source code” and “Archiving”, near
the bottom of bit.ly/1mSs5XT. More specifically, the source code may be
obtained:
• In the file SP71.zip (bit.ly/1OEVPDw).
• In “Ancillary files” with “The SP theory of intelligence: an overview”
under www.arxiv.org/abs/1306.3888. Apart from SP71.exe, all the
files should be treated as plain text files to be opened with Wordpad,
WinEdt, or the like.
• From a digital archive with: the National Museum of Computing, Block
H, Bletchley Park, Milton Keynes, MK3 6EB, UK; Phone: +44 (0)1908
374708; Email: [email protected].
• From a digital archive with: the Library and Archives Service, Bangor
University, Bangor, Gwynedd, LL57 2DG, UK; Phone: +44 (0) 1248
382981; Email: [email protected].
C
Choice of programming language for creating a computer model of SP-neural
As with the SP71 computer model, the best way to implement SP-neural
would probably be with the C++ computer language. For present purposes,
it has the following main advantages:
• It is widely used and widely available both on single-threaded and
parallel-processing machines.
39
• Like the C programming language from which it was derived, it provides for high level ‘conceptual’ programming and, at the same time,
it provides access to low-level features of the host machine if that is
required.
• As a compiled programming language, its run-time efficiency is almost
certainly higher than any of the programming languages that run interpretively. Since the SP system, like other AI-related systems, is intrinsically hungry for computer power (Section 3.1), run-time efficiency
is important.
• It provides facilities for object-oriented design including mechanisms
for the creation of hierarchies of classes with inheritance of attributes.
• There are likely to be benefits in maintaining consistency with the SP71
computer model.
References
[1] E. Davis and G. Marcus. Commonsense reasoning and commonsense
knowledge in artificial intelligence. Communications of the ACM,
58(9):92–103, 2015.
[2] M. Gold. Language identification in the limit. Information and Control,
10:447–474, 1967.
[3] D. O. Hebb. The Organization of Behaviour. John Wiley & Sons, New
York, 1949.
[4] J. Kirkpatrick. Overcoming catastrophic forgetting in neural networks.
Proceedings of the National Academy of Sciences of the United States of
America, 114(13):3521–3526, 2017.
[5] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521:436–
444, 2015.
[6] M. Li and P. Vitányi. An Introduction to Kolmogorov Complexity and
Its Applications. Springer, New York, 3rd edition, 2014.
[7] P. H. Lindsay and D. A. Norman, editors. An Introduction to Psychology.
Academic Press, New York, 1975.
[8] S. Pinker. The Language Instinct. Penguin Books, London, kindle edition, 1995.
40
[9] F. Ratliff, H. K. Hartline, and W. H. Miller. Spatial and temporal
aspects of retinal inhibitory interaction. Journal of the Optical Society
of America, 53:110–120, 1963.
[10] D. Sankoff and J. B. Kruskall. Time Warps, String Edits, and Macromolecules: the Theory and Practice of Sequence Comparisons. AddisonWesley, Reading, MA, 1983.
[11] J. Schmidhuber. Deep learning in neural networks: an overview. Neural
Networks, 61:85–117, 2015.
[12] R. J. Solomonoff. A formal theory of inductive inference. Parts I and II.
Information and Control, 7:1–22 and 224–254, 1964.
[13] L. R. Squire, D. Berg, F. E. Bloom, S. du Lac, A. Ghosh, and N. C.
Spitzer, editors. Fundamental neuroscience. Elsevier, Amsterdam,
fourth edition, 2013.
[14] G. von Békésy. Sensory Inhibition. Princeton University Press, Princeton, NJ, 1967.
[15] R. A. Wagner and M. J. Fischer. The string-to-string correction problem.
Journal of the ACM, 21(1):168–173, 1974.
[16] J. G. Wolff. Learning syntax and meanings through optimization and
distributional analysis. In Y. Levy, I. M. Schlesinger, and M. D. S.
Braine, editors, Categories and Processes in Language Acquisition, pages
179–215. Lawrence Erlbaum, Hillsdale, NJ, 1988. bit.ly/ZIGjyc.
[17] J. G. Wolff. Medical diagnosis as pattern recognition in a framework of
information compression by multiple alignment, unification and search.
Decision Support Systems, 42:608–625, 2006. bit.ly/1F366o7.
[18] J. G. Wolff. Unifying Computing and Cognition: the SP Theory and Its
Applications. CognitionResearch.org, Menai Bridge, 2006. ISBNs: 09550726-0-3 (ebook edition), 0-9550726-1-1 (print edition). Distributors,
including Amazon.com, are detailed on bit.ly/WmB1rs.
[19] J. G. Wolff. Towards an intelligent database system founded on the SP
theory of computing and cognition. Data & Knowledge Engineering,
60:596–624, 2007. bit.ly/1CUldR6.
[20] J. G. Wolff. The SP theory of intelligence: an overview. Information,
4(3):283–341, 2013. bit.ly/1NOMJ6l.
41
[21] J. G. Wolff. Application of the SP theory of intelligence to the understanding of natural vision and the development of computer vision.
SpringerPlus, 3(1):552–570, 2014. bit.ly/2oIpZB6.
[22] J. G. Wolff. Autonomous robots and the SP theory of intelligence. IEEE
Access, 2:1629–1651, 2014. bit.ly/18DxU5K.
[23] J. G. Wolff. Big data and the SP theory of intelligence. IEEE Access,
2:301–315, 2014. bit.ly/2qfSR3G. This paper, with minor revisions, is
reproduced in Fei Hu (Ed.), Big Data: Storage, Sharing, and Security
(3S), Taylor & Francis LLC, CRC Press, 2016, pp. 143–170.
[24] J. G. Wolff. The SP theory of intelligence: benefits and applications.
Information, 5(1):1–27, 2014. bit.ly/1FRYwew.
[25] J. G. Wolff. Commonsense reasoning, commonsense knowledge, and
the SP theory of intelligence. 2016. Submitted for publication.
bit.ly/2eBoE9E.
[26] J. G. Wolff. Information compression, multiple alignment, and the representation and processing of knowledge in the brain. Frontiers in Psychology, 7:1584, 2016. bit.ly/2esmYyt.
[27] J. G. Wolff. The SP theory of intelligence: its distinctive features and
advantages. IEEE Access, 4:216–246, 2016. bit.ly/2qgq5QF.
[28] J. G. Wolff. Evidence for information compression in the workings of
brains and nervous systems. Technical report, CognitionResearch.org,
2017. Submitted for publication. See bit.ly/2ruLnrV.
[29] J. G. Wolff. On the “mysterious” effectiveness of mathematics in science. Technical report, CognitionResearch.org, 2017. Submitted for
publication. bit.ly/2otrHD0. This report is also achieved in vixra.org/
and hal.archives-ouvertes.fr/hal-01534622.
[30] J. G. Wolff. Software engineering and the sp theory of intelligence.
Technical report, CognitionResearch.org, 2017. In preparation.
[31] J. G. Wolff. Towards a new mathematics for science. Technical report, CognitionResearch.org, 2017. bit.ly/2o1pr8p. This report is also
achieved in vixra.org/ and hal.archives-ouvertes.fr/hal-01534619.
42
| 2cs.AI
|
On the Skolem Problem for Continuous Linear
Dynamical Systems
Ventsislav Chonev1 , Joël Ouaknine2 , and James Worrell2
1
2
Institute of Science and Technology, Austria
University of Oxford, UK
arXiv:1506.00695v3 [cs.SY] 10 May 2016
Abstract
The Continuous Skolem Problem asks whether a real-valued function satisfying a linear differential equation has a zero in a given interval of real numbers. This is a fundamental reachability
problem for continuous linear dynamical systems, such as linear hybrid automata and continuoustime Markov chains. Decidability of the problem is currently open—indeed decidability is open
even for the sub-problem in which a zero is sought in a bounded interval. In this paper we show
decidability of the bounded problem subject to Schanuel’s Conjecture, a unifying conjecture in
transcendental number theory. We furthermore analyse the unbounded problem in terms of the
frequencies of the differential equation, that is, the imaginary parts of the characteristic roots.
We show that the unbounded problem can be reduced to the bounded problem if there is at most
one rationally linearly independent frequency, or if there are two rationally linearly independent
frequencies and all characteristic roots are simple. We complete the picture by showing that decidability of the unbounded problem in the case of two (or more) rationally linearly independent
frequencies would entail a major new effectiveness result in Diophantine approximation, namely
computability of the Diophantine-approximation types of all real algebraic numbers.
1998 ACM Subject Classification F.1.1 Models of Computation, F.2.1 Numerical Algorithms
and Problems
Keywords and phrases Differential equations, Reachability, Baker’s theorem, Schanuel’s Conjecture, Semi-algebraic sets
Digital Object Identifier 10.4230/LIPIcs.ICALP.2016.XXX
1
Introduction
The Continuous Skolem Problem is a fundamental decision problem concerning reachability in
continuous-time linear dynamical systems. The problem asks whether a real-valued function
satisfying an ordinary linear differential equation has a zero in a given interval of real numbers.
More precisely, an instance of the problem comprises an interval I ⊆ R≥0 with rational
endpoints, an ordinary differential equation
f (n) + an−1 f (n−1) + . . . + a0 f = 0
(1)
whose coefficients are real algebraic, together with initial conditions f (0), . . . , f (n−1) (0)
that are also real algebraic numbers. Writing f : R≥0 → R for the unique solution of the
differential equation subject to the initial conditions, the question is whether there exists
t ∈ I such that f (t) = 0. Decidability of this problem is currently open. Decidability of
the sub-problem in which the interval I is bounded, called the Bounded Continuous Skolem
Problem, is also open [4, Open Problem 17].
The nomenclature Continuous Skolem Problem is based on an analogy with the Skolem
Problem for linear recurrence sequences, which asks whether a given linear recurrence
© Ventsislav Chonev and Joël Ouaknine and James Worrell;
licensed under Creative Commons License CC-BY
43rd International Colloquium on Automata, Languages, and Programming (ICALP 2016).
Editors: Ioannis Chatzigiannakis, Michael Mitzenmacher, Yuval Rabani, and Davide Sangiorgi; Article No. XXX;
pp. XXX:1–XXX:20
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XXX:2
On the Skolem Problem for Continuous Linear Dynamical Systems
sequence has a zero term [12]. Whether the latter problem is decidable is an outstanding
question in number theory and theoretical computer science; see, e.g., the exposition of
Tao [20, Section 3.9].
The continuous dynamics of linear hybrid automata and the evolution of continuoustime Markov chains, amongst many other examples, are determined by linear differential
equations of the form x0 (t) = Ax(t), where x(t) ∈ Rn and A is an n × n matrix of real
numbers [1]. A basic reachability question in this context is whether, starting from an initial
state x(0), the system reaches a given hyperplane {y ∈ Rn : uT y = 0} with normal vector
u ∈ Rn . For example, one can ask whether the continuous flow of a hybrid automaton
leads to a particular transition guard being satisfied or an invariant being violated. Now
the function f (t) = uT x(t) satisfies a linear differential equation of the form (1), and it
turns out that the hyperplane reachability problem is inter-reducible with the Continuous
Skolem Problem (see [4, Theorem 6] for further details). Moreover, under this reduction
the Bounded Continuous Skolem Problem corresponds to a time-bounded version of the
hyperplane reachability problem.
The characteristic polynomial of the differential equation (1) is
χ(x) := xn + an−1 xn−1 + . . . + a0 .
Let λ1 , . . . , λm be the distinct roots of χ. Any solution of (1) has the form f (t) =
Pm
λj t
, where the Pj are polynomials with algebraic coefficients that are determined
j=1 Pj (t)e
by the initial conditions of the differential equation. We call a function f in this form
an exponential polynomial. If the roots of χ are all simple then f can be written as an
exponential polynomial in which the polynomials Pj are all constant.
The Continuous Skolem Problem can equivalently be formulated in terms of whether an
exponential polynomial has a zero in a given interval of reals. If the characteristic roots
have the form λj = rj + iωj , where rj , ωj ∈ R, then we can also write f (t) in the form
Pm
f (t) = j=1 erj t (Q1,j (t) sin(ωj t) + Q2,j (t) cos(ωj t)), where the polynomials Q1,j , Q2,j have
real algebraic coefficients. We call ω1 , . . . , ωm the frequencies of f .
Our first result is to show decidability of the Bounded Continuous Skolem Problem subject
to Schanuel’s Conjecture, a unifying conjecture in transcendental number theory that plays a
key role in the study of the exponential function on both real and complex numbers [21, 22].
Intuitively, decidability of the Bounded Continuous Skolem Problem is non-trivial because
an exponential polynomial can approach 0 tangentially. Assuming Schanuel’s Conjecture,
we show that any exponential polynomial admits a factorisation such that the zeros of each
factor can be detected using finite-precision numerical computations. Our method, however,
does not enable us to bound the precision required to find zeros, so we do not obtain a
complexity bound for the procedure.
A celebrated paper of Macintyre and Wilkie [18] obtains decidability of the first-order theory of Rexp = (R, 0, 1, <, · , +, exp) assuming Schanuel’s Conjecture over R. The proof of [17,
Theorem 3.1] mentions an unpublished result of Macintyre and Wilkie that generalises [18]
to obtain decidability when Rexp is augmented with the restricted functions sin [0,2π] and
cos [0,2π] , this time assuming Schanuel’s Conjecture over C. This result immediately implies
(conditional) decidability of the Bounded Continuous Skolem Problem. However, decidability
of latter problem is simpler and, as we show below, can be established more directly.
In the unbounded case we analyse exponential polynomials in terms of the number of
rationally linearly independent frequencies. We show that the unbounded problem can be
reduced to the bounded problem if there is at most one rationally linearly independent
frequency, or if there are two rationally linearly independent frequencies and all characteristic
V. Chonev and J. Ouaknine and J. Worrell
XXX:3
roots are simple. These two reductions are unconditional and rely on the cell decomposition
theorem for semi-algebraic sets [3] and Baker’s Theorem on linear forms in logarithms of
algebraic numbers [2].
We complete the picture by showing that decidability of the unbounded problem in the
case of two (or more) rationally linearly independent frequencies would entail a major new
effectiveness result in Diophantine approximation—namely computability of the Diophantineapproximation types of all real algebraic numbers. As we discuss in Appendix A, currently
essentially nothing is known about Diophantine-approximation types of algebraic numbers of
degree three or higher, and they are the subject of several longstanding open problems.
The question of deciding whether an exponential polynomial f has infinitely many zeros
is investigated in [7]. There the problem is shown to be decidable if f satisfies a differential
equation of order at most 7. This result does not rely on Schanuel’s Conjecture. It is also
shown in [7] that, analogously with the Continuous Skolem Problem, decidability of the
Infinite Zeros Problem in the general case would entail significant new effectiveness results in
Diophantine approximation.
2
2.1
Mathematical Background
Zero Finding
Let f : [a, b] → R be a function defined on a closed interval of reals with endpoints a, b ∈ Q.
Suppose the following two conditions hold: (i) there exists M > 0 such that f is M -Lipschitz,
i.e., |f (s) − f (t)| ≤ M |s − t| for all s, t ∈ [a, b]; (ii) given t ∈ [a, b] ∩ Q and positive error
bound ε ∈ Q, we can compute q ∈ Q such that |f (t) − q| < ε. Then given a positive
rational number δ we can compute piecewise linear functions fδ+ , fδ− : [a, b] → R such that
fδ− (t) ≤ f (t) ≤ fδ+ (t) and fδ+ (t) − fδ− (t) ≤ δ for all t ∈ [a, b]. We do this as follows:
δ
1. Pick N ∈ N such that N1 < 4(b−a)M
and consider sample points sj := a + (b−a)j
N ,
δ
j = 0, . . . , N , dividing the interval [a, b] into N sub-intervals, each of length at most 4M
.
−
δ
2. For each sample point sj compute qj ∈ Q such that |qj −f (sj )| < 4 , define fδ (sj ) = qj − 2δ ,
fδ+ (sj ) = qj + 2δ , and extend fδ− and fδ+ linearly between sample points.
Note that the Lipschitz condition on f ensures that fδ− ≤ f ≤ fδ+ .
Now suppose that f satisfies the following additional conditions: (iii) f (a) 6= 0, f (b) 6= 0;
(iv) for any t ∈ (a, b) such that f (t) = 0, f 0 (t) exists and is non-zero, i.e., f has no tangential
zeros. Then we can decide the existence of a zero of f by computing upper and lower
approximations fδ+ and fδ− for successively smaller values of δ. If fδ+ (t) < 0 for all t or
fδ− (t) > 0 for all t then we conclude that f has no zero on [a, b]; if fδ+ (s) < 0 and fδ− (t) > 0
for some s, t then we conclude that f has a zero; otherwise we proceed to a smaller value
of δ. This procedure terminates since by (iii) and (iv) either f has a zero in [a, b] or it is
bounded away from zero.
2.2
Number-Theoretic Algorithms
For the purposes of establishing decidability, we can assume that an instance of the Continuous
Pm
λj t
Skolem Problem is a real-valued exponential polynomial f (t) =
, where
j=1 Pj (t)e
λ1 , . . . , λm and the coefficients of the polynomials P1 , . . . , Pm are algebraic, see [4, Theorem
6].
For computational purposes we represent an algebraic number α by a polynomial P with
rational coefficients such that P (α) = 0, together with a numerical approximation p + qi,
where p, q ∈ Q, of sufficient accuracy to distinguish α from the other roots of P [8, Section
ICALP 2016
XXX:4
On the Skolem Problem for Continuous Linear Dynamical Systems
4.2.1]. Given this representation we can obtain approximations of α to within an arbitrarily
small additive error.
Let K be the extension field of Q generated by λ1 , . . . , λm and the coefficients of the
polynomials P1 , . . . , Pm . Note that K is closed under complex conjugation. We can compute
a primitive element of K, that is, an algebraic number θ such that K = Q(θ), together with
a representation of each characteristic root λj as a polynomial in θ with rational coefficients
(see [8, Section 4.5]). From the representation of λ1 , . . . , λm as elements of Q(θ), it is
straightforward to determine maximal Q-linearly independent subsets of {Re(λj ) : 1 ≤ j ≤ m}
and {Im(λj ) : 1 ≤ j ≤ m} (see [14, Section 1]).
Let log denote the branch of the complex logarithm defined by log(reiθ ) = log(r) + iθ
for a positive real number r and 0 ≤ θ < 2π. Recall that one can compute log z and ez to
within arbitrarily small additive error given a sufficiently precise approximation of z [6].
2.3
Laurent Polynomials
Let K be a sub-field of C that has finite dimension over Q and is closed under complex
conjugation. Fix non-negative integers r and s, and consider a single variable x and tuples
of variables y = hy1 , . . . , yr i and z = hz1 , . . . , zs i. Consider the ring of Laurent polynomials
R := K[x, y1 , y1−1 , . . . , yr , yr−1 , z1 , z1−1 , . . . , zs , zs−1 ] ,
which can be seen as a localisation1 of the polynomial ring A := K[x, y1 , . . . , yr , z1 , . . . , zs ]
in the multiplicative set generated by the set of variables {y1 , . . . , yr } ∪ {z1 , . . . , zs }. The
multiplicative units of R are the non-zero monomials in variables y1 , . . . , yr and z1 , . . . , zs . As
the localisation of a unique factorisation domain, R is itself a unique factorisation domain [9,
Theorem 10.3.7]. From the proof of this fact it moreover easily follows that R inherits from
A computability of factorisation into irreducibles (e.g., using the algorithm of [16]).
We extend the operation of complex conjugation to a ring automorphism of R as follows.
Given a polynomial
P =
n
X
aj xuj y1 vj1 . . . yr vjr z1 wj1 . . . zs wjs ,
j=1
where a1 , . . . , an ∈ K, define its conjugate to be
P :=
n
X
aj xuj y1 vj1 . . . yr vjr z1 −wj1 . . . zs −wjs .
j=1
This definition is motivated by thinking of the variables x and y1 , . . . , yr as real-valued and
the variables z1 , . . . , zs as taking values in the unit circle in the complex plane.
We will need the following proposition characterising those polynomials in P ∈ R such
that P and P are associates, i.e., such that P is equal to the product of P by a monomial.
Here we use pointwise notation for exponentiation: given a tuple of integers u = hu1 , . . . , us i,
we write z u for the monomial z1u1 . . . zsus .
I Proposition 1. Let P ∈ R be such that P = z u P for u ∈ Zs . Then either (i) P has
the form P = z u Q for some Q ∈ R with Q = Q, or (ii) there exists Q ∈ R such that
P = Q + z u Q and P does not divide Q in R.
1
Recall that the localisation of a commutative ring U in a multiplicatively closed subset S such that
0U 6∈ S is the ring of formal fractions US = {a/s : a ∈ U, s ∈ S}, with addition and multiplication
defined as usual.
V. Chonev and J. Ouaknine and J. Worrell
XXX:5
Proof. Consider a monomial M such that z u M = M . Then M has a real coefficient and the
exponent w of z in M satisfies 2w = u. Thus if z u M = M for every monomial M appearing
in P then P has the form Qz w , where 2w = u and Q is a polynomial in the variables x and
y with real coefficients. In particular Q = Q, and statement (i) of the proposition applies.
Suppose now that z u M 6= M for some monomial M appearing in P . Then the map
sending M to z u M induces a permutation of order 2 on the monomials on P . Thus we may
Pn
write P = j=1 Mj , where n = k + 2` for some k ≥ 0 and ` ≥ 1 such that z u Mj = Mj for
Pk
Pk+`
1 ≤ j ≤ k and z u Mj = Mj+` for k +1 ≤ j ≤ `. Then, writing Q := 12 j=1 Mj + j=k+1 Mj ,
we have P = Q + z u Q.
The set of monomials appearing in Q is a proper subset of the set of monomials appearing
in P (up to constant coefficients) and so Q cannot be a constant multiple of P . It also follows
that for each variable σ ∈ {x, yj , zk : 1 ≤ j ≤ r, 1 ≤ k ≤ s}, the maximum degree of σ in P
is at least its maximum degree in Q, and likewise for σ −1 . Thus Q cannot be a multiple of
P by a non-constant polynomial either. We conclude that P does not divide Q.
J
2.4
Transcendence Theory
We will use transcendence theory in our analysis of both the bounded and unbounded variants
of the Continuous Skolem Problem. In the unbounded case we will use the following classical
result.
I Theorem 2 (Gelfond-Schneider). Let a, b be algebraic numbers not equal to 0 or 1. Then
log(b)
for any branch of the logarithm function, log(a)
is either rational or transcendental.
In fact we will make use of the following corollary, which is obtained by applying Theorem 2
to the algebraic numbers a = ei(α2 −α1 ) and b = ei(β2 −β1 ) .
I Corollary 3. Let α1 6= β1 , α2 6= β2 all lie in [0, π] and suppose that cos(α1 ), cos(α2 ), cos(β1 )
2
and cos(β2 ) are algebraic. Then ββ21 −α
−α1 is either rational or transcendental.
Our results in the bounded case depend on Schanuel’s conjecture, a unifying conjecture
in transcendental number theory [15], which, if true, greatly generalises many of the central results in the field (including the Gelfond-Schneider Theorem, above). Recall that a
transcendence basis of a field extension L/K is a subset S ⊆ L such that S is algebraically
independent over K and L is algebraic over K(S). All transcendence bases of L/K have the
same cardinality, which is called the transcendence degree of the extension.
I Conjecture 4 (Schanuel’s Conjecture [15]). Let a1 , . . . , an be complex numbers that are
linearly independent over Q. Then the field Q(a1 , . . . , an , ea1 , . . . , ean ) has transcendence
degree at least n over Q.
A special case of Schanuel’s conjecture, that is known to hold unconditionally, is the
Lindemann-Weierstrass Theorem [15]: if a1 , . . . , an are algebraic numbers that are linearly
independent over Q, then ea1 , . . . , ean are algebraically independent.
We apply Schanuel’s conjecture via the following proposition.
I Proposition 5. Let {a1 , . . . , ar } and {b1 , . . . , bs } be Q-linearly independent sets of real
algebraic numbers. Furthermore, let P, Q ∈ R be two polynomials that have algebraic
coefficients and are coprime in R. Then the equations
P (t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t )
a1 t
Q(t, e
,...,e
ar t
ib1 t
,e
,...,e
ibs t
)
=
0
(2)
=
0
(3)
have no non-zero common solution t ∈ R.
ICALP 2016
XXX:6
On the Skolem Problem for Continuous Linear Dynamical Systems
Proof. Consider a solution t 6= 0 of Equations (2) and (3). By passing to suitable associates,
we may assume without loss of generality that P and Q lie in A, i.e., that all variables in P
and Q appear with non-negative exponent. Moreover, since P and Q are coprime in R, their
greatest common divisor R in A is a monomial. In particular,
R(t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t ) 6= 0 .
Thus, dividing P and Q by R, we may assume that P and Q are coprime in A and that
Equations (2) and (3) still hold.
By Schanuel’s conjecture, the extension
Q(a1 t, . . . , ar t, ib1 t, . . . , ibs t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t )/Q
has transcendence degree at least r + s. Since a1 , . . . , ar and b1 , . . . , bs are algebraic over Q,
writing
S := ht, ea1 t , . . . , ear t , eib1 t , . . . , eibs t i ,
it follows that the extension Q(S)/Q also has transcendence degree at least r + s.
From Equations (2) and (3) we can regard S as specifying a common root of P and Q.
Pick some variable σ ∈ {x, yj , zj : 1 ≤ i ≤ r, 1 ≤ j ≤ s} that has positive degree in P . Then
the component of S corresponding to σ is algebraic over the remaining components of S. We
claim that the remaining components of S are algebraically dependent and thus S comprises
at most r + s − 1 algebraically independent elements, contradicting Schanuel’s conjecture.
The claim clearly holds if σ does not appear in Q. On the other hand, if σ has positive
degree in Q then, since P and Q are coprime in A, the multivariate resultant Resσ (P, Q) is
a non-zero polynomial in the set of variables {x, yj , zj : 1 ≤ i ≤ r, 1 ≤ j ≤ s} \ {σ} which
has a root at S (see, e.g., [10, Page 163]). Thus the claim also holds in this case. In either
case we obtain a contradiction to Schanuel’s conjecture and we conclude that Equations (2)
and (3) have no non-zero solution t ∈ R.
J
3
Decidability of the Bounded Continuous Skolem Problem
Suppose that {a1 , . . . , ar } and {ib1 , . . . , ibs } are Q-linearly independent sets of real and
imaginary numbers respectively. Let the ring of Laurent polynomials R be as in Section 2.3
and consider the exponential polynomial
f (t) = P (t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t ) ,
(4)
where P ∈ R is irreducible. We say that f is a Type-1 exponential polynomial if P and P
are not associates in R, we say that f is Type-2 if P = αP for some α ∈ C, and we say that
f is Type-3 if P = U P for some non-constant unit U ∈ R.
I Example 6. The simplest example of a Type-3 exponential polynomial is g(t) = 1 + eit .
Here g(t) = P (eit ), where P (z) = 1 + z is an irreducible polynomial that is associated with
its conjugate P (z) = 1 + z −1 . Note that the exponential polynomial f (t) = 2 + 2 cos(t),
which has infinitely many tangential zeros, factors as the product of two type-3 exponential
polynomials f (t) = g(t)g(t).
In the case of a Type-2 exponential polynomial P = αP it is clear that we must have
|α| = 1. Moreover, by replacing P by βP , where β 2 = α, we may assume without loss of
V. Chonev and J. Ouaknine and J. Worrell
XXX:7
generality that P = P . Similarly, in the case of a Type-3 exponential polynomial, we can
assume without loss of generality that P = z u P for some non-zero vector u ∈ Zs .
Pm
Now consider an arbitrary exponential polynomial f (t) := j=1 Pj (t)eλj t . Assume that
the coefficient field K of R contains the coefficients of P1 , . . . , Pm . Let {a1 , . . . , ar } be a
basis of the Q-vector space spanned by {Re(λj ) : 1 ≤ j ≤ m} and let {b1 , . . . , bs } be a basis
of the the Q-vector space spanned by {Im(λj ) : 1 ≤ j ≤ m}. Without loss of generality we
may assume that each characteristic root λ is an integer linear combination of a1 , . . . , ar
and ib1 , . . . , ibs . Then eλt is a product of positive and negative powers of ea1 t , . . . , ear t and
eib1 t , . . . , eibs t , and hence there is a Laurent polynomial P ∈ R such that
f (t) = P (t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t ) .
(5)
Since P can be written as a product of irreducible factors, it follows that f can be written
as product of Type-1, Type-2, and Type-3 exponential polynomials, and moreover this
factorisation can be computed from f . Thus it suffices to show how to decide the existence
of zeros of these three special forms of exponential polynomial. We will handle all three cases
using Schanuel’s conjecture.
Pm
Writing the exponential polynomial f (t) in (5) in the form f (t) = j=1 Qj (t)eλj t , it
follows from the irreducibility of P that the polynomials Q1 , . . . , Qm have no common root.
But then by the Lindemann-Weierstrass Theorem any zero of f must be transcendental
(see [4, Theorem 8]).
I Theorem 7. The Bounded Continuous Skolem Problem is decidable subject to Schanuel’s
conjecture.
Proof. Consider an exponential polynomial
f (t) = P (t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t ) ,
(6)
where P ∈ R is irreducible. Suppose that {a1 , . . . , ar } and {ib1 , . . . , ibs } are Q-linearly
independent sets of, respectively, real and imaginary numbers lying in the coefficient field
K of R. We show how to decide whether f has a zero in a bounded interval I ⊆ R≥0 ,
considering separately the case of Type-1, Type-2, and Type-3 exponential polynomials.
Case (i): f is a type-1 exponential polynomial
Note that P and P are coprime in R since, by assumption, they are both irreducible and
are not associates. We claim that in this case the equation f (t) = 0 has no solution t ∈ R.
Indeed f (t) = 0 implies
P (t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t )
a1 t
P (t, e
,...,e
ar t
ib1 t
,e
,...,e
ibs t
)
=
0
=
0,
and the non-existence of a zero of f follows immediately from Proposition 5.
Case (ii): f is a type-2 exponential polynomial
In this case we have P = P and so f is real-valued. Our aim is to use the procedure of
Section 2.1 to determine whether or not f has a zero in [c, d], where c, d ∈ Q. To this
end, notice first that f (c), f (d) 6= 0 since any root of f must be transcendental. Moreover,
since f 0 is bounded on [c, d], f is Lipschitz on [c, d]. It remains to verify that the equations
f (t) = 0, f 0 (t) = 0 have no common solution t ∈ [c, d].
ICALP 2016
XXX:8
On the Skolem Problem for Continuous Linear Dynamical Systems
We can write f 0 (t) in the form
f 0 (t) = Q(t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t ) ,
where Q is the polynomial
r
Q=
s
X
∂P X
∂P
∂P
+
+
.
aj yj
ibj zj
∂x j=1
∂yj j=1
∂zj
We claim that P and Q are coprime in R. Indeed, since P is irreducible, P and Q can only
fail to be coprime if P divides Q.
If P has strictly positive degree k in x then Q has degree k − 1 in x and thus P cannot
divide Q. (Recall that all polynomials in R have non-negative degree in the variable x.)
On the other hand, if P has degree 0 in x then Q is obtained from P by multiplying each
Pr
Ps
monomial y u z v appearing in P by the complex-number constant j=1 aj uj + i j=1 bj vj .
Moreover, by the assumption of linear independence of {a1 , . . . , ar } and {b1 , . . . , bs }, each
monomial in P is multiplied by a different constant. Since P is not a unit, it has at least two
different monomials and so P is not a constant multiple of Q. Furthermore, for each variable
σ ∈ {yj , yj−1 : 1 ≤ j ≤ r} ∪ {zj , zj−1 : 1 ≤ j ≤ s}, its degree in P is equal to its degree in Q.
Thus P cannot be a multiple of Q by a non-constant polynomial either.
We conclude that P does not divide Q and hence P and Q are coprime. It now follows
from Proposition 5 that the equations f (t) = f 0 (t) = 0 have no solution t ∈ R.
Case (iii): f is a type-3 exponential polynomial
Suppose that f is a Type-3 exponential polynomial. Then in (6) we have that P = z u P
for some non-zero vector u ∈ Zs . By Proposition 1 we can write P = Q + z u Q for some
polynomial Q ∈ R that is coprime with P .
Now define
g1 (t) := Q(t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t )
and g2 (t) := eib1 u1 · · · eibs us g1 (t), so that f (t) = g1 (t) + g2 (t) for all t.
We show that g2 (t) 6= 0 for all t ∈ R. Indeed if g2 (t) = 0 for some t then we also have
g1 (t) = 0 and hence f (t) = 0. For such a t it follows that
P (t, ea1 t , . . . , ear t , eib1 t , . . . , eibs t )
a1 t
Q(t, e
,...,e
ar t
ib1 t
,e
,...,e
ibs t
)
=
0
=
0.
But P and Q are coprime and so these two equations cannot both hold by Proposition 5. Not
only do we have g2 (t) 6= 0 for all t ∈ R, but, applying the sampling procedure in Section 2.1
to |g2 (t)|2 (which is a differentiable function) we can compute a strictly positive lower bound
on |g2 (t)| over the interval [c, d].
Since g2 (t) 6= 0 for all t ∈ R we may define the function h : [c, d] → R by
g1 (t)
h(t) := π + i log
.
g2 (t)
Notice that h(t) = 0 if and only if f (t) = 0. Our aim is to use the procedure of Section 2.1
to decide the existence of a zero of h in the interval [c, d], and thus decide whether f has a
zero in [c, d].
V. Chonev and J. Ouaknine and J. Worrell
XXX:9
Let t ∈ (c, d) be such that h(t) = 0. Then g1 (t) = −g2 (t) and so gg12 (t)
(t) = −1 does not lie
on the branch cut of the logarithm function. It follows that h is differentiable at t and
h0 (t) = 0
iff
g2 (t) g10 (t)g2 (t) − g20 (t)g1 (t)
=0
g1 (t)
g2 (t)2
iff
g10 (t)g2 (t) − g20 (t)g1 (t) = 0 (since |g1 (t)| = |g2 (t)| =
6 0)
iff
g10 (t)g2 (t) + g20 (t)g2 (t) = 0 (since g1 (t) = −g2 (t))
iff
g10 (t) + g20 (t) = 0
iff
f 0 (t) = 0 .
Thus h(t) = h0 (t) = 0 implies f (t) = f 0 (t) = 0. But the proof in Case (ii) shows that
f (t) = f 0 (t) = 0 is impossible. (Nothing in that argument hinges on f being real-valued.)
Thus h has no tangential zeros in (c, d).
We cannot directly use the procedure in Section 2.1 to decide whether h has a zero in
[c, d] since h is not necessarily continuous: its value can jump from −π to π (or vice versa)
due to the branch cut of the logarithm along the positive real axis. However, due to the
strictly positive lower bound on |g2 (t)|, the function |h| is Lipschitz on [c, d]. Thus, applying
the sampling procedure in Section 2.1 for computing lower and upper bounds of Lipschitz
functions we can compute a set E ⊆ [c, d] such that E is a finite union of intervals with
π
rational endpoints, |f (t)| ≤ 2π
3 for t ∈ E, and |f (t)| ≥ 3 for t 6∈ E. In particular, E contains
all zeros of f in [c, d] and f is Lipschitz on E. Thus we can apply the zero-finding procedure
from Section 2.1 to the restriction h E and thereby decide whether or not h has a zero on
[c, d].
J
4
The Unbounded Case
In this section we consider the unbounded case of the Continuous Skolem Problem. For our
analysis it is convenient to present exponential polynomials in the form
f (t) =
n
X
erj t (P1,j (t) cos(ωj t) + P2,j (t) sin(ωj t)) ,
(7)
j=1
where rj , ωj are real algebraic numbers and P1,j , P2,j are polynomials with real algebraic
coefficients for j = 1, . . . , n. Our aim is to classify the difficulty of the problem in terms of
the number of rationally linear independent frequencies ω1 , . . . , ωn .
Recall that in Section 3 we have shown the bounded problem to be decidable subject to
Schanuel’s Conjecture. In Appendix C we give a reduction of the unbounded problem to the
bounded problem in case the set of frequencies spans a one-dimensional vector space over Q.
In the present section we give a reduction of the unbounded problem to the bounded problem
in case the set of frequencies spans a two-dimensional vector space over Q and the polynomials
P1,j and P2,j are all constant. (This last condition is equivalent to the assumption that f (t)
is simple.) The argument in the two-dimensional case is a more sophisticated version of that
in the one-dimensional case, although the result is not more general due the assumption of
simplicity.
In Appendix A we present a family of instances showing that obtaining decidability of
the unbounded problem in the two-dimensional case without the assumption of simplicity
would require much finer Diophantine-approximation bounds than are currently known.
ICALP 2016
XXX:10 On the Skolem Problem for Continuous Linear Dynamical Systems
4.1
Background on Semi-Algebraic Sets
A subset of Rn is semi-algebraic if it is defined by a Boolean combination of constraints
of the form P (x1 , . . . , xn ) > 0, where P is a polynomial with real algebraic coefficients. A
partial function f : Rn → R is semi-algebraic if its graph is a semi-algebraic subset of Rn+1 .
The Tarski-Seidenberg Theorem [5, Section 1] states that the semi-algebraic sets are closed
under projection and are therefore precisely the first-order definable sets over the structure
(R, <, +, ·, 0, 1).
Let (i1 , . . . , in ) be a sequence of zeros and ones of length n ≥ 1. An (i1 , . . . , in )-cell is a
subset of Rn , defined by induction on n as follows:
(i) A (0)-cell is a singleton subset of R and a (1)-cell is an open interval (a, b) ⊆ R.
(ii) Let X ⊆ Rn be a (i1 , . . . , in )-cell and f : X → R a continuous semi-algebraic function.
Then {(x, f (x)) ∈ Rn+1 : x ∈ X} is a (i1 , . . . , in , 0)-cell, while {(x, y) ∈ Rn+1 : x ∈
X ∧ y < f (x)} and {(x, y) ∈ Rn+1 : x ∈ X ∧ y > f (x)} are both (i1 , . . . , in , 1)-cells.
(iii) Let X ⊆ Rn be a (i1 , . . . , in )-cell and f, g : X → R continuous semi-algebraic functions
such that f (x) < g(x) for all x ∈ X. Then {(x, y) ∈ Rn+1 : f (x) < y < g(x)} is a
(i1 , . . . , in , 1)-cell.
A cell in Rn is a (i1 , . . . , in )-cell for some (necessarily unique) sequence (i1 , . . . , in ).
A fundamental result about semi-algebraic sets, that we will use below, is the CellDecomposition Theorem [3]: given a semi-algebraic set E ⊆ Rn one can compute a partition
of E as a disjoint union of cells E = C1 ∪ . . . ∪ Cm .
We will also need the following result, proved in Appendix B.
I Lemma 8. Let D ⊆ Rn be a semi-algebraic set, g : D → R a bounded semi-algebraic
function, and r1 , . . . , rn real algebraic numbers. Define S = {t ∈ R≥0 : (er1 t , . . . , ern t ) ∈ D}.
Then
(i) It is decidable whether or not S is bounded. If S is bounded then we can compute T0 ∈ N
such that S ⊆ [0, T0 ] and if S is unbounded then we can compute T0 ∈ N such that
(T0 , ∞) ⊆ S.
(ii) If S is unbounded then the limit g ∗ = limt→∞ g(er1 t , . . . , ern t ) exists, is an algebraic
number, and there are effective constants T1 , ε > 0 such that |g(er1 t , . . . , ern t ) − g ∗ | < e−εt
for all t > T1 .
4.2
Two Linearly Independent Frequencies
The following lemma, which is a reformulation of [4, Lemma 13], plays an instrumental role
in this section. The lemma itself relies on a powerful quantitative result in transcendence
theory—Baker’s Theorem on linear forms in logarithms of algebraic numbers [2].
I Lemma 9. Let b1 , b2 be real algebraic numbers, linearly independent over Q. Furthermore,
let ϕ1 , ϕ2 be real numbers such that eiϕ1 and eiϕ2 are algebraic. Then there exist effectively
computable constants N, T > 0 such that for all t ≥ T and all k1 , k2 ∈ Z, at least one of
|b1 t − ϕ1 − 2k1 π| > 1/tN and |b2 t − ϕ2 − 2k2 π| > 1/tN holds.
The main result of the section is the following.
Pn
I Theorem 10. Let f (t) = j=1 erj t (a1,j cos(ωj t) + a2,j sin(ωj t)) be an exponential polynomial where rj , a1,j , a2,j , ωj are real algebraic numbers and the Q-span of {ω1 , . . . , ωn } has
dimension two as a Q-vector space. Then we can decide whether or not {t ∈ R≥0 : f (t) = 0} is
bounded and, if bounded, we can compute an integer T such that {t ∈ R≥0 : f (t) = 0} ⊆ [0, T ].
V. Chonev and J. Ouaknine and J. Worrell
XXX:11
Proof. Let b1 , b2 be real algebraic numbers, linearly independent over Q, such that ωj is
an integer linear combination of b1 and b2 for j = 1, . . . , n. For each n ∈ Z, sin(nb1 t) and
cos(nb1 t) can be written as polynomials in sin(b1 t) and cos(b1 t) with integer coefficients, and
similarly for b2 . It follows that we can write f in the form
f (t) = Q(er1 t , . . . , ern t , cos(b1 t), sin(b1 t), cos(b2 t), sin(b2 t))
for some polynomial Q with real algebraic coefficients that is computable from f .
Write R++ = {t ≥ 0 : sin(b1 t) ≥ 0∧sin(b2 t) ≥ 0}, R+− = {t ≥ 0 : sin(b1 t) ≥ 0∧sin(b2 t) ≤
0}, and likewise define R−+ , R−− for the two remaining sign conditions on sin(b1 t) and
sin(b2 t). We show how to decide boundedness of {t ∈ R++ : f (t) = 0}. (The cases for
R+− , R−+ , and R−− follow mutatis mutandis.) The idea is to compute a partition of
{t ∈ R++ : f (t) = 0} into components Z1 , . . . , Zm and to separately decide boundedness of
each component Zj .
Define a semi-algebraic set
E = (u, x1 , x2 ) ∈ Rn+2 : ∃y1 , y2 ≥ 0 x21 + y12 = x22 + y22 = 1 ∧ Q(u, x1 , y1 , x2 , y2 ) = 0 .
Then for t ∈ R++ we have f (t) = 0 if and only if (ert , cos(b1 t), cos(b2 t)) ∈ E, where r =
(r1 , . . . , rn ). Now consider a cell decomposition E = C1 ∪. . .∪Cm for cells C1 , . . . , Cm ⊆ Rn+2 ,
and define
Zj = {t ∈ R++ : (ert , cos(b1 t), cos(b2 t)) ∈ Cj } ,
j = 1, . . . , m,
(8)
Then {t ∈ R++ : f (t) = 0} = Z1 ∪ . . . ∪ Zm .
Now fix j ∈ {1, . . . , m}. We show how to decide boundedness of Zj . To this end, write
Dj ⊆ Rn for the projection of the corresponding cell Cj ⊆ Rn+2 on the first n coordinates.
First suppose that {t ∈ R : ert ∈ Dj } is bounded. Then by Lemma 8 we can compute an
upper bound T of this set. But Zj ⊆ {t ∈ R≥0 : ert ∈ Dj } and so Zj ⊆ [0, T ].
On the other hand, suppose that {t ∈ R : ert ∈ Dj } is unbounded. Then, by Lemma 8,
this set contains an unbounded interval (T, ∞) for some T ∈ N. Write I = [−1, 1] and define
functions g1 , g2 , h1 , h2 : Dj → R by
g1 (u) = inf{x ∈ I : ∃y (u, x, y) ∈ Cj }
g2 (u) = inf{y ∈ I : ∃x (u, x, y) ∈ Cj }
h1 (u) = sup{x ∈ I : ∃y (u, x, y) ∈ Cj }
h2 (u) = sup{y ∈ I : ∃x (u, x, y) ∈ Cj } (10)
(9)
These functions are all semi-algebraic by quantifier elimination. Hence by Lemma 8 the
limits gi∗ = limt→∞ gi (ert ) and h∗i = limt→∞ hi (ert ) exist for i = 1, 2 and are algebraic
numbers. Clearly we have g1∗ ≤ h∗1 and g2∗ ≤ h∗2 . We now consider three cases according to
the strictness of these inequalities.
4.2.1
Case I: g1∗ = h∗1 and g2∗ = h∗2 .
We show that Zj is bounded and that we can compute T2 such that Zj ⊆ [0, T2 ].
By Lemma 8 there exist T1 , ε > 0 such that for all t > T1 and i = 1, 2,
|gi (ert ) − gi∗ | < e−εt and |hi (ert ) − h∗i | < e−εt .
(11)
Then for t ∈ R++ such that t > T1 we have
t ∈ Zj ⇐⇒ ert , cos(b1 t), cos(b2 t) ∈ Cj (by (8))
=⇒
g1 (ert ) ≤ cos(b1 t) ≤ h1 (ert ) and g2 (ert ) ≤ cos(b2 t) ≤ h2 (ert ) (by (9)(10))
=⇒
|cos(b1 t) − g1∗ | < e−εt and |cos(b2 t) − g2∗ | < e−εt (by (11))
(12)
ICALP 2016
XXX:12 On the Skolem Problem for Continuous Linear Dynamical Systems
Write g1∗ = cos(ϕ1 ) and g2∗ = cos(ϕ2 ) for some ϕ1 , ϕ2 ∈ [0, π]. Since | cos(ϕ1 + x) −
cos(ϕ1 )| ≥ x3 /3 for all x sufficiently small (by a Taylor expansion), the inequality (12) implies
that for some k1 , k2 ∈ Z,
|b1 t − ϕ1 − 2k1 π| < 3e−εt/3 and |b2 t − ϕ2 − 2k2 π| < 3e−εt/3 .
(13)
Combining the upper bounds in (13) with the polynomial lower bounds |b1 t − ϕ1 − 2k1 π| >
1/tN and |b2 t − ϕ2 − 2k2 π| > 1/tN from Lemma 9 we obtain an effective bound T2 for which
t ∈ Zj implies t < T2 .
4.2.2
Case II: g1∗ < h∗1 .
In this case we show that Zj is unbounded. The geometric intuition is as follows. We imagine
a particle in the plane whose position at time t is (cos(b1 t), cos(b2 t)), together with a “moving
target” whose extent at time t is Γt = {(x, y) : (ert , x, y) ∈ Cj }. Below we essentially argue
that such a particle is bound to hit Γt at some time t since its orbit is dense in [−1, +1]2
and Γt has positive dimension in the limit.
Proceeding formally, first notice that Cj cannot be a (. . . , 0, 1)-cell or a (. . . , 0, 0)-cell,
for then we would have g1 (u) = h1 (u) for all u ∈ Dj and hence g1∗ = h∗1 . Thus Cj must
either be a (. . . , 1, 0)-cell or a (. . . , 1, 1)-cell. In either case, Cj includes a cell of the form
{(u, x, ξ(u, x)) : u ∈ D, g1 (u) < x < h1 (u)} for some semi-algebraic function ξ.
Let c, d be real algebraic numbers such that g1∗ < c < d < h∗1 . Write c = cos(ψ 0 ) and
d = cos(ψ) for 0 ≤ ψ < ψ 0 ≤ π. By Lemma 8 the limits limt→∞ ξ(ert , c) and limt→∞ ξ(ert , d)
exist and are algebraic numbers in the interval [−1, 1]. Let θ, θ0 ∈ [0, π] be such that
cos(θ) = limt→∞ ξ(ert , d) and cos(θ0 ) = limt→∞ ξ(ert , c).
0
−θ
By Corollary 3 we know that ψθ0 −ψ
is either rational or transcendental. In particular
b2
we know that it is not equal to b1 , which is algebraic and irrational. Let us suppose that
θ 0 −θ
ψ 0 −ψ
> bb21 (the converse case is almost identical). Then there exists θ00 with θ < θ00 < θ0 ,
such that
θ < θ00 +
b2 0
(ψ − ψ) < θ0 .
b1
(14)
Since 2π, b1 , b2 are linearly independent over Q it follows from Kronecker’s approximation
theorem that {(b1 t, b2 t) mod 2π : t ∈ R≥0 } is dense in [0, 2π)2 (see [13, Chapter 23]). Thus
there is an increasing sequence t1 < t2 < . . ., with b1 tn ≡ ψ mod 2π for all n, such that
0
b2 tn mod 2π converges to θ00 . Then, defining s1 < s2 < . . . by sn = tn + ψ b−ψ
, we have
1
b1 sn ≡ ψ 0 mod 2π for all n and, by (14),
lim b2 sn = lim b2 tn +
n→∞
n→∞
b2 0
b2
(ψ − ψ) = θ00 + (ψ 0 − ψ) < θ0
b1
b1
(mod 2π)
Let η(t) = ξ(ert , cos(b1 t)) − cos(b2 t). Then for t ∈ R++ such that g(ert ) < cos(b1 t) <
h(ert ),
η(t) = 0
=⇒
cos(b2 t) = ξ(ert , cos(b1 t))
=⇒
(ert , cos(b1 t), cos(b2 t)) ∈ Cj
=⇒ t ∈ Zj (by (8)) .
Now limn→∞ η(tn ) = cos(θ) − cos(θ00 ) > 0 and limn→∞ η(sn ) < cos(θ0 ) − cos(θ0 ) = 0.
Moreover for n sufficiently large we have [tn , sn ] ⊆ R++ . It follows that η(t) has a zero in
every interval [tn , sn ] for n large enough. We conclude that Zj is unbounded.
V. Chonev and J. Ouaknine and J. Worrell
4.2.3
XXX:13
Case III: g2∗ < h∗2 .
This case is symmetric to Case II and we omit details.
J
Acknowledgements The authors wish to thank Angus Macintyre for helpful comments.
Ventsislav Chonev is supported by Austrian Science Fund (FWF) NFN Grant No S11407-N23
(RiSE/SHiNE), ERC Start grant (279307: Graph Games), and ERC Advanced Grant (267989:
QUAREM). Joël Ouaknine is supported by ERC grant AVS-ISS (648701). James Worrell is
supported by EPSRC grant EP/N008197/1.
References
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Rajeev Alur. Principles of Cyber-Physical Systems. MIT Press, 2015.
Alan Baker. Transcendental number theory. Cambridge University Press, Cambridge, 1975.
Saugata Basu, Richard Pollack, and Marie-Françoise Roy. Algorithms in Real Algebraic
Geometry (Algorithms and Computation in Mathematics). Springer-Verlag, 2006.
Paul C. Bell, Jean-Charles Delvenne, Raphaël M. Jungers, and Vincent D. Blondel. The
Continuous Skolem-Pisot Problem. Theoretical Computer Science (TCS), 411(40-42):3625–
3634, 2010.
Edward Bierstone and Pierre D. Milman. Semianalytic and subanalytic sets. Publications
Mathématiques de l’Institut des Hautes Études Scientifiques, 67(1):5–42, 1988.
Richard P. Brent. Fast multiple-precision evaluation of elementary functions. Journal of
the ACM (JACM), 23(2):242–251, 1976.
Ventsislav Chonev, Joël Ouaknine, and James Worrell. On recurrent reachability for continuous linear dynamical systems. Logic in Computer Science (LICS), 2016.
Henri Cohen. A Course in Computational Algebraic Number Theory. Springer-Verlag, 1993.
Paul M. Cohn. Basic Algebra: Groups, Rings and Fields. Springer, 2002.
David A. Cox, John Little, and Donal O’Shea. Ideals, Varieties, and Algorithms: An
Introduction to Computational Algebraic Geometry and Commutative Algebra. Springer,
2007.
Richard K. Guy. Unsolved Problems in Number Theory. Springer, 2004.
Vesa Halava, Tero Harju, Mika Hirvensalo, and Juhani Karhumäki. Skolem’s Problem – on
the border between decidability and undecidability. TUCS Technical Reports, 683, 2005.
Godfrey H. Hardy and Edward M. Wright. An Introduction to the Theory of Numbers,
volume 1. Oxford, 1999.
Bettina Just. Integer relations among algebraic numbers. In Mathematical Foundations of
Computer Science (MFCS), volume 379, pages 314–320. Springer, 1989.
Serge Lang. Introduction to Transcendental Numbers. Reading, Mass., 1966.
Arjen K. Lenstra. Factoring multivariate polynomials over algebraic number fields. SIAM
Journal on Computing, 16(3):591–598, 1987.
Angus Macintyre. Turing meets Schanuel. Preprint, to appear in the proceedings of Logic
Colloquium, 2012.
Angus Macintyre and Alex J. Wilkie. On the decidability of the real exponential field. In
(ed. Piergiorgio Odifreddi) Kreiseliana: About and Around Georg Kreisel., 1996.
Wolfgang Schmidt. Diophantine Approximation, volume 785 of Lecture Notes in Mathematics. Springer-Verlag Berlin Heidelberg, 1980.
Terence Tao. Structure and randomness: pages from year one of a mathematical blog.
American Mathematical Society, 2008.
Boris Zilber. Exponential sums equations and the Schanuel conjecture. Journal of the
London Mathematical Society, 65:27–44, 2002.
ICALP 2016
XXX:14 On the Skolem Problem for Continuous Linear Dynamical Systems
22
Boris Zilber. Pseudo-exponentiation on algebraically closed fields of characteristic zero.
Annals of Pure and Applied Logic, 132(1):67 – 95, 2005.
V. Chonev and J. Ouaknine and J. Worrell
A
XXX:15
Hardness
In this section we show that decidability of the Continuous Skolem Problem entails significant
new effectiveness results in Diophantine approximation, thereby identifying a formidable
mathematical obstacle to further progress in the unbounded case.
Diophantine approximation is a branch of number theory concerned with approximating
real numbers by rationals. A central role is played in this theory by the notion of continued
fraction expansion, which allows to compute a sequence of rational approximations to a given
real number that is optimal in a certain well-defined sense. For our purposes it suffices to
note that the behaviour of the simple continued fraction expansion of a real number a is
closely related to the (homogeneous Diophantine approximation) type of a, which is defined
to be
n
o
n
c
L(a) := inf c : a −
< 2 for some m, n ∈ Z .
m
m
Let [n1 , n2 , n3 , . . .] be the sequence of partial quotients in the simple continued fraction
expansion of a. Then, writing K(a) := supk≥0 nk , it is shown in [19, pp. 22-23] that L(a) = 0
if and only if K(a) is infinite and otherwise
K(a) ≤ L(a)−1 ≤ K(a) + 2 .
It is well known that a real number algebraic number of degree two over the rationals
has a simple continued fraction expansion that is ultimately periodic. In particular, such
numbers have bounded partial quotients. But nothing is known about real algebraic numbers
of degree three or more—no example is known with bounded partial quotients, nor with
unbounded quotients. Guy [11] asks:
Is there an algebraic number of degree greater than two whose simple continued fraction
expansion has unbounded partial quotients? Does every such number have unbounded
partial quotients?
In other words, the question is whether there is a real algebraic number a of degree at least
three such that L(a) is strictly positive, or whether L(a) = 0 for all such a.
Recall that a real number x is computable if there is an algorithm which, given any
rational ε > 0 as input, returns a rational q such that |q − x| < ε. The main result of
this section is Theorem 14, which shows that the existence of a decision procedure for the
general Continuous Skolem Problem entails the computability of L(a) for all real algebraic
numbers a. Now one possibility is that all such numbers L(a) are zero, and hence trivially
computable. However the significance of Theorem 14 is that in order to prove the decidability
of the Continuous Skolem Problem one would have to establish, one way or another, the
computability of L(a) for every real algebraic number a.
Fix positive real algebraic a, c and define the functions:
f1 (t)
= et (1 − cos(t)) + t(1 − cos(at)) − c sin(at),
f2 (t)
= et (1 − cos(t)) + t(1 − cos(at)) + c sin(at),
f (t)
= et (1 − cos(t)) + t(1 − cos(at)) − c| sin(at)| = min{f1 (t), f2 (t)}.
Then f1 (t) and f2 (t) are exponential polynomials. Moreover it is easy to check that the
function f (t) has a zero in an interval of the form (T, ∞) if and only if at least one of f1 (t),
f2 (t) has a zero in (T, ∞).
We will first prove two lemmas which show a connection between the existence zeros of
f (t) and the type L(a). We then will derive an algorithm to compute L(a) using an oracle
for the Continuous Skolem Problem, thereby demonstrating our desired hardness result.
ICALP 2016
XXX:16 On the Skolem Problem for Continuous Linear Dynamical Systems
I Lemma 11. Fix real algebraic a, c and ε ∈ Q with a, c > 0 and ε ∈ (0, 1). There exists
an effective threshold T , dependent on a, c, ε, such that if f (t) = 0 for some t ≥ T , then
L(a) ≤ c/2π 2 (1 − ε).
Proof. Suppose f (t) = 0 for some t ≥ T . Define δ1 = t − 2πm and δ2 = at − 2πn, where
m, n ∈ N and δ1 , δ2 ∈ [−π, π). Then we have
a−
|δ2 − aδ1 |
n
=
.
m
2πm
We will show that for T chosen large enough, if f (t) = 0 for t ≥ T then we can bound |δ2 |
and |aδ1 | separately from above and then apply the triangle inequality to bound |δ2 − aδ1 |,
obtaining the desired upper bound on L(a).
t−π
T −π
Define 0 < α < 1 by α2 = (1 − ε2 ). Since m ≥
≥
, for sufficiently large T we
2π
2π
have
t ≥ 2π(m − 1) ≥ 2πmα .
(15)
Furthermore, since αx2 /2 ≤ 1 − cos(x) for |x| sufficiently small, we may assume that T is
large enough such that the following is valid for |x| ≤ π:
if 1 − cos(x) ≤ cπ/T then αx2 /2 ≤ 1 − cos(x).
(16)
We have the following chain of inequalities, where (∗) follows from f (t) = 0 and et (1 −
cos(t)) ≥ 0:
(∗)
1 − cos(δ2 ) = 1 − cos(at) ≤
c| sin(at)|
c| sin(δ2 )|
c|δ2 |
=
≤
.
t
t
t
It follows that 1 − cos(δ2 ) ≤ cπ/t and so by (16) we also have
αδ22
≤ 1 − cos(δ2 ) .
2
Combining the upper and lower bounds on 1 − cos(δ2 ) and using (15), we have
|δ2 | ≤
2c
2c
c
≤
.
=
αt
2πmα2
mπ(1 − ε2 )
We next seek an upper bound on |δ1 |. To this end, let T be large enough so that
ce−t ≤
cε 2
for t ≥ T .
2aαt
(17)
Then the following chain of inequalities holds:
δ12
≤ 1 − cos(δ1 )
16
c| sin(δ2 )| − t(1 − cos(δ2 ))
=
et
≤ ce−t
cε 2
≤
2aαt
cε 2
≤
4aπα2 m
{ valid for all |δ1 | ≤ π }
{ since f (t) = 0 }
{ since | sin(δ2 )|, | cos(δ2 )| ≤ 1}
{ by (17) }
{ by (15) }
V. Chonev and J. Ouaknine and J. Worrell
XXX:17
It follows that
|aδ1 | ≤
cε
.
πm(1 − ε2 )
Finally, by the triangle inequality and the bounds on |aδ1 | and |δ2 |, we have
a−
|δ2 − aδ1 |
n
|δ2 | + |aδ1 |
c + cε
c
=
≤
≤
=
,
m
2πm
2πm
2π 2 m2 (1 − ε2 )
2π 2 m2 (1 − ε)
so the natural numbers n, m witness L(a) ≤ c/2π 2 (1 − ε).
J
I Lemma 12. Fix real algebraic a, c and ε ∈ Q with a, c > 0 and ε ∈ (0, 1). There exists
an effective threshold M , dependent on a, c, ε, such that if L(a) ≤ c(1 − ε)/2π 2 holds and is
witnessed by natural numbers n, m with m ≥ M , then f (t) = 0 for some t ≥ 2πM .
Proof. Select M large enough, so that c(1 − ε)/πM < π and
if |x| < c(1 − ε)/πM , then (1 − ε)|x| ≤ | sin(x)|.
(18)
Suppose now that L(a) ≤ c(1 − ε)/2π 2 , let this be witnessed by n, m ∈ N with m ≥ M and
define t := 2πm. We will show that f (t) ≤ 0. This suffices, because f (t) is continuous and
moreover is positive for arbitrarily large times, so it must have a zero on [t, ∞).
Since L(a) ≤ c(1 − ε)/2π 2 , we have |am − n| ≤ c(1 − ε)/2π 2 m. Therefore, we can write
at = 2πam = 2πn + δ for some δ satisfying |δ| ≤ c(1 − ε)/πm < π. We have
f (t)
= { as cos(t) = 1 }
t(1 − cos(δ)) − c| sin(δ)|
≤ { by (18) and 1 − cos(x) ≤ x2 /2 }
πmδ 2 − c(1 − ε)|δ|
≤ { by |δ| ≤ c(1 − ε)/πm }
0.
J
The following corollary is immediate:
I Lemma 13. Fix real algebraic a, c and ε ∈ Q with a, c > 0 and ε ∈ (0, 1). There exists
an effective threshold T , dependent on a, c, ε, such that if f (t) 6= 0 for all t ≥ T , then either
L(a) < c(1 − ε)/2π 2 and this is witnessed by natural numbers n, m with m < T /2π, or
L(a) ≥ c(1 − ε)/2π 2 .
We now use the above lemmas to show the central result of this section:
I Theorem 14. Fix a positive real algebraic number a. If the Continuous Skolem Problem
is decidable then L(a) may be computed to within arbitrary precision.
Proof. Suppose we know L(a) ∈ [p, q] for non-negative p, q ∈ Q. Choose a real algebraic c
with c > 0 and a rational ε ∈ (0, 1) such that
p<
c(1 − ε)
c
<
< q.
2π 2
2π 2 (1 − ε)
ICALP 2016
XXX:18 On the Skolem Problem for Continuous Linear Dynamical Systems
Write A := c(1 − ε)/2π 2 and B := c/2π 2 (1 − ε). Calculate the maximum of the thresholds T
required by Lemmas 11 and 13. Check for all denominators m ≤ T /2π whether there exists
a numerator n such that n, m witness L(a) ≤ A. If so, then continue the approximation
procedure recursively with confidence interval [p, A]. Otherwise, use the oracle for the
Continuous Skolem Problem to determine whether at least one of f1 (t), f2 (t) has a zero on
[T, ∞). If this is the case, then f (t) also has a zero on [T, ∞), so by Lemma 11, L(a) ≤ B
and we continue the approximation recursively on the interval [p, B]. If not, then L(a) ≥ A
by Lemma 13, so we continue on the interval [A, q]. Notice that in this procedure, one can
choose c, ε at each stage in such a way that the confidence interval shrinks by at least a fixed
factor, whatever the outcome of the oracle invocations. It follows therefore that L(a) can be
approximated to within arbitrary precision.
J
We conclude by remarking that the exponential polynomials f1 and f2 involved in the
proof of Theorem 14 involve only two rationally linearly independent frequencies. Thus the
theorem applies as soon as we have a decision procedure for exponential polynomials with
two rationally linear independent frequencies.
B
Proof of Lemma 8
We divide Lemma 8 into two separate results, which are proven below.
I Proposition 15. There is a procedure that, given a semi-algebraic set D ⊆ Rn+1 and real
algebraic numbers r1 , . . . , rn , returns an integer T such that {t ≥ 0 : (t, er1 t , . . . , ern t ) ∈ D}
either contains the interval (T, ∞) or is disjoint from (T, ∞). The procedure also decides
which of these two eventualities is the case.
Proof. Consider a non-zero polynomial P ∈ R[x0 , . . . , xn ] whose coefficients are real algebraic
numbers. Then we can write P (t, er1 t , . . . , ern t ) in the form
Q1 (t)eβ1 t + . . . + Qm (t)eβm t
for non-zero univariate polynomials Q1 , . . . , Qm with real algebraic coefficients and real
algebraic numbers β1 > . . . > βm . It is clear that for t sufficiently large, P (t, er1 t , . . . , arn t )
has the same sign as the leading term Q1 (t). The proposition easily follows.
J
I Lemma 16. Let g : D → R be a bounded semi-algebraic function with domain D ⊆ Rn .
Let r = (r1 , . . . , rn ) be a tuple of real algebraic numbers and T1 an integer such that
ert = (er1 t , . . . , ern t ) ∈ D for all t > T1 . Then the limit g ∗ = limt→∞ g(ert ) exists, is an
algebraic number, and there are effective constants T2 , ε > 0 such that |g(ert ) − g ∗ | < e−εt
for all t > T2 .
Proof. Since g is semi-algebraic, there is a non-zero polynomial P with real algebraic
coefficients such that P (x, g(x)) = 0 for all x ∈ D (see [3, Proposition 2.86]). In particular,
we have P (ert , g(ert )) = 0 for all t > T1 . Gathering terms, we can rewrite this equation in
the form
Q1 (g(ert ))eβ1 t + . . . + Qm (g(ert ))eβm t = 0
for non-zero univariate polynomials Q1 , . . . , Qm with real-algebraic coefficients and real
algebraic numbers β1 > . . . > βm .
If m = 1 then for all t > T1 we have Q1 (g(ert )) = 0. Thus g(ert ) is equal to some of root
of Q for all t > T1 . Then by Proposition 15 there exists T2 such that g(ert ) = g ∗ for some
fixed root g ∗ of Q1 and all t > T2 .
V. Chonev and J. Ouaknine and J. Worrell
XXX:19
If m > 1, since g is a bounded function, for all t > T1 we have
Q1 (g(ert ))
=
≤
Q2 (g(ert ))e(β2 −β1 )t + . . . + Qm (g(ert ))e(βm −β1 )t
M e(β2 −β1 )t
(19)
for some constant M . If Q1 has degree d then (19) implies that the closest root of Q1 to
g(ert ) has distance at most (M eβ2 −β1 )1/d . Hence there exists a root g ∗ of Q1 and effective
constants ε, T2 > 0 such that |g(ert ) − g ∗ | < e−εt for all t > T2 .
J
C
One Linearly Independent Frequency
Pn
I Theorem 17. Let f (t) = j=1 erj t (P1,j (t) cos(ωj t) + P2,j sin(ωj )) be an exponential polynomial such that ω1 , . . . , ωn are all integer multiples of single real algebraic number b. Then
we can decide whether or not {t ∈ R≥0 : f (t) = 0} is bounded and, if bounded, we can
compute an integer T such that {t ∈ R≥0 : f (t) = 0} ⊆ [0, T ].
Proof. Recall that for each integer n, both cos(nbt) and sin(nbt) can be written as polynomials
in sin(bt) and cos(bt) with integer coefficients. It follows that we can write f in the form
f (t) = Q(t, er1 t , . . . , ern t , cos(bt), sin(bt)) ,
for some multivariate polynomial Q with real algebraic coefficients that is computable from
f.
Write R≥0 as the union of R+ := {t ∈ R≥0 : sin(bt) ≥ 0} and R− := {t ∈ R≥0 : sin(bt) ≤
0}. We show how to decide boundedness of {t ∈ R+ : f (t) = 0}. The analogous case for
R− follows mutatis mutandis. The idea is to partition {t ∈ R+ : f (t) = 0} into components
Z1 , . . . , Zm and to show for how to decide boundedness of each Zj .
To this end define a semi-algebraic set
E := (u, x) ∈ Rn+2 : ∃y ≥ 0 x2 + y 2 = 1 ∧ Q(u, x, y) = 0 .
Then for all t ∈ R+ we have f (t) = 0 if and only if (t, eat , cos(bt)) ∈ E.
Consider a cell decomposition E = C1 ∪ . . . ∪ Cm , for cells C1 , . . . , Cm ⊆ Rn+2 , and define
Zj = {t ∈ R+ : t, ert , cos(bt) ∈ Cj }
for j = 1, . . . , m. Then {t ∈ R+ : f (t) = 0} = Z1 ∪ . . . ∪ Zm .
We now focus on Zj for some fixed j ∈ {1, . . . , m} and show how to decide whether or
not Zj is finite. Write Dj ⊆ Rn+1 for the projection of the corresponding cell Cj ⊆ Rn+2
onto the first n + 1 coordinates. Then Zj ⊆ {t ∈ R≥0 : (t, ert ) ∈ Dj }.
First, suppose that {t ∈ R≥0 : (t, ert ) ∈ Dj } is bounded. Since Dj is semi-algebraic, by
Proposition 15 we can compute an upper bound T of this set. In this case t < T whenever
t ∈ Zj .
On the other hand, suppose that {t ∈ R : (t, ert ) ∈ Dj } is unbounded. Then, by
Proposition 15, this set contains an unbounded interval (T, ∞). We claim that in this case
Zj must be unbounded. There are two cases according to the nature of the cell Cj .
Case I: Cj is a (. . . , 0)-cell.
In this case there is a continuous semi-algebraic function g : Dj → R such that Cj =
{(u, g(u)) : u ∈ Dj }. Then for t ∈ R+ ∩ (T, ∞),
g(t, ert ) = cos(bt) ⇐⇒
⇐⇒
(t, ert , cos(bt)) ∈ Cj
f (t) = 0 .
ICALP 2016
XXX:20 On the Skolem Problem for Continuous Linear Dynamical Systems
1
g(t, ert )
π
b
2π
b
3π
b
4π
b
5π
b
t
−1
T
Figure 1 Intersection points of g(t, ert ) and cos(bt) for t ∈ R+ (with R+ shown in bold).
In other words, f has a zero at each point t ∈ R+ ∩ (T, ∞) at which the graph of g(t, ert )
intersects the graph of cos(bt). Since g is a continuous function with an unbounded domain
that takes values in [−1, 1], there are infinitely many such intersection points–see Figure 1.
Case II: Cj is a (. . . , 1)-cell.
In this case Cj contains some (. . . , 0)-cell and so the argument in Case I shows that f has
infinitely many zeros in Zj .
J
| 3cs.SY
|
Object-oriented Programming Laws for Annotated Java
Programs
Gabriel Falconieri Freitas
Márcio Cornélio
Universidade de Pernambuco (UPE), Brasil
Universidade de Pernambuco (UPE), Brasil
[email protected]
Universidade Federal Rural de Pernambuco (UFRPE), Brasil
[email protected]
Tiago Massoni
Rohit Gheyi
Universidade Federal de Campina Grande (UFCG), Brasil
{massoni, rohit}@dsc.ufcg.edu.br
Object-oriented programming laws have been proposed in the context of languages that are not combined with a behavioral interface specification language (BISL). The strong dependence between
source-code and interface specifications may cause a number of difficulties when transforming programs. In this paper we introduce a set of programming laws for object-oriented languages like Java
combined with the Java Modeling Language (JML). The set of laws deals with object-oriented features taking into account their specifications. Some laws deal only with features of the specification
language. These laws constitute a set of small transformations for the development of more elaborate
ones like refactorings.
1
Introduction
Software changes constantly due to maintenance that leads to correction of fails or just to improve functionalities. However, some changes can take place to achieve quality factors like reuse and legibility. In
these cases, changes should not alter the software behavior but only its internal structure. Improving the
internal software structure is an activity known as refactoring [8]. To avoid errors due to modifications,
every change has to be done following a discipline.
This discipline can be achieved by programming laws, as guidelines to informal programming practices, establishing a basis for formal and rigorous program development. They are largely known for
imperative programming [12, 19]. Also, functional programming and logic programming have a set of
laws described by Bird and de Moor [2] and Seres [20], respectively. Laws of object-oriented programming have also been addressed in [3, 7, 6].
Design by Contract (DbC) [18] is a development methodology that aims at the construction of reliable
object-oriented systems. Its basic idea is that a contract is established among classes of a system. In this
way, software developers should formally specify what is required and ensured by methods and types.
The Java Modeling Language (JML) [16, 14] is a notation for formally specifying the behavior of Java
classes and methods.
The set of programming laws for object-oriented programming we have nowadays is designed for
program transformation with no relation to specifications languages designed for DbC. Changes in specification usually should discharge code updates, maintaining the conformance between code and specification. On the other hand, changes in program code may require changes in specifications as the
behavior implemented by code may diverge from the meaning of the original specification. For instance,
moving a redefined method to its superclass can be illegal if this transformation causes weakening of
I. Mackie and A. Martins Moreira (Eds.): Tenth International
Workshop on Rule-Based Programming (RULE 2009)
EPTCS 21, 2010, pp. 65–76, doi:10.4204/EPTCS.21.6
© Gabriel Falconieri Freitas et al.
This work is licensed under the Creative Commons
Attribution-Noncommercial-No Derivative Works License.
Object-oriented Programming Laws for Annotated Java Programs
66
1
2
3
4
p u b l i c c l a s s PositiveIntegerData {
/ / @ p r i v a t e i n v a r i a n t value . intValue ( ) > −1;
p r i v a t e Integer value ;
p u b l i c PositiveIntegerData ( ) { value = new Integer ( 0 ) ; }
5
/ * @ r e q u i r e s newValue != n u l l && newValue . intValue ( ) > −1;
@ e n s u r e s getValue ( ) . intValue ( ) == newValue . intValue ( ) ; @ * /
p u b l i c v o i d registerValue ( Integer newValue ) { /* . . . */ }
6
7
8
9
/ / @ e n s u r e s \ r e s u l t != n u l l ;
p u b l i c / * @ pure @ * / Integer getValue ( ) { /* . . . */
10
11
}
12
/ * @ r e q u i r e s getValue ( ) != n u l l ;
@ e n s u r e s ! ( \ r e s u l t ) . equals ( ” ” ) ; @ * /
p u b l i c String format ( ) { /* . . . */ }
13
14
15
16
}
Figure 1: Class PositiveIntegerData
pre-conditions and strengthening of post-conditions. Transformation of object-oriented programs with
formal contracts has already been addressed in a rather informal way [10].
A catalogue of laws (primitive transformations) to deal with Java programs annotated with JML has
been proposed in [9], which specifies about 80 laws. Here we present one law that deals only with JML
specifications and two JML-aware Java laws that deals with attributes and methods, respectively. A law
that only deals with JML can impose conditions only on JML elements present in the program, whereas
JML-aware Java laws involve both JML and Java elements for stating conditions. The three laws we
present here are catalogued in [9].
In this paper, we define laws (Section 3) of object-oriented programming for Java that are aware of
specifications written in JML, which we describe in Section 2. The laws we present here and other ones
present in a more comprehensive catalogue [9] were applied to refactoring a JML-specified version of
a core module of a Manufacturing Execution System (MES) [22]. In Section 4, we present proof of
soundness regarding the JML parts of two laws. We present an example of program transformation by
means of laws in Section 5. Final remarks appear in Section 6.
2
The Java Modeling Language
The Java Modeling Language (JML) is a behavioral interface specification language (BISL) [16, 14]
tailored to Java [11]. Thus, JML serves to describe names and static information that appear in Java
declarations and how they act, how they behave. JML specifications are written in the form of special
annotation comments that are inserted directly in source code of programs. These comments must begin
with an at-sign (@) and can be written in two ways: by using //@ ... or /*@ ... @*/.
In Figure 1, we present the class PositiveIntegerData that represents positive integers. We introduce an instance invariant (Line 2), which is a predicate that is true in all visible states of objects of a
class [16]. The invariant in the example has private visibility and establishes that the attribute value must
always be greater than -1.
JML uses the requires clause to specify the obligations of the caller of a method, what must be true
Gabriel Falconieri Freitas et al.
Law 1 hmove invariant to superclassi
class B extends A {
//@ invariant ψ1 ;
ads cnts mds
}
class C extends B {
//@ invariant ψ′1 && ψ2 ;
cnts′ mds′
}
67
=cds,Main
class B extends A {
//@ invariant ψ1 && ψ2 ;
ads cnts mds
}
class C extends B {
//@ invariant ψ′1 ;
cnts′ mds′
}
where
ψ2 ≡ this instanceof C ==> ψinv
provided
(↔) super does not appear in ψ2 .
(→) ψ2 does not contain occurrences of model fields declared in C, nor uncast occurrences of this.
to call a method. For instance, the precondition of the method registerValue requires the value of the
Integer object to be registered to be greater than -1.
A postcondition specifies the implementor’s obligation, what must be true at the end of a method,
just before it returns to the caller. In JML, the ensures clause introduces a postcondition. For instance, the
Line 7 introduces a normal postcondition that asserts that the final value of the Integer object we register
is the same as the one the method receives as argument. The JML modifier pure (Line 11) indicates that
the method doesn’t have any side effects and hence can appear in specifications. In JML, the keyword
also indicates that a method is extending the specification it inherits from its supertype.
3
Laws
Our laws extend object-oriented programming laws from other works [3, 6, 5, 7, 17]. The laws are
written in an equational style. Each side of the equation corresponds to a template of a well-formed
program. Programming laws, in which left-hand and right-hand sides are related by equality, are a
concise presentation of a pair of laws. These laws precisely indicate the modifications that can be done
to a program, stating their corresponding side-conditions. In fact, to apply a law, it is necessary to check
(syntactic or semantic) side-conditions to ensure that the transformation is behavior-preserving and also
maintains its well-formedness. We consider that we are dealing with only one package and working in a
limited open system [7], in which classes of our system can depend on external libraries.
In Java and JML context, we need to guarantee that source-code continues meeting its specifications written in JML, taking into account the semantics of JML specifications along with the notion of
specification inheritance [13]. Here we present a law for invariants written in JML.
A JML-annotated Java program has the format cds Main, where cds is the set of all classes of the
program and Main corresponds to the unique class in the program which has a static main method.
We use cnts, ads and mds inside a class to represents the class constructors, attributes and methods,
respectively. We have to emphasize that they may contain the specifications of each constructor or
68
Object-oriented Programming Laws for Annotated Java Programs
Law 2 hmove reference type attribute to superclassi
class B extends A {
ads cnts mds
}
=cds,Main
class C extends B {
/*@ nullable @*/ T a;
ads′ cnts′ mds′
}
class B extends A {
/*@ nullable @*/ T a;
ads cnts mds
}
class C extends B {
ads′ cnts′ mds′
}
provided
JML:
(←) D.a, for any D ≤ B and D C does not occur inside specifications of cds, Main, cnts, cnts’, mds
nor mds’.
Java:
(↔) T is not a primitive type.
(→) (1) a is not declared in ads; (2) The attribute name a is not declared by the subclasses of B in cds.
(←) D.a, for any D ≤ B e D C does not occur in cds, Main, cnts, cnts’, mds nor mds’.
method. It is not necessarily Java code only, we can also have the corresponding JML specifications.
In the laws, we use cds1 =cds,Main cds2 to denote the equivalence of sets of class declarations cds1 and
cds2 , where cds is a context of class declarations for cds1 and cds2 . We need to stress that this definition
takes into account only sequential programs. We write ‘→’ to indicate the condition that need to be
satisfied to apply a law from left to right. Likewise, we use ‘←’ to indicate what has to be satisfied when
applying the law from right to left. We use ‘↔ to indicate conditions that must hold in both directions.
The first law we present (Law 1) allows us to move an invariant ψ2 from a subclass C to its superclass
B. The invariant we want to move only refers to instances of C as we require the invariant to be applicable
only to instances of class C. To apply this law in any direction, we require that calls to super do not
occur in ψ2 , since after law application (in both directions) these calls may refer different elements. To
apply this law from left to right, model fields cannot appear in ψ2 and occurrences of this must be cast
otherwise the elements they refer may not be visible.
Concerning the soundness of this law, we take in account the inheritance of specifications in
JML [13], in which inherited invariants are conjoined with locally added invariants. On the left-hand
side, the invariant ψ2 , which is present in class C, is inherited by the subclasses of C and holds for all
subclasses. On the right-hand side of the law, the invariant ψ2 is inherited by all subclasses of B besides
those that are not subclasses of C. For those classes that are subclasses of B, but not subclasses of C,
the invariant holds because for objects of these classes the antecedent instanceof C fails and the whole
implication is true, not changing the meaning of any original local invariant that inherits ψ2 .
By using Law 2, we can move an attribute to a superclass if it is not already declared in the superclass
and if it does not cause name conflicts. The application of Law 2, from right to left, allows us to move an
attribute downward. In this case, we prevent access to the attribute by the expression this, and we allow
only accesses to a by C or subclasses of C, including accesses that appear in specifications.
Gabriel Falconieri Freitas et al.
69
In Law 2, we consider only attributes whose type is a reference type. There is another law for moving
an attribute of primitive type. The reason for having two disctinct laws for dealing with attributes of
primitive and reference types comes from the nullable keyword in Law 2. In JML, any declaration
(except for local variables) whose type is a reference type is implicitly declared to be non-null, except
when one adorns the declaration with a nullable annotation [16]. Thus, by default, JML always checks if
a not nullable attribute is null in all visible states of the class that declares it. When we move an attribute
to a superclass, this is not aware about the newly moved attribute and, therefore, this action can cause
a undesirable behavior. In fact, if one instantiates the superclass, JML will raise an invariant exception
reporting that the new attribute is null. To avoid this, we force attribute nullability to move it up. Then,
if one wants to move a non-null a attribute, one needs to introduce nullable annotation before moving
it. An attribute can become nullable applying a law named make attribute nullable, not presented here.
Remember that in Java only reference types can be null.
Law 3 allows us to move a redefined method from a class to its superclass. The proviso concerning
super is needed because its semantics may be affected when we move it from a subclass to a superclass,
or vice-versa. We can only move the specification of a method if it does not refer to model fields of the
class in which the method is originally declared. Furthermore, this expressions may occur in the target
method specifications only if they are cast. In fact, as in the law the method has default visibility, only
non-private elements can be referenced in its pre- and postconditions. This is similar to Java: the this
expression may appear in mbody’ if they have a cast and they mention only non-private attributes or
methods of class C. The right-hand side of Law 3 introduces type tests in each one of the specifications.
In this way we assure that the original pre- and postconditions of the redefined method of C will only be
applied to callers that are instances of C or instances of any of C’s subclasses.
Law 3 hmove redefined method to superclass: overriden method with non-default specification casei
class B extends A {
ads cnts mds
//@ requires θ1 && ψ1 ;
class B extends A {
//@ ensures θ1 && ψ2 ;
ads cnts mds
//@ also
//@ requires ψ1 ;
//@ requires θ2 && ψ′1 ;
//@ ensures ψ2 ;
//@ ensures θ2 && ψ′2 ;
rt m(pds) { mbody }
//@ also
}
//@ requires θ2 && ψ1 ;
class C extends B {
=cds,Main
//@
ensures θ2 && ψ2 ;
ads′ cnts′ mds′
rt m(pds) {
//@ also
if (!(this instanceof C))
//@ requires ψ′1 ;
{ mbody } else { mbody′ }
//@ ensures ψ′2 ;
}
rt m(pds) { mbody′ }
}
}
class C extends B {
ads′ cnts′ mds′
}
where
θ1 ≡ !(this instanceof C) and θ2 ≡ this instanceof C
Object-oriented Programming Laws for Annotated Java Programs
70
provided
JML:
(↔) super does not appear in ψ′1 nor in ψ′2 .
(→) Both ψ1 and ψ2 do not contain occurrences of model fields declared in C, nor uncast occurrences
of this.
Java:
(↔) (1) super and private attributes do not appear in mbody′ ; (2) super.m does not appear in mds’
(→) mbody′ does not contain uncast occurrences of this nor expressions of the form ((C)this).a and
of the form ((C)this).m(e) for any attribute a nor method m, in ads′ and mds′ , respectively, with
private visibility.
(←) m(pds) is not declared in mds′ .
4
Soundness
The proofs we present here are only concerned with the JML parts of the laws. Proving the Java part
is difficult due to aliasing, which can lead to representation exposure problems, for instance. In JML,
specifications present in a class are inherited by its subclasses, provided they are not private. This leads
us to two concepts: join of specifications and specification inheritance.
4.1 Join of specifications
In a program written in Java and annotated with JML, classes inherit not only attributes and methods
from superclasses, they also inherit specifications of invariants, methods, history constraints, and initialisation predicates [13, 15]. Concerning methods, a method specification may consist of several specifications cases, which are introduced by the use of clauses such as requires, assignable , ensures [16]. Each
specification case has a precondition (the default predicate is true) that states when the corresponding
specification case applies to a call. The keyword also joins specifications cases. When a precondition of
a specification case holds, the corresponding postcondition must hold also. defined earlier in this section.
The definitions we present here are taken from [15]. The notation T ⊲ (pre, post) is related to a specification case of an instance method that type checks when its receiver ( this ) has static type T . It also type
checks in contexts where this has some subtype of T . In what follows, we introduce the definition of
joint JML method specifications [15].
Definition 1 (Join of JML method specifications) Let T ′ ⊲ (pre′ , post′ ) and T ⊲ (pre, post) be specifications of an instance method m. Let U be a subtype of both T ′ and T . Then the join of (pre′ , post′ ) and
(pre, post) for U, written (pre′ , post′ ) ⊔U (pre, post), is the specification U ⊲ (p, q) with precondition p:
pre || pre′
and postcondition q:
(\ old(pre’) ==> post’) && (\ old(pre) ==> post)
In Definition 1, the precondition of joint method specifications is their disjunction. The postcondition of the join is a conjunction of implications (written ==> in JML’s notation), stating that when a
precondition holds (in the pre-state), the corresponding postcondition must hold.
Gabriel Falconieri Freitas et al.
71
ext invBLHS
= [by Definition 2 ]
V
{added invU | U ∈ supers(BLHS )}
= [by set theory]
V
{added invU | U ∈ ((supers(BLHS ) \ supers(A)) ∪ supers(A)}
= [by definition of conjunction]
V
V
( {added invU | U ∈ ((supers(BLHS ) \ supers(A))}) ∧ ( {added invW | W ∈ supers(A)})
= [by definition of added invariant in BLHS ]
V
ψ1 ∧ ( {added invW | W ∈ supers(A)})
= [by Propositional Logic]
V
ψ1 ∧ true ∧ ( {added invW | W ∈ supers(A)})
= [by Propositional Logic]
V
ψ1 ∧ ( f alse ⇒ ψinv ) ∧ ( {added invW | W ∈ supers(A)})
= [by type test for object of type B]
V
ψ1 ∧ (this instanceo f C ⇒ ψinv ) ∧ ( {added invW | W ∈ supers(A)})
= [by definition of added invariant in BRHS ]
V
V
( {added invU | U ∈ ((supers(BRHS ) \ supers(A))}) ∧ ( {added invW | W ∈ supers(A)})
= [by definition of conjunction]
V
{added invU | U ∈ ((supers(BRHS ) \ supers(A)) ∪ supers(A)}
= [by set theory]
V
{added invU | U ∈ supers(BLHS )}
= [by Definition 2 ]
B
ext invRHS
Figure 2: Proof of Law 1 - case of object of exact type B
4.2 Specification Inheritance
Subtypes in JML inherit specifications, besides attributes and methods. First, we introduce some notation
for type specification. For a type T , the invariant predicate declared in the specification of T (without
inheritance) is denoted by added invT . For a method m declared in a type T , the notation added specTm =
T ) is the join of the specification cases in type T for m. If m is declared in T
(added preTm , added postm
with no specification and is not overriding any method, then added specTm = (true, true), which is the
default specification in JML. We use supers(T ) to denote the set of all supertypes of T (including T ) and
methods(T ) to denote the set of all instance method names declared in the specifications of the types in
a set T .
Definition 2 (Extended specification) Suppose T has supertypes supers(T ), which includes T itself.
Then the extended specification of T is a specification such that:
methods: for all methods m ∈ methods(supers(T )), the extended specification of m is the join of all
added specifications for m in T and all its proper supertypes
F
ext specTm = T {added specU
m | U ∈ supers(T )}
invariant: the extended invariant of T is the conjunction of all added invariants in T and its proper
supertypes
V
ext intT = T {added invU | U ∈ supers(T )}
The definitions we present here were introduced in [15] and are the ones we use in this paper.
Object-oriented Programming Laws for Annotated Java Programs
72
BLHS
ext specm
= [by Definition 2 ]
FB
U
LHS {added specm | U ∈ supers(B)}
= [by set theory]
FB
U
LHS {added specm | U ∈ (supers(B) \ supers(A)) ∪ supers(A)}
= [by definition of join with respect to B]
F
B FA {added specW | W ∈ supers(A)})
( BLHS {added specU
m | U ∈ (supers(B) \ supers(A))}) ⊔ (
m
= [by definition of join of specification cases for BLHS ]
F
(ψ1 , \old(ψ1 ) ⇒ ψ2 ) ⊔B ( A {added specW
m | W ∈ supers(A)})
= [by Propositional Logic]
F
((ψ1 ∧ true), \old(ψ1 ∧ true) ⇒ ψ2 ) ⊔B ( A {added specW
m | W ∈ supers(A)})
= [by type test for object of type B]
((ψ1 ∧ ¬(this instanceo f C)), \old(ψ1 ∧ ¬(this instanceo f C)) ⇒ ψ2 )
F
⊔B ( A {added specW
m | W ∈ supers(A)})
= [by Propositional Logic]
((ψ1 ∧ ¬(this instanceo f C)) ∨ f alse ∨ f alse, \old(ψ1 ∧ ¬(this instanceo f C)) ⇒ ψ2 ) ∧ true ∧ true)
F
⊔B ( A {added specW
m | W ∈ supers(A)})
= [by type test for object of type B and Propositional Logic]
((ψ1 ∧ ¬(this instanceo f C)) ∨ ((this instanceo f C) ∧ ψ1 ) ∨ ((this instanceo f C) ∧ ψ′1 ),
(\old(ψ1 ∧ ¬(this instanceo f C)) ⇒ ψ2 ) ∧ (\old((this instanceo f C) ∧ ψ1 ) ⇒ ψ2 ))
F
∧(\old((this instanceo f C) ∧ ψ′1 ) ⇒ ψ′2 )) ⊔B ( A {added specW
m | W ∈ supers(A)})
= [by definition of join of specification cases for BRHS ]
FB
B F A {added specW | W ∈ supers(A)})
( RHS
{added specU
m | U ∈ (supers(B) \ supers(A))}) ⊔ (
m
= [by definition of join with respect to B]
FB
U
RHS {added specm | U ∈ (supers(B) \ supers(A)) ∪ supers(A)}
= [by set theory]
FB
U
RHS {added specm | U ∈ supers(B)}
= [by Definition 2 ]
BRHS
ext specm
Figure 3: Proof of Law 3 - case of object of exact type B
4.3 Proofs
Here we present proofs for Laws 1 and 3. Both proofs involve dealing with cases associated to the types
of objects related to the classes that are emphasized in the laws. We present the proof for just one case
of these laws. The conditions of the laws guarantee that both programs that appear in the laws are welltyped. Concerning Law 2, it is a law for attributes in which specification inheritance is not taken into
consideration.
In Figure 2, we present the proof for the case of Law 1 in which the we consider an object of exact
type B. Notice that in Law 1, on the left-hand side, an object of exact type B has to establish the (added)
invariant ψ1 . The added invariant is given by ψ1 ∧ (this instanceo f C ⇒ ψinv ), on the right-hand side.
For an object of type B, the type test is false and the whole implication results true. The whole effect is
the same of the invariant of class B on the left-hand side.
The proof for the case of Law 3 in which we consider an object of exact type B is presented in Figure 3. On the left-hand side of this law, the specification case for method m in class B has precondition ψ1
Gabriel Falconieri Freitas et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
73
p u b l i c c l a s s EvenIntegerData {
/ / @ p r i v a t e i n v a r i a n t value . intValue ( ) % 2 == 0 ;
/ / @ p r i v a t e i n v a r i a n t value . intValue ( ) > −1;
p r i v a t e Integer value ;
p u b l i c EvenIntegerData ( ) { value = new Integer ( 0 ) ; }
/ * @ r e q u i r e s newValue != n u l l ;
@ r e q u i r e s newValue . intValue ( ) % 2 == 0 && newValue . intValue ( ) > −1;
@ e n s u r e s getValue ( ) . intValue ( ) == newValue . intValue ( ) ; @ * /
p u b l i c v o i d registerValue ( Integer newValue ) { /* . . . */ }
/ / @ e n s u r e s \ r e s u l t != n u l l ;
p u b l i c / * @ pure @ * / Integer getValue ( ) { /* . . . */ }
/ * @ r e q u i r e s getValue ( ) != n u l l ;
@ e n s u r e s ! ( \ r e s u l t ) . equals ( ” ” ) ; @ * /
p u b l i c String format ( ) { /* . . . */ }
}
Figure 4: Class EvenIntegerData
and postcondition ψ2 . On the right-hand side, we enrich this specification case with type tests involving
the class name C, but with no impacts for objects with distinct types from C. The other specification
case for method m on the right-hand side also involves a type test, having no effects for classes other than
class C.
5
Application
In this section, we present an example composed of excerpts of Java classes annotated with JML, as it
is refactored by means of successive application of programming laws. Classes PositiveIntegerData
(Figure 1) and EvenIntegerData (Figure 4) represent positive and even integers, respectively. These
classes are part of a software that stores and manipulates instances of positive and even integers.
The class EvenIntegerData (Figure 4) can only hold even positive integers because of the invariant
written in Line 2. And, to reinforce the invariant, pre-conditions of method registerValue guarantee that
only even and positive values are allowed. These classes share methods that have the same functionality.
Moreover, both have an attribute called value to save the integer value of the respective data. To improve
the design of this program (e.g. by reducing the amount of duplicated code) and to accept new data types
(e.g. odd integers), we need to change the program in a disciplined way. In what follows, we present a
guideline that leads to the same structure as obtained by applying the refactoring Extract Superclass [8].
We do not present all derivation steps, we omit most of them to save space, but each step is accomplished
by the application of a law. A detailed derivation, with all the steps and their corresponding laws, can be
found elsewhere [9].
Our starting point is composed by the classes presented in Figures 1 and 4. We first introduce
a new class (IntegerData) to be the superclass of the existing classes, by applying Law hclass eliminationi, from right to left. Then, we change the superclass of classes PositiveIntegerData and
EvenIntegerData to be IntegerData, by applying Law hchange superclass: from Object to another
classi, from left to right. We prepare classes PositiveIntegerData and EvenIntegerData, for moving
attribute, invariant, and methods. First, we move the common attribute value to the superclass. This is
accomplished by the application of a sequence of laws, beginning with Law hchange attribute visibility:
Object-oriented Programming Laws for Annotated Java Programs
74
1
2
3
p u b l i c c l a s s IntegerData {
/ / @ p r o t e c t e d i n v a r i a n t value . intValue ( ) > −1;
p r o t e c t e d Integer value ;
4
// @ ensures \ r e s u l t != null; @*/
p u b l i c / * @ pure @ * / Integer getValue ( ) { /* . . . */ }
5
6
7
}
Figure 5: Class IntegerData
from private to publici, which changes the visibility of the attribute value to public, then we change
its visibility to default by the application of another law. In the sequence, we apply Law 2 to move
the attribute value from class PositiveIntegerData to class IntegerData. This is followed by the application of other laws to eliminate the attribute value from class EvenIntegerData. Then, we move
common methods to the class IntegerData. First, we apply Lawhmove original method to superclassi
to move methods from class PositiveIntegerData to IntegerData; then we apply Law 3 to move redefined methods from EvenIntegerData to IntegerData. Another law allows us to simplify conditional
commands. After moving the attribute value and methods to the superclass IntegerData, we change the
invariant of classes PositiveIntegerData and EvenIntegerData to be in the format required by Law 1.
We apply Law 1 twice, then we simplify the invariant in class IntegerData and change its visibility.
In Figure 5, we present class an excerpt of class IntegerData after the application of the
programming laws that lead to the refactoring Extract Superclass [8]. The final version of the
PositiveIntegerData and EvenIntegerData has no getValue method and does not have the common
invariant (see Figure 5, Line 2) because, now, they belong to IntegerData.
In [9], we applied the laws proposed here along with others to refactor a core module of a Manufacturing Execution System (MES) [22] software, which formalizes methods and procedures of production
in an integrated system and presents data in more useful and systematic way. To control and manipulate
data dynamically and in a highly configurable way, the MES software is built on top of a Meta Data API.
We have refactored a JML-specified version of the Meta Data API [9] by applying primitive transformations expressed by means of our laws. We applied our laws to refactor code and to accommodate new
features. For instance, we eliminate duplicate code by extracting a superclass that abstracts the behavior
of other classes present in the system. This is described by the Extract Superclass refactoring [8]. Other
refactorings presented by Fowler [8] (for instance, Replace Conditional with Polymorphism and Pull Up
Method) were also applied to the Meta Data API by means of the proposed laws.
6
Conclusion
Object-oriented programming laws were proposed by Borba et al. [3] for an object-oriented language [4].
They proposed laws for classes and commands; they also define a normal form for object-oriented programs written in their language along with a reduction strategy. They demonstrate that the set of laws
is complete with respect to this normal form. Cornélio [6, 5] proves the laws with respect to a copy
semantics [4] and formally justifies, by using programming laws and data refinement, refactoring practices documented by Fowler [8]. Silva, Sampaio, and Liu consider object-oriented programming laws in
a language with a reference semantics [21], applying such laws to code refactoring. Duarte [7] adapts
the programming laws initially proposed in [3, 5] for the Java programming and proposes other laws for
Gabriel Falconieri Freitas et al.
75
language features that are not present in the language used in [3, 5].
In this paper, we proposed laws for object-oriented programming in the presence of a behavioral
interface specification language. In the laws that deal with source-doce, we treat the transformation
considering the restrictions imposed by the specifications. These laws are based on programming laws
from previous work [3, 5, 7] that does not consider specifications. We have considered laws that address
only a subset of the JML’s Level 0 constructs [16], specially for lightweight specifications. Nevertheless,
our preliminary focus is to cover most of the JML constructs that form the core notation used in the
design by contract methodology. We have also applied our set of laws for reducing a JML-specified Java
program to a normal form [9] to address the relative completeness of the set of laws proposed.
With respect to the order of application of programming laws in program transformation process,
distinct orders may lead to different results. The conditions for application of a programming law usually
requires the application of other programming laws, defining that their applications are not commutative.
To obtain the desired target program, a proper law application sequence must be established. Commonly
used sequences of applications of laws can be registered as single transformation rules.
Concerning the refactoring process, we can view a refactoring as a target transformation that can
be reached by the application of primitive transformations expressed by means of programming laws.
The process of application of programming laws terminates when the desired structure is reached.
Cornélio [6] presents refactorings as rules constituted by a pair of programs, similar to a law, but the
transformation described is more complex than the one of a law. The program on the left-hand side
presents the class or classes before rule application; the right-hand side presents the classes after rule
application. Refactoring rules capture complex transformations. Here we have not presented refactoring
rules as in [6], but the application of the programming laws we presented here and in [9] lead us to a result
similar to that presented in [6], that is, a refactoring is derived by the application of programming laws.
Although at this moment our work does not provide a way to transform programs mechanically, it offers a
more reliable and extensible alternative to address behavior-preserving transformations like refactorings.
Moreover, since some conditions present in the laws are related to logic proofs, it is necessary to use a
theorem prover as an auxiliary tool.
Differently from laws that deal only with constructs of an object-oriented programming language, the
presence of a behavioral interface specification language requires that we be aware of issues related, for
instance, to the visibility of specification and code constructs, invariant preservation when introducing
calls to super and changing a parameter type to a supertype requires introducing casts in occurrences of
the parameter in specifications.
As future work, we intend to describe laws to support other JML clauses like initially, constraint,
represents, and model fields. Also, we intend to work on proofs for the Java parts of the laws based on
a reference semantics [1].
Acknowledgements
We are partially supported by the Brazilian Research Agency, CNPq, grant 484813/2007-2, and by the
National Institute of Science and Technology for Software Engineering (INES1 ), funded by CNPq and
FACEPE, grants 573964/2008-4 and APQ-1037-1.03/08.
1 http://www.ines.org.br
76
Object-oriented Programming Laws for Annotated Java Programs
References
[1] A. Banerjee & D. A. Naumann (2005): Ownership confinement ensures representation independence for
object-oriented programs. Journal of the ACM 52(6), pp. 894—960.
[2] R. Bird & O. de Moor (1997): Algebra of Programming. Prentice Hall.
[3] P. Borba et al. (2004): Algebraic reasoning for object-oriented programming. Science of Computer Programming 52, pp. 53–100.
[4] A. L. C. Cavalcanti & D. A. Naumann (2000): A Weakest Precondition Semantics for Refinement of
Object-oriented Programs. IEEE Transactions on Software Engineering 26(8), pp. 713–728.
[5] M. Cornélio (2004): Refactoring as Formal Refinements. Ph.D. thesis, Universidade Federal de Pernambuco.
[6] M. Cornélio et al. (2009): Sound refactorings. Science of Computer Programming Available at http://dx.
doi.org/10.1016/j.scico.2009.10.001.
[7] R. Duarte (2008): Parallelizing Java Programs Using Transformation Laws. Master’s thesis, Universidade
Federal de Pernambuco (UFPE).
[8] M. Fowler (1999): Refactoring: improving the design of existing code. Addison-Wesley Longman Publishing
Co., Inc., Boston, MA, USA.
[9] G. R. F. Freitas (2009): Refactoring Annotated Java Programs: A Rule-Based Approach. Master’s thesis,
Departamento de Sistemas e Computação, Universidade de Pernambuco. Available online.
[10] M. Goldstein, Y. A. Feldman & S. Tyszberowicz (2006): Refactoring with Contracts. In: AGILE ’06:
Proceedings of the conference on AGILE 2006. IEEE Computer Society, Washington, DC, USA, pp. 53–64.
[11] J. Gosling et al. (2005): Java Language Specification. Addison-Wesley Professional, 3rd edition edition.
[12] C. A. R. Hoare et al. (1987): Laws of programming. Commun. ACM 30(8), pp. 672–686.
[13] G. T. Leavens (2006): JML’s Rich, Inherited Specifications for Behavioral Subtypes. In: Z. Liu & J. He,
editors: ICFEM, Lecture Notes in Computer Science 4260. Springer, pp. 2–34.
[14] G. T. Leavens & Y. Cheon (2005): Design by Contract with JML. Available online.
[15] G. T. Leavens & D. A. Naumann (2006): Behavioral Subtyping, Specification Inheritance, and Modular
Reasoning. Technical Report 06-20b, Department of Computer Science, Iowa State University, Ames, Iowa,
50011.
[16] G. T. Leavens et al. (2008): JML Reference Manual. Available online.
[17] T. Massoni, R. Gheyi & P. Borba (2008): Formal Model-Driven Program Refactoring. In: J. L. Fiadeiro &
P. Inverardi, editors: FASE, Lecture Notes in Computer Science 4961. Springer, pp. 362–376. Available at
http://dx.doi.org/10.1007/978-3-540-78743-3_27.
[18] B. Meyer (1992): Applying design by contract. IEEE Computer 25, pp. 40–51.
[19] C. C. Morgan (1994): Programming from Specifications. Prentice Hall, second edition.
[20] S. Seres (2001): The Algebra of Logic Programming. Ph.D. thesis, Oxfor University Computing Laboratory.
[21] L. Silva, A. Sampaio & Z. Liu (2008): Laws of Object-Orientation with Reference Semantics. In: SEFM ’08:
Proceedings of the 2008 Sixth IEEE International Conference on Software Engineering and Formal Methods.
IEEE Computer Society, Washington, DC, USA, pp. 217–226.
[22] R. R. Zagidullin & E. B. Frolov (2008): Control of manufacturing production by means of MES systems.
Russian Engineering Research 28(2), pp. 166–168.
| 6cs.PL
|
arXiv:1608.05205v3 [cs.GR] 6 Jan 2017
Tree-decomposable and Underconstrained
Geometric Constraint Problems
Ioannis Fudos∗1 , Christoph M. Hoffmann†2 and Robert Joan-Arinyo‡3
1 Department
of Computer Science and Engineering, University of
Ioannina, Ioannina, Greece
2 Department of Computer Science, Purdue University, West Lafayette, IN,
USA
3 Department of Computer Science, Universitat Politecnica de Catalunya,
Barcelona, Catalonia
January 9, 2017
Abstract
In this paper, we are concerned with geometric constraint solvers, i.e.,
with programs that find one or more solutions of a geometric constraint problem. If no solution exists, the solver is expected to announce that no solution
has been found. Owing to the complexity, type or difficulty of a constraint
problem, it is possible that the solver does not find a solution even though
one may exist. Thus, there may be false negatives, but there should never be
false positives. Intuitively, the ability to find solutions can be considered a
measure of solver’s competence.
We consider static constraint problems and their solvers. We do not consider dynamic constraint solvers, also known as dynamic geometry programs,
in which specific geometric elements are moved, interactively or along prescribed trajectories, while continually maintaining all stipulated constraints.
However, if we have a solver for static constraint problems that is sufficiently
fast and competent, we can build a dynamic geometry program from it by
solving the static problem for a sufficiently dense sampling of the trajectory
of the moving element(s).
∗ [email protected]
† [email protected]
‡ [email protected]
1
The work we survey has its roots in applications, especially in mechanical
computer-aided design (MCAD). The constraint solvers used in MCAD took
a quantum leap in the 1990s. These approaches solve a geometric constraint
problem by an initial, graph-based structural analysis that extracts generic
subproblems and determines how they would combine to form a complete solution. These subproblems are then handed to an algebraic solver that solves
the specific instances of the generic subproblems and combines them.
1
Introduction, Concepts and Scope
A geometric constraint problem, also known as a geometric constraint system,
consists of a finite set of geometric objects, such as points, lines, circles, planes,
spheres, etc., and constraints upon them, such as incidence, distance, tangency, and
so on. A solution of a geometric constraint problem P is a coordinate assignment
for each of the geometric objects of P that places them in relation to each other such
that all constraints of P are satisfied. A problem P may have a unique solution, it
may have more than one solution, or it may have no solution.
In this chapter, we are concerned with geometric constraint solvers, i.e., with
programs that find one or more solutions of a geometric constraint problem. If no
solution exists, the solver is expected to announce that no solution has been found.
Owing to the complexity, type or difficulty of a constraint problem, it is possible
that the solver does not find a solution even though one may exist. Thus, there may
be false negatives, but there should never be false positives. Intuitively, the ability
to find solutions can be considered a measure of solver’s competence.
We consider static constraint problems and their solvers. We do not consider
dynamic constraint solvers, also known as dynamic geometry programs, in which
specific geometric elements are moved, interactively or along prescribed trajectories, while continually maintaining all stipulated constraints. However, if we have
a solver for static constraint problems that is sufficiently fast and competent, we
can build a dynamic geometry program from it by solving the static problem for a
sufficiently dense sampling of the trajectory of the moving element(s).
The work we survey has its roots in applications, especially in mechanical
computer-aided design (MCAD). The constraint solvers used in MCAD took a
quantum leap with the work by Owen [45]. Owen’s algorithm solves a geometric constraint problem by an initial, graph-based structural analysis that extracts
generic subproblems and determines how they would combine to form a complete
solution. These subproblems are then handed to an algebraic solver that solves the
specific instances of the generic subproblems and combines them. Owen’s graph
analysis is top down. A bottom-up analysis was proposed in [3]. Subsequent work
expanded the knowledge of
2
1. the structure and properties of the constraint graph, see Section 3;
2. the geometric vocabulary, see Section 4;
3. the understanding of spatial constraint systems, see Section 5.
We also look briefly at under-constrained problems in Section 6.
Restricted to points and distances, the constraint graph analysis has deep roots
in mathematics and combinatorics; see, e.g., [42, 16, 37] for some of these connections. Here, we limit the discussion to constraint problems that are motivated
by MCAD applications, and that means that the geometric vocabulary has to be
richer than points and distances between them. In the remainder of this section we
introduce informally major concepts and methods for solving geometric constraint
systems (GCS) with an application perspective in mind.
Note in particular that some definitions in this section may be provisional.
Those definitions, although formally incorrect, are so given nevertheless because
they facilitate understanding the material. They are clearly identified along with
examples that show where they fall short — and what can be done about it.
1.1
Geometric Constraint Systems (GCS)
Fix the space in which to consider a geometric constraint system (GCS). Typical
examples are the Euclidean space in 2 or 3 dimensions. Each object to be placed
by a GCS instance in that space has a specific number of degrees of freedom (dof),
i.e., a specific number of independent coordinates. In Euclidean 2-space, points
and lines each have 2 dof. In Euclidean 3-space, a point has 3 dof and a line has
4. A constraint on such objects corresponds to one or more equations expressing
the constraints on the coordinates. So, requiring a distance d between two points
A = (Ax , Ay ) and B = (Bx , By ) in 2-space would be expressed by
(Ax − Bx )2 + (Ay − By )2 = d 2
Requiring that the two points are coincident would entail two equations:
Ax = Bx
Ay = By
Accordingly, a GCS can be viewed simply as a system of equations: A solution of
the GCS, if one exists, is a valuation of the variables, the coordinates, that satisfies all equations. Viewed in this foundational way, solving a GCS boils down to
formulating a system of equations in the coordinates of the geometric entities and
solving the system by any means appropriate. The equations are almost always
algebraic.
3
C
2
1
D
2
1
1
1
1
A
B
2
(a)
2
(b)
Figure 1: Constraint problem and associated constraint graph.
1.2
Constraint Graph, Deficit, and Generic Solvability
The approach of treating a GCS as an (unstructured) system of equations is inefficient and almost always unnecessary. Given a GCS, we analyze the structure
of the corresponding equation system, and seek to identify smaller subsystems that
can be solved independently and that admit especially efficient solution algorithms.
Triangle decomposable systems, the main subject of this chapter, do this by analyzing a constraint graph that mirrors the equation structure. The constraint graph
is recursively broken down into subgraphs that correspond to independently solvable subsystems. Once subsystems have been solved, their solutions are combined
and expose, in the aggregate, additional subsystems that now can be solved separately. Given the GCS problem P = (U, F), with the set of geometric objects U
and constraints F upon them, we define the constraint graph G(P) as follows:
Definition 1.1 Given the GCS P = (U, F), its constraint graph G = (V, E) is a labeled undirected graph whose vertices are the geometric objects in U, each labeled
with its degrees of freedom. There is an edge (u, v) in E if there is a constraint between the geometric objects u0 and v0 , corresponding to u and v respectively. The
edge is labeled by the number of independent equations corresponding to the constraint between u0 and v0 .
Example 1.2 Consider the constraint problem of Figure 1 that comprises four
points in the plane, labeled A through D. A line between two points indicates a
distance constraint on the points. Thus there are 5 distance constraints, shown left.
The corresponding constraint graph is shown on the right.
Since a GCS naturally corresponds to a system of equations, we expect that
the number of independent equations should equal the number of variables. The
number of independent variables equals the sum of degrees of freedom, i.e., the
4
sum of vertex labels of the constraint graph. Moreover, the number of independent
equations equals the sum of the edge labels, in most situations. Thus, we investigate ho the structure of the constraint graph reflects the structure of the equation
system that corresponds to the GCS.
As stated, the GCS of Example 1.2 does not prescribe where on the plane to
position and orient the points after solving the GCS. Thus, while this GCS is wellconstrained, the sum of vertex labels equals the sum of edge labels plus 3. Here, the
deficit of 3 corresponds to 3 missing equations that fix where, in the plane, to place
the solution. If necessary, the remaining degrees of freedom can be determined
by placing the points with respect to a global coordinate system, for example by
adding three equations that place A at the origin and B on the (positive) x-axis.
So, if the position and orientation of the solution is undetermined, we are assured that a constraint graph where ∑ l(v) − ∑ l(e) < 3 in the plane corresponds
to an equation system in which at least one equation is not independent. Consequently, we conclude
Definition 1.3 A GCS problem P is generically over-constrained if there is an induced subgraph of the associated constraint graph, such that for the induced subgraph the following holds: ∑ l(v) − ∑ l(e) < κ and none of the constraints fixes the
geometric structure with respect to the global coordinate system, where κ = 3 in
the plane and κ = 6 in 3-space.
Extrapolating this line of reasoning, we might be led to the conclusion that we
can use this structural graph property to define
Definition 1.4 A GCS problem P in the Euclidean plane/space is generically underconstrained if it is not overconstrained and ∑ l(v) − ∑ l(e) > κ.
Definition 1.5 (Provisional)
A GCS problem P in the Euclidean plane/space is generically well-constrained if:
(i) it is not over-constrained and (ii) the sum of vertex labels of the associated
constraint graph equals the sum of the edge labels plus the deficit κ.
Note, however, for a graph to be well-constrained Definition 1.5 is not sufficient and will be refined in Section 3.7. The problem with Definition 1.5 is best
illustrated by an example. The following example exhibits a constraint graph for
which the sum of the labels of the vertices equals the sum of the labels of the edges
plus κ but it contains an overconstrained subgraph and therefore the GCS cannot
be well-constrained.
5
v2
v4
v6
v1
v3
v5
Figure 2: A GCS with an overconstrained and an underconstrained subgraph.
Example 1.6 Consider the constraint graph of Figure 2. For specificity, let the
graph vertices be points and the edges distances. The deficit is 3, so it seems that
a problem with this constraint graph is well-constrained. Now the subgraph induced by vertices v1 , ..., v4 is overconstrained; drop any of the subgraph edges and
you obtain a well-constrained subproblem with a deficit 3. On the other hand, the
subgraph induced by the vertices v3 ...v6 has a deficit of 4 and is in fact underconstrained; vertices v5 and v6 cannot be placed. There is an extra constraint in the
first subproblem that is not available for the second subproblem.
Even in 2D, dof counting fails to account for well-constrained graphs. We will
further study the concept of well-constrained graphs in Section 3.7. Analogously, a
GCS in 3-space with deficit 6 that is not overconstrained likewise does not always
have a well-constrained graph.
Note that the definitions and properties explained assume in particular that the
GCS solution does not have certain symmetries. For instance, consider placing
concentrically two circles of given radii. The constraint graph has two vertices,
labeled 2 each, and one edge labeled 2, because concentricity implies that the two
centers coincide. Here the deficit ∑ l(v) − ∑ l(e) is 2, less than κ = 3. The system
appears to be over-constrained, but the constraint system is well-constrained. The
lower deficit reflects the rotational symmetry of the solution.
In the following, we exclude GCS that fix the structure with respect to the
coordinate system, as well as GCS that exhibit symmetries that reduce κ.
1.3
Instance Solvability
Consider constructing a triangle 4(A, B,C) with the three sides a, b, c through each
of the vertex pairs, in the Euclidean plane. The three points and three lines together
have 12 degrees of freedom. Each vertex is incident to two lines, so there are six
incidence constraints, each contributing one equation. We add three angle constraints, one for each line pair. The resulting constraint graph has a deficit of three
6
C
A
D
D
A
B
C
B
D
C
A
(a)
(b)
B
(c)
Figure 3: Different solutions of the constraint problem of Example 1.2.
and thus appears to be generically well-constrained. However, the problem is not
well-constrained: if the three angle constraints add up to π, then the GCS is actually underconstrained. If they add up to π/4, say, the GCS has no solution.
The problem arises from the interdependence of the three angle constraints.
Thus generic solvability does not guarantee that problem instances are solvable. In
particular, the constraint graph does not record specific dependencies among the
equations. Such dependencies often arise from geometry theorems.
Definition 1.7 A GCS is (instance) well-constrained if the associated (algebraic)
equation system has one or more discrete real solutions. It is (instance) overconstrained if it has no solution, and is (instance) under-constrained if it has a
continuum of solutions.
The definition of generic solvability is fair in the sense that dependencies among
some of the constraints, and their equational form, can arise from geometry theorems. The example above is based on a simple theorem, but more complex theorems can arise and may be as hard to detect as solving the equations in the first
place.
1.4
Root Identification and Valid Parameter Ranges
Consider Example 1.2. With the distance constraints as drawn, the GCS has multiple solutions, some shown in Figure 3. From the equational perspective none
is distinguished. From an application perspective usually one is intended. Since
the number of distinct solutions can grow exponentially with the number of constrained geometric objects, it is not reasonable to determine all solutions and let
the application choose. This application-specific problem is the root identification
problem. Some authors use the term chirality.
Solvability generalizes to the determination of valid parameter ranges: It is
clear that the distances between the points A, B,C, and the distances between the
7
C
D
A
B
E
H
F
I
G
Figure 4: A variational constraint problem.
points B,C, D must satisfy the triangle inequality for there to be a (nondegenerate)
solution. Considering the metric constraints (distance, angle) as coordinates of
points in a configuration space, what is the manifold of points that are associated
with solvable GCS instances? Is the manifold connected or are there solutions that
cannot be reached from every starting configuration?
1.5
Variational and Serializable Constraint Problems
Some GCS have equation systems that are naturally triangular. That is, intuitively,
the geometric objects can be ordered such that they can be placed one-by-one. Such
GCS are serializable or sequential constraint problems. GCS that are not serializable have been called variational constraint problems in some of the applicationoriented literature. The examples discussed so far are all sequential. The following
example shows a variational problem.
Example 1.8 Place 9 points in the plane subject to the 15 distance constraints
indicated by the lines, in Figure 4. The following three groups of 4 points each can
be placed with respect to each other: (A, B,C, D), (D, E, F, G), and (G, H, I, A).
Each of these subproblems is sequential in nature. However, the overall placement
problem of all points is not serializable; it is variational.
Serializable constraint problems are of limited expressiveness. Due to their potential efficiency, however, they are used to great effect in some dynamic geometry
packages such as Cinderella [6] and GeoGebra [15].
1.6
Triangle-Decomposing Solvers
The overall strategy of triangle-decomposing solvers is to construct the constraint
graph and, analyzing the graph, to recursively isolate solvable subproblems. Then,
the solved subproblems are recombined. The process is recursive.
8
For planar constraint systems, triangle-decomposability is understood to be a
recursive property in which the base case consists of two vertices (2 dof each) and
one edge between them eliminating one of the four dof’s. Larger structures are
then solved by isolating three solved structures that pairwise share a single geometric element and thus can be combined by placing the solved structures relative
to each other. When this decomposition is done top-down, the solved structures are
triconnected; [45]. When the decomposition is done bottom-up, triples of solved
components are sought that pairwise share a geometric primitive; [3]. The act of
combining such triples has been called a cluster merge.
For the problem of Example 1.8, illustrated in Figure 4, the graph is first decomposed into the three subproblems discussed. Next, each of these subproblems
is further broken down. For the subproblem A, B,C, D, for example, this might be
done by placing, with respect to each other, A and B, B and C, and C and A. Those
three subsystems are then combined into the triangle A, B,C. This merging step
combines three subsystems that pairwise share a point, placing the three shared
geometric elements in one geometric construct. The triangle can then be merged
with the two subsystems C and D, and B and D. The result so obtained is the solved
subsystem with vertices A, B,C, D. Similarly, we solve subproblems D, E, F, G and
G, H, I, A, each in two merge stpdf. Finally, the entire problem is solved by merging
the three subsystem.
The recursion solving the problem of Figure 4 can be thought of either as a
top-down decomposition, or as a bottom-up reconstruction.
Note that the top-level merge step places the shared objects, here the points A,
D and G. The distances between these points are not given but can be obtained
from the three subproblem solutions. We can think of this process as a plan that is
formulated based on the constraint graph analysis. Two questions arise:
1. Since the plan so formulated is not unique, do different decompositions arrive at different solutions? This question is settled by investigating the nature of the recursive decomposition and whether it satisfies a Church-Rosser
property.
2. What specific subsystems of equations must be solved and how? This question is approached by analyzing the basic subgraph configurations that can
occur, that are allowed by the geometric constraint solver, both when decomposing and when recombining.
Concerning the second question, in our example two key operations are needed:
placing two points at a prescribed distance, and placing a third point at a required
distance from two fixed points. There is also a third operation that places a rigid
geometric configuration X such that two points of X match two given points, us9
ing translation and rotation. This operation is used when recombining the three
subproblems.
Carrying out these operations entails solving equation systems of a fixed structure. Often, these equation systems are small and can be solved very efficiently.
The simplest cases restrict the geometric repertoire to points, lines, and circles of
fixed radius. Triangle decomposable systems then require at most solving univariate quadratic polynomials.
More complicated primitives can be added. They include circles of a-priori
unknown radius, e.g., [36]; conic sections, e.g., [9]; and Bézier curves, e.g., [19].
Additions impact both the constraint graph analysis as well as the richness of algebraic equation systems that have to be solved. The impact on the graph analysis can
be limited to adding additional patterns to cluster merging. So extended, triangle
decomposition becomes a more general analysis that can be called tree decomposition. The decomposition remains a tree, but the tree may have different numbers of
subtrees associated with a growing repertoire of subgraph patterns. The number of
possible patterns is infinite. Therefore, this approach becomes self-limiting. More
than that, rigidity no longer has a simple characterization. The contributions to the
equation solving algorithms is also considerable. Section 4.3 catalogues what is
known about the equation systems required to add just circles of unknown radius.
Because of these rigors, more general techniques have to be considered. For
the extended graph analysis, the DR-type algorithms solve the problem by dynamically identifying generically solvable subgraphs; [22, 21]. Likewise, the growth
of irreducible algebraic equation systems increasingly motivates searching general
equation solvers, including numerical solving techniques, such as Newton iteration, homotopy continuation, and other procedural techniques, for instance [49].
1.7
Scope and Organization
We begin with the graph analysis of triangle-decomposable constraint systems in
2D. These systems play an important role in linkage analysis and graph rigidity,
[52, 53, 54]. But even in the Euclidean plane, applications of geometric constraint
systems argue for more expressive systems. We can increase expressiveness by
enlarging the repertoire of geometric objects, as well as by admitting more complex
cluster merging operations.
In the plane, an extended repertoire includes foremost variable-radius circles;
that is, circles whose radius is not given explicitly but must be inferred based on the
constraints. More advanced objects, such as conics and certain parametric curves,
can also be considered. Those additional geometries impact the subsystems of
equations that must be solved.
More complex cluster merging operations affect both the constraint graph struc10
tures encountered as well as the equation systems that must be solved to position
the geometric structures accordingly. We will discuss some examples.
After discussing constraint graphs and generic over-, well-, and under-constrained
graphs, we consider the equations that must be solved so as to obtain specific solutions. Here, we explain the structure of those systems as well as some tools to
transform the equation systems into simpler ones. Questions of root identification,
valid parameter ranges, and order type of solutions are to be discussed. Much is
known about these topics in the Euclidean plane. For Euclidean 3-space much less
is known. There the equation systems are in general much harder, and the number
of cases that should be considered is larger. Even simple sequential problems can
require daunting equation systems. A case in point is to find a line in 3-space that
is at prescribed distances from four given points, discussed in Section 5.
2
Geometric Constraint Systems
Definition 2.1 A geometric constraint system (GCS) is a pair (U, F) consisting
of a finite set U of geometric elements in an ambient space and a finite set F of
constraints upon them.
The ambient space typically is n-dimensional Euclidean space. The majority of
applications require n = 2 or n = 3; e.g., [45, 3, 30]. Spherical geometry may also
be considered, for instance in nautical applications.
The imposed constraints typically are binary relations. We do not consider
higher-order constraints, such as "C be the midpoint between points A and B." Note,
however, that such constraints can often be expressed by several binary constraints.
This can be done in a variety of ways, with or without variable-radius circles.
Definition 2.2 A solution of a geometric constraint system (U, F) is an assignment
of coordinates instantiating the elements of U such that the constraints F are all
satisfied.
There may be several solutions [10]. Moreover, solutions may or may not be required to be in prescribed position and orientation, in a global coordinate system.
As defined, a GCS is a static problem in that solutions fix the geometric elements with respect to each other. The dynamic geometry problem asks to maintain
constraints as some elements move with respect to each other. We consider only
static constraint problems and their solvers.
By geometric coverage we understand the diversity of geometric elements admissible in U. Points, lines and circles of given radius are adequate for many applications in Euclidean 2-space [45]. For GCS in Euclidean 3-space, an analogous
11
Geoms
Points A, B
Point A, line m
Lines m, n
Constraint
d(A, B)
d(m, A)
a(m, n)
Notes
distance not zero
zero distance allowed
lines not parallel
Table 1: Minimal GCS in the Euclidean plane; d(...) denotes distance, a(...) denotes angle.
geometric coverage could be points, lines, planes, as well as spheres and cylinders
of fixed radii. Here, the number of solutions even of simple GCS can be very large
[13].
3
Constraint Graph
The constraint graph of Definition 1.1 is an abstract representation of the equation
system equivalent to the geometric constraint problem. The analysis of the graph
yields a set of operations used to solve the equations. Ideally, those operations
are simple, for instance univariate polynomials of degree 2 [3]. To start the graph
analysis, we find minimal constraint problems. That is, constraint problems with
a minimal number of geometric objects whose solution defines a local coordinate
frame. Such problems depend on ambient space. Consider the following:
Definition 3.1 Given a constraint problem in the Euclidean plane, consisting of
two points A and B and a nonzero distance constraint between them. Such a problem is minimal. The associated constraint graph G = ({A, B}, {(A, B)}) is a minimal constraint graph.
This minimal constraint problem establishes a coordinate system of the Eu−
→
clidean plane in which A is the origin and the oriented line AB is the x-axis. This is
not the only minimal constraint problem in the plane. Table 1 shows the minimal
problems involving the basic geometric objects with 2 dof. Note that fixed-radius
circles can be used in lieu of points as long a the centers and points are not coincident. Two parallel lines, at prescribed distance in the plane, are not considered
minimal because they do not establish a coordinate frame.
Table 2 shows the main cases for Euclidean 3-space. In the case of two lines
that are skew, a third line L3 is constructed that connects the two points of closest
approach. Here, L1 and L3 define a plane that is oriented by L2 . If the lines L1 and
L2 intersect, they lie on a common plane that is coordinatized by the two lines and
12
Geoms
Points p1 , p2 , p3
Point p, line L
Lines L1 , L2
Planes E1 , E2 , E3
Constraint
d(pi , pk )
d(p, L)
d(L1 , L2 ), a(L1 , L2 )
a(Pi , Pk )
Notes
no zero distance
distance not zero
lines not parallel
no parallel planes
Table 2: Minimal GCS in Euclidean 3-space; d(...) denotes distance, a(...) denotes
angle.
is oriented by defining the third coordinate direction using a right-hand rule. If the
two lines are parallel, they define a common plane but fail to coordinatize it.
3.1
Geometric Elements and Degrees of Freedom
A geometric element has a characteristic number of degrees of freedom that depends on the type of geometry and the dimensionality of ambient space.
Definition 3.2 Degrees of freedom (dof) are the number of independent, one-dimensional
variables by which a geometric object can be instantiated and positioned.
Elementary objects are geometric objects that have a certain number of dof and
cannot be decomposed further into more elementary objects. Compound objects
can be characterized by a GCS, and consist of one or several elementary objects
placed in relation to each other according to a GCS solution instance. A complex
object is rigid if its shape cannot change, or, equivalently, if the relative position
and orientation of the elementary objects that comprise the complex object cannot
change.
The degrees of freedom for elementary geometric objects and for rigid objects
in two and three dimensions are summarized in Table 3. Singularities of coordinatization can trigger robustness issues in GCS. Therefore, the representation of
elementary geometric objects should be uniform, without singularities. For example, if we represent lines by the familiar y = mx + b formula, lines parallel to the
y-axis cannot be so represented. This problem can be avoided if we represent a line
by its distance from the origin and the direction of the normal vector of the line.
3.2
Geometric Constraints
Geometric constraints consume one or more dof and are expressed by an equal
number of independent equations. A basic set of geometric constraints in the plane
is summarized in Table 4. Constraints can be expressed directly. Alternatively,
13
Geom
Point
Line
Plane
Circle, fixed-radius
Circle, variable-radius
Rigid body
Sphere, fixed- or
variable-radius
Ellipse, variable axes
Ellipsoid, variable axes
Geometric Meaning of DoF
Variables representing coordinates
2D: distance from origin and direction
3D: distance from origin, direction in 3D
direction on the plane
Distance from origin, direction in 3D
Coordinates of center, orientation in 3D
Coordinates of center, radius, orientation in 3D
2D: 2 displacements, 1 orientation
3D: 3 displacements, 3 orientation
3 displacements or
3 displacements, radius
center, axis lengths, axis orientation
center, axis lengths, axis orientation
2D
2
3D
3
2
4
2
3
3
3
5
6
6
3
4
5
7
9
Table 3: Degrees of freedom for elementary objects and rigid objects.
Type
point-point distance
Angle between lines and planes
Point on point
Point-line distance
Line-line parallel distance
Plane-plane parallel distance
Point on line
Line on line
Point on plane
Line on plane
Fixing elementary object
Constraint
One equation representing the distance between two points p1 , p2 under the metric
|| ||: ||p1 − p2 || = d with d > 0.
Angle between two lines (can be represented by the angle between the normal
vectors).
Exceptions in 3D: Prallelism between
lines eliminates 2DoF. Line-plane orthogonality in eliminates 2 DoF.
For any metric ||p1 − p2 || = 0 is equivalent
to p1 = p2 .
One equation expressing the point-line distance.
Parallelism and distance.
Parallelism and distance.
Same as point-line distance In 3D dimension is reduced.
Same as parallel distance between lines in
2D. In 3D an additional DoF is canceled.
Same as point-plane distance.
Plane-line parallelism and zero distance.
Fixing all or some of the DoF of an elementary object.
Table 4: Types of constraints and number of dof eliminated.
14
2D
1
3D
1
1
1
2
3
1
1
2
2
1
3
3
2
2
4
Dof
1
2
DoF
some can be reduced equivalently to other constraints. For example, suppose we
stipulate that a circle of radius r be tangent to a line L in the plane. Then we
can require equivalently that the center of the circle be at distance r from L. For
a comprehensive list of constraints involving planes, points and lines in 3D see
for example [39, 56]. Finally we should note that distance 0 between two points
takes out two degrees of freedom since this causes dimension reduction (from 1dimensional to 0-dimensional). Similarly, requiring that two lines be parallel at
distance d in 3-space eliminates 3 dof, but if d = 0 the constraint eliminates 4 dof.
3.3
Compound Geometric Elements
Several geometric elements can be conceptually grouped into compound elements.
A circular arc is an example. Compound geometric elements are convenient concepts for graphical user interfaces of GCS solvers. Internally, they are decomposed
into geometric elements and implied constraints. We illustrate this dual view with
the example of a circular arc shown in Figure 5 and explain why it has overall 2
intrinsic parameters.
Since the arc does not have a fixed radius, it lies on a variable-radius circle that
has 3 dof. The two end points contribute an additional 4 dof. The two end points
are incident to the circle, reducing the overall dof to 5. Position and orientation
of the arc, in some coordinate system, requires 3 dof, leaving two intrinsic arc
parameters. They can be interpreted as 1 dof for the radius of the arc, and 1 dof for
the distance between the end points.
Note that this constraint problem has 4 solutions in general. Orienting the
end points, and connecting them with an oriented line, or circle, the arc can be
on one of two sides of the directed line. Moreover, the two end points divide the
circle into a shorter and a longer arc, in general. Which one is chosen accounts
for the other two solutions. Applications require selecting one of these solutions.
Several conventions can be followed to determine a unique segment: preservation
of the original configuration, preservation of the direction of the curve from the
first endpoint towards the second endpoint, the shortest segment and so on.
3.4
Serializable Graphs
When a GCS is serializable (Section 1.5) the geometric objects can be ordered in
such a way that they can be placed sequentially one-by-one as function of preceding, already placed elements. This idea can be formalized as follows.
Definition 3.3 Let G = (V, E) be a GCS graph and x0 , x1 , x2 . . . be elements in V .
We say that xi depends on xk , xr , . . ., written xi > xk , xr , . . ., if xi can be placed only
after the xk , xr , . . . have been placed.
15
A (2)
B (2)
on (1)
on (1)
C (3)
(a)
(b)
(c)
(d)
Figure 5: Building a complex object using one variable radius circle, two end
points and two incidence constraints.
If the constraint graph is serializable, then the pair (G, >) is a directed acyclic
graph (DAG) and admits topological sorting [1]. See Example 1.8. More formally,
Definition 3.4 Let G(V, E) be the constraint graph associated with a GCS in Euclidean 2-space. Without loss of generality assume that G is connected. We say
that the GCS is serializable if (G, >) describes a sequence of dependencies such
that, under a suitable enumeration of V ,
1. There are elements x0 and x1 ∈ V that induce a minimal constraint graph.
2. Each subsequent element xi ,
where j < i and k < i.
2 ≤ i ≤ |V | depends on elements x j and xk
In general, the enumeration is not unique and depends on the pair x0 , x1 ∈ V that
is placed first. However, as we will see later, different possible sequences derived
from a given DAG are equivalent in the sense that they lead to the same final placement for all the objects in V with respect to each other; see Section 3.7.
Example 3.5 Consider the graph in Figure 6(a). The edges have been directed to
show dependencies of placement. Choosing (A, I) as starting pair, a valid dependence relation is obtained. The list in Figure 6(b) gives a serial construction based
on the graph (G, >).
3.5
Variational Graphs
Definition 3.6 A GCS which is not serializable is called variational.
When the GCS is variational, starting with a minimal GCS and applying the
dependence relationship to the constraint graph G = (V, E) generates a sequence of
16
H > A, I
D > A, G
C > A, D
G > I, H
B > A, D
E > D, G
F > D, G
(a)
(b)
Figure 6: a) DAG derived from a serializable graph, with (A, I) as starting pair.
b) A construction sequence respecting the dependencies.
dependencies that includes only a subset of elements in V . We call it a subsequence
and the corresponding subgraph a cluster.
Assuming that the variational GCS is solvable, one repeatedly selects a minimal GCS and applies the dependence relationship, using graph edges not yet used,
resulting in a collection of subsequences.
Intuitively, the situation described means that clusters corresponding to different sequences must be merged, usually applying translations and rotations defined
by elements shared by subsequences. From an equational point of view, the existence of different subsequences reveals that there are several underlying equations
that must be solved simultaneously.
Example 3.7 Figure 7a shows a variational DAG corresponding to the variational
constraint problem in Figure 4. We choose three starting pairs (A, I), (G, F) and
(C, B). Each allows us to build a DAG from some of the graph vertices and edges.
They are listed in Figure 7(b). Notice that each subsequence identifies a serializable subgraph.
3.6
Triangle Decomposability
The strategy of triangle-decomposing solvers, as sketched in Section 1, is based on
decomposing the constraint graph recursively. Decomposition splits a (sub)graph
into 3 (sub)subgraphs that share one vertex pairwise. This is called a triangle
decomposition step. More complex splitting configurations can be considered and
called tree decomposition. Figure 8 shows a triangle and a more complex tree
decomposition step.
17
C
H > A, I
G > I, H
D
A
E > G, F
B
E
H
D > E, F
F
B > A, C
I
D > C, B
G
(a)
(b)
Figure 7: a) DAG derived from the variational graph in Figure 4 with starting pairs
(A, I), (G, F) and (C, B).
b) Three different subsequences of construction dependencies that can be identified
in the DAG.
(a)
(b)
Figure 8: a) Triangle decomposition step. b) More complex tree decomposition
step. Split subgraphs are shown as ovals, shared vertices as dots.
18
G1
C
A
D
B
E
I
F
H
G
G3
G2
Figure 9: A triangle-decomposition step for the graph shown in Figure 7(a).
From a practical point of view, in the case of GCS in Euclidean 2-space,
triangle-decomposable constraint graphs suffice for solving many problems that
arise in applications.
We define a triangle decomposition step as follows.
Definition 3.8 Let G = (V, E) be a graph. We say that three subgraphs of G,
Gi (Vi , Ei ), 1 ≤ i ≤ 3, define a triangle decomposition step of G if
1. V1 ∪V2 ∪V3 = V , E1 ∪ E2 ∪ E3 = E and Ei ∩ E j = 0,
/ i 6= j, and
2. There are three vertices, say u, v, w ∈ V , such that V1 ∩ V2 = {u}, V2 ∩ V3 =
{v} and V3 ∩V1 = {w}.
Example 3.9 Consider the graph G in Figure 7(a). As shown in Figure 9, the subgraphs G1 , G2 and G3 define a triangle decomposition step of G. Vertices pairwise
shared by the subgraphs are A, D and G.
Definition 3.10 We say that a ternary tree T is a triangle decomposition for the
graph G if
1. The root of T is the graph G.
2. Each node of T is a subgraph G0 ⊂ G which is either the root of a ternary
tree generated by a triangle decomposition step of G0 or a leaf node with a
minimal associated subgraph.
Definition 3.11 A graph for which there is a triangle decomposition is called
triangle-decomposable.
19
ABCDEFGHI
ABCDEFGHI
AC CD
AB
DEFG
AGHI
ABCD
ABD
BD
DA
AI
AH
IH
IHG
HG
GI
DE
EG
FG
FGD
GD
AC CD
DF
(a)
AB
DEFG
AGHI
ABCD
AI
ABD
BD
DA
AH
IH
IHG
HG
GI
FG
DF
DE
DEG
EG
GD
(b)
Figure 10: Two different triangle decompositions for the graph shown in Figure 7.
In general, a triangle decomposition of a graph is not unique. However, if the
graph is triangle decomposable by one sequence of decomposition stpdf, then any
legal sequence will decompose the graph [10].
Example 3.12 Consider the graph and the triangle decomposition step shown in
Figure 9. Now recursively apply decomposition stpdf to each of the subgraphs
G1 , G2 , G3 until reaching minimal subgraphs. Figure 10 shows two triangle decompositions for the graph considered that differ in the subtree rooted at node
DEFG. Notice, however, that the set of terminal nodes is the same in both triangle
decompositions.
Assume that the three subgraphs G1 , G2 , G3 are (graphs of) solvable GCS subproblems. Let u and v be the shared vertices of G2 and there is no constraint
between them. Then u and v constitute a virtual minimal GCS where the constraint
relating the two elements is not given but can be deduced from the solution of G2 .
A similar statement can be made about G1 and G3 and their shared vertices.
The triangle pattern is not the only decomposition construct [3]. Others include
the pattern shown in Figure 8b. Intuitively, a decomposition pattern represents an
equation system that must be solved simultaneously. Decomposition patterns are
infinite in number; see also SECTION 2.2
3.7
Generic Solvability and the Church-Rosser Property
Recall Definitions 3.8, 3.11, and 1.5.
Triangle-decomposable graphs, of GCS in the Euclidean plane, are generically
well-constrained, and can be solved either bottom-up [11] or top-down [35]. This
assertion is based on the shared geometric elements, of each triangle decomposition step, having 2 dof and being in general position with respect to each other.
20
L3
L3
B
B
L2
C
L2
C
A
L1
A
(a)
L1
(b)
Figure 11: Well-constrained graph with three lines, two points, and a variableradius circle. a) Constraint problem. Dashed lines represent metric constraints. b)
Constraint graph. All vertices have 2 dof except C which has three. Solid lines
represent incidence constraints.
Variable-radius circles have 3 dof and give rise to a special case illustrated in Figure 11.
Triangle decomposable graphs are also called ternary-decomposable on account of the topology of the decomposition tree. A (recursive) decomposition of
a well-constrained triangle decomposable graph is not unique. However, it can be
shown using the Church-Rosser property for reduction systems that if one triangle decomposition sequence fully reduces the graph, then all such decompositions
must succeed, [10]. The advantage of restricting to triangle-decomposable problems is that a fixed, finite repertoire of algebraic equation systems suffices to solve
this class of problems.
Triangle decomposable graphs are a subset of the set of well-constrained constraint graphs of planar GCS. The entire set has been characterized by Laman
in [37].
Definition 3.13 Let G = (V, E) be a connected, undirected graph whose vertices
represent points in 2D and edges represent distances between points. G is a wellconstrained constraint graph of a GCS iff, the deficit of G is 3 and, for every subset
U ⊂ V , the induced subgraph (U, F) has a deficit of no less than 3.
Two examples of well-constrained graphs that are not triangle-decomposable
are shown in Figure 12. In triangle decomposition, the irreducible constituents
of the constraint graph are the minimal constraint graphs, Definition 3.1. For the
general case, the set of irreducible constraint graphs has been characterized in [20]
using a network flow approach. Conceptually, irreducible constraint graphs must
be solved as a single equation system. Since the graphs can be arbitrarily complex,
so can be the equation systems.
21
(a)
(b)
Figure 12: Well-constrained graphs that are not triangle-decomposable. a) Graph
K3,3 . b) Desargue’s graph. All vertices have 2 dof, all edges consume 1 dof.
b
A
A
hb
c
b
C
c
hc
ha
a
B
C
a
B
(a)
(b)
Figure 13: Deriving the geometry of a triangle given its three altitudes. a) Formulating the three altitude problem. b) The resulting constraint graph.
22
1
2
1
2
3
5
1
2
3
6
4
4
1
2
5
3
6
4
3
2
4
Figure 14: A Laman graph in 2D that is not rigid. V1 is a variable radius circle.
V2 ,V3 ,V02 ,V03 are points. V4 , V04 are lines. e1 ,e2 ,e01 ,e02 are points. e3 ,e4 ,e03 ,e04 are on
constraints. e5 , e05 are distances. e6 , e06 are distances from cirlce ( circumference).
For the planar case, when we have only points and distances the set of triangle decomposable graphs coincides with the set of quadratically solvable graphs.
However if we extend to lines and angles, there are quadratically solvable graphs
which are not triangle decomposable. Consider the problem of finding a triangle
from its three altitudes shown in Figure 13(a). The corresponding graph is shown
in Figure 13(b) where the hexagon edges are point-on-line constraints and the diagonals are point-line distance constraints. The geometric problem is quadratically
solvable, [31], but the graph, K3,3 , is not triangle decomposable.
Laman’s theorem holds even if we extend the repertoire of geometries to any
geometry having 2 degrees of freedom and the constraints to virtually any constraint of Table 4. However, if we extend the set of geometries to include for
example variable radius circles, then the Laman condition is no longer sufficient.
Example 3.14 Consider the GCS of Figure 14. We have two rigid clusters C1 =
{V1 ,V2 ,V3 ,V4 } and C2 = {V1 ,V20 ,V30 ,V40 }, where V1 is a variable-radius circle, V2 ,V3 ,V20 ,V30
are points, and V4 ,V40 are lines. The constraints e1 , e2 , e01 , e02 are distances from the
center of V1 , e6 , e06 are distances from the circumference of V1 , e5 , e05 are distance
constraints, and e3 , e4 , e03 , e04 are incidence constraints. The two clusters share the
variable-radius circle V1 .
The graph is clearly a Laman graph but is not rigid, since it is underconstrained
— C1 and C2 can move independently around circle V1 . The problem is also overconstrained since the radius of V1 can be derived independently from cluster C1
and from cluster C2 .
23
Figure 15: Two hexahedra sharing two points.
3.8
2D and 3D Graphs
Graph analysis for spatial constraint problems is not nearly as mature as the planar
case. In the planar case, Definition 3.1 conceptualizes minimal GCS with which
a bottom-up graph analysis would start and a top-down analysis would terminate.
Table 1 lists the configurations. In 3-space, analogously, the corresponding minimal problem may consist of three points or three planes, with three constraints
between them. However, the major configurations shown in Table 2 also include
other cases, for example two lines.
In the plane, serially adding one additional point or line requires two constraints. In 3-space, adding a point or plane requires three. The inclusion of lines
in 3-space leads to serious complications. For example, since lines have 4 dof, see
Table 3, we have to be able to construct lines from given distances from 4 fixed
points. Equivalently, we have to construct a common tangent to 4 fixed spheres, a
problem known to have up to 12 real solutions, [23]. Work in [62] has explored the
optimization of the algebraic complexity of 3D subsystems.
In 3D, the Laman condition is not sufficient. Figure 15 illustrates two hexahedra sharing two vertices. If the length of the edges is given the GCS that arises is
also known as the double banana problem. The graph is a Laman graph but the
problem corresponds to two rigid bodies (each hexahedron is a rigid body) sharing two vertices and is thus non rigid in the sense that the two rigid bodies are
free to rotate around the axis defined by the two shared points. The problem is
also clearly overconstrained since the distance of the two shared geometries can be
derived independently by each of the two rigid bodies.
A necessary and sufficient condition for rigidity in 3D has been recently presented in [39] for an extended set of geometries and constraints. The authors have
extended the theory in [55, 59] to characterize systems of rigid bodies made of
points, lines and planes connected by virtually any pairwise constraint (point-point,
24
point-line, line-plane etc) except point-point coincidence. In this approach, a multigraph (V, (B, A)) is formulated, where vertices V represent rigid bodies and edges
(B, A) stand for primitive constraints that represent single equations between the
two 6-vectors that describe the rigid body motion of the two vertices. Primitive
constraints intuitively affect at most one degree of freedom. Each geometric constraint is translated to a number of primitive constraints (see Appendix C of [39]).
A distinction is made between primitive angular and blind constraints: a primitive
angular constraint may affect only a rotational degree of freedom. All other primitive constraints that may affect either a rotational or translational degree of freedom
are called blind constraints.
Therefore edges are of two types: angular (A) and blind (B). Such a scheme
is minimally rigid if and only if there is a subset B0 of the blind edges such that
(i) B − B0 is an edge disjoint union of 3 spanning trees and (ii) A ∪ B0 is an edge
disjoint union of 3 spanning trees.
Note that this characterization does not capture the cases of Figures 14 and 15.
The condition holds for both cases, but rigidity is not guaranteed since they involve
point coincidences between rigid bodies directly or indirectly.
4
Solver
After the constraint graph has been analyzed, the implied underlying equations are
to be solved. We discuss now how to do that.
4.1
2D Triangle-Decomposable Constraint Problems
We restrict to points and lines in the Euclidean plane. As discussed in Section 3.6,
triangle-decomposable constraint systems in the plane require solvers that implement three operations:
1. The two geometric elements of a minimal subgraph (Definition 3.1) are
placed consistent with the constraint between them.
2. A third geometric element is placed by two constraints on two geometric
elements already placed.
3. Given two geometric elements in fixed position, a rigid-body transformation
is done that repositions the two elements elsewhere.
These operations are applied to the decomposition tree, progressing from the leaves
of the tree to the root. The solving order is bottom-up regardless whether the
decomposition tree was built top-down or bottom-up [45, 10, 3].
25
d(p1 , p2 )
d(p, L)
a(L1 , L2 )
Set p1 = (0, 0), p2 = (d, 0)
Place L on the x-axis, p = (0, d)
Place L1 on the x-axis, L2 through the origin at angle a
Table 5: Placing minimal GCS: p represents points, L represents lines, d distance,
a angle.
Example 4.1 Consider the constraint system of Example 1.2. We choose the subgraph induced by A and B as minimal and place A at the origin and B on the
positive x-axis, at the stipulated distance from A, so executing Operation 1.
We place C by the two constraints on A and B, solving at most quadratic,
univariate equations, executing Operation 2. The triangle A, B,C is thereby constructed.
Assume that we have solved the triangle B,C, D in like manner, separately and
with B and D as vertices of the minimal subgraph. We can now assemble the two
triangles by a rigid-body transformation that moves the triangle B,C, D such that
the points B and C are matched, using Operation 3.
Note that we can extend the geometric vocabulary of points and lines, adding circles of given radius at no cost. A fixed-radius circle is replaced by its center. A
point-on-circle constraint is replaced by a distance constraint between the point and
the center, and a tangency constraint by a distance constraint between the tangent
and the center.
Operation 1: Minimal GCS Placement
This operation chooses a default coordinate system. There are three pairs of geometric elements that occur in minimal GCS: (point, point), (point, line), and (line,
line). The chosen placements are shown in Table 5. Other choices could be made.
Operation 2: Constructing one Element from two Constraints
Two elements A and B are given, a third element C is to be placed by constraints
upon them. There are six cases, with p denoting a point and L denoting a line.
(p, p) → p
(p, p) → L
(p, L) → p
(p, L) → L
(L, L) → p
(L, L) → L
The sixth case, (L, L) → L is underconstrained. See also [3]. We illustrate with the
case (p, p) → L.
26
(+,−)
(+,+)
(−,−)
B
A
(+,+)
A
B
B
(−,−)
A
(−,+)
(a)
(b)
(c)
Figure 16: Constructing a line at specific distances from two points. Equivalently,
finding common tangents of two circles with radius equal to the stipulated distances. The degenerate cases of 1 or 3 solutions are not shown.
Example 4.2 In the case (p, p) → L a line should be constructed by respective
distance from two points. Consider the two circles with the given points as center
and radius equal to the stipulated distance, as shown in Figure 16. Depending on
the three distances, there will be 4, 2 or no real solution in general.
Order the points A and B, and orient the circles centered at those points counterclockwise. Orient the line to be constructed so that the projection of A onto the line
precedes the projection of B. Then we can distinguish the up to four tangents in
a coordinate-independent way: observe whether the line orientation is consistent
with the circle orientation (+), or whether the orientations are opposite (–). See
also [3].
The degenerate cases where the two circles are tangent to each other yield three
solutions or one. In theses cases there is one double solution that represents the coincidence of two solutions, with orientations (+,–) and (–,+), or with orientations
(+,+) and (–,–).
Operation 3: Matching two Elements
The operation requires that the geometric elements to be matched be congruent.
The operation is a rigid-body transformation and is routine.
4.2
Root Identification and Order Type
We noted that constructing one element from two constraints can have multiple
solutions. As a result, a well constrained GCS has in general an exponential number of solution instances. We shall illustrate this with the simple construction
(p, p) → p.
27
Q
di
dj
Pi
Pj
Q’
Figure 17: Placing one point Q from two points Pi , Pj already known.
Example 4.3 Consider placing n points, by 2n − 3 distance constraints between
them, and assume that the distance constraints are such that we can place the
points by sequentially applying the construction (p, p) → p. In general, each new
point can be placed in two different locations: Let Pi and Pj be known points from
which the new point Q is to be placed, at distance di and d j , respectively. Draw
two circles, one about Pi with radius di , the other about Pj with radius d j as shown
in Figure 17. The intersection of the two circles are the possible locations of Q.
For n points, therefore, we could have up to 2n−2 solution instances.
Note that not all construction paths derive real solutions. If, in Example 4.3,
the distance between Pi and Pj is larger than the sum of di and d j , then there is
no real solution for placing Q and therefore any subsequent construction using this
instance of Q is not feasible. Therefore, one might argue that this pruning may
result in polynomial algorithms. However, this is unlikely since the problem of
determining whether a well constrained GCS has a real solution has been shown to
be NP-complete [11].
In general, an application will require one specific solution, usually known as
the intended solution. To identify it is not always a trivial undertaking. In [3]
finding the intended solution is called the Root Identification Problem. Notice that,
on a technical level, selecting the intended solution corresponds to selecting one
among a number of different roots of a system of nonlinear algebraic equations.
A well constrained GCS would not necessarily include enough information to
identify which solution is the intended one. Consider the following example.
Example 4.4 The well constrained GCS in Figure 18a consists of four points, four
straight segments, four point-point distances and an angle. The solution includes
four instances. Two correspond to the one shown in Figure 18b and to a symmetric
arrangement of the same shape. Solution instances in Figures 18c and 18d are
structurally different.
Clearly, the GCS sketch in Figure 18a does not include any hint on which solution instance must be chosen to be displayed on the user’s screen. Thus, additional
28
d3
P4
P3
d4
P4
3
P4
α
P3
d2
P1
3
P2
d1
(a)
P1
P1
P1
P2
P
(b)
(c)
P3
P2
(d)
Figure 18: a) GCS consisting of four points, four straight segments, four pointpoint distances and an angle. b),c) and d) Three different solution instances to the
GCS.
information must be supplied to the solver. In [3], approaches applied to overcome
this issue have been classified into five categories: Selectively moving geometric
elements, adding extra constraints to narrow down the number of possible solution
instances, placement heuristics, a dialogue with the constraint solver that identifies
interactively the intended solution, and applying a design paradigm approach based
on structuring the GCS hierarchically. Next we elaborate on each category.
Moving selected geometry
In this approach, a solution is presented graphically to the user. The user selects,
again graphically, certain geometric elements that are considered misplaced. The
user then moves those elements where they should be placed in relation to other
elements. This approach is used by the DCM solver, [7, 45].
Adding extra constraints
Adding a set of extra constraints to narrow down which is the intended solution
instance is an intuitive and simple approach to solving the root identification problem. Extra constraints could capture domain knowledge from the application or
could be just geometric — and actually over-specify the GCS. Unfortunately, both
ideas result in NP-complete problems [3]. Nevertheless, extra constraints along
with genetic algorithms have been applied to solve the root identification problem
showing a promising potential. Authors in [60] argue that the approach is both
effective and efficient in search spaces with up to 2100 solution instances. A different application of genetic algorithms to solve the root identification problem is
described in [4]. The approach mixes a genetic algorithm with a chaos optimization
method.
Example 4.5 Genetic algorithms described in [32, 40, 47], use extra geometric
29
and topological constraints defined as logical predicates on oriented geometries.
For example, assume that the polygon in Figure 18a is oriented counterclockwise.
The solution shown in Figure 18c would be selected as the intended one by requiring the following two predicates to be fulfilled
−→
PointOnSide(P3 , P1 P2 , le f t),
−→
−→
Chirality(P2 P3 , P3 P4 , cw)
Order-based heuristics
All solvers known to us derive information from the initial GCS sketch and use
it to select a specific solution. This is reasonable, because one can expect that a
user sketch is similar to what is intended. For instance, by observing on which side
of an oriented line a specific point lies in the input sketch it is often appropriate
to select solutions that preserve this sidedness. The solver described in [3] seeks
to preserve the sidedness of the geometric elements in each construction step: the
orientation of three points with respect to each other, of two lines and one point,
and of one line and two points. The work described in [3] implements an additional
heuristic for arc tangency which aims at preserving the type of tangency present in
the sketch. See Figure 20. When the rules fail, the solver opens a dialogue to allow
the user to amend the rules as the situation might require. These heuristics are also
applied in the solver described in [33].
Example 4.6 Consider placing three points, P1 , P2 and P3 , relative to each other.
The points have been drawn in the initial sketch in the position shown in Figure 19a. The order defined by the points can be determined as follows. Determine
−→
where P3 lies with respect to the line P1 P2 . If P3 is on the line, then determine
whether it lies between P1 and P2 , preceding P1 or following P2 . The solver will
preserve this orientation if possible. For this example, the solver will choose the
point P3 as shown in Figure 19b.
Dialog with the solver
A useful paradigm for user-solver interaction has to be intuitive and must account
for the fact that most application users will not be intimately knowledgeable about
the technical working of the solver. So, we need a simple but effective communication paradigm by which the user can interact with the solver and direct it to
a different solution, or even browse through a subset of solutions in case the one
shown in the user’s screen is not the intended one.
30
P3
P3
P1
P1
P2
P2
P3
(a)
(b)
Figure 19: Placing three points, P1 , P2 and P3 , relative to each other. a) Points
placed in the initial sketch and induced orientation. b) P3 and P30 are two possible
placements for the third point. Preserving the orientation defined in the sketch
leads to select P3 as the intended placement.
(a)
(b)
Figure 20: Tangency types. a) Arc and segment tangency. b) Circle-circle tangency.
Example 4.7 SolBCN, a ruler-and-compass solver described in [33] offers a simple user-solver interaction tool. Figures 21a and b respectively show the GCS
sketch input by the user and the solution instance selected by the heuristic rules
implemented. Then, the user can trigger the solution selector by clicking on a button and a list with the set of construction stpdf where a quadratic equation must be
solved is displayed, Figure 21c. The user can then change the square root sign for
each of these construction stpdf by either selecting it directly or navigating with
the next/previous pair of buttons. Figure 21d shows a solution different from the
first one so obtained.
Navigating the GCS solution space using the approach illustrated in the Example 4.7 is simple. But it has obvious drawbacks. On the one hand, the number of items in the list selector grows exponentially with the number of quadratic
construction stpdf in the GCS. On the other hand it is difficult to anticipate how
choosing a root sign for a construction step will affect the solution selected by the
next sign chosen by the user.
These problems are avoided by considering that, conceptually, all possible solution instances of a GCS can be arranged in a tree whose leaves are the different
instances, and whose internal nodes correspond to stages in the placement of individual geometric elements. The different branches from a particular node are
31
(a)
(b)
(c)
(d)
Figure 21: User-solver dialog offered by the ruler-and-compass solver described
in [33]. a) GCS sketch. b) Solution instance selected by the heuristics implemented
in the solver. c) Solution instance selector. d) Solution instance displayed after
changing the square root signs of some construction stpdf.
32
the different choices for placing the geometric element. Since the tree has depth
proportional to the number of elements in the GCS, stepping from one solution
instance to another is proportional to the tree depth only. Moreover, it is possible
to define an incremental approach by allowing the user to select at each construction step which tree branch should be used. In solvers based on the DR-planning
paradigm, [27], this tree naturally is the construction plan generated by the solver.
Example 4.8 The Equation and Solution Manager [50], features a scalable method
to navigate the solution space of GCS. The method incrementally assembles the desired solution to the GCS and avoids combinatorial explosion, by offering the user
a visual walk-through of the solution instances to recursively constructed subsystems and by permitting the user to make gradual, adaptive solution choices. Figure 22 illustrates the approach.
Design paradigm approach
One of the difficulties in selecting the intended solution of a GCS stems from the
fact that geometric elements in a problem sketch are not grouped into logical structures. Authors in [3] argue that hierarchically structuring the constraint problem
would alleviate the complexity of solving the root identification problem, for example grouping geometric elements as design features. First a basic, dimensiondriven sketch would be given. Then, subsequent dimension-driven stpdf would
modify the basic sketch and add complexity. By doing so, the design intent would
become more evident and some of the technical problems would be simplified.
Example 4.9 Consider solving the GCS in Figure 23a. The role of the arc is
clearly to round the adjacent segments, and thus it is most likely that the solution
shown in Figure 23b is the one the user meant rather than the one in Figure 23c,
when changing the angles to 30◦ . However, the solver would be unaware of the
intended meaning of the arc. Instead, the user could sketch first the quadrilateral
without the arc, and then add the arc to round a vertex. When changing some of
the dimensional constraints, the role of the arc would remain that of a round, so
preserving the user intent.
4.3
Extended Geometric Vocabulary
So far we discussed 2D constraint solvers that use only points and lines, as well
as circles of given radii. Now we will add other geometric element types. This
33
(a)
(b)
(c)
(d)
Figure 22: Incremental solution space navigation described in [50]. a) GCS problem including three circles. b) Construction plan graph for the GCS solution. c)
GCS solution instance after choosing one of the possible solutions for each construction step. d) A different GCS solution instance after rebuilding the partial
construction corresponding to the construction step labeled G14 in the tree.
(a)
(b)
(c)
Figure 23: Solution selection by the design paradigm approach: Panel (a) shows
the final GCS, panels (b) and (c) two different solution instances. If the arc is introduced as a rounding feature of a constrained quadrilateral, then selecting solution
(b) over solution (c) is a logical choice.
34
has implications for both the equation solvers as well as for the constraint graph
analysis.
The simplest addition allows GCS to include geometric elements of the new
type, but only if these elements can be constructed sequentially, from explicit constraints on a set of already placed geometric elements. A more difficult addition
would allow the use of elements of the added type in the same way as the core
vocabulary of points and lines. Note that a new element type can have more than
two degrees of freedom.
Variable-Radius Circles
Circles whose radii have not been given explicitly are arguably the most basic
extension of the 2D core solver. Variable-radius circles have three degrees of freedom. We consider two ways in which they can arise:
1. A variable-radius circle is to be constructed by a sequential step from three,
already placed geometric entities.
2. A variable-radius circle is to be determined in a step analogous to merging
three clusters (Definition 3.8) and the circle acts as a cluster.
Note that a variable-radius circle cannot be a shared element in the sense of Definition 3.8. Shared elements have already been constructed; therefore a shared
variable-radius circle has already become a fixed-radius circle.
In the following, we consider points to be circles of zero radius. Consequently,
there are four ways in which a variable-radius circle can be constructed sequentially. Table 6 summarizes the equation systems.
It is advantageous to convert a line-distance constraint into two separate tangency constraints, so simplifying the equations that must be solved. For example,
if the sought circle C is to be at distance d from line L, we work instead with two
problems, one in which the circle C is to be tangent to a parallel line L+ of L. Here
L+ is at distance d from L, on one side of L. In the other problem, C is to be
tangent to a parallel line L− also at distance d, but on the other side of L. Analogously, perimeter distance from a given circle reduces to tangency with a circle
whose radius has been increased or reduced by said distance.
When deriving the algebraic equations, we work with cyclographic maps, orienting both lines and circles. Briefly, an oriented circle C is mapped to a normal
cone µ(C), with an axis parallel to the z-axis and a half angle of π/4. The cone
intersects the xy-plane in C. Depending on the orientation, the apex of the cone is
above or below the xy-plane. Considered as zero-radius circle, the point P in the
xy-plane maps to the normal cone µ(P) with apex in the xy-plane. Oriented lines
35
Given
Elements
Equation
System
LLL
LLC
LCC
CCC
(1,1)
(1,2)
(1,1,2)
(1,1,2)
Notes
intersect two angle bisectors
intersect two planes and a cone
intersect two planes and a cone
Apollonius problem; Example 4.10
Table 6: Sequential construction of variable-radius circles. All constraints are tangent constraints. Equations formulated using cyclographic maps. For the cases
LCC and CCC the linear equation(s) are from intersecting two cones.
L in the xy-plane are mapped to planes through L at an angle of π/4 with the xyplane. This reformulates the constraint problem as a spatial intersection problem
of planes and cones. See [46, 18, 25, 24, 5] for details and further reading.
The intersection of two normal cones is a conic, in affine space, plus a shared
circle at infinity. The conic lies in a plane whose equation is readily obtained
by subtracting the two cone equations: if K1 = 0 and K2 = 0 are the two normal
cones, then K1 − K2 = 0 is the plane that contains the conic in which the two cones
intersect.
Example 4.10 Consider the sequential construction problem of finding a circle
that is tangent to three given circles. This is the classical Apollonius problem that
has eight solutions in general.
We orient the circles and require that the sought circle be oriented consistently
with the given circles at the points of tangency. After orienting the given circles,
we can map the problem to the intersection, in 3-space, of three normal cones, C1 ,
C2 , and C3 , each arising from an oriented circle. Intersect two cone pairs, say
C1 ∩ C2 and C1 ∩ C3 , obtaining two planes that, in turn, intersect in a line L in
3-space. Then intersect L with one of the cones, say C2 . Two points are obtained
that, understood as the apex of a normal cone, map each to one (oriented) circle
in the plane that is a solution; see [48]. Algebraically, the solution is obtained by
solving linear equations plus one univariate quadratic equation.
There are 8 ways to orient the three circles, but they correspond pairwise, so
only four such problems must be solved. If one or more circles are points, they
must be considered oriented both ways. So, for each zero-radius circle, the number
of solutions reduces by a factor of 2. The special cases of the Apollonius problem
have been mapped out and solved in [48] using cyclographic maps.
36
Now consider the determination of variable-radius circles in a cluster merge.
Here, there are two constraints from each cluster to the variable-radius circle to
be constructed, and the two clusters share a geometric element. The situation is
analogous to the triangle merge characterized in Definition 3.8. The various cases
and how to solve them have been studied and solved in [25, 24, 5]. Specifically,
[25, 24] map out the cases in which the constraints are on the perimeter of the
variable-radius circle; and [5] considers constraints on the center of the variableradius circle as well.
Table 7 and 8 summarize the results from those papers. The approach is conceptually as follows. Let S1 = {E0 , E1 , E2 } and S2 = {E0 , E3 , E4 } be the clusters
constraining the variable-radius circle. The two clusters share element E0 , either a
line denoted L, or a circle denoted C. The clusters can move relative to each other,
translating along the shared line if E0 = L, or rotating about the center of the shared
circle if E0 = C. Elements E1 and E2 belong to S1 and constrain the sought circle.
Likewise, E3 and E4 are the constraining elements of S2 . Proceed as follows:
1. Fix the cluster S1 that has the more difficult constraining elements E1 and
E2 ; i.e., the cluster with the larger number of circles.
2. Choose a convenient coordinate system: the shared line E0 = L as the x-axis,
or the origin as the center of the shared circle if E0 = C.
3. Construct the cyclographic map of all constraining elements. The cones and
planes of S2 are parameterized by the distance d between S1 and S2 , or else
by the angle θ between S1 and S2 .
4. Construct three planes from the constraining elements E1 , . . . , E4 . They are
either cyclographic maps of lines, or normal cone intersections. The elements of S2 give rise to parameterized coefficients, by distance d for translation along the x-axis, or by angle θ for rotation around the origin, of the
moving cluster S2 . Intersect the planes, so obtaining a point with parameterized coordinates.
5. Substitute the parameterized point into the equation of the element E1 of the
fixed cluster, so obtaining a univariate polynomial that finds the intersection
point(s) of the four cyclographic objects; a polynomial in d or θ .
6. Solve the polynomial as described, each obtained by a particular configuration of orientations.
Some of the constraints can be on the center c of the variable-radius circle,
and [5] considers those cases. Note that there can be at most two constraints on the
37
Constraint type
E(LL, LL)
E(CL, LL)
E(CL,CL)
Planes
Polynomial degree
[L2 ], [L3 ]t , [L4 ]t
[L2 ], [L3
]t ,
[L2 ], [L4
]t ,
[L4
(1,2)
]t
(2,4)
([C1 ], [C3
]t )
(4,4)
]t ,
]t
(2,4)
E(CC, LL)
([C1 ], [C2 ]), [L3
[L4
E(CC,CL)
([C1 ], [C2 ])
([C1 ], [C3 ]t )
[L4 ]
(4,4)
E(CC,CC)
([C1 ], [C2 ])
([C1 ], [C3 ]t )
([C3 ], [C4 ])t
(4,4)
Table 7: Cluster cases; all constraints on circle perimeter. E(...) denotes whether
clusters share a line L, the translational case, or share a circle C, rotational case. [X]
denotes the cyclographic map equation µ(X) of X; [X]t denotes the equation with
coefficients parameterized by distance t (translation case) or by angle θ (rotation
case). (X,Y ) denotes the intersection plane equation of X and Y . The parameterized point is substituted into the equation [C1 ], except for the first case where
it is substituted into [L1 ]. (m, n) denotes the equation degrees, namely m for the
translation case E = L, and n for the rotation case E = C.
38
center of the variable-radius circle, for otherwise the relative position of S2 to S1
would be determined and the role of the variable-radius circle would be curtailed.
The problem is again solved in the same conceptual manner, but with a twist.
When a constraint is placed on the center c of the variable-radius circle, the constraint can be expressed by extending cyclographic maps with a map τ(X). Here,
τ(L) is a vertical plane through the line L. Moreover, τ(C) is a cylinder through
the circle C with axis parallel to the z-axis. The results so obtain are summarized
in Table 8.
A problem is denoted E0 (E1 E2 , E3 E4 ). E0 is the shared element by the two
clusters, a line L or a circle C. E1 and E2 are the two elements of the fixed cluster constraining the variable-radius circle. E3 and E4 are the two elements of the
moving cluster constraining the variable-radius circle. The numbers (m,n) are the
equation degrees when E0 = L (m), and E0 = C (n). An element Ek0 constrains the
center, an element E the circumference, of the variable-radius circle.
5
Spatial Geometric Constraints
Compared to constraint problems in the plane, our knowledge of spatial constraint
systems is relatively modest. The constraint graph analysis applies with some notable caveats. For example, Laman’s characterization of rigidity does not apply in
3-space, not even when restricting to points only, and distances between them; see
Section 3.8. Furthermore, the subsystems isolated by constraint graph decomposition can be complex, especially if lines are admitted to the geometric vocabulary.
We illustrate the latter point with a few examples.
Points and Planes
Points and planes comprise the most elementary vocabulary in spatial constraint
solving. Both have three degrees of freedom and are dual of each other. In analogy
to the minimal constraint graph in 2D (Definition 3.1), a minimal constraint graph
in 3D consists of three elements, points or planes, and three constraints, forming
a triangle. The initial placement for the four combinations places the elements in
canonical order, planes first, points second. Table 9 summarizes the method, [8].
Note that the constraint between two planes is an angle, and the constraint between
two points or a point and a plane is a distance. The initial placement fails for the
exceptional angles 0 and π, as well as for distance 0 between two points.
Sequential constructions of points and planes are straightforward. The locus
of a third point p, at respective distances from two known points p1 and p2 , is the
intersection of two spheres centered at p1 and p2 . It is a circle that is contained in
39
One center constraint
Problem
Degree
Two center constraints
Problem
Degree
E(LL,LL’)
(1,2)
E(LL,L’L’)
E(LL’,LL’)
(1,2)
(1,2)
E(CL,LL’)
E(CL’,LL)
E(C’L,LL)
(2,4)
(2,4)
(2,4)
E(CL,L’L’)
E(CL’,LL’)
E(C’L,LL’)
E(C’L’,LL)
(2,4)
(2,4)
(2,4)
(2,4)
E(CL,CL’)
E(CL,C’L)
(4,4)
(4,16)
E(CL,C’L’)
E(CL’,CL’)
E(C’L,CL’)
E(C’L,C’L)
(4,8)
(4,4)
(4,16)
(4,4)
E(CC,LL’)
E(CC’,LL)
(2,4)
(4,4)
E(CC,L’L’)
E(CC’,LL’)
E(C’C’,LL)
(2,4)
(4,4)
(2,4)
E(CC,CL’)
E(CC,C’L)
E(CC’,CL)
(4,4)
(4,32)
(8,32)
E(CC,C’L’)
E(CC’,CL’)
E(CC’,C’L)
E(C’C’,CL)
(4,8)
(8,32)
(8,32)
(4,8)
E(CC,CC’)
(16,64)
E(CC,C’C’)
E(CC’,CC’)
(2,8)
(16,64)
Table 8: Cluster cases with constraints on the center of the variable-radius circle.
(m, n) denotes the equation degree for E = L and E = C, respectively. L0 denotes a
constraint between a line and the center of the variable-radius circle; C0 denotes a
constraint between a circle and the center of the variable-radius circle.
40
Canonical Placement of three Points or Planes
P1 , P2 , P3
P1 placed as the plane z = 0
P2 placed to intersect P1 in the y-axis
P3 placed to contain the origin
P1 , P2 , p3
P1 placed as the plane z = 0
P2 placed to intersect P1 in the y-axis
p3 is placed on the xz-plane
P1 , p2 , p3
P1 placed as the plane z = 0
p2 is placed on the positive z-axis
p3 is placed on the xz-plane with z ≥ 0
p1 , p2 , p3
p1 is placed at the origin
p2 is placed on the positive x-axis
p3 is placed on the xz-plane with z ≥ 0
Table 9: Placement of three entities that are mutually constrained. P denotes a
plane p a point.
a plane perpendicular to the line through p1 ad p2 . As before, this fact can be used
to simplify the algebra.
The simplest, nonsequential constraint system is the octahedron, consisting of
6 elements and 12 constraints, [28, 29]. The name derives from the constraint
graph that has the topology of the octahedron. There are 7 major configurations
according to the number of planes. The configurations with 5 and with 6 planes are
structurally underconstrained. Solutions of the octahedron constraint system have
been proposed in [28, 29, 43, 8]. The number of distinct solutions is up to 16.
Points, Lines and Planes
A line in 3-space has four degrees of freedom. Usually lines are represented with
6 coordinates, using Plücker coordinates, or with 8, using dual quaternions; e.g.,
[2]. Consequently, the implicit relationships between the coordinates have to be
made explicit by additional equations when solving constraint systems with lines
in 3-space.
The sequential construction of lines can be trivial, for instance determining a
line by distance from two intersecting planes. But it can also be hard, for instance,
when determining a line by distance from four given points in space. The latter
problem, in geometric terms, asks for common tangents to four fixed spheres. This
problem has up to 12 real solutions. The upper bound of 12 was shown in [30]
41
Figure 24: The two variational constraint problems with 4 lines. The 4 double
lines represent both angle and distance, the solid diagonals distance, and the dashed
diagonal angle constraints.
using elementary algebra. The lower bound of 12 was established in [41] with
an example. In contrast to the restricted problem of only points and planes, this
sequential constraint problem involving lines is therefore much more demanding.
We can constrain two lines by distance from each other, by angle, and also by
both distance and angle. Small constraint configurations that involve lines have
been investigated, [14, 13]. The papers show that there are 2 variational configurations of 4 lines. They are shown in Figure 24. The papers also show that the
number of configurations with 5 geometric elements, including lines, is 17. Moreover, when 6 elements are considered, the number of distinct configurations grows
to 683. Because of this daunting growth pattern, it is natural to seek alternatives.
One such alternative has been proposed in [58, 57] where, instead of lines, only
segments of lines are allowed. Consequently, many constraints can be formulated
as constraints on end points. Note that in many applications this is perfectly adequate.
6
Under-constrained Geometric Constraint Problems
In general, existing geometric constraint solving techniques have been developed
under the assumption that problems are well-constrained, that is, they adhere to
Definition 1.5 given in Section 1. Put differently, the number of constraints and
their placement on the geometric elements define a problem with a finite number of
solution instances for non-degenerate configurations. However, there are a number
of scenarios where the assumption of well-constrained does not apply. Examples
are early stages of the design process when only a few parameters are fixed or in
cooperative design systems where different activities in product design and manufacture examine different subsets of the information in the design model, [26].
The problem then is under-constrained, that is, there are infinitely many solution
instances for non-degenerate configurations.
42
R
r
α
d1
d2
Figure 25: Hook of a car trunk locker.
Example 6.1 Consider the hook of a car trunk locker shown in Figure 25. Once
distances d1 and d2 have placed the center of the exterior circle of the hook with
respect to the hook’s axis of rotation, the designer is mainly interested in finding
a value for the angle α where the exterior circle is tangent to the small circle
transitioning to the inner circle of the hook. The angle α has to be such that the
hook smoothly rotates while closing and opening the hood. At this design stage,
the stem shape of the hook is irrelevant.
This, and other simple examples taken from Computer-Aided Design, illustrate the need for efficient and reliable techniques to deal with under-constrained
systems. The same need is found in other fields where geometric constraint solving plays a central role, such as kinematics, dynamic geometry, robotics, as well as
molecular modeling applications.
Recent work on under-constrained GCS with one degree of freedom, has brought
significant progress in understanding and formalizing generically under-constrained
systems; [51, 53, 54]. The work focuses on GCS restricted to points and distances,
also generically called linkages.
The goal of geometric constraint solving is to effectively determine realizations or embeddings of geometric objects in the ambient space in which the GCS
problem is formulated. Thus, the current trend is that solving an under-constrained
GCS should be understood as solving some well-constrained GCS derived from
the given one.
There are two ways to transform an under-constrained GCS into a well-constrained
one: adding to the GCS as many extra constraints as needed or removing from
the GCS unconstrained geometric entities. Note that removing constrained entities
makes little sense. Accordingly, the literature on under-constrained GCS advocates
to transform an under-constrained GCS into a well-constrained one by adding new
constraints. This technique was formally defined in [34] as follows.
43
G
Over−constrained
111111111
000000000
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
Well−constrained
Under−constrained
Tree decomposable
Solvable
Completable
Figure 26: A partition of the geometric constraint graphs set, G . The set of treedecomposable graphs straddles over the sets of well- and over-constraint graphs.
Definition 6.2 Let G(V, E) be an under-constrained graph associated with a GCS
problem. Let E 0 be a set of additional edges each bounded by two distinct vertices
in V such that the graph G0 (V, E ∪ E 0 ) is well-constrained. We say that E 0 is a
completion for G and that G0 is the completed graph of G.
Let G denote the set of geometric constraint graphs. Definitions 1.5, 1.4 and 1.3
given in Section 1 induce in G a partition as shown in Figure 26. The set of treedecomposable graphs straddles over the sets of well- and over-constraint graphs.
As described in Section 3.6, the set of well-constrained, tree-decomposable graphs
are solvable by the tree decomposition approach.
Within the set of under-constrained graphs we can distinguish two families:
Those which are not tree-decomposable and those which are. It is easy to see
that there is no completion for a graph in the first family that could transform the
graph into a tree-decomposable one. Considering graphs in the second family,
Definition 1.5 fixes the number of extra constraints that must be added. However deciding which constraints should actually be added to the graph is not a
straightforward matter because the resulting graph could be either over-constrained
or well-constrained but not tree-decomposable.
Example 6.3 Figure 27a shows an under-constrained graph. To see that it is
tree-decomposable just consider as the first decomposition step the subgraphs
induced by the sets of vertices {E, F, G}, {A, G} and {A, B,C, D, F}. Finding
the stpdf needed to complete a tree-decomposition is routine. The completion
E = {(A, G), (G, B), (G, D), (E, F)} generates the well-constrained graph shown
in Figure 27b. To see that the graph is tree-decomposable take as a first decomposition step the two minimal constraint graphs with edges {(E, D)} and {(E, F)}
plus the subgraph induced by edges {(A, B), (A,C), (B,C), (B, D), (C, F), (D, F)}.
However, completion E = {(A, E), (E, G), (E, D), (G, D)} results in the graph de44
A
C
A
C
B
A
C
B
B
G
G
G
D
E
D
F
(a)
E
D
F
(b)
F
E
(c)
Figure 27: a) An under-constrained, tree-decomposable graph G. b) A treedecomposable completion of G. c) A non tree-decomposable completion of G.
picted in Figure 27c which is no tree-decomposable.
In what follows we restrict to the completion problem for triangle decomposable GCS.
Definition 6.4 An under-constrained graph G(V, E) is said to be completable if
there is a set of edges E 0 which is a completion for G.
Notice that completability does not require triangle-decomposability of the underconstrained graph, nor does it imply that a well-constrained, completed graph
be tree-decomposable.
Definition 6.5 Let G(V, E) be a triangle-decomposable, under-constrained graph
and let E 0 be a completion for G. We say that E 0 is a triangle-decomposable completion, (td-completion in short) for G if G0 (V, E ∪ E 0 ) is triangle-decomposable.
Reported techniques dealing with under-constrained GCS differ mainly in the
way they figure out completions as well as whether they aim at figuring out tdcompletions or just completions. The work in [38] describes an algorithm where
the constraint graph is captured as a bipartite connectivity graph whose nodes are
either geometric entities or constraints. Each edge connects a constraint node with
the constrained geometric node. In analogy to sequential solvers the graph edges
are directed to indicate which constraints are used to fix (incident) geometric objects. The connectivity graph is analyzed according to the degrees of freedom of
under-constrained geometric nodes. Each under-constrained geometric node is a
candidate to support an additional edge to a new constraint, or if there is an edge
45
A
C
ABCDEFG
B
ABCDG
CF
DEF
G
AG
ABC
AB
BC
AC
DE
DBG
DB
BG
DF
DG
D
EF
F
E
(a)
(b)
Figure 28: a) A triangle-decomposition for the under-constrained graph G(V, E)
in Figure 27a. Pairs of vertices {A, G}, {D, E}, {B, G} and {D, G} in leaf nodes
do not define graph edges. b) A set of additional edges defined on V (G). Edges
(A, E), (A, G), (B, G), (D, E), (D, G) and (E, G) do not belong to E.
that heads a propagation path to an existing constraint node that has an unused
condition. When there are several candidates on which the new constraint can be
established, the selection is left to the user. A similar approach to solve underconstrained GCS based on degrees of freedom analysis is described in [44].
Two different notions of td-completion were introduced in [34]. The first one
is called free completion and is computed in three stpdf. First a triangle decomposition for the given graph is figured out. Then the set of under-constrained leaf
nodes in the decomposition is identified. Notice that each node in this set stores a
subgraph G(V, E) where |V | = 2 and E is the empty set. Thus one edge is missing. Finally the completion is computed as the set of missing edges in the underconstrained leaf nodes of the triangle-decomposition.
Example 6.6 Figure 28a shows a triangle-decomposition for the under-constrained
graph G(V, E) in Figure 27a. We have |V | = 7 and |E| = 7. For 2D problems, the
general property of a well-constrained graph described in Section 1 is the Laman
theorem, [37], 2|V | − |E| = 3. Hence the number of additional edges required to
complete G to a well-constrained graph is |E 0 | = 2|V | − |E| − 3 = 4. The set of
leaves in the triangle-decomposition corresponding to under-constrained minimal
graphs includes exactly four elements: {A, G}, {B, G}, {D, E} and {D, G}. Thus
E 0 = {(A, G), (B, G), (D, E), (D, G)} is a free completion for G. The completed
well-constrained graph G0 (V, E ∪ E 0 ) is shown in Figure 27b.
The second td-completion is called conditional completion. The first and second stpdf are the same as in the free completion. However, in the third step, edges
46
to complete under-constrained leaf nodes in the decomposition are drawn from an
additional graph defined over a subset of vertices of the given graph. If the number
of those edges that are not in the original graph is smaller than the number required by Definition 1.5, then the completed graph will remain under-constrained.
However a free completion can eventually be applied to get a well-constrained
completion.
Example 6.7 Consider again the under-constrained graph G(V, E) in Figure 27a
and its triangle-decomposition shown in Figure 28a. The set of under-constrained
pairs of vertices in the triangle-decomposition is {(A, G), (B, G), (D, E), (D, G)}.
Assume that the set of additional edges defined on V (G) is
{(A, B), (A, E), (A, G), (B, G), (C, F), (D, F), (E, G), (E, D), (G, D)}
as shown in Figure 28b. Now, additional edges for the completion must be drawn
from a set E ∗ such that E ∩ E ∗ = 0.
/ In the case at hand,
E ∗ = {(A, E), (A, G), (B, G), (D, E), (D, G), (E, G)}
. Thus, a completion for G(V, E) is
E 0 = {(A, G), (B, G), (D, E), (D, G)} ⊂ E ∗
. Figure 27b shows the completed graph G0 (V, E ∪ E 0 ).
A technique to complete general under-constrained graphs is described in [34].
The approach is based on transforming the problem of computing a completion
into a combinatorial optimization problem. Edges in the graph and in the additional set are assigned different weights. Then a greedy algorithm generates a
well-constrained problem provided that there are enough edges in the additional
set. A variant of this approach is reported in [61].
There is a class of 2D, triangle-decomposable, under-constrained GCS that
occur in a number of fields like dynamic geometry or mechanical computer aided
design. In these GCS the geometries are points, the constraints are usually pointpoint distances, and exactly one edge is missing in the associated constraint graph.
Such GCS are known as linkages.
Example 6.8 Figure 29a shows an illustration of a crankshaft and connecting rod
in a reciprocating piston engine. The crankshaft and connecting rod can be abstracted as the GCS shown in Figure 29b. The GCS includes four points Pi , where
0 ≤ i ≤ 3, two lines L1 , L2 , three point-point distances, di , 0 ≤ i ≤ 2, one line-line
angle, α. Moreover, the points P0 , P2 and P3 must be on the line L1 , and the points
P0 , P1 must be on the line L2 . If, for example, values of the distance d1 or of the
angle α are freely assigned, then the GCS can be considered a linkage.
47
111
000
000
111
000
111
000
111
000
111
000
111
1111
0000
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
11
00
00
11
P1
on
L1
α
L2
on
on
P2
d3
P3
L1
α
P4
(a)
P2
d2
d1
d1
P5
on
P3
d2
P0
d0
on
P1
d4
(b)
(c)
Figure 29: a) A crankshaft and connecting rod in a reciprocating piston engine. b)
The crankshaft and connecting rod abstracted as a GCS. c) Geometric constraint
graph.
Reachability is an important problem in fields such as dynamic geometry or
conformational molecular geometry. It can be formalized as follows:
Let Rs and Re be two realizations of a well defined geometric construction where Rs is called the starting instance and Re the ending
instance. Are there continuous transformations that preserve the incidence relationships established in the geometric construction and
transform Rs to Re?
The well defined geometric construction in the reachability problem can be understood as a linkage, that is, a well-constrained GCS problem where values assigned
to one specific constraint can freely change. In [17] the reachability problem is
solved assuming that the underlying GCS is triangle-decomposable. The approach
first computes the set of intervals of values that the free constraint can take for
which the linkage is realizable. This set of intervals is known as the linkage Cayley
configuration space, [12]. When both Rs and Re are realizations with the free constraint taking values within one Cayley interval, the path is an interval arc. When
Rs and Re realizations belong to different intervals, finding a path entails figuring out whether there are continuous transitions between consecutive intervals that
permit the linkage to reach Re when starting at Rs. If more than one such path is
found, one is chosen according to some predefined strategy. In [17] the minimal
arc length path is the one chosen.
Example 6.9 The crankshaft and connecting rod GCS in Figure 29b is triangledecomposable, ruler-and-compass solvable. A construction plan that places each
geometric element with respect to each other is
48
− π2
π
2
σ
σ0 (s1 ) = +1;
σ1 (s1 ) = +1;
σ2 (s1 ) = −1;
σ3 (s1 ) = −1;
σ0 (s2 ) = +1
σ1 (s2 ) = −1
σ2 (s2 ) = +1
σ3 (s2 ) = −1
A
B
C
D
A
0 2.2
10.6 17.4
25.8
30
B
C
A → [− π2 , 0]
B → [0, π2 ]
C → [− π2 , π2 ]
d1
0
(a)
(b)
Figure 30: a) Cayley configuration for distance constraint d2 . b) Cayley configuration for angle constraint α.
1.
2.
3.
4.
P0
P3
L1
C0
=
=
=
=
origin()
distPP(P0 , d3 )
line2P(P0 , P3 )
circleCR(P3 , d1 )
5.
6.
7.
8.
P2
L2
C1
P1
=
=
=
=
intLC(L1 ,C0 , s1 )
linePA(P0 , α)
circleCR(P0 , d0 )
intLC(L2 ,C1 , s2 )
When a construction step has more than one solution, an orientation parameter
is needed to select the desired solution instance. Parameters s1 and s2 in stpdf 5
and 8 respectively allow to select one of the two points in a line-circle intersection.
Figure 30a shows the Cayley configuration space when the point-point distance
constraint value d1 changes. Intervals labeled A, B, C and D define orientations
and parameter values for which the GCS is realizable. Intervals A and C yield
realizations consistent with the one depicted in Figure 30a. Intervals B and D
yield realizations where point P2 would be placed on the line L1 opposite to P1
with respect to P0 . The Cayley configuration space when the varying parameter
is the angle α is shown in Figure 30b. Now orientations are represented along a
radial axis and angle values for which the solution is realizable are represented as
circular intervals.
Linkages are extensively studied in [52, 53, 54]. The object of these works is to
lay sound theoretical foundations for a reliable and efficient computation of Cayley configuration spaces for general tree decomposable linkages. New concepts
like size and computational complexity are introduced and efficient algorithms are
developed to answer a number of questions on linkages like effectively computing
Cayley configuration spaces and solving the reachability problem. Methods so far
applied to compute Cayley configuration spaces, like the one used in [17], suffer
from potential combinatorial growth. The work in [53, 54] shows that for low Cayley complexity GCS problems, computing the configuration space is polynomial in
49
the number of geometric elements of the problem.
Acknowledgement
Hoffmann gratefully acknowledges partial support by the National Science Foundation under award 1361783.
References
[1] A. Aho, J. Hopcroft, and J. Ullman. The Design and Analysis of Computer
Algorithms. Addison-Wesley, 1974.
[2] W. Blaschke. Kinematik und Quaternionen. VEB Verlag der Wissenschaften,
Berlin, Germany, 1960.
[3] W. Bouma, I. Fudos, C. Hoffmann, J. Cai, and R. Paige. Geometric constraint
solver. Computer Aided Design, 27(6):487–501, June 1995.
[4] C.H. Cao, W. Fu, and W. Li. The research of a new geometric constraint
solver. In X.-T Yan, C.-Y Jiang, and N.P. Juster, editors, Perspectives from
Europe, and Asia on engineering design and manufacturing, A comparison
of engineering design and manufacture in Europe and Asia, pages 205–214.
Springer Scinence+Business Media, LLC., 2004.
[5] C.-S. Chiang and R. Joan-Arinyo. Revisiting variable radius circles in constructive geometric constraint solving. Computer Aided Geometric Design,
221(4):371–399, 2004 2004.
[6] Cinderella. The interactive geometry software Cinderella, June 2007.
http://cinderella.de/tiki-index.php.
[7] D-Cubed. The Dimensional Constraint Manager. Cambridge, England, 2003.
2D DCM Version 44.0, 3D DCM Version 28.0.
[8] C. Durand and C. Hoffmann. A systematic framework for solving geometric
constraints analytically. JSC, 30:493–519, 2000.
[9] I. Fudos and C.M. Hoffmann. Constraint-based parametric conics for CAD.
Computer Aided Design, 28(2):91–100, 1996.
[10] I. Fudos and C.M. Hoffmann. Correctness proof of a geometric constraint
solver. International Journal of Computational Geometry and Applications,
6(4):405–420, 1996.
50
[11] I. Fudos and C.M. Hoffmann. A graph-constructive approach to solving systems of geometric constraints. ACM Transactions on Graphics, 16(2):179–
216, April 1997.
[12] H. Gao and M. Sitharam. Combinatorial classification of 2D underconstrained sytems. In Proceedings of the Seventh Asian Symposium on Computer Mathematics (ASCM 2005), pages 118–127, September 6 2005.
[13] X.-S. Gao, C. M. Hoffmann, and W. Yang. Solving spatial basic geometric
constraint configurations with locus intersection. CAD, 36:111–122, 2004.
[14] X.-S Gao, C.M. Hoffmann, and W.-Q Yang. Solving spatial basic geometric
constraint configurations with locus intersection. In Proceedings of Solid
Modeling SM’02, Saarbrucken, Germany, June, 17-21 2002. ACM Press.
[15] GeoGebra. http://www.geocebra.org/cms, July 2007.
[16] L. Henneberg. Die graphische Statik der starren Körper. Springer, 1908.
[17] M. Hidalgo and R. Joan-Arinyo. The reachability problem in constructive
geometric constraint solving based dynamic geometry. Journal of Automated
Reasoning, 44(7):709–720, 2013. DOI:10.1007/s10817-013-9280-y.
[18] C. M. Hoffmann. Computer vision, descriptive geometry and classical mechanics. In B. Falcidieno, I. Hermann, and C. Pienovi, editors, Computer
Graphics and Mathematics, pages 229–224. Springer Verlag, Eurographics
Series, 1992.
[19] C. M. Hoffmann and J. Peters. Geometric constraints for CAGD. In
M. Daehlen, T. Lyche, and L. Schumaker, editors, Mathematical Methods
for Curves and Surfaces, pages 237–254. Vanderbilt University Press, 1995.
[20] Christoph M. Hoffmann, Andrew Lomonosov, and Meera Sitharam. Finding
solvable subsets of constraint graphs. In Principles and Practice of Constraint
Programming - CP97, Third International Conference, Linz, Austria, October
29 - November 1, 1997, Proceedings, pages 463–477, 1997.
[21] Christoph M. Hoffmann, Andrew Lomonosov, and Meera Sitharam. Decomposition plans for geometric constraint problems, part ii: new algorithms.
Journal of Symbolic Computation, 31(4):409–427, 2001.
[22] Christoph M. Hoffmann, Andrew Lomonosov, and Meera Sitharam. Decomposition plans for geometric constraint systems, part i: Performance measures
for cad. Journal of Symbolic Computation, 31(4):367–408, 2001.
51
[23] Christoph M. Hoffmann and Bo Yuan. On spatial constraint solving approaches. In Revised Papers from the Third International Workshop on Automated Deduction in Geometry, ADG ’00, pages 1–15, London, UK, UK,
2001. Springer-Verlag.
[24] C.M. Hoffmann and C.-S. Chiang. Variable-radius circles of cluster merging
in geometric constraints. Part II: Rotational clusters. Computer Aided Design,
34:799–805, October 2002.
[25] C.M. Hoffmann and C.-X. Chiang. Variable-radius circles of cluster merging in geometric constraints. Part I: Translational clusters. Computer Aided
Design, 34:787–797, October 2002.
[26] C.M. Hoffmann and R. Joan-Arinyo. Distributed maintenance of multiple
product views. Computer-Aided Design, 32(7):421–431, June 2000.
[27] C.M. Hoffmann, A. Lomonosov, and M. Sitharam. Decompostion Plans for
Geometric Constraint Systems, Part I: Performance Measurements for CAD.
Journal of Symbolic Computation, 31:367–408, 2001.
[28] C.M. Hoffmann and P.J. Vermeer. Geometric constraint solving in R2 and R3 .
In D.-Z. Du and F. Hwang, editors, Computing in Euclidean Geometry, pages
266–298. World Scientific Publishing, 1995.
[29] C.M. Hoffmann and P.J. Vermeer. A spatial constraint problem. In J.P. Merlet
and B. Ravani, editors, Computational Kinematics’95, pages 83–92. Kluwer
Academic Publ., 1995.
[30] C.M. Hoffmann and B. Yuan. On spatial constrint solving approaches. In
J. Richter-Gebert and D. Wand, editors, Proceedings of ADG’2000, Zurich,
Switzerland, 2000.
[31] R. Joan-Arinyo. Triangles, ruler and compass. Technical Report LSI-95-6-R,
Department LiSI, Universitat Politècnica de Catalunya, 1995.
[32] R. Joan-Arinyo, M.V. Luzón, and A. Soto. Genetic algorithms for root multiselection in constructive geometric constraint solving. Computer & Graphics,
27(1):51–60, 2003.
[33] R. Joan-Arinyo and A. Soto. A ruler-and-compass geometric constraint
solver. In M.J. Pratt, R.D. Sriram, and M.J. Wozny, editors, Product Modeling for Computer Integrated Design and Manufacture, pages 384 – 393.
Chapman and Hall, London, 1997.
52
[34] R. Joan-Arinyo, A. Soto-Riera, S. Vila-Marta, and J. Vilaplana. Transforming
an underconstrained geometric constraint problem into a wellconstrained one.
In G. Elber and V.Shapiro, editors, Eight Symposium on Solid Modeling and
Applications, pages 33–44, Seattle (WA) USA, June 16-20 2003. ACM Press.
[35] R. Joan-Arinyo, A. Soto-Riera, S. Vila-Marta, and J. Vilaplana-Pasto. Revisiting decomposition analysis of geometric constraint graphs. In Proceedings
of the Seventh ACM Symposium on Solid Modeling and Applications, SMA
’02, pages 105–115, New York, NY, USA, 2002. ACM.
[36] R.R. Kavasseri. Variable radius circle computations in geometric constraint
solving. Master’s thesis, Computyer Sciences. Purdue University, August
1966.
[37] G Laman. On graphs and the rigidity of plane skeletal structures. J. Engineering Mathematics, 4:331–340, 1970.
[38] R.S. Latham and A.E. Middleditch. Connectivity analysis: a tool for processing geometric constraints. Computer Aided Design, 28(11):917–928, November 1996.
[39] Audrey Lee-St.John and Jessica Sidman. Combinatorics and the rigidity of
CAD systems. Computer-Aided Design, 45(2):473 – 482, 2013. Solid and
Physical Modeling 2012.
[40] M.V. Luzón, A. Soto, J.F. Gálvez, and R. Joan-Arinyo. Searching the solution
space in constructive geometric constraint solving with genetic algorithms.
Applied Intelligence, 22:109–124, 2005.
[41] I. Macdonald, J. Pach, and T. Theobald. Common tangents to four unit balls
in R 3. Discrete and Comp. Geometry, 26:1–17, 2001.
[42] James Clerk Maxwell. On reciprocal figures and diagrams of forces. Philosophical Magazine, 27:250–261, 1864.
[43] D. Michelucci and S. Foufou. Using the Cayley-Menger determinants for
geometric constraint solving. In Solid Modeling and Applications, pages 285–
290. ACM, New York, 2004.
[44] A. Noort, M. Dohem, and W.F. Bronsvoort. Solving over and underconstrained geometric models. In Geometric Constraint Solving. SpringerVerlag, Berlin, Heidelberg, 1998.
53
[45] J.C. Owen. Algebraic solution for geometry from dimensional constraints.
In R. Rossignac and J. Turner, editors, Symposium on Solid Modeling Foundations and CAD/CAM Applications, pages 397–407, Austin, TX, June 5-7
1991. ACM Press.
[46] H. Pottmann and M. Peternell. Applications of Laguerre geometry in cagd.
CAGD, 15:165–188, 1998.
[47] E. Yeguas R. Joan-Arinyo, M.V. Luzón. Search space pruning to solve the
root identification problem in geometric constraint solving,. Computer-Aided
Design and Applications, 6(1):15–25, 2009.
[48] K. Ramanathan. Variable radius circle computations in geometric constraint
solving. Master’s thesis, Department of Computer Science. Purdue University, 1996.
[49] E. C. Sherbrooke and N. M. Patrikalakis. Computation of the solutions of
nonlinear polynomial systems. CAGD, 10:379–405, 1993.
[50] M. Sitharam, A. Arbree, Y. Zhou, and N. Kohareswaran. Solution space
navigation for geometric constraint systems. ACM Transactions on Graphics,
25(2):194–213, April 2006.
[51] M. Sitharam and H. Gao. Characterizing graphs with convex and connected Cayley configuration spaces. Discrete & Computational Geometry,
43(3):594–625, 2010.
[52] M. Sitharam and M. Wang. How the beast really moves: Cayley analysis
of mechanisms realization spaces using CayMos. Computer Aided Design,
46:205–210, January 2014.
[53] M. Sitharam, M. Wang, and H. Gao. Cayley configuration spaces of
1-dof tree-decomposable linkages, Part I: Structure and extreme points.
arXiv:1112.6008v7[cs.CG], 26 Feb 2014.
[54] M. Sitharam, M. Wang, and H. Gao. Cayley configuration spaces of 1-dof
tree-decomposable linkages, Part II: Combinatorial characterization of complexity. arXiv:1112.6009v4[cs.CG], 7 Nov 2012.
[55] Tiong-Seng Tay. Rigidity of multi-graphs. i. linking rigid bodies in n-space.
Journal of Combinatorial Theory, Series B, 36(1):95 – 112, 1984.
[56] Gilles Trombettoni and Marta Wilczkowiak. Gpdof — a fast algorithm to
decompose under-constrained geometric constraint systems: Application to
54
3d modeling. International Journal of Computational Geometry and Applications, 16(05n06):479–511, 2006.
[57] H.A. van der Meiden. Semantics of Families of Objects. PhD thesis, Delft
Technical University, 2008.
[58] H.A. van der Meiden and W.F. Bronsvoort. A constructive approach to calculate parameter ranges for systems of geometric constraints. Computer-Aided
Design, 38(4):275–283, 2006.
[59] Neil White and Walter Whiteley. The algebraic geometry of motions of barand-body frameworks. SIAM J. Algebraic Discrete Methods, 8(1):1–32, January 1987.
[60] E. Yeguas, R. Joan-Arinyo, and M.V. Luzón. Modeling the performance
of evolutionary algorithms on the root identification problem: A case study
with PBIL and CHC algorithms. Evolutionary Computation, 19(1):107–135,
2011.
[61] G.-F. Zhang and X.-S Gao. Well-constrained completion and decomposition
for under-constrained geometric consraint problems. Int. Jour. of Computational Geometry & Applications, 16(5-6):461–478, 2006.
[62] Y. Zhou. Combinatorial decomposition, generic independence and algebraic
complexity of geometric constraint systems. Applications in biology and engineering. PhD thesis, University of Florida, 2006.
55
| 8cs.DS
|
arXiv:1512.00982v5 [stat.ME] 23 Jan 2017
Bayesian non-parametric inference for Λ-coalescents: posterior
consistency and a parametric method
Jere Koskela
[email protected]
Paul A. Jenkins
[email protected]
Institut für Mathematik
Technische Universität Berlin
Straße des 17. Juni 136, 10623 Berlin
Germany
Department of Statistics
University of Warwick
Coventry CV4 7AL
UK
Dario Spanò
[email protected]
Department of Statistics
University of Warwick
Coventry CV4 7AL
UK
January 24, 2018
Abstract
We investigate Bayesian non-parametric inference of the Λ-measure of Λ-coalescent processes with recurrent mutation, parametrised by probability measures on the unit interval.
We give verifiable criteria on the prior for posterior consistency when observations form a
time series, and prove that any non-trivial prior is inconsistent when all observations are
contemporaneous. We then show that the likelihood given a data set of size n ∈ N is constant across Λ-measures whose leading n−2 moments agree, and focus on inferring truncated
sequences of moments. We provide a large class of functionals which can be extremised using finite computation given a credible region of posterior truncated moment sequences, and
a pseudo-marginal Metropolis-Hastings algorithm for sampling the posterior. Finally, we
compare the efficiency of the exact and noisy pseudo-marginal algorithms with and without
delayed acceptance acceleration using a simulation study.
1
Introduction
The Λ-coalescent family is a class of coalescent processes parametrised by probability measures
on the unit interval, Λ ∈ M1 ([0, 1]), introduced by Donnelly and Kurtz [1999], Pitman [1999]
and Sagitov [1999]. We focus in particular on the family of Λ-coalescents with recurrent, finite
sites, finite alleles mutation, which we refer to simply as “Λ-coalescents” throughout this paper.
This class of processes will be introduced formally in Section 2.
In recent years, Λ-coalescents have gained prominence as population genetic models for species
with a highly skewed family size distribution, particularly among marine species [Boom et al.,
1994, Árnason, 2004, Eldon and Wakeley, 2006, Birkner and Blath, 2008], and have also been
suggested as models of evolution under natural selection [Neher and Hallatschek, 2013]. The
Λ-measure models skewness of the family size distribution. Thus, Λ represents an important
confounding factor for inference from genetic data, as well as being a quantity of interest in its
own right. Failure to properly account for uncertainty in Λ could lead to model misspecification
1
and incorrect inference. Consequently, likelihood-based inference for Λ-coalescents has also been
an active area of research [Birkner and Blath, 2008, Birkner et al., 2011, Koskela et al., 2015].
A review of Λ-coalescents and their use in population genetic inference can be found in [Birkner
and Blath, 2009], and Steinrücken et al. [2013] present a review of Beta-coalescent models for
marine species. In this paper we make the following contributions:
1. We provide the first non-parametric analysis of inferring Λ-measures from genetic data.
Our method is also the first instance of inferring Λ using the Bayesian paradigm.
2. We prove inconsistency of the posterior in full generality when data is contemporaneous,
and give verifiable criteria for consistency when the data set forms a time series.
3. We present an implementable parametrisation of the non-parametric inference problem by
quotienting the infinite dimensional space M1 ([0, 1]) in a suitable, data-driven way. We
believe this quotienting approach to have utility in infinite dimensional inference beyond
the context of this work.
4. We implement a pseudo-marginal MCMC algorithm for sampling the posterior, and provide an illustrative simulation study which demonstrates the feasibility of the algorithm
for inference.
The usual approach to inference is to focus on a parametric family of Λ-measures and infer
the parameters from observations. Common choices of parametric family are Λ(dr) = 2(2 +
ψ 2 )−1 δ0 (dr)+ψ 2 (2+ψ 2 )−1 δψ (dr) where ψ ∈ (0, 1] [Eldon and Wakeley, 2006], Λ = Beta(2−α, α)
where α ∈ (1, 2) [Birkner and Blath, 2008, Birkner et al., 2011] and Λ(dr) = cδ0 (dr) + 2−1 (1 −
c)rdr where c ∈ [0, 1] [Durrett and Schweinsberg, 2005]. We adopt the Bayesian non-parametric
approach to circumvent restrictive, finite-dimensional parametrisations. We show in Theorem 1
that the posterior is inconsistent in the typical setting of sampling from a stationary population
at a fixed time under any non-trivial prior (i.e. one which does not assign full mass to the truth),
including parametric families. In Theorem 2 we adapt a result of [Koskela et al., 2017] to provide
verifiable criteria on the prior for posterior consistency when time series data is available. We
also show in Section 5 that the popular Dirichlet process mixture model prior [Lo, 1984] satisfies
these conditions. Recent advances in sequencing technology have made genetic time series data
available [Drummond et al., 2002, Beaumont, 2003, Anderson, 2005, Drummond et al., 2005,
Bollback et al., 2008, Minin et al., 2008, Malaspinas et al., 2012, Mathieson and McVean, 2013],
and our results provide strong motivation for its continued use and development.
In Section 4 we make use of the fact that a sample of size n ∈ N carries information about
Λ ∈ M1 ([0, 1]) only via its first n−2 moments to make progress towards an implementable, nonparametric algorithm. This fact is implicit in well known sampling recursions for Λ-coalescent
processes [Möhle, 2006, Birkner and Blath, 2008, 2009] and made explicit in Lemma 4. It has
two important consequences. Firstly, it is natural to parametrise the inference problem with
truncated moment sequences because any variation in the posterior between two Λ-measures
whose first n − 2 moments coincide is due solely to the prior. Thus we obtain parametric inference algorithms requiring no discretisation or truncation, which are nevertheless as general as
any non-parametric method. Secondly, while the conditions imposed on the prior by Theorem 2
are restrictive when viewed in M1 ([0, 1]) (e.g. they rule out all of the parametric families listed
above), they are sufficiently mild that priors whose support contains an arbitrarily good approximation of the truncated moment sequence of any desired Λ ∈ M1 ([0, 1]) can be constructed.
Posterior consistency of finite moment sequences is readily inherited from posterior consistency
of the Λ-measure.
The truncated moment sequence approach can be thought of as regularising an underdetermined
inference problem by identifying an appropriate, data-driven quotient space of parameters. This
is reminiscent of the method of likelihood-informed subspaces [Cui et al., 2014], which is a
2
recently developed tool for efficient MCMC for inverse problems. The difference between existing
subspace approaches and our quotient space is that existing work has focused on projections
onto finite dimensional subspaces which preserve “most” of the likelihood in some sense, while
our method captures the likelihood function exactly. Hence we believe our work will have wider
utility beyond the Λ-coalescent setting.
We show in Theorem 3 that identifying n − 2 moments of Λ does not correspond to identifying
smaller regions in M1 ([0, 1]) with increasing n, when measured in either total variation or
Kullback-Leibler divergence. Hence the straightforward approach of computing a maximum a
posteriori moment sequence, identifying a candidate Λ∗ ∈ M1 ([0, 1]) with that moment sequence
and treating Λ∗ as a point estimator is inappropriate. Instead we use results by Winkler [1988]
to provide a broad class of functionals of Λ which can be maximised or minimised over credible
regions of the posterior. As a simple example, it is often of great interest whether the Kingman
coalescent [Kingman, 1982], Λ = δ0 , provides an adequate model for observed genetic data.
Any parametric family containing δ0 as a special case could be used to assess the Kingman
assumption, but such an approach requires justification of the parametric family. It is also
possible that δ0 is a good model within the family, but a poor fit in some broader one. Our
method avoids both of these problems.
Finally, we provide a pseudo-marginal Metropolis-Hastings algorithm [Beaumont, 2003, Andrieu and Roberts, 2009] for sampling truncated moment sequences from the posterior distribution. We compare the performance of the standard pseudo-marginal algorithm, the noisy
pseudo-marginal algorithm, as well as delayed acceptance accelerated versions of both algorithms
[Christen and Fox, 2005] in order to reduce the number of expensive likelihood estimations and
improve the computational feasibility of our inference. We find that the noisy algorithm does
not improve upon standard pseudo-marginal inference in this setting, especially given that twice
as many likelihood evaluations are required. In contrast, an off-the-shelf delayed acceptance
step dramatically speeds up computations. These results are illustrated with a simulation study
comparing the Kingman hypothesis, Λ = δ0 , based on simulated data from both the Kingman
and Bolthausen-Sznitman (Λ = U (0, 1)) coalescents. The Kingman coalescent is the classical
null model of neutral evolution, while the Bolthausen-Sznitman coalescent has recently emerged
as an alternative model in the presence of selection [Schweinsberg, 2015] or population expansion into uninhabited territory [Berestycki et al., 2013]. In particular, the Bolthausen-Sznitman
coalescent has been suggested as a model for the genetic ancestry of microbial populations such
as influenza or HIV [Neher and Hallatschek, 2013].
The rest of the paper is laid out as follows. Section 2 provides an introduction to Λ-coalescents,
Λ-Fleming-Viot jump-diffusions and a duality relation connecting the two. In Section 3 we
state and prove our consistency results. Section 4 presents our parametrisation via moments
and shows that it preserves all information in the data. Section 5 contains example families of
priors which satisfy both consistency and tractable push-forward priors on moment sequences.
Section 6 contains our results on inferring functionals of Λ based on finitely many moments. In
Section 7 we present the pseudo-marginal Metropolis-Hastings algorithm for sampling posterior
distributions of moment sequences, and an accompanying simulation study which demonstrates
the practicality of the method. Section 8 concludes with a discussion.
2
Preliminaries
Let [d] := {1, . . . , d} represent d genetic types (e.g. d = 4l for l loci of DNA), M = (Mij )di,j=1 be
a stochastic matrix specifying mutation probabilities between types, θ > 0 be the mutation rate
and Λ ∈ M1 ([0, 1]) denote the Λ-measure. Here and throughout we assume M has a unique
stationary distribution m ∈ M1 ([d]).
3
Let X := {Xt }t≥0 denote the Λ-Fleming-Viot process with recurrent, finite sites, finite alleles mutation: a jump-diffusion on the d-dimensional probability simplex Sd := {x ∈ [0, 1]d :
P
d
i=1 xi = 1} with generator
G Λ f (x) =
d
d
X
∂2
∂
Λ({0}) X
f (x) + θ
f (x)
xi (δij − xj )
xj (Mji − δij )
2
∂xi ∂xj
∂xi
i,j=1
+
d
X
i=1
Z
xi
i,j=1
[f ((1 − r)x + rei ) − f (x)] r−2 Λ(dr)
(1)
(0,1]
acting on functions f ∈ C 2 (Sd ). This process is a model for the distribution of genetic types [d]
in a large population undergoing recurrent mutation and random mating with high-fecundity
reproduction events, in which a single individual becomes ancestral to a non-trivial fraction
r ∈ (0, 1] of the whole population. As with Λ-coalescents, we will abbreviate “Λ-Fleming-Viot
process with recurrent, finite sites, finite alleles mutation” to just “Λ-Fleming-Viot process”
throughout this paper.
The first term on the right hand side (henceforth R.H.S.) of (1) models diffusion of the allele
frequencies due to random mating, or genetic drift in the terminology of population genetics. The
second term models mutation, and the third (jump) term models high-fecundity reproductive
events. Without the jump term, i.e. when Λ = δ0 , {Xt }t≥0 reduces to the classical d-dimensional
Wright-Fisher diffusion with recurrent mutation. We denote the law of a Λ-Fleming-Viot process
Λ
with initial condition x ∈ Sd by PΛ
x (·), and expectation with respect to this law by Ex [·]. We
suppress dependence on initial conditions whenever the stationary process is meant. For bounded
Λ
f : Sd 7→ R, let PtΛ f (x) := EΛ
x [f (Xt )] be the associated transition semigroup, pt (x, y) be the
transition density and π Λ (x) be the corresponding stationary density on Sd , assumed unique.
The transition semigroup is Feller for any Λ ∈ M1 ([0, 1]) [Bertoin and Le Gall, 2003], and all
densities are assumed to exist with respect to the d-dimensional Lebesgue measure dx.
A realisation of a Λ-Fleming-Viot process specifies the relative frequencies of genetic types
in an infinite population across time. It is natural to imagine sampling n ∈ N individuals
from the population at a fixed time, and ask about the ancestral tree connecting the sampled
individuals. This ancestral tree is a random object, and is described by the Λ-coalescent, denoted
by Π := {Πt }t≥0 , taking values in Pnd , the set of [d]-labelled partitions of [n] [Donnelly and Kurtz,
1999, Section 5]. The process is started from the unlabelled partition ψn := {{1}, {2}, . . . , {n}},
and propagates backwards in time from the point of view of the reproductive evolution of the
population. When the process has p ∈ N blocks, any 2 ≤ k ≤ p of them merge at rate
Z
λp,k := Λ({0})1{2} (k) +
(1 − r)p−k rk−2 Λ(dr),
(0,1]
where 1A (k) is the indicator function of the set A. We will abbreviate λk,k ≡ λk throughout
the paper. Once the process has merged into a single block, known as the most recent common
ancestor (MRCA), an ancestral type is sampled from an initial law ν ∈ M1 ([d]). We focus
on the case of a stationary population, in which case ν = m. This type is inherited along the
branches of the Λ-coalescent tree, with mutations occurring at rate θ and mutant types sampled
from M . The result is a random labelling of partitions for all times 0 ≤ t ≤ T , where T denotes
Λ
the hitting time of the MRCA. Let PΛ
n denote the law of Π started from ψn , and En denote
the corresponding expectation. For the entirety of the paper we assume θ and M are known,
and focus on inferring Λ. This assumption is crucial for the proof of consistency of Bayesian
inference from time series data (Theorem 2 in Section 3), but not needed for correctness of the
finitely many moments parametrisation in Section 4, or of the algorithms in Section 7.
The following relationship between Λ-Fleming-Viot processes and corresponding Λ-coalescents
is classical [Bertoin and Le Gall, 2003], and will be useful in the (in)consistency proofs in the
4
next section:
"
Λ
E
d
Y
#
ni
Xt (i)
"
=
EnΛ
i=1
d
Y
#
m(i)
|Πt (i)|
,
(2)
i=1
where ni denotes the number of observed individuals with label i ∈ [d] sampled i.i.d. from
the random measure Xt , and |Πt (i)| denotes the number of blocks in partition Πt with label
i ∈ [d]. Formula (2) is an example of so called moment duality between stochastic processes
(see e.g. [Möhle, 1999], and references therein for details): Intuitively, (2) states that the law of
the type frequencies of Λ-coalescent leaves coincides with a multinomial sample from a random
measure drawn from the corresponding Λ-Fleming-Viot process.
3
Posterior consistency
d
Let Q ∈ M1 (M1 ([0, 1])) be
Pda prior distribution for Λ, and n = (n1 , . . . , nd ) ∈ N denote observed
type frequencies of n := i=1 ni [d]-labelled lineages generated by the Λ-coalescent. For Borel
sets A ∈ B(M1 ([0, 1])), define the posterior as
R
PΛ (n)Q(dΛ)
.
Q(A|n) = R A n Λ
M1 ([0,1]) Pn (n)Q(dΛ)
Informally, posterior consistency holds when Q(·|n) concentrates on a neighbourhood of the
Λ0 ∈ M1 ([0, 1]) which generated n as n → ∞. This is a natural requirement for statistical
inference as it ensures the truth can be learned from a sufficient amount of data. For an
overview of Bayesian non-parametric statistics, the reader is directed at [Hjort et al., 2010] and
references therein.
It is well known that non-parametric posterior consistency is highly sensitive to the details of
the topology defining the neighbourhood system as well as the mode of convergence [Diaconis
and Freedman, 1986]. This will also be the case for our consistency result for time series data,
Theorem 2. In contrast, the inconsistency result for contemporaneous observations, Theorem 1,
is very universal. Hence we postpone specification of these details until after Theorem 1.
Remark 1. In Theorem 1 below, we give a formal statement of inconsistency from the point of
view of Bayesian estimators, which are our interest in this paper. Elements of the proof of Theorem 1 will also be useful in the proving Theorem 2 later in the paper. However, we emphasize
that the negative result also holds for frequentist estimators based on contemporaneous data.
Essentially the same argument used to prove Theorem 1 shows that the limiting likelihood
lim EΛ [q(n|X)] = π Λ (x)
n→∞
is positive for any observation x and any Λ, at least provided {π Λ }Λ is a bounded family of
stationary densities so that the Dominated Convergence Theorem holds without the regularising
effect of the prior. Hence, estimators cannot converge to the true Λ-measure generating the
data. Ultimately, the problem is that the d-dimensional vector corresponding to one exact
draw from a stationary Λ-Fleming-Viot process cannot be expected to uniquely identify an
infinite dimensional object. Similar problems of identifiability also occur in estimation of lifetime
distributions in branching processes [Höpfner et al., 2002, Hoffmann and Olivier, 2016], for which
identifiability issues can be overcome by letting the length of the observation window grow to
infinity, analogously to Theorem 2 below.
Theorem 1. Let n ∈ Nd denote the observed type frequencies in a sample of size n ∈ N generated
by a Λ-coalescent started from ψn at a fixed time, and let x := limn→∞ nn ∈ M1 ([d]) denote the
5
limiting observed relative type frequencies. Then the limiting posterior is given by
R Λ
π (x)Q(dΛ)
lim Q(A|n) = R A
.
Λ
n→∞
M1 ([0,1]) π (x)Q(dΛ)
In particular, the R.H.S. is positive for any A ∈ B(M1 ([0, 1])) which intersects the support of
Q, regardless of the Λ ∈ M1 ([0, 1]) generating the data.
Proof. Conditioning on the ancestral tree of the observed sample give the following representation for the posterior:
R
R
EnΛ 1{n} (Π0 ) Q(dΛ)
PΛ
n (n)Q(dΛ)
A
A
.
Q(A|n) = R
=R
Λ
Λ
M1 ([0,1]) Pn (n)Q(dΛ)
M1 ([0,1]) En 1{n} (Π0 ) Q(dΛ)
Using (2) we can write
Λ
A E [q(n|X)] Q(dΛ)
,
Λ
M1 ([0,1]) E [q(n|X)] Q(dΛ)
R
Q(A|n) = R
Qd
ni
n
where q(n|X) := n1 ,...,n
i=1 Xi is the multinomial sampling probability. We will show the
d
requisite convergence of the numerator and denominator separately, and the result will follow
by the algebra of limits. We begin by considering the numerator.
By Fubini’s theorem
Z
Z
Λ
E [q(n|X)] Q(dΛ) =
Z
q(n|y)
Sd
A
Z
Λ
π (y)Q(dΛ)dy =:
q(n|y)FQ;A (y)dy,
Sd
A
where FQ;A (y) is a sub-probability density on Sd since it is a mixture of probability densities.
Hence FQ;A (y)dy defines a finite measure on Sd , and q(n|y) ≤ 1 so that by the Dominated
Convergence theorem
Z
Z
lim q(n|y)FQ;A (y)dy.
q(n|y)FQ;A (y)dy =
lim
Sd n→∞
n→∞ S
d
By the Law of Large Numbers n ∼ bnxc so that q(n|y) → q(bnxc|y), and by Stirling’s formula
q(bnxc|y) ∼
d
Y
yi nxi
i=1
or
log(q(bnxc|y)) ∼ n
d
X
i=1
xi log
yi
xi
= −n
xi
d
X
,
xi log
i=1
xi
yi
= −n KL(x, y),
where KL(x, y) denotes the Kullback-Leibler divergence between the probability mass functions
x and y. By Gibbs’ inequality KL(x, y) ≥ 0 and KL(x, y) = 0 if and only if x = y, so that
q(n|·) → δx (·) almost surely. Hence
Z
Z
lim q(n|y)FQ;A (y)dy = FQ;A (x) =
π Λ (x)Q(dΛ),
Sd n→∞
A
as required. The argument for the denominator is identical after substituting M1 ([0, 1]) for the
domain of integration A.
6
Remark 2. There is an apparent contradiction between the negative conclusion of Theorem 1
and recent positive results [Spence et al., 2016, Theorems 2, 3, 4 and 5] showing that Λ-measures
can often be identified from their site frequency spectra. The contradiction is resolved by noting
that Spence et al. [2016] work directly with the expected site frequency spectrum, thereby
sidestepping both the randomness of the ancestral tree and the randomness of the mutation
process given the tree. Numerical investigations by Spence et al. [2016] show that their method
is unreliable unless a modest number (10-100) of independent realisations of ancestral trees is
available. Independent trees cannot be sampled from populations whose ancestry is described
by any non-Kingman Λ-coalescent, even in the idealised scenario of an infinitely long genome
in the presence of recombination. However, as noted by [Spence et al., 2016], in practice the
decay of correlations with increasing genome length is determined by the prelimiting model of
evolution, and not necessarily the limiting Λ-coalescent. For example, the selective sweep model
of Durrett and Schweinsberg [2005] can allow for asymptotically independent trees across a
genome in the presence of multiple mergers for some combinations of parameters, in which case
the identifiability results of Spence et al. [2016] hold.
The following example is an extension of a result by Der and Plotkin [2014], and demonstrates
that the lack of consistency can have dramatic consequences for statistical identifiability even
in the case of very simple priors.
Example 1. Consider d = 2, Mij = 1/2 for i, j ∈ {1, 2}, and set Q(dΛ) = 12 δδ0 (dΛ) + 12 δδ1 (dΛ).
The stationary law π Λ (x) is known in the parent-independent, two-allele case for both of these
atoms [Der and Plotkin, 2014]:
Γ(2θ) θ−1
x (1 − x)θ−1
[Γ(θ)]2
1−θ
1
π δ1 (x) = |1 − 2x| θ ,
θ
π δ0 (x) =
so the expected limiting posterior probabilities can be computed assuming either data-generating
measure. These are listed in Table 1 for some candidate values of θ, while Figure 1 depicts limiting posterior probabilities as functions of the observed allele frequencies. The extreme sensitivity
of the posterior probabilities in Figure 1 is akin to the “Bayesian brittleness” investigated by
Owhadi et al. [2015], resulting in inferences which are not robust to small changes in the observed
allele frequencies, prior probabilities or latent parameters.
θ
0.04
0.1
0.5
1
5
10
17
Eδ0 [Q(δ0 |X)]
0.84
0.73
0.54
0.50
0.65
0.75
0.82
Eδ1 [Q(δ0 |X)]
0.16
0.27
0.46
0.50
0.35
0.25
0.18
Table 1: Expected posterior probabilities given an infinite number of contemporaneous observations in the parent-independent, two-allele model.
The fact that π δ0 (x) = π δ1 (x) when θ = 1 in Example 1 was pointed out by Der and Plotkin
[2014] as proof of the fact that Λ-measures cannot in general be uniquely identified from independent draws from π Λ (x). Our calculations illustrate that inference suffers from low power and
poor stability even when θ 6= 1 if all observations are contemporaneous.
The inconsistency result of Theorem 1 holds for essentially arbitrary priors. Our next aim is
to show that the posterior can be consistent when the data set is a time series of increasing
7
3.0
2.5
2.0
Density
0.0
0.5
1.0
1.5
2.0
1.5
0.0
0.5
1.0
Density
2.5
3.0
3.5
Theta = 10
3.5
Theta = 0.1
0.0
0.2
0.4
0.6
0.8
1.0
0.0
x
0.2
0.4
0.6
0.8
1.0
x
Figure 1: Limiting posterior probabilities as functions of the observed allele frequency x for
θ = 0.1 and 10 in the parent-independent, two-allele model. Q(δ0 |x) is dotted, Q(δ1 |x) is dashdotted, π δ0 (x) is solid and π δ1 (x) is dashed. Note the extreme sensitivity of the posterior to the
observed allele frequencies near x = 0.5 when θ = 10 and near x = 0 or 1 when θ = 0.1.
length. This does not contradict the unidentifiability claim of Der and Plotkin [2014], because
the authors only consider independent draws from π Λ . In contrast, in our setting it is information
about transition densities pΛ
∆ (x, y) which facilitates posterior consistency.
We begin by defining the topology and weak posterior consistency following the set up of van der
Meulen and van Zanten [2013], who considered similar time series data for one dimensional
diffusions. In addition to topological details, posterior consistency is highly sensitive to the
support of the prior, which should not exclude the truth. This is usually guaranteed by insisting
that the prior places positive mass on all neighbourhoods of the truth, typically measured in
terms of Kullback-Leibler divergence. In our setting such a support condition is provided by (5)
below.
Definition 1. Fix η > 0 and let Dη be a collection of Lebesgue probability densities on [η, 1]
satisfying inf r∈[η,1] φ(r) > 0 and supr∈[η,1] φ(r) < ∞ for each φ ∈ Dη . We assume that Λ(dr) =
φ(r)dr for φ ∈ Dη , and denote the data generating density by φ0 .
Restricting the support of φ to [η, 1] ensures that the Λ-coalescent can have no Kingman component, and that the Λ-Fleming-Viot process is a compound Poisson process with drift. Furthermore, most previously studied parametric families of Λ-measures are ruled out, including all
those mentioned in Section 1. However, we will see in Section 4 that the prior can be chosen to
satisfy the conditions of Definition 1 and place mass arbitrarily close to any desired Λ-measure,
or family of Λ-measures, in a way we will make precise in Example 2.
Before we define and prove consistency of Bayesian nonparametric inference, we need to first
establish that Λ is identifiable from discrete observations. This is done in Lemma 1 below.
Lemma 1. For any pair φ0 6= φ ∈ Dη and any ∆ > 0 there exists x ∈ Sd and a test function f
φ0
φ
φ
such that P∆
f (x) 6= P∆
f (x). In particular, identifying P∆
is equivalent to identifying φ.
Proof. Let φ0 ∈ Dη and φ ∈ Dη agree on their first n−3 moments, and suppose that the (n−2)th
moments λ0n and λn differ, as must be the case for some n ≥ 3 if φ0 6= φ. We begin by using the
spectral representation of Griffiths [2014] to show that certain eigenvalues of the corresponding
Λ-Fleming-Viot generators G φ0 and G φ are different.
By [Griffiths, 2014, Theorem 5], the eigenvalues are naturally indexed with multi-indices n ∈
Nd . Let qn0 and qn denote the eigenvalues of G φ0 and G φ with index n, respectively. Define
8
independent random variables (U, Y, V ) ∼ 2u1[0,1] ⊗φ⊗U (0, 1), and let W := U Y . By [Griffiths,
2014, equation (41)], we have that qn can be written as
n
qn =
E[(1 − W )n−2 ] + C(θ, M, n),
2
P
for each n such that di=1 ni = n, where the expectation is with respect to the law of W , and
C(θ, M, n) is a constant depending only on its arguments. A binomial expansion yields
n−2
n−2
X n − 2
X n − 2
2
n−2
k
k
E[(1 − W )
]=
(−1) E[W ] =
(−1)k
λk+2 .
k
k
k+2
k=2
k=2
All terms on the R.H.S. coincide for qn and qn0 except the k = n − 2 term, for which
(−1)n−2 2
(−1)n−2 2 0
λn 6=
λn
n
n
0 and q for two densities φ and φ coincide for multiby assumption. Hence, the eigenvalues qm
m
0
indices of total degree up to m ≤ n − 1, and eigenvalues for multi-indices of total degree n differ
when n − 2 is the order of the first moment which is distinct between φ0 and φ.
Next, we use the above result on eigenvalues, in conjunction with [Griffiths, 2014, Theorem 5],
to show that G φ0 and G φ also have some distinct right eigenfunctions. To that end, let h0n (x)
and hn (x) be the respective eigenfunctions of G φ0 and G φ with eigenvalues qn0 and qn . Applying
the representation of G φ in [Griffiths, 2014, equation (39)] to the monomial test function
q(m|x) :=
d
Y
i
xm
i ,
i=1
as well as a binomial expansion to the resulting [(1 − W )xi + W V ]mi −1−δij terms yields
mi −2
d
X
mi X
(1 − W )m−k−2 (W V )k q(m − (k + 1)ei |x)
G φ q(m|x) =
2
i=1
−
d
X
k=0
mi −1−δij
mi (mj − δij )
i,j=1
X
k=0
mi − 1 − δij
(1 − W )m−k−2 (W V )k q(m − kei |x)
k
d
d
X
X
∂
+θ
Mji xj − xi
q(m|x).
∂xi
i=1
(3)
j=1
The fact that the R.H.S. depends on φ only via powers of W up to order n − 2 makes it clear
that the action of the generators G φ0 and G φ coincide on polynomials of degree m < n, which
the eigenfunctions h0m (x) and hm (x) are for m < n by [Griffiths, 2014, Theorem 5].
Now consider h0n (x) and hn (x) for a fixed n of total degree n. Again by [Griffiths, 2014, Theorem
5], the total degree n terms of both equal ξ n , and hence coincide. See [Griffiths, 2014, Theorem
5] for the definition of ξ as a linear, M -dependent transformation of x. We will focus instead
on terms of total degree n − 1, and to that effect introduce the representations
hn (x) = ξ n +
d
X
an(n−ei ) hn−ei (x) +
d
X
d
X
X
bn(n−ei ) hn−ei (x) +
bnm hm (x)
m:m<n−1
i=1
−
anm hm (x)
m:m<n−1
i=1
G φ hn (x) = −qn hn (x) +
X
an(n−ei ) qn−ei hn−ei (x) −
X
m:m<n−1
i=1
9
anm qm hm (x)
(4)
where the coefficients anm and bnm must satisfy anm qm = bnm in order for hn (x) to be an
eigenfunction, as first three terms on the R.H.S. of (4) arise as the definition of an eigenfunction
expansion of G φ ξ n . Let a0nm and b0nm be the analogous coefficients for the same representations
of the polynomials h0n and G φ0 h0n .
For each m, the only term of total degree m in hm (x) is ξ m . Therefore, if b0n(n−ei ) 6= bn(n−ei )
0
for some i ∈ [d], then a0n(n−ei ) 6= an(n−ei ) since qn−e
= qn−ei , and we are done because then
i
hn (x) contains at least one term which has a coefficient different to that of the corresponding
term in h0n (x). Terms of total degree n − 1 arise in the first line of (3) by taking k = 0, in the
second by taking k = 1, and do not arise in the third. In particular, the coefficient of q(n − ei |x)
on the R.H.S. of (3) is
ni
ni
n−2
E[(1 − W )n−3 W V ]
E[(1 − W )
] − (n − 2)
2
2
#
"
n−2
X n − 2
ni
k2+k
1+
(−1)
=
λk+2
2
k
2
k=1
These coefficients all differ for G φ0 and G φ because they can be written as the same linear
combinations of the first n − 2 moments of φ0 and φ, respectively. The lower degree terms with
coefficients anm with m < n − 1 cannot contribute to the coefficients of terms of degree n − 1.
Finally, the coefficients of xn−ei and ξ n−ei coincide because ξ is a linear transformation of x.
In short, the eigenfunction hn (x) has the form
i
h
P
n−2
ni
d
(−1)k 2+k
1 + n−2
λk+2
X
k=1
k
2
2
hn (x) = ξ n +
ξ n−ei + lower order terms,
n−1
n−3
] + Cn−ei ,θ,M
2 E[(1 − W )
i=1
and the coefficients of the degree n − 1 terms all differ between h0n (x) and hn (x).
φ0
φ
hn (x) = e−qn ∆ hn (x). The former is a
Finally, fix ∆ > 0, and consider P∆
hn (x) and P∆
polynomial, but cannot be a scalar multiple of the latter because otherwise hn (x) would be an
φ0
eigenfunction of G φ0 , which we have shown is not the case. Hence, the two polynomials P∆
hn (x)
φ
and P∆
hn (x) are distinct, and thus differ on some non-empty, open set, which concludes the
proof.
Definition 2. Fix a sampling interval ∆ > 0 and a finite Borel measure ν on Sd placing positive
mass in all non-empty open sets. A weak topology on Dη is generated by open sets of the form
0
φ
φ
φ
Uf,ε
:= {φ0 : kP∆
f (x) − P∆
f (x)k1,ν < ε},
for any φ ∈ Dη , ε > 0 and f ∈ Cb (Sd ), the set of continuous and bounded functions on Sd , where
k · kp,ν is the Lp (Sd , ν)-norm. The Lebesgue measure is meant whenever no measure is specified.
φ
Lemma 1 of [Koskela et al., 2017] yields that the topology generated by Uf,ε
is Hausdorff, and
hence separates points.
Definition 3. Let n0 , . . . , nm denote m + 1 samples observed at times 0, ∆, . . . , ∆m from a
stationary Λ-coalescent, with each sample being of size n ∈ N. See e.g. [Beaumont, 2003] for
details of how temporally structured samples can be generated. Weak posterior consistency
holds if Q(Uφc0 |n0 , . . . , nm ) → 0 Pφ0 -a.s. as n, m → ∞, where Uφ0 is any open neighbourhood of
φ0 ∈ Dη .
Theorem 2. Let n0 , . . . , nm be as in Definition 3 and x0 , . . . , xm denote the observed limiting
type frequencies as n → ∞, i.e. xi = limn→∞ ni /n. Suppose that Dη , the support of Q, satisfies
10
the conditions of Definition 1. Furthermore, for any ε > 0 and any φ0 ∈ Dη suppose that
Z 1
φ0 (r)
φ0 (r)
−2
Q φ ∈ Dη :
log
+
− 1 r φ0 (r)dr < ε > 0.
φ(r)
φ(r)
η
(5)
Then weak posterior consistency holds for Q on Dη .
Remark 3. A similar result for jump diffusions with unit diffusion coefficient was established
in [Koskela et al., 2017, Theorem 1], and our proof will follow a similar structure. Before
presenting the proof, let us highlight how the present result differs from the jump diffusion case.
Both proofs of consistency require verification of a Kullback-Leibler condition for the prior, and
uniform equicontinuity of the family of semigroups corresponding to densities supported by the
prior. The former result is immediate by the same argument used to prove [Koskela et al., 2017,
Lemma 2], whose statement is provided below in Lemma 2 in the interest of a self-contained
proof. The latter, Lemma 3 below, is different to its counterpart, [Koskela et al., 2017, Lemma
3], which relies on positive definiteness of the diffusion coefficient.
Proof. For fixed m ∈ N, the same argument used to prove Theorem 1 yields that the following
convergence holds Pφ0 -a.s. as n → ∞:
lim Q(dφ|n0 , . . . , nm ) ∝ π φ (x0 )
n→∞
m
Y
pφ∆ (xi−1 , xi )Q(dφ).
i=1
Hence it is sufficient to establish posterior consistency for m + 1 exact observations from a
stationary Λ-Fleming-Viot process as m → ∞. We achieve this by adapting the proof of [Koskela
et al., 2017, Theorem 1], which entails verifying two conditions. The first is that the prior places
sufficient mass in Kullback-Leibler neighbourhoods of φ0 , i.e. that for any ε > 0 we have
Q(φ ∈ Dη : KL(φ0 , φ) < ε) > 0,
(6)
where KL is Kullback-Leibler divergence between pφ∆0 and pφ∆ :
!
Z Z
pφ∆0 (x, y)
KL(φ0 , φ) :=
log
pφ∆0 (x, y)π φ0 (x)dydx.
φ
p∆ (x, y)
Sd Sd
φ
: φ ∈ Dη }:
The second is establishing uniform equicontinuity of {P∆
sup
sup
φ∈Dη kx−yk2 <δ
φ
φ
f (y)| < ε
|P∆
f (x) − P∆
for each ∆ > 0 and f ∈ Cb (Sd ).
Condition (6) follows from a straightforward modification of [Koskela et al., 2017, Lemma 2].
A statement of this result, adapted to the present context and notation, is provided below. Its
proof follows the same structure that of as [Koskela et al., 2017, Lemma 2], and is omitted.
Lemma 2. Condition (5) implies condition (6) for any ε > 0.
φ
It remains to establish uniform equicontinuity on the semigroup {P∆
f : φ ∈ Dη } for f ∈ Cb (Sd ).
This can be done by verifying the hypotheses of [Koskela et al., 2017, Lemma 3], which we will
do below.
φ
Lemma 3. For each ∆ > 0 and f ∈ Cb (Sd ), the collection {P∆
f : φ ∈ Dη } is uniformly
equicontinuous: for any ε > 0 there exists δ := δ(ε, f, ∆) > 0 such that
sup
sup
φ∈Dη kx−yk2 <δ
φ
φ
|P∆
f (x) − P∆
f (y)| < ε.
11
Proof. We begin by showing that the required uniform equicontinuity is true for f ∈ Lip(Sd ),
the set of Lipschitz functions on Sd .
By [Wang, 2010, Proposition 2.2, in particular equation (2.2)], a sufficient condition for equicontinuity for a fixed φ ∈ Dη is that for kx − yk2 ≤ δ we have
θ(x − y)T (M − Id )(x − y)
+
kx − yk2 (1 + kx − yk2 )
Z
d
1X
η
|xi − yi |r−2 φ(r)dr ≤ Cδ kx − yk2 ,
(7)
i=1
where Id is the d × d identity√matrix. Now the first term is trivially bounded by θkM − Id k2 kx −
yk2 , and the second by η −2 dkx − yk2 using the Cauchy-Schwarz inequality. Still by [Wang,
2010, Proposition 2.2], for f ∈ Lip(Sd ) and kx − yk2 ≤ δ we have the bound
φ
φ
|P∆
f (x) − P∆
f (y)| ≤ (2 + δ −1 )eCδ ∆ C(f )kx − yk2 ,
where C(f ) is a constant depending only on f . Uniformity in φ ∈ Dη now follows from the fact
that the constant Cδ in (7) is independent of φ.
Now consider a general test function f ∈ Cb (Sd ). The simplex Sd is compact, meaning that
Lipschitz functions are dense in Cb (Sd ). Thus for any β > 0, there exists g ∈ Lip(Sd ) be such
that kg − f k∞ < β. The triangle inequality then yields the elementary bound
sup
sup
φ∈Dη kx−yk2 <δ
≤ sup
φ
φ
|P∆
f (x) − P∆
f (y)|
sup
φ∈Dη kx−yk2 <δ
o
n
φ
φ
φ
φ
φ
φ
g(y)| .
g(x) − P∆
g(y)| + |P∆
f (y) − P∆
g(x)| + |P∆
f (x) − P∆
|P∆
Now, the first term on the R.H.S. can be bounded by
φ
φ
g(x)| = Eφx [f (Xt ) − g(Xt )] ≤ Eφx [|f (Xt ) − g(Xt )|] < β
f (x) + P∆
|P∆
by construction of g. The second term is bounded analogously. The last term can be made
arbitrarily small by choice of sufficiently small δ. For fixed f , all three bounds are uniform in
φ, which concludes the proof.
The remainder of the proof follows as in [Koskela et al., 2017]. It suffices to show that for
φ
φ0
f ∈ Cb (Sd ) and B := {φ ∈ Dη : kP∆
f − P∆
f k1,ν > ε} we have Q(B|x0 , . . . , xm ) → 0 Pφ0 -almost
surely. To that end we fix f ∈ Lip(Sd ), ε > 0 and thus the set B. Condition (6) implies that
[van der Meulen and van Zanten, 2013, Lemma 5.2] holds, so that if for a measurable collection
of subsets Cm ⊂ Dη there exists c > 0 such that
e
mc
Z
φ
π (x0 )
Cm
m
Y
pφ∆ (xi−1 , xi )Q(dφ) → 0
i=1
Pφ0 -almost surely, then Q(Cm |x0 , . . . , xm ) → 0 Pφ0 -almost surely as well. Likewise, Lemma
3 implies [van der Meulen and van Zanten, 2013, Lemma 5.3]: there exists a compact subset
K ⊂ Sd , N ∈ N and compact, connected sets I1 , . . . , IN that cover K such that
B⊂
N
[
Bj+ ∪
j=1
where
Bj±
N
[
Bj− ,
j=1
φ
φ0
:= φ ∈ Dη : P∆
f (x) − P∆
f (x) > ±
12
ε
for every x ∈ Ij
4ν(K)
.
Thus it is only necessary to show Q(Bj± |x0 , . . . , xm ) → 0 Pφ0 -almost surely. Define the stochastic
process
!1/2
Z
m
Y
Dm :=
π φ (x0 )
pφ∆ (xi−1 , xi )Q(dφ)
.
Bj+
i=1
Now Dm → 0 exponentially fast as m → ∞ by an argument identical to that used to prove
[van der Meulen and van Zanten, 2013, Theorem 3.5 ]. The same is also true of the analogous
stochastic process defined by integrating over Bj− , which completes the proof.
In the next section we show that given a data set of size n, it is natural to infer the first
n − 2 moments of Λ because they fully capture the signal in the data set. Example 2 at the
end of Section 4 provides a family of priors which satisfy the hypotheses of Theorem 2, and
whose support can be chosen to contain arbitrarily close approximations to truncated moment
sequences of any Λ ∈ M1 ([0, 1]).
Remark 4. The hypotheses of Theorem 2 are strong, and thus it would be desirable to obtain a
posterior contraction rate in addition to just consistency. In fact, methods akin to that employed
in the proof have been extended to provide rates for compound Poisson processes [Gugushvili
et al., 2015] and scalar diffusions on compact intervals [Nickl and Söhl, 2015]. However, extending
either approach to our setting would require bounds of the form
kπ φ − π φ0 k2 ≤ Cn−β ,
kπ φ /π φ0 − 1k2 ≤ C̃n−β̃
for some constants β, β̃, C, C̃ > 0. Since the Λ-Fleming-Viot stationary density is intractable in
nearly all cases, it does not seem possible to extend our approach to obtain rates of posterior
consistency.
4
Parametrisation by finitely many moments
Consider a set of type frequencies n ∈ Nd of size n :=
finite alleles mutation started from ψn .
Pd
i=1 ni
generated by a Λ-coalescent with
λ3 ,λ4 ,...,λn
Lemma 4. The likelihood satisfies PΛ
(n). That is, Λ is conditionally inden (n) = Pn
n
pendent of n given {λk }k=3 .
P
n
Proof. Let −qnn = n−1
k=1 n−k+1 λn,n−k+1 be the total merger rate of the Λ-coalescent with n
blocks. It is well known that the Λ-coalescent likelihood is the unique solution to the recursion
[Möhle, 2006, Birkner and Blath, 2008, 2009]:
PΛ
n (n)
d
X
θ
=
(nj − 1 + δij )Mji PΛ
n (n − ei + ej )
nθ − qnn
i,j=1
ni
X X
1
n
ni − k + 1 Λ
+
λn,k
P
(n − (k − 1)ei ),
nθ − qnn
k
n − k + 1 n−k+1
(8)
i:ni ≥2 k=2
with boundary condition PΛ
1 (ei ) = m(i). Repeated application of this recursion yields a closed
system of linear equations for the likelihood because all sample sizes on the R.H.S. are equal
to or smaller than the one on the left hand side. This system is far too large to solve for all
but very small sample sizes, but it is clear that the solution can depend on Λ only through the
polynomial moments {λm,k }nk≤m=2 .
13
Polynomial moments can be written as a linear combination of monomial moments:
λm,k =
m−k
X
j=0
m−k
(−1)j λk+j ,
j
(9)
meaning that only the monomial moments {λk }nk=2 are required. Since λ2 =
moments {λk }nk=3 are sufficient.
R1
0
Λ(dx) = 1, the
Motivated by Lemma 4 we make the following definition:
Definition 4. Let ∼n be the equivalence relation on M1 ([0, 1]) defined via
(1)
(2)
Λ1 ∼n Λ2 if λk = λk for k ∈ {3, . . . , n}
(i)
where λk :=
R1
0
xk−2 Λi (dx). We call the equivalence classes of ∼n moment classes of order n.
In view of Lemma 4 it is natural to consider the problem of inferring Λ from n in the quotient
space M1 ([0, 1])/ ∼n , not in M1 ([0, 1]). Moreover, requiring all linear combinations of the form
(9) to be non-negative guarantees a unique solution to the Hausdorff moment problem, so that
each completely monotonic moment sequence bounded by 1 corresponds to some Λ ∈ M1 ([0, 1]).
Hence we parametrise the space M1 ([0, 1])/ ∼n by truncated, completely monotonic moment
sequences of length n − 2 with leading term λ3 ≤ 1. This approach yields a compact, finitedimensional parameter space which nevertheless captures all the signal in the data. Table 2 lists
some moment sequences corresponding to popular families of Λ-measures.
Λ
λk
δ0
0
δ1
1
Beta(2 − α, α)
(2−α)k−2
(2)k−2
U (0, 1)
2
δ
2+ψ 2 0
1
k−1
+
ψ2
δ
2+ψ 2 ψ
ψk
2+ψ 2
cδ0 +
(1−c)
2 rdr
1−c
2k
Table 2: Moment sequences of particular Λ-coalescents. Here (a)k := a(a + 1) . . . (a + k − 1)
denotes the rising factorial.
Naturally, the prior Q ought to be chosen to yield a tractable push-forward prior on truncated
moment sequences. These push-forward priors inherit posterior consistency whenever Q satisfies
the conditions of Theorem 2 because truncated moment sequences can be written as bounded
functionals of Λ.
5
Prior distributions
In this section we provide an example family of priors which satisfies the consistency criteria of
Theorem 2 and has tractable push-forward distributions on truncated moment sequences.
Definition 5. Let Q ∈ M1 (M1 ([0, 1])) be a prior distribution for Λ. Then the moments {λk }nk=3
have joint prior Qn on the space of completely monotonic sequences of length n − 2 given by
!
Z
Z
n
Y
k−2
Qn (λ3 ∈ dy3 , . . . , λn ∈ dyn ) :=
1{dyk }
r Λ(dr) Q(dΛ).
(10)
M1 ([0,1]) k=3
(0,1]
The prior Q should to be chosen such that the R.H.S. of (10) is tractable, and the following
example illustrates that such a choice is possible.
14
Example 2. Fix η > 0 and α ∈ M([η, 1]) with finite mass and a strictly positive Lebesque
density α(r). Suppose Dη satisfies the conditions of Definition 1, and in addition that every
φ ∈ Dη is continuous. Let R(dτ ) be a probability measure on (0, ∞) placing positive mass in all
non-empty open sets. For x ∈ [η, 1] and τ > 0 let
qx,τ (r) :=
1[η,1] (r − x)hτ (r − x)
,
hτ ([η, 1])
where hτ is the Gaussian density on R with mean 0 and variance τ −1 .
Let DP (α) be the law of a Dirichlet process centred on α [Ferguson, 1973] and let Q be given by
the Dirichlet process mixture distribution [Lo, 1984] with mixing distribution DP (α) ⊗ R and
mixture components qx,τ . In other words, let F ∼ DP (α) and F (r1 ), F (r2 ), . . . be the weights
of ordered atoms of F . Let {τi }∞
i=1 be i.i.d. draws from R. Then a draw from Q is given by
φ(r)|F, {τi }∞
i=1 =
∞
X
F (ri )qri ,τi (r).
i=1
The prior Q places full mass on equivalent densities bounded from above and away from 0 by
construction. We assume Q satisfies Lemma 1, at which point it remains to check (5) to verify
Q has posterior consistency. Theorem 1 of [Bhattacharya and Dunson, 2012] yields that for any
δ > 0 the prior places positive mass in all open balls: Q(φ ∈ Dη : kφ − φ0 k∞ < δ) > 0 for any
φ0 ∈ Dη under the above assumptions on α, R and (qx,τ )x∈[η,1],τ >0 . Now fix ε > 0, φ0 ∈ Dη , as
well as 0 < δ < c such that
c
c
δ
log
∨ log
+
< η 2 ε.
c−δ
c+δ
c−δ
Then
Z
Q φ ∈ Dη :
1
log
η
φ0 (r)
φ(r)
φ0 (r)
+
−1
φ(r)
r
−2
φ0 (r)dr < ε
≥ Q(φ ∈ Dη : kφ − φ0 k∞ < δ) > 0,
because for any φ satisfying kφ − φ0 k∞ < δ we have
)
Z 1(
φ0 (r)
φ0 (r)
log
+
− 1 r−2 φ0 (r)dr
φ(r)
φ(r)
η
Z 1
φ0 (r)
φ0 (r)
φ0 (r)
≤
log
∨ log
+
− 1 r−2 φ0 (r)dr
φ0 (r) − δ
φ0 (r) + δ
φ0 (r) − δ
η
c
c
δ
≤ log
∨ log
+
η −2 < ε.
c−δ
c+δ
c−δ
We now use the machinery of Regazzini et al. [2002] to give an explicit system of equations for
the distribution function of Qn under this choice of Q. Define the family of functions
Z 1
gp (x) :=
rp (c + (1 − c(1 − η))qx,τ (r))dr
η
for p ∈ N and x ∈ [η, 1], as well as the vectors gn (x) := (g1 (x), g2 (x), . . R. , gn (x)) and sn :=
(s1 , . . . , sn ) ∈ Rn . For brevity, for a measure ν and a function f let ν(f ) := f dν whenever the
integral exists.
Let γα be a Gamma random measure with parameter α, that is, a random finite measure on [η, 1]
such that for any measurable partition {A1 , . . . , An } the random variables (γα (A1 ), . . . , γα (An ))
15
are independent and gamma distributed with common scale parameter 1 and respective shape
parameters α(Ak ). Let
hn (sn ; gn ; α) := E [exp (isn · γα (gn ))]
be the characteristic function of γα (gn ) := (γα (g1 ), . . . , γα (gn )). Note that hn (sn ; gn ; α) =
hn (1; sn · gn ; α) and by [Regazzini et al., 2002, Proposition 10]
Z 1
log(1 − isn · gn )dα .
(11)
hn (sn ; gn ; α) = exp −
0
Now let Fn (σ, gn , α) be the joint distribution function of F (gn ) := (F (g1 ), . . . , F (gn )) under
DP (α). The following trick was introduced in [Hannum et al., 1981, equation (2.9)]:
Fn (σ, gn , α) = Fn (0, γα (gn − σ), α)
for any σ ∈ Rn , so that it is sufficient to invert hn at the origin to obtain Fn . This can be done
using the multidimensional version of the Gurland inversion formula [Gurland, 1948, Theorem
3]:
Let C0 , . . . , Cn ∈ Rn+1 solve
Cn = −1
n−r−1
X
k=0
n−r
Cr+k = 1 for r ∈ {0, . . . , n − 1}.
k
(12)
Then
(−1)n+1 2n Fn (σ, gn , α) = C0
n
X
X
Ck
+
(πi)k
Z
∞
Z
...
1≤j1 <...<jk ≤n 0
k=1
0
∞
hk (sk ; gj1 − σj1 , . . . , gjk − σjk ; α)
dsk .
s1 × · · · × sk
The characteristic functions hk and the constants Ck can be computed from (11) and (12) respectively, so that the R.H.S. can be evaluated numerically for practical applications. Numerical
methods are discussed in [Regazzini et al., 2002, Section 6].
Finally, we demonstrate that the restrictive assumptions of Theorem 2 still allow inference for
broad classes of moment sequences with arbitrarily small approximation errors. Let β(r) be any
non-negative probability density on [0, 1], and define the truncation β̄(r) := κ(η, c, K)−1 (β(r) ∨
c ∧ K), where κ is the normalising constant
Z 1
κ(η, c, K) :=
β(r) − (β(r) − K)+ + (c − β(r))+ dr.
η
Note that Q(φ ∈ Dη : kφ − β̄k∞ < δ) > 0 for any δ > 0, and fix such a φ. Now consider the
error on the k th moment:
Z 1
Z 1
rk β(r)dr −
rk φ(r)dr
0
η
≤ β((0, η)) + (1 − η)δ +
(κ(η, c, K) − 1)β([η, 1]) + c(1 − η)
+
κ(η, c, K)
Z
η
1
(β(r) − K)+
dr.
κ(η, c, K)
Each term on the R.H.S. can be made small by choosing η, c and δ sufficiently small, and K
sufficiently large because κ → 1 as η → 0, c → 0 and K → ∞, and
Z 1
Z 1
(β(r) − K)+ dr = 1 −
β(r) ∧ Kdr → β((0, η))
η
η
as K → ∞ by the Monotone Convergence Theorem. A further approximation step also enables
consideration of atoms by choosing β which places all of its mass in neighbourhoods of the
desired locations for atoms. Hence it is possible to ensure the support of Q extends arbitrarily
close to any desired moment sequences despite the restrictive assumptions on Dη in Theorem 2.
16
6
Robust bounds on functionals of Λ
Having established consistency criteria for the posterior and a finite parametrisation via n − 2
leading moments, we now turn to what can be said about Λ based on inferring these moments.
It would be ideal if the diameter of moment classes shrunk with increasing n, as then it would
be possible to fix a representative Λ ∈ M1 ([0, 1]) with specified n − 2 leading moments and
control the remaining within-moment-class error. In Theorem 3 we show that such shrinking
does not happen, and devote the remainder of the section to presenting quantities which can be
controlled based on n − 2 moments alone. We begin by recalling some standard results from the
theory of orthogonal polynomials.
m
Definition 6. Suppose n is odd. Let m := n−3
2 and {φk }k=0 be the first m + 1 Λ-orthogonal
m
polynomials. Let {ξk }k=1 be the zeros of φm .
m
Remark 5. It is a standard result that {φk }m−1
k=0 and {ξk }k=1 are constant within moment
classes of order at least n.
The following bounds on Λ in terms of its leading n − 2 moments are classical:
Lemma 5 (Chebyshev-Markov-Stieltjes (CMS) inequalities). Define
ρm−1 (z) :=
m−1
X
!−1
|φk (z)|2
.
k=0
Then the following inequalities are sharp:
Λ([0, ξj ]) ≤
j
X
ρm−1 (ξk ) ≤ Λ([0, ξj+1 )) for j = {1, . . . , m},
k=1
where ξm+1 := 1.
We are now in a position to prove the following theorem:
Theorem 3. For any n ∈ N and any completely monotonic sequence of moments {λk,k }nk=3
with λ3,3 ≤ 1 there exist uncountably many measures Λ ∈ M1 ([0, 1]), all with leading moments
{λk,k }nk=3 and all satisfying the CMS inequalities, such that for any pair Λx and Λy of them we
have dT V (Λx , Λy ) = 2, where dT V is the total variation distance.
Proof. It will be convenient to write the CMS inequalities in the following, equivalent form:
0 ≤ Λ([0, ξ1 )) ≤ ρm−1 (ξ1 )
0 ≤ Λ([ξj , ξj+1 )) ≤ ρm−1 (ξj ) + ρm−1 (ξj+1 ) for j ∈ {1, . . . , m − 1}
0 ≤ Λ([ξm , 1]) ≤ ρm−1 (ξm )
Pm
where the
last
inequality
follows
from
the
fact
that
k=1 ρm−1 (ξk ) = 1. This equality holds
Pm
because k=1 ρm−1 (ξk ) is the sum of all order m Gauss quadrature weights, or equivalently the
quadrature applied to the constant function 1, which is a polynomial of degree 0. The equality
follows by recalling that Gauss quadrature is exact for polynomials of order up to 2m − 1.
Now let the measures Λx and Λy be described by sequences of (m + 1) weights (x0 , x1 , . . . , xm )
and (y0 , y1 , . . . , ym ), with the j th weight denoting the mass that the corresponding measure
places in the interval [ξj , ξj+1 ) (with obvious adjustments for the right hand boundary terms).
17
For brevity let ζj := ρm−1 (ξj ). Suppose first that m is odd, and let the vectors of weights be
given as
(x0 , x1 , x2 , x3 , x4 , . . . , xm−1 , xm ) = (ζ1 , 0, ζ2 + ζ3 , 0, ζ4 + ζ5 , . . . , 0, ζm )
(y0 , y1 , y2 , y3 , y4 , . . . , ym−1 , ym ) = (0, ζ1 + ζ2 , 0, ζ3 + ζ4 , 0, . . . , ζm−1 + ζm , 0)
P
Both measures have total mass m
j=1 ζj = 1, and the interlacing masses have no overlap so
dT V (Λx , Λy ) = 2. The case where m is even is similar.
Remark 6. The same result holds in Kullback-Leibler divergence due to Pinsker’s inequality:
for probability measures P and Q such that P ∼ Q, and letting KL(P, Q) := EP [log(dP/dQ)]
we have
r
1
dT V (P, Q) ≤
KL(P, Q),
2
so that KL(Λx , Λy ) ≥ 8 for Λx and Λy as in Theorem 3.
Despite this seemingly disappointing result, it is possible to make some conclusions about Λ
based on n − 2 moments. For example, the Kingman hypothesis can be tested in a robust way
by checking whether the vector (0, 0, . . . , 0) lies in a desired credible region of the posterior
Qn (·|n), and the plausibility of any other Λ of interest can be assessed similarly. More generally,
it is possible to maximise/minimise a certain class of functionals subject to moment constraints
obtained from a credible region to obtain robust bounds for quantities of interest. We begin by
recalling some relevant definitions.
m
Definition 7. Let m, n ∈ N, and fix R-valued constants {ck }m
k=1 , a sequence {ik }k=1 of
m
{3, . . . , n}-valued indices and a binary sequence {jk }k=1 of zeros and ones. Let
Cm := Λ ∈ M1 ([0, 1]) : (−1)jk λik ≤ ck for k ∈ {1, . . . , m}
(13)
be a subset of M1 ([0, 1]) with leading n − 2 moments in a desired region specified by m linear
inequalities. Let ext(Cm ) be the extremal points in Cm , i.e. those which cannot be written as
non-trivial convex combinations of elements in Cm , and
(
)
p
p
X
X
wk δxk where 1 ≤ p ≤ m + 1, wk ≥ 0, xk ∈ [0, 1] and
CD := ν ∈ Cm : ν =
wk = 1
k=1
k=1
be the set of discrete probability measures on [0, 1] with at most m + 1 atoms.
Example 3. The extremal points of M1 ([0, 1]) are the Dirac measures:
ext (M1 ([0, 1])) = {δx : x ∈ [0, 1]}.
For our purposes Cm should be thought of as an envelope containing a desired credible region
of truncated, completely monotonic moment sequences expressed using finitely many linear
constraints. We postpone discussion of how an approximate credible region can be obtained to
the next section, and simply assume one is available. The importance of Definition 7 is that
maxima and minima of certain functionals of Λ coincide on Cm and CD , and that CD is finitedimensional so that these extrema can be found numerically. The class of functionals for which
this can be done is given below.
Definition 8. TheR functional F : C →
7 R is measure-affine if, for every ν ∈ C and p ∈ M1 (ext(C))
such that ν(E) = ext(C) γ(E)p(dγ) for every E ∈ B([0, 1]), F is p-integrable and
Z
F (ν) =
F (γ)p(dγ).
ext(C)
18
Intuitively, ν is a barycentre of C with weights on extremal points given by p, and F is measureaffine if it commutes with the operation of expressing ν as the weighted sum of extremal points.
If C consists of finitely many points, this definition coincides with the usual definition of affine
functions.
The following two results originate from [Winkler, 1988, Proposition 3.1 and Theorem 3.2]:
Lemma 6. If q : [0, 1] 7→ R is bounded on at least one side then F : ν 7→ Eν [q] is measure-affine.
Lemma 7. Let Cm be as in Definition 7 and F : Cm 7→ R be measure-affine. Then
inf F (ν) = inf F (ν)
(14)
sup F (ν) = sup F (ν).
(15)
ν∈Cm
ν∈CD
ν∈Cm
ν∈CD
The purpose of Lemma 7 is that the optimisation problems on the R.H.S. of (14) and (15)
are finite-dimensional and can be solved numerically. Hence tight bounds for measure-affine
functionals F (Λ) over credibility regions can be computed in an assumption-free manner.
In order to specify Cm it remains to be able to approximate the posterior, which we achieve via
MCMC. This will be detailed in the next section. Before that, we conclude this section with a
simple example computation.
Example 4. Suppose a posterior credible region is specified via two linear constraints as
C2 = {Λ ∈ M1 ([0, 1]) : λ3 ≤ 0.5 and 0.3 ≤ λ4 },
and that the measure-affine functional of interest is the exponential:
Z 1
F (Λ) :=
e−r Λ(dr).
0
Then the finite dimensional subspace CD ⊂ C2 consists of discrete probability measures on [0, 1]
with at most three atoms:
(
)
p
p
X
X
CD = Λ ∈ C2 : Λ =
wk δrk where 1 ≤ p ≤ 3, wk ≥ 0, rk ∈ [0, 1] and
wk = 1 .
k=1
k=1
This yields three maximisation/minimisation problems, one corresponding to each number of
atoms, though in practice only the largest needs to be solved since the two others can be
recovered as special cases. In this case, the constrained optimisation problem is
Maximise/Minimise: ae−x + be−y + (1 − a − b)e−z
Subject to: ax + by + (1 − a − b)z ≤ 0.5
− ax2 − by 2 − (1 − a − b)z 2 ≤ −0.3
a, b ≤ 1
− a, −b ≤ 0
a+b≤1
x, y, z ≤ 1
− x, −y, −z ≤ 0.
Numerical evaluation in Mathematica yields the bounds F (Λ) ∈ (0.620, 0.810).
19
7
Simulation study
Efficient methods for approximating the Λ-coalescent likelihood pointwise exist [Birkner et al.,
2011, Koskela et al., 2015] and can be readily adapted to the form developed by Beaumont
[2003] for time series data. These likelihood estimators can then be used in the pseudo-marginal
Metropolis-Hastings algorithm [Beaumont, 2003, Andrieu and Roberts, 2009], in which the likelihood evaluations required in a standard Metropolis-Hastings algorithm are replaced with unbiased estimators. The resulting algorithm still targets the correct posterior and inherits the
efficient exploration of parameter space of MCMC methods. Thus it is well-suited to highdimensional situations with intractable likelihood.
Let S(n) denote the space of completely monotonic sequences of length n − 2, and for λ ∈ S(n)
b
let L(λ; n) be the likelihood function and L(λ;
n) be an unbiased estimator. We recall the
pseudo-marginal Metropolis-Hastings algorithm in Algorithm 1 below.
Algorithm 1 Pseudo-marginal Metropolis-Hastings for finite moment sequences
Require: Prior Pn , observation n, transition kernel K : S(n) × S(n) 7→ R+ and N ∈ N.
1: Initialise sample S ← ∅ and moment sequence λ ← λ0 .
b
2: Compute likelihood estimator L(λ;
n).
3: for j = 1, . . . , N do
4:
Sample λ0 ∼ K(λ, ·).
b 0 ; n).
5:
Compute likelihood estimator L(λ
6:
7:
8:
9:
10:
11:
12:
13:
14:
Set a ← 1 ∧
b 0 ;n)Pn (λ0 )
K(λ0 ,λ)L(λ
.
0
b
K(λ,λ )L(λ;n)P
n (λ)
Sample U ∼ U (0, 1).
if U < a then
b
b 0 ) and λ ← λ0 .
Set S ← S ∪ {λ0 }, L(λ)
← L(λ
else
Set S ← S ∪ {λ}.
end if
end for
return S
Algorithm 1 returns a sample of moment sequences S, whose limiting distribution is the posterior.
A credible region C can be approximated from MCMC output, and used to form Cm as per
(13). Measure-affine quantities of interest can then be maximised or minimised using finite
computation by making use of Lemma 7.
By way of demonstration we perform a simulation study on simulated data, focusing on assessing
the Kingman hypothesis, Λ = δ0 , which can be robustly evaluated based upon whether or
not λ3 = 0. The type space consists of 10 binary loci, or 210 types, with mutations flipping
a uniformly chosen locus, i.e. the matrix M has 10 identical, non-zero entries in each row
corresponding to flipping one locus each. The total mutation rate is set at θ = 0.1. Samples
of 20 lineages were simulated at each of five time points from both the Kingman (λ3 = 0) and
Bolthausen-Sznitman (Λ = U (0, 1), λ3 = 0.5) coalescents. The data sets are summarised in
Table 3. Both data sets come from independent simulations, and are sampled from a population
at stationarity. The Kingman coalescent is a classical model of genetic ancestry, while the
Bolthausen-Sznitman coalescent has recently been suggested as a ancestral model for influenza
and HIV [Neher and Hallatschek, 2013].
We set η = 10−6 and specify the prior for the density of Λ on [η, 1] as a Dirichlet process mixture
model of truncated Gaussian kernels. The base measure is the uniform measure on [η, 1], with
total mass scaled to 0.1. Finally, the prior for τ −1/2 , the standard deviations of the truncated
Gaussian kernels was chosen to be the Beta(1.0, 3.0) distribution on [0, 1]. Truncating the
20
Time
0.0
0.5
1.0
Bolthausen-Sznitman
20 x 1001001111
19 x 1001001111
1 x 1101001111
20 x 1001001111
1.5
19 x 1001001111
1 x 1001101111
2.0
19 x 1001001111
1 x 1001001110
20
15
5
8
6
6
10
6
4
16
4
Kingman
x 0010000000
x 0010000000
x 0000000000
x 0010000000
x 0000000000
x 0010001000
x 0010000000
x 0000000000
x 0010001000
x 0010000000
x 0010001000
Table 3: Simulated observed sequences from the two models with θ = 0.1 and M as described
in the paragraph above.
maximal standard deviation at 1 excludes some very flat densities from the support of the prior,
but the standard normal density is already very flat across [η, 1] and the truncation was found to
yield substantial gains in speed of convergence of algorithms. Note that neither data generating
model lies in the support of this prior, but both can be well approximated by members of the
support. The choice of hyperparameters was made because it yields a relatively flat marginal
prior for λ3 , the quantity of interest (c.f. Figure 3), and the prior satisfies the requirements of
the consistency result in Theorem 2.
We make use of the Sethumaran stick-breaking construction of the Dirichlet process [Sethuraman, 1994] and truncate our prior after the first four atoms. For our choice of base measure and concentration parameter this results in a total variation truncation error of order
400e−30 ≈ 3.7 × 10−11 [Ishwaran and James, 2001, Theorem 2]. Any truncation error could be
avoided by pushing forward the prior directly onto the space of moment sequences as illustrated
in Section 5. The cost is a more computationally expensive prior to sample and evaluate, as
well as a higher dimensional parameter space consisting of 98 moments for our data sets. We
do not investigate this strategy further in this paper.
The four atom truncation results in 11 parameters: four locations and standard deviations of
truncated Gaussian kernels, and three stick break points. The fourth break point is set to fulfil
the constraint of the weights summing to 1. We propose updates to these parameters using a
truncated Gaussian random walk on [η, 1]4 × [0, 1]4 × [0, 1]3 with covariance matrix 0.0025 Id.
This scaling was found to result in a reasonable balance of acceptance probability and jump size
for the first moment λ3 .
We approximate the likelihoods required for computing the acceptance probability a using a
straightforward adaptation of the optimised importance sampling method of Koskela et al. [2015]
to the time series setting of [Beaumont, 2003], but need to specify the number of particles to use.
More particles will result in more accurate approximations, but at greater computational cost.
In [Doucet et al., 2015] the authors show that tuning the variance of the log likelihood estimator
to 1.44 results in efficient algorithms under a wide range of assumptions. Preliminary simulations
showed this was achieved in our setting by choosing 75 particles for Bolthausen-Sznitman data,
and 180 particles for Kingman data.
Remark 7. In the context of real data, when the true data generating parameters are not
known, optimising the number of particles using trial runs may require an infeasible amount of
computation. In practice, adaptive algorithms, which optimise parameters online, can be used
to circumvent this problem. Recent work by Sherlock et al. [2015] has shown how to implement
such adaptivity to all four algorithms discussed below, while maintaining ergodicity and the
21
correct stationary distribution of the algorithm.
It is well known that the standard, exact pseudo-marginal algorithm suffers from “sticking”
behaviour, where an unusually high likelihood estimator prevents the algorithm from moving for
a macroscopic number of steps [Andrieu and Roberts, 2009]. The usual solution is to use a noisy
version of the algorithm, in which the likelihood estimator is recomputed at each stage. This
doubles the number of required likelihood evaluations and biases the algorithm into an incorrect
stationary distribution, but can greatly reduce the variance of estimates. We compare both
the exact and noisy versions of the pseudo-marginal algorithm in Figure 2. We also investigate
the effect of delayed acceptance acceleration [Christen and Fox, 2005], in which proposed moves
are first subjected to an accept-reject decision based on an approximate likelihood function
that is cheap to compute. Only samples which are accepted at this first stage are subjected
to an accept-reject decision based on the full likelihood estimates, or more specifically a slight
modification to ensure that the delayed acceptance mechanism does not affect the stationary
distribution of the algorithm. In the Λ-coalescent setting approximate likelihoods are readily
available in the form of Product of Approximate Conditionals (or PAC) methods [Koskela et al.,
2015, Section 4.3], which we use to implement delayed acceptance chains.
Figure 2 shows trace plots of 20 000 steps from the four algorithms introduced above. The
exact pseudo-marginal algorithm exhibits sticking behaviour as might be expected, but it is
surprising to see that the noisy algorithm does not completely eliminate it. We conjecture that
the remaining stickiness in the noisy trace plot is due to multiple, narrow modes in the 11
dimensional posterior. It is also clear that the bias in the noisy algorithm is confounding the
signal in the data, as the traces are much more intermixed than those of the exact algorithm.
Both the noisy and exact pseudo-marginal algorithm are very computationally expensive to run,
particularly for the Kingman data set due to the larger number of particles used to estimate
likelihoods. Delayed acceptance acceleration reduces these run times as expected, particularly
for the Kingman case. Both delayed acceptance algorithms also suffer from sticking, and show
less clear separation of the traces than the exact algorithm. They also look very similar to each
other.
Since it appears to be difficult to eliminate sticking behaviour in this case, we chose to leverage
the speed up obtained by making use of delayed acceptance and ran a further exact, delayed
acceptance pseudo-marginal algorithm for 200 000 steps. A trace plot is shown in Figure 3.
Sticking behaviour is still present, but on a much shorter scale relative to the run length. Run
times are comparable to the noisy algorithm without delayed acceptance, and the BolthausenSznitman trace is again clearly centred at a higher level than the Kingman trace. We thinned
the output of this long run by a factor of 4 000 to reduce the effect of sticking points to obtain
50 samples of first moments, which were used to plot the histograms in Figure 3.
It is clear from both the trace plots and histograms in Figure 3 that the run length is still
not sufficient for fully converged estimates. However, both plots already show a clear shift of
posterior modes toward the values generating the data. The red histogram is consistent with the
Kingman coalescent, while the blue one is consistent with the Bolthausen-Sznitman coalescent.
Moreover, approximate 95% credible intervals are λ3 ∈ [0.1, 0.6] for the Bolthausen-Sznitman
posterior, and λ3 ∈ [η, 0.5] for the Kingman posterior. This suggests the relatively short time
series is nevertheless sufficiently informative to reject the incorrect model in both cases.
8
Discussion
In this paper we have presented a robust framework for Bayesian non-parametric inference under
Λ-coalescent processes for time series data, and studied the feasibility of implementable families
22
0.8
0.6
0.0
0.2
0.4
1st Moment
0.6
0.4
0.0
0.2
1st Moment
0.8
1.0
Noisy
1.0
Exact
0
5000
10000
15000
Kingman (dotted), 5.2 days, acc. pr. = 8%
Bolthausen−Sznitman (solid), 1.9 days, acc. pr. = 9%
20000
0
5000
10000
15000
Kingman (dotted), 9.3 days, acc. pr. = 35%
Bolthausen−Sznitman (solid), 3.1 days, acc. pr. = 49%
0.8
0.6
0.0
0.2
0.4
1st Moment
0.6
0.4
0.0
0.2
1st Moment
0.8
1.0
Noisy Delayed Acceptance
1.0
Exact Delayed Acceptance
20000
0
5000
10000
15000
20000
Kingman (dotted), 1.1 days, acc. pr. = 14%, 35% (overall 5%)
Bolthausen−Sznitman (solid), 0.8 days, acc. pr. = 27%, 24% (overall 6%)
0
5000
10000
15000
20000
Kingman (dotted), 1.5 days, acc. pr. = 11%, 67% (overall 7%)
Bolthausen−Sznitman (solid), 0.9 days, acc. pr. = 13%, 54% (overall 7%)
Figure 2: Trace plot of the pseudo-marginal algorithm (top left), the noisy algorithm (top right)
and corresponding delayed acceptance algorithms (bottom row). Also shown are computation
times (on a mid-range Toshiba laptop with an Intel i5 processor) and acceptance probabilities.
Delayed acceptance runs show acceptance probabilities for both stages, as well as an overall
probability. All runs are independent and initialised from the prior.
2.0
Density
0.0
0.0
0.5
0.2
1.0
1.5
0.6
0.4
1st Moment
2.5
0.8
3.0
3.5
Prior and posterior densities
1.0
Exact Delayed Acceptance
0
50000
100000
150000
200000
Kingman (dotted), 11.7 days, acc. pr. = 16%, 27% (overall 4%)
Bolthausen−Sznitman (solid), 6.2 days, acc. pr. = 17%, 21% (overall 4%)
0.0
0.2
0.4
0.6
0.8
1.0
1st Moment
Figure 3: (Left) Trace plot of the delayed acceptance exact pseudo-marginal algorithm, along
with computation times (on a mid-range Toshiba laptop with an Intel i5 processor) and acceptance probabilities for both stages as well as overall. Both runs are independent and initialised
from the prior. (Right) Histograms of both Kingman (dashed) and Bolthausen-Sznitman (solid)
posteriors estimated from 50 MCMC samples obtained by thinning the runs shown on the left.
The estimated prior density is shown (dotted), based on 10 000 independent samples from the
prior.
23
of algorithms for practical inference. We demonstrated that time series data is necessary for
consistent inference, and gave verifiable conditions for posterior consistenct. As seen in Example
1, lack of consistency can lead to very low statistical power and high sensitivity of inference
both to confounding parameters, such as mutation rate, and the observed allele frequencies.
A theoretical guarantee of consistency is crucial as expressions for statistical power rely on
intractable stationary distributions and transition densities of Λ-Fleming-Viot jump-diffusions,
making the reliability of experiments without time series data very difficult to evaluate.
Efficient methods for importance sampling Λ-coalescent trees are available [Birkner et al., 2011,
Koskela et al., 2015], and these can be used to generalise the pseudo-marginal MCMC algorithms
of [Beaumont, 2003] for temporally spaced data. The consistency conditions of Theorem 2 on
the prior are sufficiently mild to permit the use of Dirichlet process mixture model priors, which
can be readily truncated for implementable algorithms. Alternatively, we have shown that
parametrising the inference problem via truncated moment sequences leads to implementable
algorithms with no discretisation or truncation error. This work provides a strong indication
that time series data, and accompanying inference methods such as the one outlined above,
should be adopted as standard whenever the coalescent generating the data cannot be assumed
to be known.
Generalising of consistency result within the Λ-Fleming-Viot process class to include unknown
drift, which can be used to model e.g. mutation, recombination and selection, as well as more
general Λ-measures is of great interest. However, it is difficult for a number of reasons. Firstly,
relaxing conditions on Λ near 0 while ensuring the integral in (5) remains finite is challenging.
Likewise, it is well known that equivalent changes of measure for Lévy processes necessitate
equivalent Lévy measures (see e.g. [Sato, 1999], Theorem 33.1), and this is also the condition
needed for the jump-diffusions considered in [Cheridito et al., 2005]. The way in which the
drift can be transformed while maintaining absolute continuity in [Cheridito et al., 2005] is
also restrictive, and depends on the diffusion coefficient and Lévy compensator. Finally, any
difference in diffusion coefficients will obviously destroy absolute continuity outright, so if there
were an atom Λ({0}) > 0, its size would have to be known with certainty.
It would also be of great interest to obtain contraction rates of the posterior under verifiable
conditions. Obtaining rates is a challenging problem in non-i.i.d. Bayesian non-parametric inference, and existing results by Gugushvili et al. [2015] for compound Poisson processes and Nickl
and Söhl [2015] for scalar diffusions do not seem generalisable. A different approach by Nguyen
[2013] for mixing measures of infinite mixture models could present a promising directions of
future work by viewing the Λ-coalescent tree as a mixture of merger events, but adaptation into
the present setting is a formidable task and is beyond the scope of this paper.
The method of parametrising the unknown Λ-measure by its first n − 2 moments when the data
set is of size n ∈ N reflects the limited amount of signal in finite data. More precisely, the
likelihood given a sample of size n ∈ N is constant within moment classes of order n, so that
any variation in the posterior within these moment classes is due solely to the prior. Hence
this parametrisation can be seen as regularising an under-determined inference problem in an
infinite dimensional space by identifying an appropriate, data-driven, finite dimensional quotient
space in which to conduct inference. We believe this approach to have more broad applicability
in non-parametric statistics as well as an alternative to direct regularisation by a prior in the
infinite dimensional space, or to approximate projections onto finite dimensional subspaces [Cui
et al., 2014].
The algorithms used to approximate the posterior and maximise/minimise quantities of interest
given the posterior are highly computationally intensive, and we do not expect our approach to
be competitive with well-chosen parametric families when the number of observed lineages or loci
is large. However, the simulations in Section 7 demonstrate that our assumption-free framework
24
can be used to empirically evaluate the modelling fit of parametric families given moderately
sized pilot data, for instance by ensuring that the family contains a candidate Λ which matches
the MAP estimators of some small number of moments. Such parametric families can then be
confidently used to process larger data sets. The pseudo-marginal method can also be adapted
to incorporate unknown mutation parameters, recombination and other forces not considered in
this paper, albeit at the cost of greater computational cost and lower parameter identifiability.
This cost can be alleviated to a large extent by modern GPU and cluster computing approaches,
because the importance sampling algorithm used to estimate likelihoods is readily parallelisable.
For example, up to 500 fold speed up was reported by Lee et al. [2010] when computations were
parallelised on GPUs instead of being run in serial on CPUs. Such gains in computation speed
would make the algorithms employed in Section 7 practical for many realistic genetic data sets.
Acknowledgements
The authors are grateful to Tim Sullivan for insight into orthogonal polynomials as well as
Bayesian brittleness, to Felipe Medina Aguayo for fruitful conversations about pseudo-marginal
methods, to Yun Song for helpful comments on Remark 2, and to Matthias Birkner for assistance
with identifiability conditions. Jere Koskela was supported by EPSRC as part of the MASDOC
DTC at the University of Warwick. Grant No. EP/HO23364/1. Paul Jenkins is supported in
part by EPSRC grant EP/L018497/1.
References
E. C. Anderson. An efficient Monte Carlo method for estimating Ne from temporally spaced
samples using a coalescent-based likelihood. Genetics, 170(2):955–967, 2005.
C. Andrieu and G. O. Roberts. The pseudo-marginal approach for efficient Monte Carlo computations. Ann. Stat., 37(2):697–725, 2009.
E. Árnason. Mitochondrial cytochrome b DNA variation in the high-fecundity Atlantic cod:
trans-Atlantic clines and shallow gene genealogy. Genetics, 166:1871–1885, 2004.
M. A. Beaumont. Estimation of population growth or decline in genetically monitored populations. Genetics, 164:1139–1160, 2003.
J. Berestycki, N. Berestycki, and J. Schweinsberg. The genealogy of branching Brownian motion
with absorption. Ann. Probab., 41(2):527–618, 2013.
J. Bertoin and J.-F. Le Gall. Stochastic flows associated to coalescent processes. Probab. Theory
Related Fields, 126:261–288, 2003.
A. Bhattacharya and D. B. Dunson. Strong consistency of nonparametric Bayes density estimation on compact metric spaces with applications to specific manifolds. Ann. Inst. Stat. Math.,
64:687–714, 2012.
M. Birkner and J. Blath. Computing likelihoods for coalescents with multiple collisions in the
infinitely many sites model. J. Math. Biol., 57(3):435–463, 2008.
M. Birkner and J. Blath. Measure-valued diffusions, general coalescents and population genetic
inference. in J. Blath, P. Mörters, M. Scheutzow (Eds.), Trends in Stochastic Analysis, LMS
351:329–363, 2009.
M. Birkner, J. Blath, and M. Steinrücken. Importance sampling for Lambda–coalescents in the
infinitely many sites model. Theor. Popln Biol., 79(4):155–173, 2011.
25
J. P. Bollback, T. L. York, and R. Nielsen. Estimation of 2Ne s from temporal allele frequency
data. Genetics, 179(1):497–502, 2008.
J. D. G. Boom, E. G. Boulding, and A. T. Beckenback. Mitochondrial DNA variation in introduced populations of Pacific oyster, Crassostrea gigas, in British Columbia. Can. J. Fish.
Aquat. Sci., 51:1608–1614, 1994.
P. Cheridito, D. Filipović, and M. Yor. Equivalent and absolutely continuous measure changes
for jump-diffusion processes. Ann. Appl. Probab., 15(3):1713–1732, 2005.
J. A. Christen and C. Fox. Markov chain Monte Carlo using an approximation. J. Comput.
Graph. Stat., 14(4):795–810, 2005.
T. Cui, J. Martin, Y. M. Marzouk, A. Solonen, and A. Spantini. Likelihood-informed dimension
reduction for nonlinear inverse problems. Inverse Problems, 30(11):114015, 2014.
R. Der and J. B. Plotkin. The equilibrium allele frequency distribution for a population with
reproductive skew. Genetics, 196(4):1199–1216, 2014.
P. Diaconis and D. Freedman. On the consistency of Bayes estimates. Ann. Statist., 14(1):1–26,
1986.
P. Donnelly and T. Kurtz. Particle representations for measure-valued population models. Ann.
Probab., 27(1):166–205, 1999.
A. Doucet, M. Pitt, G. Deligiannidis, and R. Kohn. Efficient implementation of Markov chain
Monte Carlo when using an unbiased likelihood estimator. Biometrika, 102(2):295–313, 2015.
A. J. Drummond, G. K. Nicholls, A. G. Rodrigo, and W. Solomon. Estimating mutation parameters, population history and genealogy simultaneously from temporally spaced sequence
data. Genetics, 161(3):1307–1320, 2002.
A. J. Drummond, A. Rambaut, B. Shapiro, and O. G. Pybus. Bayesian coalescent inference of
past population dynamics from molecular sequences. Mol. Biol. Evol., 22:1185–1192, 2005.
R. Durrett and J. Schweinsberg. A coalescent model for the effect of advantageous mutations
on the genealogy of a population. Stoch. Proc. Appl., 115:1628–1657, 2005.
B. Eldon and J. Wakeley. Coalescent processes when the distribution of offspring number among
individuals is highly skewed. Genetics, 172:2621–2633, 2006.
T. S. Ferguson. A Bayesian analysis of some nonparametric problems. Ann. Stat., 1(2):209–230,
1973.
R. C. Griffiths. The Λ-Fleming-Viot process and a connection with Wright-Fisher diffusion.
Adv. Appl. Probab., 46(4):1009–1035, 2014.
S. Gugushvili, F. van der Meulen, and P. Spreij. Nonparametric Bayesian inference for multidimensional compound Poisson processes. Mod. Stoch. Theory Appl., 2(1):1–15, 2015.
J. Gurland. Inversion formulae for the distribution of ratios. Ann. Math. Statist., 19:228–237,
1948.
R. C. Hannum, M. Hollander, and N. A. Langberg. Distributional results for random functionals
of a Dirichlet process. Ann. Probab., 9:665–670, 1981.
N. L. Hjort, C. Holmes, P. Müller, and S. G. Walker, editors. Bayesian nonparametrics. Cambridge series in statistical and probabilistic mathematics. Cambridge University Press, 2010.
M. Hoffmann and A. Olivier. Nonparametric estimation of the division rate of an age dependent
branching process. Stoch. Proc. Appl., 126(5):1433–1471, 2016.
26
R. Höpfner, M. Hoffmann, and E. Löcherbach. Nonparametric estimation of the death rate in
branching diffusions. Scand. J. Stat., 29:665–690, 2002.
H. Ishwaran and L. F. James. Gibbs sampling methods for stick-breaking priors. J. Amer.
Statist. Assoc., 96(453):161–173, 2001.
J. F. C. Kingman. The coalescent. Stochast. Process. Applic., 13(3):235–248, 1982.
J. Koskela, P. A. Jenkins, and D. Spanò. Computational inference beyond Kingman’s coalescent.
J. Appl. Probab., 52(2):519–537, 2015.
J. Koskela, D. Spanò, and P. A. Jenkins. Consistency of Bayesian nonparametric inference for
discretely observed jump diffusions. Preprint, arXiv:1506.04709, 2017.
A. Lee, C. Yau, M. B. Giles, A. Doucet, and C. C. Holmes. On the utility of graphics cards to
perform massively parallel simulation of advanced Monte Carlo methods. J. Comp. Graph.
Stat., 19(4):769–789, 2010.
A. Y. Lo. On a class of Bayesian nonparametric estimates. 1. density estimates. Ann. Statist.,
12:351–357, 1984.
A.-S. Malaspinas, O. Malaspinas, S. N. Evans, and M. Slatkin. Estimating allele age and selection
coefficient from time-serial data. Genetics, 192(2):599–607, 2012.
I. Mathieson and G. McVean. Estimating selection coefficients in spatially structured populations
from time series data of allele frequencies. Genetics, 193(3):973–984, 2013.
V. N. Minin, E. W. Bloomquist, and M. A. Suchard. Smooth skyride through a rough skyline:
Bayesian coalescent-based inference of population dynamics. Mol. Biol. Evol., 25:1459–1471,
2008.
M. Möhle. The concept of duality and applications to Markov processes arising in neutral
population genetics models. Bernoulli, 5(5):761–777, 1999.
M. Möhle. On sampling distributions for coalescent processes with simultaneous multiple collisions. Bernoulli, 12(1):35–53, 2006.
R. A. Neher and O. Hallatschek. Genealogies of rapidly adapting populations. Proc. Natl Acad.
Sci., 110(2):437–442, 2013.
X. Nguyen. Convergence of latent mixing measures in finite and infinite mixture models. Ann.
Stat., 41(1):370–400, 2013.
R. Nickl and J. Söhl. Nonparametric Bayesian posterior contraction rates for discretely observed
scalar diffusions. Preprint, arXiv:1510.05526, 2015.
H. Owhadi, C. Scovel, and T. Sullivan. Brittleness of Bayesian inference under finite information
in a continuous world. Electron. J. Statist., 9(1):1–79, 2015.
J. Pitman. Coalescents with multiple collisions. Ann. Probab., 27(4):1870–1902, 1999.
E. Regazzini, A. Guglielmi, and G. Di Nunno. Theory and numerical analysis for exact distributions of functionals of a Dirichlet process. Ann. Stat., 30(5):1376–1411, 2002.
S. Sagitov. The general coalescent with asynchronous mergers of ancestral lineages. J. Appl.
Probab., 36(4):1116–1125, 1999.
K.-I. Sato. Lévy processes and infinitely divisible distributions. Cambridge University Press,
1999.
J. Schweinsberg. Rigorous results for a population model with selection II: genealogy of the
population. Preprint, arXiv:1507.00394, 2015.
27
J. Sethuraman. A constructive definition of Dirichlet priors. Stat. Sinica, 4:639–650, 1994.
C. Sherlock, A. Golightly, and D. A. Henderson. Adaptive, delayed-acceptance MCMC for
targets with expensive likelihoods. Preprint, arXiv:1509.00172, 2015.
J. P. Spence, J. A. Kamm, and Y. S. Song. The site frequency spectrum for general coalescents.
Genetics, 202:1549–1561, 2016.
M. Steinrücken, M. Birkner, and J. Blath. Analysis of DNA sequence variation within marine
species using Beta–coalescents. Theor. Popln Biol., 87:15–24, 2013.
F. van der Meulen and H. van Zanten. Consistent nonparametric Bayesian inference for discretely
observed scalar diffusions. Bernoulli, 19(1):44–63, 2013.
J. Wang. Regularity of semigroups generated by Lévy type operators via coupling. Stoch. Proc.
Appl., 120(9):1680–1700, 2010.
G. Winkler. Extreme points of moment sets. Math. Oper. Res., 30(4):581–587, 1988.
28
| 10math.ST
|
MARKOV-NASH EQUILIBRIA IN MEAN-FIELD GAMES WITH
DISCOUNTED COST ∗
arXiv:1612.07878v2 [cs.SY] 14 Jan 2017
NACI SALDI, TAMER BAŞAR, AND MAXIM RAGINSKY
†
Abstract. In this paper, we consider discrete-time dynamic games of the mean-field type with
a finite number N of agents subject to an infinite-horizon discounted-cost optimality criterion. The
state space of each agent is a locally compact Polish space. At each time, the agents are coupled
through the empirical distribution of their states, which affects both the agents’ individual costs
and their state transition probabilities. We introduce a new solution concept of the Markov-Nash
equilibrium, under which a policy is player-by-player optimal in the class of all Markov policies.
Under mild assumptions, we demonstrate the existence of a mean-field equilibrium in the infinitepopulation limit N → ∞, and then show that the policy obtained from the mean-field equilibrium
is approximately Markov-Nash when the number of agents N is sufficiently large.
Key words. Mean-field games, Nash equilibrium, discounted cost.
AMS subject classifications. 91A15, 91A10, 91A13, 93E20.
1. Introduction. Mean-field game theory has been developed independently by
Huang, Malhamé, and Caines [25] and Lasry and Lions [29] to study continuoustime non-cooperative differential games with a large number of identical agents. The
key underlying idea is that, under a particular equilibrium condition, called the Nash
certainty equivalence principle [25], the decentralized game problem can be reduced to
a single-agent decision problem. This principle essentially says that the state evolution
of an individual agent should be consistent with the total population behavior.
The optimal solution of this single-agent decision problem provides an approximation to Nash equilibria of games with large (but finite) population sizes. However,
in contrast to the standard single-agent optimal control problem, the characterization of this optimal solution in the continuous-time setting leads to a Fokker-Planck
equation evolving forward in time, and a Hamilton-Jacobi-Bellman equation evolving
backward in time. We refer the reader to [24, 38, 23, 5, 12, 13, 17, 4, 34, 33] for studies
of continuous-time mean-field games with different models and cost functions, such
as games with major-minor players, risk-sensitive games, robust mean field games,
games with jump parameters, and LQG games.
By contrast, there are relatively few results on discrete-time mean-field games.
Existing works have mostly studied the setup where the state space is discrete (finite
or countable) and the agents are coupled only through their cost functions; that is,
the mean-field term does not affect the evolution of the states of the agents. Gomes
et al. [16] study a discrete-time mean-field game with a finite state space over a
finite horizon. Adlakha et al. [1] consider a discrete-time mean-field game where the
state space is a countable subset of a d-dimensional lattice, under an infinite-horizon
discounted cost criterion. This setup is the closest to the one studied here. However,
in addition to the state space being at most countable, Adlakha et al. [1] also assume
that, for each agent, every stationary policy induces a Markov chain that has at least
one invariant distribution. In this case, the optimal control problem in the mean-field
∗ This research was supported in part by the U.S. Air Force Office of Scientific Research (AFOSR)
under MURI grant FA9550-10-1-0573, and in part by the Office of Naval Research under (ONR)
MURI grant N00014-16-1-2710 and grant N00014-12-1-0998.
† The authors are with the Coordinated Science Laboratory, University of Illinois at UrbanaChampaign, Urbana, IL 61801, Email: {nsaldi,basar1,maxim}@illinois.edu
1
limit corresponds to a standard homogeneous Markov decision problem. Biswas [9]
considers the average-cost setting, where the state space is a σ-compact Polish space
and the transition probability of an individual agent does not depend on the empirical
distribution of the states. Under the regularity conditions imposed in [9], it can be
shown that, for each stationary policy, the corresponding Markov process for each
agent has a unique invariant measure, which is a standard technique when studying
average-cost problems. However, these regularity conditions are stated in terms of a
specific metric topology on the set of stationary policies, and appear to be too strong
to hold under reasonable assumptions. Discrete-time mean-field games with linear
individual dynamics are studied in [15, 31, 35, 32].
In this paper, we consider discrete-time mean-field games with a Polish state
space, under the infinite-horizon discounted-cost optimality criterion. In such a game,
the agents are coupled through the empirical distribution of their states at each time,
which affects both the individual costs and the state transition probabilities of each
agent. In Section 2 we formulate the finite-agent discrete-time game problem of
the mean-field type and introduce the solution concept of Markov-Nash equilibrium,
under which a policy is player-by-player optimal in the class of all Markov policies.
In Section 3, we introduce the infinite-population mean-field game and prove the
existence of a mean-field equilibrium, which we use in Section 4 to deduce the existence
of approximate Markov-Nash equilibria for N -agent games when N is sufficiently
large. In Section 5 we present an example to illustrate our results. Section 6 lists
some directions for future research. To the best of our knowledge, this is the first
result demonstrating the existence of an (exact or approximate) equilibrium policy
for a general discrete-time mean-field game with finitely many agents.
Notation. For a metric space E, we let Cb (E) denote the set of all bounded continuous
real functions on E. Let P(E) denote the set of all Borel probability measures on E.
For any E-valued random element x, L(x)( · ) ∈ P(E) denotes the distribution of
x.
R A sequence {µRn } of measures on E is said to converge weakly to a measure µ if
g(e)µn (de) → E g(e)µ(de) for all g R∈ Cb (E). For any ν ∈ P(E) and measurable
E
real function g on E, we define ν(g) := gdν. For any subset B of E, we let ∂B and
B c denote the boundary and complement of B, respectively. The notation v ∼ ν
means that the random element v has distribution ν. For real numbers a and b, the
notation a ∨ b denotes the maximum of a and b. Unless otherwise specified, the term
“measurable” will refer to Borel measurability.
2. Finite Player Game with Discounted Cost. We consider a discrete-time
N -agent stochastic game with a Polish state space X and a Polish action space A.
The dynamics of the game are specified by an initial state distribution µ0 ∈ P(X) and
a stochastic state transition kernel p : X × A × P(X) → P(X) as follows. For every
N
t ∈ {0, 1, 2, . . .} and every i ∈ {1, 2, . . . , N }, let xN
i (t) ∈ X and ai (t) ∈ A denote the
state and the action of Agent i at time t, and let
(N )
et
( · ) :=
N
1 X
δ N ( · ) ∈ P(X)
N i=1 xi (t)
denote the empirical distribution of the state configuration at time t, where δx ∈ P(X)
is the Dirac measure at x. The initial states xN
i (0) are independent and identically
distributed according to µ0 , and, for each t ≥ 0, the next-state configuration (xN
1 (t +
2
1), . . . , xN
N (t + 1)) is generated at random according to the probability law
N
Y
(N )
N
N
p dxN
i (t + 1) xi (t), ai (t), et
i=1
(2.1)
,
where p( · |x, a, µ) denotes the image of the triple (x, a, µ) ∈ X × A × P(X) in P(X)
under the state transition kernel p.
To complete the description of the game dynamics, we must specify how the
agents select their actions at each time step. To that end, we introduce the history
spaces H0 = X × P(X) and Ht = (X × A × P(X))t × (X × P(X)) for t = 1, 2, . . ., all
endowed with product Borel σ-algebras.1 A policy for a generic agent is a sequence
π = {πt } of stochastic kernels on A given Ht ; we say that such a policy is Markov if
each πt is a Markov kernel on A given X. The set of all policies for Agent i is denoted
by Πi , and the subset consisting of all Markov policies by Mi . Furthermore, we let
Mci denote the set of all Markov policies for Agent i that are weakly continuous; that
is, π = {πt } ∈ Mci if for all t ≥ 0, πt : X → P(A) is continuous when P(A) is endowed
with the weak topology.
QN
QN
QN
Let Π(N ) = i=1 Πi , M(N ) = i=1 Mi , and M(N,c) = i=1 Mci . We let π(N ) :=
(π 1 , . . . , π N ), π i ∈ Πi denote the N -tuple of policies for all the agents in the game.
We will refer to π (N ) ∈ Π(N ) simply as a ‘policy.’ Under such a policy, the action
configuration at each time t ≥ 0 is generated at random according to the probability
law
N
Y
i=1
N
πti daN
i (t) hi (t) ,
(N )
(2.2)
(N )
N
N
N
N
N
where hN
) for t ≥ 1
i (0) = (xi (0), e0 ) and hi (t) = (hi (t − 1), xi (t), ai (t − 1), et
are the histories observed by Agent i at each time step. When π (N ) ∈ M(N ) , Eq. (2.2)
becomes
N
Y
N
πti (daN
i (t)|xi (t)).
i=1
The stochastic update rules in Eqs. (2.1) and (2.2), together with the initial state
distribution µ0 , uniquely determine the probability law of all the states and actions
for all i ∈ {1, . . . , N } and all t ≥ 0. We will denote expectations with respect to this
(N )
probability law by E π
· .
We now turn to the question of optimality. The one-stage cost function for a
generic agent is a measurable function c : X × A × P(X) → [0, ∞). For Agent i,
the infinite-horizon discounted cost under the initial distribution µ0 and a policy
π (N ) ∈ Π(N ) is given by
∞
(N ) X
(N )
(N )
N
Ji (π (N ) ) = E π
β t c(xN
(t),
a
(t),
e
)
,
t
i
i
t=0
where β ∈ (0, 1) is the discount factor. The standard notion of optimality is a playerby-player one:
Definition 2.1. A policy π (N ∗) = (π 1∗ , . . . , π N ∗ ) constitutes a Nash equilibrium
1 We
endow the set P(X) with the topology of weak convergence, which makes it a Polish space.
3
if
(N )
Ji
(N )
(π (N ∗) ) = inf Ji
π i ∈Πi
(N ∗)
(π −i , π i )
(N ∗)
for each i = 1, . . . , N , where π −i := (π j∗ )j6=i .
There are two challenges pertaining to Nash equilibria in the class of games considered here. The first challenge is the (almost) decentralized nature of the information
structure of the problem. Namely, the agents have access only to their local state
(N )
information xN
, and there is no
i (t) and the empirical distribution of the states et
general theory (of existence and characterization of Nash equilibria) for such games
even with special structures for the transition probabilities. The second difficulty
arises because of the so-called curse of dimensionality; that is, the solution of the
problem becomes intractable when the number of states/actions and agents is large.
The existence of Nash equilibria in this case is a challenging problem even when the
N
agents have access to full state information {xN
i (t)}i=1 (see [26, 3, 30] and references
therein). Therefore, it is of interest to find an approximate decentralized equilibrium
with reduced complexity. To that end, we adopt the following solution concept:
Definition 2.2. A policy π(N ∗) ∈ M(N ) is a Markov-Nash equilibrium if
(N )
Ji
(N )
(π (N ∗) ) = inf Ji
π i ∈Mi
(N ∗)
(π −i , π i )
for each i = 1, . . . , N , and an ε-Markov-Nash equilibrium (for a given ε > 0) if
(N )
Ji
(N )
(π (N ∗) ) ≤ inf Ji
π i ∈Mi
(N ∗)
(π −i , π i ) + ε
for each i = 1, . . . , N .
The main contribution of this paper is the proof of existence of ε-Markov-Nash
equilibria for games with sufficiently many agents. To this end, we first consider a
mean-field game that arises in the infinite-population limit N → ∞ and prove the
existence of an appropriately defined mean-field equilibrium for this limiting meanfield game. Then we pass back to the finite-N case and show that, if each agent
in the original problem adopts the mean-field eqiulibrium policy, then the resulting
policy will be an approximate Markov-Nash equilibrium for all sufficiently large N .
It is important to note that, although the policy in the mean-field equilibrium is an
approximate Markov-Nash equilibrium for the finite-agent game problem, it is indeed
a true Nash equilibrium in the infinite population regime. This follows from the fact
that the set of Markov policies is sufficiently rich for optimality in the limiting case,
as each agent is faced with a single-agent decision problem.
2.1. Assumptions. In this section, we state all assumptions that will be made
on the game model for easy reference. They will be imposed when needed in the
remainder of the paper.
Let w : X → [1, ∞) be a continuous moment function; that is, there exists an
increasing sequence of compact subsets {Kn }n≥1 of X such that
lim
inf
n→∞ x∈X\Kn
w(x) = ∞.
We assume that w(x) ≥ 1 + dX (x, x0 )p for some p ≥ 1 and x0 ∈ X, where dX is the
metric on X. In order to study bounded and unbounded one-stage cost functions c
4
simultaneously, we define the following function:
(
1, if c is bounded
v :=
w, if c is unbounded
For any g : X → R, define the v-norm of g as
|g(x)|
.
x∈X v(x)
kgkv := sup
Let Bv (X) denote the Banach space of all real valued measurable functions g on X
with finite v-norm and let Cv (X) denote the Banach space of all real valued continuous
functions in Bv (X).
Analogously, for any finite signed measure µ on X, let us define the v-norm of µ
as
Z
kµkv := sup
g(x)µ(dx) .
g∈Bv (X):
kgkv ≤1
X
Note that if v = 1, then kµkv corresponds to the total variation distance [21, Section
7.2]. Let
Pv (X) := µ ∈ P(X) : kµkv < ∞
Z
v(x)µ(dx) < ∞ .
= µ ∈ P(X) :
X
It is known that weak topology on P(X) can be metrized using the following metric:
ρ(µ, ν) :=
∞
X
2−(m+1) |µ(fm ) − ν(fm )|,
m=1
where {fm }m≥1 is an appropriate sequence of continuous and bounded functions such
that kfm k ≤ 1 for all m ≥ 1 (see [36, Theorem 6.6, p. 47]). We define
ρv (µ, ν) := ρ(µ, ν) + |µ(v) − ν(v)|
which is a metric on Pv (X). Note that
ρv (µn , µ) → 0 ⇐⇒ µn (g) → µ(g), ∀g ∈ Cv (X).
Indeed, (⇐) is trivial. For (⇒), let ρv (µn , µ) → 0 which means that µn → µ weakly
and µn (v) → µ(v). Let g ∈ Cv (X). Define non-negative continuous function h as
h := kgkv v + g. It is known that [20, Proposition E.2]
Z
Z
h(x)µ(dx).
h(x)µn (dx) ≥
lim inf
n→∞
X
X
But, since µn (v) → µ(v), we should have
Z
Z
g(x)µ(dx).
g(x)µn (dx) ≥
lim inf
n→∞
X
X
5
Conversely, define non-negative continuous function u as u := kgkv v − g. Then, we
have
Z
Z
u(x)µ(dx).
u(x)µn (dx) ≥
lim inf
n→∞
X
X
But, since µn (v) → µ(v), we should also have
Z
Z
g(x)µ(dx).
lim sup g(x)µn (dx) ≤
n→∞
X
X
Thus, µn (g) → µ(g), which establishes the result. Let us call the topology induced
by metric ρv on Pv (X) as v-topology. It can be proved that Pv (X) with metric ρv is
a Polish space.
Suppose that v = w (i.e., c is unbounded). Define the Wasserstein distance of
order p ≥ 1 on Pv (X) as follows [39, Definition 6.1]:
1
Wp (µ, ν) := inf E[dX (X, Y )p ] p : L(X) = µ and L(Y ) = ν .
Note that Wp is a metric on Pv (X) since v(x) ≥ 1 + dX (x, x0 )p for all x ∈ X. Furthermore, Wp (µn , µ) → 0 if and only if µn (g) → µ(g) for all continuous g with
|g(x)| ≤ 1 + dX (x, x0 )p [39, Definition 6.8, Theorem 6.9]. The last observation implies
that the following metric βv (µ, ν) := Wp (µ, ν) + |µ(v) − ν(v)| metrizes the v-topology
on Pv (X). Using the dual formulation of Wp [39, Theorem 5.10], we can write βv as
follows:
βv (µ, ν) =
1
|µ(h) − ν(g)| p + |µ(v) − ν(v)|,
sup
(2.3)
(h,g)∈L1 (µ)×L1 (ν):
h(x)−g(y)≤dX (x,y)p
where L1 (λ) denotes the set of all λ-integrable real functions on X.
Remark 1. In the remainder of the paper, P(X) is always equipped with the
weak topology while Pv (X) is always equipped with the v-topology. In other words,
when we say that a function over Pv (X) is continuous, it should be understood that it
is continuous with respect to v-topology. Similarly, a function over P(X) is continuous
if it is continuous with respect to weak topology.
Assumption 1:
(a) The one-stage cost function c is continuous.
(b) A is compact and X is locally compact.
(c) There exists a non-negative real number α such that
Z
sup
w(y)p(dy|x, a, µ) ≤ αw(x).
(a,µ)∈A×P(X)
X
(d) The stochastic kernel p( · |x, a, µ) is weakly continuous; that is, if
(xn , an , µn ) → (x, a, µ) in X × A R× P(X), then p( · |xn , an , µn ) → p( · |x, a, µ)
weakly. In addition, the function X w(y)p(dy|x, a, µ) is continuous in (x, a, µ).
(e) The initial probability measure µ0 satisfies
Z
v(x)µ0 (dx) =: M < ∞.
X
6
Remark
2. Note that Assumption 1-(c) implies that the range of p lies in Pv (X);
that is, p( · |x, a, µ) : (x, a, µ) ∈ X×A×P(X) ⊂ Pv (X). Therefore, Assumption 1-(d)
is equivalent to the following condition: if (xn , an , µn ) → (x, a, µ) in X × A × P(X),
then p( · |xn , an , µn ) → p( · |x, a, µ) with respect to the v-topology on Pv (X).
For each t ≥ 0, let us define
Z
t
t
w(x)µ(dx) ≤ α M .
Pv (X) := µ ∈ Pv (X) :
X
(f) There exist γ ≥ 1 and a non-negative real number R such that for each t ≥ 0,
if we define Mt := γ t R, then
c(x, a, µ) ≤ Mt v(x).
sup
(a,µ)∈A×Pvt (X)
(g) We assume that αβγ < 1.
Assumption 2:
Define the following moduli of continuity:
ωp (r) :=
sup
(x,a)∈X×A
ωc (r) :=
sup
(x,a)∈X×A
sup
kp( · |x, a, µ) − p( · |x, a, ν)kv
sup
|c(x, a, µ) − c(x, a, ν)|,
µ,ν:
ρ̃v (µ,ν)≤r
µ,ν:
ρ̃v (µ,ν)≤r
where ρ̃v = βv (see (2.3)) if c is unbounded, and ρ̃v = ρ if c is bounded.
For any function g : Pv (X) → R, we define the v-norm of g as follows:
kgk∗v :=
|g(µ)|
.
µ∈Pv (X) µ(v)
sup
(h) We assume that ωp (r) → 0 and ωc (r) → 0 as r → 0. Moreover, for any
µ ∈ Pv (X), the functions
ωp (ρ̃v ( · , µ)) : Pv (X) → R
and
ωc (ρ̃v ( · , µ)) : Pv (X) → R
have finite v-norm.
(i) There exists a non-negative real number B such that
Z
sup
v 2 (y)p(dy|x, a, µ) ≤ Bv 2 (x).
(a,µ)∈A×Pv (X)
X
Remark 3. Suppose that v = w. Define a metric λX on X as follows:
(
0,
if x = y
λX (x, y) :=
v(x) + v(y), if x 6= y
7
For any real-valued measurable function g on X, define the Lipschitz seminorm of g
as:
kgkλ := sup
x6=y
|g(x) − g(y)|
.
λX (x, y)
Let Lipλ (1, R) := {g : kgkλ ≤ 1}. Then, for any µ, ν ∈ Pv (X), we have [19, Lemma
2.1]
kµ − νkv =
sup
g∈Lipλ (1,R)
Z
g(x)µ(dx) −
Z
g(x)ν(dx) .
(2.4)
X
X
This alternative formulation of v-norm will be useful when verifying Assumption 2-(h)
for specific examples.
Remark 4. In the remainder of this paper, all the proofs are obtained under the
assumption that the cost function c is unbounded. The bounded case can be covered
by slight modification of the proofs for the unbounded case.
Before proceeding to the next section, we prove an important (but straightforward) result which will be used in the sequel. It basically states that there is no loss of
generality in restricting the infima in Def. 2.2 to weakly continuous Markov policies.
Theorem 2.3. Suppose Assumption 1 holds. Then, for any policy π (N ) ∈ M(N ) ,
we have
(N )
inf Ji
π i ∈Mi
(N )
(N )
(π −i , π i ) = iinf c Ji
π ∈Mi
(N )
(π −i , π i )
for each i = 1, . . . , N .
Proof. The proof is given in Appendix A.
3. Mean-field games and mean-field equilibria. We begin by considering a
mean-field game that can be interpreted as the infinite-population limit N → ∞ of
the game introduced in the preceding section. This mean-field game is specified by
the quintuple X, A, p, c, µ0 , where, as before, X and A denote the state and action
spaces, respectively, p( · |x, a, µ) is the transition probability, and c is the one-stage
cost function. We also define the history spaces as G0 = X and Gt = (X × A)t × X
for t = 1, 2, . . ., which are endowed with their product Borel σ-algebras. A policy is
a sequence π = {πt } of stochastic kernels on A given Gt . The set of all policies is
denoted by Π. A Markov policy is a sequence π = {πt } of stochastic kernels on A
given X. The set of Markov policies is denoted by M.
Remark 5. It is important to note that mean-field games are not games in the
strict sense. As will be shown below, they are single-agent stochastic control problems
with a constraint on the distribution of the state at each time step.
In this section, we impose Assumption 1 on the components X, A, p, c, µ0 of the
mean-field game model.
Instead of N agents in the original game, here we have a single agent and model
the collective behavior of (a large population of) other agents by an exogenous statemeasure flow µ := (µt )t≥0 ⊂ P(X) with a given initial condition µ0 . We say that a
policy π ∗ ∈ Π is optimal for µ if
Jµ (π ∗ ) = inf Jµ (π),
π∈Π
8
where
Jµ (π) := E π
X
∞
t=0
β t c(x(t), a(t), µt )
is the infinite-horizon discounted cost of policy π with the measure flow µ. Here, the
evolution of the states and actions is given by
x(0) ∼ µ0 ,
x(t) ∼ p( · |x(t − 1), a(t − 1), µt−1 ), t = 1, 2, . . .
a(t) ∼ πt ( · |g(t)), t = 0, 1, . . . ,
where g(t) ∈ Gtis the state-action history up to time t.
Let M := µ ∈ P(X)∞ : µ0 is fixed be the set of all state-measure flows with
a given initial condition µ0 . Define the set-valued mapping Φ : M → 2Π as Φ(µ) =
{π ∈ Π : π is optimal for µ}. Conversely, we define a mapping Λ : Π → M as follows:
given π ∈ Π, the state-measure flow µ := Λ(π) is constructed recursively as
Z
p( · |x(t), a(t), µt )Pπ (da(t)|x(t))µt (dx(t)),
µt+1 ( · ) =
X×A
π
where P (da(t)|x(t)) denotes the conditional distribution of a(t) given x(t) under π
and (µτ )0≤τ ≤t . Note that if π is a Markov policy (i.e., πt (da(t)|g(t)) = πt (da(t)|x(t))
for all t), then Pπ (da(t)|x(t)) = πt (da(t)|x(t)).
We are now in a position to introduce the notion of an equilibrium for the meanfield game:
Definition 3.1. A pair (π, µ) ∈ Π × M is a mean-field equilibrium if π ∈ Φ(µ)
and µ = Λ(π).
The following structural result shows that the restriction to Markov policies entails
no loss of optimality:
Proposition 3.2. For any state measure flow µ ∈ M, we have
inf Jµ (π) = inf Jµ (π).
π∈Π
π∈M
Furthermore, we have Λ(Π) = Λ(M); that is, for any π ∈ Π, there exists π̂ ∈ M such
that µπt = µπ̂t for all t ≥ 0.
Proof. The proof is given in Appendix B.
In other words, we can restrict ourselves to Markov policies in the definitions of Φ and
Λ without loss of generality — that is, we have Φ(M) = 2M and Λ(M) = M. This
implies that, unlike the finite-population case, the policy in the mean-field equilibrium,
if it exists, constitutes a true Nash equilibrium of Markovian type for the infinitepopulation game problem. Put differently, this policy is player-by-player optimal in
the class of all admissible (not only Markov) policies.
The main result of this section is the existence of a mean-field equilibrium under
Assumption 1.
Theorem 3.3. Under Assumption 1, the mean-field game (X, A, p, c, µ0 ) admits
a mean-field equilibrium (π, ν) ∈ M × M.
Proof of Theorem 3.3. For each t ≥ 0, let us define
Lt :=
∞
X
(βα)k−t Mk ,
k=t
9
where α and Mk are constants defined in Assumption 1-(c) and (f), respectively. Note
that
Lt = Mt + (βα)Lt+1 .
For each t ≥ 0, we define
Cvt (X) := u ∈ Cv (X) : kukv ≤ Lt
and
Pvt (X × A) := µ ∈ P(X × A) : µ1 ∈ Pvt (X) ,
where for any ν ∈ P(X × A), ν1 denotes the marginal of ν on X; that is,
ν1 ( · ) := ν( · × A).
Moreover, we define
C :=
Ξ :=
∞
Y
t=0
∞
Y
Cvt (X)
Pvt (X × A).
t=0
We equip C with the following metric:
ρ(u, v) :=
∞
X
σ −t kut − vt kv ,
t=0
where σ > 0 is chosen so that σ > γ and σβα < 1. The first condition and Assumption 1-(g) guarantee that ρ(u, v) < ∞ for all u, v ∈ C. It can also be proved that C is
complete with respect to ρ.
For any ν ∈ Ξ and t ≥ 0, we define the operator Ttν as
Z
ν
Tt u(x) = min c(x, a, νt,1 ) + β u(y)p(dy|x, a, νt,1 ) ,
a∈A
X
where u : X → R.
Lemma 3.4. Let ν ∈ Ξ be arbitrary. Then, for all t ≥ 0, Ttν maps Cvt+1 (X) into
t
Cv (X). In addition, for any u, r ∈ Cvt+1 (X), we have
kTtν u − Ttν rkv ≤ αβku − rkv .
(3.1)
Proof. Let u ∈ Cvt+1 (X). By [6, Proposition 7.32], Ttν u is continuous. We also
have
c(x, a, νt,1 ) + β
kTtν ukv ≤
≤
sup
R
X
u(y)p(dy|x, a, νt,1 )
v(x)
(x,a)∈X×A
Mt v(x) + βαLt+1 v(x)
v(x)
(x,a)∈X×A
sup
10
= Mt + βαLt+1 = Lt ,
where the last inequality follows from Assumption 1-(c) and (f). This completes the
proof of the first statement. The proof of the second statement is straightforward, so
we omit the details.
Using operators {Ttν }t≥0 , let us define the operator T ν : C → C as
T ν u t = Ttν ut+1 , for t ≥ 0.
(3.2)
By Lemma 3.4, T ν is a well defined operator; that is, it maps C into itself.
Since Ttν satisfies (3.1) for all t ≥ 0, it can be shown that T ν is a contraction
operator on C with modulus σβα < 1. Hence, T ν has a unique fixed point by the
Banach fixed point theorem, as C is complete with respect to ρ.
ν
For any ν ∈ Ξ, we let J∗,t
: X → R denote the discounted-cost value function
at time t of the nonhomogeneous Markov decision process
with the one-stage cost
functions c(x, a, νt,1 ) t≥0 and the transition probabilities p( · |x, a, νt,1 ) t≥0 . Under
ν
Assumption 1, it can be proved that, for all t ≥ 0, J∗,t
is continuous. Let J∗ν :=
ν
J∗,t t≥0 .
Lemma 3.5. For any ν ∈ Ξ, we have J∗ν ∈ C.
Proof. Let π be arbitrary Markov policy. Note that for all t ≥ 0, we have
ν
(y) ≤
J∗,t
∞
X
k=t
≤
∞
X
k=t
≤
∞
X
β k−t E π c(x(k), a(k), νk,1 ) x(t) = y
β k−t Mk E π v(x(k)) x(t) = y (by Assumption 1-(f))
β k−t Mk αk−t v(y) (by Assumption 1-(c))
k=t
= Lt v(y).
ν
Hence, J∗,t
∈ Cvt (X). This completes the proof.
The following theorem is a known result in the theory of nonhomogeneous Markov
decision processes (see [22, Theorems 14.4 and 17.1]).
Theorem 3.6. For any ν ∈ Ξ, the collection of value functions J∗ν is the unique
fixed point of the operator T ν . Furthermore, π ∈ M is optimal if and only if
Z
π
ν
(y)p(dy|x, a, νt,1 )
νt
(x, a) : c(x, a, νt,1 )+β J∗,t+1
X
ν ν
= Tt J∗,t+1 (x)
= 1,
(3.3)
where νtπ = L x(t), a(t) under π and ν.
To prove the existence of a mean-field equilibrium, we adopt the technique of Jovanovic
∞
and Rosenthal [27]. Define the set-valued mapping Γ : Ξ → 2P(X×A) as follows:
Γ(ν) = C(ν) ∩ B(ν),
where
C(ν) :=
′
ν ′ ∈ P(X × A)∞ : ν0,1
= µ0 and
11
′
νt+1,1
(·) =
Z
X×A
p( · |x, a, νt,1 )νt (dx, da)
and
B(ν) :=
(x, a) : c(x, a, νt,1 )
ν ′ ∈ P(X × A)∞ : ∀t ≥ 0, νt′
Z
ν
ν
+ β J∗,t+1
(y)p(dy|x, a, νt,1 ) = Ttν J∗,t+1
(x)
=1 .
X
The following proposition implies that the image of Ξ under Γ is contained in 2Ξ .
Proposition 3.7. For any ν ∈ Ξ, we have Γ(ν) ⊂ Ξ.
Proof. Fix any ν ∈ Ξ. It is sufficient to prove that C(ν) ⊂ Ξ. Let ν ′ ∈ C(ν). We
′
prove by induction that νt,1
∈ Pvt (X) for all t ≥ 0. The claim trivially holds for t = 0
′
as ν0,1 = µ0 . Assume the claim holds for t and consider t + 1. We have
Z
′
w(y)νt+1,1
(dy) =
X
Z
X×A
≤
Z
w(y)p(dy|x, a, νt,1 )νt (dx, da)
X
αw(x)νt,1 (dx) (by Assumption 1-(c))
X
t+1
≤α
Z
M (as νt,1 ∈ Pvt (X)).
′
Hence, νt+1,1
∈ Pvt+1 (X). This completes the proof.
We say that ν ∈ Ξ is a fixed point of Γ if ν ∈ Γ(ν). The following proposition
makes the connection between mean-field equilibria and the fixed points of Γ.
Proposition 3.8. Suppose that Γ has a fixed point ν = (νt )t≥0 . Construct a
Markov policy π = (πt )t≥0 by disintegrating each νt as νt (dx, da) = νt,1 (dx)πt (da|x),
and let ν 1 = (νt,1 )t≥0 . Then the pair (π, ν 1 ) is a mean-field equilibrium.
Proof. If ν ∈ Γ(ν), then corresponding Markov policy π satisfies (3.3) for ν.
Therefore, by Theorem 3.6, π ∈ Φ(ν 1 ). Moreover, since ν ∈ C(ν), we have Λ(π) = ν1 .
This completes the proof.
By Proposition 3.8, it suffices to prove that Γ has a fixed point in order to establish the
existence of a mean-field equilibrium. To that end, we will use Kakutani’s fixed point
theorem [2, Corollary 17.55]. Note that, since w is a continuous moment function, the
set Pvt (X) is compact with respect to the weak topology [20, Proposition E.8, p. 187],
and so, Pvt (X × A) is tight as A is compact. Furthermore, Pvt (X × A) is closed with
respect to the weak topology. Hence, Pvt (X × A) is compact with respect to the weak
topology. Therefore, Ξ is compact with respect to the product topology. In addition,
Ξ is also convex.
Note that it can be proved in the same way as in [27, Theorem 1] that C(ν) ∩
B(ν) 6= ∅ for any ν ∈ Ξ. Furthermore, we can show that both C(ν) and B(ν) are
convex, and thus their intersection is also convex. Moreover, Ξ is a convex compact
subset of a locally convex topological space M(X × A)∞ , where M(X × A) denotes
the set of finite signed measures on X × A. The final piece we need in order to deduce
the existence of a fixed point of Γ by an appeal to Kakutani’s fixed point theorem is
the following:
Proposition 3.9. The graph of Γ, i.e., the set
Gr(Γ) := {(ν, ξ) ∈ Ξ × Ξ : ξ ∈ Γ(ν)} ,
12
is closed.
Proof. Let (ν (n) , ξ (n) )
n≥1
⊂ Ξ × Ξ be such that ξ (n) ∈ Γ(ν (n) ) for all n and
(ν (n) , ξ (n) ) → (ν, ξ) as n → ∞ for some (ν, ξ) ∈ Ξ × Ξ. To prove Gr(Γ) is closed, it
is sufficient to prove ξ ∈ Γ(ν).
We first prove that ξ ∈ C(ν). For all n and t, we have
Z
(n) (n)
(n)
(3.4)
p( · |x, a, νt,1 )νt (dx, da).
ξt+1,1 ( · ) =
X×A
(n+1)
Since ξ (n) → ξ in Ξ, ξt+1 → ξt+1 weakly. Let g ∈ Cb (X). Then, by [37, Theorem
3.3], we have
Z
Z
Z
Z
(n) (n)
lim
g(y)p(dy|x, a, νt,1 )νt (dx, da) =
g(y)p(dy|x, a, νt,1 )νt (dx, da)
n→∞
X×A
X
X×A
X
R
R
(n)
(n)
since ν t → ν t weakly and X g(y)p(dy|x, a, νt,1 ) converges to X g(y)p(dy|x, a, νt,1 )
continuously2 (see [37, p. 388]). This implies
R that the sequence of measures on the
right-hand side of (3.4) converges weakly to X×A p( · |x, a, νt,1 )νt (dx, da). Therefore,
we have
Z
ξt+1,1 ( · ) =
p( · |x, a, νt,1 )νt (dx, da),
X×A
from which we deduce that ξ ∈ C(ν).
It remains to prove that ξ ∈ B(ν). For each n and t, let us define
Z
(n)
(n)
(n)
ν (n)
(y)p(dy|x, a, νt,1 )
Ft (x, a) = c(x, a, νt,1 ) + β J∗,t+1
X
and
Ft (x, a) = c(x, a, νt,1 ) + β
Z
ν
J∗,t+1
(y)p(dy|x, a, νt,1 ).
X
Recall that, by definition,
(n)
(n)
ν
J∗,t
(x) = min Ft
a∈A
ν
(x, a) and J∗,t
(x) = min Ft (x, a).
a∈A
By assumption, we have
(n)
(n)
ν (n)
1 = ξt
(x, a) : Ft (x, a) = J∗,t
(x) , for all n.
(n)
ν (n)
ν (n)
:= (x, a) : Ft(n) (x, a) = J∗,t
are continu(x) . Since both Ft and J∗,t
(n)
ν
:=
(x, a) : Ft (x, a) = J∗,t (x) which is also closed as
ous, At is closed. Define At
ν
both Ft and J∗,t
are continuous.
(n)
Let At
(n)
Suppose that Ft
tinuously, as n → ∞.
(n)
ν
converges to Ft continuously and J∗,t
ν
converges to J∗,t
con-
2 Suppose g, g (n ≥ 1) are measurable functions on metric space E. The sequence g is said to
n
n
converge to g continuously if limn→∞ gn (en ) = g(e) for any en → e where e ∈ E.
13
ν
(x) + ǫ(M ) which is
For each M ≥ 1, define BtM := (x, a) : Ft (x, a) ≥ J∗,t
ν
closed, where ǫ(M ) → 0 as M → ∞. Since both Ft and J∗,t is continuous, we can
S∞
choose {ǫ(M )}M≥1 so that ξt (∂BtM ) = 0 for each M . Since Act = M=1 BtM and
BtM ⊂ BtM+1 , we have by monotone convergence theorem
(n)
ξt
(n)
Act ∩ At
Hence, we have
(n)
= lim inf ξt
M→∞
(n)
BtM ∩ At ).
(n)
(n)
(n)
(n)
M
+ ξt Bt ∩ At
1 = lim sup lim inf ξt At ∩ At
n→∞ M→∞
(n)
(n)
(n)
(n)
M
+ ξt Bt ∩ At
≤ lim inf lim sup ξt At ∩ At
.
M→∞
n→∞
For fixed M , let us evaluate the limit of the second term in the last expression as
(n)
n → ∞. First, note that ξt converges weakly to ξt as n → ∞ when both measures are
restricted to BtM , as BtM is closed and ξt (∂BtM ) = 0 [10, Theorem 8.2.3]. Furthermore,
1A(n) ∩B M converges continuously to 0: if (x(n) , a(n) ) → (x, a) in BtM , then
t
t
(n)
lim Ft
n→∞
(x(n) , a(n) ) = Ft (x, a)
≥ J∗,t (x) + ǫ(M )
(n)
= lim J∗,t (x(n) ) + ǫ(M ).
n→∞
(n)
Hence, for large enough n’s, we have Ft
(x(n) , a(n) ) 6∈
(n)
At .
(n)
(x(n) , a(n) ) > J∗,t (x(n) ) which implies that
Then, by [37, Theorem 3.3], for each M we have
(n)
(n)
BtM ∩ At
lim sup ξt
n→∞
Therefore, we obtain
(n)
1 ≤ lim sup ξt
n→∞
= 0.
(n)
At ∩ At
(n)
≤ lim sup ξt (At )
n→∞
≤ ξt (At ),
where the last inequality follows from Portmanteau theorem [8, Theorem 2.1] and the
fact that At is closed. Hence, ξt (At ) = 1. Since t is arbitrary, this is true for all t.
This means that ξ ∈ B(ν). Therefore, ξ ∈ Γ(ν) which completes the proof under
(n)
ν
ν (n)
the assumption that Ft converges to Ft continuously and J∗,t
converges to J∗,t
continuously, as n → ∞, which we prove next.
Note that, for continuous functions, continuous convergence coincides with the
uniform convergence over compact sets (see [28, Lemma 2.1]). Therefore, it is equiva(n)
ν (n)
lent to establish that Ft uniformly converges to Ft over compact sets and J∗,t
uniν
formly converges to J∗,t
over compact sets, as these functions are all continuous. Fur(n)
ν
thermore, if J∗,t
(n)
ν
converges to J∗,t
continuously for all t, then Ft
(n)
Ft continuously for all t. Indeed, let (x
,a
14
(n)
) → (x, a). Since
also converges to
ν (n)
J∗,t+1
(y)
≤ Lt+1 v(y)
for all n ≥ 1 and
rem 3.3] we have
R
X
(n)
v(y)p(dy|x(n) , a(n) , νt,1 ) →
(n)
lim F (x(n) , a(n) )
n→∞ t
(n)
= lim c(x
n→∞
,a
(n)
(n)
, νt,1 )
R
X
+β
v(y)p(dy|x, a, νt,1 ), by [37, TheoZ
X
= c(x, a, νt,1 ) + β
Z
(n)
ν (n)
J∗,t+1
(y)p(dy|x(n) , a(n) , νt,1 )
ν
J∗,t+1
(y)p(dy|x, a, νt,1 )
X
= Ft (x, a).
(n)
ν
ν
Therefore, it is sufficient to prove that J∗,t
uniformly converges to J∗,t
over compact
sets, for all t.
Proposition 3.10. For any compact K ⊂ X, we have
(n)
lim sup J∗,t (x) − J∗,t (x)| = 0
n→∞ x∈K
(n)
for all t ≥ 0. Therefore, J∗,t converges to J∗,t continuously as n → ∞, for all t.
Proof. The proof of the proposition is given in Appendix C.
Theorem 3.3 is now a consequence of the following:
Theorem 3.11. Under Assumption 1, there exists a fixed point ν of the set
valued mapping Γ : Ξ → 2Ξ . Therefore, the pair (π, ν 1 ) is a mean field equilibrium,
where π and ν 1 are constructed as in the statement of Proposition 3.8.
Proof. Recall that Ξ is a compact convex subset of the locally convex topological
space M(X × A)∞ . Furthermore, Γ has closed graph by Proposition 3.9, and it takes
nonempty convex values. Therefore, by Kakutani’s fixed point theorem [2, Corollary
17.55], Γ has a fixed point. The second statement follows from Proposition 3.8.
4. Existence of Approximate Markov-Nash Equilibria. Now we are in a
position to prove the main result of the paper — namely, the existence of approximate
Markov-Nash equilibria in games with sufficiently many agents. Let (π, µ) denote the
mean-field equilibrium, which exists by Theorem 3.3. In a nutshell, the proof boils
down to showing that, if each of the N agents adopts the mean-field equilibrium policy
π, then the resulting policy π (N ) = {π, π, . . . , π} is an ε-Markov-Nash equilibrium for
all sufficiently large N .
In addition to Assumption 1, we impose Assumption 2 in this section. Furthermore, we assume that
(j) For each t ≥ 0, πt : X → P(A) is weakly continuous.
The following theorem is the main result of the paper:
Theorem 4.1. Suppose that Assumptions 1 and 2, and (j) hold. Then, for any
ε > 0, there exists a positive integer N (ε), such that, for each N ≥ N (ε), the policy
π (N ) = {π, π, . . . , π} is an ε-Markov-Nash equilibrium for the game with N agents.
The remainder of the section is devoted to the proof of Theorem 4.1. In a nutshell,
the logic of the proof can be described as follows: We first show that, as N → ∞, the
empirical distribution of the agents’ states at each time t converges to a deterministic
limit given by the mean-field equilibrium distribution of the state at time t. This
allows us to deduce that the evolution of the state of a generic agent closely tracks
the equilibrium state-measure flow in the infinite-population limit. We then show
that the infinite-population limit is insensitive to individual-agent deviations from
the mean-field equilibrium policy.
15
We start by defining a sequence of stochastic kernels Ptπ ( · |x, µ)
X × P(X) as
Z
p( · |x, a, µ)πt (da|x).
Ptπ ( · |x, µ) :=
t≥0
on X given
A
Since πt is assumed to be weakly continuous, Ptπ ( · |x, µ) is also weakly continuous in
π
(x, µ). In the sequel, to ease the notation, we will also write Ptπ ( · |x, µ) as Pt,µ
( · |x).
Recall that measure flow µ in the mean-field equilibrium satisfies
Z
π
( · ).
µt+1 ( · ) =
Ptπ ( · |x, µt )µt (dx) = µt Pt,µ
t
X
For each N ≥ 1, let xN
i (t)
1≤i≤N
(N )
denote the state configuration at time t in the N (N )
person game under the policy π , and let et denote the corresponding empirical
distribution.
′
Lemma 4.2. Fix an arbitrary N ≥ 1. Let π (N ) be arbitrary policy for the
′
N -agent game problem. Then, under π (N ) , for all t ≥ 0 and i = 1, . . . , N , we have
t
L(xN
i (t)) ∈ Pv (X).
Proof. The proof of the lemma is similar to the proof of Proposition 3.7.
Recall the Polish space (Pv (X), ρv ). Define the Wasserstein distance of order 1 on
the set of probability measures P(Pv (X)) over Pv (X) as follows [39, Definition 6.1]:
W1 (Φ, Ψ) := inf E[ρv (X, Y )] : L(X) = Φ and L(Y ) = Ψ .
Moreover, define the following spaces:
Z
P1 (Pv (X)) := Φ ∈ P(Pv (X)) :
Pv (X)
ρv (µ, µ0 )Φ(dµ) < ∞
and
Cv (Pv (X)) := Υ : Pv (X) → R; Υ is continuous and kΥk∗v < ∞ .
(N )
(N )
Lemma 4.2 implies that L(et )(Pv (X)) = 1 and L(et )( · ) ∈ P1 (Pv (X)), for all
N ≥ 1 and t ≥ 0.
Lemma 4.3. Let {Φn }n≥1 ⊂ P1 (Pv (X)) and δµ ∈ P1 (Pv (X)). Then the following
are equivalent.
(i) W1 (Φn , δµ ) → 0 as n → ∞.
(ii) E |F (Xn ) − F (X)| → 0 as n → ∞, for any F ∈ Cv (Pv (X)) and for any
sequence of Pv (X)-valued random elements {Xn }n≥1 and a Pv (X)-valued random
element X such
that L(Xn ) = Φn and L(X) = δµ .
(iii) E |Xn (f ) − X(f )| → 0 as n → ∞, for any f ∈ Cb (X) ∪ {v} and for any
sequence of Pv (X)-valued random elements {Xn }n≥1 and a Pv (X)-valued random element X such that L(Xn ) = Φn and L(X) = δµ .
Proof. (i) ⇒ (iii)
Fix any {Xn }n≥1 and X that satisfy the hypothesis of (iii). We first prove the result
16
for f ∈ Cb (X). Define ∆n ( · ) := L(Xn , X)( · ). Since L(X) = δµ , the only coupling
between Φn and δµ is given by Φn ⊗ δµ . Therefore, ∆n = Φn ⊗ δµ . Then, we have
Z
lim E |Xn (f ) − X(f )| = lim
|ν(f ) − ζ(f )|∆n (dν, dζ)
N →∞
N →∞ P (X)2
Z v
= lim
|ν(f ) − µ(f )|Φn (dν)
N →∞ P (X)
v
Z
=
|ν(f ) − µ(f )|δµ (dν)
(4.1)
Pv (X)
= 0,
where (4.1) follows from the fact that h(ν) := |ν(f ) − µ(f )| ∈ Cb (Pv (X)) and Φn
converges to δµ weakly. This establishes the result for f ∈ Cb (X). For f = v, the
result follows from the definition of ρv and the fact that W1 (Φn , Φ) = E ρv (Xn , X)
since there is only one coupling between Φn and Φ.
Note that (iii), (ii) ⇒ (i) is clear by definition of ρv .
(i) ⇒ (ii)
Let F ∈ Cv (Pv (X)). Since the coupling between Φn and δµ is unique, we can write
Z
|F (ν) − F (µ)|Φn (dν).
E |F (Xn ) − F (X)| =
Pv (X)
Define l(ν) := |F (ν) − F (µ)|. Since Φn → δµ weakly, we have [20, Proposition E.2]
Z
l(ν) + klk∗v ν(v)Φn (dν) ≥ l(µ) + klk∗v µ(v)
lim inf
n→∞
Pv (X)
and
lim inf
n→∞
Z
Pv (X)
−l(ν) + klk∗v ν(v)Φn (dν) ≥ −l(µ) + klk∗v µ(v)
as both l(ν) + klk∗v ν(v) and −l(ν) + klk∗v ν(v) are nonnegative continuous functions on
Pv (X). Note that
Z
lim
ν(v)Φn (dν) − µ(v) = 0
n→∞
Pv (X)
by definition of ρv and (i). Thus, we have
Z
l(ν)Φn (dν) ≥ l(µ)
lim inf
n→∞
Pv (X)
≥ lim sup
n→∞
Z
l(ν)Φn (dν).
Pv (X)
Therefore,
lim
n→∞
Z
l(ν)Φn (dν) = l(µ) = |F (µ) − F (µ)| = 0.
Pv (X)
This completes the proof.
17
The following proposition states that, at each time t, the sequence of random
(N )
measures et converges to the mean-field equilibrium distribution µt of the state at
time t as N → ∞:
(N )
Proposition 4.4. For all t ≥ 0, limN →∞ W1 L(et ), δµt = 0 in P1 Pv (X) .
Proof. We prove that
(N )
lim E |et (f ) − µt (f )| = 0
N →∞
for any f ∈ Cv (X) by induction on t. Since Cv (X) ⊃ Cb (X) ∪ {v}, this will complete
the proof by Lemma 4.3.
QN
Note that since {xN
i (0)}1≤i≤N ∼
i=1 µ0 , the claim is true for t = 0 as any
f ∈ Cv (X) is µ0 -integrable by Assumption 1-(e). Suppose the claim holds for t and
consider t + 1. Fix any g ∈ Cv (X). Then, we have
(N )
(N )
(N )
|et+1 (g) − µt+1 (g)| ≤ |et+1 (g) − et
(N )
π
Pt,e
(N ) (g)| + |et
t
π
π
Pt,e
(N ) (g) − µt Pt,µt (g)|, (4.2)
t
π
µt Pt,µ
t
since (µt )t≥0 is the measure flow in the mean field equilibrium.
where µt+1 =
First, let us consider the second term in (4.2). Define F : Pv (X) → R as
Z Z
π
(g) :=
g(y)Ptπ (dy|x, µ)µ(dx).
F (µ) = µPt,µ
X
X
Note that
kF k∗v =
=
|F (µ)|
µ∈Pv (X) µ(v)
R R
sup
X X
sup
µ∈Pv (X)
=
µ∈Pv (X)
=
|g(y)|Ptπ (dy|x, µ)µ(dx)
µ(v)
R R
kgkv v(y)Ptπ (dy|x, µ)µ(dx)
µ(v)
X X
sup
µ∈Pv (X)
= kgkv
R R
X X
sup
g(y)Ptπ (dy|x, µ)µ(dx)
µ(v)
αµ(v)
= αkgkv .
µ∈Pv (X) µ(v)
sup
Hence, kF k∗v < ∞. If we can prove that F is also continuous, then by Lemma 4.3
the expectation of the second term in (4.2) goes to zero. To this end, let µn → µ in
Pv (X) with respect to v-topology. Define
Z
g(y)Ptπ (dy|x, µn )
ln (x) :=
X
and
l(x) :=
Z
g(y)Ptπ (dy|x, µ).
X
R
R
Note that F (µn ) = X ln (x)µn (dx) and F (µ) = X l(x)µ(dx). We first prove that ln
converges continuously to l. Let xn → x. Since πt is assumed to be continuous, we
have πt ( · |xn ) → πt ( · |x) weakly. Note that
Z
sup
g(y)p(dy|xn , a, µn ) ≤ kgkv α sup v(xn ) < ∞
n≥1
n≥1
X
18
since v is continuous and {xn } is convergent. Therefore, by [37, Theorem 3.3] we have
Z Z
g(y)p(dy|xn , a, µn )πt (da|xn )
lim ln (xn ) = lim
n→∞
n→∞ A X
Z Z
=
g(y)p(dy|x, a, µ)πt (da|x)
A
X
= l(x).
(4.3)
Thus, ln converges to l continuously. In addition, we have |ln (x)| ≤ αkgkv v(x) for all
x ∈ X and µn (v) → µ(v). Then, again by [37, Theorem 3.3], we have
Z
ln (x)µn (dx)
lim F (µn ) = lim
n→∞
n→∞ X
Z
l(x)µ(dx)
=
X
= F (µ).
Hence, F ∈ Cv (Pv (X)) and so the expectation of the second term in (4.2) goes to
zero.
Now, consider the first term in (4.2). Let us write the expectation of this term as
(N )
(N ) π
N
N
E E |et+1 (g) − et Pt,e(N ) (g)| x1 (t), . . . , xN (t) .
t
Then, by Lemma 6.2 in Appendix D, we have
2
(N )
(N ) π
N
N
E E |et+1 (g) − et Pt,e(N ) (g)| x1 (t), . . . , xN (t)
t
2
(N )
(N ) π
N
N
(g)|
x
(t),
.
.
.
,
x
(t)
≤ E E |et+1 (g) − et Pt,e
(N )
1
N
t
≤E
1
N2
N Z
X
i=1
π
N
g 2 (y)Pt,e
(N ) (dy|xi (t))
t
X
+
Z
X
≤E
kgk2v
N2
N Z
X
i=1
X
+
π
N
v 2 (y)Pt,e
(N ) (dy|xi (t))
t
Z
X
≤
kgk2v
N2
2
π
N
g(y)Pt,e
(N ) (dy|xi (t))
t
2
N
π
v(y)Pt,e
(N ) (dy|xi (t))
t
N
X
2 2 N
E Bv 2 (xN
i (t)) + α v (xi (t)) .
i=1
The last expression follows from Assumption 1-(c) and Assumption 2-(i). For any
t ≥ 0, one can prove that supN ≥1 supi=1,...,N E[v 2 (xN
i (t))] ≤ H for some H ∈ R by
Assumption 2-(i). Therefore, the expectation of the first term in (4.2) also converges
to zero as N → ∞. Since g is arbitrary, this completes the proof.
From the above proposition and from the assumed continuity of the transition
probability p( · |x, a, µ) in µ, we now deduce that the evolution of the state of a generic
19
agent in the original game with sufficiently many agents should closely track the
evolution of the state in the mean-field game:
Proposition 4.5. If (π, µ) is a mean-field equilibrium, then
(N )
lim J1
N →∞
(π (N ) ) = Jµ (π) = inf
Jµ (π ′ ).
′
π ∈Π
Proof. For each t ≥ 0, let us define
Z
c(x, a, µ)πt (da|x).
cπt (x, µ) :=
A
N
Note that, for any permutation σ which is independent of xN
1 (t), . . . , xN (t), we have
(N )
N
L xN
1 (t), . . . , xN (t), et
Therefore, we can write
(N )
N
= L xN
σ(1) (t), . . . , xσ(N ) (t), et
.
N
1 X N
(N )
(N )
N
E c(xi (t), aN
)
)
=
E c(xN
(t),
a
(t),
e
i (t), et
1
t
1
N i=1
(N )
(N )
cπt (x, et ) .
= E et
Define F : Pv (X) → R as
F (µ) :=
Z
cπt (x, µ)µ(dx).
X
(N )
(N )
N
Hence, E c(xN
) = E F (et ) . First, note that
1 (t), a1 (t), et
kF k∗v
=
sup
µ∈Pv (X)
≤
sup
µ∈Pv (X)
≤
R
X×A
c(x, a, µ)πt (da|x)µ(dx)
µ(v)
R
R
− X cπt (x, µt )µ(dx)
c (x, µt )µ(dx)
X πt
+ sup
µ(v)
µ(v)
µ∈Pv (X)
R
c (x, µ)µ(dx)
X πt
Mt µ(v)
ωc (ρ̃v (µ, µt ))
+ sup
µ(v)
µ∈Pv (X) µ(v)
µ∈Pv (X)
sup
= kω(ρ̃v ( · , µt ))k∗v + Mt < ∞. (by Assumption 2-(h))
Hence, F has finite v-norm. For continuity, let µn → µ in v-topology. Then, we have
|F (µn ) − F (µ)| ≤
Z
Z
cπt (x, µn )µn (dx) − cπt (x, µ)µn (dx)
X
Z
Z
+
cπt (x, µ)µn (dx) − cπt (x, µ)µ(dx)
X
X
Z
Z
≤ ωc (ρ̃v (µn , µ)) +
cπt (x, µ)µn (dx) − cπt (x, µ)µ(dx) .
X
X
20
X
Note that cπt (x, µ) ∈ Cv (X) since πt is weakly continuous and µ ∈ Pvl (X) for some
l ≥ 1. Therefore, the last expression goes to zero as µn → µ in v-topology. Hence,
F ∈ Cv (Pv (X)). Therefore, by Proposition 4.4 we have
(N )
(N )
(N )
N
lim E c(xN
) = lim E et
cπt (x, et )
1 (t), a1 (t), et
N →∞
N →∞
(N )
= lim E[F (et
N →∞
)]
= F (µt )
= µt (cπt (x, µt )).
(4.4)
Since t is arbitrary, this is true for all t ≥ 0. Recall that the pair (π, µ) is a mean
field equilibrium, and so we can write
Jµ (π) =
∞
X
β t µt (cπt (x, µt )).
t=0
Note that for all N ≥ 1
(N )
J1
(π (N ) ) =
∞
X
(N )
N
)
β t E c(xN
1 (t), a1 (t), et
t=0
∞
X
≤
t=0
≤
∞
X
t=0
≤
∞
X
t=0
(N )
N
N
) − c(xN
β t E |c(xN
1 (t), a1 (t), µ0 )|
1 (t), a1 (t), et
N
+ E c(xN
1 (t), a1 (t), µ0 )
(N )
(t))
β t E ωc (ρ̃v (et , µ0 )) + E M0 v(xN
1
β
t
kωc (ρ̃v ( · , µ0 ))k∗v E
(N )
et (v)
∞
X
≤ kωc (ρ̃v ( · , µ0 ))k∗v + M0 M
(βα)t
+ M0 E v(xN
1 (t))
t=0
< ∞.
Therefore, by (4.4) and the dominated convergence theorem, we obtain
(N )
lim J1
N →∞
(π (N ) ) = Jµ (π)
which completes the proof.
Now let {π̃ (N ) }N ≥1 ⊂ Mc1 be an arbitrary
sequence of weakly continuous Markov
policies for Agent 1. For each N ≥ 1, let x̃N
i (t) 1≤i≤N be the state configuration
at time t in the N -person game under the policy π̃ (N ) := {π̃ (N ) , π, . . . , π} (i.e., when
Agents 2 through N stick to the mean-field equilibrium policy π, while Agent 1
(N )
deviates with π̃ (N ) ), and let ẽt denote the corresponding empirical distribution. The
following result, whose proof is a slight modification of the proof of Proposition 4.4,
states that, in the infinite-population limit, the law of the empirical distribution of the
states at each time t is insensitive to local deviations from the mean-field equilibrium
policy:
(N )
Proposition 4.6. For each t ≥ 0, limN →∞ W1 L(ẽt ), δµt = 0 in P1 (Pv (X)).
21
Proof. We follow the same technique that was employed in the proof of Proposition 4.4. Namely, we prove that
(N )
lim E |ẽt (f ) − µt (f )| = 0
N →∞
for any f ∈ Cv (X) by induction on t.
QN
Since x̃N
i (0) 1≤i≤N ∼
i=1 µ0 , the claim holds for t = 0. Suppose the claim
holds for t and consider t + 1. Choose any g ∈ Cv (X) and write
(N )
(N )
(N −1)
|ẽt+1 (g) − µt+1 (g)| ≤ |ẽt+1 (g) − ẽt+1
+
+
(N −1) π
|ẽt
Pt,ẽ(N ) (g)
−
(N ) π
|ẽt Pt,ẽ
(N ) (g)
π
(g)|,
µt Pt,µ
t
−
(N −1) π
ẽt
Pt,ẽ(N ) (g)|
t
t
+
t
(N −1)
where ẽt+1
(N −1)
and ẽt
(g)|
(N −1)
|ẽt+1 (g)
−
(N ) π
ẽt Pt,ẽ
(N ) (g)|
t
(4.5)
are given by
N
(N −1)
ẽt+1
:=
(N −1)
:=
1 X
δ (N )
N − 1 i=2 x̃i (t+1)
and
N
ẽt
1 X
δ (N ) .
N − 1 i=2 x̃i (t)
In the remainder of the proof, we show that expectation of each term in (4.5) converges
to zero as N → ∞, which will complete the proof as g is arbitrary.
First, let us consider the first term in (4.5). We have
(N )
(N −1)
|ẽt+1 (g) − ẽt+1
(g)| ≤
N
N
1 X
1 X
g(x̃N
(t
+
1))
−
g(x̃N
i
i (t + 1))
N i=1
N i=2
∨
N
N
1 X
1 X
g(x̃N
(t
+
1))
−
g(x̃N
i
i (t + 1))
N i=1
N − 1 i=1
When we take the expectation of above term, we obtain
(N )
(N −1)
E |ẽt+1 (g) − ẽt+1 (g)|
≤
≤
N
X
1
1
1
E |g(x̃N
E |g(x̃N
(t
+
1))|
+
−
i (t + 1))|
1
N
N
N − 1 i=1
kgkv t+1
N
kgkv αt+1 M.
α M + 1−
N
N −1
Note that the last expression goes to zero as N → ∞.
We can write the expectation of the second term in (4.5) as
(N −1)
(N −1) π
N
N
E E |ẽt+1 (g) − ẽt
Pt,ẽ(N ) (g)| x̃1 (t), . . . , x̃N (t) .
t
22
Using Lemma 6.2 in Appendix D, we can prove that this term converges to zero as in
the proof of Proposition 4.4.
For the expectation of the third term in (4.5), we have
(N −1) π
(N ) π
E |ẽt
Pt,ẽ(N ) (g) − ẽt Pt,ẽ
(N ) (g)|
t
t
Z Z
(N −1)
π
=E
(dx)
g(y)Pt,ẽ
(N ) (dy|x)ẽt
t
X X
Z Z
(N )
π
−
g(y)Pt,ẽ(N ) (dy|x)ẽt (dx)
t
X X
Z
1
π
N
g(y)Pt,ẽ(N ) (dy|x̃1 (t))
≤ E
N
t
X
Z
N
X
1
1
π
N
E
−
g(y)Pt,ẽ(N ) (dy|x̃i (t))
+
N
N − 1 i=1
t
X
≤
≤
N
X
N
kgkv α N
E v(x̃N
E v(x̃1 (t)) + 1 −
kgkv α
i (t))
N
N −1
i=1
kgkv t+1
N
α M + 1−
kgkv αt+1 M.
N
N −1
Again, the last expression goes to zero as N → ∞.
π
For the last term in (4.5), let us define F : Pv (X) → R as F (µ) := µPt,µ
(g).
Since πt is weakly continuous, it can be proven as in the proof of Proposition 4.4 that
F ∈ CV (Pv (X)). Then, by Lemma 4.3, the expectation of the last term converges
(N )
to zero since L(ẽt ) → δµt with respect to W1 distance in P1 Pv (X) by induction
hypothesis. This completes the proof.
For each N ≥ 1, let {x̂N (t)}t≥0 denote the states of the non-homogenous Markov
chain that evolves as follows:
(N )
π̃
( · |x̂N (t)).
x̂N (0) ∼ µ0 and x̂N (t + 1) ∼ Pt,µ
t
Note that
Jµ (π̃ (N ) ) =
∞
X
t=0
β t E c(x̂N (t), âN (t), µt ) .
Lemma 4.7. Let F be a family of equicontinuous functions on Pv (X) with respect
to v-topology. For any t ≥ 0, define Ft : Pv (X) → [0, +∞] as:
Ft (µ) := sup |Ψ(µ) − Ψ(µt )|.
Ψ∈F
Suppose that Ft is real valued and kFt k∗v < ∞. Then Ft ∈ Cv (Pv (X)). Therefore,
(N )
lim E sup Ψ(ẽt ) − Ψ(µt ) = 0.
N →∞
Ψ∈F
Proof. It is sufficient to prove that Ft is continuous with respect to v-topology.
Let µn → µ in v-topology. Then we have
|Ft (µn ) − Ft (µ)| = sup |Ψ(µn ) − Ψ(µt )| − sup |Ψ(µ) − Ψ(µt )|
Ψ∈F
Ψ∈F
23
≤ sup |Ψ(µn ) − Ψ(µ)| → 0
Ψ∈F
as N → ∞ since F is equicontinuous. This completes the proof.
Using Lemma 4.7 we now prove the following result.
Lemma 4.8. Fix any t ≥ 0. Suppose that
N
lim L(x̃N
1 (t))(gN ) − L(x̂1 (t))(gN ) = 0
N →∞
for any sequence {gN }N ≥1 ⊂ Cv (X) with supN ≥1 kgN kv < ∞. Then, we have
(N )
lim L(x̃N
1 (t), ẽt
N →∞
)(TN ) − L(x̂N
1 (t), δµt )(TN ) = 0
for any sequence {TN} of functions on X × Pv (X) satisfying:
(i) The family TN (x, · ) : x ∈ X, N ≥ 1) is equicontinuous with respect to
v-topology.
(ii) TN ( · , µ) ∈ Cv (X) for all µ ∈ Pv (X) and N ≥ 1.
(iii) supN ≥1 kTN ( · , µ)kv < ∞ for all µ ∈ Pv (X).
(iv) The function Ft (µ) := sup x∈X |TN (x, µ) − TN (x, µt )| on Pv (X) is real valued
N ≥1
and kFt k∗v < ∞.
Proof. Fix any sequence {TN }N ≥1 satisfying the hypothesis of the lemma. Then,
we have
(N )
L(x̃N
1 (t), ẽt
Z
≤
)(TN ) − L(x̂N
1 (t), δµt )(TN )
(N )
X×Pv (X)
TN (x, µ)L(x̃N
1 (t), ẽt
−
Z
X×Pv (X)
+
Z
X×Pv (X)
−
Z
)(dx, dµ)
TN (x, µ)L(x̃N
1 (t), δµt )(dx, dµ)
TN (x, µ)L(x̃N
1 (t), δµt )(dx, dµ)
X×Pv (X)
TN (x, µ)L(x̂N
1 (t), δµt )(dx, dµ) .
(4.6)
First, let us consider the second term in (4.6). We have
Z
Z
lim
TN (x, µt )L(x̃N
(t))(dx)
−
TN (x, µt )L(x̂N
1
1 (t))(dx) = 0
N →∞
X
X
as {TN ( · , µt )}N ≥1 ⊂ Cv (X) with supN ≥1 kTN ( · , µt )kv < ∞.
Now, consider the first term in (4.6). To this end, let us define F := TN (x, · ) :
x ∈ X, N ≥ 1) . Then, we have
lim
N →∞
Z
(N )
X×Pv (X)
TN (x, µ)L(x̃N
1 (t), ẽt
−
Z
X×Pv (X)
≤ lim
N →∞
Z Z
X
TN (x, µ)L(x̃N
1 (t), δµt )(dx, dµ)
(N )
Pv (X)
)(dx, dµ)
TN (x, µ)L(ẽt
24
|x̃N
1 (t))(dµ|dx)
−
Z
Pv (X)
TN (x, µ)L(δµt )(dµ) L(x̃N
1 (t))(dx)
(N )
N
N
N
≤ lim E E TN (x̃1 (t), ẽt ) − TN (x̃1 (t), µt ) x̃1 (t)
N →∞
(N )
≤ lim E sup Ψ(ẽt ) − Ψ(µt )
N →∞
Ψ∈F
= 0 (by Lemma 4.7).
This completes the proof.
For any sequence {gN }N ≥1 ⊂ Cv (X) with supN ≥1 kgN kv =: L < ∞ and for any
t ≥ 0, let us define
Z
(N )
gN (y)p(dy|x, a, µ)π̃t (da|x).
lN,t (x, µ) :=
A×X
Note that for any (µ, ν) ∈ Pv (X)2 , we have
Z
(N )
sup |lN,t (x, µ) − lN,t (x, ν)| ≤ L sup kp( · |x, a, µ) − p( · |x, a, ν)kv π̃t (da|x)
x∈X
x∈X
N ≥1
A
≤ Lωp (ρ̃v (µ, ν)).
Since ωp (r) → 0 as r → 0 by Assumption 2-(h), the family {lN,t (x, · ) : x ∈ X, N ≥ 1}
is equicontinuous. Moreover, the function
Ft (µ) := sup |lN,t (x, µ) − lN,t (x, µt )| ≤ Lωp (ρ̃v (µ, µt ))
x∈X
N ≥1
is real-valued and kFt k∗v < ∞ since kωp (ρ̃v ( · , µt ))k∗v < ∞ by again Assumption 2(h). Note also that lN,t ( · , µ) ∈ Cv (X) for all µ ∈ Pv (X) and N ≥ 1, and
supN ≥1 klN,t( · , µ)kv < ∞ for all µ ∈ Pv (X).
Indeed, we have |lN,t (x, µ)| ≤ Lαv(x), and so, lN,t ( · , µ) ∈ Bv (X) for all µ ∈
Pv (X). Furthermore, if xn → x in X, then since
Z
rn (a) :=
gN (y)p(dy|xn , a, µ)
X
continuously converges to
r(a) :=
Z
gN (y)p(dy|x, a, µ)
X
and krn k ≤ Lv(xn ) ≤ L supn≥1 v(xn ) < ∞, we have limn→∞ lN,t (xn , µ) = lN,t (x, µ)
(N )
by [37, Theorem 3.3] as π̃t
Cv (X). We also have
is assumed to be weakly continuous. Hence, lN,t ( · , µ) ∈
|lN,t (x, µ)|
sup sup
= sup sup
v(x)
N ≥1 x∈X
N ≥1 x∈X
R
≤ L sup
(N )
g (y)p(dy|x, a, µ)π̃t (da|x)
A×X N
R
v(x)
(N )
A×X
v(y)p(dy|x, a, µ)π̃t
v(x)
x∈X
25
(da|x)
≤ Lα < ∞.
Hence, supN ≥1 klN,t ( · , µ)kv < ∞. Using these observations, we now prove the following result.
Proposition 4.9. For any sequence {gN }N ≥1 ⊂ Cv (X) with supN ≥1 kgN kv < ∞
and for any t ≥ 0, we have
N
lim L(x̃N
1 (t))(gN ) − L(x̂1 (t))(gN ) = 0.
N →∞
Therefore, by Lemma 4.8, for any t ≥ 0, we have
(N )
lim L(x̃N
1 (t), ẽt
N →∞
)(TN ) − L(x̂N
1 (t), δµt )(TN ) = 0
for any sequence {TN } of functions on X × Pv (X) satisfying hypothesis (i)-(iv) in
Lemma 4.8.
Proof. We prove the result by induction on t. The claim trivially holds for t = 0
N
as L(x̃N
1 (0)) = L(x̂1 (0)) = µ0 for all N ≥ 1. Suppose the claim holds for t and
consider t + 1. We can write
Z
(N )
N
N
L(x̃1 (t + 1))(gN ) − L(x̂1 (t + 1))(gN ) =
)(dx, dµ)
lN,t (x, µ)L(x̃N
1 (t), ẽt
X×Pv (X)
−
Z
X×Pv (X)
lN,t (x, µ)L(x̂N
1 (t), δµt )(dx, dµ) .
Since the family {lN,t}N ≥1 satisfies the hypothesis of Lemma 4.8 as we showed above,
the last term converges to zero as N → ∞. This completes the proof.
Recall that we have
Jµ (π̃ (N ) ) =
∞
X
t=0
β t E c(x̂N (t), âN (t), µt ) .
Theorem 4.10. Let {π̃ (N ) }N ≥1 ⊂ Mc1 be an arbitrary sequence of policies for
Agent 1. Then, we have
(N )
lim J1
N →∞
(π̃ (N ) , π, . . . , π) − Jµ (π̃ (N ) ) = 0.
Proof. Fix any t ≥ 0 and define
TN,t (x, µ) :=
Z
(N )
c(x, a, µ)π̃t
(da|x).
A
Note that for any (µ, ν) ∈ Pv (X)2 , we have
Z
Z
(N )
(N )
c(x, a, µ)π̃t (da|x) − c(x, a, ν)π̃t (da|x)
sup TN,t (x, µ) − TN,t (x, ν) = sup
x∈X
N ≥1
x∈X
N ≥1
A
A
≤ ωc (ρ̃v (µ, ν)).
Since ωc (r) → 0 as r → 0, the family TN,t (x, · ) : x ∈ X, N ≥ 1 is equicontinuous.
Moreover, the function
Ft (µ) := sup TN,t (x, µ) − TN,t (x, µt )
x∈X
N ≥1
26
is real-valued and kFt k∗v < ∞.
One can also prove that TN ( · , µ) ∈ Cv (X) for all N ≥ 1 and µ ∈ Pv (X), and
supN ≥1 kTN ( · , µ)kv < ∞ for all µ ∈ Pv (X). Therefore, by Proposition 4.9, we have
N (N )
N
) − E c(x̂N
lim E c(x̃N
1 , ã1 , ẽt
1 , â1 , µt ) = 0.
N →∞
Since t is arbitrary, above result holds for all t ≥ 0. By using the same method as in
the last part of Proposition 4.5, we can conclude that
(N )
lim J1
N →∞
(π̃ (N ) , π, . . . , π) − Jµ (π̃ (N ) ) = 0.
by the dominated convergence theorem.
As a corollary of Proposition 4.5 and Theorem 4.10, we obtain the following result.
Corollary 4.11. We have
(N )
lim J1
N →∞
(π̃ (N ) , π, . . . , π) ≥ inf
Jµ (π ′ )
′
π ∈Π
= Jµ (π)
(N )
= lim J1
N →∞
(π, π, . . . , π).
Now we are ready to prove the main result of the paper:
Proof. [Proof of Theorem 4.1] For sufficiently large N , we need to prove that
(N )
Ji
(N )
(π (N ) ) ≤ inf c Ji
π i ∈M
(N )
(π −i , π i ) + ε
(4.7)
i
for each i = 1, . . . , N . Since the transition probabilities and the one-stage cost functions are the same for all agents, it is sufficient to prove (4.7) for Agent 1 only. Given
ǫ > 0, for each N ≥ 1, let π̃ (N ) ∈ Mc1 be such that
(N )
J1
(N )
(π̃ (N ) , π, . . . , π) < ′inf c J1
π ∈M1
ε
(π ′ , π, . . . , π) + .
3
Then, by Corollary 4.11, we have
(N )
lim J1
N →∞
(π̃ (N ) , π, . . . , π) = lim Jµ (π̃ (N ) )
N →∞
≥ inf′ Jµ (π ′ )
π
= Jµ (π)
(N )
= lim J1
N →∞
(π, π, . . . , π).
Therefore, there exists N (ε) such that
(N )
inf J1
π ′ ∈Mc1
(N )
(π ′ , π, . . . , π) + ε > J1
(π̃ (N ) , π, . . . , π) +
ε
3
(N )
≥ J1 (π, π, . . . , π).
≥ Jµ (π) +
for all N ≥ N (ε).
27
2ε
3
5. Example. In this section, we consider an example, the additive noise model,
in order to illustrate our results. In this model, the dynamics of a generic agent for
N -agent game problem are given by
xN
i (t + 1) =
N
1 X
N
N
N
N
N
f (xN
i (t), ai (t), xj (t)) + g(xi (t), ai (t))wi (t),
N j=1
(5.1)
N
N
N
where xN
i (t), xj (t) ∈ X, ai (t) ∈ A, and wi (t) ∈ W. Here, we assume that X =
W = R, A ⊂ R, and the ‘noise’ {wiN (t)} is a sequence of i.i.d. random variables with
standard normal distribution. Note that we can write (5.1) as
Z
(N )
N
N
N
N
f (xN
(dy) + g(xN
xi (t + 1) =
i (t), ai (t), y) et
i (t), ai (t))wi (t)
X
(N )
N
N
N
) + g(xN
= F (xN
i (t), ai (t))wi (t),
i (t), ai (t), et
R
where F (x, a, µ) := X f (x, a, y)µ(dy). The one-stage cost function of a generic agent
is given by
(N )
N
c(xN
i (t), ai (t), et
N
1 X
N
N
d(xN
i (t), ai (t), xj (t))
N j=1
Z
(N )
N
d(xN
(dy),
=
i (t), ai (t), y) et
)=
X
for some measurable function d : X × A × X → [0, ∞).
For this model, Assumption 1 holds with w(x) = v(x) = 1+x2 under the following
conditions: (i) A is compact, (ii) g is continuous, and f is bounded and continuous,
(iii) supa∈A g 2 (x, a) ≤ Lx2 for some L > 0, (iv) sup(x,a,y)∈K×A×X d(x, a, y) < ∞ for
any compact K ⊂ X, (v) d(x, a, y) ≤ Rw(x)w(y) for some R > 0, (vi) ωd (r) → 0 as
r → 0, where
ωd (r) = sup
y∈X
|d(x, a, y) − d(x, a, y)|,
sup
(x,a),(x′ ,a′ ):
|x−x′ |+|a−a′ |≤r
(vii) α2 β < 1, where α = max{1 + kf k, L}.
Indeed, we have
Z
Z
2
q(y)m(dy)
(1 + y 2 )p(dy|x, a, µ) =
1 + F (x, a, µ) + g(x, a)y
X
X
≤ 1 + kf k2 + g 2 (x, a) ≤ α(1 + x2 ),
where q is density of the standard normal distribution and m is the Lebesgue measure. Hence, Assumption 1-(c) holds. In order to verify Assumption 1-(d), suppose
(xn , an , µn ) → (x, a, µ) and let g ∈ Cb (X). Then, we have
Z
Z
lim
g(y)p(dy|xn , an , µn ) = lim
g F (xn , an , µn ) + g(xn , an )z q(z)m(dz)
n→∞ X
n→∞ X
Z
g F (x, a, µ) + g(x, a)z q(z)m(dz)
=
X
28
since g and F are continuous, where the continuity of F follows from [37, Theorem
3.3] and the fact that f is bounded and continuous. Therefore, the transition probability
p( · |x, a, µ) is weakly continuous. Similarly, one can verify that the function
R
(1
+
y 2 )p(dy|x, a, µ) is continuous. Thus, Assumption 1-(d) holds. Note that AsX
sumption 1-(e) holds if the initial distribution µ0 has a finite second moment. Finally,
we will verify Assumption 1-(a) and (f). For (a), suppose (xn , an , µn ) → (x, a, µ).
Then, we have
Z
Z
|c(xn , an , µn ) − c(x, a, µ)| =
d(xn , an , y)µn (dy) − d(x, a, y)µ(dy)
X
X
Z
Z
≤
d(xn , an , y)µn (dy) − d(x, a, y)µn (dy)
X
X
Z
Z
d(x, a, y)µn (dy) − d(x, a, y)µ(dy)
X
X
Z
Z
≤ ωd |xn − x| + |an − a| +
d(x, a, y)µn (dy) − d(x, a, y)µ(dy) .
X
X
The last expression goes to zero as N → ∞, since d(x, a, · ) ∈ Cb (X) and ωd (r) → 0
as r → 0. Thus, c is continuous. For (f), we have
Z
sup
c(x, a, µ) =
sup
d(x, a, y)µ(dy)
(a,µ)∈A×Pvt (X)
(a,µ)∈A×Pvt (X)
≤ Rv(x)
X
sup
(a,µ)∈A×Pvt (X)
Z
v(y)µ(dy) ≤ Rαt v(x).
X
Hence, Assumption 1-(f) holds with γ = α. Therefore, under (i)-(vii), there exists a
mean-field equilibrium for the mean-field game that arises from the above finite-agent
game problem.
For the same model, Assumption 2 holds under the following conditions: (viii)
d(x, a, y) is (uniformly) Hölder continuous in y with exponent p and Hölder constant
Kd , (ix) f (x, a, y) is (uniformly) Hölder continuous in y with exponent p and Hölder
constant Kf , (x) g is bounded and inf (x,a)∈X×A |g(x, a)| =: θ > 0.
Indeed, we have
Z
Z
ωc (r) = sup
sup
d(x, a, y)µ(dy) − d(x, a, y)ν(dy)
(x,a)∈X×A
µ,ν:
ρ̃v (µ,ν)≤r
sup
sup
≤
(x,a)∈X×A
µ,ν:
ρ̃v (µ,ν)≤r
X
X
Kd Wp (µ, ν)p
≤ Kd r p .
Hence, ωc (r) → 0 as r → 0 and ωc (ρ̃( · , µ)) has a finite v-norm. Therefore, Assumption 2-(h) holds for ωc . For ωp , we have
ωp (r) =
=
sup
sup
(x,a)∈X×A
µ,ν:
ρ̃v (µ,ν)≤r
sup
sup
(x,a)∈X×A
kp( · |x, a, µ) − p( · |x, a, ν)kv
sup
µ,ν:
l∈Lipλ (1,R)
ρ̃v (µ,ν)≤r
Z
X
l F (x, a, µ) + g(x, a)z q(z)m(dz)
29
−
Z
l(F (x, a, ν) + g(x, a)z)q(z)m(dz)
Z
F (x, a, µ)
m(dz)
sup
sup
l g(x, a)z q z −
= sup
µ,ν:
g(x, a)
l∈Lipλ (1,R) X
(x,a)∈X×A
ρ̃v (µ,ν)≤r
Z
F (x, a, ν)
−
m(dz)
l g(x, a)z q z −
g(x, a)
X
Z
F (x, a, µ)
sup
≤ klk sup
sup
h(z)q z −
m(dz)
µ,ν:
g(x, a)
h∈Lipλ (1,R) X
(x,a)∈X×A
ρ̃v (µ,ν)≤r
Z
F (x, a, ν)
− h(z)q z −
m(dz) ,
(5.2)
g(x, a)
X
X
where (5.2) follows from the fact that l(g(x, a) · ) ∈ Lipλ (klk, R). Note that for any
compact interval K = [−k, k] ⊂ X, we can upper bound (5.2) as follows:
Z
F (x, a, µ)
m(dz)
sup
(5.2) ≤ klk sup
sup
h(z)q z −
µ,ν:
g(x, a)
h∈Lipλ (1,R)
(x,a)∈X×A
K
ρ̃v (µ,ν)≤r
Z
F (x, a, ν)
−
h(z)q z −
m(dz)
g(x, a)
K
Z
Z
F (x, a, µ)
F (x, a, ν)
|h(z)|q z −
+
|h(z)|q z −
m(dz) +
m(dz)
g(x, a)
g(x, a)
c
Kc
Z K
F (x, a, µ)
m(dz)
sup
≤ klk sup
sup
h(z)q z −
µ,ν:
g(x, a)
h∈Lipλ (1,R)
(x,a)∈X×A
K
ρ̃v (µ,ν)≤r
Z
F (x, a, ν)
m(dz)
h(z)q z −
−
g(x, a)
K
Z
Z
F (x, a, µ)
F (x, a, ν)
(1 + z 2 )q z −
+
m(dz) +
(1 + z 2 )q z −
m(dz)
g(x, a)
g(x, a)
Kc
Kc
(5.3)
Note that the last two terms in the last expression go to zero (uniformly in (x, a, µ, ν))
as k → ∞, since F and g are bounded, and inf (x,a)∈X×A |g(x, a)| > 0. For any ε > 0,
let Kε = [−kε , kε ] ⊂ X so that the sum of these terms is less than ε for all (x, a, µ, ν).
Let us define the Lipschitz seminorm of q on Kε as
Tε :=
sup
(x,z)∈Kε ×Kε
x6=z
|q(x) − q(z)|
.
|x − z|
Then, we have
F (x, a, µ)
h(z)q z −
m(dz)
µ,ν:
g(x, a)
h∈Lipλ (1,R) Kε
(x,a)∈X×A
ρ̃v (µ,ν)≤r
Z
F (x, a, ν)
h(z)q z −
−
m(dz) + klkε
g(x, a)
Kε
Z
F (x, a, µ) F (x, a, ν)
−
m(dz)
sup
≤ klkTε sup
sup
|h(z)|
µ,ν:
g(x, a)
g(x, a)
h∈Lipλ (1,R) Kε
(x,a)∈X×A
(5.3) ≤ klk
sup
sup
sup
Z
ρ̃v (µ,ν)≤r
30
+ klkε
klkTε
θ
Z
(1 + z 2 )m(dz)
≤
klkTε
θ
Z
(1 + z 2 )m(dz)
≤
klkTε
θ
Z
(1 + z 2 )m(dz)Kf rp + klkε.
≤
Kε
Kε
sup
µ,ν:
ρ̃v (µ,ν)≤r
|F (x, a, µ) − F (x, a, ν)| + klkε
(x,a)∈X×A
sup
sup
Kf Wp (µ, ν)p + klkε
sup
µ,ν:
(x,a)∈X×A
ρ̃v (µ,ν)≤r
Kε
Since ε is arbitrary, we have ωp (r) → 0 as r → 0 and ωp (ρ̃v ( · , µ)) has finite vnorm. Thus, Assumption 2-(h) holds. Finally, it is straightforward to prove that
Assumption 2-(i) holds, since the noise has a standard normal distribution. Therefore,
under (viii)-(x), Assumption 2 holds.
The final piece in order to deduce the existence of approximate Markov-Nash
equilibria in the above game with sufficiently many agents is to prove condition (j).
However, verification of (j) is highly dependent on the systems components, and so,
it is quite difficult to find a global assumption in order to satisfy (j). One way to
establish this is as follows. For any h ∈ Cv (X) and ν ∈ Pv (X), define
Z
Rν,h (x, a) := c(x, a, ν) + β h(y)p(dy|x, a, ν)
ZX
= c(x, a, ν) + β h F (x, a, ν) + g(x, a)z q(z)m(dz).
X
Using this, let us state the following condition:
(j’) For any h ∈ Cv (X) and ν ∈ Pv (X), there a exists unique minimizer ax ∈ A of
Rν,h (x, · ), for each x ∈ X.
Under assumption (j’), we can establish that the policy in the mean field equilibrium
satisfies the weak continuity condition (j).
Indeed, fix any t ≥ 0 and consider the policy πt at time t in π, where π is the
policy in the mean-field equilibrium. By (j’), we must have πt ( · |x) = δft (x) ( · ) for
some ft : X → A which minimizes some function Rν,h (x, a) of the above form; that
is, mina∈A Rν,h (x, a) = Rν,h (x, ft (x)) for all x ∈ X. If ft is continuous, then πt
is also continuous. Hence, in order to prove (j), it is sufficient to prove that ft is
continuous. Suppose xn → x in X. Note that l( · ) = mina∈X Rν ( · , a) is continuous.
Therefore, every accumulation point of the sequence {ft (xn )}n≥1 must be a minimizer
for Rν,h (x, · ). Since there exists a unique minimizer ft (x) of Rν,h (x, · ), the set of
all accumulation points of {ft (xn )}n≥1 must be {ft (x)}. This implies that ft (xn )
converges to ft (x) since A is compact. Hence, ft is continuous.
Remark 6. We note that, using similar ideas, the finite-horizon cost criterion;
that is,
E
X
T
t=0
c(x(t), a(t), µt ) for some T < ∞,
(5.4)
can be handled with the same quantitative results. The only part that needs to verified
differently from the infinite-horizon case is Proposition 3.10. Namely, we have to
(n)
ν (n)
ν
establish that Ft and J∗,t
continuously converge to Ft and J∗,t
, respectively, since
31
we cannot use the fixed-point argument anymore (see the proof of Proposition 3.10).
Note that, in finite-horizon case, for each n and t < T , these functions are given by
Z
(n)
(n)
(n)
ν (n)
Ft (x, a) = c(x, a, νt,1 ) + J∗,t+1
(y)p(dy|x, a, νt,1 ),
ZX
ν
(y)p(dy|x, a, νt,1 ),
Ft (x, a) = c(x, a, νt,1 ) + J∗,t+1
X
and
(n)
(n)
ν
J∗,t
(x) = min Ft
a∈A
ν
(x, a) and J∗,t
(x) = min Ft (x, a).
a∈A
Observe that the discount factor β is missing in the above equations. For t = T , we
(n)
(n)
have FT (x, a) = c(x, a, νT,1 ) and FT (x, a) = c(x, a, νT,1 ). Since c is continuous and
(n)
(n)
ν
ν
continuously converges to J∗,T
by [6,
νT,1 weakly converges to νT,1 , we have that J∗,T
(n)
Proposition 7.32]. But this implies that FT −1 continuously converges to FT −1 by the
(n)
ν
ν
discussion before Proposition 3.10, and so, J∗,T
−1 continuously converges to J∗,T −1
again by [6, Proposition 7.32]. Then, by induction hypothesis, we can conclude that
(n)
ν (n)
ν
Ft and J∗,t
continuously converge to Ft and J∗,t
, respectively, for each t ≤ T .
Therefore, Theorems 3.3 and 4.1 hold for the finite-horizon cost criterion under the
same set of assumptions. Furthermore, if we start the mean-field
game at time τ > 0
with initial measure µτ , then the pair {πt }τ ≤t≤T , {µt }τ ≤t≤T in Theorem 3.3 is still
a mean-field equilibrium for the sub-game. Similar conclusion can be reached for
the finite-agent game problem; that is, the policy {πt }τ ≤t≤T is an ε-Markov-Nash
equilibrium for the finite-agent game problem starting at time τ with i.i.d. initial
measures drawn according to µτ .
6. Conclusion. Using the mean-field approach, we have shown that there exist nearly Markov-Nash equilibria for finite-population game problems with infinitehorizon discounted costs when the number of agents is sufficiently large. Under mild
technical conditions, we have first established the existence of a Nash equilibrium
in the limiting mean-field game problem. We have then applied this policy to the
finite population game and have demonstrated that it constitutes an ε-Markov-Nash
equilibrium for games with a sufficiently large number of agents.
One interesting future direction is to study mean-field games with imperfect information. In this case, one possible approach is to use the theory developed for
partially observed Markov decision processes (POMDPs). Finally, average-cost and
risk-sensitive optimality criteria are also worth studying. In particular, using the vanishing discount approach in MDP theory (i.e., β → 1), we may be able to establish
similar results for the average cost case.
Appendix.
A. Proof of Theorem 2.3. Since the transition probabilities and the one-stage
cost functions are the same for all agents, it is sufficient to prove the result for Agent 1.
To this end, choose an arbitrary policy π (N ) ⊂ M(N ) . We prove that for any ε > 0,
there exists π 1,ε ∈ Mc1 such that
(N )
J1
(N )
(π 1,ε , π 2 , . . . , π N ) < J1
which will complete the proof.
32
(π 1 , π 2 , . . . , π N ) + ε,
To achieve this, we first describe the overall N -agent game as an Markov Decision Process (MDP) with state space XN × P(X), action space AN , and transition
probability
Q(dy, dµ|x, ν, a) := δFN (y) (dµ)
N
Y
p(dyi |xi , ai , ν),
i=1
P
where FN (y) := N1 N
i=1 δyi . The one-stage cost function is given by C(x, µ, a) =
c(x1 , a1 , µ) which corresponds to the one-stage cost function of Agent 1. The cost
function for this MDP is β-discounted cost and denoted by J(π 1 , . . . , π N ) for the
policy π (N ) = (π 1 , . . . , π N ). Observe that for any π̃ ∈ M, we have
(N )
J(π̃, π 2 , . . . , π N ) = J1
(π̃, π 2 , . . . , π N ).
Let (x(t), νt ) and a(t) denote the state and action of this MDP at time t, respectively.
Let x(t) = (x1 (t), . . . , xN (t)).
Recall that µ0 = L(x1 (0)). By Lusin’s theorem [14, Theorem 7.5.2], for any
δ0 > 0, there exists a closed set F0 ⊂ X such that µ0 (F0 ) < δ0 and π01 is weakly
continuous on F0 . Since P(A) is a convex subset of a locally convex vector space (i.e.,
the set of finite signed measures over A), by Dugundji extension theorem [18, Theorem
1
7.4] we can extend π01 : F0 → P(A) to X continuously. Let π0,δ
denote this extended
1,1
1
1
1
continuous function and define πδ := (π0,δ , π1 , π2 , . . .). Then, it is straightforward
to prove that
|J(πδ1,1 , π 2 , . . . , π N ) − J(π 1 , π 2 , . . . , π N )|
(N ) 1,1
(N ) 1
2
N
2
N
= |J1 (π , π , . . . , π ) − J1 (π , π , . . . , π )|
Z
≤ 2L0
v(x0 )µ0 (dx0 ).
F0c
Note that the last expression can be made arbitrarily small by choosing arbitrarily
small δ0 since v is µ0 -integrable.
We can apply the same method to the policy πδ1,1 ; that is, we replace π11 with a con1
tinuous π1,δ
that agrees with π11 on some closed subset F1 of X with probability 1 − δ1 .
1
1
, π1,δ
, π21 , . . .). Therefore, |J(πδ1,2 , π 2 , . . . , π N )− J(πδ1,1 , π 2 , . . . , π N )| ≤
Let πδ1,2 := (π0,δ
R
2L1 F c v(x1 )µ1 (dx1 ), where µ1 = L(x1 (1)) under πδ1,1 or πδ1,2 . Continuing in this way,
1
we will obtain πδ1,∞ ∈ Mc1 such that
|J(π 1 , π 2 , . . . , π N ) − J(π 1,∞ , π 2 , . . . , π N )|
(N )
(N )
= |J1 (π 1 , π 2 , . . . , π N ) − J1 (π 1,∞ , π 2 , . . . , π N )|
≤
∞
X
|J(π 1,t , π 2 , . . . , π N ) − J(π 1,t+1 , π 2 , . . . , π N )|
t=0
≤
∞
X
t=0
2Lt β
t
Z
v(xt )µt (dxt ).
Ftc
The last expression goes to zero with the proper choice of the sequence {δt }t≥0 . This
completes the proof.
33
B. Proof of Proposition 3.2. Let µ ∈ M and π ∈ Π be arbitrary. According
to the Ionescu-Tulcea theorem, an initial distribution µ0 on X, a policy π, and state
measure flow µ define a unique probability measure Pπ on G∞ = (X × A)∞ . Define
µπt := Pπ (dx(t)) for each t ≥ 0.
Let π̂ be the Markov policy given by
π̂t (da(t)|x(t)) := Pπ (da(t)|x(t)).
We first prove that Pπ̂ (dx(t)) =: µπ̂t = µπt for all t. We prove this by induction. For
t = 0 the claim clearly holds as the initial distribution µ0 is fixed. Assume the claim
is true for t ≥ 0. Then we have
Z
π̂
p( · |x(t), a(t), µt )π̂t (da(t)|x(t))µπ̂t (dx(t))
µt+1 ( · ) =
X×A
Z
=
p( · |x(t), a(t), µt )Pπ (da(t)|x(t))µπt (dx(t))
X×A
= µπt+1 ( · ).
This proves the claim.
To complete the proof of the first statement, take any t ≥ 0. Then we have
Z
π̂
π̂
c(x(t), a(t), µπ̂t )π̂(da(t)|x(t))µπ̂t (dx(t))
E c(x(t), a(t), µt ) =
X×A
Z
=
c(x(t), a(t), µπt )Pπ (da(t)|x(t))µπt (dx(t))
X×A
= E π c(x(t), a(t), µπt ) .
Since t is arbitrary, the above equality holds for all t ≥ 0. Thus, by Tonelli’s theorem
[7, Theorem 18.3], Jµ (π̂) = Jµ (π).
The proof of the second statement can be done similarly, so we omit the details.
C. Proof of Proposition 3.10. We use successive approximations to prove
(n)
(n)
(n)
Proposition 3.10. To ease the notation, let T (n) := T ν , J∗ := J∗ν , T := T ν , and
(n)
J∗ := J∗ν . Let u0 = u0 = 0 and
(n)
(n)
uk+1 = T (n) uk
and uk+1 = T uk .
Since T (n) and T are contractive operators on C with modulus σβα, it can be proved
that
(n)
(n)
ρ(uk , J∗ ), ρ(uk , J∗ ) ≤ (σβα)k L0 .
(6.1)
Lemma 6.1. For any compact K ⊂ X, we have
(n)
lim sup uk,t (x) − uk,t (x) = 0,
n→∞ x∈K
(n)
for all k ≥ 0 and t ≥ 0, where uk
(n)
and uk
t≥0
(n)
Since u0 = u0 =
= uk,t
= uk,t
t≥0
.
0, the claim trivially holds for
Proof. Fix any compact K ⊂ X.
k = 0 and for all t ≥ 0. Suppose that the claim holds for k and for all t ≥ 0. Then,
consider k + 1 and arbitrary t. For these indices, we have
(n)
sup uk+1,t (x) − uk+1,t (x)
x∈K
34
Z
(n)
(n)
(n)
= sup min c(x, a, νt,1 ) + β uk,t+1 (y)p(dy|x, a, νt,1 )
x∈K a∈A
X
Z
− min c(x, a, νt,1 ) + β uk,t+1 (y)p(dy|x, a, νt,1 )
a∈A
≤
X
(n)
sup
(x,a)∈K×A
c(x, a, νt,1 ) − c(x, a, νt,1 )
+β
sup
(x,a)∈K×A
Z
(n)
X
(n)
uk,t+1 (y)p(dy|x, a, νt,1 )
−
Z
uk,t+1 (y)p(dy|x, a, νt,1 ) .
X
Note that c(· , · , ν (n) ) converges to c(· , · , ν) continuously as n → ∞. Furthermore,
(n)
(n)
since uk,t+1 converges to uk,t+1 continuously (as uk,t+1 , uk,t+1 are continuous and
(n)
(n)
uk,t+1 uniformly converges to uk,t+1 over compact sets by assumption), uk,t+1 (y) ≤
R
R
(n)
(n)
Lt+1 v(y) for all n ≥ 1, and X v(y)p(dy|x(n) , a(n) , νt,1 ) → X v(y)p(dy|x, a, νt,1 ) for
R
(n)
(n)
any (x(n) , a(n) ) → (x, a) in X × A, by [37, Theorem 3.3] X uk,t+1 (y)p(dy|· , · , νt,1 )
R
converges to X uk,t+1 (y)p(dy|· , · , νt,1 ) continuously as n → ∞. Since continuous
convergence is equivalent to uniform convergence over compact sets for continuous
functions, the last expression goes to zero as n → ∞. This completes the proof.
Using previous lemma, we now complete the proof of Proposition 3.10.
Fix any compact K ⊂ X. For all k ≥ 0, we have
(n)
sup
x∈K
(n)
uk,t (x) − uk,t (x)
J∗,t (x) − J∗,t (x)|
(n)
(n)
≤ kJ∗,t − uk,t kv + sup
+ kuk,t − J∗,t kv
v(x)
v(x)
x∈K
(n)
(n)
(n)
≤ σ t ρ(J∗ , uk ) + sup
uk,t (x) − uk,t (x)
v(x)
x∈K
+ σ t ρ(J∗ , uk )
(n)
≤ 2σ t (σβα)k L0 + sup
x∈K
uk,t (x) − uk,t (x)
v(x)
(by (6.1)).
This last expression can be made arbitrary small by first choosing large enough k and
then large enough n. This completes the proof since supx∈K v(x) < ∞.
D. Auxiliary Lemma. In this section, we state and prove an auxiliary lemma
which is a generalization of [11, Lemma A.2] to unbounded real-valued functions.
Lemma 6.2. Let P : X → P(X) be a transition probability on X given X. Fix
N ≥ 1 and let y1 , . . . , yN ∈ X. Let X1 , . . . , XN be independent random variables such
PN
PN
(N )
(N )
that L(Xi ) = P ( · |yi ). Let e0 ( · ) := N1 i=1 δyi and e1 ( · ) := N1 i=1 δXi . Then,
we have
N
(N )
2
1 X
(N )
E[Zi2 ] + E[Zi ]2 ,
E |e1 (g) − e0 P (g)| ≤ 2
N i=1
where Zi = g(Xi ).
Proof. By Jensen’s inequality, we have
(N )
2
(N )
E |e1 (g) − e0 P (g)| = E
2
N
N Z
1 X
1 X
g(Xi ) −
g(y)P (dy|yi )
N i=1
N i=1 X
35
≤E
N
N Z
2
1 X
1 X
g(Xi ) −
g(y)P (dy|yi ) .
N i=1
N i=1 X
The result follows by expanding and simplifying the term inside the expectation.
REFERENCES
[1] S. Adlakha, R. Johari, and G.Y. Weintraub, Equilibria of dynamic games with many
players: Existence, approximation, and market structure, Journal of Economic Theory,
156 (2015), pp. 269–316.
[2] C.D. Aliprantis and K.C. Border, Infinite Dimensional Analysis, Berlin, Springer, 3rd ed.,
2006.
[3] P. Barelli and J. Duggan, A note on semi-Markov perfect equlibria in discounted stochastic
games, J. Econ. Theory, 15 (2014), pp. 596–604.
[4] D. Bauso, H. Tembine, and T. Başar, Robust mean field games, Dyn. Games Appl., 6 (2016),
pp. 277–303.
[5] A. Bensoussan, J. Frehse, and P. Yam, Mean Field Games and Mean Field Type Control
Theory, Springer, New York, 2013.
[6] D. P. Bertsekas and S. E. Shreve, Stochastic optimal control: The discrete time case,
Academic Press New York, 1978.
[7] P. Billingsley, Probability and Measure, Wiley, 3rd ed., 1995.
, Convergence of Probability Measures, New York: Wiley, 2nd ed., 1999.
[8]
[9] A. Biswas, Mean field games with ergodic cost for discrete time markov processes.
arXiv:1510.08968, 2015.
[10] V.I. Bogachev, Measure Theory: Volume II, Springer, 2007.
[11] A. Budhiraja and A.P. Majumder, Long time results for a weakly interacting particle system
in discrete time, Stochastic Analysis and Applications, 33 (2015), pp. 429–463.
[12] P. Cardaliaguet, Notes on mean-field games, 2011.
[13] R. Carmona and F. Delarue, Probabilistic analysis of mean-field games, SIAM J. Control
Optim., 51 (2013), pp. 2705–2734.
[14] R. M. Dudley, Real Analysis and Probability, Cambridge University Press, 2004.
[15] R. Elliot, X. Li, and Y. Ni, Discrete time mean-field stochastic linear-quadratic optimal
control problems, Automatica, 49 (2013), pp. 3222–3233.
[16] D.A. Gomes, J. Mohr, and R.R. Souza, Discrete time, finite state space mean field games,
J. Math. Pures Appl., 93 (2010), pp. 308–328.
[17] D.A. Gomes and J. Saúde, Mean field games models - a brief survey, Dyn. Games Appl., 4
(2014), pp. 110–154.
[18] A. Granas and J. Dugundji, Fixed Point Theory, Springer, 2003.
[19] M. Hairer and J.C. Mattingly, Yet Another Look at Harris’ Ergodic Theorem for Markov
Chains, Springer Basel, 2011, pp. 109–117.
[20] O. Hernández-Lerma and J.B. Lasserre, Discrete-Time Markov Control Processes: Basic
Optimality Criteria, Springer, 1996.
[21]
, Further Topics on Discrete-Time Markov Control Processes, Springer, 1999.
[22] K. Hinderer, Foundations of Non-stationary Dynamic Programming with Discrete Time Parameter, Springer-Verlag, 1970.
[23] M. Huang, Large-population LQG games involving major player: The Nash certainity equivalence principle, SIAM J. Control Optim., 48 (2010), pp. 3318–3353.
[24] M. Huang, P.E. Caines, and R.P. Malhamé, Large-population cost coupled LQG problems
with nonuniform agents: Individual-mass behavior and decentralized ǫ-Nash equilibria,
IEEE. Trans. Autom. Control, 52 (2007), pp. 1560–1571.
[25] M. Huang, R.P. Malhamé, and P.E. Caines, Large population stochastic dynamic games:
Closed loop McKean-Vlasov sysyems and the Nash certainity equivalence principle, Communications in Information Systems, 6 (2006), pp. 221–252.
[26] A. Jaśkiewicz and A. Nowak, Stationary almost Markov perfect equlibria in discounted
stochastic games, Math. Oper. Res., 41 (2016), pp. 430–441.
[27] B. Jovanovic and R.W. Rosenthal, Anonymous sequential games, Journal of Mathematical
Economics, 17 (1988), pp. 77–87.
[28] H.J. Langen, Convergence of dynamic programming models, Math. Oper. Res., 6 (1981),
pp. 493–512.
[29] J. Lasry and P.Lions, Mean field games, Japan. J. Math., 2 (2007), pp. 229–260.
36
[30] Y.J. Levy and A. McLennan, Corrigendum to: Discounted stochastic games with no stationary Nash equilibrium: Two examples, Econometrica, 83 (2015), pp. 1237–1252.
[31] J. Moon and T. Başar, Discrete-time decentralized control using the risk-sensitive performance criterion in the large population regime: a mean field approach, in ACC 2015,
Chicago, Jul. 2015.
[32] J. Moon and T. Başar, Discrete-time mean field Stackelberg games with a large number of
followers, in CDC 2016, Las Vegas, Dec. 2016.
, Robust mean field games for coupled Markov jump linear systems, International Journal
[33]
of Control, 89 (2016), pp. 1367–1381.
[34]
, Linear quadratic risk-sensitive and robust mean field games, IEEE. Trans. Autom.
Control, 62 (2017).
[35] M. Nourian and G.N. Nair, Linear-quadratic-Gaussian mean field games under high rate
quantization, in CDC 2013, Florence, Dec. 2013.
[36] K.R. Parthasarathy, Probability Measures on Metric Spaces, AMS Bookstore, 1967.
[37] R. Serfozo, Convergence of Lebesgue integrals with varying measures, Sankhya Ser.A, (1982),
pp. 380–402.
[38] H. Tembine, Q. Zhu, and T. Başar, Risk-sensitive mean field games, IEEE. Trans. Autom.
Control, 59 (2014), pp. 835–850.
[39] C. Villani, Optimal transport: Old and New, Springer, 2009.
37
| 3cs.SY
|
Published as a conference paper at ICLR 2018
E FFICIENT S PARSE -W INOGRAD
C ONVOLUTIONAL N EURAL N ETWORKS
Xingyu Liu∗, Jeff Pool†, Song Han‡ §, William J. Dally∗†
∗
Stanford University, † NVIDIA, ‡ Massachusetts Institute of Technology, § Google Brain
{xyl, dally}@stanford.edu
arXiv:1802.06367v1 [cs.CV] 18 Feb 2018
A BSTRACT
Convolutional Neural Networks (CNNs) are computationally intensive, which
limits their application on mobile devices. Their energy is dominated by the
number of multiplies needed to perform the convolutions. Winograd’s minimal
filtering algorithm (Lavin, 2015) and network pruning (Han et al., 2015) can reduce
the operation count, but these two methods cannot be directly combined – applying
the Winograd transform fills in the sparsity in both the weights and the activations.
We propose two modifications to Winograd-based CNNs to enable these methods
to exploit sparsity. First, we move the ReLU operation into the Winograd domain
to increase the sparsity of the transformed activations. Second, we prune the
weights in the Winograd domain to exploit static weight sparsity. For models on
CIFAR-10, CIFAR-100 and ImageNet datasets, our method reduces the number of
multiplications by 10.4×, 6.8× and 10.8× respectively with loss of accuracy less
than 0.1%, outperforming previous baselines by 2.0×-3.0×. We also show that
moving ReLU to the Winograd domain allows more aggressive pruning.
1
I NTRODUCTION
Deep Convolutional Neural Networks (CNNs) have shown significant improvement in many machine
learning applications. However, CNNs are compute-limited. Their performance is dominated by the
number of multiplies needed to perform the convolutions. Moreover, the computational workload
of CNNs continues to grow over time. LeCun et al. (1998) proposed a CNN model with less than
2.3 × 107 multiplies for handwritten digit classification. Later, Krizhevsky et al. (2012) developed
AlexNet, an ImageNet-winning CNN with more than 1.1 × 109 multiplies. In 2014, ImageNetwinning and runner up CNNs increased the number of multiplies to 1.4 × 109 (Szegedy et al., 2015)
and 1.6 × 1010 (Simonyan & Zisserman, 2015) respectively. Despite the powerful representational
ability of large scale CNNs, their computational workload prohibits deployment on mobile devices.
Two research directions have been explored to address the problem. Lavin (2015) proposed using
Winograd’s minimal filtering algorithm (Winograd, 1980) to reduce the number of multiplies needed
to perform 3 × 3 kernel convolutions. On the other end, pruning the model (Han et al., 2015; 2016b)
and exploiting the dynamic sparsity of activations due to ReLU also reduces the required multiplies.
Unfortunately, the above two directions are not compatible: the Winograd transformation fills in the
zeros in both the weights and the activations (Figure 1(a)) – eliminating the gain from exploiting
sparsity. Thus, for a pruned network, Winograd’s algorithm actually increases the number of
multiplies; the loss of sparsity more than offsets the reduced operation count.
In this paper, we introduce two modifications to the original Winograd-based convolution algorithm to
eliminate this problem. First, we move the ReLU operation to be after the Winograd transform to also
make the activations sparse at the point where the multiplies are performed. Second, we prune the
weights after (rather than before) they are transformed. Thus, the weights are sparse when the elementwise multiply is performed — reducing the operation count. Together, these two modifications enable
the gains of Winograd’s algorithm and of exploiting sparsity to be combined. We open-source our
code and models at https://github.com/xingyul/Sparse-Winograd-CNN.
1
Published as a conference paper at ICLR 2018
Train
Activation
Layer 𝑖
Activation
Layer 𝑖
Kernel
Layer 𝑖
ReLU
Prune
Activation
Layer 𝑖
ReLU
Train
Transformed
Kernel
Layer 𝑖
𝐺𝑇 ⋅ ( ) ⋅ 𝐺
𝐵𝑇 ⋅ ( ) ⋅ 𝐵
Eltwise
Muliply
ReLU-ed
Transformed
Activation
Layer 𝑖
Winograd
domain
𝐵𝑇 ⋅
Transformed
Kernel
Layer 𝑖
Eltwise
Product
Channel wise
Summation
𝐴𝑇 ⋅ ( ) ⋅ 𝐴
⋅𝐵
Prune
Transformed
Activation
Layer 𝑖
ReLU-ed
Transformed
Activation
Layer 𝑖
Winograd
domain
Train
Transformed
Kernel
Layer 𝑖
ReLU
Eltwise
Muliply
Prune
Eltwise
Muliply
Eltwise
Product
Channel wise
Summation
𝐴𝑇 ⋅ ( ) ⋅ 𝐴
Activation
Layer 𝑖 + 1
(a)
𝐵𝑇 ⋅ ( ) ⋅ 𝐵
Transformed
ReLU-ed
Activation
Layer 𝑖
Winograd
domain
Eltwise
Product
Channel wise
Summation
𝐴𝑇 ⋅ ( ) ⋅ 𝐴
Activation
Layer 𝑖 + 1
(b)
Activation
Layer 𝑖 + 1
(c)
Figure 1: Combining Winograd convolution with sparse weights and activations. (a) Conventional
Winograd-based convolution fills in the zeros in both the weights and activations. (b) Pruning the
4 × 4 transformed kernel restores sparsity to the weights. (c) Our proposed Winograd-ReLU CNN.
Moving the ReLU layer after Winograd transformation also restores sparsity to the activations.
2
R ELATED W ORK
Linear Algebra property in Convolution: Previous research proposes using the linear algebra
property of convolution to reduce the number of multiplies by trading additions for multiplies. Cong
& Xiao (2014) convert convolution into matrix multiplies and utilize the linear algebra property at
the sub-matrix block level. This approach achieves a 47% saving in multiplies. Lavin (2015) exploits
the element-level linear algebra property of convolution, i.e. Winograd’s minimal filtering algorithm
(Winograd, 1980). This approach reduces the number of multiplies by 2.25× to 4×, depending on
the image patch size used in the algorithm. Winograd’s algorithm is also used in a state-of-the-art
deep learning library, cuDNN (Chetlur et al., 2014), to improve computation efficiency.
Model Compression: Model compression reduces the number of multiplies of CNNs by pruning
network parameters (LeCun et al., 1990; Hassibi et al., 1993) and exploiting weight sparsity. Han
et al. (2015; 2016b) proposed learning the sparsity pattern of network weights by eliminating weights
whose absolute value is less than an empirical threshold. This approach can prune the convolutional
layers of the model to only 30% − 50% of the original size and reduce the number of multiplies
required. Liu et al. (2017) first proposed pruning and re-training the weights in Winograd domain for
conventional Winograd convolution. Li et al. (2017) later showed promising results on large datasets
and reported 90% sparsity in the Winograd parameters of AlexNet with less than 0.1% accuracy loss.
Dynamic Activation Sparsity: The ReLU non-linearity sets activations whose values are negative to
zero, causing dynamic sparsity in activations. Model compression can work in tandem with dynamic
activation sparsity and reduce multiplication workload. Han et al. (2015) showed that exploiting
sparsity of both weights and activations can reduce the number of multiplies by 4 − 11×. Huan
et al. (2016) further proposed to manually set a small positive ReLU threshold at test time to exploit
greater sparsity in activation without losing testing accuracy. Research in novel architectures also
led to optimizations for deep learning accelerators to exploit the sparsity in activations. Han et al.
(2016a) proposed using a Leading Non-zero Detection unit (LNZD) for their fully-connected layer
accelerator to efficiently skip zeros in input activations. Albericio et al. (2016) proposed a similar
mechanism for a convolution layer accelerator.
2
Published as a conference paper at ICLR 2018
3
S PARSE W INOGRAD C ONVOLUTION
We first introduce the conventional Winograd convolution and show how sparsity of weights or
activations is lost during the dataflow of the algorithm. We then present the novel Winograd-ReLU
CNN architecture. It preserves sparsity in both weights and activations before multiplies are performed
and significantly reduces the computational workload.
3.1
S PARSITY IN C ONVENTIONAL S PATIAL AND W INOGRAD CNN
The basic block of the conventional Winograd convolution algorithm works on an p×p patch (denoted
by d) extracted with stride of (p − 2) × (p − 2) from an H × W input feature map. With “valid”
padding, the p×p patch is convolved with a 3×3 kernel (denoted by g) to produce an (p−2)×(p−2)
output patch (denoted by S). The output patches are assembled into an output feature map.
Input activation patch d and kernel g (spatial-domain activation and weights) are transformed using
matrices B and G to be B T dB and GgGT (Winograd-domain activation and weights) respectively,
both with shape p × p. After element-wise product in Winograd-domain, the output activation S is
obtained using matrix A (equation (1)). Matrices B, G and A are p-specific. When p = 4, B and A
consists of 1, −1 and 0, so the multiplication with B and A only requires addition. It reduces the
number of multiplies from 9(p − 2)2 to p2 . Lavin (2015) gives details of the algorithm.
S = AT [[GgGT ]
[B T dB]]A
(1)
Spatial Baseline Network: When using a “vanilla” pruned network, as introduced by Han et al.
(2015), a ReLU non-linear operation is performed by the previous layer on spatial-domain input d
and spatial-domain weight g is pruned. The output activation patch S is obtained from equation (2).
This is illustrated in Figure 1(a) for p = 4. Though g and d may both be sparse due to pruning and
ReLU respectively, the element-wise multiply is dense due to G(·)GT and B(·)B T transformations
filling the spatial-domain zeros. Sparsity does not reduce the number of multiplies in Winograd’s
algorithm.
S = AT [[GPrune(g)GT ] [B T ReLU(d)B]]A
(2)
Winograd Native Pruned Network: When using the Winograd-domain pruned network introduced
by Liu et al. (2017) and Li et al. (2017), the spatial-domain input d is ReLU-ed by the previous layer
while the Winograd-domain weight GgGT is pruned. The output activation patch S is obtained from
equation (3). The algorithm when p = 4 is also illustrated in Figure 1(b). Though Winograd-domain
weights are sparse due to pruning, Winograd-domain activations are still dense due to B(·)B T
transforms. The sparsity in spatial activations due to ReLU does not reduce the number of multiplies.
S = AT [[Prune(GgGT )]
3.2
[B T ReLU(d)B]]A
(3)
W INOGRAD -R E LU CNN
To address the above problems, we introduce the Winograd-ReLU Network. Instead of applying
ReLU to the activations in the spatial domain, we apply ReLU to the activations in the Winograd
domain, as in equation (4) and Figure 1(c). The ReLU operation zeros all negative transformed
activations, reducing the number of multiplies in the Winograd domain.
S = AT [[Prune(GgGT )]
[ReLU(B T dB)]]A
(4)
In the Winograd-ReLU CNN, we eliminate the spatial-domain kernel entirely. Because this ReLU is
really associated with the previous layer, we perform this transformed ReLU starting with the second
layer. We point out that the proposed new CNN architecture is not mathematically equivalent to the
vanilla CNN nor the conventional Winograd CNN. Due to the change of network architecture, the
training and pruning should also be changed. Our method operates in three phases: dense training,
pruning, and retraining.
Dense training: we train a dense p × p kernel directly in the transform domain. The transformed kernel is initialized and trained directly by back-propagation through the inverse transform — eliminating
the need to maintain a kernel in the spatial domain or to transform a spatial kernel.
3
Published as a conference paper at ICLR 2018
Pruning: we prune the transformed kernel by computing the threshold t required to achieve a desired
pruning rate r and setting all weights whose absolute value less than t to zero. In our experiments,
we used the same r for all Winograd-ReLU layers. Because sensitivity varies from layer to layer, we
expect that better performance could be achieved by varying the pruning rate ri for each layer i.
Re-training: we re-train the model using a “sparsity mask” to force the weights that were pruned
to remain zero. The sparsity mask is computed during the pruning step and is kept constant during
re-training. The gradient of the network’s loss, L, with respect to the input activation and Winograd
weights can be derived using the chain rule. Equation (5) shows the calculation of input activation
gradient ∇d L and Winograd weight gradient ∇GgGT L using the loss gradient passed from upstream
layers ∇S L.
∇GgGT L = (A∇S LAT ) (B T dB) mask
(5)
∇d L = B[(A∇S LAT ) (GgGT ) mask]B T
4
E XPERIMENTS
We applied the methodology described above to several different CNNs on different datasets. The
original network models are chosen such that the majority of the convolution layers have 3 × 3
kernels. This ensures the largest portion of layers can be converted to Winograd convolution layers
and ReLU be put in Winograd domain. We used image classification datasets of different scales:
CIFAR-10, CIFAR-100 (Krizhevsky & Hinton, 2009) and ImageNet 2012 (Russakovsky et al., 2015).
For network architectures, we chose VGG-nagadomi (Nagadomi, 2014), ConvPool-CNN-C model
(Springenberg et al., 2015) and a variation of ResNet-18 (He et al., 2016a) respectively on three
datasets. Using the Tensorflow (Abadi et al., 2016) framework, we trained the spatial baseline CNN,
corresponding conventional Winograd CNN, and Winograd-ReLU CNN models from scratch. Then
the three models are iteratively pruned and re-trained. For a specific dataset, we used the same data
augmentation for the training of all models on the dataset.
4.1
CIFAR-10
We used VGG-nagadomi (Nagadomi, 2014) on the CIFAR-10 dataset. VGG-nagadomi is a lightweight version of VGGNet (Simonyan & Zisserman, 2015). It contains 8 convolution layers with 3×3
kernels. The best reported validation set accuracy it achieves on CIFAR-10 is 93.31% (Nagadomi,
2014). We trained three models from scratch. The corresponding conventional Winograd CNN
model and Winograd-ReLU CNN model can achieve validation set accuracy of 93.30% and 93.43%
respectively. The first convolution layer is most sensitive to pruning and we set its density to a
constant of 80%. We iteratively pruned and re-trained other convolution layers with density from
80% down to 20%.
Test Accuracy
94
Spatial Pruning
93
Winograd Native
Pruning
92
91
Winograd-ReLU
Pruning (ours)
90
20
30
40
50
60
70
80
(%)
Weight Density
Figure 2: Test accuracy vs density for the three models in Figure 1 on VGG-nagadomi.
Figure 2 shows test accuracy as a function of weight density for the three models. The two baseline
models can only be pruned to 60% density before accuracy falls significantly (> 0.1%). Our
Winograd-ReLU CNN model can be pruned to 40% density before falling to the same accuracy.
Table 1 shows the input activation density and compares the workloads for each pruned convolution
layer in three models. Pruning two baseline models reduces the convolution layer workload by 5.1×
and 3.7× 1 respectively. Pruning the Winograd-ReLU model reduces the convolution layer workload
by 13.3×, a 2.6× and 3.6× improvement respectively over the two baselines. The improvement of
overall network workload reduction is 2.2× and 3.0× respectively over two baselines.
1
All Winograd CNN model workload reduction results include the intrinsic 2.25× reduction.
4
Published as a conference paper at ICLR 2018
Table 1: VGG-nagadomi weight and activation density on CIFAR-10.
Spatial Baseline CNN
Winograd CNN Native
Winograd-ReLU CNN
Pruning (Han et al., 2015)
Pruning (Li et al., 2017)
Pruning (ours)
Layer
Density
Density
Density
Workload
Workload
Workload
Weight Act
Weight Act
Weight Act
conv0
80% 100%
80%
80% 100%
80%
80% 100%
80%
conv1
60%
50%
30%
60% 100%
27%
40%
46%
8%
conv2
60%
19%
12%
60% 100%
27%
40%
39%
7%
conv3
60%
37%
22%
60% 100%
27%
40%
40%
7%
conv4
60%
18%
11%
60% 100%
27%
40%
40%
7%
conv5
60%
26%
15%
60% 100%
27%
40%
38%
7%
conv6
60%
24%
14%
60% 100%
27%
40%
35%
6%
conv7
60%
35%
21%
60% 100%
27%
40%
36%
6%
conv total
20%(5.1×)
27%(3.7×)
8%(13.3×)
overall
21%(4.7×)
29%(3.5×)
10%(10.4×)
4.2
CIFAR-100
We used the ConvPool-CNN-C (Springenberg et al., 2015) model on on the CIFAR-100 dataset.
ConvPool-CNN-C contains 9 convolution layers, out of which 7 have 3 × 3 kernels. We trained three
models from scratch. The spatial baseline CNN model and conventional Winograd CNN model can
achieve single model validation accuracy of 69.34% and 69.32% respectively. The corresponding
Winograd-ReLU network model can achieve validation set accuracy of 69.75%. We pruned the first
convolution layer to a constant density of 80%. We iteratively pruned and re-trained the other layers
to densities from 80% down to 20%.
Test Accuracy
70
69
Spatial Pruning
68
Winograd Native
Pruning
Winograd-ReLU
Pruning (ours)
67
66
65
20
30
40
50
60
70
80
(%)
Weight Density
Figure 3: Test accuracy vs density for the three models in Figure 1 on ConvPool-CNN-C.
Figure 3 shows the accuracy as a function of density for spatial baseline and Winograd-ReLU models.
The spatial-baseline and Winograd-ReLU models can be pruned to 60% density without significant
(> 0.1%) loss of accuracy. In contrast, the conventional Winograd CNN model can only be pruned to
70% density. At a given density, the Winograd-ReLU model has the highest accuracy.
Table 2: ConvPool-CNN-C weight and activation density on CIFAR-100.
Spatial Baseline CNN
Winograd CNN Native
Winograd-ReLU CNN
Pruning (Han et al., 2015)
Pruning (Li et al., 2017)
Pruning (ours)
Layer
Density
Density
Density
Workload
Workload
Workload
Weight Act
Weight Act
Weight Act
conv0
80% 100%
80%
80% 100%
80%
80% 100%
80%
conv1
60%
53%
33%
70% 100%
31%
60%
54%
14%
conv2
60%
52%
32%
70% 100%
31%
60%
53%
14%
conv3
60%
77%
46%
70% 100%
31%
60%
54%
14%
conv4
60%
35%
21%
70% 100%
31%
60%
54%
14%
conv5
60%
32%
19%
70% 100%
31%
60%
42%
11%
conv6
60%
56%
33%
70% 100%
31%
60%
43%
11%
conv total
29%(3.5×)
31%(3.2×)
14%(7.1×)
overall
30%(3.4×)
32%(3.1×)
15%(6.8×)
Table 2 shows the input activation density and compares the workloads for each pruned convolution
layer in three models. Pruning two baseline models reduces the convolution layer workload by 3.5×
and 3.2× respectively. Pruning the Winograd-ReLU model reduces the workload by 7.1×, a 2.1×
and 2.2× improvement respectively over the two baselines. The improvement of overall network
workload reduction is 2.0× and 2.2× respectively over two baselines.
5
Published as a conference paper at ICLR 2018
4.3
I MAGE N ET
We used a variation of the full pre-activation version (He et al., 2016b) of ResNet-18 (He et al., 2016a)
on the ImageNet 2012 dataset. We used this version because it performs the best among various
ResNet versions and its structure suits our Winograd-ReLU approach – its ReLU units are located
before convolutions in the residual modules. The variation is different from original ResNet-18
by replacing all 2 × 2-stride 3 × 3 convolution layers with a 2 × 2 max-pooling layer followed by
a 1 × 1-stride 3 × 3 convolution layer. Such difference ensure most of convolution layers can be
converted to Winograd convolution layer. Another difference is that it doesn’t have the last max
pooling layer so the last group of residual modules has spatial size of 14 × 14, in order to keep the
spatial size even instead of odd. This setting suits Winograd convolution with p = 4 best in that even
spatial size is required for even p values.
Top-1 Test Accuracy
We trained three models from scratch. For single model and single central 224 × 224 cropping, the
spatial baseline CNN model and conventional Winograd CNN model can achieve single model top1/top-5 validation accuracy of 66.67%/87.42% and 66.84%/87.47%. The corresponding WinogradReLU CNN model can achieve validation top-1/top-5 accuracy of 66.78%/87.43%. We kept the first
convolution layer intact. We iteratively pruned other convolution layers with density rate from 80%
down to 10%.
68
67
Spatial Pruning
66
Winograd
Native Pruning
Winograd-ReLU
Pruning (ours)
65
64
63
10
20
30
40
50
60
70
80
(%)
Top-5 Test Accuracy
Weight Density
88
Spatial Pruning
87.5
Winograd
Native Pruning
Winograd-ReLU
Pruning (ours)
87
86.5
86
85.5
10
20
30
40
50
60
70
80
(%)
Weight Density
Figure 4: Top-1 and top-5 validation accuracy vs density for three models on a variation of ResNet-18.
Figure 4 shows the accuracy as a function of density for three models. The spatial baseline CNN
model and conventional Winograd CNN model can be pruned to 60% and 50% respectively without
significant (> 0.1%) loss of top-1 or top-5 accuracy. The Winograd-ReLU model can be pruned
much further, to 30%/35% density without significant (> 0.1%) loss of top-1/top-5 accuracy. At
these densities, top-1 accuracies are 66.53%, 66.45% and 66.61% for three models respectively,
with a dense spatial baseline of 66.67%; top-5 accuracies are 87.29%, 87.30% and 87.35% for three
models respectively, with a dense spatial baseline of 87.42%.
Table 3 shows the input activation density and compares the workloads for each pruned convolution
layer in three models. Pruning the two baseline models reduces the convolution layer workload by
5.1× and 4.5× respectively. Pruning the Winograd-ReLU model reduces the workload by 13.2×,
a 2.6× and 2.9× improvement respectively over the two baselines. The improvement of overall
network workload reduction is 2.3× and 2.6× respectively over two baselines.
5
D ISCUSSION
In this section, we summarize the experiment results and compare the three models in terms of a)
weight and activation dimensions and b) the dynamic density of activations. We then visualize the
kernels to illustrate the pattern of the proposed Winograd-ReLU model kernel.
6
Published as a conference paper at ICLR 2018
Table 3: ResNet-18 variation weight and activation density on ImageNet.
Spatial Baseline CNN
Winograd CNN Native
Pruning (Han et al., 2015)
Pruning (Li et al., 2017)
Layer
Density
Density
Workload
Workload
Weight Act
Weight Act
res2a_2a
60% 90%
54%
50% 100%
22%
res2a_2b
60% 64%
39%
50% 100%
22%
res2b_2a
60% 71%
43%
50% 100%
22%
res2b_2b
60% 53%
32%
50% 100%
22%
res3a_2a
60% 94%
56%
50% 100%
22%
res3a_2b
60% 35%
21%
50% 100%
22%
res3b_2a
60% 47%
28%
50% 100%
22%
res3b_2b
60% 29%
17%
50% 100%
22%
res4a_2a
60% 88%
53%
50% 100%
22%
res4a_2b
60% 23%
14%
50% 100%
22%
res4b_2a
60% 36%
22%
50% 100%
22%
res4b_2b
60% 21%
13%
50% 100%
22%
res5a_2a
60% 45%
27%
50% 100%
22%
res5a_2b
60% 14%
9%
50% 100%
22%
res5b_2a
60% 16%
10%
50% 100%
22%
res5b_2b
60% 12%
7%
50% 100%
22%
conv total
20%(5.1×)
22%(4.5×)
overall
21%(4.7×)
24%(4.2×)
5.1
Winograd-ReLU CNN
Pruning (ours)
Density
Workload
Weight Act
35% 48%
8%
35% 50%
8%
35% 50%
8%
35% 50%
8%
35% 49%
8%
35% 50%
8%
35% 49%
8%
35% 49%
8%
35% 49%
8%
35% 50%
8%
35% 50%
8%
35% 49%
8%
35% 50%
8%
35% 48%
7%
35% 48%
7%
35% 49%
8%
8%(13.2×)
9%(10.8×)
W EIGHT AND ACTIVATION D IMENSION
In a convolutional neural network, a convolution-ReLU pair acts as a classifier on a spatial patch of
an input feature. The dimension of the space being classified is the total number of elements passing
through the ReLU layer. The decision boundaries of the classifier are determined by the weights.
Insufficient non-zero weights or insufficient activations results in too simple a decision boundary and
causes accuracy loss.
Experimental results have shown that Winograd-ReLU CNN can reach the same accuracy as both
vanilla spatial baseline CNN and conventional Winograd CNN without pruning, and that WinogradReLU CNN is more robust to aggressive pruning. In this subsection we provide an explanation for
the latter observation from the aspect of activation and weight dimensions. We provide a summary on
dimensions in Table 4.
Table 4: Comparison of ReLU dimension and weight dimension in three types of networks. Assume
the convolution-ReLU pair operates on input activation of spatial size of H × W and the number of
input and output channels are C and K respectively.
Spatial Baseline
Winograd native pruned Winograd-ReLU
CNN (Han et al., 2015) CNN (Li et al., 2017)
CNN (ours)
Weight dimension K × C × 3 × 3
K ×C ×p×p
K ×C ×p×p
p
p
ReLU dimension H × W × K
H ×W ×K
p−2 H × p−2 W × K
Weight Dimension Increase: Compared to a vanilla 3 × 3 CNN, a conventional Winograd CNN
uses (p × p)-dimension Winograd kernels. Training a Winograd CNN from scratch allows higher
dimension (p × p) for Winograd kernels, and a Winograd-ReLU CNN shares these characteristics.
ReLU Dimension Increase: A major difference between our Winograd-ReLU CNN and conventional Winograd CNN is that the ReLU layers in Winograd-ReLU CNN have higher dimension. The
dimension increase comes from the Winograd transformation extracting p × p feature patches with
(p − 2) × (p − 2) strides from H × W activations. The total number of extracted Winograd-domain
p
p
activations is p−2
H × p−2
W , an increase from the spatial domain’s H × W .
We can see that our Winograd-ReLU architecture has an advantage on the dimensions of weights
and activations over other two models. This means Winograd-ReLU CNNs classify on a higher
dimension with more complex decision boundaries, which forms a stronger representational ability in
high dimensional image feature space.
7
Published as a conference paper at ICLR 2018
5.2
DYNAMIC ACTIVATION D ENSITY
As is shown in the ImageNet results in the previous section, dynamic activation density of spatial
baseline CNN model varies significantly among layers. Layers at earlier stages typically have higher
density in activation than later stages. In Winograd-ReLU CNN model, the dynamic activation
densities vary little among layers and are all close to 50%.
An explanation is that the nature of image convolution ensures activations d to be spatially smooth.
Thus, due to the structure of matrix B (Lavin, 2015), 15 of 16 elements in the 4 × 4 matrix of
Winograd-domain activation patch B T · d · B have a mean close to zero. This benefits classification
within a patch since ReLU layer is most powerful when half of activations are positive.
5.3
K ERNEL V ISUALIZATION
We visualize the kernels of the proposed Winograd-ReLU model. We selected the first 6 input and
output channels of layer res2a_2a of ResNet-18 at three different pruning densities. Unlike spatial
domain kernels, Winograd-ReLU kernels do not show clear physical meanings such as edge or corner
detectors. However, we observe that values of the (2, 2) elements (from top-left, 1-based indices) in
each kernel are typically distinct in a kernel and are most likely kept during aggressive pruning. A
possible reason for this is that the (2, 2) elements of Winograd-domain activation in a 4 × 4 patch
are special: interested readers can calculate B T · d · B symbolically and will realize that (2, 2)
elements are the only elements that are transformed with a linear combination of only adding and no
subtraction. In a spatially smooth activation patch, this means the (2, 2) elements are the ones and
the only ones with a non-zero mean.
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0.0
0.0
0.0
0.1
0.1
0.1
0.2
0.2
0.2
0.3
0.3
0.3
0.4
0.4
0.4
Figure 5: Kernels of ResNet-18 Winograd-ReLU model res2a_2a layer with density of 100% (left,
87.43% top-5 accuracy), 35% (middle, 87.36% top-5 accuracy) and 15% (right, 86.57% top-5
accuracy). Positive, negative and pruned weights are in red, blue and black respectively.
6
C ONCLUSION AND F UTURE W ORK
We have shown that we can combine the computational savings of sparse weights and activations with
the savings of the Winograd transform by making two modifcations to conventional CNNs. To make
the weights sparse at the point of multiplication, we train and prune the weights in the transform
domain. This simple approach does not reduce the workload with respect to spatial pruning, though,
so we move the ReLU non-linear operation after the Winograd transform to make the activations
sparse at the point of multiplication. Moving ReLU to the Winograd domain also allows the weights
to be more aggressively pruned without losing accuracy. With a 2 × 2 output patch (p = 4), the
net result is a reduction of 10.4×, 6.8× and 10.8× in computation on three datasets: CIFAR-10,
CIFAR-100 and ImageNet.
We plan to extend this work in the following directions. First, we expect that even greater savings on
computation can be realized by using larger patch sizes (e.g., p = 6), and there may be benefit in
exploring different Winograd transformation matrices (B,G and A). Second, we expect that using
different pruning rates ri for each network layer will help maintain accuracy and improve overall
workload reduction. Finally, we expect that combining our Winograd-ReLU network with other
network simplification techniques, e.g. quantization of weights and/or activations (Courbariaux et al.,
2015; Lin et al., 2016; Rastegari et al., 2016), will reduce the energy of computation even further.
8
Published as a conference paper at ICLR 2018
R EFERENCES
Martín Abadi et al. Tensorflow: A system for large-scale machine learning. In Proceedings of the
12th USENIX Conference on Operating Systems Design and Implementation (OSDI), 2016.
Jorge Albericio, Patrick Judd, Tayler Hetherington, Tor Aamodt, Natalie Enright Jerger, and Andreas
Moshovos. Cnvlutin: Ineffectual-neuron-free Deep Neural Network Computing. In Proceedings of
the 43rd International Symposium on Computer Architecture (ISCA), 2016.
Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro,
and Evan Shelhamer. cuDNN: Efficient primitives for deep learning. CoRR, abs/1410.0759, 2014.
URL http://arxiv.org/abs/1410.0759.
Jason Cong and Bingjun Xiao. Minimizing computation in convolutional neural networks. In
International Conference on Artificial Neural Networks (ICANN), pp. 281–290. Springer, 2014.
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural
networks with binary weights during propagations. In Advances in neural information processing
systems (NIPS), 2015.
Song Han, Jeff Pool, John Tran, and William J. Dally. Learning both weights and connections for
efficient neural networks. In Advances in neural information processing systems (NIPS), 2015.
Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A. Horowitz, and William J.
Dally. EIE: Efficient inference engine on compressed deep neural network. In Proceedings of the
43rd International Symposium on Computer Architecture (ISCA), 2016a.
Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks
with pruning, trained quantization and huffman coding. In International Conference on Learning
Representations (ICLR), 2016b.
Babak Hassibi, David G Stork, et al. Second order derivatives for network pruning: Optimal brain
surgeon. Advances in Neural Information Processing Systems (NIPS), pp. 164–164, 1993.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016a.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual
networks. In European Conference on Computer Vision (ECCV), pp. 630–645. Springer, 2016b.
Yuxiang Huan, Yifan Qin, Yantian You, Lirong Zheng, and Zhuo Zou. A multiplication reduction
technique with near-zero approximation for embedded learning in IoT devices. In System-on-Chip
Conference (SOCC), 29th IEEE International, pp. 102–107. IEEE, 2016.
Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (NIPS), 2012.
Andrew Lavin. Fast algorithms for convolutional neural networks. CoRR, abs/1509.09308, 2015.
URL http://arxiv.org/abs/1509.09308.
Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. In Advances in neural
information processing systems (NIPS), pp. 598–605, 1990.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Sheng R. Li, Jongsoo Park, and Ping Tak Peter Tang. Enabling sparse Winograd convolution by native
pruning. CoRR, abs/1702.08597, 2017. URL http://arxiv.org/abs/1702.08597.
Zhouhan Lin, Matthieu Courbariaux, Roland Memisevic, and Yoshua Bengio. Neural networks with
few multiplications. In Proceedings of International Conference on Learning Representations
(ICLR), 2016.
9
Published as a conference paper at ICLR 2018
Xingyu Liu, Song Han, Huizi Mao, and William J. Dally. Efficient sparse-winograd convolutional
neural networks. International Conference on Learning Representations (ICLR) Workshop, 2017.
Nagadomi. Code for kaggle-cifar10 competition. 5th place.
nagadomi/kaggle-cifar10-torch7, 2014.
https://github.com/
Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. XNOR-Net: Imagenet
classification using binary convolutional neural networks. In European Conference on Computer
Vision (ECCV), 2016.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang,
Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet
Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV),
2015.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image
recognition. In Proceedings of International Conference on Learning Representations (ICLR),
2015.
Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for
simplicity: The all convolutional net. International Conference on Learning Representations
(ICLR) Workshop, 2015.
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov,
Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In
2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9, June 2015.
Shmuel Winograd. Arithmetic complexity of computations, volume 33. Siam, 1980.
10
| 9cs.NE
|
ONLINE CIRCLE AND SPHERE PACKING
arXiv:1708.08906v1 [cs.DS] 29 Aug 2017
CARLA NEGRI LINTZMAYER, FLÁVIO KEIDI MIYAZAWA,
AND EDUARDO CANDIDO XAVIER
Abstract. In this paper we consider the Online Bin Packing Problem in three variants:
Circles in Squares, Circles in Isosceles Right Triangles, and Spheres in Cubes. The two
first ones receive an online sequence of circles (items) of different radii while the third
one receive an online sequence of spheres (items) of different radii, and they want to
pack the items into the minimum number of unit squares, isosceles right triangles of
leg length one, and unit cubes, respectively. For Online Circle Packing in Squares,
we improve the previous best-known competitive ratio for the bounded space version,
when at most a constant number of bins can be open at any given time, from 2.439
to 2.3536. For Online Circle Packing in Isosceles Right Triangles and Online Sphere
Packing in Cubes we show bounded space algorithms of asymptotic competitive ratios
2.5490 and 3.5316, respectively, as well as lower bounds of 2.1193 and 2.7707 on the
competitive ratio of any online bounded space algorithm for these two problems. We also
considered the online unbounded space variant of these three problems which admits a
small reorganization of the items inside the bin after their packing, and we present
algorithms of competitive ratios 2.3105, 2.5094, and 3.5146 for Circles in Squares, Circles
in Isosceles Right Triangles, and Spheres in Cubes, respectively.
1. Introduction
In two or three-dimensional (offline) bin packing problems, one receives a list of items
and wants to pack them into the minimum number of bins so that two items do not overlap
and each item must be totally contained in the bin. In this paper, we are interested in
problems where the items can be circles or spheres. Thus, packing a circle i of radius ri
means finding coordinates xi and yi to its center such that they respect the boundaries of
the bin and, for every other circle j of radius rj and coordinates xj and yj , (xi − xj )2 +
(yi − yj )2 ≥ (ri + rj )2 . For spheres this is similar, but we also have coordinates zi .
As applications for these problems where items can be circles or spheres, we can mention
origami design [4], crystallography, error-correcting codes, coverage of a geographical area
with cell transmitters, storage of cylindrical barrels, and packaging bottles or cans [15].
In the online variant, one item arrives at a time and must be packed as soon as it
arrives, without knowledge of further items. Also, after packing an item, it cannot be
Date: Wednesday 30th August, 2017, 00:19.
This work was supported by São Paulo Research Foundation (grants 2016/14132-4, 2015/11937-9) and
National Counsel of Technological and Scientific Development (grants 306358/2014-0, 311499/2014-7, and
425340/2016-3).
1
2
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
moved to another bin. At any moment we consider that a bin is open or closed. We can
only pack items into open bins, and once a bin is closed, it cannot be opened again. With
this, we can classify algorithms as having bounded space, when the number of open bins
can be bounded by a constant, or unbounded space, when there is no guarantee on the
number of open bins. In this paper, we do not allow items packed in a same bin to be
reorganized, except for the unbounded space algorithms, which may do this on a constant
number of items.
We consider the Online Circle Packing in Squares problem, the Online Circle Packing in
Isosceles Right Triangles problem, and the Online Sphere Packing in Cubes problem. The
two first ones receive an online sequence of circles of different radii and the objective is to
pack them into the minimum number of unit squares and isosceles right triangles of leg
length one, respectively. Similarly, the third one receives an online sequence of spheres of
different radii and the objective is to pack them into the minimum number of unit cubes.
Demaine et al. [4] showed that it is NP-hard to decide whether a given set of circles
can be packed into a rectangle, an equilateral triangle, or a unit square. Still, we are
interested in designing algorithms that can give some guarantee on the number of bins
used. We say that an online algorithm A has competitive ratio ρ if A(I) ≤ ρOPT(I) for
every instance I, where A(I) is the number of bins used by the algorithm and OPT(I) is
the number of bins of an optimal offline solution. If A(I) ≤ ρOPT(I) + c for a constant
c, then we say the algorithm has asymptotic competitive ratio ρ.
When the items of the packing problem are not circles or spheres, there are a number
of results in the literature. Friedman [7], for instance, considers packing squares into
squares, while Kamali et al. [11] consider packing equilateral triangles into squares. More
common results, though, consider packing rectangles into unit squares. We highlight the
best-known algorithms for this problem, for which we can yet consider the case when
orthogonal rotations of the items are allowed or not. For the offline version with or
without rotations, the best asymptotic approximation has factor ln(1.5) + 1 ≈ 1.405 [2].
For the online version, the best asymptotic competitive ratio is 2.5467 for the bounded
space case [5] and 2.25 for the unbounded space case [6]. For three dimensions, there
exists an asymptotic polynomial time approximation scheme for packing cubes into unit
cubes [1] and a competitive algorithm of asymptotic ratio 2.6161 for the online version [8].
A recent survey by Christensen et al. [3] presents approximation and online algorithms
for some variations of bin packing in two or three dimensions. A survey by Lodi et al. [12]
also considers two-dimensional variations, contemplating heuristics and exact algorithms.
To the best of our knowledge, competitive algorithms for the Online Circle Packing in
Squares were considered only by Hokama et al. [10], who showed a lower bound of 2.292
on the competitive ratio of any bounded space algorithm and gave one such algorithm
with asymptotic competitive ratio 2.439. We did not find results within this approach for
Online Circle Packing in Isosceles Right Triangles and Online Sphere Packing in Cubes.
ONLINE CIRCLE AND SPHERE PACKING
3
For the offline Circle Packing in Squares, Miyazawa et al. [13] showed an asymptotic
polynomial time approximation scheme when one can augment one direction of the unit
bin by a small constant. There are also several results that involve packing circles of same
radii into squares [15] or packing spheres of same radii into cubes [16]. We also refer the
reader to [9] for a review on circle and sphere packing related results.
Our contributions. We show a general idea of bounded space and unbounded space
algorithms for packing circles/spheres which works for the three problems that we are considering. The bounded space algorithms are in part based on the one given by Hokama et
al. [10], but they have a simpler analysis. Furthermore, we were able to improve the occupation ratio related to the class of small circles, we found a simpler way of subdividing a
bin, and we used another method to find the numerical results. With that, our bounded
space algorithm for Online Circle Packing in Squares has asymptotic competitive ratio
2.3536, for Online Circle Packing in Isosceles Right Triangles it has asymptotic competitive ratio 2.5490, and for Online Sphere Packing in Cubes it has asymptotic competitive
ratio 3.5316. The unbounded space algorithms are a modification of the bounded space
ones using an idea presented by Epstein [5]. Our unbounded space algorithm for Online
Circle Packing in Squares has asymptotic competitive ratio 2.3105, for Online Circle Packing in Isosceles Right Triangles it has asymptotic competitive ratio 2.5094, and for Online
Sphere Packing in Cubes it has asymptotic competitive ratio 3.5146. These algorithms
allow a small reorganization of the items after their packing inside the bin (they do not
allow items to be exchanged between bins). Lastly, we give lower bounds of 2.1193 and
2.7707 on the competitive ratio of any bounded space algorithm for Online Circle Packing
in Isosceles Right Triangles and Online Sphere Packing in Cubes, respectively. Our results
are summarized in Table 1. We highlight that the asymptotic competitive ratio given by
our bounded space algorithm for Online Circle Packing in Squares is an improvement on
a previous result while all other results we give are new.
Table 1. Summary of the results presented in this paper.
Competitive ratio
Problem
Lower bound
Bounded space
Unbounded space
Circle Packing in Squares
2.292 [10]
2.3536 (Thm. 3.3)
2.3105 (Thm. 3.4)
Circle Packing in Isosceles Right Triangle
2.1193 (Thm. 4.5)
2.5490 (Thm. 4.2)
2.5094 (Thm. 4.3)
Sphere Packing in Cubes
2.7707 (Thm. 5.5)
3.5316 (Thm. 5.2)
3.5146 (Thm. 5.3)
Section 2 describes the general idea of our algorithms. Sections 3, 4, and 5 specify the
algorithms for each of the three problems.
4
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
2. Preliminaries and general ideas
In this section, we use the word “item” to denote a circle or a sphere, according to the
problem. A “unit bin” means a unit square, an isosceles right triangle of leg length one,
or a unit cube, also according to the problem.
The general idea of the bounded and unbounded space algorithms that we will present
is the following. We classify items into large ones and small ones and pack them separately
so that we can differentiate Lbins, which are bins that pack large items, from Sbins, which
are bins that pack small items. Nevertheless, Lbins and Sbins are unit bins. We will
further classify the items and bins in types so that large items of type i are packed into
Lbins of type i and small items of type i are packed into Sbins of type i, for some integer
i. After explaining the algorithms, we will analyse their asymptotic competitive ratios
using the weighting method [5], which will be explained later. For that, the concept of
occupation ratio of a closed bin, which is the minimum total area/volume occupied by the
items packed in such bin, will be notably used.
For an integer M given as a parameter, an item of radius r is large if r > 2/M and it
is small if r ≤ 2/M. For a positive integer i ≥ 1, let ρi ∗ be the largest value of radius
such that i items of radius ρi ∗ can be packed into a unit bin. Let ρi be ρi ∗ if the latter
is known or the best-known lower bound for ρi ∗ [14, 15, 16], otherwise. See Figure 1 for
some examples.
Figure 1. Values of ρ1 , ρ2 , and ρ3 for unit squares and isosceles right
triangles of leg length one.
Let K be an integer such that ρK+1 < 2/M ≤ ρK . We classify a large item of radius r
as type i, for 1 ≤ i < K, if ρi+1 < r ≤ ρi and as type K if 2/M < r ≤ ρK . We will use Ii
to denote a large item of type i.
Let C be an integer given as a parameter. Given a small item of radius r, we find the
largest integer p such that C p r ≤ 2/M. Then we classify such item as type i and subtype
p if 2/(i + 1) < C p r ≤ 2/i, where M ≤ i < CM. For simplicity, we will say that the item
is of type (i, p).
The hexagonal packing is the densest one for circles of equal radii and the hexagonal
√
close packing is the densest one for spheres of equal radii. The first one, of density π/ 12,
can be seen as if we tessellate the two dimensional space with hexagons and place a circle
√
in each one of them. The second one, of density π/ 18, can be seen as if we tessellate the
ONLINE CIRCLE AND SPHERE PACKING
5
three dimensional space with rhombic dodecahedra and place a sphere in each one of them.
We will use these tessellations to pack small items in two or three dimensions, accordingly.
√
For that, note that a regular hexagon of side length 2r/ 3 has an inscribed circle of radius
√
at most r and a rhombic dodecahedron of side length 3r/ 6 has an inscribed sphere of
radius at most r.
Both bounded and unbounded space algorithms pack small items in the same way, so
we start by describing this part of them. We define a q-bin(i,p) as a polytope that has
the same form of the bin, but has side length 1/C p+1. For each i, with M ≤ i < CM, at
most one Sbin of type i is kept open at a time to pack small circles of type (i, p), for any
p ≥ 0. When an Sbin of type i is opened, it is divided into C d q-bins(i, 0), where d is the
dimension of the space. See Figure 2.
Figure 2. Division of an Sbin of type i into q-bins(i, 0) for C = 4.
√
√
Let ℓ = 4/(C p i 3) if the items are circles, let ℓ = 6/(C p i 6) if they are spheres, and
let d be the dimension of the space. We will use the word “cell” to generally denote a
hexagon or a rhombic dodecahedron of a tessellation. A q-bin(i, p) is either subdivided
into C d q-bins(i, p + 1) or it is tessellated with cells of side ℓ to pack small items of type
(i, p). Note that we must have 1/C p+1 ≥ 2/(iC p ) for this to work, but since i < CM, it
suffices to choose a M > 2. Also, note that ℓ is chosen precisely so that the cell can pack
an item of type (i, p). Figure 3 shows the measurements of the cells.
Figure 3. Measurements of the cells used to tessellate a q-bin(i, p).
For two dimensions, the tessellation of the q-bin(i, p) is done by placing a hexagon on
the left bottom corner of the bin with two of its sides parallel to the bottom of the bin.
For three dimensions, the tessellation is done by placing a rhombic dodecahedron in the
left bottom corner of the bin with four of its faces, each one sharing two nonadjacent
vertices with two of the other, parallel to the sides of the bin. See Figure 4. Note that
hexagons and rhombic dodecahedra are space-filling polytopes.
6
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
Figure 4. Tesselation of a q-bin(i, p).
When a small item of type (i, p) arrives, the algorithms will simply try to pack it into
a cell of a q-bin(i, p). If there is no such cell, they will subdivide a larger q-bin(i, p′ ), i.e.,
find an empty q-bin(i, p′ ) with the largest p′ such that p′ < p and subdivide it, until a
q-bin(i, p) is found. This is formally described below:
if there is no tessellated q-bin(i, p) with empty cells then
if there is no empty q-bin(i, p) or empty q-bin(i, p′ ) with p′ < p then
close the current Sbin of type i (if any) and open a new one;
while there is no empty q-bin(i, p) do
let p′ be the largest integer < p such that there is an empty q-bin(i, p′ );
subdivide one empty q-bin(i, p′ ) in C d q-bins(i, p′ + 1);
tessellate an empty q-bin(i, p);
pack the item in one empty cell of a tessellated q-bin(i, p).
Now we consider packing large circles in the bounded space algorithms. For each i, with
1 ≤ i ≤ K, at most one Lbin of type i is kept open at a time to pack large items of type
i. When an Lbin of type i is opened, it is divided into i c-bins, which are circles/spheres
of radii ρi . See Figure 5.
When an Ii arrives, it is either packed into an empty c-bin or the current Lbin of type
i (if any) is closed and a new one is opened. Note that at most K Lbins are kept open by
the algorithm at any given time. Together with the (C − 1)M Sbins, we can see that this
algorithm has bounded space.
Packing large circles in the unbounded space algorithms follows an idea given by Epstein [5], which uses waiting bins. A waiting bin packs an I1 with other items, instead
ONLINE CIRCLE AND SPHERE PACKING
7
Figure 5. Three first types of Lbins for unit squares and isosceles right
triangles of leg length one.
of packing it alone, if their radii are related with a given parameter D. All large items
which are not packed in waiting bins are packed as in the bounded space algorithms. As
we mentioned, the unbounded space algorithms allow some reorganization of the items
after their packing. We highlight here that this happens only for items I1 that are inside
open waiting bins. Since the specific items and possible configurations of waiting bins
depend on the format of the bin, we will better describe such algorithms in each of the
following sections. Also, we cannot guarantee how many waiting bins are kept open by
the algorithms at any given time, which is why they have unbounded space.
Each of the next sections analyses the asymptotic competitive ratios of the algorithms
considering Online Circle Packing in Squares, Online Circle Packing in Isosceles Right
Triangles, and Online Sphere Packing into Cubes. We specify parameters M, C, and D
for each problem as well as give more details on the unbounded space algorithms. We
also present lower bounds on the competitive ratios of any bounded space algorithm for
Online Circle Packing in Isosceles Right Triangles and Online Sphere Packing into Cubes.
As we mentioned, the analysis of the asymptotic competitive ratio for the bounded
space algorithms uses the weighting method [5], in which we first have to show a weighting
function w over the items such that the sum of weights of items in any bin generated by
the algorithms is at least 1, except for a constant number of bins. So, if B is the set of
P
items in a closed bin, w(B) = i∈B w(i) ≥ 1. Afterwards, we find every possible feasible
configuration, i.e., sets of items that can be packed into a bin, calculate their sum of
weights, and find the supremum β of such sum of weights. As a result, the asymptotic
competitive ratio of the algorithm is bounded from above by β because of the following.
If S is the total number of bins used by the algorithm, X is the number of closed bins at
the end, and O is the maximum number of open bins at the end, then S − O = X ≤ w(I),
where I is the set of items of the input. If OPT(I) is the number of bins used by an
optimal offline solution, then w(I)/β ≤ OPT(I), since every feasible configuration has
weight at most β. Thus, we have S ≤ w(I) + O ≤ βOPT(I) + O.
The analysis of the asymptotic competitive ratio for the unbounded space algorithms
is made using a generalized form of the weighting method [5], in which we have to show
weighting functions w1 and w2 such that the sum of weights of items in any bin generated
by the algorithms is at least 1 on average for at least one of the functions, except perhaps
8
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
for a constant number of bins. Again, the supremum β of the sums of weights of feasible
configurations, considering both weight functions, is an upper bound for the asymptotic
competitive ratio.
Note that both analysis depend on finding the supremum of the sum of weights of
feasible configurations. However, it is not reasonable to list every possible such set, so
we use the following fact. Suppose that a bin of maximum weight has some known-items
I ′ ⊆ I whose sum of weights is W and whose sum of areas/volumes is V , and let OR be a
lower bound on the occupation ratio of the unknown items of the bin, i.e., items in I\I ′ .
The weighting functions used in the next sections are such that the weight given to any
item i not in I ′ of area/volume v(i) is w(i) ≤ v(i)/OR. Thus, the total sum of weights in
such bin is at most W + (S − V )/OR, where S is the area/volume of the bin.
This means that we can concentrate on finding a few configurations and the lower bound
on the occupation ratio of items that are not on them, in order to find an upper bound
on the asymptotic competitive ratio of our algorithms. For that, we use a two-phase
program. The first phase uses constraint programming to test if a given set of items is a
feasible or an infeasible configuration. Given a set of items I = {i1 , i2 , . . . , in }, let I ′ =
{i′1 , i′2 , . . . , i′n } be a set of items such that r(i′j ) = µr(ij ) for some 0 < µ < 1, where r(i)
is the radius of item i. We show in Theorems 2.1, 2.2, and 2.3 that the packing of I ′ and
I are related for each of the three problems we are considering.
Theorem 2.1. Consider the Online Sphere Packing in Cubes problem. Let S = {s1 , s2 ,
. . . , sn } be a given set of spheres and S ′ = {s′1 , s′2 , . . . , s′n } be a set of spheres such that
2
r(s′i ) = µr(si ) for some 0 < µ < 1. Let rmin = min{r(si ) : si ∈ S} and δ ≤ (1−µ 3)rmin . If
S ′ cannot be packed in a bin using a regular grid of granularity δ, then S cannot be packed
in a bin with general positions.
Proof. We will prove the contrapositive. Suppose S can be packed in general positions
of the bin and let ((x1 , y1 , z1 ), (x2 , y2 , z2 ), . . . , (xn , yn , zn )) be one of its packings. Let
((x′1 , y1′ , z1′ ), (x′2 , y2′ , z2′ ), . . . , (x′n , yn′ , zn′ )) be a rounding of the previous packing to the
nearest coordinates of the regular grid of granularity δ. We will show that ((x′1 , y1′ , z1′ ),
(x′2 , y2′ , z2′ ), . . . , (x′n , yn′ , zn′ )) is a packing of S ′ . Note that, for every pair si and sj of
spheres, |x′i − x′j | ≥ |xi − xj | − δ, |yi′ − yj′ | ≥ |yi − yj | − δ, and |zi′ − zj′ | ≥ |zi − zj | − δ,
because every coordinate has changed by at most δ/2.
Let d be the distance between two spheres si and sj . Thus, d2 = (xi − xj )2 + (yi −
yj )2 + (zi − zj )2 , d ≥ r(si ) + r(sj ) ≥ 2rmin , |xi − xj | ≤ d, |yi − yj | ≤ d, and |zi − zj | ≤ d.
There are four main cases. First, consider that |xi − xj | ≥ δ, |yi − yj | ≥ δ, and
|zi − zj | ≥ δ. We have
|x′i − x′j |2 + |yi′ − yj′ |2 + |zi′ − zj′ |2 ≥ (|xi − xj | − δ)2 + (|yi − yj | − δ)2 + (|zi − zj | − δ)2
= d2 − 2δ(|xi − xj | + |yi − yj | + |zi − zj |) + 3δ 2 ≥ d2 − 6δd + 3δ 2 ≥ d2 − 6δd
ONLINE CIRCLE AND SPHERE PACKING
2
≥d −6
(1 − µ2 )rmin
3
9
d = d2 − 2(1 − µ2 )rmin d
≥ d2 − (1 − µ2 )d2 = µ2 d2 ≥ µ2 (r(si ) + r(sj ))2 = (r(s′i ) + r(s′j ))2 ,
and thus s′i and s′j do not intersect in this case.
Now consider, without loss of generality, that |xi − xj | < δ, but |yi − yj | ≥ δ and
|zi − zj | ≥ δ. We have
|x′i − x′j |2 + |yi′ − yj′ |2 + |zi′ − zj′ |2 ≥ (|yi − yj | − δ)2 + (|zi − zj | − δ)2
= d2 − |xi − xj |2 − 2δ(|yi − yj | + |zi − zj |) + 2δ 2 ≥ d2 − δ 2 − 4δd + 2δ 2 ≥ d2 − 4δd
(1 − µ2 )rmin
2
≥d −4
d > d2 − 2(1 − µ2 )rmin d
3
≥ d2 − (1 − µ2 )d2 = µ2 d2 ≥ µ2 (r(si ) + r(sj ))2 = (r(s′i ) + r(s′j ))2 ,
and thus s′i and s′j do not intersect in this case.
For the third case, consider, without loss of generality, that |xi −xj | < δ and |yi −yj | < δ,
but |zi − zj | ≥ δ. We have
|x′i − x′j |2 + |yi′ − yj′ |2 + |zi′ − zj′ |2 ≥ (|zi − zj | − δ)2
= d2 − |xi − xj |2 − |yi − yj |2 − 2δ|zi − zj | + δ 2 ≥ d2 − 2δ 2 − 2δd + δ 2 ≥ d2 − δ 2 − δd
2
(1 − µ2 )2 rmin
(1 − µ2 )d2 (1 − µ2 )d2
(1 − µ2 )rmin
2
≥d −
d ≥ d2 −
−2
−
9
3
36
3
13
= d2 − (1 − µ2 )d2 > d2 − (1 − µ2 )d2 = µ2 d2 ≥ µ2 (r(si ) + r(sj ))2 = (r(s′i ) + r(s′j ))2 ,
36
′
and thus si and s′j do not intersect in this case either.
For the last case, consider that |xi − xj | < δ, |yi − yj | < δ, and |zi − zj | < δ. We have
2
2
(1 − µ2 )2 rmin
(1 − µ2 )rmin
2
2
2
2
2
|xi − xj | + |yi − yj | + |zi − zj | < 3δ ≤ 3
≤
≤ rmin
.
9
3
On the other hand,
|xi − xj |2 + |yi − yj |2 + |zi − zj |2
=
d2
≥
(ri + rj )2
≥
2
4rmin
,
so we reach a contradiction, which means this case cannot happen.
Now, it remains to show that every sphere s′i does not intercept the borders of the bin.
2
Since (1−µ 3)rmin ≤ 2(1 − µ)rmin for µ ≤ 1, we have that δ ≤ 2(1 − µ)rmin . Thus,
x′i ≤ xi +
δ
δ
≤ 1 − r(si ) + ≤ 1 − r(si ) + (1 − µ)rmin ≤ 1 − r(si ) + (1 − µ)r(si ) = 1 − µr(si ) = 1 − r(s′i )
2
2
and
x′i ≥ xi −
δ
δ
≥ r(si ) − ≥ r(si ) − (1−µ)rmin ≥ r(si ) − (1−µ)r(si ) = µr(si ) = r(s′i ) .
2
2
The same is valid for coordinates y and z.
10
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
Therefore, ((x′1 , y1′ , z1′ ), (x′2 , y2′ , z2′ ), . . . , (x′n , yn′ , zn′ )) is a packing of S ′ .
Theorem 2.2. Consider the Online Circle Packing in Squares problem. Let C = {c1 , c2 ,
. . . , cn } be a given set of circles and C ′ = {c′1 , c′2 , . . . , c′n } be a set of circles such that
2
r(c′i ) = µr(ci ) for some 0 < µ < 1. Let rmin = min{r(ci ) : ci ∈ C} and δ ≤ (1−µ 2)rmin . If
C ′ cannot be packed in a bin using a regular grid of granularity δ, then C cannot be packed
in a bin with general positions.
Proof. Similar to the proof of Theorem 2.1.
Theorem 2.3. Consider the Online Circle Packing in Isosceles Right Triangles problem.
Let C = {c1 , c2 , . . . , cn } be a given set of circles and C ′ = {c′1 , c′2 , . . . , c′n } be a set of
circles such that r(c′i ) = µr(ci ) for some 0 < µ < 1. Let rmin = min{r(ci ) : ci ∈ C} and δ
2
≤ (1−µ 2)rmin . If C ′ cannot be packed in a bin using a regular grid of granularity δ, then C
cannot be packed in a bin with general positions.
Proof. Similar to the proof of Theorem 2.1.
Consider the notation given in the previous theorems. In the first phase, to test for
the feasibility of a packing, we rescale the instance such that the size of the bin used
in the test is W = ⌈1/δ⌉ and the radius of each item is also multiplied by W . As we
use an integer solver, we have to assure that all grid positions are integers. So, we take
r(i′j ) = ⌊r(ij )µW ⌋ and we test if I ′ can be packed into a bin of side length W . If not,
then I does not have a feasible packing into the bin of side length 1; otherwise there is no
guarantee of infeasibility. Also, we take r(i′j ) = ⌈r(ij )W ⌉ and we test if I ′ can be packed
into the bin of side length W . If it can, then I can also be packed into the bin of side
length 1; otherwise there is no guarantee of feasibility.
In the first phase we find a set F of feasible configurations that contain large items.
Each configuration I ∈ F is associated with an integer f (I) which indicates that all large
items from type 1 to f (I) were tested to be part of I. For example, if I = (1, 4, 4, 5) and
f (I) = 14, then we know that a configuration with one large item of type 1, two large
items of type 4, and one large item of type 5 is feasible, and that any large item of type
up to 14 makes I infeasible if added to it. Note that it may be the case that large items
of type greater than 14 do not make I infeasible (the first phase was simply not able to
decide it). In the second phase the idea is to add items of type greater than f (I) in the
remaining space of the bin by using a criterion of space: if the item’s area/volume is at
most the remaining area/volume of the bin, such item will be considered. We may create
an infeasible configuration at this point, however this is not an obstacle since our goal is to
find a configuration with maximum total weight. Specifically, in the second phase we have
an integer programming which simulates a knapsack problem, described next. Let I ∈ F
and let I ′ be the set we want to build in the second phase. Let OR be the occupation
ratio of bins containing small items, let OR(i) be the occupation ratio of Lbins of type i,
ONLINE CIRCLE AND SPHERE PACKING
11
and let k be an integer such that OR(i) > OR for all i > k. Our program will maximize
P
P
1≤i≤k w(i)x(i) + (V −
1≤i≤k v(i)x(i))/OR, where w(i) is the weight of large item i,
v(i) is the area/volume of item i (according to the dimension), x(i) is a binary variable
which indicates if item i is in I ′ or not, and V is the area/volume of the bin (according to
the dimension). This program is subject to x(i) = 1 for all i ∈ I, x(i) = 0 for all i ≤ f (I)
P
such that i ∈
/ I, and 1≤i≤k v(i)x(i) ≤ V .
This two-phase program is used in the following sections during the analysis of each
algorithm.
3. Online Circle Packing in Squares
For the Online Circle Packing in Squares problem, optimal values of ρi are known for i ≤
30, and good lower bounds are known for all i up to 9996 [14]. Since ρ9996 < 0.005076143
and ρi < ρ9996 for i < 9996, we chose M = 360, which makes 2/M ≤ 0.005555556. This
means that we can find K such that ρK+1 < 2/M ≤ ρK , and so we can classify all large
items. Recall that Ii denotes a large circle of type i.
We start by showing the occupation ratio of a closed Sbin, which is a result valid for
both bounded space and unbounded space algorithms. Parameter C is chosen to be 5 in
order to maximize this ratio, as Theorem 3.2 shows.
Theorem 3.1. The occupation ratio of a closed Sbin of type i, for M ≤ i < CM, is at
π M2
√
+ 18.48
.
least 1 − C 21−1 − C12 1 − 8.62C
M
M2
12 (M +1)2
Proof. Note that the area loss in a closed Sbin is due to (i) packing circles into hexagons,
(ii) tessellating a q-bin(i, p) with hexagons, and (iii) non-full q-bins(i, p).
Consider (i). We pack one circle of type (i, p) into an hexagon of side length C p4i√3 . Thus,
2
√
2
4√
3 3
2
and
the
hexagon
has
area
area
.
the circle has area at least π (i+1)C
p
2
Cpi 3
2
2
i
M
√π
because
This gives an occupation ratio of at least √π12 (i+1)
2 , which is at least
12 (M +1)2
2
2
i > M and x /(x + 1) is increasing for x ≥ 0.
Consider (ii). When we tessellate a q-bin(i, p), we lose a strip of height at most
once the height of the hexagon at the top border plus a strip of width at most twice
the side of the hexagon at the right border, as depicted in Figure 6. Also, note that
these strips intersect, so the length of one is 1/C p+1 but the length of the other is
not. See Figure
3 for the measurements
of the hexagon.
loss is at
Thus, the
area
8 1
8.62
4
1
8
1
8
32
1
−
most C p+1 C p i√3 + C p+1 − C p i√3 C p i = C 2p+1 i √3 + 4 − C 2p i2 √3 < C 2p+1
M
18.48
. If such
C 2p+2 M 2
1
8.62
− C 2p+1
C 2p+2
M
2
tessellated q-bin(i, p) is full, then its occupation ratio is at least √π12 (MM+1)2
2
18.48
1
+ C 2p+2
+ 18.48
/ C 2p+2
= √π12 (MM+1)2 1 − 8.62C
.
M2
M
M2
Now, consider (iii). When an Sbin closes, there can be only one q-bin(i, p) for each
p ≥ 0 which is tessellated but not full. Also, no q-bin(i, 0) can be empty, but at most
12
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
Figure 6. For the Online Circle Packing in Squares, the darker hexagons
are lost due to tiling one q-bin(i, p) into hexagons. The borders of the qbin(i, p) are the thick lines. The area between the dashed lines and the
border is the one being counted as lost. Note that it incorporates small
pieces of hexagons lost in the left and bottom borders.
C 2 − 1 q-bins(i, p) for each p ≥ 1 can. Thus, the area loss due to non-full q-bins(i, p) is
P
P
2 −1
1
at most p≥0 ( C 2p+2
) + p≥1 ( CC2p+2
) = C 21−1 + C12 .
Putting everything together, we have that the occupation ratio of the closed Sbin is at
18.48 √π
M2
+
.
least 1 − C 21−1 − C12 1 − 8.62C
2
M
M
12 (M +1)2
Theorem 3.2. The occupation ratio of a closed Sbin of type i, for M ≤ i < 5M, is at
π
M2
551
√
least 600
+ 18.48
.
1 − 43.1
M
M2
12 (M +1)2
Proof. Given the result of Theorem 3.1 and using M = 360, we found that the maximum
π M2
√
+ 18.48
, approximately 0.729189, happens
value of 1 − C 21−1 − C12 1 − 8.62C
M
M2
12 (M +1)2
when C = 5.
Regarding large items, for the bounded space algorithm, note that each closed Lbin of
type i packs i circles of type i, which have radii at least ρi+1 . Its occupation ratio is at
least iπρ2i+1 .
Now we give the weighting function w which will be used for the analysis of the bounded
space algorithm. We define w(Ii ) = 1i , so a closed Lbin of type i (it has i items) has total
P
551
18.48 √π
M2
1 − 43.1
+
. For a small circle c of type
weight ij=1 1i = 1. Let OR = 600
1
M
M
12 (M +1)2
(i, p) and area a(c), we define w(c) = a(c)
, so a closed Sbin B of type i has total weight
OR
P
a(c)
1
c∈B OR = OR a(B) ≥ 1, where a(B) is the sum of areas of items in B and, according to
Theorem 3.2, a(B) ≥ OR.
Theorem 3.3 concludes with the asymptotic competitive ratio of the bounded space
algorithm.
Theorem 3.3. The algorithm for packing circles in squares with bounded space has asymptotic competitive ratio strictly below 2.3536.
ONLINE CIRCLE AND SPHERE PACKING
13
i
iπρ2i+1
i
iπρ2i+1
i
iπρ2i+1
i
iπρ2i+1
1 0.269506 38 0.790380 75 0.798059 112 0.819406
2 0.406430 39 0.768280 76 0.799114 113 0.823647
3 0.589049 40 0.773389 77 0.805433 114 0.822212
4 0.539012 41 0.779668 78 0.810417 115 0.824644
5 0.553297 42 0.768956 79 0.816878 116 0.827210
6 0.573695 43 0.775801 80 0.812840 117 0.831850
7 0.639593 44 0.771901 81 0.812682 118 0.837983
8 0.698132 45 0.779857 82 0.811604 119 0.844561
9 0.621032 46 0.772500 83 0.813538 120 0.828951
10 0.637038 47 0.774662 84 0.818145 121 0.826427
11 0.676929 48 0.775070 85 0.824701 122 0.823354
12 0.676860 49 0.784267 86 0.808254 123 0.821206
13 0.683131 50 0.792801 87 0.809045 124 0.823527
14 0.711252 51 0.806713 88 0.804577 125 0.824775
15 0.736311 52 0.799272 89 0.807785 126 0.825403
16 0.690400 53 0.784604 90 0.809183 127 0.825023
17 0.712728 54 0.785721 91 0.812688 128 0.821568
18 0.712713 55 0.788024 92 0.809623 129 0.822238
19 0.740519 56 0.789861 93 0.814375 130 0.822885
20 0.717484 57 0.785549 94 0.811480 131 0.824018
21 0.736604 58 0.789094 95 0.815584 132 0.825242
22 0.730430 59 0.783854 96 0.819112 133 0.829802
23 0.742673 60 0.788237 97 0.825048 134 0.831057
24 0.753982 61 0.791144 98 0.831812 135 0.834326
25 0.729297 62 0.798101 99 0.821731 136 0.838223
26 0.743707 63 0.797034 100 0.819641 137 0.834469
27 0.744288 64 0.803190 101 0.818015 138 0.837264
28 0.752047 65 0.806944 102 0.818325 139 0.837932
29 0.765618 66 0.812209 103 0.820286 140 0.838330
30 0.752223 67 0.822742 104 0.824771 141 0.841156
31 0.751754 68 0.804939 105 0.825170 142 0.845329
32 0.764948 69 0.795970 106 0.816158 143 0.830805
33 0.753806 70 0.794231 107 0.815139 144 0.831470
34 0.758906 71 0.797673 108 0.816638 145 0.828671
35 0.763582 72 0.797184 109 0.814799 146 0.826412
36 0.762132 73 0.800550 110 0.817711 147 0.827501
37 0.776068 74 0.795449 111 0.818423 148 0.830048
Table 2. For circles in square, these are the first 148 values of i and the
respective lower bound on the occupation ratio associated with Ii .
Proof. First note that, since M = 360, we have OR ≥ 0.729189. We show some values
of iπρ2i+1 in Table 2, from where we can note that for i ≥ 21 we have iπρ2i+1 > 0.729189.
Thus, in the description of the two-phase program mentioned in the end of Section 2,
the second phase has k = 20, i.e., items of type up to 20 will be used to augment each
configuration found during phase one.
14
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
Now consider a bin of maximum weight. If it does not have any I1 or I2 , phase two of the
program finds a configuration with four I4 and four I5 . Thus, the asymptotic competitive
ratio is given by the sum of the weights of such circles (4 14 + 4 51 ) plus the remaining area
(1 − 4πρ25 − 4πρ26 ) divided by OR, and this gives a result which is less than 1.825165. Now
we need only to consider the configurations when these two types of circles are present.
In the two-phase program, we tested all possible variations of configurations with I1
and I2 , and we found the following ones:
(1) I1 , I2 , two I4 ;
(2) I1 , I3 , two I4 ;
(3) I1 , three I4 , I12 ;
(4) two I2 , I3 , I5 , two I6 ,
I7 ;
(5) two I2 , two I4 , nine
I20 .
For example, during phase one we find that two I2 , one I3 , and one I5 can be packed into
a square. For such case, phase one is not able to say if adding an I6 makes it infeasible or
not, so phase two tries do add circles from type 6 to type 20 on it, forming configuration 4.
One can easily notice that removing items from the configurations above keeps them
feasible (for instance a configuration with just one I1 ), will give lower values of competitive
ratios and this is why they are not considered in these calculations.
In the following, we calculate the total sum of weights in each of these configurations,
considering the formula W + (1 − V )/OR, as used by phase two of the program:
(1)
(2)
(3)
(4)
(5)
1 + 12 + 2 14 + (1 − πρ22 − πρ23 − 2πρ25 )/OR < 2.353507;
1 + 31 + 2 14 + (1 − πρ22 − πρ24 − 2πρ25 )/OR < 2.196255;
1
+ +(1 − πρ22 − 3πρ25 − πρ213 )/OR < 2.203375;
1 + 3 14 + 12
2 21 + 31 + 51 + 2 61 + 17 + (1 − 2πρ23 − πρ24 − πρ26 − 2πρ27 − πρ28 )/OR < 2.014954;
1
+ (1 − 2πρ23 − 2πρ25 − 9πρ221 )/OR < 1.951641.
2 21 + 2 41 + 9 20
Thus, the highest value we achieve is at most 2.353507.
Now we analyse the unbounded space algorithm. We will use waiting bins that are
supposed to pack either one I1 with one I2 or one I1 with two I4 , as depicted in Figure 7.
All other circles are packed as in the bounded space algorithm.
Consider the notation of Figure 7. In the left drawing, the circle centered at point
F is an I1 while the circle centered at point J is an I2 . We fix that the radius of the
I1 is D and we want to find what is the maximum radius x that the I2 can have as a
√
function of D. Note that, since ∠EGF = 90◦ and ∠GEF = 45◦ , we have EF = D 2.
√
√
√
√
Similarly, JK = x 2. Since EF + V J + JK = D 2 + D + x + x 2 = 2, we must have
√
2
x = √2+1
− D.
For the right drawing of Figure 7, the circle centered at point L is an I1 while the circle
centered at point K is an I4 . Again, we fix that the radius of the I1 is D and we want to
find what is the maximum radius x that the I4 can have as a function of D. Note that
the triangle LZK is rectangle, GL = 21 − x because GE = F K = x, LK = D + x, and
ONLINE CIRCLE AND SPHERE PACKING
15
Figure 7. For the Online Circle Packing in Squares, a waiting bin either
places the I1 at corner E and the I2 at corner K (left drawing) or it places
the I1 centered to the left with one I4 in corner J and the other in the right
bottom corner (right drawing).
KG = 1−(D +x) because T E = D and F J = x. From (D +x)2 = (1−(D +x))2 +( 21 −x)2
√
we have that x = 32 − 2D + 1.
√
2
Now note that we need to have ρ2 < D ≤ ρ1 , ρ3 < √2+1
− D ≤ ρ2 , and ρ5 <
√
3
− 2D + 1 ≤ ρ4 so that the circles are of the desired types. This is true for ρ2 < D <
2
0.331553. We fixed D = 0.325309.
Finally, the algorithm works as follows. If an I1 with radius r > D arrives, then we
pack it as in the bounded space algorithm (one per bin). Otherwise, r ≤ D and we pack it
in an already opened waiting bin containing either one I2 or at least one I4 , if they exist,
or we open a new one, pack the I1 there, and let it open waiting for an I2 or two I4 . Note
that this last case, in which we do not know if the next item that will fill the waiting bin
that contains one I1 is one I2 or one I4 , is the reason why we allow the reorganization of
items I1 inside waiting bins.
√ √
Let γ = 2/( 2 + 1) − D. If an I2 with radius r > γ arrives, then we pack it as in the
bounded space algorithm (two per bin). When r ≤ γ, the circles are labeled according to
their arrival so that the following steps can be repeated at every sequence of 72 of them.
If the I2 is among the first 70 of the sequence, then it is packed as in the bounded space
algorithm (two per bin); if it is one of the last 2 of the sequence, then it is packed in a
waiting bin: either there is one with an I1 or we open a new one, pack the circle there,
and let it open waiting for an I1 .
√
Let λ = 3/2 − 2D + 1. If an I4 with radius r > λ arrives, then we pack it as in the
bounded space algorithm (four per bin). When r ≤ λ, the circles are labeled according to
their arrival so that the following steps are repeated at every sequence of 34 such circles.
If the I4 is among the first 32 of the sequence, then it is packed as in the bounded space
algorithm (four per bin); if it is one of the last 2, then it is packed in a waiting bin.
We will use two weighting functions w1 and w2 for the analysis. When the algorithm
ends we have two possibilities: there are open waiting bins with I1 , in which case we apply
w1 over all circles, or there are not, in which case we apply w2 . Both w1 and w2 differ
16
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
from w only regarding I1 , I2 , and I4 if their radii are at most D, γ, and λ, respectively.
If I1 has radius r ≤ D, then w1 (I1 ) = 1 and w2 (I1 ) = 0. If I2 has radius r ≤ γ, then
35
37
8
9
w1 (I2 ) = 72
and w2 (I2 ) = 72
. If I4 has radius r ≤ λ, then w1 (I4 ) = 34
and w2 (I4 ) = 34
.
Note that, for both functions, the sum of weights in any bin is at least 1 if we do not
consider I1 , I2 , and I4 , because this is true for w. Now consider that w1 was applied
over such items. All waiting bins have total weight at least 1 because they contain
an I1 . For every set of 72 I2 , we have 70 of them packed into 35 bins and the last 2
packed into 2 waiting bins with one I1 each. Thus, the average weight of these bins is
35
+ 2 1 + 35
/37 = 1. This is similar for I4 and w2 .
35 2 72
72
Theorem 3.4 concludes with the asymptotic competitive ratio of our algorithm.
Theorem 3.4. The algorithm for packing circles in squares with unbounded space has
asymptotic competitive ratio strictly below 2.3105.
Proof. First consider w1 was applied (this means there are open waiting bins with one I1
each). An I1 always has weight 1 and we consider its area at least πρ22 ; an I2 either has
and area at least πρ23 ; and an I4 either
weight 21 and area at least πγ 2 or it has weight 35
72
8
has weight 41 and area at least πλ2 or it has weight 34
and area at least πρ25 .
If w2 was applied, an I1 either has weight 1 and area at least πD 2 or it has weight 0
and area at least πρ32 ; an I2 either has weight 21 and area at least πγ 2 or it has weight 37
72
and area at least πρ23 ; and an I4 either has weight 41 and area at least πλ2 or it has weight
9
and area at least πρ25 .
34
Now consider a bin of maximum weight. We again consider all configurations presented
in the proof of Theorem 3.3 and recalculate the total sums of weights, but now with all
possible combinations of weights and areas, as mentioned above. The last configuration
is the one without circles of type 1 or type 2:
(1) 1 + 21 + 2 14 + (1 − πρ22 − πλ2 − 2πγ 2 )/OR < 2.310294,
35
+ 2 14 + (1 − πρ22 − πρ23 − 2πγ 2 )/OR < 2.310033,
1 + 72
8
+ (1 − πρ22 − πλ2 − 2πρ25 )/OR < 2.310467,
1 + 21 + 2 34
8
35
+ 2 34
+ (1 − πρ22 − πρ23 − 2πρ25 )/OR < 2.310206,
1 + 72
1 + 21 + 2 14 + (1 − πD 2 − πλ2 − 2πγ 2 )/OR < 2.223957,
37
+ 2 14 + (1 − πD 2 − πρ23 − 2πγ 2 )/OR < 2.251474,
1 + 72
9
+ (1 − πD 2 − πλ2 − 2πρ25 )/OR < 2.282954,
1 + 21 + 2 34
9
37
+ 2 34
+ (1 − πD 2 − πρ23 − 2πρ25 )/OR < 2.310470;
1 + 72
(2) 1 + 31 + 2 14 + (1 − πρ22 − πρ24 − 2πγ 2 )/OR < 2.166671,
8
+ (1 − πρ22 − πρ24 − 2πρ25 )/OR < 2.166843,
1 + 31 + 2 34
1 + 31 + 2 14 + (1 − πD 2 − πρ24 − 2πγ 2 )/OR < 2.080334,
9
1 + 31 + 2 34
+ (1 − πD 2 − πρ24 − 2πρ25 )/OR < 2.139330;
1
+ (1 − πρ22 − 3πγ 2 − πρ213 )/OR < 2.158998,
(3) 1 + 3 14 + 12
8
1
1 + 3 34
+ 12
+ (1 − πρ22 − 3πρ25 − πρ213 )/OR < 2.159257,
ONLINE CIRCLE AND SPHERE PACKING
17
1
1 + 3 14 + 12
+ (1 − πD 2 − 3πγ 2 − πρ213 )/OR < 2.072661,
1
9
+ 12
+ (1 − πD 2 − 3πρ25 − πρ213 )/OR < 2.161156;
1 + 3 34
(4) 2 21 + 31 + 51 + 2 61 + 17 + (1 − 2πλ2 − πρ24 − πρ26 − 2πρ27 − πρ28 )/OR < 1.987698,
35
+ 13 + 51 + 2 16 + 71 + (1 − 2πρ23 − πρ24 − πρ26 − 2πρ27 − πρ28 )/OR < 1.987176,
2 72
37
+ 13 + 51 + 2 16 + 71 + (1 − 2πρ23 − πρ24 − πρ26 − 2πρ27 − πρ28 )/OR < 2.042731;
2 72
1
+ (1 − 2πλ2 − 2πγ 2 − 9πρ221 )/OR < 1.894801,
(5) 2 21 + 2 41 + 9 20
35
1
2 72
+ 2 14 + 9 20
+ (1 − 2πρ23 − 2πγ 2 − 9πρ221 )/OR < 1.894279,
8
1
+ 9 20
+ (1 − 2πλ2 − 2πρ25 − 9πρ221 )/OR < 1.894974,
2 21 + 2 34
8
1
35
+ 2 34
+ 9 20
+ (1 − 2πρ23 − 2πρ25 − 9πρ221 )/OR < 1.894452,
2 72
1
37
+ 2 14 + 9 20
+ (1 − 2πρ23 − 2πγ 2 − 9πρ221 )/OR < 1.949835,
2 72
9
1
+ 9 20
+ (1 − 2πλ2 − 2πρ25 − 9πρ221 )/OR < 1.953797,
2 21 + 2 34
37
9
1
2 72
+ 2 34
+ 9 20
+ (1 − 2πρ23 − 2πρ25 − 9πρ221 )/OR < 2.008831;
(6) 4 41 + 4 51 + (1 − 4πγ 2 − 4πρ26 )/OR < 1.765996,
8
+ 4 15 + (1 − 4πρ25 − 4πρ26 )/OR < 1.766342,
4 34
9
+ 4 15 + (1 − 4πρ25 − 4πρ26 )/OR < 1.883989.
4 34
And the result follows.
4. Online Circle Packing in Isosceles Right Triangles
For the Online Circle Packing in Isosceles Right Triangles problem, optimal values and
good lower bounds of ρi are known for i ≤ 299 [14]. Since ρ299 < 0.0211636617 and
ρi < ρ299 for i < 299, we chose M = 93, which makes 2/M ≤ 0.021505376. Recall that Ii
denotes a large circle of type i.
We also start by showing the occupation ratio of a closed Sbin in Theorem 4.1, which
is a result valid for both bounded space and unbounded space algorithms. Parameter C
is chosen as 3 in order to maximize this ratio.
Theorem 4.1. The occupation ratio of a closed Sbin of type i, for M ≤ i < 3M, is at
π
M2
55
√
1 − 25.86
+ 32.17
.
least 144
M
M2
12 (M +1)2
Proof. Note that the area loss in a closed Sbin is due to (i) packing circles into hexagons,
(ii) tessellating a q-bin(i, p) with hexagons, and (iii) non-full q-bins(i, p).
2
2
Consider (i). Packing one circle of type (i, p) of area at least π (i+1)3
into an
p
2
√
i2
hexagon of area 3 2 3 3p i4√3 gives an occupation ratio of at least √π12 (i+1)
2 , which is at
2
least √π12 (MM+1)2 .
For (ii), consider Figure 8. Note that line r is perpendicular to two sides of the hexagons.
Moving from the point where r intersects the diagonal border of the triangle inwards, we
notice that we lose a strip of width at most 34p i (see Figure 3 for the measurements of the
hexagon), whose extreme is depicted by dashed line t. Now note that s is perpendicular
to the diagonal border of the bin as well as to line t. From this, we note that the distance
18
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
AB, from the intersection A of t and the leg to the extreme B of the triangle is
√
3+ √3
3p i 3
√
3+ √3
.
p
3 i 3
The
1
−
is the area
area of the q-bin(i, p) minus the area of the triangle of leg length 3p+1
of such lost strip. Also, note that we lose a strip of height at most half the height of the
hexagon at the bottom border plus a strip of width at most half the side of the hexagon
at the left border.
Figure 8. For Online Circle Packing in Isosceles Right Triangle, the darker
hexagons are lost due to tiling one q-bin(i, p) into hexagons. The borders
of the q-bin(i, p) are the thick lines. The area between the dashed lines and
the borders is the one being counted as lost.
√
√
2
1
1
1
3+ √3
√3 )2
+
Thus, the area loss is at most 12 32p+2
− 21 ( 3p+1
− 33+
−
+
pi
pi 3
p
3p+1
3
3 i 3
√
√
8.62
32.17
1
1
2√
1
√3
√8 + 4 − 2p
− 33+
16 + 283 3 < 32p+1
− 32p+2
. If such
= 32p+1
pi 3
3p+1
i
3 i2
M
M2
3p i 3
3
2
1
tessellated q-bin(i, p) is full, then its occupation ratio is at least √π12 (MM+1)2 ( 12 32p+2
−
2
32.17
1 1
π
32.17
25.86
8.62
M
+ 32p+2 M 2 )/( 2 32p+2 ) = √12 (M +1)2 1 − M + M 2 .
32p+1 M
Now, consider (iii). When an Sbin closes, there can be only one q-bin(i, p) for each
p ≥ 0 which is tessellated but not full. Also, no q-bin(i, 0) can be empty, but at most
32 − 1 q-bins(i, p) for each p ≥
1 can. Thus, the area loss due to non-full q-bins(i, p) is at
P 1 32 −1 1 1
P
1 1
most p≥0 2 32p+2 + p≥1 2 32p+2 = 2 32 −2 + 21 312 .
Putting everything together, we have that the occupation ratio of the closed Sbin is at
32.17 √π
M2
+
.
least 21 − 12 321−1 − 21 312 1 − 25.86
2
M
M
12 (M +1)2
Regarding the bounded space algorithm, note that each closed Lbin of type i keeps i
circles of type i. Thus, its occupation ratio is at least iπρ2i+1 .
Now we give the weighting function w which will be used for the analysis of the bounded
space algorithm. We define w(Ii ) = 1/i, so a closed Lbin of type i has total weight
π
M2
55
√
1 − 25.86
+ 32.17
. For a small circle c of type (i, p)
exactly 1. Let OR = 144
M
M2
12 (M +1)2
and area a, we define w(c) = a/OR, so a closed Sbin of type i has total weight at least 1
since at least OR of the bin’s area is occupied, according to Theorem 4.1.
ONLINE CIRCLE AND SPHERE PACKING
19
i
iπρ2i+1
i
iπρ2i+1
i
iπρ2i+1
i
iπρ2i+1
1 0.134753 14 0.337587 27 0.360436 40 0.374184
2 0.214343 15 0.333221 28 0.360441 41 0.373902
3 0.241844 16 0.340169 29 0.362313 42 0.375378
4 0.246066 17 0.339481 30 0.364711 43 0.375743
5 0.285752 18 0.345218 31 0.365624 44 0.375958
6 0.281614 19 0.348218 32 0.366048 45 0.378311
7 0.291069 20 0.349180 33 0.367387 46 0.379287
8 0.305438 21 0.350563 34 0.368083 47 0.378933
9 0.319025 22 0.354141 35 0.369609 48 0.379238
10 0.310476 23 0.353191 36 0.370060 49 0.381510
11 0.318106 24 0.354329 37 0.371387 50 0.380398
12 0.323289 25 0.358628 38 0.371382 51 0.382243
13 0.329000 26 0.358762 39 0.374478 52 0.382699
Table 3. For circles in isosceles right triangle, these are the first 50 values
of i and the respective lower bound on the occupation ratio associated with
Ii .
Theorem 4.2 concludes with the asymptotic competitive ratio of the bounded space
algorithm.
Theorem 4.2. The algorithm for packing circles in isosceles right triangle with bounded
space has asymptotic competitive ratio strictly below 2.5490.
Proof. Note that, since M = 93, we have OR ≥ 0.246036. We show some values of iπρ2i+1
in Table 3, from where we can note that for i ≥ 4 we have iπρ2i+1 > 0.246036. Thus, in the
two-phase program mentioned in the end of Section 2, the second phase can try to add
items of type up to 3 to augment each configuration found during phase one. However,
phase one can easily verify if configurations of items of type up to 3 are feasible or not,
so phase two is not necessary.
Now consider a bin of maximum weight. If it does not have any I1 or I2 , phase two
of the program finds a configuration with three I3 . Thus, the asymptotic competitive
ratio is given by the sum of the weights of such circles (3 31 ) plus the remaining area
(0.5 − 3πρ24 ) divided by OR, which is less than 2.049261. Now we only need to consider
the configurations when these two types of circles are present.
With the first phase of the two-phase program, we tested all possible variations of
configurations with items I1 , I2 , and I3 , and we found the following ones:
(1) I1 , I2 ;
(2) I1 , two I3 ;
(3) two I2 , I3 .
In the following, we calculate the total sum of weights in each of these configurations,
considering the formula W + ( 21 − V )/OR:
(1) 1 +
1
2
+ ( 12 − πρ22 − πρ23 )/OR < 2.548932;
20
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
(2) 1 + 2 13 + ( 21 − πρ22 − 2πρ24 )/OR < 2.495885;
(3) 2 21 + 31 + ( 12 − 2πρ23 − πρ24 )/OR < 2.166716.
And so the result follows.
Now we analyse the unbounded space algorithm. As circles of type greater than 3
already have an occupation ratio greater than the occupation ratio of small circles, we
will use waiting bins only to pack one I1 with one I2 , as depicted in Figure 9. All other
circles are packed as in the bounded space algorithm.
Figure 9. For Online Circle Packing in Isosceles Right Triangle, a waiting
bin places the I1 at corner E and the I2 at corner F .
Consider the notation of Figure 9. The circle centered at point E is an I1 while the
circle centered at point F is an I2 . We fix that the radius of the I1 is D and we want
to find what is the maximum radius x that
√ the I2 can have as a function of D. Note
√
2− 2
, and triangle KEH is rectangle, we have
that, since ∠KEH = 67.5◦ , cos(67.5) =
q √
q √ 2
2+√2
√2 . Similarly, JG = x
HK = D 2+
. Since triangle EF L is rectangle, EL = D−x,
2− 2
2− 2
√
√
EF = D + x, and LF = KJ, we have
KJ = 2 Dx. At last, since HK + KJ + JG = 2,
q
√
√
√
√
√
we must have x = (5 − 4 2)D − 2 2(7 − 5 2)D − 7 2 + 10 D − 2 + 2.
q
√
√
√
√
√
Now, for simplicity, let γ = (5 − 4 2)D −2 2(7 − 5 2)D − 7 2 + 10 D − 2 + 2.
Note that we need to have ρ2 < D ≤ ρ1 and ρ3 < γ ≤ ρ2 so that the circles are of the
desired types. This is true for 0.2071068 < D < 0.230249 and so we fixed D = 0.229661.
The algorithm works as follows. If an I1 with radius r > D arrives, then we pack it as
in the bounded space algorithm (one per bin). Otherwise, r ≤ D and we pack it in an
already opened waiting bin containing one I2 , if it exists, or we open a new one, pack the
I1 there, and let it open waiting for an I2 .
If an I2 with radius r > γ arrives, then we pack it as in the bounded space algorithm
(two per bin). When r ≤ γ, the circles are labeled according to their arrival so that the
following steps can be repeated at every sequence of 63 of them. The first 58 are packed
in 29 bins (2 per bin) and the last five ones are packed in five waiting bins when they
ONLINE CIRCLE AND SPHERE PACKING
21
arrive: either there is one with an I1 or we open one, pack the circle there, and let it open
waiting for an I1 .
Again, we will use two weighting functions w1 and w2 for the analysis. When the
algorithm ends we also have two possibilities: there are open waiting bins with I1 , in
which case we apply w1 over all circles, or there are not, in which case we apply w2 . Both
w1 and w2 differ from w only regarding I1 and I2 , if their radii are at most D and γ,
respectively. If I1 has radius r ≤ D, w1 (I1 ) = 1 and w2 (I1 ) = 0. If I2 has radius r ≤ γ,
29
34
w1 (I2 ) = 63
and w2 (I2 ) = 63
.
It remains to show that, for both functions, the sum of weights in any bin is at least
1 on average. Consider w1 over items I1 and I2 . All waiting bins have total weight at
least 1 because they contain an I1 . For every set of 63 I2 , we have 58 of them packed into
29 bins and the last five ones packed into five waiting bins with an I1 each. Thus, the
29
average weight of these bins is 29(2 29
)
+
5(1
+
)
/34 = 1. This is similar for w2 .
63
63
Theorem 4.3 concludes with the asymptotic competitive ratio of our algorithm.
Theorem 4.3. The algorithm for packing circles in isosceles right triangles with unbounded space has asymptotic competitive ratio strictly below 2.5094.
Proof. First consider w1 was applied. An I1 always has weight 1 and we consider its area
29
and area
at least πρ22 , and an I2 either has weight 12 and area at least πγ 2 or has weight 63
2
at least πρ3 .
If w2 was applied, an I1 either has weight 1 and area at least πD 2 or has weight 0 and
34
area at least πρ32 , and an I2 either has weight 12 and area at least πγ 2 or has weight 63
and area at least πρ23 .
Now consider a bin of maximum weight. We again consider all configurations presented
in the proof of Theorem 4.2 and recalculate the total sums of weights, but now with
all possible combinations of weights and areas, as mentioned above. The asymptotic
approximation ratio when no I1 or I2 exist is still bounded above by 2.049261. For the
configurations with items I1 or I2 we now have:
(1) 1 + 21 + ( 12 − πρ22 − πγ 2 )/OR < 2.509320,
29
+ ( 12 − πρ22 − πρ23 )/OR < 2.509249,
1 + 63
1 + 21 + ( 12 − πD 2 − πγ 2 )/OR < 2.430039,
34
+ ( 12 − πD 2 − πρ23 )/OR < 2.509334.
1 + 63
(2) 1 + 2 31 + ( 21 − πρ22 − 2πρ24 )/OR < 2.495885,
1 + 2 31 + ( 21 − πD 2 − 2πρ24 )/OR < 2.416605;
(3) 2 21 + 31 + ( 12 − 2πγ 2 − πρ24 )/OR < 2.087492,
29
2 63
+ 13 + ( 12 − 2πρ23 − πρ24 )/OR < 2.087351,
34
+ 13 + ( 12 − 2πρ23 − πρ24 )/OR < 2.246081.
2 63
And the result follows.
22
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
4.1. A lower bound for bounded space algorithms. Consider the following notation.
A circle c, of area A(c), has weight w(c) = 1/i if we can pack i copies of c in a bin but
P
we cannot pack i + 1 copies of c. For a set C of circles, define w(C) = c∈C w(c) and
P
A(C) = c∈C A(c). Hokama et al. [10] showed that, for Online Circle Packing in Squares,
√
every bounded space online algorithm has competitive ratio at least w(C) + π12 (1 − A(C)),
where C is a set of circles that can be packed into a unit square. We can show the same
result for Online Circle Packing in Isosceles Right Triangles.
Theorem 4.4. Let C be a set of circles that can be packed into an isosceles right triangle
of leg length one. Every bounded space online algorithm has competitive ratio at least
√
w(C) + π12 (1 − A(C)).
With the first phase of the program mentioned in the end of Section 2, we verified
that we can pack one circle of each type 1, 2, 4, 9, 11, and 28 in the same bin. Using
Theorem 4.4, we have the following result.
Theorem 4.5. Any bounded space online approximation algorithm for packing circles in
isosceles right triangles has competitive ratio at least 2.1193.
5. Online Sphere Packing in Cubes
For Online Sphere Packing in Cubes, optimal values and good lower bounds of ρi
are known for i up to 1024192 [14]. Since ρ1024192 < 0.00554967 and ρi < ρ1024192 for
i < 1024192, we chose M = 300, which makes 2/M ≤ 0.006666667. Recall that Ii denotes
a large sphere of type i.
We start by showing the occupation ratio of a closed Sbin, which is a result valid for
both bounded space and unbounded space algorithms. Parameter C is chosen as 3 in
order to maximize this ratio.
Theorem 5.1. The occupation ratio of a closed Sbin of type i, for M ≤ i < 3M, is at
π
M3
649
√
1 − 40.98
+ 61.26
− 2443.77
.
least 702
M
M2
M3
18 (M +1)3
Proof. Note that the volume loss in a closed Sbin is due to three factors: (i) packing spheres
into rhombic dodecahedra, (ii) tessellating a q-bin(i, p) with rhombic dodecahedra, and
(iii) non-full q-bins(i, p).
3
2
into
Consider (i). Packing one sphere of type (i, p) of volume at least 4π
3
3p (i+1)
√
3
a rhombic dodecahedron of volume 169 3 3p i6√6
gives an occupation ratio of at least
i3
√π
18 (i+1)3
3
which is at least √π18 (MM+1)3 .
For (ii), consider Figure 10. When we tessellate a q-bin(i, p), we lose three blocks of
1
1
, and width 3p+1
is
the bin. From face GG′ F ′ F , a block of depth at most 34p i , height 3p+1
1
1
4
8√
lost. From face HEF G, a block of depth at most 3p i 2 , height 3p+1 , and width 3p+1 − 3p i
(because it intersects the last block) is lost. And from face EF F ′ E ′ , a block of depth at
ONLINE CIRCLE AND SPHERE PACKING
23
1
1
most 34p i , height 3p+1
− 3p i8√2 (because it intersects the last block), and width 3p+1
− 34p i
(because it intersects the first block) is lost. See Figure 3 for the measurements of the
rhombic dodecahedron.
Figure 10. For the Online Sphere Packing in Cubes, the darker rhombic
dodecahedra are lost due to tessellating one q-bin(i, p) into hexagons. The
borders of the q-bin(i, p) are the thick lines. The areas between the dashed
lines and the borders are the ones being counted as lost. Note that they
incorporate small pieces of rhombic dodecahedra in the other borders.
Thus, the volume loss due to the borders of one tessellated q-bin(i, p) is at most ( 34p i )
1
1
1
1
1
1
1
) ( 3p+1
) + ( 3p i8√2 )( 3p+1
)( 3p+1
− 34p i ) + ( 34p i )( 3p+1
− 3p i8√2 )( 3p+1
− 34p i ) = 33p+2
(4+ √82 +4) −
( 3p+1
i
13.66
1
61.26
√ ) < 3p+2
( √322 +16+ √322 ) + 33p1 i3 ( 128
− 33p+3
+ 390.51
3p M 3 . If such tessellated q-bin(i, p) is
33p+1 i2
3
M
M2
2
3
13.66
61.26
1
1
π
M
+ 33p+3
− 390.51
/ 33p+3
full, then its occupation ratio is at least √18 (M +1)3 33p+3 − 33p+2
3p M 3
M
M2
3
61.26
2443.77
+
−
.
= √π18 (MM+1)3 1 − 40.98
2
3
M
M
M
Now, consider (iii). When an Sbin closes, there can only be one q-bin(i, p) for each
p ≥ 0 which is tessellated but not full. Also, no q-bin(i, 0) can be empty, but at most
33 − 1 q-bins(i, p) for each p ≥ 1 can. Thus, the volume loss due to non-full q-bins(i, p) is
P
P
at most p≥0 (1/33p+3 ) + p≥1 ((33 − 1)/33p+3) = 1/(33 − 1) + 1/33 .
Putting everything together, we have that the occupation ratio of the closed Sbin of
π
61.26
2443.77
M3
√
+
−
.
type i is at least 1 − 331−1 − 313 1 − 40.98
2
3
M
M
M
18 (M +1)3
Regarding the bounded space algorithm, note that each closed Lbin of type i keeps at
most i spheres of type i. Thus, its occupation ratio is at least i(4/3)πρ3i+1 , since the radius
of any of its spheres is at least ρi+1 .
The weighting function w which will be used for the analysis is the same as the ones
given for the previous problems. A large sphere Ii receives weight w(Ii ) = 1/i. For
π M3
649
√
1 − 40.98
+ 61.26
− 2443.77
, a small sphere s of type (i, p) and radius
OR = 702
M
M2
M3
18 (M +1)3
3
r receives weight w(s) = 4πr /(3OR).
Theorem 5.2 concludes with the asymptotic competitive ratio of the bounded space
algorithm.
Theorem 5.2. The algorithm for packing spheres in cubes with bounded space has asymptotic competitive ratio strictly below 3.5316.
24
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
Proof. Since M = 300, we have OR ≥ 0.585590. We show some values of i(4/3)πρ3i+1 in
Table 4, from where we can note that, for i ≥ 236, we have i(4/3)πρ3i+1 > 0.585590. We use
the two-phase program mentioned in the end of Section 2 to find possible configurations,
which are shown below. All configurations considered large spheres of type up to 235 for
the second phase.
Consider a bin of maximum weight. If it does not have I1 or I2 , then phase two of
the program finds a configuration with three I3 , four I4 , five I5 , and one I27 . Thus,
the asymptotic competitive ratio is given by the sum of the weights of such spheres
1
(3 13 +4 41 +5 15 + 27
) plus the remaining volume (1−4πρ34 −(16/3)πρ35 −(20/3)πρ36 −(4/3)πρ328 )
divided by OR, which is at most 3.040409.
All possible variations of configurations with I1 and I2 found by the two-phase program
are the following:
(2) I1 , three I4 , one I8 ,
five I9 , ten I10 ;
(1) I1 , I2 , nine I9 , ten
I10 , I33 ;
(3) two I2 , two I4 , five I5 ,
four I9 , three I10 .
In the following, we calculate the total sum of weights in each of these configurations,
considering the formula W + (1 − V )/OR and using OR ≥ 0.585590 as OR for all configurations:
1
1
+ 33
+ (1 − (4/3)πρ32 − (4/3)πρ33 − (36/3)πρ310 − (40/3)πρ311 −
(1) 1 + 21 + 9 19 + 10 10
(4/3)πρ334 )/OR < 3.531580;
1
+ (1 − (4/3)πρ32 − (12/3)πρ35 − (4/3)πρ39 − (20/3)πρ310 −
(2) 1 + 3 41 + 81 + 5 91 + 10 10
(40/3)πρ311)/OR < 3.434356;
1
+ (1 − (8/3)πρ33 − (8/3)πρ35 − (20/3)πρ36 − (16/3)πρ310 −
(3) 2 21 + 2 14 + 5 15 + 4 91 + 3 10
(12/3)πρ311)/OR < 3.246382.
For the unbounded space algorithm, we also have that a waiting bin can pack one I1
with either one I2 or two I4 , as depicted in Figure 11. All other spheres are packed as in
the bounded space algorithm.
E
H
F
E
H
F
H
L
F
H
G
E
K
G
K
J
E'
H'
F' H'
G'
L
J
F'
E'
H'
F'
G'
G'
Figure 11. For Online Sphere Packing in Cubes, a waiting bin either
places the I1 at corner H ′ and the I2 at corner F (left drawings) or it
places the I1 centered at face HGG′ H ′ with one I4 at corner E and the
other at corner F ′ (right drawings).
F'
ONLINE CIRCLE AND SPHERE PACKING
25
i i(4/3)πρ3i+1 i i(4/3)πρ3i+1 i i(4/3)πρ3i+1 i i(4/3)πρ3i+1
1
0.133417 38 0.512017
75
0.545362 211 0.592174
2
0.210497 39 0.519548
76
0.537212 212 0.594981
77
0.537329 213 0.597788
3
0.315745 40 0.509090
4
0.308052 41 0.502683
78
0.537011 214 0.600594
5
0.357008 42 0.500044
79
0.540779 215 0.603401
6
0.393341 43 0.503414
80
0.538053 216 0.606207
7
0.458148 44 0.512617
81
0.541092 217 0.609014
82
0.540091 218 0.611820
8
0.418723 45 0.519972
9
0.370946 46 0.528259
83
0.546666 219 0.614627
10 0.374895 47 0.539743
84
0.553252 220 0.590387
11 0.409321 48 0.510538
85
0.559839 221 0.588235
86
0.566425 222 0.583545
12 0.446532 49 0.515781
13 0.483743 50 0.515564
87
0.561817 223 0.584662
14 0.417067 51 0.511943
88
0.542814 224 0.583463
15 0.422827 52 0.515117
89
0.541957 225 0.584316
90
0.541448 226 0.582069
16 0.450217 53 0.512968
17 0.470755 54 0.511075
91
0.542754 227 0.582964
18 0.463482 55 0.513913
92
0.546913 228 0.583569
19 0.451938 56 0.522789
93
0.547168 229 0.585194
20 0.466283 57 0.531881
94
0.549273 230 0.585598
95
0.553838 231 0.587916
21 0.457616 58 0.541209
22 0.467420 59 0.550540
96
0.558158 232 0.584947
23 0.477859 60 0.559871
97
0.563516 233 0.584738
98
0.567314 234 0.582294
24 0.475015 61 0.569203
25 0.486897 62 0.578534
99
0.573024 235 0.584336
26 0.504206 63 0.537287 100 0.561194 236 0.586733
27 0.464907 64 0.538601 200 0.581381 237 0.589219
28 0.482104 65 0.541275 201 0.579463 238 0.591670
29 0.499323 66 0.533264 202 0.579480 239 0.594156
30 0.516541 67 0.527819 203 0.580509 240 0.596642
31 0.533759 68 0.531857 204 0.581582 241 0.599128
32 0.495350 69 0.538335 205 0.578124 242 0.601614
33 0.486560 70 0.546027 206 0.579691 243 0.604100
34 0.496882 71 0.553801 207 0.581678 244 0.606586
35 0.488672 72 0.544790 208 0.583755 245 0.609072
36 0.499446 73 0.546273 209 0.586561 246 0.611558
37 0.513083 74 0.553519 210 0.589368 247 0.614044
Table 4. For spheres in cubes, these are the values of i and the respective
lower bound on the occupation ratio associated with i items Ii for 1 ≤ i ≤
100 and 200 ≤ i ≤ 247.
Consider the notation of Figure 11. In the left drawings, the sphere centered at point
J is an I1 while the circle centered at point K is an I2 . We fix that the radius of the I1 is
D and we want to find what is the maximum radius x that the I2 can have as a function
26
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
√
√
of D. Because HF = 2 and HH ′ , we know F H ′ = 3. So, sin(∠JH ′ L) = √13 , which
√
√
√
3
− D.
means H ′J = D 3. Similarly, KF = x 3. Thus, we have x = √3+1
For the right drawings of Figure 11, the sphere centered at point J is an I1 while the
√
spherep
centered at point K is an I4 . Because HE = 1 and T G′ = 2 we can find that
√
x = − 2D + 2 + √12 + 1.
p
√
√
3
Let γ = √3+1
− D and λ = − 2D + 2 + √12 + 1. Again, we need to have ρ2 < D ≤ ρ1 ,
ρ3 < γ ≤ ρ2 , and ρ5 < λρ4 so that the spheres are of the desired types. This is true for
0.316987299 < D < 0.3342699108. We fixed D = 0.33213.
The unbounded space algorithm works in the same way of the previous ones. If an I1
arrives and it has radius r > D, then we pack it as in the bounded space algorithm (one
per bin). Otherwise, if r ≤ D, then we pack it in a waiting bin (favoring already opened
ones, as before).
If an I2 with radius r > γ arrives, then we pack it as in the bounded space algorithm
(two per bin). Spheres of type 2 with r ≤ γ are labeled according to their arrival so that
the following procedure can be repeated at every sequence of 29 of them. The first 28
spheres are packed in 14 bins (two per bin) and the last one is packed in one waiting bin.
At last, if an I4 with radius r > λ arrives, then we pack it as in the bounded space
algorithm (four per bin). Spheres of type 4 with r ≤ λ are labeled according to their
arrival so that the following procedure can be repeated at every sequence of 18 such
spheres. The first 16 are packed in 4 bins (4 per bin) and the last 2 is packed in a waiting
bin.
Again, when the algorithm finishes there can only be two possibilities: there are open
waiting bins with I1 or not. If the former happens, we apply function w1 over all spheres,
and if the latter happens, we apply w2 .
Both w1 and w2 differ from function w only regarding spheres of type 1, 2, and 4 of
radii at most D, γ, and λ, respectively. For I1 of radius r ≤ D, w1 (s) = 1 and w2 (s) = 0.
15
4
14
and w2 (s) = 28
. For I4 of radius r ≤ λ, w1 (s) = 18
and
For I2 of radius r ≤ γ, w1 (s) = 28
5
w2 (s) = 18 .
One can observe that, for both these functions, the sum of weights in any closed bin
is at least 1 on average. Theorem 5.3 concludes with the asymptotic competitive ratio of
our algorithm.
Theorem 5.3. The algorithm for packing spheres in cubes with unbounded space has
asymptotic competitive ratio strictly below 3.5146.
Proof. If w1 was applied, then an I1 always has volume at least (4/3)πρ32 and it always has
weight 1 while if function w2 was applied, then it has volume at least (4/3)πD 3 and also
or 15
and volume at
weight 1. Regarding I2 , we can have either a sphere with weight 14
28
28
least (4/3)πρ33 or spheres with weight 21 and volume at least (4/3)πγ 3 . And regarding I4 ,
ONLINE CIRCLE AND SPHERE PACKING
27
4
5
we can have either spheres with weight 18
or 18
and volume at least (4/3)πρ35 or spheres
with weight 14 and volume at least (4/3)πλ3.
Now consider a bin of maximum weight. We recalculate the total sums of weights,
with all possible combinations of weights and volumes, as mentioned above, over the
configurations presented in the proof of Theorem 5.2. The last configuration is the one
without spheres of type 1 or 2:
1
1
+ 33
+ (1 − (4/3)πρ32 − (4/3)πγ 3 − (36/3)πρ310 − (40/3)πρ311 −
(1) 1 + 12 + 9 19 + 10 10
(4/3)πρ334 )/OR < 3.514593,
1
1
1 + 14
+ 9 91 + 10 10
+ 33
+ (1 − (4/3)πρ32 − (4/3)πρ33 − (36/3)πρ310 − (40/3)πρ311 −
28
(4/3)πρ334 )/OR < 3.514339,
1
1
+ 33
+ (1 − (4/3)πD 3 − (4/3)πγ 3 − (36/3)πρ310 − (40/3)πρ311 −
1 + 21 + 9 19 + 10 10
(4/3)πρ334 )/OR < 3.480357, or
1
1
+ 9 91 + 10 10
+ 33
+ (1 − (4/3)πD 3 − (4/3)πρ33 − (36/3)πρ310 − (40/3)πρ311 −
1 + 15
28
(4/3)πρ334 )/OR < 3.514585;
1
+ (1 − (4/3)πρ32 − (12/3)πλ3 − (4/3)πρ39 − (20/3)πρ310 −
(2) 1 + 3 41 + 81 + 5 19 + 10 10
(40/3)πρ311)/OR < 3.446663,
4
1
1 + 3 18
+ 18 + 5 19 + 10 10
+ (1 − (4/3)πρ32 − (12/3)πρ35 − (4/3)πρ39 − (20/3)πρ310 −
(40/3)πρ311)/OR < 3.370020,
1
+ (1 − (4/3)πD 3 − (12/3)πλ3 − (4/3)πρ39 − (20/3)πρ310 −
1 + 3 14 + 81 + 5 91 + 10 10
(40/3)πρ311)/OR 3.412427, or
1
5
+ 18 + 5 19 + 10 10
+ (1 − (4/3)πD 3 − (12/3)πρ35 − (4/3)πρ39 − (20/3)πρ310 −
1 + 3 18
(40/3)πρ311)/OR < 3.502450;
1
(3) 2 21 + 2 41 + 5 15 + 4 91 + 3 10
+ (1 − (8/3)πγ 3 − (8/3)πλ3 − (20/3)πρ36 − (16/3)πρ310 −
(12/3)πρ311)/OR < 3.207947,
1
14
+ 2 14 + 5 51 + 4 19 + 3 10
+ (1 − (8/3)πρ33 − (8/3)πλ3 − (20/3)πρ36 − (16/3)πρ310 −
2 28
(12/3)πρ311)/OR < 3.207439,
4
1
+ 5 51 + 4 19 + 3 10
+ (1 − (8/3)πγ 3 − (8/3)πρ35 − (20/3)πρ36 − (16/3)πρ310 −
2 21 + 2 18
(12/3)πρ311)/OR < 3.156852,
14
4
1
2 28
+ 2 18
+ 5 15 + 4 19 + 3 10
+ (1 − (8/3)πρ33 − (8/3)πρ35 − (20/3)πρ36 − (16/3)πρ310 −
(12/3)πρ311)/OR < 3.156344,
1
15
+ 2 14 + 5 51 + 4 19 + 3 10
+ (1 − (8/3)πρ33 − (8/3)πλ3 − (20/3)πρ36 − (16/3)πρ310 −
2 28
(12/3)πρ311)/OR < 3.276405,
5
1
+ 5 51 + 4 19 + 3 10
+ (1 − (8/3)πγ 3 − (8/3)πρ35 − (20/3)πρ36 − (16/3)πρ310 −
2 21 + 2 18
(12/3)πρ311)/OR < 3.267963, or
5
1
15
+ 2 18
+ 5 15 + 4 19 + 3 10
+ (1 − (8/3)πρ33 − (8/3)πρ35 − (20/3)πρ36 − (16/3)πρ310 −
2 28
(12/3)πρ311)/OR < 3.336421;
28
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
1
(4) 3 31 + 4 14 + 5 51 + 27
+ (1 −4πρ34 −(16/3)πλ3 −(20/3)πρ36 −(4/3)πρ328 )/OR < 3.031488,
5
1
+5 15 + 27
+(1−4πρ34 −(16/3)πρ35 −(20/3)πρ36 −(4/3)πρ328 )/OR < 2.929298,
3 13 +4 18
5
1
+5 15 + 27
+(1−4πρ34 −(16/3)πρ35 −(20/3)πρ36 −(4/3)πρ328 )/OR < 3.151520.
3 31 +4 18
And the result follows.
5.1. A lower bound for bounded space algorithms. In this section we will consider
the following notation. A sphere s has radius r(s), weight w(s) = 1/i if we can pack i
copies of s in a bin but we cannot pack i+1 copies of s, surface area A(s), and volume V (s).
P
P
P
For a set S of spheres, define r(S) = s∈S r(s), w(S) = s∈S w(s), A(S) = s∈S A(s),
P
and V (S) = s∈S V (s).
Theorem 5.4. Let S be a set of spheres that can be packed into a unit cube. Every
√
bounded space online algorithm has competitive ratio at least w(S) + π18 (1 − V (S)).
Proof. This proof is equal to the proof given by Hokama et al. [10] in a similar result for
packing circles in unit squares, with the proper alterations for spheres.
√
Let 0 < ǫ < 1 be a constant, δ = πǫ/(4 18), and ϕ < δ. We will build a sequence of
sets of spheres S0 , S1 , . . . until we obtain an Sk such that V (Sk ) ≥ (1 − ϕ)2 .
We start with S0 = S and, for each n ≥ 1, the idea is to build Sn by using the spheres
√
in Sn−1 plus some small spheres of radius 6ℓn /3 where ℓn < ℓn−1 and ℓ1 < mins∈S0 r(s).
We describe the details next.
√
2+44
32π
√ |Sn−1 | + 12 √
). Suppose that
For n ≥ 1, let ℓn ≤ ϕ/( 3√6 3 A(Sn−1 ) + 16π
r(S
)
+
n−1
3
9 3
3 3
Sn−1 can be packed in one bin and fix one of its packings. Consider a tessellation of
rhombic dodecahedra of side ℓn on the bin in which Sn−1 is packed. Add to the packing
√
spheres of radii 6ℓn /3 in every feasible rhombic dodecahedron, i.e., rhombic dodecahedra
that do not intersect spheres of Sn−1 or the borders of the bin.
As in Theorem 5.1, the total volume of√ infeasible rhombic
dodecahedra
that intersect
√
√
√
2 √2ℓn
4ℓ
2 √2ℓn
2 √2ℓn
4ℓ
2 √2ℓn
n
n
√
√
the border of a unit cubic bin is at most
+
(1
−
)
+
(1
−
)(1
−
)
3
3
3
3
3
3
√
√ )ℓ3 . If a rhombic dodecahedron intersects the interior of a sphere
)ℓn + ( 332
< ( 4 √2+4
3
3 n
of radius√r centered at a point p, then it is properly contained in the sphere of radius
r + 2ℓn / 3 centered at p (see Figure 3 for the measurements of rhombic dodecahedra).
Thus, the total volume of feasible rhombic dodecahedra is at least
!
√
3 !
X
4
32
2ℓ
4 2+4
n
√
ℓn − √ ℓ3n −
π r(s) + √
1−
3
3
3 3
3
s∈Sn−1
!
√
X 4
32
8 3
6
4 2+4
3
2
2
3
√
ℓn − √ ℓn −
π r(s) + √ ℓn r(s) + 4ℓn r(s) + √ ℓn
= 1−
3
3
3 3
3
3 3
s∈Sn−1
!
√
4 2+4
6
32
16π 2
32π
√
= 1−
ℓn − √ ℓ3n − V (Sn−1 ) − √ ℓn A(Sn−1 ) −
ℓn r(Sn−1 ) − √ ℓ3n |Sn−1 |
3
3
3 3
3 3
9 3
!
√
6
32
32π
4 2+4
16π
√
ℓn − √ ℓn − V (Sn−1 ) − √ ℓn A(Sn−1 ) −
> 1−
ℓn r(Sn−1 ) − √ ℓn |Sn−1 | (1)
3
3
3 3
3 3
9 3
ONLINE CIRCLE AND SPHERE PACKING
29
≥ 1 − V (Sn−1 ) − ϕ
where (1) follows because ℓn ≤ 1.
Since every new sphere occupies a volume of √π18 of the volume of the rhombic dodecahedron, we have that V (Sn ) ≥ V (Sn−1 )+(1−V (Sn−1 )−ϕ) √π18 . It follows, from induction,
n
n
that V (Sn ) ≥ (1 − ϕ) 1 − 1 − √π18
. Note that 1 − √π18 < ϕ as n grows, so there
exists k ≥ 0 such that V (Sk ) ≥ (1 − ϕ)2 . From transitivity, Sk can be packed in a bin
and S0 ⊆ Sk .
Now consider an instance composed by N disjoint copies of Sk and let S0 have qi spheres
of type i, i.e., of radii between ρi+1 and ρi . Note that an optimal offline solution uses N
bins to pack such instance and consider that, for the online algorithm, the spheres arrive
in non-increasing order of radii. Any online algorithm with bounded space B uses at least
Nqi /i − B bins for every type of sphere in S0 . Suppose that nj spheres, each of volume vj ,
were added to Sj−1 in order to construct Sj . Since the best packing of spheres has density
√
√
π/ 18, any online algorithm with bounded space B uses at least Nnj vj 18/π − B bins
to pack the nj spheres.
Let N ≥ 2(t + k)B/ǫ, where t is the number of different types of spheres in S. By the
analysis above, any algorithm with bounded space B uses at least
X
type i in S0
!
√
√
X
k
k
X
qi
18
18 X
N qi
N nj vj
(nj vj ) − kB
−B +
−B =N
− tB + N
i
π
i
π j=1
j=1
type i in S0
!
√
√
X
18
18
q
i
2
= N
(1 − ϕ) −V (S) −(t+k)B
+
(V (Sk ) − V (S0 )) −(t+k)B ≥ N w(S)+
i
π
π
type i in S0
!
!
√
√
√
18
18
18
(1−2ϕ−V (S)) −(t+k)B = N w(S)+
(1−V (S)) −N
2ϕ−(t+k)B
> N w(S)+
π
π
π
!
!
√
√
18
18
Nǫ
> N w(S)+
(1−V (S)) −
−(t+k)B ≥ N w(S)+
(1−V (S)) −N ǫ ,
π
2
π
and so the result follows.
With the first phase of the program mentioned in the end of Section 2, we verified that
we can pack one sphere of type 1, one of type 2, four of type 9, and one of type 18 in the
same bin. Using Theorem 5.4, we have the following result.
Theorem 5.5. Any bounded space online approximation algorithm for packing spheres in
cubes has competitive ratio at least 2.7707.
References
[1] Nikhil Bansal, Jos R. Correa, Claire Kenyon, and Maxim Sviridenko. Bin Packing in Multiple Dimensions: Inapproximability Results and Approximation Schemes. Mathematics of Operations Research,
31(1):31–49, 2006.
30
C. N. LINTZMAYER, F. K. MIYAZAWA, AND E. C. XAVIER
[2] Nikhil Bansal and Arindam Khan. Improved Approximation Algorithm for Two-dimensional Bin
Packing. In Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA’2014), pages 13–25. SIAM, 2014.
[3] Henrik I. Christensen, Arindam Khan, Sebastian Pokutta, and Prasad Tetali. Approximation and
online algorithms for multidimensional bin packing: A survey. Computer Science Review, 24:63–79,
2017.
[4] Erik D. Demaine, Sándor P. Fekete, and Robert J. Lang. Circle Packing for Origami Design Is Hard,
pages 609–626. A K Peters/CRC Press, Singapore, 2010.
[5] Leah Epstein. Two-dimensional Online Bin Packing with Rotation. Theoretical Computer Science,
411(31):2899–2911, 2010.
[6] Leah Epstein and Rob van Stee. This Side Up! ACM Transactions on Algorithms, 2(2):228–243,
2006.
[7] Erich Friedman. Packing Unit Squares in Squares: A Survey and New Results. The Electronic Journal
of Combinatorics, DS7:24, 2009.
[8] Xin Han, Deshi Ye, and Yong Zhou. A note on online hypercube packing. Central European Journal
of Operations Research, 18(2):221–239, 2010.
[9] M. Hifi and R. M’Hallah. A Literature Review on Circle and Sphere Packing Problems: Models and
Methodologies. Advances in Operations Research, 2009:1–22, 2009.
[10] Pedro Hokama, Flávio Keidi Miyazawa, and Rafael Crivellari Saliba Schouery. A Bounded Space
Algorithm for Online Circle Packing. Information Processing Letters, 116(5):337–342, 2016.
[11] Shahin Kamali, Alejandro López-Ortiz, and Zahed Rahmati. Online Packing of Equilateral Triangles.
In Proceedings of the 27th Canadian Conference on Computational Geometry (CCCG’2015), 2015.
[12] Andrea Lodi, Silvano Martello, Michele Monaci, and Daniele Vigo. Two-Dimensional Bin Packing
Problems, pages 107–129. John Wiley & Sons, Inc., 2013.
[13] Flávio Keidi Miyazawa, Lehilton Lelis Chaves Pedrosa, Rafael Crivellari Saliba Schouery, Maxim
Sviridenko, and Yoshiko Wakabayashi. Polynomial-Time Approximation Schemes for Circle and
Other Packing Problems. Algorithmica, 76(2):536–568, 2016.
[14] Eckard Specht. Packomania. http://www.packomania.com/. Accessed: 2017-03-06.
[15] Péter Gábor Szabó, Mihaly Csaba Markót, Tibor Csendes, Eckard Specht, Leocadio G. Casado, and
Inmaculada Garcı́a. New Approaches to Circle Packing in a Square. Springer Optimization and Its
Applications. Springer US, New York, USA, 2007.
[16] Milos Tatarevic. On Limits of Dense Packing of Equal Spheres in a Cube. The Electronic Journal of
Combinatorics, 22(1):35, 2015.
Instituto de Computação, Universidade Estadual de Campinas, Campinas, Brazil
E-mail address: {carlanl,fkm,eduardo}@ic.unicamp.br
| 8cs.DS
|
arXiv:1707.05581v4 [math.GR] 8 Sep 2017
ON STRONGLY QUASICONVEX SUBGROUPS
HUNG CONG TRAN
Abstract. We develop a theory of strongly quasiconvex subgroups of
an arbitrary finitely generated group. Strong quasiconvexity generalizes quasiconvexity in hyperbolic groups and is preserved under quasiisometry. Strongly quasiconvex subgroups are more reflexive than stable
subgroups defined by Durham-Taylor, and they also capture the geometry of ambient groups. We prove that strongly quasiconvex subgroups
have many properties analogous to those of quasiconvex subgroups of
hyperbolic groups. We show a strong connection between strong quasiconvexity and lower relative divergence, and we suggest a classification
of strongly quasiconvex subgroups by using lower relative divergence.
We give complete descriptions of strong quasiconvexity and stability
in relatively hyperbolic groups. Additionally, we characterize strongly
quasiconvex special subgroups and stable special subgroups of two dimensional right-angled Coxeter groups, and we use these results to study
Morse boundary of right-angled Coxeter groups. We prove that two
notions of strong quasiconvexity and stability are equivalent in a rightangled Artin group AΓ (except for the case of finite index subgroups)
and we also characterize non-trivial strongly quasiconvex subgroups of
infinite index in AΓ (i.e. non-trivial stable subgroups in AΓ ) by quadratic lower relative divergence. These results strengthen the work
of Koberda-Mangahas-Taylor on characterizing purely loxodromic subgroups of right-angled Artin groups.
1. Introduction
One method to understand the structure of a group G is to investigate subgroups of G. Using this approach, one often investigates subgroup H ≤ G
whose geometry reflects that of G. Quasiconvex subgroup of hyperbolic
group is a successful application of this approach. This is well-known that a
quasiconvex subgroup of a hyperbolic group is also hyperbolic. Generalizing
quasiconvex subgroups of hyperbolic groups, Dahmani introduced the first
definition of relatively quasiconvex subgroups of relatively hyperbolic groups
in [Dah03]. Then Osin proposed a substantially different definition of relatively quasiconvex subgroups of relatively hyperbolic groups in [Osi06]. In
[Hru10], Hruska proved these two definitions are equivalent to each other. In
the same paper, Hruska also proved that a relatively quasiconvex subgroup
of a relatively hyperbolic group is also relatively hyperbolic with peripheral
Date: September 11, 2017.
2000 Mathematics Subject Classification. 20F67, 20F65.
1
ON STRONGLY QUASICONVEX SUBGROUPS
2
structure inherited from the one of the ambient group but the subgroup is
not necessarily hyperbolic.
Quasiconvexity is not as useful for arbitrary finitely generated groups since
quasiconvexity depends on a choice of generating set and, in particular, is
not preserved under quasi-isometry. In [DT15b], Durham-Taylor introduce a
strong notion of quasiconvexity in finitely generated groups, called stability,
which is preserved under quasi-isometry. Stability agrees with quasiconvexity when ambient groups are hyperbolic. Moreover, the stable subgroups of
mapping class groups are precisely the convex cocompact subgroups defined
by Farb-Mosher in [FM02] and the such type of subgroups of mapping class
groups is a primary motivation for the concept stable subgroups of a finitely
generated group (see [DT15b]). The study of stable subgroups has received
much attention in recent years, for examples: Koberda-Mangahas-Taylor
[KMT], Aougab-Durham-Taylor [ADT], Antolı́n-Mj-Sisto-Taylor [AMST],
Abbott-Behrstock-Durham [ABD], and others.
However, a stable subgroup of a finitely generated group is always hyperbolic no matter the ambient group is hyperbolic or not (see [DT15b]). In
some sense, the geometry of a stable subgroup does not reflect completely
that of the ambient group. Therefore, we introduce another concept of quasiconvexity, called strong quasiconvexity, which is strong enough to be preserved under quasi-isometry and reflexive enough to capture the geometry
of ambient groups.
Definition 1.1. Let G be a finite generated group and H a subgroup of G.
We say H is strongly quasiconvex in G if for every K ≥ 1, C ≥ 0 there is some
M = M (K, C) such that every (K, C)–quasigeodesic in G with endpoints
on H is contained in the M –neighborhood of H.
Outside hyperbolic setting, there are many strongly quasiconvex subgroups that are not stable. For example, non-hyperbolic peripheral subgroups of a relatively hyperbolic group and non-hyperbolic hyperbolically
embedded subgroups of a finitely generated group are non-stable strongly
quasiconvex subgroups (see [DS05] and [Sis16]). In Section 7, we also provide
many examples of non-stable strongly quasiconvex subgroups in right-angled
Coxeter groups.
However, there is a strong connection between strong quasiconvexity and
stability. More precisely, a subgroup is stable if and only if it is strongly quasiconvex and hyperbolic (see Theorem 4.8). In some sense, we define strong
quasiconvexity from stability but taking out the hyperbolicity to make subgroups better reflect the geometry of the ambient group. Additionally, we
can study stable subgroups in a finitely generated group via investigating its
strongly quasiconvex subgroups. More specifically, in this paper, we study
strong quasiconvexity in relatively hyperbolic groups, right-angled Coxeter
groups, and right-angled Artin groups and that leads to many results on stability in these such groups that were unknown before (see Sections 6, 7 and
8). We remark that the study of stability of right-angled Coxeter groups in
ON STRONGLY QUASICONVEX SUBGROUPS
3
this paper also help us understand connectedness of Morse boundary of these
groups. We characterize a strongly quasiconvex subgroup via the completely
super linear lower relative divergence of the ambient group with respect to
the subgroup. This characterization helps us characterize a stable subgroup
via lower relative divergence and hyperbolicity.
It is well-known that ordinary quasiconvex subgroups fail to satisfy many
properties analogous to those of quasiconvex subgroups of hyperbolic groups.
However, we prove that strongly quasiconvex subgroups have many those
such properties (see Sections 4 and 5). Especially, there is a nice connection
between the boundary of a strongly quasiconvex subgroup and the boundary
of the ambient group that is analogous to one in hyperbolic setting. We also
prove that the lower relative divergence of a group with respect to a strongly
quasiconvex subgroup is diverse. Therefore, this suggests a classification of
strongly quasiconvex subgroups by using lower relative divergence.
1.1. Some properties of strongly quasiconvex subgroups. It is wellknown that a quasiconvex subgroup H of a hyperbolic group G is finitely
generated and undistorted (see [Sho91]) and H has finite index in its commensurator (see [KS96]). In addition, the intersection between two quasiconvex subgroups of a hyperbolic group is also a quasiconvex subgroup (see
[Sho91]) and the collection of quasiconvex subgroups of a hyperbolic group
has finite height and finite width (see [GMRS98]). In this paper, we also
prove that strongly quasiconvex subgroups of arbitrary finitely generated
groups also have analogous properties.
Theorem 1.2. Let G be an arbitrary finitely generated group. Then:
(1) If H is a strongly quasiconvex subgroup of G, then H is finitely
generated, undistorted subgroup and H has finite index in its commensurator.
(2) If H1 and H2 are arbitrary strongly quasiconvex subgroups of G, then
H1 ∩ H2 is strongly quasiconvex in G, H1 , and H2 .
(3) If H = {H1 , · · · , Hℓ } is a finite collection of strongly quasiconvex
subgroups of G, then H has finite height, finite width, and bounded
packing.
In [AMST], Antolı́n-Mj-Sisto-Taylor prove that if H is a finite collection
of stable subgroups of a finitely generated group G, then H has finite height,
finite width, and bounded packing in G. The above theorem strengths their
work to strongly quasiconvex subgroups. Due to the lack of hyperbolicity
of an arbitrary finitely generated group G and its strongly quasiconvex subgroups, the proof for these properties is more challenging. Combining Theorem 1.2 with the work of Sageev [Sag95, Sag97] and Hruska-Wise [HW09],
we have as an immediate Corollary:
Corollary 1.3. Suppose H is a strongly quasiconvex codimension 1 subgroup of a finitely generated group G. Then the corresponding CAT(0) cube
complex is finite dimensional.
ON STRONGLY QUASICONVEX SUBGROUPS
4
In hyperbolic setting, the inclusion map H ֒→ G of a quasiconvex subgroup H into G induces a topological embedding of the Gromov boundary
of H into the Gromov boundary of G and the image of this embedding is
also the limit set of H in the Gromov boundary of G. Moreover, if H1 and
H2 are two quasiconvex subgroups of a hyperbolic group G, then the limit
set Λ(H1 ∩ H2 ) of the subgroup H1 ∩ H2 is the intersection ΛH1 ∩ ΛH2 of
limit sets of H1 and H2 (see [Gro93]).
In order to prove a strongly quasiconvex subgroup has similar properties
as a quasiconvex subgroup of a hyperbolic group has as above, we need a
boundary on finitely generated groups. In [Cor], Cordes introduces Morse
boundary of proper geodesic metric spaces which generalizes the contracting
boundary on CAT(0) spaces (see [CS15]). More importantly, Morse boundary is a quasi-isometry invariant and it agrees with Gromov boundary on
hyperbolic spaces (see [Cor]). In this paper, we also prove a connection between strong quasiconvexity and Morse boundary that is similar to the one
in hyperbolic setting.
Theorem 1.4. Let G be a finitely generated group. Then
(1) If H is a strongly quasiconvex subgroup of G, then the inclusion
i : H ֒→ G induces a topological embedding î : ∂M H → ∂M G of
the Morse boundary of H into the Morse boundary of G such that
î(∂M H) = ΛH, where ΛH is the limit set of H in the Morse boundary ∂M G of G.
(2) If H1 and H2 are strongly quasiconvex subgroups of G, then H1 ∩ H2
is strongly quasiconvex in G and Λ(H1 ∩ H2 ) = ΛH1 ∩ ΛH2 .
1.2. Strong quasiconvexity and lower relative divergence. Lower relative divergence was introduced by the author in [Tra15] to study geometric
connection between a geodesic space and its subspaces. Roughly speaking, relative divergence measures the distance distortion of the complement
of the r–neighborhood of a subspace in the whole space when r increases.
Lower relative divergence is proved to be preserved under quasi-isometry
(see Proposition 4.9 in [Tra15]). Therefore, we can define lower relative
divergence on a pair of groups (G, H), where G is finitely generated and
H ≤ G, via some (any) Cayley graph of G. We refer the reader to Section
2 for precise definitions.
In [CS15], Charney-Sultan introduce the concept of lower divergence of
quasi-geodesic to study Morse quasi-geodesics in CAT(0) spaces. Later,
Arzhantseva-Cashen-Gruber-Hume in [ACGH] use the concept of CharneySultan to study Morse quasi-geodesics in arbitrary geodesic metric spaces.
We remark that lower relative divergence introduced by the author agrees
with lower divergence introduced by Charney-Sultan on quasi-geodesics (see
Theorem 4.12 in [Tra15]). In this paper, we generalize the work of CharneySultan and Arzhantseva-Cashen-Gruber-Hume to characterize Morse subsets in arbitrary geodesic metric spaces via lower relative divergence.
ON STRONGLY QUASICONVEX SUBGROUPS
5
Theorem 1.5. Let X be a geodesic space and A a subset of X of infinite
diameter. Then A is Morse in X if and only if the lower relative divergence
of X with respect to A is completely super linear.
Applying the above theorem, we also characterize strongly quasiconvex
subgroups via lower relative divergence.
Theorem 1.6. Let G be a finitely generated group and H infinite subgroup
of G. Then H is strongly quasiconvex in G if and only if the lower relative
divergence of G with respect to H is completely super linear.
It is proved in [Tra15] the lower relative divergence of a hyperbolic group
with respect to a quasiconvex subgroup is at least exponential. The author
in [Tra] also proves that lower relative divergence of a non-hyperbolic group
with respect to a stable subgroup can be any polynomials. In this paper, we
also prove a similar result but applying on non-stable strongly quasiconvex
subgroups.
Theorem 1.7. For each d ≥ 2 there is a non-hyperbolic group Gd together
with a non-stable strongly quasiconvex subgroup whose lower relative divergence is a polynomial of degree d.
The above theorem shows that the lower relative divergence with respect
to strongly quasiconvex subgroups are diverse and, therefore, the theorem,
suggests a classification of strongly quasiconvex subgroups using lower relative divergence of ambient groups with respect to them.
Question 1.8. Is lower relative divergence of a CAT(0) group with respect
to an infinite strongly quasiconvex subgroup at least quadratic?
1.3. Strong quasiconvexity and stability in relatively hyperbolic
groups. In this paper, we fully describe strongly quasiconvex subgroups
and stable subgroups of relatively hyperbolic groups. We remark that in
relatively hyperbolic group setting, strongly quasiconvex subgroups and stable subgroups are distinct from relatively quasiconvex subgroups defined by
Dahmani [Dah03] and Osin [Osi06]. We first come up with the description
of strongly quasiconvex subgroups of relatively hyperbolic groups.
Theorem 1.9. Let (G, P) be a finitely generated relatively hyperbolic group
and H finitely generated undistorted subgroup of G. Then the following are
equivalent:
(1) The subgroup H is strongly quasiconvex in G.
(2) The subgroup H ∩P g is strongly quasiconvex in P g for each conjugate
P g of peripheral subgroup in P.
(3) The subgroup H ∩ P g is strongly quasiconvex in G for each conjugate
P g of peripheral subgroup in P.
In [ADT], Aougab-Durham-Taylor characterize stability in relatively hyperbolic groups whose peripheral subgroups are one-ended and have linear
ON STRONGLY QUASICONVEX SUBGROUPS
6
divergence. In this paper, we give a complete characterization of stable
subgroups of arbitrary relatively hyperbolic groups.
Corollary 1.10. Let (G, P) be a finitely generated relatively hyperbolic group
and H finitely generated undistorted subgroup of G. Then the following are
equivalent:
(1) The subgroup H is stable in G.
(2) The subgroup H ∩ P g is stable in P g for each conjugate P g of peripheral subgroup in P.
(3) The subgroup H ∩ P g is stable in G for each conjugate P g of peripheral subgroup in P.
1.4. Strong quasiconvexity, stability and Morse boundary of rightangled Coxeter groups. In this paper, we also establish a characterization
of strongly quasiconvex special subgroups of two-dimensional right-angled
Coxeter groups.
Theorem 1.11. Let Γ be a simplicial, triangle free graph with vertex set S
and K subgroup of GΓ generated by some subset S1 of S. Then the following
conditions are equivalent:
(1) The subgroup K is strongly quasiconvex in GΓ .
(2) If S1 contains two non-adjacent vertices of an induced 4–cycle σ,
then S1 contains all vertices of σ.
(3) Either |K| < ∞ or the lower relative divergence of GΓ with respect
to K is at least quadratic.
An immediate corollary of the above theorem is a characterization of
stable special subgroups of two-dimensional right-angled Coxeter groups.
Corollary 1.12. Let Γ be a simplicial, triangle free graph with vertex set S
and K subgroup of GΓ generated by some subset S1 of S. Then the following
conditions are equivalent:
(1) The subgroup K is stable in GΓ .
(2) The set S1 does not contain a pair of non-adjacent vertices of an
induced 4–cycle in Γ.
The above corollary can be used to prove some connectedness of Morse
boundary of right-angled Coxeter groups:
Corollary 1.13. If the simplicial, triangle free graph Γ contains an induced
loops σ of length greater than 4 such that the vertex set of σ does not contain
a pair of non-adjacent vertices of an induced 4–cycle in Γ, then the Morse
boundary of right-angled Coxeter group GΓ is not totally disconnected.
We remark that Behrstock in [Beh] constructs example of certain rightangled Coxeter groups which contain a closed hyperbolic surface stable subgroup and prove that the Morse boundaries of these groups are not totally
disconnected. We believe that the approach of Behrstock can also be generalize to proofs of Corollaries 1.12 and 1.13. If so, our proofs of Corollaries
ON STRONGLY QUASICONVEX SUBGROUPS
7
1.12 and 1.13 can be considered as alternative ways. However, we remark
that the method in [Beh] can not be used to study strongly quasiconvex
subgroups in right-angled Coxeter groups as in Theorem 1.11.
Conjecture 1.14. The Morse boundary of a right-angled Coxeter group GΓ
is not totally disconnected if and only if the defining graph Γ contains an
induced loops σ of length greater than 4 such that the vertex set of σ does
not contain a pair of non-adjacent vertices of an induced 4–cycle in Γ.
1.5. Strong quasiconvexity, stability, and lower relative divergence
in right-angled Artin groups. As we discussed above, two notions of
strong quasiconvexity and stability are equivalent in hyperbolic setting.
Outside hyperbolic setting, there are many strongly quasiconvex subgroups
that are not stable. A natural question arises on the existence of nonhyperbolic ambient group on which two notions of strongly quasiconvex
subgroup of infinite index and stable subgroup are equivalent:
Question 1.15. Is there any non-hyperbolic ambient group setting on which
two notions of strongly quasiconvex subgroup of infinite index and stable
subgroup are equivalent? In other word, is there any non-hyperbolic ambient
group setting on which all strongly quasiconvex subgroups of infinite index
are hyperbolic?
The following theorem answers above question:
Theorem 1.16. Let Γ be a simplicial, finite, connected graph such that Γ
does not decompose as a nontrivial join. Let H be a non-trivial, infinite
index subgroup of the right-angled Artin group AΓ . Then the following are
equivalent:
(1) H is strongly quasiconvex.
(2) H is stable.
(3) The lower relative divergence of AΓ with respect to H is quadratic.
(4) The lower relative divergence of AΓ with respect to H is completely
super linear.
The above theorem also shows that linear and quadratic functions are
the only functions which are lower relative divergence in right-angled Artin
groups in the non-trivial case (i.e. the defining graph of the ambient group
is simplicial, finite, connected and the subgroup is non-trivial and of infinite index). The most surprising fact obtained from the above theorem is
the equivalence between strong quasiconvexity and stability (except for the
trivial case of finite index subgroups) in non-hyperbolic settings.
In [KMT], Koberda-Mangahas-Taylor give several characterizations of
purely loxodromic subgroups in right-angled Artin groups. One of these is
the equivalence between purely loxodromic subgroups and stable subgroups.
The above theorem also strengthens the work of Koberda-Mangahas-Taylor
in [KMT] to provide several alternative characterizations of purely loxodromic subgroups in right-angled Artin groups.
ON STRONGLY QUASICONVEX SUBGROUPS
8
Corollary 1.17. Let Γ be a simplicial, finite, connected graph such that
Γ does not decompose as a nontrivial join. Let H be a non-trivial, finitely
generated, infinite index subgroup of the right-angled Artin group AΓ . Then
the following are equivalent:
(1) H is purely loxodromic.
(2) H is strongly quasiconvex.
(3) The lower relative divergence of AΓ with respect to H is quadratic.
(4) The lower relative divergence of AΓ with respect to H is completely
super linear.
We conjecture that similar characterizations can be obtained for convex
cocompact subgroups of mapping class groups.
1.6. Some questions. Hierarchically hyperbolic groups were recently introduced by Behrstock, Hagen, and Sisto [BHS17] to provide a uniform
framework in which to study many important families of groups, including
mapping class groups, right-angled Coxeter groups, most 3-manifold groups,
right-angled Artin groups and many others. In [ABD], Abbott-BehrstockDurham prove a complete characterization of stability in hierarchically hyperbolic groups.
As we discussed above, strong quasiconvexity and stability both generalize the notion quasiconvexity of hyperbolic groups but they are two different
notions in general. So far we obtained several results about strong quasiconvexity in right-angled Artin groups and right-angled Coxeter groups and
we remark again that these such groups are also hierarchically hyperbolic.
Therefore, it would be nice if we can characterize strong quasiconvexity in
hierarchically hyperbolic groups.
Question 1.18. Characterizing strong quasiconvexity in hierarchically hyperbolic groups?
We also hope there is a connection between strong quasiconvexity and
stability in some certain hierarchically hyperbolic groups which is analogous
to the one in right-angled Artin groups.
Question 1.19. What are conditions on a hierarchically hyperbolic group
so that all strongly quasiconvex subgroups of infinite index are stable?
There is also a natural question that arises on the existence of hierarchically hyperbolic structure on a strongly quasiconvex subgroup of a hierarchically hyperbolic group.
Question 1.20. Is a strongly quasiconvex subgroup of a hierarchically hyperbolic group hierarchically hyperbolic?
In [Sis16], Sisto proves that a hyperbolically embedded subgroup is strongly
quasiconvex. We hope to obtain a converse statement.
Question 1.21. Under what conditions, a strongly quasiconvex subgroup
is hyperbolically embedded?
ON STRONGLY QUASICONVEX SUBGROUPS
9
In [Osi16], Osin characterizes acylindrically hyperbolic groups using hyperbolically embedded subgroups. Therefore, the answer to the above question may give some connection between two notions of strongly quasiconvex
subgroups and acylindrically hyperbolic groups.
Acknowledgments. I want to thank Prof. Christopher Hruska, Prof. Ruth
Charney, Prof. Jason Behrstock, Prof. Thomas Koberda, Prof. François Dahmani, and Hoang Thanh Nguyen for their very helpful conversations. I also
thank Matthew Haulmark for advice that improved the exposition of the
paper. Lastly, I want to thank Anthony Genevois for many helpful discussions.
2. Preliminaries
In this section, we review some well-known concepts in geometric group
theory: geodesic spaces, geodesics, quasigeodesics, quasi-isometry, quasiisometric embedding, and Morse subsets. We also discuss some recently
developed concepts: relative divergence, geodesic divergence, and geodesic
lower divergence.
Definition 2.1. Let (X, dX ) and (Y, dY ) be two metric spaces and the map
Φ from X to Y a (K, L)–quasi-isometric embedding if for all x1 , x2 in X the
following inequality holds:
(1/K) dX (x1 , x2 ) − L ≤ dY Φ(x1 ), Φ(x2 ) ≤ K dX (x1 , x2 ) + L.
If, in addition, NL Φ(X) = Y , then Φ is called a (K, L)–quasi-isometry.
Two spaces X and Y are quasi-isometric if there is a (K, L)–quasi-isometry
form X to Y .
The special case of a quasi-isometric embedding where the domain is a
connected interval in R (possibly all of R is called a (K, L)–quasi-geodesic.
A geodesic is a (1, 0)–quasi-geodesic. The metric space X is a geodesic space
if any pair of points in X can be joined by a geodesic segment.
Definition 2.2. Let X be a geodesic space and A a subset of X. Let r be
any positive number.
(1) Nr (A) = x ∈ X dX (x, A) < r
(2) ∂Nr (A) = x ∈ X dX (x, A) = r
(3) Cr (A) = X − Nr (A).
(4) Let dr,A be the induced length metric on the complement of the r–
neighborhood of A in X. If the subspace A is clear from context, we
can use the notation dr instead of using dr,A .
Definition 2.3. A subset A of a geodesic metric space X is Morse if for
every K ≥ 1, C ≥ 0 there is some M = M (K, C) such that every (K, C)–
quasigeodesic with endpoints on A is contained in the M –neighborhood of
A. We call the function M a Morse gauge.
ON STRONGLY QUASICONVEX SUBGROUPS
10
Before we define the concepts of relative divergence, geodesic divergence,
and geodesic lower divergence, we need to build the notions of domination
and equivalence which are the tools used by these concepts.
Definition 2.4. Let M be the collection of all functions from [0, ∞) to
[0, ∞]. Let f and g be arbitrary elements of M. The function f is dominated
by the function g, denoted f g, if there are positive constants A, B, C
and D such that f (x) ≤ Ag(Bx) + Cx for all x > D. Two functions f and g
are equivalent, denoted f ∼ g, if f g and g f . The function f is strictly
dominated by the function g, denoted f ≺ g, if f is dominated by g and they
are not equivalent.
We say a function f in M is completely super linear if for every choice of
C > 0 the collection of x ∈ [0, ∞) such that f (x) ≤ Cx is bounded.
Remark 2.5. The relations and ≺ are transitive. The relation ∼ is an
equivalence relation on the set M. If f g in M and f is completely super
linear, then g is also completely super linear. Therefore, if f ∼ g in M and
one of them is completely super linear, then the other is also completely
super linear.
Let f and g be two polynomial functions in the family M. We observe
that f is dominated by g if and only if the degree of f is less than or equal
to the degree of g and they are equivalent if and only if they have the same
degree. All exponential functions of the form abx+c , where a > 1, b > 0 are
equivalent. Therefore, a function f in M is linear, quadratic or exponential...
if f is respectively equivalent to any polynomial with degree one, two or any
function of the form abx+c , where a > 1, b > 0.
Definition 2.6. Let {δρn } and {δρ′n } be two families of functions of M,
indexed over ρ ∈ (0, 1] and positive integers n ≥ 2. The family {δρn } is
dominated by the family {δρ′n }, denoted {δρn } {δρ′n }, if there exists constant
n δ ′M n . Two families
L ∈ (0, 1] and a positive integer M such that δLρ
ρ
n
′n
n
{δρ } and {δρ } are equivalent, denoted {δρ } ∼ {δρ′n }, if {δρn } {δρ′n } and
{δρ′n } {δρn }.
We say a family {δρn } of functions in M is completely super linear if there
is some n0 ≥ 3 such that δρn0 are completely super linear for all ρ ∈ (0, 1].
Remark 2.7. A family {δρn } is dominated by (or dominates) a function f
in M if {δρn } is dominated by (or dominates) the family {δρ′n } where δρ′n = f
for all ρ and n. The equivalence between a family {δρn } and a function f
in M can be defined similarly. Thus, a family {δρn } is linear, quadratic,
exponential, etc. if {δρn } is equivalent to the function f where f is linear,
quadratic, exponential, etc..
If {δρn } {δρ′n } and {δρn } is completely super linear, then {δρ′n } is also
completely super linear. Therefore, if {δρn } ∼ {δρ′n } and one of them is
completely super linear, then the other is also completely super linear.
ON STRONGLY QUASICONVEX SUBGROUPS
11
We now review the concept of relative divergence which was originally
introduced by the author in [Tra15].
Definition 2.8 (Relative divergence in spaces). Let X be a geodesic space
and A a subset of X of infinite diameter. For each ρ ∈ (0, 1] and positive
integer n ≥ 2, we define two functions σρn , δρn : [0, ∞) → [0, ∞] as follows:
For each positive r, if there is no pair of x1 , x2 ∈ ∂Nr (A) such that
dr (x1 , x2 ) < ∞ and d(x1 , x2 ) ≥ nr, we define σρn (r) = ∞. Otherwise, we
define σρn (r) = inf dρr (x1 , x2 ) where the infimum is taken over all x1 , x2 ∈
∂Nr (A) such that dr (x1 , x2 ) < ∞ and d(x1 , x2 ) ≥ nr.
For each r, let δρn (r) = sup dρr (x1 , x2 ) where the supremum is taken over
all x1 , x2 ∈ ∂Nr (A) such that dr (x1 , x2 ) < ∞ and d(x1 , x2 ) ≤ nr.
The family of functions {σρn } is the lower relative divergence of X with
respect A, denoted div(X, A) and the family of functions {δρn } is the upper
relative divergence of X with respect A, denoted Div(X, A).
Convention 2.9. Let X be a geodesic space and A a subset of X of infinite
diameter. Let {σρn } be the lower relative divergence of X with respect to
A. Assume that Nr0 (A) = X for some r0 . Therefore, ∂Nr (A) = ∅ for each
r > r0 . This implies that for each ρ ∈ (0, 1] and positive integer n ≥ 2
σρn (r) = ∞ for r > r0 .
Definition 2.10 (Relative divergence in groups). Let G be a finitely generated group and H its subgroup. We define the lower relative divergence of G
with respect to H, denoted div(G, H), to be the lower relative divergence of
the Cayley graph Γ(G, S) with respect to H for some (any) finite generating
set S.
Similarly, we define the upper relative divergence of G with respect to H,
denoted Div(G, H), to be the upper relative divergence of the Cayley graph
Γ(G, S) with respect to H for some (any) finite generating set S.
Remark 2.11. We first remark that the subgroup H in the above definition
is not required to be finitely generated. Moreover, the relative divergence is
a pair quasi-isometry invariant concept (see Proposition 4.3 and Proposition
4.9 in [Tra15]). This implies that the relative divergence on a finitely generated group does not depend on the choice of finite generating sets of the
ambient group. Also, the lower relative divergence div(G, H) is dominated
by the upper relative divergence Div(G, H) with a mild condition on the
pair (G, H) (see Theorem 4.14 in [Tra15]).
We now come up with geodesic divergence and geodesic lower divergence.
Definition 2.12 (Geodesic divergence). The divergence of two geodesic
rays α and β with the same initial point x0 in a geodesic space X, denoted
Divα,β , is a function g : (0, ∞) → (0, ∞] defined as follows. For each positive
r, if there is no path outside the open ball with radius r about x0 connecting
α(r) and β(r), we define g(r) = ∞. Otherwise, we define g(r) is the infimum
ON STRONGLY QUASICONVEX SUBGROUPS
12
on the lengths of all paths outside the open ball with radius r about x0
connecting α(r) and β(r).
The divergence of a bi-infinite geodesic γ, denoted Divγ , is the divergence
of the two geodesic rays obtained from γ with the initial point γ(0).
Definition 2.13 (Geodesic lower divergence). Let α be a bi-infinite geodesic. For any r > 0 and t ∈ R, if there is no path from α(t − r) to
α(t + r) that lies outside the open ball of radius r about α(t), we define
ρα (r, t) = ∞. Otherwise, we let ρα (r, t) denote the infimum of the lengths
of all paths from α(t − r) to α(t + r) that lies outside the open ball of radius
r about α(t). Define the lower divergence of α to be the growth rate of the
following function:
ldivα (r) = inf ρα (r, t)
t∈R
Remark 2.14. The geodesic divergence is a classical notion related to
curvature and the geodesic lower divergence was originally introduced by
Charney-Sultan in [CS15]. In general, these concepts measure different aspects of a bi-infinite geodesic in a geodesic space. More precisely, let α
be a bi-infinite geodesic in a geodesic space X. It is not hard to see that
ldivα ≤ Divα and two functions are not the same in general. However,
ldivα ∼ Divα if α is a periodic geodesic (i.e. there is an isometry g of X
such that gα = α).
Also, the lower divergence ldivα of α in X is equivalent to the lower
relative divergence div(X, α) of X with respect to α. The proof of this fact
is similar to the proof of Proposition 6.6 in [Tra15] and we leave it to the
reader.
3. Characterizing Morse subsets via lower relative divergence
In most parts of this section, we characterize Morse subsets with infinite
diameter in term of the lower divergence related to them. This result will
be used to characterize strongly quasiconvex subgroups defined in Section
4. We also study behaviors of geodesic rays in finite neighborhood of some
Morse subsets and this result will be applied to study the height of strongly
quasiconvex subgroups later.
Proposition 3.1 (Completely super linear relative divergence implies Morse).
Let X be a geodesic space and A a subset of X of infinite diameter. If the
lower relative divergence {σρn } of X with respect to A is completely super
linear, then A is Morse in X.
Proof. Let n0 ≥ 3 and ρ0 ∈ (0, 1] such that σρn00 is a completely super linear
function. Let γ : [a, b] → X be an arbitrary (K, L)–quasi-geodesic in X
connecting two points in A. By Lemma 1.11 of [BH99] III.H, we can find a
continuous quasi-geodesic c : [a, b] → X such that:
(1) c(a) = γ(a) and c(b) = γ(b).
ON STRONGLY QUASICONVEX SUBGROUPS
13
(2) ℓ(c|[t,t′ ] ) ≤ K1 d c(t), c(t′ ) + L1 , where K1 ≥ 1, L1 ≥ 0 only depend
on K and L.
(3) The Hausdorff distance between γ and c is less than C, where C only
depends on K and L.
Therefore, it is sufficient to prove that c lies in some M –neighborhood of
A, where M only depends on K1 , L1 , n0 and the function σρn00 . Let t0 ∈ [a, b]
such that c(t0 ) is the farthest point from A and let m = d c(t0 ), A . If
m ≤ 2L1 , then c obviously lies in some M –neighborhood of A, where M
only depends on L1 . Therefore, we can assume that m > 2L1 . We claim
that if m > 2L1 , then σρn00 2Km1 n0 ≤ (4K1 +2)m. Let m1 = 2Km1 n0 , t1 ∈ [a, t0 ]
and t2 ∈ [t0 , b] such that c(t1 ), c(t2 ) both lie in ∂Nm1 (A) and c[t1 , t2 ] lies
outside Nm1 (A). We consider two cases.
Case 1: ℓ(c|[t1 ,t2 ] ) ≤ 4K1 m
Since d c(t0 ), A) = m and d c(t1 ), A) = 2Km1 n0 , then the length ℓ(c|[t1 ,t0 ] )
is bounded below 1 − 2K11 n0 m. Similarly, the length ℓ(c|[t0 ,t2 ] ) is bounded
below 1 − 2K11 n0 m. Therefore, the length ℓ(c|[t1 ,t2 ] ) is bounded below by
2 − K11n0 m. This implies that
L1
1
ℓ(c|[t1 ,t2 ] ) −
d c(t1 ), c(t2 ) ≥
K1
K1
1
L1
1
2−
m−
≥
K1
K1 n0
K1
1 m L1
m
+ 1−
−
≥
K1
K1 n0 K1
K1
m
L1
m
m
m
+
−
≥
≥ n0
≥
K1 2K1 K1
K1
2K1 n0
Also, the path c([t1 , t2 ]) lies outside
fore,
σρn00
m
2K1 n0
–neighborhood of A. There-
m
≤ ℓ(c|[t1 ,t2 ] ) ≤ 4K1 m.
2K1 n0
Case 2: ℓ(c|[t1 ,t2 ] ) > 4K1 m
We can choose t3 and t4 in [t1 , t2 ] (t3 < t4 ) such that ℓ(c|[t3 ,t4 ] ) = 4K1 m.
Since m1 ≤ d c(t3 ), A ≤ m, we can choose a point u in ∂Nm1 (A) such
that the geodesic α1 connecting u, c(t3 ) lies outside Nm1 (A) and its length
is bounded above by m. Similarly, we can choose a point v in ∂Nm1 (A)
such that the geodesic α2 connecting v, c(t
Nm1 (A) and its
S4 ) lies outside
S
length is bounded above by m. Let β = α1 c[t1 , t2 ] α2 . Then β is a path
outside Nm1 (A) that connects two points u and v on ∂Nm1 (A). Moreover,
the length of β is bounded above by (4K1 + 2)m by the construction. Also,
ON STRONGLY QUASICONVEX SUBGROUPS
14
1
L1
d(u, v) ≥ d c(t3 ), c(t4 ) − 2m ≥
ℓ(c|[t3 ,t4 ] ) −
− 2m
K1
K1
L1
L1
m
≥ 4m −
− 2m ≥ 2m −
≥ m ≥ n0
.
K1
K1
2K1 n0
Therefore,
σρn00
m
≤ ℓ(β) ≤ (4K1 + 2)m.
2K1 n0
Since σρn00 is a completely super linear function, then there is an upper bound on m depending on K1 and n0 . Therefore, c lies in some M –
neighborhood of A, where M only depends on K1 , L1 , n0 and the function
σρn00 .
The following two lemmas will be used to compute the lower relative
divergence of geodesic spaces with respect to their Morse subsets. Most
techniques for the proof of these lemmas were studied from [ACGH].
Lemma 3.2. Suppose γ is a concatenation of two (L, 0)–quasi-geodesics γ1
and γ2 , where γ1 , γ2 , and γ are all parametrized by arc length. Let C > 1 and
r > 0 such that the distance between two endpoints a, c of γ is at least r and
the length of γ is at most Cr. For each ρ ∈ (0, 1] and L′ > L + C + C/ρ + 1,
there exist an (L′ , 0)–quasi-geodesic α parametrized by arc length connecting
two points a, c such that the image of α lies ρr–neighborhood the image of
γ and the length of α is bounded above by the length of γ.
Proof. For each path β parametrized by arc length we will abuse notation
to identify β to its image. Moreover, for each pair of points u and v on β
we denote dβ (u, v) to be the length of the subpath of β connecting u and v.
If γ is an (L′ , 0)–quasi-geodesic, then α = γ is a desired quasi-geodesic. We
now assume that γ is not an (L′ , 0)–quasi-geodesic.
Since d(a, c) ≥ r ≥ dγ (a, c)/C > dγ (a, c)/L′ , then there is a maximal
subsegment [x, y]γ of γ such that d(x, y) = dγ (x, y)/L′ by the continuity
of γ. It is obvious that x and y can not both lie in the same path γ1 or
γ2 . Therefore, we can assume x ∈ γ1 and y ∈ γ2 . Let [x, y] be a geodesic
connecting x, y and α = [a, x]γ1 ∪ [x, y] ∪ [y, c]γ2 . We parametrized α by arc
length and we will prove that α is an (L′ , 0)–quasi-geodesic.
Let u and v be an arbitrary two points in α. If u and v both lie in the same
one of these segments [a, x]γ1 , [x, y], or [y, c]γ2 , then dα (u, v) ≤ L′ d(u, v). We
now consider the remaining three cases.
Case 1: One of these points u, v lies in [a, x]γ1 and the other lies in [x, y].
We can assume that u lies in [a, x]γ1 and v lies in [x, y]. We observe that
d(u, y) > dγ (u, y)/L′ by the maximality of [x, y]γ and the continuity of γ.
Therefore,
ON STRONGLY QUASICONVEX SUBGROUPS
15
dγ (u, y)
− d(v, y) =
L′
dγ (u, x)
dγ (u, x) + dγ (x, y)
− d(v, y) =
+ d(x, y) − d(v, y) =
=
L′
L′
dγ (u, x)
dγ (u, x) + d(x, v)
dα (u, v)
=
+ d(x, v) ≥
=
L′
L′
L′
′
or dα (u, v) ≤ L d(u, v).
Case 2: One of these points u, v lies in [y, c]γ2 and the other lies in [x, y].
Using symmetric argument as in Case 1, we can prove dα (u, v) ≤ L′ d(u, v).
Case 3: One of these points u, v lies in [a, x]γ1 and the other lies in
[y, c]γ2 . We can assume that u lies in [a, x]γ1 and v lies in [y, c]γ2 . We
observe again that d(u, v) > dγ (u, v)/L′ by the maximality of [x, y]γ and the
continuity of γ. Also, it is not hard to see dα (u, v) ≤ dγ (u, v). This implies
that d(u, v) > dα (u, v)/L′ or dα (u, v) ≤ L′ d(u, v).
Therefore, α is an (L′ , 0)-quasi-geodesic parametrized by arc length connecting a and c. It is not hard to see the length of α is less than or equal
to the length of γ. In order to see that the path α lies completely inside
ρr–neighborhood of γ, we observe that
d(u, v) ≥ d(u, y) − d(v, y) ≥
d(x, y) =
Cr
dγ (x, y)
<
= ρr.
′
L
C/ρ
Lemma 3.3. For each C > 1 and ρ ∈ (0, 1] there is a constant L =
L(C, ρ) ≥ 1 such that the following holds. Let r be an arbitrary positive
number and γ a continuous path with the length less than Cr. Assume the
distance between two endpoints x and y is at least r. Then there is an (L, 0)–
quasi-geodesic α connecting two points x, y such that the image of α lies in
the ρr–neighborhood of γ and the length of α is less than or equal to the
length of γ.
Proof. Let x0 = x and x1 be the last point on γ such that d(x0 , x1 ) = ρr/4.
Similarly, we define xi = y if d(xi−1 , y) < ρr/2 or to be the last point on γ
such that d(xi−1 , xi ) = ρr/4. By construction, we observe that d(xi , xj ) ≥
ρr/4 if i 6= j and y = xn for some n ≤ 4C/ρ. Let α1 be a concatenation of
geodesics [x0 , x1 ][x1 , x2 ] · · · [xn−1 , xn ]. Then α1 connects two points x, y, the
length of α1 is less than or equal to the length of γ, and α1 lies completely
inside (ρr/2)–neighborhood of γ.
Using Lemma 3.2 for each integer 1 ≤ i ≤ ⌊n/2⌋, there are (L2 , 0)–quasigeodesics (where L2 only depends on C and ρ) from x2(i−1) to x2i such that
the concatenation α2 of these with subpath [x2⌊n/2⌋ , xn ]α1 of α1 satisfying
the following conditions:
(1) α2 lies completely inside (ρr/4)–neighborhood of α1 .
(2) The length of α2 is less than or equal to the length of α1 .
ON STRONGLY QUASICONVEX SUBGROUPS
16
Repeating this process at most d times (d ≤ n) until we get an (Ld , 0)–
quasi-geodesic αd (where Ld only depends on C and ρ) connecting two points
x and y satisfying the following conditions:
(1) αd lies completely inside (ρr/2d )–neighborhood of αd−1 .
(2) The length of αd is less than or equal to the length of αd−1 .
Therefore, the length of the (Ld , 0)–quasi-geodesic αd is less than the
length of γ. Also,
ρr
ρr ρr
+
+ · · · + d < ρr.
2
4
2
This implies that αd lies in the ρr–neighborhood of γ. Therefore, L = Ld is
the desired number only depending on C and ρ.
Proposition 3.4 (Morse implies completely super linear relative divergence). Let X be a geodesic space and A a subset of X of infinite diameter.
If A is Morse in X, then the lower relative divergence {σρn } of X with respect
to A is completely super linear.
Proof. We will prove the stronger result that σρn is completely super linear
for all n ≥ 5 and ρ ∈ (0, 1]. Let µ be the Morse gauge of A and assume for
the contradiction that σρn00 is not completely super linear for some n0 ≥ 5
and ρ0 ∈ (0, 1]. Then there is C > 0 for which there is an unbounded
sequence of numbers rm and paths γm satisfying
(1) Each path γm lies outsides the (ρ0 rm )–neighborhood of A and their
endpoints xm , ym lie in ∂Nrm (A).
(2) d(xm , ym ) ≥ n0 rm and ℓ(γm ) ≤ Crm .
′ be a geodesic of length less than 2r that connects x
For each m, let γm
m
m
′′ be a geodesic of length less than
and some point um in A. Similarly, let γm
′ ∪ γ ∪ γ ′′ .
2rm that connects ym and some point vm in A. Let γ m = γm
m
m
Then γ m is a path of length at most (C + 4)rm that connects two points
um , vm on A with d(um , vm ) ≥ (n0 − 4)rm ≥ rm .
By Lemma 3.3, for each m there is an (L, 0)–quasi-geodesic αm (L does
not depend on m) connecting two points um , vm satisfying
(1) ℓ(αm ) ≤ ℓ(γ m ) ≤ (C + 4)rm .
(2) αm lies in the (ρ0 rm /2)–neighborhood of γ m .
′ is contained
By triangle inequality, each (ρ0 rm /2)–neighborhood of γm
′′
in the open ball B(xm , 2rm + ρ0 rm /2) and (ρ0 rm /2)–neighborhood of γm
is contained in the open ball B(ym , 2rm + ρ0 rm /2). Since d(xm , ym ) ≥
n0 rm ≥ 5rm , the intersection between two open balls B(xm , 2rm + ρ0 rm /2)
and B(ym , 2rm + ρ0 rm /2) is empty. Therefore, two (ρ0 rm /2)–neighborhoods
′ and γ ′′ have an empty intersection. Since α
of γm
m is connected and αm
m
′ and γ ′′ , α can not lie
intersects with both (ρ0 rm /2)–neighborhoods of γm
m
m
′ ∪ γ ′′ . This implies that
completely inside the (ρ0 rm /2)–neighborhood of γm
m
there is a point sm on αm that lies in the (ρ0 rm /2)–neighborhood of γm .
Since each path γm lies outside the (ρ0 rm )–neighborhood of A, the point sm
must lie outside the (ρ0 rm /2)–neighborhood of A.
ON STRONGLY QUASICONVEX SUBGROUPS
17
Recall that each αm is an (L, 0)–quasi–geodesic with endpoints on A and
A is Morse in X with Morse gauge µ. Each path γm must lie completely
inside the µ(L, 0)–neighborhood of A. In particular, each point sm must lie
inside the µ(L, 0)–neighborhood of A. Therefore, µ(L, 0) ≥ ρ0 rm /2 for all m
which contradicts to the choice of sequence rm . Therefore, σρn is completely
super linear for all n ≥ 5 and ρ ∈ (0, 1].
The following theorem is deduced from Propositions 3.1 and 3.4.
Theorem 3.5 (Characterization of Morse subsets). Let X be a geodesic
space and A a subset of X of infinite diameter. Then A is Morse in X if
and only if the lower relative divergence of X with respect to A is completely
super linear.
We now study behavior of geodesic rays in finite neighborhood of some
Morse subset. We first start with a technical lemma that basically refers to
“quasigeodesic approximation” of some certain concatenations of geodesics.
Lemma 3.6. Let γ = γ1 γ2 γ3 be the concatenation of three geodesics γ1 , γ2 ,
and γ3 . If there is a number r > 0 such that ℓ(γ1 ) < r, ℓ(γ2 ) = 20r, and
ℓ(γ3 ) < r, then there is a (5, 0)–quasigeodesic α with the same endpoints of
γ and α ∩ γ2 6= ∅.
Proof. For each path β parametrized by arc length we will abuse notation
to identify β to its image. Moreover, for each pair of points u and v on
β we denote dβ (u, v) to be the length of the subpath of β connecting u
and v. Let a and b be the endpoints of γ1 , let b and c be the endpoints
of γ2 , and let c and d be the endpoints of γ3 . We claim that there are
points x1 ∈ γ1 and y1 ∈ γ2 such that d(x1 , y1 ) ≤ 5r and the concatenation
α1 = [a, x1 ]γ1 ∪ [x1 , y1 ] ∪ [y1 , c]γ2 is a (5, 0)–quasigeodesic, where [x1 , y1 ] is a
geodesic connecting x1 , y1 .
If the concatenation β = γ1 γ2 is a (5, 0)–quasigeodesic, then we can choose
x1 = y1 = b and α1 = β. We now assume that β is not a (5, 0)–quasigeodesic.
Since d(a, c) ≥ d(b, c) − d(a, b) ≥ 19r and dβ (a, c) ≤ d(a, b) + d(b, c) ≤ 21r,
d(a, c) > dβ (a, c)/5. Therefore, there is a maximal subsegment [x1 , y1 ]β of β
such that d(x1 , y1 ) = dβ (x1 , y1 )/5 by the continuity of β. It is obvious that
x1 and y1 can not both lie in the same path γ1 or γ2 . Therefore, we can
assume x1 ∈ γ1 and y1 ∈ γ2 . Obviously, d(x1 , y1 ) = dβ (x1 , y1 )/5 ≤ ℓ(β)/5 ≤
5r. Moreover, α1 is a (5, 0)–quasigeodesic by using the same argument as in
Lemma 3.2.
Similarly, there are points x2 ∈ γ3 and y2 ∈ γ2 such that d(x2 , y2 ) ≤
5r and the concatenation α2 = [b, y2 ]γ2 ∪ [y2 , x2 ] ∪ [x2 , d]γ3 is a (5, 0)–
quasigeodesic, where [y2 , x2 ] is a geodesic connecting y2 , x2 . Let
α = [a, x1 ]γ1 ∪ [x1 , y1 ] ∪ [y1 , y2 ]γ2 ∪ [y2 , x2 ] ∪ [x2 , d]γ3 .
ON STRONGLY QUASICONVEX SUBGROUPS
18
Then α ⊂ α1 ∪ α2 and α ∩ γ2 6= ∅. We now prove that α is a (5, 0)–
quasigeodesic. Let u and v be arbitrary two points in α. We consider
several cases.
If two points u and v both lies in α1 , then dα1 (u, v) ≤ 5d(u, v) because
α1 is a (5, 0)–quasigeodesic. Also, dα (u, v) = dα1 (u, v) by the construction
of α and α1 . This implies that dα (u, v) ≤ 5d(u, v). We also obtain a similar
inequality if two points u and v both lies in α2 . In the remaining case, u
and v does not lie in the same path α1 or α2 . Therefore, one of two points
u and v must lie in [a, x1 ]γ1 ∪ [x1 , y1 ] (say u) and the remaining point must
lie in [y2 , x2 ] ∪ [y2 , d]γ3 (say v). Since the length of [a, x1 ]γ1 ∪ [x1 , y1 ] is less
than 6r, d(a, u) ≤ 6r. Similarly, d(d, v) ≤ 6r. Therefore,
d(u, v) ≥ d(a, d) − d(a, u) − d(d, v)
≥ d(b, c) − d(a, b) − d(c, d) − 12r
≥ 20r − r − r − 12r ≥ 6r.
Also, dα (u, v) ≤ ℓ(α) ≤ ℓ(γ) ≤ 22r. Therefore, dα (u, v) ≤ 5d(u, v). Therefore, α is a (5, 0)–quasigeodesic.
Proposition 3.7. Let X be a geodesic space and A a µ–subset of X. Then
there is a constant D > 0 such that the following hold. If γ : [0, ∞) → X is
a geodesic ray that lies in some finite neighborhood of A, then there is C > 0
such that γ|[C,∞) lies in the D–neighborhood of A.
Proof. Assume that γ lies in the r–neighborhood of A for some r > 0. For
each i ≥ 1 let γi = γ|[20(i−1)r,20ir] . Then each γi is a geodesic of length 20r.
Since γ lies in the r–neighborhood of A, for each i ≥ 0 there is a geodesic βi
of length less than r that connects γ(20ir) and A. By Lemma 3.6 for each
i ≥ 1 there is a (5, 0)–quasigeodesic αi with the same endpoints with the
concatenation βi−1 γi βi such that αi ∩ γi 6= ∅.
Since each αi is a (5, 0)–quasigeodesic with endpoints in A, αi lies entirely
inside the µ(5, 0)–neighborhood of A. We recall that αi ∩ γi 6= ∅. Therefore,
there is ti ∈ [20(i − 1)r, 20ir] such that γ(ti ) lies in the µ(5, 0)–neighborhood
of A. Therefore, for each i ≥ 1 each geodesic γ|[ti ,ti+1 ] has endpoints in
the µ(5, 0)–neighborhood of A. This implies that γ|[ti ,ti+1 ] lies in some D–
neighborhood of A, where D only depends on µ. Set C = t1 then γ|[C,∞)
lies in the D–neighborhood of A obviously.
4. Strong quasiconvexity, stability, and lower relative
divergence
In this section, we proposed concept of strongly quasiconvex subgroups.
We show the connection between strongly quasiconvex subgroups and stable
subgroups and we characterize these subgroups using lower relative divergence. We also study some basic properties of strongly quasiconvex subgroups that are analogous to properties of quasiconvex subgroups in hyperbolic groups.
ON STRONGLY QUASICONVEX SUBGROUPS
19
Definition 4.1. Let Φ : A → X be a quasi-isometric embedding between
geodesic metric spaces. We say A is strongly quasiconvex in X if the image
Φ(A) is Morse in X. We say A is stable in X if for any K ≥ 1, L ≥ 0 there
is an R = R(K, L) ≥ 0 so that if α and β are two (K, L)–quasi-geodesics
with the same endpoints in Φ(A), then the Hausdorff distance between α
and β is less than R.
Note that when we say A is strongly quasiconvex (stable) in X we mean
that A is strongly quasiconvex (stable) in X with respect to a particular
quasi-isometric embedding Φ : A → X. Such a quasi-isometric embedding
will always be clear from context, for example an undistorted subgroup H of
a finitely generated group G. The following proposition provides with some
basic facts about strong quasiconvexity.
Proposition 4.2. Suppose that A, B, X are geodesic metric spaces and
f
g
B → A → X are quasi-isometric embeddings.
(1) If B is strongly quasiconvex in X via g ◦ f , then B is strongly quasiconvex in A via f .
(2) If B is strongly quasiconvex in A via f and A is strongly quasiconvex
in X via g, then B is strongly quasiconvex in X via g ◦ f .
Proof. We first prove Statement (1) of the above proposition. More precisely,
we are going to prove that f (B) is Morse in A. For each K ≥ 1 and L ≥ 0
let γ be an arbitrary (K, L)–quasi-geodesic in A that connects two points
in f (B). Then
g(γ) is (K ′ , L′ )–quasigeodesic in X that connects two points
in g f (B) , where K ′ , L′ depend only on K, L, and the quasi-isometry
embedding g. Since B is strongly quasiconvex
in X via g ◦ f , then g(γ)
lies in some M –neighborhood of g f (B) , where M depend only on K ′
and L′ . Again, g is a quasi-isometry embedding. Therefore, γ lies in some
M ′ –neighborhood of f (B), where M ′ depends only on K, L, f and the
quasi-isometry embedding g. This implies that B is strongly quasiconvex
(stable) in A via f .
We now prove Statement (2) of the above proposition. More precisely,
we are going to prove that (g ◦ f )(B) is Morse in X. For each K ≥ 1 and
L ≥ 0 let α be an
in X that connects two
arbitrary (K, L)–quasi-geodesic
points in g f (B) . Since g f (B) ⊂ g(A) and g(A) is Morse subset in X,
then α lies in some M –neighborhood of g(A), where M depends only on K
and L. Also, g : A → X is a quasi-isometry embedding. Then there is a
(K1 , L1 )–quasigeodesic β in A that connects two points in f (B) such that
the Hausdorff distance between g(β) and α is bounded above by C, where
K1 , L1 , C depend only on K, L, M , and the quasi-isometry embedding g.
Also, f (B) is a Morse subset of A. Then β lies in some M1 –neighborhood
f (B), where M1 depends only on K1 and L1 . Again, g : A → X is a quasi
isometry embedding. Then g(β) lies in some M2 –neighborhood of g f (B) ,
where M2 depends only on M1 and the map g. Therefore, α lies in the
ON STRONGLY QUASICONVEX SUBGROUPS
20
(M2 + C)–neighborhood of g f (B) . Therefore, B is strongly quasiconvex
in X via g ◦ f .
The following proposition gives the exact relationship between strongly
quasiconvex and stable subspaces.
Proposition 4.3. Let Φ : A → X be a quasi-isometric embedding between
geodesic metric spaces. Then A is stable in X if and only if A is strongly
quasiconvex and hyperbolic.
Proof. One direction of the above proposition is deduced from Remark 3.1
and Lemma 3.3 in [DT15b]. We now prove that if A is strongly quasiconvex
and hyperbolic, then A is stable in X. For any K ≥ 1, L ≥ 0 let α and
β are two (K, L)–quasi-geodesics with the same endpoints in Φ(A). Since
Φ(A) is Morse in X, there is a constant C ≥ 0 not depending on α and
β such that α and β both lies in the C–neighborhood of Φ(A). This is
an easy exercise that there are constants K1 ≥ 1, L1 ≥ 0, D ≥ 0 not
depending on α and β and two (K1 , L1 )–quasi-geodesics α1 and β1 with the
same endpoints in A such that the Hausdorff distance between Φ(α1 ), α
and the Hausdorff distance between Φ(β1 ), β are both bounded above by
D. Since A is a hyperbolic space and Φ is a quasi-isometric embedding,
the Hausdorff distance between Φ(α1 ) and Φ(β1 ) is bounded above by some
constant D1 which does not depend on α and β. Therefore, the Hausdorff
distance between α and β is bounded above by D1 + 2D. Therefore, A is
stable in X.
We now come up with concepts of strongly quasiconvex subgroups and
stable subgroups.
Definition 4.4. Let G be a finite generated group and S an arbitrary finite
generating set of G. Let H be a finite generated subgroup of G and T an
arbitrary finite generating set of H. The subgroup H is undistorted in G if
the natural inclusion i : H → G induces a quasi-isometric embedding from
the Cayley graph Γ(H, T ) into the Cayley graph Γ(G, S). We say H is stable
in G if Γ(H, T ) is stable in Γ(G, S).
We remark that stable subgroups were proved to be independent of the
choice of finite generating sets (see Section 3 in [DT15b]).
Definition 4.5. Let G be a finite generated group and H a subgroup of G.
We say H is quasiconvex in G with respect to some finite generating set S
of G if there exists some C > 0 such that every geodesic in the Cayley graph
Γ(G, S) that connects a pair of points in H lies inside the C–neighborhood
of H. We say H is strongly quasiconvex in G if H is a Morse subset in the
Cayley graph Γ(G, S) for some (any) finite generating set S.
Remark 4.6. If H is a quasiconvex subgroup of a group G with respect
to some finite generating set S, then H is also finitely generated and undistorted in G (see Lemma 3.5 of [BH99] III.Γ). However, we emphasis that the
ON STRONGLY QUASICONVEX SUBGROUPS
21
concept of quasiconvex subgroups depend on the choice of finite generating
set of the ambient group.
This is clear that if H is a Morse subset in the Cayley graph Γ(G, S) with
some finite generating set S, then H is also a quasiconvex subgroup of G
with respect to S. In particular, H is finitely generated and undistorted in
G. Therefore, the strong quasiconvexity of a subgroup does not depend on
the choice of finite generating sets by the Statement (1) in Proposition 4.2.
Moreover, if a finitely generated group G acts properly and cocompactly in
some space, then H is a strongly quasiconvex (stable) subgroup of G if and
only if H is strongly quasiconvex (stable) in X via some (any) orbit map
restricted on H.
The following theorem is a direct consequence of Theorem 3.5
Theorem 4.7 (Characterizing strongly quasiconvex subgroups). Let G be
a finitely generated group and H infinite subgroup of G. Then H is strongly
quasiconvex in G if and only if the lower relative divergence of G with respect
to H is completely super linear.
We also obtain several characterizations of stable subgroups by the following theorem and the proof can be deduced from Proposition 4.3 and
Theorem 4.7.
Theorem 4.8 (Characterizing stable subgroups). Let G be a finitely generated group and H infinite subgroup of G. Then the following are equivalent.
(1) H is stable in G.
(2) H is hyperbolic and strongly quasiconvex in G
(3) H is hyperbolic and the lower relative divergence of G with respect
to H is completely super linear.
We now study some basic results about strongly quasiconvex subgroups
related to subgroup inclusion and subgroup intersection.
Lemma 4.9. (Proposition 9.4, [Hru10]). Let G be a group with a finite
generating set S. Suppose xH and yK are arbitrary left cosets of subgroups
of G. For each constant L there is a constant L′ = L′ (G, S, xH, yK, L) so
that in the metric space (G, dS ) we have
NL (xH) ∩ NL (yK) ⊂ NL′ (xHx−1 ∩ yKy −1 ).
Proposition 4.10. Let G be a finitely generated group, A a subgroup of G,
and B a subgroup of A. Then
(1) If A is finitely generated, undistorted in G and B is strongly quasiconvex in G, then B is strongly quasiconvex in A.
(2) If B is strongly quasiconvex in A and A is strongly quasiconvex in
G, then B is strongly quasiconvex in G
Proof. We note that if subgroups A is undistorted in G, then B is undistorted
in A if and only if B is undistorted in G. Therefore, the above proposition
is a direct result of Proposition 4.2.
ON STRONGLY QUASICONVEX SUBGROUPS
22
We remark that Statement (1) in Proposition 4.10 can be strengthen by
the following proposition.
Proposition 4.11. Let G be a finitely generated group and A undistorted
subgroup of G. If H is a strongly quasiconvex subgroup of G, then H1 =
H ∩ A is a strongly quasiconvex subgroup of A. In particular, H1 is finitely
generated and undistorted in A.
Proof. We fix finite generating sets S and T for G and A respectively. Let
f : Γ(A, T ) → Γ(G, S) be a quasi-isometry embedding which is an extension
of the inclusion A ֒→ G. We will prove that H1 is a Morse subset of Γ(A, T ).
For each K ≥ 1 and L ≥ 0 let α be an arbitrary (K, L)–quasi-geodesic in
Γ(A, T ) that connects two points in H1 . Since f : Γ(A, T ) → Γ(G, S) is a
quasi-isometry embedding which is an extension of the inclusion A ֒→ G,
then f (α) is a (K ′ , L′ )–quasigeodesic in Γ(G, S) that connects two points in
H1 and f (α) lies in some D–neighborhood of A in Γ(G, S), where K ′ , L′ , D
depend only on K, L and the map f . Since H is a Morse subset in Γ(G, S),
f (α) also lies in some D1 –neighborhood of H in Γ(G, S), where D1 depends
only on K ′ and L′ . By Lemma 4.9, there is M = M (G, S, A, H, D, D1 ) such
that f (α) lies in the M –neighborhood of H1 = H ∩ A in Γ(G, S). Again,
f : Γ(A, T ) → Γ(G, S) is a quasi-isometry embedding which is an extension
of the inclusion A ֒→ G. Then, α lies in some M ′ –neighborhood of H1 in
Γ(A, T ), where M ′ depend only on M and the map f .
The following proposition is a direct result of Propositions 4.10 and 4.11.
Proposition 4.12. Let G be a finitely generated group and H1 , H2 strongly
quasiconvex subgroups of G. Then H1 ∩ H2 is strongly quasiconvex in H1 ,
H2 , and G.
Quasiconvex subgroups in hyperbolic groups have well-known properties
that they have finite height and therefore, they have finite index in their
commensurators. We will show that strongly quasiconvex subgroups in general also have these properties.
Definition 4.13. Let G be a group and H a subgroup.
(1) Conjugates g1 Hg1−1 , · · · gk Hgk−1 are essentially distinct if the cosets
g1 H, · · · , gk H are distinct.
(2) H has height at most n in G if the intersection of any (n + 1) essentially distinct conjugates is finite. The least n for which this is
satisfied is called the height of H in G.
(3) The width of H is the maximal cardinality of the set {gi H : |gi Hgi −1 ∩ gj Hgj −1 | =
∞}, where {gi H} ranges over all collections of distinct cosets.
Similarly, given a finite collection H = {H1 , · · · , Hl }.
(1) Conjugates g1 Hσ(1) g1−1 · · · gk Hσ(k) gk−1 are essentially distinct if the
cosets g1 Hσ(1) , · · · , gk Hσ(k) are distinct.
ON STRONGLY QUASICONVEX SUBGROUPS
23
(2) The finite collection H of subgroups of G has height at most n if the
intersection of any (n + 1) essentially distinct conjugates is finite.
The least n for which this is satisfied is called the height of H in G.
(3) The width of H is the maximal cardinality of the set
−1
−1
| = ∞},
∩ gσ(j) Hσ(j) gσ(j)
{gσ(i) Hσ(i) : |gσ(i) Hσ(i) gσ(i)
where {gσ(i) Hσ(i) } ranges over all collections of distinct cosets.
Definition 4.14. Let G be a group and H a subgroup. The commensurator
of H in G, denoted CommG (H) is defined as
CommG (H) = { g ∈ G | [H : H ∩ gHg−1 ] < ∞, [gHg−1 : H ∩ gHg−1 ] < ∞ }.
Theorem 4.15. Let H = {H1 , · · · , Hℓ } be a finite collection of strongly
quasiconvex subgroups of a finitely generated group G. Then H has finite
height.
Proof. By Proposition 3.7 there is a constant D > 0 such that the following
hold. If γ : [0, ∞) → Γ(G, S) is a geodesic ray in the Cayley graph Γ(G, S)
that lies in some finite neighborhood of Hi , then there is C > 0 such that
γ|[C,∞) lies in the D–neighborhood of Hi . Let g1 Hσ(1) g1−1 · · · gk Hσ(k) gk−1 be
T
essential distinct conjugates with infinite intersection gi Hσ(i) gi−1 .
T
Since the Cayley graph Γ(G, S) is proper and gi Hσ(i) gi−1 is an infinite strongly quasiconvex subgroups, there is a geodesic ray γ : [0, ∞) →
Γ(G,
the Cayley graph Γ(G, S) that lies in some finite neighborhood of
T S) in −1
gi Hσ(i) gi . Therefore, for each i the ray γ lies in some finite neighborhood
of gi Hσ(i) . This implies that the ray gi−1 γ lies in some finite neighborhood
of Hσ(i) . Thus, there is some Ci > 0 such that gi−1 γ|[Ci ,∞) lies in the D–
neighborhood of Hσ(i) . In other word, γ|[Ci ,∞) lies in the D–neighborhood
of gi Hσ(i) . Therefore, all left cosets gi Hσ(i) intersect some ball B(a, D). But
there is a uniform bound N on the number of cosets of subgroups in H
intersecting any metric ball of radius D. Thus, the collection H has finite
height.
Corollary 4.16. If H is an infinite strongly quasiconvex subgroup of a
finitely generated group G, then H has finite index in its commensurator
CommG (H).
We know prove that a strongly quasiconvex subgroup have bounded packing and finite widith. We first recall the concept of bounded packing subgroups in [HW09].
Definition 4.17. Let G be a finitely generated group and Γ a Cayley graph
with respect to a finite generating set. A subgroup H has bounded packing
in G if, for all D ≥ 0, there exist N ∈ N such that for any collection of N
distinct cosets gH in G, at least two are separated by a distance of at least
D.
ON STRONGLY QUASICONVEX SUBGROUPS
24
Similarly, a finite collection H = {H1 , · · · , Hℓ } of subgroups of G has
bounded packing if, for all D ≥ 0, there exist N ∈ N such that for any
collection of N distinct cosets gHi with Hi ∈ H, at least two are separated
by a distance of at least D.
Lemma 4.18. Let (X, d) be a geodesic space. Let A, B, and C be Morse
subsets of X with a Morse gauge µ. For each D > 0 there is a number D ′
depending on D and µ such that if A, B, and C are pairwise D-close, then
the intersection ND′ (A) ∩ ND′ (B) ∩ ND′ (C) is not empty.
Proof. Let
x ∈ ND (A) ∩ ND (B)
y ∈ ND (B) ∩ ND (C)
z ∈ ND (C) ∩ ND (A).
Let u be a point in a geodesic [y, z] connecting y and z such that d(x, u) =
d x, [y, z] . Let γ1 be the concatenation of two geodesics [x, u] and [u, y].
Similarly, let γ2 be the concatenation of two geodesics [x, u] and [u, z]. Then
γ1 and γ2 are both (3, 0)–quasigeodesic. The proof of this claim is quite
elementary and the reader can also see the proof of this claim in the proof
of Lemma 2.2 in [Cor]. Since γ1 , γ2 , and [y, z] are all (3, 0)–quasigeodesic
and A, B, and C are all µ–Morse subsets of X, then there is a number D ′
depending on D and µ such that
γ1 ⊂ ND′ (B)
γ2 ⊂ ND′ (A)
[y, z] ⊂ ND′ (C)
In particular, u belongs to the intersection ND′ (A) ∩ ND′ (B) ∩ ND′ (C).
Therefore, this intersection is not empty.
Lemma 4.19 (Lemma 4.2 in [HW09]). Suppose H ≤ G has height 0 < n <
∞ in G. Choose g ∈ G so that gH 6= H, and let K = H ∩ gHg −1 . Then K
has the height less than n in H.
Theorem 4.20. Let H be a strongly quasiconvex subgroup of a finitely generated group G. Then H has bounded packing in G.
The proof of the above theorem follows the same line of argument as
Theorem 4.8 in [HW09]. At some point in the proof of Theorem 4.8 in
[HW09], Hruska-Wise need to use the thinness of a triangle in a hyperbolic
space. However, due to the lack of hyperbolicity in group G and subgroup
H in the above theorem, we need to use Lemma 4.18 instead.
Proof. Fix finite generating sets S and T for G and H respectively. By
Theorem 4.15, we know that the height of H in G is finite. We will prove
the theorem by induction on height. If the height of H in G is zero, then
H is a finite group. Therefore, H has bounded packing by Corollary 2.6
ON STRONGLY QUASICONVEX SUBGROUPS
25
in [HW09]. We now assume by induction that the theorem holds for every
finitely generated group G0 and strongly quasiconvex subgroup H0 with the
height of H0 in G0 less than the height of H in G.
Let H be a set of left cosets gH which are pairwise D-close. We are going
to prove that the cardinality of H is bounded by a number depending on
D. We can translating H so we may assume that H ∈ H. Observe that if
d(gH, H) < D then gH = hxH for some h ∈ H and x ∈ B(1, D). It follows
that the left cosets gH intersecting ND (H) lie in at most |B(1, D)| distinct
H–orbits. Thus it suffices to bound the number of elements of H in the
orbit H(gH) for each fixed g ∈
/ H.
If we let K = H ∩ gHg−1 , then Lemma 4.19 shows that the height of K
in H is less than the height of H in G. Since K is a strongly quasiconvex
subgroup H, the inductive hypothesis applied to K ≤ H gives for each D ′
a number M ′ = M ′ (D ′ ) so that any collection of M ′ distinct cosets hK in
H contains a pair separated by a dT –distance at least D ′ . Furthermore,
the proof of Lemma 4.2 in [HW09] shows that there is a well-defined map
hgH → hK taking left cosets of H in the orbit of gH to left cosets of K. A
similar argument shows that this map is bijective.
In order to complete the proof, we will show that D–closeness of distinct
cosets h1 gH and h2 gH in (G, dS ) implies D ′ –closeness of the corresponding
cosets h1 K and h2 K in (H, dT ) for some D ′ depending on D. The proof
of this step in Theorem 4.8 in [HW09] requires the thinness of a triangle.
However, we need to use Lemma 4.18 for our situation. In fact, there is
D1 depending on D such that ND1 (H) ∩ ND1 (h1 gH) ∩ ND1 (h2 gH) contains
an element u. By Lemma 4.9, there is number D2 depending on D1 but is
independent of the choice of hi ∈ H such that
ND1 (H) ∩ ND1 (hi gH) ⊂ ND2 (hi K).
Therefore, u ∈ ND2 (h1 K) ∩ ND2 (h2 K). In other word, dS (h1 K, h2 K) < D2 .
Since H is an undistorted subgroup of G, then dT (h1 K, h2 K) < D ′ for some
D ′ depending on D, as desired.
Theorem 4.21. Let H = {H1 , · · · , Hℓ } be a finite collection of stable subgroups of a finitely generated group G. Then H has bounded packing in G.
Further, H has finite width.
Proof. First, let M be a collection of coset of subgroups in H which are
pairwise D–close. By Theorem 4.20, for each i = 1, 2, · · · , ℓ there is a Ni ≥ 0
such that the number of coset of Hi in M is at most Ni . Therefore, the
number of elements of M is at most ΣNi . Thus, H has bounded packing.
Arguing as in the proof of Theorem 4.15, we see that any two conjugates
of subgroups in H with infinite intersection have cosets uniformly close together. Therefore, H has finite width.
ON STRONGLY QUASICONVEX SUBGROUPS
26
5. Morse boundaries and strong quasiconvexity
We recall basic properties of a quasiconvex subgroup in a hyperbolic group
that the subgroup is also hyperbolic groups and the subgroup inclusion
induces a topological embedding of the Gromov boundary of the subgroup
into the Gromov boundary of the ambient group. Moreover, the image of
this topological embedding is identical to the limit set of the subgroup inside
the ambient group. Also, the intersection of two quasiconvex subgroups is
again a quasiconvex subgroup and the limit set of the intersection of two
quasiconvex subgroups is equal to the intersection of limit sets of the two
subgroups. In this section, we prove some analogous properties for strongly
quasiconvex subgroups in general.
We first review the concept of Morse boundary in [Cor].
Definition 5.1. Let M be the set of all Morse gauges. We put a partial
ordering on M so that for two Morse gauges N, N ′ ∈ M, we say N ≤ N ′ if
and only if N (K, L) ≤ N ′ (K, L) for all K, L.
Definition 5.2. Let X be a proper geodesic space. The Morse boundary of
X with basepoint p, denoted ∂M Xp , is defined to be the set of all equivalence
classes of Morse geodesic rays in X with initial point p, where two rays
α, α′ : [0, ∞) → X are equivalent if there exists a constant K such that
dX α(t), α(t) < K for all t > 0. We denote the equivalence class of a ray
α in ∂M Xp by [α].
On ∂M Xp , we could build a topology as follows:
Consider the subset of the Morse boundary
N
∂M
Xp = { x | The class x contains at least an N –Morse geodesic rays α with α(0) = p }.
N X by: x → x as n → ∞ if and only if there
We define convergence in ∂M
p
n
exists N –Morse geodesic rays αn with αn (0) = p and [αn ] = xn such that
every subsequence of αn contains a subsequence that converges uniformly
on compact sets to a geodesic ray α with [α] = x. The closed subsets F in
∂M Xp are those satisfying the condition
{xn } ⊂ F and xn → x ⇒ x ∈ F.
We equip the the Morse boundary ∂M Xp with the direct limit topology
∂M Xp = lim
∂N X .
−
→ M p
M
Let A be a subset of X with basepoint p. The limit set ΛA of A in ∂M Xp
is the set of all points c in ∂M Xp represented by rays based at p that lie in
some finite neighborhood of A.
Remark 5.3. The direct limit topology on ∂M Xp is independent of basepoint p (see Proposition 3.5 in [Cor]). Therefore, we can assume the basepoint is fixed, suppress it from the notation and write ∂M X. Moreover,
Morse boundary is a quasi-isometry invariant (see Proposition 3.7 [Cor]).
Therefore, we define the Morse boundary of a finitely generated group G,
ON STRONGLY QUASICONVEX SUBGROUPS
27
denoted ∂M G, as the Morse boundary of its Cayley graph. We also define
the limit set of a subgroup H of G in ∂M G accordingly.
We define an action of G on ∂M G as follows. For each element g in G
and [α] in ∂M G, g[α] = [β], where α and β are two rays at the basepoint in
some Cayley graph of G such that the Hausdorff distance between gα and
β is finite.
Definition 5.4. Let X and Y be proper geodesic metric spaces and p ∈ X,
p′ ∈ Y . We say that f : ∂M Xp → ∂M Yp′ is Morse preserving if given N in
N X to ∂ N ′ Y ′ .
M there exists an N ′ in M such that f injectively maps ∂M
p
M p
The following proposition shows that if a quasi-isometric embedding defines a strongly quasiconvex subspace then it induces a Morse preserving
map. This proposition is a key lemma for the proof of the fact that inclusion map of a strongly quasiconvex subgroup into a finitely generated group
induces a topological embedding on Morse boundaries.
Proposition 5.5. Let X, A be two proper geodesic spaces and A is strongly
quasiconvex in X via the map f . Then for each Morse gauge N there is another Morse gauge N ′ such that for every N –Morse geodesic ray γ : [0, ∞) →
A there is an N ′ –Morse geodesic ray with basepoint f γ(0) in X bounded
Hausdorff distance from f (γ) (.i.e. f induces a Morse preserving map).
Proof. Since f is a quasi-isometry embedding, f (γ) is a (K, L)–quasi-geodesic,
where K, L only depends on the quasi-isometry embedding constants of f .
Also f (γ) is Morse with the Morse gauge N1 depending only on the Morse
gauge N and the map f (see the proof of Statement (2) of Proposition
4.2). Since X is a proper geodesic space and f (γ) is a (K, L)–Morse quasi
geodesic with Morse gauge N1 , there is a geodesic ray α based at f γ(0) in
X bounded Hausdorff distance D from f (γ) where D only depends on K,
L, and the Morse gauge N1 . This implies that D only depends on the Morse
gauge N and the map f . Since the Hausdorff distance between f (γ) and α
is D and f (γ) is N1 –Morse, the ray α is N ′ –Morse where N ′ only depends
on D and the Morse gauge N1 . Therefore, N ′ only depends on the Morse
gauge N and the map f .
Theorem 5.6. Let X, A be two proper geodesic spaces and A is strongly
quasiconvex in X via the map f . Then f induces a topological embedding
∂M f : ∂M A → ∂M X such that ∂M f (∂M A) = Λf (A).
Proof. The fact that f induces a topological embedding ∂M f : ∂M A → ∂M X
is a direct result of Proposition 5.5 via Proposition 4.2 in [Cor]. However,
we need to check that ∂M f (∂M A) = Λf (A).
We first recall the construction of the map ∂M f from Proposition 4.2
in [Cor]. Fix a point p in A and let q = f (p). By Proposition 5.5, for
each Morse gauge N there is another Morse gauge N ′ such that for every
N –Morse geodesic ray α based at p there is an N ′ –Morse geodesic ray β
with basepoint q in X bounded Hausdorff distance from f (α). We define
ON STRONGLY QUASICONVEX SUBGROUPS
28
∂M f [α] = [β], then ∂M f : ∂M A → ∂M X is a topological embedding (see
the proof of Proposition 4.2 in [Cor]). Obviously, ∂M f (∂M A) ⊂ Λf (A) by
the construction. We now prove the opposite inclusion.
Let c be an arbitrary element in Λf (A). There is a Morse geodesic ray
β1 based at q in X such that c = [β1 ] and β1 lies in some finite neighborhood
of f (A). Since f is a quasi-geometric embedding, there is a quasi-geodesic
α1 based at p in A such that the Hausdorff distance between f (α1 ) and
β1 is finite. Therefore, f (α1 ) is also a Morse quasi-geodesic. This implies
that α1 is also a Morse quasi-geodesic in A by Statement (1) of Proposition
4.2. Also, A is a proper geodesic space. Then there is a Morse geodesic ray
α based at p in X such that the Hausdorff distance between α and α1 is
finite. Thus, the Hausdorff distance between f (α) and f (α1 ) is also finite.
This implies that the Hausdorff distance between
f (α) and β1 is finite. By
the construction of ∂M f we have ∂M f [α] = [β] for some geodesic ray β
based at q such that the Hausdorff distance between f (α) and β is finite.
Therefore, Hausdorff distance
between β and β1 is finite. This implies that
c = [β1 ] = [β] = ∂M f [α] ∈ ∂M f (∂M A). Therefore, ∂M f (∂M A) = Λf (A).
We now come up with the main theorem of this section.
Theorem 5.7. Let G be a finitely generated group. Then
(1) If H is a finitely generated strongly quasiconvex subgroup of G, then
the inclusion i : H ֒→ G induces a topological embedding î : ∂M H →
∂M G such that î(∂M H) = ΛH.
(2) If H1 and H2 are finitely generated strongly quasiconvex subgroups
of G, then H1 ∩ H2 is strongly quasiconvex in G and Λ(H1 ∩ H2 ) =
ΛH1 ∩ ΛH2 .
Proof. The Statement (1) is a direct result of Theorem 5.6. Therefore, we
only need to prove Statement (2). The fact that H1 ∩H2 is strongly quasiconvex in G is a result of Proposition 4.12. Since H1 ∩ H2 ⊂ H1 , Λ(H1 ∩ H2 ) ⊂
ΛH1 . Similarly, Λ(H1 ∩ H2 ) ⊂ ΛH2 . Therefore, Λ(H1 ∩ H2 ) ⊂ ΛH1 ∩ ΛH2 .
We now let c be an arbitrary element in ΛH1 ∩ ΛH2 . Then there is geodesic
ray γ based at e in some Cayley graph Γ(G, S) such that γ lies in some
finite neighborhoods of H1 and H2 . We need to prove that γ lies in some
finite neighborhood of H1 ∩ H2 . Since γ lies in some finite neighborhoods
of H1 and H2 , there is an infinite sequence (gn ) of G–vertices on γ and
two group elements u, v such that gn u ∈ H1 and gn v ∈ H2 . Therefore,
gn g1−1 = (gn u)(g1 u)−1 = (gn v)(g1 v)−1 ∈ H1 ∩ H2 . This implies that (gn ) lies
in some finite neighborhood of H1 ∩ H2 . Also H1 ∩ H2 is Morse in Γ(G, S)
and all elements in (gn ) are pairwise different. Therefore, γ lies in some finite
neighborhood of H1 ∩ H2 . Thus, c is also an element in Λ(H1 ∩ H2 ).
ON STRONGLY QUASICONVEX SUBGROUPS
29
6. Strong quasiconvexity and stability in relatively
hyperbolic groups
In this section, we investigate strongly quasiconvex subgroups and stable
subgroups in relatively hyperbolic groups. We first recall the concepts of
coned off Cayley graphs and relative hyperbolic groups.
Definition 6.1. Given a finitely generated group G with Cayley graph
Γ(G, S) equipped with the path metric and a finite collection P of subgroups
of G, one can construct the coned off Cayley graph Γ̂(G, S, P) as follows: For
each left coset gP where P ∈ P, add a vertex vgP , called a peripheral vertex,
to the Cayley graph Γ(G, S) and for each element x of gP , add an edge
e(x, gP ) of length 1/2 from x to the vertex vgP . This results in a metric
space that may not be proper (i.e. closed balls need not be compact).
Definition 6.2 (Relatively hyperbolic group). A finitely generated group
G is hyperbolic relative to a finite collection P of subgroups of G if the coned
off Cayley graph is δ–hyperbolic and fine (i.e. for each positive number n,
each edge of the coned off Cayley graph is contained in only finitely many
circuits of length n).
Each group P ∈ P is a peripheral subgroup and its left cosets are peripheral
left cosets and we denote the collection of all peripheral left cosets by Π.
We now review some known results on relatively hyperbolic subgroups including undistorted subgroups and peripheral subgroups in relatively hyperbolic groups, relations between quasigeodesics in Cayley graphs and coned
off Cayley graphs.
Theorem 6.3. [Hru10, Theorem 1.5 and Theorem 9.1] Let (G, P) be a
finitely generated relatively hyperbolic group and let H be a finitely generated
undistorted subgroup of G. Let S be some (any) finite generating set for G.
Then
(1) There is a constant A = A(S) such that for each geodesic c in the
coned off Cayley graph Γ̂(G, S, P) connecting two points of H, every
G–vertex of c lies within the A–neighborhood of H with respect to
the metric dS .
(2) All subgroups of H of the form H ∩ P g where g ∈ G, P ∈ P and
|H ∩ P g | = ∞ lie in only finitely many conjugacy classes in H. Furthermore, if O is a set of representatives of these conjugacy classes
then (H, O) is relatively hyperbolic.
Theorem 6.4. [DS05, Corollary 1.14] A relatively hyperbolic group (G, P)
is hyperbolic if each peripheral subgroup in P is hyperbolic.
Lemma 6.5. [Hru10, Lemma 8.8] Let (G, P) be a finitely generated relatively
hyperbolic group with a finite generating set S. For each K ≥ 1 and L ≥
0 there is A = A(K, L) > 0 such that the following holds. Let α be a
(K, L)–quasigeodesic in Γ(G, S) and c a geodesic in Γ̂(G, S, P) with the same
ON STRONGLY QUASICONVEX SUBGROUPS
30
endpoints in G. Then each G–vertex of c lies in the A–neighborhood of some
vertex of α with respect to the metric dS .
Lemma 6.6. [DS05, Lemma 4.15] Let (G, P) be a finitely generated relatively
hyperbolic group. Then each conjugate P g of peripheral subgroup in P is
strongly quasiconvex in G. Moreover, each D–neighborhood of peripheral
left coset is M –Morse in Cayley graph of G where the Morse gauge M only
depends on D and the choice of finite generating set of G.
The following theorem provides with characterizations of strongly quasiconvex subgroups in relatively hyperbolic groups and this is the main theorem of this section.
Theorem 6.7. Let (G, P) be a finitely generated relatively hyperbolic group
and H finitely generated undistorted subgroup of G. Then the following are
equivalent:
(1) The subgroup H is strongly quasiconvex in G.
(2) The subgroup H ∩P g is strongly quasiconvex in P g for each conjugate
P g of peripheral subgroup in P.
(3) The subgroup H ∩ P g is strongly quasiconvex in G for each conjugate
P g of peripheral subgroup in P.
Proof. The implications “(1) ⇒ (2) ⇒ (3)” are direct results from Propositions 4.10, 4.12, and Lemma 6.6. We now prove the implication “(3) ⇒ (1)”.
Let S be a finite generating set of G and let A0 = A0 (S) be the constant in
Statement (1) of Theorem 6.3.
Let K ≥ 1 and L ≥ 0 arbitrary. Let A = A(K, L) be constant in Lemma
6.5. By Lemma 4.9 there is a constant A1 > 0 such that for each t ∈
B(e, A0 ) ∩ G and P ∈ P
NA (tP ) ∩ NA+A0 (H) ⊂ NA1 (tP t−1 ∩ H).
Let µ be the maximum of all Morse gauge functions of all set of the form
NA1 (tP t−1 ∩ H) where |t|S < A0 and P ∈ P.
Let
D = K(2A + 1) + L + µ(K, L) + A0 + A + A1 + 1.
Let α : [a, b] → Γ(G, S) be an arbitrary (K, L)–quasigeodesic in Γ(G, S)
that connects two points h1 and h2 in H. By Lemma 1.11 of [BH99] III.H,
we can assume that α is continuous and
ℓ(α|[t,t′ ] ) ≤ KdS α(t), α(t′ ) + L.
We will show that α lies in the D–neighborhood of H. Let c be a geodesic in
Γ̂(G, S, P) that connects two points h1 and h2 and let h1 = s0 , s1 , · · · , sn =
h2 be all G–vertices of c. By the choice of A0 , all G–vertices si lie in the
A0 –neighborhood
of H. Also by the choice of A, there is ti ∈ [a, b] such that
dS si , α(ti ) < A. We consider t0 = s0 and tn = sn . Therefore, each point
α(ti ) lies in the (A0 + A)–neighborhood of H. It is sufficient to show that
ON STRONGLY QUASICONVEX SUBGROUPS
31
each α(Ii ) lies in the D–neighborhood of H where Ii is a closed subinterval
of [a, b] with endpoints ti−1 and ti . We consider two cases.
Case 1: Two G-vertices si−1 and si are adjacent in Γ(G, S). Then the
distance between two points α(ti−1 ) and α(ti ) with respect to the metric dS
is bounded above by 2A + 1. Therefore,
ℓ(α|Ii ) ≤ KdS α(ti−1 ), α(ti ) + L ≤ K(2A + 1) + L.
Also, α(ti ) lies in the (A + A0 )–neighborhood of H. Therefore, α(Ii ) lies in
the D–neighborhood of H by the choice of D.
Case 2: Two G-vertices si−1 and si are not adjacent in Γ(G, S). Therefore, si−1 and si both lie in the same peripheral left coset gP . Since
dS (gP, H) ≤ dS (si , H) < A0 , gP = htP for some h ∈ H and |t|S < A0 .
Observe that the endpoints of α(Ii ) both lies in NA (htP ) ∩ NA+A0 (H).
Therefore, the endpoints of h−1 α(Ii ) both lies in NA (tP ) ∩ NA+A0 (H) ⊂
NA1 (tP t−1 ∩ H). Since h−1 α is also a (K, L)–quasigeodesic, h−1 α(Ii ) lies
entirely in the µ(K, L)–neighborhood of the set NA1 (tP t−1 ∩ H). Therefore,
h−1 α(Ii ) lies in (µ(K, L) + A1 )–neighborhood of the
set H. Translating by
h, we see that α(Ii ) also lies in the µ(K, L) + A1 –neighborhood of the set
H. By the choice of D, α(Ii ) lies in D–neighborhood of the set H.
Therefore, subgroup H is strongly quasiconvex in G.
An immediate corollary is characterizations of stable subgroups in relatively hyperbolic groups.
Corollary 6.8. Let (G, P) be a finitely generated relatively hyperbolic group
and H finitely generated undistorted subgroup of G. Then the following are
equivalent:
(1) The subgroup H is stable in G.
(2) The subgroup H ∩ P g is stable in P g for each conjugate P g of peripheral subgroup in P.
(3) The subgroup H ∩ P g is stable in G for each conjugate P g of peripheral subgroup in P.
Proof. We can see that (1) implies (2) and (1) implies (3) easily. In fact, if
the subgroup H is stable in G, then H is strongly quasiconvex in G and H
is hyperbolic (see Theorem 4.8). Therefore, H ∩ P g is strongly quasiconvex
in H, P g , and G by Proposition 4.12 and Lemma 6.6. In particular, H ∩ P g
is hyperbolic because H is a hyperbolic group. Therefore, H ∩ P g is stable
in P g and G by Theorem 4.8. The equivalence between (2) and (3) can also
be seen easily from Theorems 4.8 and 6.7.
We now prove that (3) implies (1). By the hypothesis, H ∩ P g is strongly
quasiconvex in G for each conjugate P g of peripheral subgroup in P. Therefore, H is strongly quasiconvex in G by Theorem 6.7. Also H ∩ P g is
hyperbolic for each conjugate P g of peripheral subgroup in P. Therefore,
the subgroup H is also hyperbolic by Statement (2) of Theorem 6.3 and
Theorem 6.4. Therefore, the subgroup H is stable in G by Theorem 4.8.
ON STRONGLY QUASICONVEX SUBGROUPS
32
7. Strong quasiconvexity, stability in right-angled Coxeter
groups and related lower relative divergence
In this section, we investigate strongly quasiconvex subgroups and stable subgroups in right-angled Coxeter groups. More precisely, we characterize strongly quasiconvex (stable) special subgroups in two dimensional
right-angled Coxeter groups. Also for each integer d ≥ 2 we construct a
right-angled Coxeter group together with a non-stable strongly quasiconvex
subgroup whose lower relative divergence is exactly polynomial of degree
d. We remark that these are also first examples of groups together with
non hyperbolic subgroups whose lower relative divergence are polynomial of
degree d (d ≥ 2).
7.1. Some backgrounds in right-angled Coxeter groups. This subsection provides with background on right-angled Coxeter groups.
Definition 7.1. Given a finite, simplicial graph Γ, the associated rightangled Coxeter group GΓ has generating set S the vertices of Γ, and relations
s2 = 1 for all s in S and st = ts whenever s and t are adjacent vertices.
Let S1 be a subset of S. The subgroup of GΓ generated by S1 is a rightangled Coxeter group GΓ1 , where Γ1 is the induced subgraph of Γ with
vertex set S1 (i.e. Γ1 is the union of all edges of Γ with both endpoints
in S1 ). The subgroup GΓ1 is called a special subgroup of GΓ . Any of its
conjugates is called a parabolic subgroup of GΓ .
Definition 7.2. Given a finite, simplicial graph Γ, the associated Davis
complex ΣΓ is a cube complex constructed as follows. For every k–clique,
T ⊂ Γ, the special subgroup GT is isomorphic to the direct product of k
copies of Z2 . Hence, the Cayley graph of GT is isomorphic to the 1–skeleton
of a k–cube. The Davis complex ΣΓ has 1–skeleton the Cayley graph of GΓ ,
where edges are given unit length. Additionally, for each k–clique, T ⊂ Γ,
and coset gGT , we glue a unit k–cube to gGT ⊂ ΣΓ . The Davis complex ΣΓ
is a CAT(0) cube complex and the group GΓ acts properly and cocompactly
on the Davis complex ΣΓ (see [Dav08]).
Remark 7.3. Let Γ be a finite simplicial graph with vertex set S and Γ1
the induced subgraph of Γ generated by some subset T of S. Then the Davis
complex ΣΓ1 embeds isometrically in the Davis complex ΣΓ . Moreover, the
(1)
Cayley graph ΣΓ1 of the group GΓ1 embeds isometrically in the Cayley
(1)
graph ΣΓ of the group GΓ .
We also remark that a Davis complex ΣΓ is a two dimensional cube complex if and only if the defining graph Γ is triangle free and contains at least
one edge.
Definition 7.4. Let Σ be a CAT(0) cube complex. In Σ, we consider the
equivalence relation on the set of midcubes of cubes generated by the rule
that two midcubes are related if they share a face. A hyperplane, H, is the
ON STRONGLY QUASICONVEX SUBGROUPS
33
union of the midcubes in a single equivalence class. We define the support
of a hyperplane H, denoted N (H), to be the union of cubes which contain
midcubes of H.
Remark 7.5. Let Γ be a finite simplicial graph with vertex set S and ΣΓ
the associated Davis complex. Each hyperplane in ΣΓ separates ΣΓ into
two convex sets. It follows that the distance between a pair of vertices with
respect to the metric dS equals the number of hyperplanes in ΣΓ separating
those vertices.
For a generator v, let ev denote the edge from the basepoint 1 to the
vertex v. Any edge in ΣΓ determines a unique hyperplane, namely the hyperplane containing the midpoint of that edge. Denote by Hv the hyperplane
containing the midpoint of ev .
For a cube in ΣΓ , all of the parallel edges are labeled by the same generator
v. It follows that all of the edges crossing a hyperplane H have the same label
v, and we call this a hyperplane of type v. Obviously, if two hyperplanes
with the types v1 and v2 cross, then v1 and v2 commute. Since GΓ acts
transitively on edges labeled v, a hyperplane is of type v if and only if it
is a translate of the standard hyperplane Hv . Obviously, the star subgroup
GSt(v) is the stabilizer of the hyperplane Hv and GSt(v) can also be considered
as vertices of the support N (Hv ) of Hv . Therefore, the subgroup gGSt(v) g−1
is the stabilizer of the hyperplane gHv and gGSt(v) can also be considered
as vertices of the support N (gHv ) of gHv .
7.2. Strongly quasiconvex special subgroup and stable special subgroup characterizations in right-angled Coxeter groups. In this subsection, we establish characterizations of strongly quasiconvex special subgroups and stable special subgroups in two dimensional right-angled Coxeter
groups in term of defining graphs. As a consequence, we give a criterion on
simplical, triangle free graphs that guarantee whose associated right-angled
Coxeter groups possess Morse boundaries which are not totally disconnected.
We first develops some results in right-angled Coxeter groups that help
establish the characterizations of strongly quasiconvex (stable) special subgroups in two dimensional right-angled Coxeter groups.
Lemma 7.6. Let Γ be a simplicial graph with vertex set S. Let K be a special
subgroup of GΓ generated by some subset S1 of S and Γ1 the subgraph of Γ
(1)
induced by S1 . For each g ∈ GΓ γ is the shortest path in ΣΓ that connects g
(1)
to some point in K if and only if γ is a geodesic in ΣΓ and each hyperplane
(1)
crossed by some edge of γ does not intersect the Cayley graph ΣΓ1 of K.
(1)
Proof. We first assume that γ is the shortest path in ΣΓ that connects g
(1)
to some point h in K. Obviously, γ must be a geodesic in ΣΓ connecting
h and g. Let n be the length of γ. Then, the geodesic γ along the direction
from h to g is the concatenation of n edges e1 e2 · · · en where each edge ei is
labelled by some vertex si in S. Therefore, g = h(s1 s2 · · · sn ).
ON STRONGLY QUASICONVEX SUBGROUPS
34
Assume for the contradiction that there is a hyperplane Hi crossed by the
(1)
edge ei intersects ΣΓ1 . We choose the smallest number i with that property.
In particular, si ∈ S1 ⊂ K. We claim that for each j < i each hyperplane
Hj crossed by edge ej of γ intersects Hi . Indeed, each hyperplane Hj (j < i)
(1)
(1)
does not intersect ΣΓ1 by the choice of i. Moreover, ΣΓ1 is connected then
(1)
(1)
there is a path α in ΣΓ1 that connects h and a point u in Hi ∩ ΣΓ1 . Let
β is the path in Hi that connects u to the midpoint of ei . Therefore, each
hyperplane Hj (j < i) must intersect the concatenation αβ. However, each
hyperplane Hj (j < i) does not intersect α because it does not intersect
(1)
ΣΓ1 . This implies that each hyperplane Hj (j < i) intersects Hi . Therefore,
sj commutes to si for each j < i. Thus,
g = h(s1 s2 · · · si−1 si si+1 · · · sn ) = (hsi )(s1 s2 · · · si−1 si+1 · · · sn ).
The above equality shows that there is path of length n − 1 that connects g
to the point hsi ∈ K. This is a contradiction. Therefore, each hyperplane
(1)
crossed by some edge of γ does not intersect the Cayley graph ΣΓ1 .
(1)
We now assume that γ is a geodesic in ΣΓ and each hyperplane crossed
(1)
by some edge of γ does not intersect the Cayley graph ΣΓ1 . Let γ ′ be an
(1)
(1)
arbitrary path in ΣΓ that connects g to some point h1 in K. Since ΣΓ1
(1)
is connected, there is a path α′ in ΣΓ1 that connects h and h1 . Therefore,
hyperplanes crossed by edges of γ intersect the concatenation α′ γ ′ . Also,
(1)
these hyperplanes do not intersect α′ since they do not intersect ΣΓ1 . Therefore, all these hyperplanes must intersect γ ′ . This implies that ℓ(γ) ≤ ℓ(γ ′ ).
(1)
Therefore, γ is the shortest path in ΣΓ that connects g to some point in
K.
Lemma 7.7. Let Γ be a simplicial graph with vertex set S. Let K be a special
subgroup of GΓ generated by some subset S1 of S. Let σ be an induced
4–cycle with two pairs of non-adjacent vertices (a1 , a2 ) and (b
1 , b2 ) such
that a1 , a2 both lie in S1 but b1 does not. Then dS (b1 b2 )n , K = 2n and
dS (b1 b2 )n b1 , K = 2n + 1.
Proof. We will only prove that dS (b1 b2 )n , K = 2n and the argument for
the remaining equality is identical. Let Γ1 the subgraph of Γ induced by
(1)
S1 . Let g = (b1 b2 )n and γ the geodesic in ΣΓ connecting the identity
element e and g. Therefore, the geodesic γ along the direction from e to
g traces the word (b1 b2 )n and let Hi be the hyperplane crossed by the ith
edge of γ. Since the hyperplane H1 is labelled by b1 ∈
/ S1 , H1 does not
(1)
intersect ΣΓ1 . Also the geodesic γ is labelled by non-commuting elements
b1 and b2 alternatively. This implies that Hi does not intersect Hj for i 6= j.
In particular, Hi does not intersect H1 for i 6= 1. This implies that Hi
(1)
does not intersect ΣΓ1 for i 6= 1 either. By Lemma 7.6, the geodesic γ is
the shortest path that connects (b1 b2 )n and some point in K. Therefore,
ON STRONGLY QUASICONVEX SUBGROUPS
35
dS (b1 b2 )n , K = ℓ(γ) = 2n. Using a similar argument we also obtain the
equality dS (b1 b2 )n b1 , K = 2n + 1.
Proposition 7.8. Let Γ be a simplicial graph with vertex set S. Let K be
a special subgroup of GΓ generated by some subset S1 of S. Assume that
there is an induced 4–cycle σ with two pairs of non-adjacent vertices (a1 , a2 )
and (b1 , b2 ) such that a1 , a2 both lie in S1 but b1 does not. Then K is not
strongly quasiconvex in GΓ .
Proof. By Theorem 4.7, this is sufficient to prove that the lower relative
divergence of GΓ with respect to K is not completely super linear. Let {σρn }
(1)
be the lower relative divergence of ΣΓ with respect to K. We claim that
for each n ≥ 2 and ρ ∈ (0, 1]
σρn (r) ≤ (4n + 2)r for each r > 1.
Indeed, for each r > 1 choose an integer m in [r, 2r]. If m = 2k for some
integer k, we choose x = (b1 b2 )k . Otherwise, m = 2k + 1 for some integer k
and we choose x = (b1 b2 )k b1 . Therefore, dS (x, K) = m by Lemma 7.7. We
choose y = (a1 a2 )mn x. Then
dS (y, K) = dS (a1 a2 )mn x, K = dS x, (a1 a2 )−mn K = dS (x, K) = m.
Since x commutes with both a1 and a2 , y = x(a1 a2 )mn . This implies that
there is a geodesic γ connecting x, y and γ traces the word (a1 a2 )mn . Obviously, dS (x, y) = ℓ(γ) = 2mn.
By the construction,
each vertex of γ is element of the form x(a1 a2 )i or
i
x (a1 a2 ) a1 in GΓ . Again, x commutes with both a1 and a2 . Therefore,
dS x(a1 a2 )i , K = dS (a1 a2 )i x, K = dS x, (a1 a2 )−i K = dS (x, K) = m > r
and dS x (a1 a2 )i a1 , K = m > r similarly. Therefore, the path γ lies
outside the r–neighborhood of K.
(1)
Since r ≤ dS (x, K) = m ≤ 2r, there is a path γ1 in ΣΓ connecting x and
some point u in ∂Nr (K) such that γ1 lies outside the r–neighborhood of K
and its length is bounded above by r. In particular, dS (x, u) ≤ r. Similarly,
(1)
there is a path γ2 in ΣΓ connecting y and some point v in ∂Nr (K) such that
γ2 lies outside the r–neighborhood of K and its length is bounded above by
r. In particular, dS (x, u) ≤ r. Therefore,
dS (u, v) ≥ dS (x, y) − dS (x, u) − dS (y, v) ≥ 2mn − r − r ≥ (2n − 2)r ≥ nr.
We observe that γ = γ1 ∪ γ ∪ γ2 is the path outside the r–neighborhood
of K that connects two points u, v in ∂Nr (K). Also,
ℓ(γ) = ℓ(γ1 ) + ℓ(γ) + ℓ(γ2 ) ≤ r + 2mn + r ≤ (4n + 2)r.
Therefore, σρn (r) ≤ (4n + 2)r for each r > 1. This implies that the lower
relative divergence of GΓ with respect to K is not completely super linear.
Therefore, K is not strongly quasiconvex in GΓ by Theorem 4.7.
ON STRONGLY QUASICONVEX SUBGROUPS
36
Proposition 7.9. Let Γ be a simplicial, triangle free graph with vertex set
S and K subgroup of GΓ generated by some subset S1 of S. We assume that
if S1 contains two non-adjacent vertices of an induced 4–cycle σ, then S1
contains all vertices of σ. If K is an infinite subgroup of GΓ , then the lower
relative divergence of GΓ with respect to K is at least quadratic
(1)
Proof. Let {σρn } be the lower relative divergence of ΣΓ with respect to K.
We claim that for each n ≥ 3 and ρ ∈ (0, 1]
σρn (r) ≥ (r − 1)(ρr − 1) for each r > 0.
In fact, if σρn (r) = ∞, then the above inequality is true obviously. Otherwise,
let γ be an arbitrary path outside Nρr (K) that connects two points x and y
(1)
in ∂Nr (K) such that dS (x, y) ≥ nr. Let γ1′ be a shortest geodesic in ΣΓ that
connects x and some point h1 in K. Similarly, let γ2′ be a shortest geodesic
(1)
in ΣΓ that connects y and some point h2 in K. Therefore, dS (x, h1 ) =
dS (y, h2 ) = r obviously. This implies that
dS (h1 , h2 ) ≥ dS (x, y) − dS (x, h1 ) − dS (y, h2 ) ≥ (n − 2)r ≥ r.
Let Γ1 the subgraph of Γ induced by S1 . We connect h1 and h2 by
(1)
a geodesic α = e1 e2 · · · em of length m in the Cayley graph ΣΓ1 of K.
Obviously m ≥ r and each edge ei of α is labelled by some vertex si in S1 .
Let Hi be the hyperplane crossed by the edge ei of α. Then each Hi must
intersect γ1′ ∪ γ ∪ γ2′ . However, Hi does not intersect γ1′ ∪ γ2′ by Lemma 7.6.
Therefore, each hyperplane Hi must intersect some edge fi of γ.
Let h0 , h1 , h2 , · · · , hm be all vertices of α such that hi−1 and hi are endpoints of ei . For 1 ≤ i ≤ m let gi be the endpoint of the edge fi of γ that lies
in the same component of ΣΓ − Hi containing gi . Let αi be a the geodesic in
the support of Hi that connects hi and gi . By the convexity of the graph of
(1)
(1)
ΣΓ1 in the graph ΣΓ each αi is the concatenation of two geodesic βi and βi′
(1)
(1)
such that βi lies entirely in ΣΓ1 and βi′ intersects ΣΓ1 only at the common
endpoint h′i ∈ K of βi and βi′ .
We set mi = ℓ(βi′ ) for 1 ≤ i ≤ m. Then mi ≥ ρr since gi lies outside the
ρr–neighborhood of K. We assume that βi′ = ei1 ei2 · · · eimi then each edge eij
of βi′ is labelled by an element in Lk(si ) and the first edge ei1 is labelled by
a vertex not in S1 . We claim that for each 1 ≤ i ≤ m − 1 there is at least
(mi − 1) hyperplanes crossed by βi′ intersect the subpath γi of γ connecting
gi , gi+1 . Since Γ is triangle free, no pair of different vertices in Lk(si ) are
adjacent. This implies that no pairs of hyperplanes crossed by two different
edges of βi′ intersects. Since the first edge ei1 of βi′ is labelled by a vertex not
(1)
in S1 , the hyperplane crossed by ei1 does not in intersect ΣΓ1 . Therefore, no
(1)
subsequent hyperplane crossed by βi′ intersects ΣΓ1 either. We now consider
′ .
the possibility a hyperplane crossed by βi′ intersects βi+1
ON STRONGLY QUASICONVEX SUBGROUPS
37
If the hyperplane crossed by the first edge ei1 of βi′ does not in intersect
′
no subsequent hyperplane crossed by βi′ intersects βi+1
either. There′
fore, all mi hyperplanes crossed by βi intersect the subpath γi of γ. If the
′ , then edge ei is labelled by a vertex
hyperplane crossed by ei1 intersect βi+1
1
a in Lk(si ) ∩ LK(si+1 ). Since Γ is triangle free, two vertices si and si+1 are
not adjacent. We claim that the hyperplane crossed by the second edge ei2
′ . Otherwise, the second edge ei of β ′ is labelled
of βi′ does not intersect βi+1
2
i
by a vertex b other than a in Lk(si ) ∩ LK(si+1 ). Therefore, four points a,
b, si , and si+1 are all vertices of an induced 4–cycle σ in Γ. Moreover, two
non-adjacent vertices si and si+1 both lie in S1 but a does not. This is a
contradiction. Therefore, the hyperplane crossed by the second edge ei2 of
′ . This also implies that hyperplanes crossed edge
βi′ does not intersect βi+1
′
′ . Thus, all these hyperplanes
i
ej (2 ≤ j ≤ m) of βi does not intersect βi+1
must intersect the subpath γi of γ. Therefore, there is exactly (mi − 1)
hyperplanes crossed by βi′ intersect the subpath γi of γ.
For 1 ≤ i ≤ m − 1 the length of the subpath γi of γ is at least (mi − 1)
because there is at least (mi − 1) hyperplanes crossed by βi′ intersect the
subpath γi . This implies that
′ ,
βi+1
ℓ(γ) ≥ Σm−1
i=1 ℓ(γi ) ≥ (m − 1)(mi − 1) ≥ (r − 1)(ρr − 1).
Therefore, σρn (r) ≥ (r − 1)(ρr − 1) for each r > 0. This implies that the
lower relative divergence of GΓ with respect to K is at least quadratic.
The following theorem establishes characterizations of strongly quasiconvex subgroups in two dimensional right-angled Coxeter groups. This theorem is a direct result of Propositions 7.8, 7.9 and Theorem 4.7. This is also
the main theorem of this section.
Theorem 7.10. Let Γ be a simplicial, triangle free graph with vertex set S
and K subgroup of GΓ generated by some subset S1 of S. Then the following
conditions are equivalent:
(1) The subgroup K is strongly quasiconvex in GΓ .
(2) If S1 contains two non-adjacent vertices of an induced 4–cycle σ,
then S1 contains all vertices of σ.
(3) Either |K| < ∞ or the lower relative divergence of GΓ with respect
to K is at least quadratic.
Question 7.11. Is lower relative divergence of a CAT(0) group with respect
to an infinite strongly quasiconvex subgroup at least quadratic?
The following corollary is a direct result of Theorem 7.10. The corollary
establishes characterization of stable subgroups in two dimensional rightangled Coxeter groups.
Corollary 7.12. Let Γ be a simplicial, triangle free graph with vertex set S
and K subgroup of GΓ generated by some subset S1 of S. Then the following
conditions are equivalent:
ON STRONGLY QUASICONVEX SUBGROUPS
38
(1) The subgroup K is stable in GΓ .
(2) The set S1 does not contain a pair of non-adjacent vertices of an
induced 4–cycle in Γ.
Proof. Assume that S1 does not contain any pair of non-adjacent vertices of
an induced 4-cycle of Γ. Then the subgroup K is hyperbolic (see Corollary
12.6.3 in [Dav08]). Also the subgroup K is strongly quasiconvex in GΓ by
Theorem 7.10. Therefore, K is stable subgroup by Theorem 4.8.
We now assume that S1 contains a pair of non-adjacent vertices of an
induced 4-cycle σ of Γ. We will prove that K is not stable in GΓ . If K
is not strongly quasiconvex in GΓ , then K is not stable in GΓ by Theorem
4.8. Otherwise, S1 contains all vertices of the induced 4–cycle σ by Theorem
7.10. This implies that K is not hyperbolic by Corollary 12.6.3 in [Dav08].
Thus, K is not stable in GΓ .
Question 7.13. Can we characterize all strongly quasiconvex (stable) subgroups of right-angled Coxeter groups using defining graphs?
We guess that the recent work of Abbott-Behrstock-Durham [ABD] can
help us characterize all stable subgroups of right-angled Coxeter groups using defining graphs. However, characterizing all strongly quasiconvex subgroups of right-angled Coxeter groups using defining graphs is still a difficult
question.
The following corollary give a criterion on two dimensional right-angled
Coxeter groups to possess Morse boundaries which are not totally disconnected. We remark that the first example of a right-angled Coxeter group
whose Morse boundary is not totally disconnected was constructed by Behrstock in [Beh]. The following corollary generalizes his example in some sense.
Corollary 7.14. If the simplicial, triangle free graph Γ contains an induced
loops σ of length greater than 4 such that the vertex set of σ does not contain
a pair of non-adjacent vertices of an induced 4–cycle in Γ, then the Morse
boundary of right-angled Coxeter group GΓ is not totally disconnected.
Proof. Since the vertex set S1 of the loop σ satisfies hypothesis of Corollary
7.12, the subgroups Gσ generated by S1 is a stable subgroup in GΓ . Also,
the Morse boundary of Gσ is topologically embedded in the Morse boundary
of GΓ (see Theorem 5.6) and the Morse boundary of Gσ is a topological
circle. This implies that the Morse boundary of GΓ also contains a circle
and therefore, it is not totally disconnected.
Conjecture 7.15. The Morse boundary of a right-angled Coxeter group GΓ
is not totally disconnected if and only if the defining graph Γ contains an
induced loops σ of length greater than 4 such that the vertex set of σ does
not contain a pair of non-adjacent vertices of an induced 4–cycle in Γ.
7.3. Higher relative lower divergence in right-angled Coxeter groups.
In this subsection, we construct right-angled Coxeter groups together with
non-stable strongly quasiconvex subgroups whose lower relative divergence
ON STRONGLY QUASICONVEX SUBGROUPS
39
are arbitrary polynomials of degrees at least 2. We remark that the author
in [Tra] also constructed various examples of right-angled Coxeter groups
whose lower relative divergence with respect to some subgroups are arbitrary polynomials of degrees at least 2. However, all subgroups on those
examples are stable and in this subsection we work on the cases of nonstable subgroups.
We first establish connection between lower relative divergence of certain
pairs of right-angled Coxeter groups and divergence of some certain geodesics
in their Davis complexes. This connection is a key ingredient for main
examples in this subsection.
Lemma 7.16. Let Γ be a simplicial graph with the vertex S and let u, v
be two non-adjacent vertices of Γ. Let Ω be a new graph by coning off two
points u, v of Γ with a new vertex t and let S = S ∪ {t}. Let x, y be
(1)
endpoints of a path γ in ΣΩ − N1 (GΓ ). If g1 , g2 be two elements in GΓ
such that dS (x, GΓ ) = dS (x, g1 ) and dS (y, GΓ ) = dS (x, g2 ), then g1−1 g2 lies
in the subgroup I generated by u and v.
Proof. We can assume that x and y both lie in GΩ . Let γ1 be a geodesic
(1)
(1)
in ΣΩ connecting x, g1 and γ2 be a geodesic in ΣΩ connecting y, g2 . We
can assume that γ1 along the direction from g1 to x is the concatenation
(1)
γ1 = e1 e2 · · · em of edges of ΣΩ . Similarly, γ2 along the direction from g2 to
(1)
y is the concatenation γ2 = f1 f2 · · · fn of edges of ΣΩ . We claim that two
edges e1 and f1 are crossed by the same hyperplane which is labelled by t.
We first recall that GΓ is a subgroup generated by S and S − S = {t}.
Also γ1 is the shortest geodesic that connects g1 and GΓ . Then, the first
edge e1 of γ1 is labelled by t and we call H the hyperplane crossed by e1 .
By Remark 7.5 the vertex set of the support of H is g1 GSt(t) . Also, each
(1)
element in g1 GSt(t) has distance 0 or 1 from GΓ . This implies that H ∩ ΣΩ
lies in the 1–neighborhood of GΓ . In particular, H does not intersect γ.
(1)
Also, H does not intersect the Cayley graph ΣΓ of GΓ . Thus, H intersects
the path γ2 . Moreover, H must intersect the first edge f1 of γ2 because
(1)
H ∩ ΣΩ lies in the 1–neighborhood of GΓ . This fact implies that g1 , g2
are both vertices of the support of H. In other word, g1 and g2 lie in the
same left coset of GSt(t) or g1−1 g2 lies in GSt(t) . Also, g−1 g2 lies in GΓ and
GΓ ∩ GSt(t) = I which is the subgroup generated by u and v. Thus, g1−1 g2
lies in the subgroup I.
Lemma 7.17. Let Γ be a simplicial graph with the vertex S and let u, v
be two non-adjacent vertices of Γ. Let Ω be a new graph by coning off two
points u, v of Γ with a new vertex t and let S = S ∪{t}. Let I be the subgroup
(1)
generated by u and v. Then for each x in ΣΓ dS (tx, GΓ ) = dS (x, I) + 1.
Proof. We assume that x lies in GΓ and let n = dS (x, I). Let γ be a geodesic
(1)
of length n in ΣΓ that connects x to some point h in I. We can express
ON STRONGLY QUASICONVEX SUBGROUPS
40
(1)
γ = e1 e2 · · · en as a concatenation of edges in ΣΓ , where each edge ei is
labelled by some element si ∈ S. Therefore, x = h(s1 s2 · · · sn ). Let γ1 = tγ
then γ1 is a geodesic that connects th and tx. Moreover, we can express
γ1 = f1 f2 · · · fn where each edge fi is the translation of edge ei of γ by t. In
particular, each edge fi is also labelled by si . Since th = ht, there is an edge
e′1 labelled by t with endpoints h and th. We construct γ = e′1 ∪ γ1 and we
claim that γ is the shortest geodesic that connects tx and GΓ .
(1)
(1)
(1)
We first prove that γ is a geodesic in ΣΩ . Since the inclusion ΣΓ ֒→ ΣΩ
(1)
is an isometric embedding, γ is also a geodesic in ΣΩ . Therefore, γ1 = tγ
(1)
is a geodesic in ΣΩ . Since γ1 is labelled by elements in S and the first edge
e′1 of γ is labelled by t ∈
/ S, the hyperplane crossed by e′1 does not intersect
(1)
γ1 . Therefore, γ is a goedesic in ΣΩ .
We now prove that γ is the shortest geodesic that connects tx and GΓ .
By Lemma 7.6, it is sufficient to prove that each hyperplane crossed by some
(1)
edge of γ does not intersect the Cayley graph ΣΓ of GΓ . Since GΓ is the
subgroup generated by S and the hyperplane H crossed by the first edge e′1
(1)
of γ is labelled by t, H does not intersect ΣΓ . Assume for the contradiction
(1)
that hyperplane Hi crossed by some edge fi of γ intersects ΣΓ . We choose i
is the smallest number with that property. In particular, Hi must intersect
H and therefore, si commutes to t. This implies that si = u or si = v.
Therefore, si is an element in I.
By the choice of i, each hyperplane Hj crossed by edge fj (j < i) does
(1)
not intersect ΣΓ . Therefore, each Hj must intersect Hi . This implies that
si commutes to all sj for j < i. Therefore,
x = h(s1 s2 · · · si−1 si si+1 · · · sn ) = (hsi )(s1 s2 · · · si−1 si+1 · · · sn ).
(1)
This implies that there is a path in ΣΓ with length n − 1 that connects
x to the element hsi ∈ I which is a contradiction. Thus, each hyperplane
(1)
crossed by some edge of γ does not intersect the Cayley graph ΣΓ of GΓ .
Therefore, the path γ is the shortest geodesic that connects tx and GΓ . This
implies that
dS (tx, GΓ ) = ℓ(γ) = ℓ(γ1 ) + 1 = ℓ(γ) + 1 = dS (x, I) + 1.
Proposition 7.18. Let Γ be a simplicial graph with the vertex S and let
u, v be two non-adjacent vertices of Γ. Let Ω be a new graph by coning off
two points u, v of Γ with a new vertex t and let S = S ∪ {t}. Let α be the
bi-infinite geodesic containing the identity e labelled by u and v alternatively.
Then
Σ
(1)
Σ
(1)
Divα Ω div(GΩ , GΓ ) Divα Γ ,
Σ
(1)
Σ
(1)
where Divα Ω and Divα Γ
respectively.
(1)
(1)
are geodesic divergence of α in ΣΩ and ΣΓ
ON STRONGLY QUASICONVEX SUBGROUPS
41
Proof. Let I be a subgroup generated by u and v. Then the geodesic α
(1)
(1)
is the Cayley graph of I in ΣΩ and ΣΓ . Therefore, the lower relative
(1)
(1)
divergence of ΣΩ and ΣΩ with respect to the path α are equivalent to the
lower relative divergence of these spaces with respect to I. Also by Remark
2.14, the lower relative divergence of a geodesic space with respect to a
periodic bi-infinite geodesic is equivalent to the divergence of the geodesic
in the space. Therefore, this is sufficient to prove
(1)
(1)
(1)
div(ΣΩ , I) div(ΣΩ , GΓ ) div(ΣΓ , I).
(1)
Let {σρn } be the lower relative divergence of ΣΩ with respect to GΓ , let
(1)
{σρn } be the lower relative divergence of ΣΩ with respect to I, and let {σρn }
(1)
be the lower relative divergence of ΣΓ with respect to I. We will prove that
for each n ≥ 2 and ρ ∈ (0, 1]
σρn (r) ≤ σρn (r) ≤ σρ2n (r) + 2r for each r > 1.
We first prove the left inequality. If σρn (r) = ∞, the left inequality is true
obviously. Otherwise, let x, y be an arbitrary point in ∂Nr (GΓ ) such that
(1)
there is a path outside Nr (GΓ ) in ΣΩ that connects x and y and dS (x, y) ≥
nr. Let g1 , g2 be two elements in GΓ such that dS (x, GΓ ) = dS (x, g1 ) and
dS (y, GΓ ) = dS (x, g2 ). Then g1 , g2 both lie in the same left coset gI of the
subgroup I by Lemma 7.16.
(1)
(1)
(1)
Since element g in GΩ acts isometrically on ΣΩ , div(ΣΩ , gI) = div(ΣΩ , I) =
{σρn }. This is obvious that
dS (x, gI) = dS (x, GΓ ) = r and dS (y, gI) = dS (y, GΓ ) = r.
Also, any path that lies outside some s–neighborhood of GΓ also lies outside
the s–neighborhood of gI because gI is a subset of GΓ . Therefore, σρn (r) ≤
(1)
(1)
σρn (r). This implies that div(ΣΩ , I) div(ΣΩ , GΓ ).
We now prove the second inequality σρn (r) ≤ σρ2n (r)+ 2r for each r > 1. If
σρ2n (r) = ∞, then the inequality is true obviously. Otherwise, let x1 , y1 be
(1)
arbitrary two points in ∂Nr (I) ⊂ ΣΓ such that there is a path outside Nr (I)
(1)
(1)
connecting x1 , y1 and dS (x1 , y1 ) ≥ (2n)r. Since the inclusion ΣΓ ֒→ ΣΩ
is an isometric embedding, dS (x1 , y1 ) = dS (x1 , y1 ) ≥ (2n)r.
(1)
Let β be an arbitrary path outside Nρr (I) in ΣΓ connecting x1 and y1 .
Then the path β1 = tβ connects two points tx1 and ty1 . By Lemma 7.17,
(1)
β1 lies outside the (ρr + 1)–neighborhood of GΓ in ΣΩ and
dS (tx1 , GΓ ) = dS (x1 , I) = r + 1 and dS (ty1 , GΓ ) = dS (y1 , I) = r + 1.
This implies that there is a geodesic β2 with length 1 that lies outside the
(1)
r–neighborhood of GΓ in ΣΩ and β2 connects tx1 with some point x2 ∈
Nr (GΓ ). Similarly, there is a geodesic β3 with length 1 that lies outside
ON STRONGLY QUASICONVEX SUBGROUPS
a0
a0
a1 a2 a3
b1
42
ad−1 ad
b0
a1 a2 a3
b1
ad−1
ad
b0
bd−1 bd
b2 b3 b4
bd−1
b2 b3 b4
t
bd
Γd
Ωd
Figure 1. Graph Ωd is constructed from graph Γd by coning
off two non-adjacent vertices ad , bd by a new vertex t.
(1)
the r–neighborhood of GΓ in ΣΩ and β3 connects ty1 with some point
y2 ∈ Nr (GΓ ). By Lemma 7.17, this is not hard to see there is a path outside
(1)
the r–neighborhood of GΓ in ΣΩ that connects tx1 and ty1 . Therefore, we
(1)
can construct another path outside the r–neighborhood of GΓ in ΣΩ that
connects x2 and y2 obviously. Moreover,
dS (x2 , y2 ) ≥ dS (tx1 , ty1 ) − dS (tx1 , x2 ) − dS (ty1 , y2 )
≥ dS (x1 , y1 ) − dS (tx1 , x2 ) − dS (ty1 , y2 )
≥ 2nr − 1 − 1 ≥ nr.
Also β = β2 ∪ β1 ∪ β3 is the path outside the ρr–neighborhood of GΓ connecting x2 and y2 . Therefore,
σρn (r) ≤ ℓ(β) ≤ ℓ(β2 ) + ℓ(β1 ) + ℓ(β3 ) ≤ 1 + ℓ(β) + 1 ≤ ℓ(β) + 2r.
(1)
(1)
This implies that σρn (r) ≤ σρ2n (r)+2r. Therefore, div(ΣΩ , GΓ ) div(ΣΓ , I).
Thus,
Σ
(1)
Σ
(1)
Divα Ω div(GΩ , GΓ ) Divα Γ .
We are now ready to construct the main examples for this subsection and
we start with defining graphs. More precisely, for each integer d ≥ 2 a graph
Ωd and its subgraph Γd are constructed as in Figure 1. We remark that the
graph Γd was originally constructed by Dani-Thomas [DT15a] to study divergence of right-angled Coxeter groups and the graph Ωd is a variation of
Γd . However, we are going to use these pairs of graphs to construct corresponding pairs of right-angled Coxeter groups with desired lower relative
divergence in this paper.
Proposition 7.19 (Proposition 3.19 in [Tra16]). For each d ≥ 2, let Γd be
the graph as in Figure 1. Let αd be a bi-infinite geodesic containing e and
(1)
labeled by · · · ad bd ad bd · · · . Then the divergence of αd in ΣΓd is equivalent
to the polynomial of degree d.
ON STRONGLY QUASICONVEX SUBGROUPS
43
We remark that the above proposition was proved in [Tra16] but the proof
based mostly on the work of Dani-Thomas in [DT15a]. Using a similar
argument as in Proposition 3.19 in [Tra16] we also obtain an analogous
proposition as follows.
Proposition 7.20. For each d ≥ 2, let Ωd be the graph as in Figure 1.
Let αd be a bi-infinite geodesic containing e and labeled by · · · ad bd ad bd · · · .
(1)
Then the divergence of αd in ΣΩd is equivalent to the polynomial of degree
d.
We now state the main theorem of this subsection.
Theorem 7.21. For each d ≥ 2 we construct the graph Ωd and the subgraph
Γd as in Figure 1. Then GΓd is a non-stable strongly quasiconvex subgroup of
GΩd and the lower relative divergence of GΩd with respect to GΓd is equivalent
to the polynomial of degree d.
Proof. By Propositions 7.18, 7.19, and 7.20, the lower relative divergence
of GΩd with respect to GΓd is equivalent to the polynomial of degree d.
In particular, GΓd is a strongly quasiconvex subgroup of GΩd by Theorem
4.7. However, the subgroup GΓd is not hyperbolic because the graph Γd
contains an induced 4–cycle. This implies that GΓd is not a stable subgroup
of GΩd .
8. Strong quasiconvexity, stability, and lower relative
divergence in right-angled Artin groups
In this section, we prove that two notions of strong quasiconvexity and stability are equivalent in the right-angled Artin group AΓ (except for the case
of finite index subgroups). We also characterize non-trivial strongly quasiconvex subgroups of infinite index in AΓ (i.e. non-trivial stable subgroups
in AΓ ) by quadratic lower relative divergence. These results strengthen the
work of Koberda-Mangahas-Taylor [KMT] on characterizing purely loxodromic subgroups of right-angled Artin groups.
8.1. Some backgrounds in right-angled Artin groups.
Definition 8.1. Given a finite simplicial graph Γ, the associated rightangled Artin group AΓ has generating set S the vertices of Γ, and relations
st = ts whenever s and t are adjacent vertices.
Let S1 be a subset of S. The subgroup of AΓ generated by S1 is a rightangled Artin group AΓ1 , where Γ1 is the induced subgraph of Γ with vertex
set S1 (i.e. Γ1 is the union of all edges of Γ with both endpoints in S1 ). The
subgroup AΓ1 is called a special subgroup of AΓ .
A reduced word for a group element g in AΓ is a minimal length word in
the free group F (S) representing g.
Definition 8.2. Let Γ1 and Γ2 be two graphs, the join of Γ1 and Γ2 is a
graph obtained by connecting every vertex of Γ1 to every vertex of Γ2 by an
edge.
ON STRONGLY QUASICONVEX SUBGROUPS
44
Let J be an induced subgraph of Γ which decomposes as a nontrivial join.
We call AJ a join subgroup of AΓ . A reduced word w in AΓ is called a join
word if w represents element in some join subgroup. If β is a subword of w,
we will say that β is a join subword of w when β is itself a join word.
Definition 8.3. Let Γ be a simplicial, finite, connected graph such that Γ
does not decompose as a nontrivial join. A group element g in AΓ is loxodromic if g is not conjugate into a join subgroup. If every nontrivial group
element in a subgroup H of AΓ is loxodromic, then H is purely loxodromic.
Definition 8.4. Let Γ be a finite simplicial graph with the set S of vertices.
Let T be a torus of dimension |S| with edges labeled by the elements of S.
Let XΓ denote the subcomplex of T consisting of all faces whose edge labels
span a complete subgraph in Γ (or equivalently, mutually commute in AΓ ).
XΓ is called the Salvetti complex.
Remark 8.5. The fundamental group of XΓ is AΓ . The universal cover
X̃Γ of XΓ is a CAT(0) cube complex with a free, cocompact action of AΓ .
(1)
Obviously, the 1-skeleton X̃Γ of X̃Γ is the Cayley graph of AΓ with respect
to the generating set S. Consequently, reduced words in AΓ correspond
(1)
to geodesics in X̃Γ , which we call combinatorial geodesics. We refer to
(1)
distance in X̃Γ as combinatorial distance. If Γ1 is an induced subgraph of
(1)
a Γ, then there is a natural isometric embeddings of the 1-skeleton X̃Γ1 of
(1)
X̃Γ1 into the 1-skeleton X̃Γ of X̃Γ .
Theorem 8.6 (Theorem 1.1, Theorem 5.2, and Corollary 6.2 in [KMT]).
Let Γ be a simplicial, finite, connected graph such that Γ does not decompose
as a nontrivial join. Let H be a finitely generated subgroup of AΓ . Then the
following are equivalent.
(1) H is a purely loxodromic subgroup.
(2) H is stable.
(3) There exists a positive number N = N (H) such that for any reduced
word w representing h ∈ H, and any join subword w′ of w, we have
ℓ(w′ ) ≤ N .
The following corollary is a direct consequence of the above theorem.
Corollary 8.7. Let Γ be a simplicial, finite, connected graph with the vertex
set S such that Γ does not decompose as a nontrivial join. Let H be a nontrivial finitely generated purely loxodromic subgroup. Then there is a positive
(1)
constant M such that every geodesic in the Cayley graph X̃Γ connecting
points in H lies in the M –neighborhood of H.
The properties of hyperplanes of X̃Γ in the following remark were observed
by Behrstock-Charney in [BC12].
Remark 8.8. Each hyperplane in X̃Γ separates X̃Γ into two convex sets.
It follows that the combinatorial distance between a pair of vertices equals
the number of hyperplanes in X̃Γ separating those vertices.
ON STRONGLY QUASICONVEX SUBGROUPS
45
For a generator v, let ev denote the edge from the basepoint 1 to the
vertex v. Any edge in X̃Γ determines a unique hyperplane, namely the hyperplane containing the midpoint of that edge. Denote by Hv the hyperplane
containing the midpoint of ev .
For a cube in X̃Γ , all of the parallel edges are labeled by the same generator v. It follows that all of the edges crossing a hyperplane H have the same
label v, and we call this a hyperplane of type v. Since AΓ acts transitively
on edges labeled v, a hyperplane is of type v if and only if it is a translate
of the standard hyperplane Hv .
For a vertex v of the graph Γ let lk(v) denote the subgraph of Γ spanned
by the vertices adjacent to v and let st(v) denote the subgraph spanned by
v and lk(v). Obviously, ASt(v) is a join subgroup and Alk(v) is a subgroup
of ASt(v) . We call such subgroup ASt(v) star subgroup.
Lemma 8.9 (Lemma 3.1 in [BC12]). Let H1 = g1 Hv and H2 = g2 Hw . Then
(1) H1 intersects H2 if and only if v, w commute and g1−1 g2 ∈ Alk(v) Alk(w) .
(2) There is a hyperplane H3 intersecting both H1 and H2 if and only if
there is u in st(v) ∩ st(w) such that g1−1 g2 ∈ Alk(v) Alk(u) Alk(w) .
8.2. Lower relative divergence of a right-angled Artin group with
respect to a purely loxodromic subgroup. In this section, we compute
the lower relative divergence of a right-angled Artin group with respect to
a purely loxodromic subgroup (i.e. a stable subgroup). We first prove the
quadratic upper bound for the lower relative divergence of a right-angled
Artin group with respect to a loxodromic subgroup and we need the following
lemmas.
Lemma 8.10. Let Γ be a simplicial, finite, connected graph with the vertex
set S such that Γ does not decompose as a nontrivial join. Let H be a
non-trivial, finitely generated purely loxodromic subgroup of AΓ . There is
a positive number K such that for each element g in AΓ and each pair of
commuting generators (s1 , s2 ) in S
|i| + |j|
− |g|S − 1.
K
Proof. By Theorem 8.6, there is a positive integer N such that for any
reduced word w representing h ∈ H, and any join subword w′ of w, we have
ℓ(w′ ) ≤ N . Let K = N + 1 and we will prove that
dS (gsi1 sj2 , H) ≥
dS (gsi1 sj2 , H) ≥
|i| + |j|
− |g|S − 1.
K
Let m = dS (gsi1 sj2 , H). Then there is an element g1 in AΓ with |g1 |S = m
and h in H such that h = gsi1 sj2 g1 . Since si1 sj2 is an element in some join
subgroup of AΓ and |g1 |S = m, then h can be represented by a reduced word
w that is a product of at most (|g|S + 1 + m) join subwords. Also, the length
of each join subword of w is bounded above by N . Therefore, the length of
ON STRONGLY QUASICONVEX SUBGROUPS
46
w is bounded above by N |g|S + m + 1 . Also,
ℓ(w) ≥ |si1 sj2 |S − |g1 |S − |g|S ≥ |i| + |j| − m − |g|S .
This implies that
|i| + |j| − m − |g|S ≤ N |g|S + m + 1 .
Therefore,
dS (gsi1 sj2 , H) = m ≥
N
|i| + |j|
|i| + |j|
− |g|S −
≥
− |g|S − 1.
N +1
N +1
K
Lemma 8.11. Let Γ be a simplicial, finite, connected graph with the vertex
set S such that Γ does not decompose as a nontrivial join. Let H be a
on-trivial, finitely generated purely loxodromic subgroup of AΓ and h an
arbitrary element in H. There is a number L ≥ 1 such that for each positive
integer m ≥ L2 and generator s in S there is a path α outside the (m/L−L)–
neighborhood of H connecting sm and hsm with the length bounded above by
Lm.
Proof. Let M = diam Γ, K the positive integer as in Lemma 8.10 and k =
|h|S . Let L = 2(k + 1)(M + 1) + K + k + M + 1. Choose a reduced word
w = sǫ11 sǫ22 · · · sǫkk , where si ∈ S and ǫi ∈ {−1, 1}
that represents element h. For each i ∈ {1, 2, · · · , k} let ti be a vertex in
St(si ) and wi = sǫ11 sǫ22 · · · sǫi i . Then the length of each word wi is bounded
above by k, wi+1 = wi sǫi i , and wk = w that represents element h.
We now construct a path α0 outside the (m/L − L)–neighborhood of H
connecting sm and w1 tm
1 with the length bounded above by 2(M +1)m. Since
M = diam Γ, we can choose positive integer n ≤ M and n + 1 generators
u0 , u1 , · · · , un in S such that the following conditions hold:
(1) u0 = s and un = t1 .
(2) uj and uj+1 commutes where j ∈ {0, 1, 2, · · · , n − 1}.
m
For each j ∈ {0, 1, 2, · · · , n − 1} let βj be a path connecting um
j and uj+1 of
length 2m with vertices
m−1 m
m
m 2
m m
m
um
uj+1 , um−2
um
j , uj uj+1 , uj uj+1 , · · · , uj uj+1 , uj
j+1 , · · · , uj+1 .
j
By Lemma 8.10, the above vertices must lie outside the (m/K−1)–neighborhood
of H. Therefore, these vertices also lies outside the (m/L−L)–neighborhood
of H. Therefore, βj is a path outside the (m/L − L)–neighborhood of H
ǫ1 m
m ǫ1
m
m
connecting um
j and uj+1 . Since w1 t1 = s1 t1 = t1 s1 , then we can connect
m
tm
1 and w1 t1 by an edge βn labelled by s1 . Let α0 = β0 ∪ β1 ∪ · · · ∪ βn .
Then, it is obvious that the path α0 outside the (m/L − L)–neighborhood of
H connecting sm and w1 tm
1 with the length bounded above by 2(M + 1)m.
By similar constructions as above, for each i ∈ {1, 2, · · · , k − 1} there
is a path αi outside the (m/L − L)–neighborhood of H connecting wi tm
i
with
the
length
bounded
above
by
2(M
+
1)m.
We
can
also
and wi+1 tm
i+1
ON STRONGLY QUASICONVEX SUBGROUPS
47
construct a path αk outside the (m/L − L)–neighborhood of H connecting
m with the length bounded above by 2M m. Let α = α ∪
htm
0
k and hs
α1 ∪ · · · ∪ αk . Then, it is obvious that the path α outside the (m/L − L)–
neighborhood of H connecting sm and hsm with the length bounded above
by 2(k + 1)(M + 1)m. By the choice of L we observe that the length of α is
also bounded above by Lm.
Proposition 8.12. Let Γ be a simplicial, finite, connected graph with the
vertex set S such that Γ does not decompose as a nontrivial join. Let H be
a non-trivial, finitely generated purely loxodromic subgroup of AΓ . Then the
lower relative divergence of AΓ with respect to H is at most quadratic.
Proof. Let h be an arbitrary group element in H and L ≥ 1 a constant as in
Lemma 8.11. Since each cyclic subgroup in a CAT(0) group is undistorted
(see Corollary III.Γ.4.8 and Theorem III.Γ.4.10 in [BH99]), there is a positive
integer L1 such that
|hk |S ≥
|k|
− L1 for each integer k.
L1
(1)
Let {σρn } be the lower relative divergence of X̃Γ with respect to H. We
will prove that function σρn (r) is bounded above by some quadratic function
for each n ≥ 2 and ρ ∈ (0, 1].
Choose a positive integer m ∈ L(L + r), 2L(L + r) and a generator s
in S. Then, there is a path α0 outside the (m/L − L)–neighborhood of H
connecting sm and hsm with the length bounded above by Lm. It is obvious
that the path α0 also lies outside the r–neighborhood of H by the choice of
m. Choose a positive integer k which lies between L1 nr + 8L(L + r) + L1
and L1 nr + 8L(L + r) + L1 + 1 . Let α = α0 ∪ hα0 ∪ h2 α0 ∪ · · · ∪ hk−1 α0 .
Then, α is a path outside the r–neighborhood of H connecting sm , hk sm
with the length bounded above by kLm. By the choice of k and m, the
length of α is bounded above by 2L1 L2 (L + r) nr + 8L(L + r) + L1 + 1 .
Since r ≤ dS (sm , H) ≤ m, then there is a path γ1 outside Nr (H) connecting sm and some point x ∈ ∂Nr (H) such that the length of γ1 is
bounded above by m. By the choice of m, the length of γ1 is also bounded
above by 2L(L + r). Similarly, there is a path γ2 outside Nr (H) connecting hk sm and some point y ∈ ∂Nr (H) such that the length of γ2 is
bounded above by 2L(L + r). Let α = γ1 ∪ α ∪ γ2 then α is a path
outside Nr (H) connecting x, y and
the length of α is bounded above by
2L1 L2 (L+r) nr+8L(L+r)+L1 +1 +4L(L+r). Therefore, for each ρ ∈ (0, 1]
dρr (x, y) ≤ 2L1 L2 (L + r) nr + 8L(L + r) + L1 + 1 + 4L(L + r).
ON STRONGLY QUASICONVEX SUBGROUPS
48
Also,
dS (x, y) ≥ dS (sm , hk sm ) − dS (sm , x) − dS (hk sm , y)
k
≥ |hk |S − 2m − 2L(L + r) − 2L(L + r) ≥
− L1 − 8L(L + r)
L1
≥ nr + 8L(L + r) − 8L(L + r) ≥ nr.
Thus, for each ρ ∈ (0, 1]
σρn (r) ≤ 2L1 L2 (L + r) nr + 8L(L + r) + L1 + 1 + 4L(L + r).
This implies that the lower relative divergence of AΓ with respect to H
is at most quadratic.
We now establish the quadratic lower bound for the lower relative divergence of a right-angled Artin group with respect to a loxodromic subgroup.
Proposition 8.13. Let Γ be a simplicial, finite, connected graph with the
vertex set S such that Γ does not decompose as a nontrivial join. Let H be
a non-trivial, finitely generated purely loxodromic subgroup of AΓ . Then the
lower relative divergence of AΓ with respect to H is at least quadratic.
Proof. By Corollary 8.7, there is a positive integer M such that every geo(1)
desic in the Cayley graph X̃Γ connecting points in H lies in M –neighborhood
of H. By Theorem 8.6, there is another positive integer N such that for any
reduced word w representing h ∈ H, and any join subword w′ of w, we have
(1)
ℓ(w′ ) ≤ N . Let {σρn } be the lower relative divergence of X̃Γ with respect
to H. We will prove that for each n ≥ 9 and ρ ∈ (0, 1]
2M + 3N + 2
r−1
.
ρr − 3N ) − 2r for each r >
σρn (r) ≥
3N + 1
ρ
Let u and v be an arbitrary pair of points in ∂Nr (H) such that dr (u, v) <
∞ and dS (u, v) ≥ nr. Let γ be an arbitrary path that lies outside the ρr–
neighborhood of H connecting
u and v. We will prove that the length of γ
r−1
is bounded below by 3N
+1 ρr − 3N ) − 2r.
(1)
Let γ1 be a geodesic of length r in X̃Γ connecting u and some point x
(1)
in H. Let γ2 be another geodesic of length r in X̃Γ connecting v and some
(1)
point y in H. Let α be a geodesic in X̃Γ connecting x and y. Then α
lies in the M –neighborhood of H. Choose a positive integer m such that
r < (3N + 1)m + 1 < 2r.
Since dS (x, y) ≥ dS (u, v) − 2r ≥ (n − 2)r ≥ 7r, there is a subpath α1 with
length (3N +1)m+1 of α such that α1 ∩ B(x, 2r)∪B(y, 2r) = ∅. Also, the
lengths of γ1 and γ2 are both r. This implies that (γ1 ∪ γ2 ) ∩ Nr (α1 ) = ∅.
Since γ ∩ Nρr (H) = ∅ and α1 ⊂ NM (H), then γ ∩ Nρr−M (α1 ) = ∅. Also,
ρr − M > ρr/2. Thus, γ ∩ Nρr/2 (α1 ) = ∅. Let γ = γ1 ∪ γ ∪ γ2 . Then,
γ ∩ Nρr/2 (α1 ) = ∅.
ON STRONGLY QUASICONVEX SUBGROUPS
a
c
b
49
d
Figure 2
We assume that α1 = e0 w1 e1 w2 · · · em wm , where each ei is an edge labelled by some generator ai in S, each wi is a subpath of α1 of length
exactly 3N . For each i ∈ {0, 1, 2, · · · , m} let Hi be the hyperplane intersecting ei and vi a point in Hi ∩ γ. For each i ∈ {1, 2, · · · , m} let βi be the
subpath of γ connecting vi−1 and vi .
If there is some hyperplane of type b that intersects two hyperplanes
Hi−1 and Hi for some i ∈ {1, 2, · · · , m}, then ei−1 wi corresponds to a word
w that represents an element in Alk(ai−1 ) Alk(b) Alk(ai ) by Lemma 8.9. Also,
the length of w is 3N + 1. Then w has some reduced spelling that is a
product of 3 join words and so one of the join words in this latter spelling
has length greater than N . This contradicts the choice of N . Therefore, no
hyperplane intersects both Hi−1 and Hi for each i ∈ {1, 2, · · · , m}. Also,
γ lies outside the (ρr/2)–neighborhood of α1 . We have the following basic
fact:
ℓ(βi ) ≥ ρr/2 + ρr/2 − ℓ(wi ) ≥ ρr − 3N.
Therefore,
ℓ(γ) = ℓ(γ)−2r ≥
m
X
i=1
ℓ(βi )−2r ≥ m(ρr−3N )−2r ≥
r−1
ρr−3N )−2r.
3N + 1
Thus,
σρn (r) ≥
r−1
2M + 3N + 2
.
ρr − 3N ) − 2r for each r >
3N + 1
ρ
or the lower relative divergence of G with respect to H is at least quadratic.
The following theorem is obtained from Propositions 8.12 and 8.13.
Theorem 8.14. Let Γ be a simplicial, finite, connected graph with the vertex
set S such that Γ does not decompose as a nontrivial join. Let H be a nontrivial, finitely generated purely loxodromic subgroup of AΓ . Then the lower
relative divergence of AΓ with respect to H is exactly quadratic.
We now construct an example of a right-angled Artin group AΓ together
with a non-virtually cyclic subgroup H such that the lower relative divergence of the pair (AΓ , H) is exactly quadratic.
Corollary 8.15. Let Γ be the graph as in Figure 2. Let AΓ be the associated
right-angled Artin group and H the subgroup of AΓ generated by two elements
ada and dad. Then H is a free subgroup of rank 2, and the lower relative
divergence of AΓ with respect to H is exactly quadratic.
ON STRONGLY QUASICONVEX SUBGROUPS
50
Proof. This is obvious that the subgroup K generated by two elements a
and d is a free group of rank 2. Since H is a subgroup of K generated by
two elements ada and dad, H is also a free group of rank 2. It is obvious
that any join subword of a reduced word in AΓ representing a non-trivial
group element in H is an element in the set {a, a−1 , a2 , a−2 , d, d−1 , d2 , d−2 }.
Therefore, H is a loxodromic subgroup by Theorem 8.6. This implies that
the lower relative divergence of AΓ with respect to H is exactly quadratic
by the main theorem.
8.3. Main theorem.
Lemma 8.16. Let K be a simplicial, finite graph (K is not necessarily
connected) and Γ the coned-off graph of K with cone point v. If H is a
strongly quasiconvex subgroup of the right-angled Artin group AΓ , then H is
trivial or H has finite index in AΓ .
Proof. We assume that H is nontrivial and we will prove that H has finite
index in AΓ . Since AΓ = AK × hvi, it is sufficient to prove that H ∩ AK
has finite index in AK and v n lies in H for some n 6= 0. We first assume for
the contradiction that H ∩ hvi is a trivial group. Since v commutes with all
vertices of Γ, v commutes with all elements in H. Therefore, v n Hv −n = H
for all n. Also, v i H 6= v j H for i 6= j. This is a contradiction by Theorem
4.15. Therefore, v n lies in H for some n 6= 0.
We now assume again for the contradiction that H ∩ AK has infinite
index in AK . Then there is a infinite sequence (gn ) of distinct elements in
AK such that gi (H ∩ AK ) 6= gj (H ∩ AK ) for i 6= j. This is not difficult
to see that gi H 6= gj H for i 6= j. Since v n lies in H and v n commutes
with all gi , then the infiniteTcyclic group generated by v n is a subgroup
of each gi Hgi−1 . Therefore, gi Hgi−1 is an infinite subgroup. This again
contradicts to Theorem 4.15. Therefore, H ∩ AK has finite index in AK .
This implies that H has finite index in AΓ .
Lemma 8.17. Let Γ be a simplicial, finite, connected graph and K be an induced star subgraph st(v). For each g1 and g2 in AΓ there is a finite sequence
of conjugates of star subgroups g1 AK g1−1 = Q0 , Q1 , · · · , Qm = g2 AK g2−1 such
that Qi−1 ∩ Qi is infinite for each i ∈ {1, 2, · · · , m}.
Proof. Let S be the vertex set of Γ and n = |g1−1 g2 |S . We will prove the
above lemma by induction on n. If n = 0, then g1 = g2 . Therefore, the conclusion is true obviously. We now assume that n = 1. Then there is a vertex
u such that g2 = g1 uǫ , where ǫ = 1 or ǫ = −1. Since Γ is finite and connected, there is a finite sequence of vertices of u = u0 , u1 , u2 , · · · , uℓ = v such
that ui−1 is adjacent to ui for each i ∈ {1, 2, 3, · · · , ℓ}. Let Pi = g1 ASt(ui ) g1−1
and Pi′ = g2 ASt(ui ) g2−1 for each i ∈ {0, 1, 2, 3, · · · , ℓ}. This is not hard to
′
check Pi−1 ∩ Pi and Pi−1
∩ Pi′ are infinite for each i ∈ {1, 2, 3, · · · , ℓ}. Moreover, P0 = P0′ because g2 = g1 uǫ , where ǫ = 1 or ǫ = −1. Therefore, the
conclusion is true for n = 1 obviously.
ON STRONGLY QUASICONVEX SUBGROUPS
51
Assume the conclusion is true for all n ≤ k for some k ≥ 1. We will prove
that the conclusion is true for n = k + 1. In fact, if |g1−1 g2 |S = k + 1, then
there is g3 in AΓ such that |g1−1 g3 |S = k and |g3−1 g2 |S = 1. By the inductive hypothesis, there is a finite sequence of conjugates of star subgroups
g1 AK g1−1 = L0 , L1 , · · · , Lm1 = g3 AK g3−1 such that Li−1 ∩ Li is infinite for
each i ∈ {1, 2, · · · , m1 }. Similarly, there is a finite sequence of conjugates of
star subgroups g3 AK g3−1 = L′0 , L′1 , · · · , L′m2 = g2 AK g2−1 such that L′i−1 ∩ L′i
is infinite for each i ∈ {1, 2, · · · , m2 }. Therefore, there is a finite sequence of
conjugates of star subgroups g1 AK g1−1 = Q0 , Q1 , · · · , Qm = g2 AK g2−1 such
that Qi−1 ∩ Qi is infinite for each i ∈ {1, 2, · · · , m}. This implies that the
conclusion is true for n = k + 1.
Proposition 8.18. Let Γ be a simplicial, finite, connected graph such that Γ
does not decompose as a nontrivial join. Let H be a non-trivial, infinite index
subgroup of the right-angled Artin group AΓ . If H is a strongly quasiconvex
subgroup, then H is free.
Proof. We first observe that H is finitely generated and each conjugate of H
is also a strongly quasiconvex subgroup. We now prove that for each vertex
v of Γ and g ∈ AΓ the subgroup gHg−1 ∩ Ast(v) is trivial. We assume for
the contradiction that g0 Hg0−1 ∩ Ast(v) is not trivial for some vertex v and
some g0 ∈ AΓ . We claim that gHg−1 ∩ Ast(v) has finite index in Ast(v) for
all g ∈ AΓ
In fact, since g0 Hg0−1 is a strongly quasiconvex subgroup and Ast(v) is
an undistorted subgroup, then g0 Hg0−1 ∩ Ast(v) is a strongly quasiconvex
subgroup of Ast(v) by Proposition 4.11. Therefore, g0 Hg0−1 ∩ Ast(v) has
finite index in Ast(v) by Lemma 8.16.
We now prove that gHg−1 ∩ Ast(v) has finite index in Ast(v) for all g ∈ AΓ .
By Lemma 8.17, there is a finite sequence of conjugates of star subgroups
g0−1 Ast(v) g0 = Q0 , Q1 , · · · , Qm = g−1 Ast(v) g such that Qi−1 ∩ Qi is infinite
for each i ∈ {1, 2, · · · , m}. Since g0 Hg0−1 ∩ Ast(v) has finite index in Ast(v) ,
H ∩g0−1 Ast(v) g0 has finite index in Q0 = g0−1 Ast(v) g0 . Also, subgroup Q0 ∩Q1
is infinite. Then, H ∩ Q1 is not trivial. Using a similar argument as above,
we obtain H ∩ Q1 has finite index in Q1 . Repeating this process, we have
H ∩ g −1 Ast(v) g has finite index in g−1 Ast(v) g. In other word, gHg−1 ∩ Ast(v)
has finite index in Ast(v) .
By Theorem 4.15, there is a number n such that the intersection of any
(n+1) essentially distinct conjugates of H is finite. Since H has infinite index
in AΓ , there is n + 1 distinct element g1 , g2 , · · · gn+1 such that gi H 6= gj H for
each i 6= j. Also, gi Hgi−1 ∩ Ast(v) has finite index in Ast(v) for each i. Then
T
T
(∩gi Hgi−1 ) Ast(v) also has finite index in Ast(v) . In particular, gi Hgi−1
is infinite which is a contradiction. Therefore, for each vertex v of Γ and
g ∈ AΓ the subgroup gHg−1 ∩ Ast(v) is trivial. This implies that H is a free
group by Theorem 1.2 in [KMT].
ON STRONGLY QUASICONVEX SUBGROUPS
52
We are now ready for the main theorem in this section.
Theorem 8.19. Let Γ be a simplicial, finite, connected graph such that Γ
does not decompose as a nontrivial join. Let H be a non-trivial, infinite
index subgroup of the right-angled Artin group AΓ . Then the following are
equivalent:
(1) H is strongly quasiconvex.
(2) H is stable.
(3) The lower relative divergence of AΓ with respect to H is quadratic.
(4) The lower relative divergence of AΓ with respect to H is completely
super linear.
Proof. The implication “(1) ⇒ (2)” is obtained from Theorem 4.8 and
Proposition 8.18. The implication “(2) ⇒ (3)” is deduced from Theorem
8.14. The implication “(3) ⇒ (4)” is trivial and the implication “(4) ⇒ (1)”
follows Theorem 4.8.
References
[ABD]
[ACGH]
[ADT]
[AMST]
[BC12]
[Beh]
[BH99]
[BHS17]
[Cor]
[CS15]
[Dah03]
[Dav08]
[DS05]
Carolyn Abbott, Jason Behrstock, and Matthew Gentry Durham. Largest
acylindrical actions and stability in hierarchically hyperbolic groups. Submitted. arXiv:1705.06219.
Goulnara N. Arzhantseva, Christopher H. Cashen, Dominik Gruber, and David
Hume. Characterizations of Morse quasi-geodesics via superlinear divergence
and sublinear contraction. Submitted. arXiv:1601.01897.
Tarik Aougab, Matthew Gentry Durham, and Samuel J. Taylor. Middle recurrence and pulling back stability. Submitted. arXiv:1609.06698.
Yago Antolı́n, Mahan Mj, Alessandro Sisto, and Samuel J. Taylor. Intersection properties of stable subgroups and bounded cohomology. Submitted.
arXiv:1612.07227.
Jason Behrstock and Ruth Charney. Divergence and quasimorphisms of rightangled Artin groups. Math. Ann., 352(2):339–356, 2012.
Jason Behrstock. A counterexample to questions about boundaries, stability,
and commensurability. Submitted. arXiv:1705.03984.
Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature, volume 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999.
Jason Behrstock, Mark F. Hagen, and Alessandro Sisto. Hierarchically hyperbolic spaces, I: Curve complexes for cubical groups. Geom. Topol., 21(3):1731–
1804, 2017.
Matthew Cordes. Morse boundaries of proper geodesic metric spaces. Preprint.
arXiv:1502.04376.
Ruth Charney and Harold Sultan. Contracting boundaries of CAT(0) spaces.
J. Topol., 8(1):93–117, 2015.
François Dahmani. Combination of convergence groups. Geom. Topol., 7:933–
963, 2003.
Michael W. Davis. The geometry and topology of Coxeter groups, volume 32 of
London Mathematical Society Monographs Series. Princeton University Press,
Princeton, NJ, 2008.
Cornelia Druţu and Mark Sapir. Tree-graded spaces and asymptotic cones of
groups. Topology, 44(5):959–1058, 2005. With an appendix by Denis Osin and
Sapir.
ON STRONGLY QUASICONVEX SUBGROUPS
53
[DT15a]
Pallavi Dani and Anne Thomas. Divergence in right-angled Coxeter groups.
Trans. Amer. Math. Soc., 367(5):3549–3577, 2015.
[DT15b] Matthew Gentry Durham and Samuel J. Taylor. Convex cocompactness and
stability in mapping class groups. Algebr. Geom. Topol., 15(5):2839–2859,
2015.
[FM02]
Benson Farb and Lee Mosher. Convex cocompact subgroups of mapping class
groups. Geom. Topol., 6:91–152, 2002.
[GMRS98] Rita Gitik, Mahan Mitra, Eliyahu Rips, and Michah Sageev. Widths of subgroups. Trans. Amer. Math. Soc., 350(1):321–329, 1998.
[Gro93]
M. Gromov. Asymptotic invariants of infinite groups. In Geometric group theory, Vol. 2 (Sussex, 1991), volume 182 of London Math. Soc. Lecture Note
Ser., pages 1–295. Cambridge Univ. Press, Cambridge, 1993.
[Hru10]
G.C. Hruska. Relative hyperbolicity and relative quasiconvexity for countable
groups. Algebr. Geom. Topol., 10(3):1807–1856, 2010.
[HW09]
G. Christopher Hruska and Daniel T. Wise. Packing subgroups in relatively
hyperbolic groups. Geom. Topol., 13(4):1945–1988, 2009.
[KMT]
Thomas Koberda, Johanna Mangahas, and Samuel J. Taylor. The geometry of
purely loxodromic subgroups of right-angled Artin groups. To appear in Trans.
Amer. Math. Soc.
[KS96]
Ilya Kapovich and Hamish Short. Greenberg’s theorem for quasiconvex subgroups of word hyperbolic groups. Canad. J. Math., 48(6):1224–1244, 1996.
[Osi06]
Denis V. Osin. Relatively hyperbolic groups: Intrinsic geometry, algebraic
properties, and algorithmic problems. Mem. Amer. Math. Soc., 179(843):1–
100, 2006.
[Osi16]
D. Osin. Acylindrically hyperbolic groups. Trans. Amer. Math. Soc.,
368(2):851–888, 2016.
[Sag95]
Michah Sageev. Ends of group pairs and non-positively curved cube complexes.
Proc. London Math. Soc. (3), 71(3):585–617, 1995.
[Sag97]
Michah Sageev. Codimension-1 subgroups and splittings of groups. J. Algebra,
189(2):377–389, 1997.
[Sho91]
Hamish Short. Quasiconvexity and a theorem of Howson’s. In Group theory
from a geometrical viewpoint (Trieste, 1990), pages 168–176. World Sci. Publ.,
River Edge, NJ, 1991.
[Sis16]
Alessandro Sisto. Quasi-convexity of hyperbolically embedded subgroups.
Math. Z., 283(3-4):649–658, 2016.
[Tra]
Hung Cong Tran. Purely loxodromic subgroups in right-angled Coxeter groups.
Submitted. arXiv:1703.09032.
[Tra15]
Hung Cong Tran. Relative divergence of finitely generated groups. Algebr.
Geom. Topol., 15(3):1717–1769, 2015.
[Tra16]
Hung Cong Tran. Divergence of Morse geodesics. Geom. Dedicata, 180:385–
397, 2016.
Department of Mathematics, The University of Georgia, 1023 D. W. Brooks
Drive, Athens, GA 30605, USA
E-mail address: [email protected]
| 4math.GR
|
Deep Learning with Topological Signatures
Roland Kwitt
Department of Computer Science
University of Salzburg, Austria
[email protected]
arXiv:1707.04041v3 [cs.CV] 16 Feb 2018
Christoph Hofer
Department of Computer Science
University of Salzburg, Austria
[email protected]
Marc Niethammer
UNC Chapel Hill, NC, USA
[email protected]
Andreas Uhl
Department of Computer Science
University of Salzburg, Austria
[email protected]
Abstract
Inferring topological and geometrical information from data can offer an alternative
perspective on machine learning problems. Methods from topological data analysis,
e.g., persistent homology, enable us to obtain such information, typically in the form
of summary representations of topological features. However, such topological
signatures often come with an unusual structure (e.g., multisets of intervals) that is
highly impractical for most machine learning techniques. While many strategies
have been proposed to map these topological signatures into machine learning
compatible representations, they suffer from being agnostic to the target learning
task. In contrast, we propose a technique that enables us to input topological
signatures to deep neural networks and learn a task-optimal representation during
training. Our approach is realized as a novel input layer with favorable theoretical
properties. Classification experiments on 2D object shapes and social network
graphs demonstrate the versatility of the approach and, in case of the latter, we
even outperform the state-of-the-art by a large margin.
1
Introduction
Methods from algebraic topology have only recently emerged in the machine learning community,
most prominently under the term topological data analysis (TDA) [7]. Since TDA enables us to
infer relevant topological and geometrical information from data, it can offer a novel and potentially
beneficial perspective on various machine learning problems. Two compelling benefits of TDA
are (1) its versatility, i.e., we are not restricted to any particular kind of data (such as images,
sensor measurements, time-series, graphs, etc.) and (2) its robustness to noise. Several works have
demonstrated that TDA can be beneficial in a diverse set of problems, such as studying the manifold
of natural image patches [8], analyzing activity patterns of the visual cortex [28], classification of 3D
surface meshes [27, 22], clustering [11], or recognition of 2D object shapes [29].
Currently, the most widely-used tool from TDA is persistent homology [15, 14]. Essentially1 ,
persistent homology allows us to track topological changes as we analyze data at multiple “scales”.
As the scale changes, topological features (such as connected components, holes, etc.) appear and
disappear. Persistent homology associates a lifespan to these features in the form of a birth and
a death time. The collection of (birth, death) tuples forms a multiset that can be visualized as a
persistence diagram or a barcode, also referred to as a topological signature of the data. However,
leveraging these signatures for learning purposes poses considerable challenges, mostly due to their
1
We will make these concepts more concrete in Sec. 2.
Death
(2) Transform & Project
(µ2 , σ 2 )
x = ρ(p)
(µ1 , σ 1 )
sµ,σ,ν ((x0 , x1 ))
ν
Output:
(y1 , y2 )> ∈ R0 + ×R0
(1) Rotate points in D by π/4
Death-Birth
(persistence)
Input Layer
Input: D ∈ D
p = (b, d)
−1
Param.: θ = (µi , σ i )N
i=0
Birth
∆
(x0 , x1 )
Birth+Death
ν
Figure 1: Illustration of the proposed network input layer for topological signatures. Each signature, in the
form of a persistence diagram D ∈ D (left), is projected w.r.t. a collection of structure elements. The layer’s
learnable parameters θ are the locations µi and the scales σ i of these elements; ν ∈ R+ is set a-priori and
meant to discount the impact of points with low persistence (and, in many cases, of low discriminative power).
The layer output y is a concatenation of the projections. In this illustration, N = 2 and hence y = (y1 , y2 )> .
unusual structure as a multiset. While there exist suitable metrics to compare signatures (e.g., the
Wasserstein metric), they are highly impractical for learning, as they require solving optimal matching
problems.
Related work. In order to deal with these issues, several strategies have been proposed. In [2] for
instance, Adcock et al. use invariant theory to “coordinatize” the space of barcodes. This allows to
map barcodes to vectors of fixed size which can then be fed to standard machine learning techniques,
such as support vector machines (SVMs). Alternatively, Adams et al. [1] map barcodes to so-called
persistence images which, upon discretization, can also be interpreted as vectors and used with
standard learning techniques. Along another line of research, Bubenik [6] proposes a mapping
of barcodes into a Banach space. This has been shown to be particularly viable in a statistical
context (see, e.g., [10]). The mapping outputs a representation referred to as a persistence landscape.
Interestingly, under a specific choice of parameters, barcodes are mapped into L2 (R2 ) and the
inner-product in that space can be used to construct a valid kernel function. Similar, kernel-based
techniques, have also recently been studied by Reininghaus et al. [27], Kwitt et al. [20] and Kusano
et al. [19].
While all previously mentioned approaches retain certain stability properties of the original representation with respect to common metrics in TDA (such as the Wasserstein or Bottleneck distances),
they also share one common drawback: the mapping of topological signatures to a representation that
is compatible with existing learning techniques is pre-defined. Consequently, it is fixed and therefore
agnostic to any specific learning task. This is clearly suboptimal, as the eminent success of deep
neural networks (e.g., [18, 17]) has shown that learning representations is a preferable approach.
Furthermore, techniques based on kernels [27, 20, 19] for instance, additionally suffer scalability
issues, as training typically scales poorly with the number of samples (e.g., roughly cubic in case of
kernel-SVMs). In the spirit of end-to-end training, we therefore aim for an approach that allows to
learn a task-optimal representation of topological signatures. We additionally remark that, e.g., Qi et
al. [25] or Ravanbakhsh et al. [26] have proposed architectures that can handle sets, but only with
fixed size. In our context, this is impractical as the capability of handling sets with varying cardinality
is a requirement to handle persistent homology in a machine learning setting.
Contribution. To realize this idea, we advocate a novel input layer for deep neural networks that
takes a topological signature (in our case, a persistence diagram), and computes a parametrized
projection that can be learned during network training. Specifically, this layer is designed such that
its output is stable with respect to the 1-Wasserstein distance (similar to [27] or [1]). To demonstrate
the versatility of this approach, we present experiments on 2D object shape classification and the
classification of social network graphs. On the latter, we improve the state-of-the-art by a large
margin, clearly demonstrating the power of combining TDA with deep learning in this context.
2
Background
For space reasons, we only provide a brief overview of the concepts that are relevant to this work and
refer the reader to [16] or [14] for further details.
Homology. The key concept of homology theory is to study the properties of some object X by
means of (commutative) algebra. In particular, we assign to X a sequence of modules C0 , C1 , . . .
2
which are connected by homomorphisms ∂n : Cn → Cn−1 such that im ∂n+1 ⊆ ker ∂n . A structure
of this form is called a chain complex and by studying its homology groups Hn = ker ∂n / im ∂n+1
we can derive properties of X.
A prominent example of a homology theory is simplicial homology. Throughout this work, it is
the used homology theory and hence we will now concretize the already presented ideas. Let K
be a simplicial complex and Kn its n-skeleton. Then we set Cn (K) as the vector space generated (freely) by Kn over Z/2Z2 . The connecting homomorphisms ∂n : Cn (K) → Cn−1 (K) are
called
∂n (σ) =
Pn boundary operators. For a simplex σ = [x0 , . . . , xn ] ∈ Kn , we definePthem asP
σi ) =
∂n (σi ).
i=0 [x0 , . . . , xi−1 , xi+1 , . . . , xn ] and linearly extend this to Cn (K), i.e., ∂n (
Persistent homology. Let K be a simplicial complex and (K i )m
i=0 a sequence of simplicial complexes such that ∅ = K 0 ⊆ K 1 ⊆ · · · ⊆ K m = K. Then, (K i )m
i=0 is called a filtration of K. If we
use the extra information provided by the filtration of K, we obtain the following sequence of chain
complexes (left),
···
∂3
C22
···
∂3
∂2
C11
∂2
C12
ι
C01
∂1
C02
ι
ι
C2m
∂1
C1m
∂0
0
ι
∂1
C0m
C01 = [[v1 ], [v2 ]]Z2
K1
0
ι
ι
∂2
∂0
∂0
Example
C21
C11 = 0
0
C02 = [[v1 ], [v2 ], [v3 ]]Z2
K2
⊆
∂3
⊆
···
C12 = [[v1 , v3 ], [v2 , v3 ]]Z2
v2
K3
v3
v4
v1
C02 = [[v1 ], [v2 ], [v3 ], [v4 ]]Z2
C12 = [[v1 , v3 ], [v2 , v3 ], [v3 , v4 ]]Z2
C21 = 0
C22 = 0
C23 = 0
where Cni = Cn (Kni ) and ι denotes the inclusion. This then leads to the concept of persistent
homology groups, defined by
j
Hni,j = ker ∂ni /(im ∂n+1
∩ ker ∂ni ) for
i≤j .
The ranks, βni,j = rank Hni,j , of these homology groups (i.e., the n-th persistent Betti numbers),
capture the number of homological features of dimensionality n (e.g., connected components for
n = 0, holes for n = 1, etc.) that persist from i to (at least) j. In fact, according to [14, Fundamental
Lemma of Persistent Homology], the quantities
i,j−1
µi,j
− βni,j ) − (βni−1,j−1 − βni−1,j ) for
n = (βn
i<j
(1)
encode all the information about the persistent Betti numbers of dimension n.
Topological signatures. A typical way to obtain a filtration of K is to consider sublevel sets of a
function f : C0 (K) → R. This function can be easily lifted to higher-dimensional chain groups of
K by
f ([v0 , . . . , vn ]) = max{f ([vi ]) : 0 ≤ i ≤ n} .
−1
Given m = |f (C0 (K))|, we obtain (Ki )m
((−∞, ai ]) for
i=0 by setting K0 = ∅ and Ki = f
1 ≤ i ≤ m, where a1 < · · · < am is the sorted sequence of values of f (C0 (K)). If we construct
a multiset such that, for i < j, the point (ai , aj ) is inserted with multiplicity µi,j
n , we effectively
encode the persistent homology of dimension n w.r.t. the sublevel set filtration induced by f . Upon
adding diagonal points with infinite multiplicity, we obtain the following structure:
Definition 1 (Persistence diagram). Let ∆ = {x ∈ R2∆ : mult(x) = ∞} be the multiset of the
diagonal R2∆ = {(x0 , x1 ) ∈ R2 : x0 = x1 }, where mult denotes the multiplicity function and let
R2? = {(x0 , x1 ) ∈ R2 : x1 > x0 }. A persistence diagram, D, is a multiset of the form
D = {x : x ∈ R2? } ∪ ∆ .
We denote by D the set of all persistence diagrams of the form |D \ ∆| < ∞ .
For a given complex K of dimension nmax and a function f (of the discussed form), we can interpret
persistent homology as a mapping (K, f ) 7→ (D0 , . . . , Dnmax −1 ), where Di is the diagram of
dimension i and nmax the dimension of K. We can additionally add a metric structure to the space of
persistence diagrams by introducing the notion of distances.
Simplicial homology is not specific to Z/2Z, but it’s a typical choice, since it allows us to interpret n-chains
as sets of n-simplices.
2
3
Definition 2 (Bottleneck, Wasserstein distance). For two persistence diagrams D and E, we define
their Bottleneck (w∞ ) and Wasserstein (wqp ) distances by
! p1
w∞ (D, E) = inf sup ||x − η(x)||∞ and wqp (D, E) = inf
η x∈D
η
X
x∈D
||x − η(x)||pq
,
(2)
where p, q ∈ N and the infimum is taken over all bijections η : D → E.
Essentially, this facilitates studying stability/continuity properties of topological signatures w.r.t.
metrics in the filtration or complex space; we refer the reader to [12],[13], [9] for a selection of
important stability results.
Remark. By setting µni,∞ = βni,m −βni−1,m , we extend Eq. (1) to features which never disappear, also
referred to as essential. This change can be lifted to D by setting R2? = {(x0 , x1 ) ∈ R × (R ∪ {∞}) :
x1 > x0 }. In Sec. 5, we will see that essential features can offer discriminative information.
3
A network layer for topological signatures
In this section, we introduce the proposed (parametrized) network layer for topological signatures
(in the form of persistence diagrams). The key idea is to take any D and define a projection w.r.t. a
collection (of fixed size N ) of structure elements.
In the following, we set R+ := {x ∈ R : x > 0} and R+
0 := {x ∈ R : x ≥ 0}, resp., and start by
rotating points of D such that points on R2∆ lie on the x-axis, see Fig. 1. The y-axis can then be
interpreted as the persistence of features. Formally, we let b0 and b1 be the unit vectors in directions
(1, 1)> and (−1, 1)> and define a mapping ρ : R2? ∪ R2∆ → R × R+
0 such that x 7→ (hx, b0 i, hx, b1 i).
This rotates points in R? ∪ R2∆ clock-wise by π/4. We will later see that this construction is beneficial
for a closer analysis of the layers’ properties. Similar to [27, 19], we choose exponential functions as
structure elements, but other choices are possible (see Lemma 1). Differently to [27, 19], however,
our structure elements are not at fixed locations (i.e., one element per point in D), but their locations
and scales are learned during training.
Definition 3. Let µ = (µ0 , µ1 )> ∈ R × R+ , σ = (σ0 , σ1 ) ∈ R+ × R+ and ν ∈ R+ . We define
sµ,σ,ν : R × R+
0 →R
as follows:
sµ,σ,ν
2
2
2
2
e−σ0 (x0 −µ0 ) −σ1 (x1 −µ1 ) ,
x1 ∈ [ν, ∞)
x1
2
2
2
2
(x0 , x1 ) =
e−σ0 (x0 −µ0 ) −σ1 (ln( ν )ν+ν−µ1 ) , x1 ∈ (0, ν)
0,
x1 = 0
A persistence diagram D is then projected w.r.t. sµ,σ,ν via
X
Sµ,σ,ν : D → R,
D 7→
sµ,σ,ν (ρ(x)) .
(3)
(4)
x∈D
Remark. Note that sµ,σ,ν is continuous in x1 as
x
ν + ν and lim sµ,σ,ν (x0 , x1 ) = 0 = sµ,σ,ν (x0 , 0)
lim x = lim ln
x→ν
x→ν
x1 →0
ν
and e(·) is continuous. Further, sµ,σ,ν is differentiable on R × R+ , since
∂x1
1 = lim
(x) and
x→ν + ∂x1
lim
∂ ln
x1
ν
∂x1
x→ν −
ν+ν
(x) = lim
x→ν −
ν
=1 .
x
Also note that we use the log-transform in Eq. (4) to guarantee that sµ,σ,ν satisfies the conditions of
Lemma 1; this is, however, only one possible choice. Finally, given a collection of structure elements
Sµi ,σi ,ν , we combine them to form the output of the network layer.
4
−1
+
+
+
Definition 4. Let N ∈ N, θ = (µi , σ i )N
i=0 ∈ (R × R ) × (R × R )
N
Sθ,ν : D → (R+
0)
N
and ν ∈ R+ . We define
N −1
D 7→ Sµi ,σi ,ν (D) i=0 .
as the concatenation of all N mappings defined in Eq. (4).
Importantly, a network layer implementing Def. 4 is trainable via backpropagation, as (1) sµi ,σi ,ν is
differentiable in µi , σ i , (2) Sµi ,σi ,ν (D) is a finite sum of sµi ,σi ,ν and (3) Sθ,ν is just a concatenation.
4
Theoretical properties
In this section, we demonstrate that the proposed layer is stable w.r.t. the 1-Wasserstein distance wq1 ,
see Eq. (2). In fact, this claim will follow from a more general result, stating sufficient conditions on
q
functions s : R2? ∪ R2∆ → R+
0 such that a construction in the form of Eq. (3) is stable w.r.t. w1 .
Lemma 1. Let
s : R2? ∪ R2∆ → R+
0
have the following properties:
(i) s is Lipschitz continuous w.r.t. k · kq and constant Ks
(ii) s(x = 0, for x ∈ R2∆
Then, for two persistence diagrams D, E ∈ D, it holds that
X
x∈D
s(x) −
X
y∈E
s(y) ≤ Ks · wq1 (D, E) .
(5)
Proof. see Appendix B
Remark. At this point, we want to clarify that Lemma 1 is not specific to sµ,σ,ν (e.g., as in Def. 3).
Rather, Lemma 1 yields sufficient conditions to construct a w1 -stable input layer. Our choice of
sµ,σ,ν is just a natural example that fulfils those requirements and, hence, Sθ,ν is just one possible
representative of a whole family of input layers.
With the result of Lemma 1 in mind, we turn to the specific case of Sθ,ν and analyze its stability
properties w.r.t. wq1 . The following lemma is important in this context.
Lemma 2. sµ,σ,ν has absolutely bounded first-order partial derivatives w.r.t. x0 and x1 on R × R+ .
Proof. see Appendix B
Theorem 1. Sθ,ν is Lipschitz continuous with respect to wq1 on D.
Proof. Lemma 2 immediately implies that sµ,σ,ν from Eq. (3) is Lipschitz continuous w.r.t || · ||q .
Consequently, s = sµ,σ,ν ◦ ρ satisfies property (i) from Lemma 1; property (ii) from Lemma 1 is
satisfied by construction. Hence, Sµ,σ,ν is Lipschitz continuous w.r.t. wq1 . Consequently, Sθ,ν is
Lipschitz in each coordinate and therefore Liptschitz continuous.
Interestingly, the stability result of Theorem 1 is comparable to the stability results in [1] or [27]
(which are also w.r.t. wq1 and in the setting of diagrams with finitely-many points). However, contrary
to previous works, if we would chop-off the input layer after network training, we would then have a
mapping Sθ,ν of persistence diagrams that is specifically-tailored to the learning task on which the
network was trained.
5
b5
shift due to noise
b8
S1
a2
a3
b1
b2,3,4
b6
b7
b7
b8
ν
b5
b3
Artificially added noise
Filtration directions
Persistence diagram
(0-dim. features)
a3
b9
a2
a1
b9
b1
b2 a1 b4
b6
Figure 2: Height function filtration of a “clean” (left, green points) and a “noisy” (right, blue points) shape
along direction d = (0, −1)> . This example demonstrates the insensitivity of homology towards noise, as the
added noise only (1) slightly shifts the dominant points (upper left corner) and (2) produces additional points
close to the diagonal, which have little impact on the Wasserstein distance and the output of our layer.
5
Experiments
To demonstrate the versatility of the proposed approach, we present experiments with two totally
different types of data: (1) 2D shapes of objects, represented as binary images and (2) social network
graphs, given by their adjacency matrix. In both cases, the learning task is classification. In each
experiment we ensured a balanced group size (per label) and used a 90/10 random training/test
split; all reported results are averaged over five runs with fixed ν = 0.1. In practice, points in input
diagrams were thresholded at 0.01 for computational reasons. Additionally, we conducted a reference
experiment on all datasets using simple vectorization (see Sec. 5.3) of the persistence diagrams in
combination with a linear SVM.
Implementation. All experiments were implemented in PyTorch3 , using DIPHA4 and Perseus [23].
Source code is publicly-available at https://github.com/c-hofer/nips2017.
5.1
Classification of 2D object shapes
We apply persistent homology combined with our proposed input layer to two different datasets of
binary 2D object shapes: (1) the Animal dataset, introduced in [3] which consists of 20 different
animal classes, 100 samples each; (2) the MPEG-7 dataset which consists of 70 classes of different
object/animal contours, 20 samples each (see [21] for more details).
Filtration. The requirements to use persistent homology on 2D shapes are twofold: First, we need
to assign a simplicial complex to each shape; second, we need to appropriately filtrate the complex.
While, in principle, we could analyze contour features, such as curvature, and choose a sublevel set
filtration based on that, such a strategy requires substantial preprocessing of the discrete data (e.g.,
smoothing). Instead, we choose to work with the raw pixel data and leverage the persistent homology
transform, introduced by Turner et al. [29]. The filtration in that case is based on sublevel sets of
the height function, computed from multiple directions (see Fig. 2). Practically, this means that we
directly construct a simplicial complex from the binary image. We set K0 as the set of all pixels
which are contained in the object. Then, a 1-simplex [p0 , p1 ] is in the 1-skeleton K1 iff p0 and p1
are 4–neighbors on the pixel grid. To filtrate the constructed complex, we define by b the barycenter
of the object and with r the radius of its bounding circle around b. Finally, we define, for [p] ∈ K0
and d ∈ S1 , the filtration function by f ([p]) = 1/r · hp − b, di. Function values are lifted to K1 by
taking the maximum, cf. Sec. 2. Finally, let di be the 32 equidistantly distributed directions in S1 ,
starting from (1, 0). For each shape, we get a vector of persistence diagrams (Di )32
i=1 where Di is the
0-th diagram obtained by filtration along di . As most objects do not differ in homology groups of
higher dimensions (> 0), we did not use the corresponding persistence diagrams.
Network architecture. While the full network is listed in the supplementary material (Fig. 6), the
key architectural choices are: 32 independent input branches, i.e., one for each filtration direction.
Further, the i-th branch gets, as input, the vector of persistence diagrams from directions di−1 , di
and di+1 . This is a straightforward approach to capture dependencies among the filtration directions.
We use cross-entropy loss to train the network for 400 epochs, using stochastic gradient descent
(SGD) with mini-batches of size 128 and an initial learning rate of 0.1 (halved every 25-th epoch).
3
4
https://github.com/pytorch/pytorch
https://bitbucket.org/dipha/dipha
6
MPEG-7
Animal
Skeleton paths
Class segment sets
†
ICS
†
BCF
86.7
90.9
96.6
97.2
67.9
69.7
78.4
83.4
Ours
91.8
69.5
‡
‡
Figure 3: Left: some examples from the MPEG-7 (bottom) and Animal (top) datasets. Right: Classification
results, compared to the two best (†) and two worst (‡) results reported in [30].
Results. Fig. 3 shows a selection of 2D object shapes from both datasets, together with the obtained
classification results. We list the two best (†) and two worst (‡) results as reported in [30]. While,
on the one hand, using topological signatures is below the state-of-the-art, the proposed architecture
is still better than other approaches that are specifically tailored to the problem. Most notably, our
approach does not require any specific data preprocessing, whereas all other competitors listed in
Fig. 3 require, e.g., some sort of contour extraction. Furthermore, the proposed architecture readily
generalizes to 3D with the only difference that in this case di ∈ S2 . Fig. 4 (Right) shows an exemplary
visualization of the position of the learned structure elements for the Animal dataset.
5.2
Classification of social network graphs
In this experiment, we consider the problem of graph classification, where vertices are unlabeled
and edges are undirected. That is, a graph G is given by G = (V, E), where V denotes the set of
vertices and E denotes the set of edges. We evaluate our approach on the challenging problem of
social network classification, using the two largest benchmark datasets from [31], i.e., reddit-5k
(5 classes, 5k graphs) and reddit-12k (11 classes, ≈12k graphs). Each sample in these datasets
represents a discussion graph and the classes indicate subreddits (e.g., worldnews, video, etc.).
Filtration. The construction of a simplicial complex from G = (V, E) is straightforward: we set
K0 = {[v] ∈ V } and K1 = {[v0 , v1 ] : {v0 , v1 } ∈ E}. We choose a very simple filtration based on
the vertex degree, i.e., the number of incident edges to a vertex v ∈ V . Hence, for [v0 ] ∈ K0 we get
f ([v0 ]) = deg(v0 )/ maxv∈V deg(v) and again lift f to K1 by taking the maximum. Note that chain
groups are trivial for dimension > 1, hence, all features in dimension 1 are essential.
Network architecture. Our network has four input branches: two for each dimension (0 and 1) of
the homological features, split into essential and non-essential ones, see Sec. 2. We train the network
for 500 epochs using SGD and cross-entropy loss with an initial learning rate of 0.1 (reddit_5k), or
0.4 (reddit_12k). The full network architecture is listed in the supplementary material (Fig. 7).
Results. Fig. 5 (right) compares our proposed strategy to state-of-the-art approaches from the
literature. In particular, we compare against (1) the graphlet kernel (GK) and deep graphlet kernel
(DGK) results from [31], (2) the Patchy-SAN (PSCN) results from [24] and (3) a recently reported
graph-feature + random forest approach (RF) from [4]. As we can see, using topological signatures
in our proposed setting considerably outperforms the current state-of-the-art on both datasets. This is
an interesting observation, as PSCN [24] for instance, also relies on node degrees and an extension of
the convolution operation to graphs. Further, the results reveal that including essential features is key
to these improvements.
5.3
Vectorization of persistence diagrams
Here, we briefly present a reference experiment we conducted following Bendich et al. [5]. The idea
is to directly use the persistence diagrams as features via vectorization. For each point (b, d) in a
persistence diagram D we calculate its persistence, i.e., d − b. We then sort the calculated persistences
by magnitude from high to low and take the first N values. Hence, we get, for each persistence
diagram, a vector of dimension N (if |D \ ∆| < N , we pad with zero). We used this technique
on all four data sets. As can be seen from the results in Table 4 (averaged over 10 cross-validation
runs), vectorization performs poorly on MPEG-7 and Animal but can lead to competitive rates on
reddit-5k and reddit-12k. Nevertheless, the obtained performance is considerably inferior to our
proposed approach.
7
MPEG-7
Animal
reddit-5k
reddit-12k
5
10
20
40
80
160
81.8
48.8
37.1
24.2
82.3
50.0
38.2
24.6
79.7
46.2
39.7
27.9
74.5
42.4
42.1
29.8
68.2
39.3
43.8
31.5
64.4
36.0
45.2
31.6
Ours
91.8
69.5
54.5
44.5
0.8
Death
1.0
N
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
Birth
1.0
Figure 4: Left: Classification accuracies for a linear SVM trained on vectorized (in RN ) persistence diagrams
(see Sec. 5.3). Right: Exemplary visualization of the learned structure elements (in 0-th dimension) for the
Animal dataset and filtration direction d = (−1, 0)> . Centers of the learned elements are marked in blue.
1
1
1
G = (V, E)
f −1 ((−∞, 2])
1
5
3
1
2
2
1
f −1 ((−∞, 3])
f −1 ((−∞, 5])
reddit-5k
reddit-12k
GK [31]
DGK [31]
PSCN [24]
RF [4]
41.0
41.3
49.1
50.9
31.8
32.2
41.3
42.7
Ours (w/o essential)
Ours (w/ essential)
49.1
54.5
38.5
44.5
Figure 5: Left: Illustration of graph filtration by vertex degree, i.e., f ≡ deg (for different choices of ai , see
Sec. 2). Right: Classification results as reported in [31] for GK and DGK, Patchy-SAN (PSCN) as reported in
[24] and feature-based random-forest (RF) classification. from [4].
Finally, we remark that in both experiments, tests with the kernel of [27] turned out to be computationally impractical, (1) on shape data due to the need to evaluate the kernel for all filtration directions
and (2) on graphs due the large number of samples and the number of points in each diagram.
6
Discussion
We have presented, to the best of our knowledge, the first approach towards learning task-optimal
stable representations of topological signatures, in our case persistence diagrams. Our particular
realization of this idea, i.e., as an input layer to deep neural networks, not only enables us to learn with
topological signatures, but also to use them as additional (and potentially complementary) inputs to
existing deep architectures. From a theoretical point of view, we remark that the presented structure
elements are not restricted to exponential functions, so long as the conditions of Lemma 1 are met.
One drawback of the proposed approach, however, is the artificial bending of the persistence axis (see
Fig. 1) by a logarithmic transformation; in fact, other strategies might be possible and better suited
in certain situations. A detailed investigation of this issue is left for future work. From a practical
perspective, it is also worth pointing out that, in principle, the proposed layer could be used to handle
any kind of input that comes in the form of multisets (of Rn ), whereas previous works only allow
to handle sets of fixed size (see Sec. 1). In summary, we argue that our experiments show strong
evidence that topological features of data can be beneficial in many learning tasks, not necessarily to
replace existing inputs, but rather as a complementary source of discriminative information.
8
A
Technical results
Lemma 3. Let α ∈ R+ , β ∈ R, γ ∈ R+ . We have
i)
lim
x→0
ln(x)
x
2
2
1
x→0 x
· e−α(ln(x)γ+β) = 0
· e−α(ln(x)γ+β) = 0 .
ii) lim
Proof. We omit the proof for brevity (see supplementary material for details), but remark that only
(i) needs to be shown as (ii) follows immediately.
B
Proofs
Proof of Lemma 1. Let ϕ be a bijection between D and E which realizes wq1 (D, E) and let D0 =
D \ ∆, E0 = E \ ∆. To show the result of Eq. (5), we consider the following decomposition:
D = ϕ−1 (E0 ) ∪ ϕ−1 (∆)
= (ϕ−1 (E0 ) \ ∆) ∪ (ϕ−1 (E0 ) ∩ ∆) ∪ (ϕ−1 (∆) \ ∆) ∪ (ϕ−1 (∆) ∩ ∆)
|
{z
} |
{z
} |
{z
} |
{z
}
A
B
C
(6)
D
Except for the term D, all sets are finite. In fact, ϕ realizes the Wasserstein distance wq1 which implies
ϕ D = id. Therefore, s(x) = s(ϕ(x)) = 0 for x ∈ D since D ⊂ ∆. Consequently, we can ignore D
in the summation and it suffices to consider E = A ∪ B ∪ C. It follows that
X
x∈D
s(x) −
X
s(y) =
y∈E
X
x∈D
=
X
x∈E
≤ Ks ·
s(x) −
X
x∈D
x∈E
s(x) − s(ϕ(x)) ≤
X
x∈E
X
s(ϕ(x)) =
X
x∈E
s(x) −
X
s(ϕ(x))
x∈E
|s(x) − s(ϕ(x))|
||x − ϕ(x)||q = Ks ·
X
x∈D
||x − ϕ(x)||q = Ks · wq1 (D, E) .
Proof of Lemma 2. Since sµ,σ,ν is defined differently for x1 ∈ [ν, ∞) and x1 ∈ (0, ν), we need to
distinguish these two cases. In the following x0 ∈ R.
(1) x1 ∈ [ν, ∞): The partial derivative w.r.t. xi is given as
∂
∂ −σi2 (xi −µi )2
sµ,σ,ν (x0 , x1 ) = C ·
e
(x0 , x1 )
∂xi
∂xi
=C ·e
−σi2 (xi −µi )2
·
(−2σi2 )(xi
(7)
− µi ) ,
where C is just the part of exp(·) which is not dependent on xi . For all cases, i.e., x0 → ∞, x0 →
−∞ and x1 → ∞, it holds that Eq. (7) → 0.
(2) x1 ∈ (0, ν): The partial derivative w.r.t. x0 is similar to Eq. (7) with the same asymptotic
behaviour for x0 → ∞ and x0 → −∞. However, for the partial derivative w.r.t. x1 we get
∂
∂ −σ12 (ln( x1 )ν+ν−µ1 )2
ν
sµ,σ,ν (x0 , x1 ) = C ·
e
(x0 , x1 )
∂x1
∂x1
x
ν
1
= C · e( ... ) · (−2σ12 ) · ln
ν + ν − µ1 ·
ν
x1
(8)
x
ν
1
1
= C 0 · e( ... ) · ln
·
+(ν − µ1 ) · e( ... ) ·
.
ν
x1
x
| {z 1}
|
{z
}
(a)
(b)
As x1 → 0, we can invoke Lemma 4 (i) to handle (a) and Lemma 4 (ii) to handle (b); conclusively,
Eq. (8) → 0. As the partial derivatives w.r.t. xi are continuous and their limits are 0 on R, R+ , resp.,
we conclude that they are absolutely bounded.
9
References
[1] H. Adams, T. Emerson, M. Kirby, R. Neville, C. Peterson, P. Shipman, S. Chepushtanova, E. Hanson,
F. Motta, and L. Ziegelmeier. Persistence images: A stable vector representation of persistent homology.
JMLR, 18(8):1–35, 2017.
[2] A. Adcock, E. Carlsson, and G. Carlsson. The ring of algebraic functions on persistence bar codes. CoRR,
2013. https://arxiv.org/abs/1304.0530.
[3] X. Bai, W. Liu, and Z. Tu. Integrating contour and skeleton for shape classification. In ICCV Workshops,
2009.
[4] I. Barnett, N. Malik, M.L. Kuijjer, P.J. Mucha, and J.-P. Onnela. Feature-based classification of networks.
CoRR, 2016. https://arxiv.org/abs/1610.05868.
[5] P. Bendich, J.S. Marron, E. Miller, A. Pieloch, and S. Skwerer. Persistent homology analysis of brain artery
trees. Ann. Appl. Stat, 10(2), 2016.
[6] P. Bubenik. Statistical topological data analysis using persistence landscapes. JMLR, 16(1):77–102, 2015.
[7] G. Carlsson. Topology and data. Bull. Amer. Math. Soc., 46:255–308, 2009.
[8] G. Carlsson, T. Ishkhanov, V. de Silva, and A. Zomorodian. On the local behavior of spaces of natural
images. IJCV, 76:1–12, 2008.
[9] F. Chazal, D. Cohen-Steiner, L. J. Guibas, F. Mémoli, and S. Y. Oudot. Gromov-Hausdorff stable signatures
for shapes using persistence. Comput. Graph. Forum, 28(5):1393–1403, 2009.
[10] F. Chazal, B.T. Fasy, F. Lecci, A. Rinaldo, and L. Wassermann. Stochastic convergence of persistence
landscapes and silhouettes. JoCG, 6(2):140–161, 2014.
[11] F. Chazal, L.J. Guibas, S.Y. Oudot, and P. Skraba. Persistence-based clustering in Riemannian manifolds.
J. ACM, 60(6):41–79, 2013.
[12] D. Cohen-Steiner, H. Edelsbrunner, and J. Harer. Stability of persistence diagrams. Discrete Comput.
Geom., 37(1):103–120, 2007.
[13] D. Cohen-Steiner, H. Edelsbrunner, J. Harer, and Y. Mileyko. Lipschitz functions have Lp -stable persistence.
Found. Comput. Math., 10(2):127–139, 2010.
[14] H. Edelsbrunner and J. L. Harer. Computational Topology : An Introduction. American Mathematical
Society, 2010.
[15] H. Edelsbrunner, D. Letcher, and A. Zomorodian. Topological persistence and simplification. Discrete
Comput. Geom., 28(4):511–533, 2002.
[16] A. Hatcher. Algebraic Topology. Cambridge University Press, Cambridge, 2002.
[17] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016.
[18] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural
networks. In NIPS, 2012.
[19] G. Kusano, K. Fukumizu, and Y. Hiraoka. Persistence weighted Gaussian kernel for topological data
analysis. In ICML, 2016.
[20] R. Kwitt, S. Huber, M. Niethammer, W. Lin, and U. Bauer. Statistical topological data analysis - a kernel
perspective. In NIPS, 2015.
[21] L. Latecki, R. Lakamper, and T. Eckhardt. Shape descriptors for non-rigid shapes with a single closed
contour. In CVPR, 2000.
[22] C. Li, M. Ovsjanikov, and F. Chazal. Persistence-based structural recognition. In CVPR, 2014.
[23] K. Mischaikow and V. Nanda. Morse theory for filtrations and efficient computation of persistent homology.
Discrete Comput. Geom., 50(2):330–353, 2013.
[24] M. Niepert, M. Ahmed, and K. Kutzkov. Learning convolutional neural networks for graphs. In ICML,
2016.
[25] C.R. Qi, H. Su, K. Mo, and L.J. Guibas. PointNet: Deep learning on point sets for 3D classification and
segmentation. In CVPR, 2017.
[26] S. Ravanbakhsh, S. Schneider, and B. Póczos. Deep learning with sets and point clouds. In ICLR, 2017.
[27] R. Reininghaus, U. Bauer, S. Huber, and R. Kwitt. A stable multi-scale kernel for topological machine
learning. In CVPR, 2015.
[28] G. Singh, F. Memoli, T. Ishkhanov, G. Sapiro, G. Carlsson, and D.L. Ringach. Topological analysis of
population activity in visual cortex. J. Vis., 8(8), 2008.
10
[29] K. Turner, S. Mukherjee, and D. M. Boyer. Persistent homology transform for modeling shapes and
surfaces. Inf. Inference, 3(4):310–344, 2014.
[30] X. Wang, B. Feng, X. Bai, W. Liu, and L.J. Latecki. Bag of contour fragments for robust shape classification.
Pattern Recognit., 47(6):2116–2125, 2014.
[31] P. Yanardag and S.V.N. Vishwanathan. Deep graph kernels. In KDD, 2015.
This supplementary material contains technical details that were left-out in the original submission
for brevity. When necessary, we refer to the submitted manuscript.
C
Additional proofs
In the manuscript, we omitted the proof for the following technical lemma. For completeness, the
lemma is repeated and its proof is given below.
Lemma 4. Let α ∈ R+ , β ∈ R and γ ∈ R+ . We have
(i) lim
x→0
ln(x)
x
1
x→0 x
(ii) lim
2
· e−α(ln(x)γ+β) = 0
2
· e−α(ln(x)γ+β) = 0 .
Proof. We only need to prove the first statement, as the second follows immediately. Hence, consider
2
ln(x) −α(ln(x)γ+β)2
·e
= lim ln(x) · e− ln(x) · e−α(ln(x)γ+β)
x→0
x→0
x
2
= lim ln(x) · e−α(ln(x)γ+β) −ln(x)
x→0
−1
2
= lim ln(x) · eα(ln(x)γ+β) +ln(x)
x→0
−1
∂
∂ α(ln(x)γ+β)2 +ln(x)
(∗)
= lim
ln(x) ·
e
x→0 ∂x
∂x
−1
γ
1
1
α(ln(x)γ+β)2 +ln(x)
· 2α(ln(x)γ + β) +
= lim · e
x→0 x
x x
−1
2
= lim eα(ln(x)γ+β) +ln(x) · (2α(ln(x)γ + β)γ + 1)
lim
x→0
=0
where we use de l’Hôpital’s rule in (∗).
D
Network architectures
2D object shape classification. Fig. 6 illustrates the network architecture used for 2D object shape
classification in [Manuscript, Sec. 5.1]. Note that the persistence diagrams from three consecutive
filtration directions di share one input layer. As we use 32 directions, we have 32 input branches.
The convolution operation operates with kernels of size 1 × 1 × 3 and a stride of 1. The max-pooling
operates along the filter dimension. For better readability, we have added the output size of certain
layers. We train with the network with stochastic gradient descent (SGD) and a mini-batch size of
128 for 300 epochs. Every 20th epoch, the learning rate (initially set to 0.1) is halved.
11
Graph classification. Fig. 7 illustrates the network architecture used for graph classification in
Sec. 5.2. In detail, we have 3 input branches: first, we split 0-dimensional features into essential
and non-essential ones; second, since there are only essential features in dimension 1 (see Sec. 5.2,
Filtration) we do not need a branch for non-essential features. We train the network using SGD with
mini-batches of size 128 for 300 epochs. The initial learning rate is set to 0.1 (reddit_5k) and 0.4
(reddit_12k), resp., and halved every 20th epochs.
D.1
Technical handling of essential features
In case of of 2D object shapes, the death times of essential features are mapped to the max. filtration
value and kept in the original persistence diagrams. In fact, for Animal and MPEG-7, there is always
only one connected component and consequently only one essential feature in dimension 0 (i.e., it
does not make sense to handle this one point in a separate input branch).
In case of social network graphs, essential features are mapped to the real line (using their birth time)
and handled in separate input branches (see Fig. 7) with 1D structure elements. This is in contrast to
the 2D object shape experiments, as we might have many essential features (in dimensions 0 and 1)
that require handling in separate input branches.
12
di−1 di di+1 (Filtration directions)
,...,
,...,
(for 32 directions, we have 32 3-tuple of persistence diagrams as input)
Input layer
N=75
Output: 3 × 75
...
this branch is repeated 32 times
Convolution
Filters: 3->16
Kernel: 1
Stride: 1
Convolution
Filters: 16->4
Kernel: 1
Stride: 1
Output: 16 × 75
Max-pooling
Output: 4 × 75
Linear
Output: 1 × 75
...
75->25
BatchNorm
Linear
25->25
ReLU
Concatenate
Linear
32· 25->100
BatchNorm
Linear
100->70 (in case of Animal, 100->20 )
Dropout
Cross-entropy loss
Figure 6: 2D object shape classification network architecture.
13
Input data (obtained by filtrating the graph by vertex degree)
0-dim. homology
1-dim. homology
es
ial
ial
nt
nt
se
se
es
essential features
Input layer
2D
Input layer
N=150
Linear
1D
Input layer
N=50
Linear
BatchNorm
Linear
Linear
25->25
25->25
75->75
ReLU
N=50
50->25
50->25
BatchNorm
Linear
1D
Linear
150->75
BatchNorm
essential features
ReLU
ReLU
Concatenate
Linear
125->200
BatchNorm + ReLU
Linear
200->100
BatchNorm
ReLU
Linear
100->50
BatchNorm + ReLU
Linear
100->5
BatchNorm
Dropout
Cross-entropy loss
Figure 7: Graph classification network architecture.
14
| 1cs.CV
|
A likelihood ratio approach to sequential change point
detection
Holger Dette, Josua Gösmann
arXiv:1802.07696v1 [math.ST] 21 Feb 2018
Ruhr-Universität Bochum
Fakultät für Mathematik
44780 Bochum
Germany
February 22, 2018
Abstract
In this paper we propose a new approach for sequential monitoring of a parameter
of a d-dimensional time series. We consider a closed-end-method, which is motivated
by the likelihood ratio test principle and compare the new method with two alternative
procedures. We also incorporate self-normalization such that estimation of the longrun variance is not necessary. We prove that for a large class of testing problems the
new detection scheme has asymptotic level α and is consistent. The asymptotic theory
is illustrated for the important cases of monitoring a change in the mean, variance and
correlation. By means of a simulation study it is demonstrated that the new test performs
better than the currently available procedures for these problems.
AMS subject classification: 62M10, 62G10, 62G20, 62L99
Keywords and phrases: change point analysis, self-normalization, sequential monitoring,
likelihood ratio principle
1
Introduction
An important problem in statistical modeling of time series is the problem of testing for structural stability as changes in the data generating process may have a substantial impact on
statistical inference developed under the assumption of stationarity. Because of its importance
there exists a large amount of literature, which develops tests for structural breaks in various
models and we refer to Aue and Horváth (2013) and Jandhyala et al. (2013) for more recent
reviews of the literature. There are essentially to ways how the problem of change point analysis is addressed. A large portion of the literature discusses a-posteriori change point analysis,
1
where the focus is on the detection of structural breaks given a historical data set [see Davis
et al. (1995), Csörgö and Horváth (1997), Aue et al. (2009b), Jirak (2015) among many others].
On the other hand, in many applications, such as engineering, medicine or risk management
data arrives steadily and therefore several authors have addressed the problem of sequentially
monitoring changes in a parameter. Page (1954), Moustakides (1986) and Lai (1995) among
others developed detection schemes for models with an infinite time horizon. These always stop
and different methods are compared in terms of the average run length. An alternative monitoring ansatz, which on the one hand allows to control (asymptotically) the type I error (if no
changes occur) and on the other hand provides the possibility of power analysis was introduced
by Chu et al. (1996) in the context of testing the structural stability of the parameters in a
linear model. Horváth et al. (2004), Fremdt (2014) extended this approach for linear models
with infinite time horizon, while Aue et al. (2012), Wied and Galeano (2013), Pape et al. (2016)
developed monitoring procedures for changes in a capital asset pricing model, correlation and
variance under the assumption of a finite time horizon.
In this paper we propose a general alternative sequential test in this context, which is
applicable for change point analysis of a p-dimensional parameter of a d-dimensional time series
if a historical data set from a stable phase is available and then data arrives consecutively. Our
approach differs from the current methodology as it is motivated by a likelihood ratio test for
a structural change. To be precise, Wied and Galeano (2013) proposed to compare estimates
from the historical data set, say X1 , . . . , Xm , with estimators from the sequentially observed
data Xm+1 , . . . , Xm+k [see also Chu et al. (1996); Horváth et al. (2004); Aue et al. (2012);
Pape et al. (2016)], while Fremdt (2014) and Kirch and Weber (2017) suggested to compare
the estimate from the historical data set with estimates from the sequentially observed data
Xm+j+1 , . . . , Xm+k (for all j = 0, . . . , k − 1). In contrast, motivated by the likelihood ratio
principle, our approach sequentially compares the estimates from the samples X1 , . . . , Xm+j
and Xm+j+1 , . . . , Xm+k (for all j = 0, . . . , k − 1).
Moreover, we also propose a self-normalized test, which avoids the problem of estimating
the long-run variance. While the concept of self-normalization has been studied intensively
for a-posteriori change point analysis [see Shao and Zhang (2010) and Shao (2015) among
many others], to our best knowledge self-normalization has not been studied in the context of
sequential monitoring.
The statistical model and the change point problem for a general parameter of the marginal
distribution are introduced in Section 2, where we also provide the motivation for the statistic
used in the sequential scheme (see Example 2.1) and a discussion of the alternative methods.
The asymptotic properties of the new monitoring scheme are investigated in Section 3. In particular we prove a result, which allows to control (asymptotically) the probability of indicating
a change in the parameter although there is in fact structural stability (type I error). Moreover,
we also show that the new test is consistent and investigate the concept of self-normalization in
this context. These asymptotic considerations require several assumptions, which are stated in
the general context and verified in Section 4 fro the case of monitoring changes in the mean and
2
variance matrix. In Section 5 the finite sample properties of the new procedure are investigated
by means of a simulation study, and we also demonstrate the (empirical) superiority of our
approach. Finally, all proofs are deferred to a technical appendix (see Section A).
2
Sequential change point testing
Consider a d-dimensional time series {Xj }j∈Z , where Xj has distribution function Fj , and
denote by θj = θ(Fj ) a p-dimensional parameter of interest of the distribution of Fj . We are
taking the sequential point of view and assume that there exists a historical period of length,
say m ∈ N, such that the process is stable in the sense
θ1 = θ2 = · · · = θm .
(2.1)
We are interested to monitor if the parameter θt changes in the future m + k ≥ m + 1. The
sequence X1 , . . . , Xm is usually referred to a historical or initial training data set, see for example
Chu et al. (1996), Horváth et al. (2004), Wied and Galeano (2013) or Kirch and Weber (2017),
among many others. Based on this stretch of “stable” observations a sequential procedure
should be conducted to test the hypotheses
H0 : θ1 = · · · = θm = θm+1 = θm+2 = . . . ,
(2.2)
against the alternative that the parameter θm+k? changes for some k ? ≥ 1, that is
H1 : ∃k ? ∈ N : θ1 = · · · = θm+k? −1 6= θm+k? = θm+k? +1 = . . . ,
(2.3)
In order to motivate our approach in particular the test statistic used in the detection scheme,
which will be used in the proposed sequential test, we begin with a very simple example of a
change in the mean.
Example 2.1 Consider a sequence {Xt }t∈Z of independent, d-dimensional normal distributed
random variables with (positive definite) variance matrix Σ and mean vectors
Z
µt = E[Xt ] = θ(Ft ) =
xdFt (x) ∈ Rd , j = 1, 2, . . . .
(2.4)
Rd
During the monitoring procedure, we propose to successively test the hypotheses
H0 : µ1 = · · · = µm = µm+1 = · · · = µm+k ,
versus
(k)
HA : ∃ j ∈ {0, . . . , k − 1} : µ1 = · · · = µm+j 6= µm+j+1 = · · · = µm+k
(2.5)
based on the sample X1 , . . . , Xm+k . Under the assumptions made in this example we can easily
derive the likelihood ratio
m+k
Q
sup
f (Xt , µ)
µ∈Rd t=1
Λm (k) =
sup
m+j
Q
f (Xt , µ(1) ) ·
j∈{0,...,k−1} t=1
µ(1) ,µ(2) ∈Rd
m+k
Q
t=m+j+1
3
,
f (Xt , µ(2) )
where f (·, µ) denotes the density of a normal distribution with mean µ and variance matrix Σ
(note that the first m observations are assumed to be mean-stable). A careful calculation now
proves the identity
> −1 m+j
k−1 (m + j)(k − j)
m+k
m+k
µ̂
−
µ̂
Σ
−2 log Λm (k) = max
µ̂m+j
−
µ̂
1
1
m+j+1
m+j+1
j=0
m+k
> −1 m+j
k−1 (m + k)(m + j)
,
= max
Σ µ̂1 − µ̂m+k
µ̂m+j
− µ̂m+k
1
1
1
j=0
(k − j)
(2.6)
where v > denotes the transposed of the vector v (usually considered as a column vector) and
j
µ̂ji
X
1
=
Xt
j − i + 1 t=i
(k)
is the mean of the observations Xi , . . . , Xj . Consequently the null hypothesis H0
(k)
rejected in favor of the alternative HA for large values of the statistic
k−1
max(m + j)(k − j) µ̂m+j
− µ̂m+k
1
m+j+1
>
j=0
should be
Σ−1 µ̂m+j
− µ̂m+k
1
m+j+1 .
(2.7)
However, as pointed out in Csörgö and Horváth (1997), the asymptotic properties of a likelihood
ratio type statistic of the type (2.7) are difficult to study. For this reason we propose to use a
weighted version of (2.7) and consider the statistic
k−1
max(m + j)2 (k − j)2 µ̂m+j
− µ̂m+k
1
m+j+1
>
j=0
m+j
k−1
= max (k − j)
j=0
X
m+k
X
Xt − (m + j)
t=1
Σ−1 µ̂m+j
− µ̂m+k
1
m+j+1
m+j
Xt Σ
−1
(k − j)
X
Xt − (m + j)
t=1
t=m+j+1
(2.8)
m+k
X
Xt ,
t=m+j+1
for which (after appropriate normalization) weak convergence of a corresponding sequential
empirical process can be established. Note that the right-hand side in (2.8) corresponds to the
well known CUSUM statistic, which has become a standard tool for change point detection in
a retrospective setting.
Motivated by the previous example we propose to use the statistic
k−1
m+k
D̂m (k) = m−3 max (m + j)2 (k − j)2 θ̂1m+j − θ̂m+j+1
j=0
>
m+j
m+k
− θ̂m+j+1
,
Σ̂−1
m θ̂1
(2.9)
for monitoring changes in the parameter θj , where θ̂ij = θ(F̂ij ) denotes the estimator obtained
from the empirical distribution function
j
F̂ij (z)
X
1
I{Xt ≤ z}
=
j − i + 1 t=i
4
of the observations Xi , . . . , Xj and the matrix Σ̂m corresponds to an estimator of a long-run
variance based on X1 , . . . , Xm . The scaling by m−3 will be necessary to obtain weak convergence
in the sequel.
We use the sequence {D̂m (k)}k∈N in combination with an increasing threshold function w(·)
as a monitoring scheme. More precisely, let T ∈ N denote a constant factor defining the window
of monitoring, then we reject the null hypothesis in (2.2) at the first time k ∈ {1, 2, . . . , T m}
for which the detector D̂m (k) exceeds the threshold function w : [0, T ] → R+ , that is
D̂m (k) > w(k/m) .
(2.10)
This definition yields a stopping rule defined by
n
o
τm = inf 1 ≤ k ≤ T m | D̂m (k) > w(k/m) ,
(if the set {1 ≤ k ≤ T m | D̂m (k) > w(k/m)} is empty we define τm = ∞). The threshold
function has to be chosen such that the test has asymptotic level α, that is
D̂m (k)
mT
>1 ≤α,
(2.11)
lim sup PH0 τm < ∞ = lim sup PH0 max
k=1 w(k/m)
m→∞
m→∞
and is consistent, i.e.
lim PH1 τm < ∞ = 1 .
m→∞
(2.12)
Following Aue et al. (2012) we call this procedure a closed-end method, because monitoring is
only performed in the interval [m + 1, mT ].
To our best knowledge, the detection scheme defined by (2.10) has not been considered in
the literature so far. However, our approach is related to the work of Chu et al. (1996); Aue et al.
(2012); Wied and Galeano (2013) and Pape et al. (2016), who investigated sequential monitoring
schemes for various parameters (such as the correlation, the variance or the parameters of the
capital asset pricing model). In the general situation considered in this section their approach
uses
Q̂m (k) =
k2 m
m+k > −1 m
m+k
θ̂1 − θ̂m+1
Σ̂m (θ̂1 − θ̂m+1
m
(2.13)
as a basic statistic in the sequential procedure. Note that a sequential scheme based on this
statistic measures the differences between the estimator θ̂1m from the initial data and the estimam+k
tor θ̂m+1
from all observations excluding the training sample. As a consequence - in particular
m+k
may be corrupted by observations
in the case of a rather late change - the estimator θ̂m+1
before the change point, which might lead to a loss of power. Another related procedure uses
the statistic
k−1
P̂m (k) = max
j=0
> −1 m
(k − j)2 m
m+k
m+k
θ̂1 − θ̂m+j+1
Σ̂m θ̂1 − θ̂m+j+1
m
5
(2.14)
and was recently suggested by Fremdt (2014) and reconsidered by Kirch and Weber (2017).
These authors compare the estimate from the data X1 , . . . , Xm with estimates from the data
Xm+j+1 , . . . , Xm+k (for different values of j). This method may lead to a loss in power in
problems with a small sample of historical data and a rather late change point. In contrast our
approach compares the estimates of the parameters before and after all potential positions of
a change point j ∈ {m + 1, . . . , m + k}.
In this paper, we argue that the performance of the change point tests can be improved by
replacing θ̂1m by θ̂1m+j inside the maximum, which would directly lead to a scheme of the form
(2.9). Here, we would like to point to our simulation study in Section 5, which contains many
cases where a sequential detection scheme based on the statistic D̂m outperforms schemes based
on Q̂m or P̂m .
3
Asymptotic properties
In the subsequent discussion we use the following notation. We denote by `∞ (V1 , V2 ) the space
of all bounded functions f : V1 → V2 equipped with sup-norm, where V1 , V2 are normed linear
P
D
spaces. The symbols −→ and =⇒ mean convergence in probability and weak convergence (in the
space under consideration), respectively. The process {W (s)}s∈[0,T +1] will usually represent a
Pd 2 1/2
standard p-dimensional Brownian motion. For a vector v ∈ Rd , we denote by |v| =
i=1 vi
its euclidean norm.
3.1
Weak convergence
Throughout this paper we denote by θ a p-dimensional functional of the d-dimensional distribution function F and define its influence function (assuming its existence) by
θ((1 − ε)F + εδx ) − θ(F )
,
ε&0
ε
IF(x, F, θ) = lim
(3.1)
where δx (z) = I{x ≤ z} is the distribution function of the Dirac measure at the point x ∈ Rd
and the inequality in the indicator is understood component-wise. Throughout this section
we make the following assumptions, which will be verified for several important examples in
Section 4 [see also Shao and Zhang (2010) for similar regularity conditions].
Assumption 3.1 Under the null hypothesis (2.2) we assume that the times series {Xt }t∈N is
strictly stationary with E[IF(X1 , F, θ)] = 0 and that the weak convergence
bmsc
p
1 X
D
√
IF(Xt , F, θ) =⇒ ΣF W (s) ,
m t=1
6
(3.2)
holds in the space `∞ ([0, T + 1], Rp ) as m → ∞, where the long-run variance matrix is defined
by (assuming convergence of the series)
X
ΣF =
Cov IF(X0 , F, θ), IF(Xt , F, θ) ∈ Rp×p
(3.3)
t∈Z
and {W (s)}s∈[0,T +1] is a p-dimensional (standard) Brownian motion.
Assumption 3.2 The remainder terms
j
Ri,j =
θ̂ij
X
1
IF(Xt , F, θ)
− θ(F ) −
j − i + 1 t=i
(3.4)
in the linearization of θ̂ij − θ(F ) satisfy
sup (j − i + 1)|Ri,j | = oP (n1/2 ) .
(3.5)
1≤i<j≤n
For the statement of our first result we introduce the notation (throughout this paper we use
the convention θ̂zu = 0, whenever z > u)
z
u
− θ̂z+1
Ũ(`, z, u) := (u − z)(z − `) θ̂`+1
,
(3.6)
u
U(z, u) := Ũ(0, z, u) = (u − z)z θ̂1z − θ̂z+1
,
(3.7)
and denote by
∆2 = {(s, t) ∈ [0, T + 1]2 | s ≤ t} ,
∆3 = (r, s, t) ∈ [0, T + 1]3 | r ≤ s ≤ t .
(3.8)
(3.9)
the 2- and 3-dimensional triangle in [0, T + 1]2 and [0, T + 1]3 , respectively.
Theorem 3.3 Let Assumptions 3.1 and 3.2 be satisfied. If the null hypothesis in (2.2) holds,
then as m → ∞
−3/2
m
Ũ(bmrc, bmsc, bmtc)
D
(r,s,t)∈∆3
1/2
=⇒ ΣF
B(s, t) + B(r, s) − B(r, t)
(r,s,t)∈∆3
(3.10)
in the space `∞ (∆3 , Rp ), where the process {B(s, t)}(s,t)∈∆2 is defined by
B(s, t) = tW (s) − sW (t) , (s, t) ∈ ∆2 ,
(3.11)
and {W (s)}s∈[0,T +1] denotes a p-dimensional Brownian motion on the interval [0, T + 1].
7
Remark 3.4 As a by-product of Theorem 3.3 and the representation (3.7) we obtain the weak
convergence of the double-indexed CUSUM-process (3.7), that is (m → ∞)
m−3/2 · U(bmsc, bmtc)
D
(s,t)∈∆2
1/2
=⇒ {ΣF B(s, t)}(s,t)∈∆2 ,
(3.12)
where ∆2 denotes the 2-dimensional triangle in [0, T +1]2 and the process B is defined in (3.11).
In particular the covariance structure of the process B is given by
Cov B(s1 , t1 ), B(s2 , t2 ) = t1 t2 (s1 ∧ s2 ) − t1 s2 (s1 ∧ t2 ) − s1 t2 (t1 ∧ s2 ) + s1 s2 (t1 ∧ t2 ) .
Consequently, the process {B(s, t)}(s,t)∈∆2 can be considered as a natural extension of the
standard Brownian bridge as for fixed t the process {B(s, t)}s∈[0,t] is a Brownian bridge on the
interval [0, t].
Observing the definition (3.7) the statistic (2.9) allows the representation
k−1
D̂m (k) = m−3 max |U> (m + j, m + k)Σ̂−1
m U(m + j, m + k)| ,
j=0
(3.13)
and we obtain the following Corollary as a consequence of Theorem 3.3.
Corollary 3.5 Let the assumptions of Theorem 3.3 be satisfied. If the null hypothesis in (2.2)
holds, and Σ̂m denotes a consistent estimator of the long-run variance ΣF defined in (3.3), then
as m → ∞
Tm
max
k=1
D̂m (k) D
B(s, t)> B(s, t)
=⇒ sup sup
.
w(k/m)
w(t − 1)
t∈[1,T +1] s∈[1,t]
for any threshold function w : [0, T ] → R+ , which is increasing.
By the result in Corollary 3.5 it is reasonable to choose for a given level α a threshold function
wα (·), such that
B(s, t)> B(s, t)
P
sup sup
>1 =α
(3.14)
wα (t − 1)
t∈[1,T +1] s∈[1,t]
and to reject the null hypothesis H0 in (2.2) at time k, if
D̂m (k) > wα (k/m) .
(3.15)
By Corollary 3.5 this test has asymptotic level α, that is
T m D̂ (k)
m
lim PH0 max
>1 =α
m→∞
k=1 wα (k/m)
(if the assumptions of Theorem 3.3 hold and wα satisfies (3.14)). The choice of of wα (·) has
been investigated by several authors [see Chu et al. (1996), Aue et al. (2009b) and Wied and
8
Galeano (2013) among others] and we will compare different options by means of a simulation
study in Section 5. Note that one can take any function (which is increasing and bounded from
below by a positive constant) and multiply an appropriate constant such that (3.14) is fulfilled.
Next we discuss the consistency of the monitoring scheme (3.15). For this purpose we consider
the alternative hypothesis in (2.3), where the location of the change point is increasing with
the length of the training sample, that is m + k ∗ = bmcc for some 1 < c < T + 1. Recalling the
definition of D̂m (k) in (2.9) and observing the inequality
2
bmcc
T (m+1) >
bmcc
T (m+1)
θ̂1 − θ̂bmcc+1 Σ̂−1
− θ̂bmcc+1
bmcc2 T (m + 1) − bmcc
Tm
m θ̂1
max D̂m (k) ≥
·
k=1
m3
wα (T )
bmcc
m
it is intuitively clear that the statistic maxTk=1
D̂m (k) converges to infinity, provided that θ̂1
T (m+1)
and θ̂bmcc+1 are consistent estimates of the parameter θ before and after the change point
m + k ∗ = bmcc and Σ̂m converges to a positive definite p × p matrix. The following Theorem
3.8 makes these heuristic arguments more precise. Its proof requires several assumptions, which
are stated first. The result might be even correct under slightly weaker assumptions. However,
in the form stated below we can also prove consistency of a sequential scheme based on a
self-normalized version of D̂m (k) (see Theorem 3.10 in Section 3.2).
Assumption 3.6 If the alternative hypothesis H1 defined in (2.3) holds we assume that the
change occurs at position m + k ∗ = bmcc for some c ∈ (1, T + 1). Moreover, let {Zt (1)}t∈Z and
{Zt (2)}t∈Z denote (strictly) stationary Rd -valued processes with marginal distribution functions
F (1) and F (2) , respectively, such that
θ(F (1) ) 6= θ(F (2) ) ,
and that for each m ∈ N
D
(X1 , X2 , . . . , Xbmcc ) = (Z1 (1), . . . , Zbmcc (1)),
D
(Xbmcc+1 , . . . , XbmT c ) = (Zbmcc+1 (2), . . . , ZbmT c (2)) .
(3.16)
(3.17)
Note, that formally the process {Xt }t=1,...bmT c is a triangular array, that is Xt = Xm,t , but we do
not reflect this in our notation. Further, we assume that there exist two (standard) Brownian
motions W1 and W2 such that the joint weak convergence
!
!
1 Pbmsc
p
(1)
√
IF(Z
(1),
F
,
θ)
Σ
W
(s)
(1)
t
1
F
D
t=1
m
s∈[0,c]
s∈[0,c]
p
1 P
=⇒
bmsc
(2)
√
Σ
W
(s)
−
W
(c)
(2)
IF(Z
(2),
F
,
θ)
2
2
t
F
t=bmcc+1
s∈[c,T +1]
m
s∈[c,T +1]
holds, where ΣF (1) and ΣF (2) denote positive definite matrices defined in the same way as (3.3),
that is
X
ΣF (`) =
Cov IF(Z0 (`), F (`) , θ), IF(Zt (`), F (`) , θ) ∈ Rp×p , ` = 1, 2
(3.18)
t∈Z
and both phases in (3.16) and (3.17) fulfill Assumption 3.2 for the corresponding expansion.
9
Remark 3.7
(a) The interpretation of Assumption 3.6 is as follows: There exist two regimes and the
process under consideration switches from one regime to the other.
(b) The assumption of two stationary phases before and after the change point is commonly
made in the literature to analyze change point tests under the alternative [see for example
Aue et al. (2009b), Dette and Wied (2016), Kirch and Weber (2017) among others]. Note,
that we do not assume that the two limiting processes W1 and W2 are independent.
(c) Often Assumption 3.6 is directly implied by the underlying change point problem. For
example, in the situation of the mean vector introduced in (2.1) it is usually assumed
that Xt = µt + εt , where {εt }t∈Z is a stationary process and µt = µ(1) if t ≤ bmcc and
µt = µ(2) if t ≥ bmcc + 1. In this case Assumption 3.6 is obviously satisfied. Further
examples are discussed in Section 4.
Theorem 3.8 Let Assumption 3.6 be satisfied and let the threshold function wα satisfy (3.14).
Further assume that Σ̂m denotes a consistent estimator of the long-run variance ΣF (1) based on
the observations X1 , . . . , Xm . Under the alternative hypothesis H1 we have
D̂m (k)
Tm
lim P max
>1 =1.
m→∞
k=0 wα (k/m)
Remark 3.9 We can establish similar results for the statistics (2.13) and (2.14) proposed by
Wied and Galeano (2013) and Fremdt (2014), respectively. For example, if the assumptions of
Theorem 3.3 are satisfied we obtain the weak convergence
Tm
max
k=1
Q̂m (k) D
B(t, 1)> B(t, 1)
=⇒ sup
,
w(k/m)
w(t − 1)
t∈[1,T +1]
>
B(s, 1) + B(t, 1)
B(s, 1) + B(t, 1)
P̂m (k) D
max
=⇒ sup sup
k=1 w(k/m)
w(t − 1)
t∈[1,T +1] s∈[1,t]
Tm
under the null hypothesis, which can be used to construct an asymptotic level α monitoring
scheme based on the statistics Q̂m and P̂m , respectively. Consistency of the corresponding tests
follows along the arguments given in the proof of Theorem 3.8. The details are omitted for the
sake of brevity. The finite sample properties of the three different tests will be investigated by
means of a simulation study in Section 5.
3.2
Self-Normalization
The test proposed in Section 3.1 requires an estimator of the long-run variance Σ̂m , and we
discuss commonly used estimates for this purpose in Section 4. However, it has been pointed
10
out by several authors that this problem is not an easy one as the common estimates depend
sensitively on a regularization parameter (for example a bandwidth), which might be difficult
to select in practice. An alternative to long-run variance estimation is the concept of selfnormalization, which will be investigated in this section. This approach has been studied
intensively for a-posteriori change point analysis [see Shao and Zhang (2010) and Shao (2015)
among many others], but - to our best knowledge - self-normalization has not been considered
in the context of sequential monitoring. In the following we discuss a self-normalized version of
the statistic D̂m proposed in this paper (see equation (2.9)). Self-normalization of the statistic
P̂m in (2.14) will be briefly discussed in Remark 3.12.
To be precise, we define a self-normalizing matrix
V(z, u) =
+
z
X
>
j
z
z
θ̂1 − θ̂j+1
j 2 (z − j)2 θ̂1j − θ̂j+1
j=1
u
X
(3.19)
2
2
(u − j) (j − z)
j
θ̂z+1
−
u
θ̂j+1
j
θ̂z+1
−
>
u
θ̂j+1
.
j=z+1
and replace the estimate Σ̂m of the long-run variance in (2.9) by the matrix
This yields the self-normalized statistic
k−1
m+k
D̂SN (k) = m max (m + j)2 (k − j)2 θ̂1m+j − θ̂m+j+1
>
j=0
1
V(m + j, m + k).
m4
m+k
V−1 (m + j, m + k) θ̂1m+j − θ̂m+j+1
,
(3.20)
for which we prove the following result in Section A.
Theorem 3.10 Let w : [0, T ] → R+ denote any threshold function, which is increasing and
let the assumptions of Theorem 3.3 be satisfied. If the null hypothesis in (2.2) holds, then as
m→∞
Tm
max
k=1
SN
D̂m
(k) D
|B̃(s, t)|
=⇒ sup sup
,
w(k/m)
t∈[1,T +1] s∈[1,t] w(t − 1)
(3.21)
where
−1
B̃(s, t) = B > (s, t) N1 (s) + N2 (s, t) B(s, t) ,
the process {B(s, t)}s∈∆2 is defined in (3.11) and {N1 (s)}s∈[0,T +1] and {N2 (s, t)}(s,t)∈∆2 are given
by
Zs
N1 (s) =
B(r, s)B > (r, s)dr ,
(3.22)
0
Z
N2 (s, t) =
t
>
B(r, t) + B(s, r) − B(s, t) B(r, t) + B(s, r) − B(s, t) dr .
s
11
The monitoring rule is now defined in the same way as described in Section 3.1 determining a
threshold function wα (·), such that
|B̃(s, t)|
>1 =α
(3.23)
P
sup sup
t∈[1,T +1] s∈[1,t] wα (t − 1)
for a given level α, and rejecting the null hypothesis H0 in (2.2) at the time k, if
SN
D̂m
(k) > wα (k/m) .
(3.24)
By Theorem 3.10 this test has asymptotic level α and our next result shows that this procedure
is also consistent.
Theorem 3.11 Let Assumption 3.6 be satisfied and let the threshold function wα satisfy (3.23).
Under the alternative hypothesis H1 we have
T m D̂SN (k)
lim P max m
> 1 = 1.
m→∞
k=0 wα (k/m)
Remark 3.12 The statistic P̂m (k) defined (2.14) can be self-normalized in a similar manner,
that is
> −1
k
m+k
m+k
P̂mSN (k) = m3 max(k − j)2 θ̂1m − θ̂m+j+1
V (m + j, m + k) θ̂1m − θ̂m+j+1
.
(3.25)
j=0
If the the null hypothesis holds and the assumptions of Theorem 3.3 are satisfied it can be
shown using (3.12) and similar arguments as given in the proof of Theorem 3.10 that
Tm
max
k=1
P̂mSN (k) D
|B̃ SN (s, t)|
=⇒ sup sup
,
w(k/m)
t∈[1,T +1] s∈[1,t] w(t − 1)
(3.26)
where the process B̃ SN is defined by
B̃ SN (s, t) = B(s, 1) + B(t, 1)
>
−1
N1 (s) + N2 (s, t)
B(s, 1) + B(t, 1) .
Similarly, consistency follows along the lines given in the proof of Theorem 3.11. The details
are omitted for the sake of brevity.
On the other hand a statistic like Q̂m (k) defined in (2.13) cannot be self-normalized in a
straightforward manner as it does not employ a maximum, which is necessary to separate
points before and after the (possible) change point. In particular one cannot use the matrix V
in (3.19), which is based on such a separation. Obviously, a self-normalization approach without
separation could be constructed but this would lead to a severe loss in power and is therefore not
discussed here. We refer the reader to Shao and Zhang (2010) for a comprehensive discussion of
SN
this problem. The finite sample properties of both self-normalized methods D̂m
(k) and P̂mSN (k)
will be compared by means of a simulation study in Section 5 .
12
3.3
Implementation
We will close this section with a description of the algorithm to detect changes in the functional θ(F ) employing self-normalization. In the following paragraph Ŝm (k) denotes any of the
SN
(k), P̂m (k), P̂mSN (k) and Q̂m (k) discussed in Section 3.1 and 3.2.
statistics D̂m (k), D̂m
Algorithm 3.13 Let {X1 , . . . , Xm } denote the “stable” training data satisfying (2.1).
(Initialization) Choose a threshold function wα such that the probability of type I error
is asymptotically α. Further choose the factor T to determine how much longer the
monitoring can be performed.
(Monitoring) If Xm+k has been observed, compute the the statistic Ŝm (k) and reject the
null hypothesis of no change in the parameter θ(F ) if Ŝm (k) > wα (k/m). In this case
stop monitoring. Otherwise, repeat this comparison with the next observation Xm+k+1 .
(Stop) If there has been no rejection at time m + mT , stop monitoring with observation
Xm+mT and conclude that no change has occurred within the monitoring period.
4
Some specific change point problems
In this section we illustrate how the assumptions of Section 3 can be verified for concrete
functionals. Exemplarily we consider the mean and variance, but similar arguments could be
given for other functionals under consideration. To be precise consider a time series {Xt }t∈Z ,
which forms a physical system in the sense of Wu and Rosenblatt (2005), that is
(
g(εt , εt−1 , . . . ) if t < bmcc
Xt =
(4.1)
h(εt , εt−1 , . . . ) if t ≥ bmcc ,
where {εt }t∈Z denotes a sequence of i.i.d. random variables with values in some measure space
S such that the functions g, h : SN → Rd are measurable. The functions g and h determine
the phases of the physical system before and after the change at position bmcc with c > 1,
respectively. Under the null hypothesis we will always assume that g and h coincide, which
yields that the (whole) times series {Xt }t∈Z is strictly stationary. In the case g 6= h the random
variables Xt form a triangular array, but for the sake of readability we do not reflect this in our
notation. In order to adapt the concept of physical dependence to the situation considered in
this paper, let ε00 be an independent copy of ε0 and define
(
g(εt , εt−1 , . . . , ε1 , ε00 , ε−1 , . . . ) if t < bmcc
0
Xt =
(4.2)
h(εt , εt−1 , . . . , ε1 , ε00 , ε−1 , . . . ) if t ≥ bmcc .
The distance
δt,q := E[|Xt − Xt0 |q ]
13
1/q
,
(4.3)
between Xt and its counterpart Xt0 is used to quantify the (temporal) dependence of the physical
system and δt,q measures the influence of ε0 on the random variable Xt . Further let
Θm,q =
∞
X
δt,q
(4.4)
t=m
denote the tail sum of the coefficients (which might diverge). Additionally, we define the
(ordinary) long-run variance matrix of the phases before and after the change by
X
Γ(g) =
Cov g(ε0 , ε−1 , . . . ), g(εt , εt−1 , . . . ) ,
t∈Z
Γ(h) =
X
Cov h(ε0 , ε−1 , . . . ), h(εt , εt−1 , . . . ) .
(4.5)
t∈Z
4.1
Sequential testing for changes in the mean vector
In this section we are interested in detecting changes in the mean
Z
xdFt (x) , t = 1, 2, . . . .
µ(Ft ) = E[Xt ] =
(4.6)
Rd
of a d-dimensional time series {Xt }t∈Z . Sequential detection schemes for a change in the mean
have been been investigated by Chu et al. (1996), Horváth et al. (2004) and Aue et al. (2009b)
among others. We consider the closed-end-procedure developed in Section 3 and assume that
the first m observations X1 , . . . , Xm to be mean-stable.
As the mean functional (4.6) is linear the influence function is given by
IF(x, F, θ) = x − µ(F ) = x − EF [X] ,
and therefore Assumption 3.2 is obviously satisfied (note that Ri,j = 0 for all i, j). Assumption
3.1 reduces to Donsker’s invariance principle, that is
o
n 1 bmsc
X
D 1/2
√
(Xt − E[Xt ])
=⇒ ΣF W (s) s∈[0,T +1]
(4.7)
m t=1
s∈[0,T +1]
P
in `∞ ([0, T + 1], Rd ) with ΣF = t∈Z Cov X0 , Xt , which has been derived by Wu and Rosenblatt (2005) for physical systems under the assumption that Θm,2 < ∞ (see See Theorem 3 in
this reference). Note that for this functional the (ordinary) long-run variance matrix Γ(g) and
ΣF coincide (under stationarity).
If the alternative of a change in the mean at position bmcc for some c ∈ (1, T + 1) holds,
we may assume that
h = g + ∆µ ,
(4.8)
where ∆µ = E[Xbmcc ] − E[Xbmcc−1 ]. Consequently, if Θm,2 < ∞, Assumption 3.6 is also satisfied
with W1 = W = W2 (see also the discussion at the end of Remark 3.7). We summarize these
observations in the following proposition.
14
Proposition 4.1 Assume that (4.1) holds with Θm,2 < ∞ and further let Σ̂m denote a consistent estimator of the (positive definite) long-run variance matrix ΣF (1) (before the change)
based on the observations X1 , . . . , Xm .
(a) If g = h, then the assumptions of Theorem 3.5 and 3.10 are satisfied for the functional
(4.6). In other words: The sequential tests for a change in the mean based on the statistics
R
D̂ or D̂SN with θ(Ft ) = Rd xdFt (x) have asymptotic level α.
(b) Let representation (4.8) hold with ∆µ 6= 0, then the assumptions of Theorem 3.8 and 3.11
are satisfied for the functional (4.6). In other words: The sequential tests for a change in
the mean based on the statistics D̂ or D̂SN are consistent.
The finite sample properties of this test will be investigated in Section 5.1.
4.2
Sequential testing for changes in the variance
In this section, we focus on detecting changes in the variance. Following Aue et al. (2009a),
who investigated this problem in the non-sequential case, we consider a time series {Xt }t∈Z
with common mean µ = EF [X] and define the functional
Z
Z
Z
>
V (F ) =
xx dF (x) −
xdF (x)
x> dF (x) .
(4.9)
Rd
Rd
Rd
A careful but straightforward calculation shows that the corresponding influence function is
given by
IF(x, F, V ) = −EF [XX > ] + 2EF [X]EF [X > ] + xx> − EF [X]x> − xEF [X]>
= (x − EF [X])(x − EF [X])> − V (F ) .
(4.10)
Hence the remainder term (under stationarity with X1 ∼ F ) in equation (3.4) is given by
j
Ri,j = V
Z
X
1
− V (F ) −
IF(Xt , F, V )
j − i + 1 t=i
Z
Z
j
j
>
xx dF̂i (x) −
xdF̂i (x)
x> dF̂ij (x) −
(F̂ij )
j
X
1
(Xi − E[X1 ])(Xi − E[X1 ])>
j
−
i
+
1
d
d
d
R
R
R
Z
Z
Z
Z t=i
=−
xdF̂ij (x)
x> dF̂ij (x) +
xE[X1> ]dF̂ij (x) +
E[X1 ]x> dF̂ij (x) − E[X1 ]E[X1> ]
d
d
d
d
R
RZ
Z R
R
=−
x − E[X1 ]dF̂ij (x)
x> − E[X1> ]dF̂ij (x) .
=
Rd
Rd
(4.11)
15
Define vech(·) to be the operator that stacks the columns of a symmetric d × d-matrix above
the diagonal as a vector of dimension d(d + 1)/2. As this operator is linear, it is obvious that
expansion (4.11) is equivalent to
j
vech(Ri,j ) = vech(V
(F̂ij ))
X
1
− vech(V ) −
IF v (Xt , F, V ) ,
j − i + 1 t=i
(4.12)
where IF v is defined as
IF v (Xt , F, V ) = vech IF(Xt , F, V ) = IF Xt , F, vech(V )
(4.13)
(Xt,1 − E[Xt,1 ])2 − E (Xt,1 − E[Xt,1 ])2
(X − E[X ])(X − E[X ]) − E(X − E[X ])(X − E[X ])
t,1
t,2
t,2
t,1
t,2
t,1
t,2
t,1
(Xt,2 − E[Xt,2 ])2 − E (Xt,2 − E[Xt,2 ])2
(Xt,1 − E[Xt,1 ])(Xt,3 − E[Xt,3 ]) − E (Xt,1 − E[Xt,1 ])(Xt,3 − E[Xt,3 ]) ,
=
(Xt,2 − E[Xt,2 ])(Xt,3 − E[Xt,3 ]) − E (Xt,2 − E[Xt,2 ])(Xt,3 − E[Xt,3 ])
2
2
(X
−
E[X
])
−
E
(X
−
E[X
])
t,3
t,3
t,3
t,3
..
.
and Xt,h denotes the h-th component of the vector Xi . We now provide sufficient conditions
such that the general theory in Section 3 is applicable for the functional vech(V ). Assumption
3.1 is satisfied if the time series {Xt }t∈Z is stationary and the invariance principle
bmsc
p
1 X
D
√
IF v (Xt , F, V ) =⇒ ΣF W (s) ,
m t=1
(4.14)
∗
holds in the space `∞ (R, Rd ), where W is a d∗ = d(d + 1)/2-dimensional Brownian motion and
ΣF is defined in (3.3) with IF = IF v . Invariance principles of the form (4.14) are well known
for many classes of weakly dependent time series. The required assumptions for the underlying
time series {Xt }t∈Z are typically the same as for the mean - except for some extra moment
conditions to cover the product structure of the random variables in (4.13). Condition (3.5) in
Assumption 3.2 reads as follows
j
j
X
X
1
Xt,k − E[Xt,k ]
sup
Xt,` − E[Xt,` ] = oP (n1/2 )
j
−
i
+
1
1≤i<j≤n
t=i
t=i
(4.15)
(1 ≤ k, ` ≤ d∗ ). The validity of this assumption depends on the underlying dependence
structure, in particular of the properties of the functions g and h in (4.1), and exemplarily we
give sufficient conditions in the following result, which is proved in the appendix.
Proposition 4.2 Assume that (4.1) holds with bounded functions g and h and δt,4 = O(ρt ) for
some ρ ∈ (0, 1). Let Σ̂m denote a consistent estimator of the long-run variance ΣF (1) (before
16
the change) based on the observations X1 , . . . , Xm . Further assume, that there exists a constant
ν > 0, such that Γ(g) − ν · Id and Γ(h) − ν · Id are positive definite matrices, where Γ(g) and
Γ(h) are defined in (4.5).
(a) If g = h, then the assumptions of Theorem 3.5 and 3.10 are satisfied for the functional
(4.9). In other words: The sequential tests for a change in the variance based on the
statistics D̂ or D̂SN have asymptotic level α.
(b) If h = A · g for some non-singular matrix A ∈ Rd×d with A · V (F (1) ) · A> 6= V (F (1) ), then
the assumptions of Theorem 3.8 and 3.11 are satisfied for the mean functional (4.6). In
other words: The sequential tests for a change in the variance based on the statistics D̂
or D̂SN are consistent.
Remark 4.3 The assumption of bounded observations in Proposition 4.2 is crucial to prove
the estimate (4.15). Essentially a proof of such a statement requires a version of Theorem 1 in
Shao (1995) for dependent random variables. The main ingredient for a proof of Shao’s result
is an Erdös-Renyi-Law of large numbers in the case of dependent random variables, which to the authors best knowledge - is only known for bounded random variables [see Kifer (2017)
for example]. On the other hand the assumption of bounded functions g and h in (4.1) is not
necessary for the functional (4.9) in the case of M -dependent time series.
5
Finite sample properties
In this section, we investigate the finite sample properties of the new detection schemes based
on the statistics D̂ and D̂SN in (2.9) and (3.20) and also provide a comparison to the detection
schemes based on Q̂, P̂ and P̂ SN , which are defined in (2.13), (2.14) and (3.25), respectively.
For the choice of the threshold function we follow the ideas of Horváth et al. (2004), Aue et al.
(2009b) and Wied and Galeano (2013) and consider the parametric family
2
w(t) = (t + 1) · max
t
t+1
2γ
,δ ,
(5.1)
where the parameter γ varies in the interval [0, 1/2) and δ > 0 is a small constant introduced to
avoid problems in the denominator of the ratio considered in (3.14). For the statistic Q̂ these
threshold functions are motivated by the law of iterated logarithm and are used to reduce the
stopping delay under the alternative hypothesis see Aue et al. (2009b) or Wied and Galeano
(2013)].
Note that we use the squared versions of the thresholds from the cited references, since we
consider statistics in terms of quadratic forms. To be precise consider three different threshold
functions
(T1) w(t) = cα ,
17
(T2) w(t) = cα (t + 1)2 ,
n t 1/2
o
−10
(T3) w(t) = cα (t + 1) · max
, 10
,
t+1
2
where the constant cα is chosen by Monte-Carlo simulations, such that
L(t)
P
sup
=α.
t∈[1,T +1] w(t − 1)
(5.2)
Here L denotes the limit process corresponding to the test statistic under consideration. For
the estimation of the long-run variance we employ the well-known quadratic spectral kernel
[see Andrews (1991)] and its implementation contained in the R-package ’sandwich’ [see Zeileis
(2004)]. For the sake of brevity, we will only display situations where the parameter T is fixed
as T = 1, i.e the monitoring period will always have the same size as the historical data set.
For the case T ≥ 2 we obtained a similar picture, which will not be displayed here. All results
that are presented below are based on 1000 simulation runs.
5.1
Changes in the mean
For the analysis of the new procedures in the problem of detecting changes in the mean we look
at independent data, a MA(3)- and an AR(1)-process, that is
(M1) Xt ∼ εt ,
(M2) Xj = 0.1Xt−1 + εt ,
(M3) Xj = εt + 0.3εt−1 + 0.1εt−2 ,
where {εt }t∈Z is a sequence of independent standard Gaussian random variables. In the case
of the alternative hypothesis we consider the sequence
(
Xt
if t < m + b m2 c
µ
Xt =
Xt + µ if t ≥ m + b m2 c
for various values of µ. For all discussed detection schemes the empirical rejection probabilities
for the models (M1) - (M3) and threshold functions (T1) - (T3) are shown in Figure 1 and 2
corresponding to the choice m = 50 and m = 100 as initial sample size. The results can be
summarized as follows. The statistic D̂ outperforms P̂ and Q̂ with respect to the power for
all combinations of the model and threshold function. Further the statistic P̂ shows a better
performance with respect to power as Q̂ in all cases under consideration. For example, the plot
in the left-upper corner of Figure 1 shows, that D̂ already has empirical power close to 1 (0.94)
for a change of size 1, while P̂ and Q̂ have only empirical power of 0.85 and 0.73, respectively.
This relation is basically the same in all plots contained in Figures 1 and 2. Moreover the
18
results show no substantial differences between the different threshold functions and for this
reason we will only consider the constant threshold (T1) in the remaining part of this paper.
In Figures 3 we compare the power of the tests based on the self-normalized statistics D̂SN
and P̂ SN . The results are similar as before and the empirical power obtained by the use of D̂SN
is considerably higher. Only for model (M3) and m = 50 the statistic based on P̂ SN is slightly
more powerful for small values of µ, but this superiority is achieved by a to high nominal level
(see the right panel in Figure 3). A comparison of the results with the corresponding rejection
probabilities in Figures 1 and 2 shows, that self-normalization yields a substantial loss of power
in the sequential detection schemes and we also observe this loss in the examples considered in
Section 5.2 and 5.3 (these results are not displayed for the sake of brevity).
On the other hand the approximation of the nominal level is more stable with respect to
different dependence structures for self-normalized methods. To illustrate this fact we display
in Table 1 the type-I error for all five sequential monitoring schemes based on the statistics D̂,
P̂ and Q̂. The results provide some empirical evidence that the self-normalized statistics yield
a more stable approximation of the nominal level.
(M1)
(M2)
(M3)
(T1)
(T2)
(T3)
Figure 1: Empirical rejection probabilities of the sequential tests for a change in the mean based
on the statistics D̂ (solid line), P̂ (dashed line) , Q̂ (dotted line). The initial and total sample
size are m = 50 and m(T + 1) = 100, respectively, and the change occurs at observation 75.
The level is α = 0.05. Different rows correspond to different threshold functions, while different
columns correspond to different models.
19
(M1)
(M2)
(M3)
(T1)
(T2)
(T3)
Figure 2: Empirical rejection probabilities of the sequential tests for a change in the mean based
on the statistics D̂ (solid line), P̂ (dashed line), Q̂ (dotted line). The initial and total sample
size are m = 100 and m(T + 1) = 200, respectively, and the change occurs at observation 150.
The level is α = 0.05. Different rows correspond to different threshold functions, while different
columns correspond to different models.
20
(M1)
(M2)
(M3)
Figure 3: Empirical rejection probabilities of the sequential tests for a change in the mean based
on the self-normalized tests statistics D̂SN (solid line), P̂ SN (dashed line). The initial and total
sample size are m = 50 and m(T + 1) = 100 (upper panel, change at observation 75) and
m = 100 and m(T + 1) = 200 (lower panel, change at observation 150). The level is α = 0.05
and the threshold function is given by (T1).
m
50
model \ statistic
(M1)
(M2)
(M3)
D̂
7.0%
8.1%
3.9%
P̂
6.0%
7.5%
4.0%
Q̂
6.8%
8.2%
4.3%
D̂SN
5.4%
5.7%
5.5%
P̂ SN
6.0%
6.7%
6.8%
(M1)
(M2)
(M3)
5.8%
5.9%
2.8%
6.0%
6.0%
3.4%
6.2%
6.0%
3.8%
5.2%
5.2%
5.1%
6.6%
6.8%
5.3%
100
Table 1: Simulated type I error (level α = 0.05) of the sequential tests for a change in the mean
based on the statistics D̂, P̂ , Q̂, D̂SN and P̂ SN . The threshold function is (T1).
5.2
Changes in the variance
In this subsection we present a small simulation study investigating the performance of the
detection schemes for a change in the variance matrix. We consider the following models
(V1) Xj = εj ,
(V2) Xj = A1 Xj−1 + εj ,
(V3) Xj = εj + A2 εj−1 + A3 εj−2 .
21
(V1)
(V2)
(V3)
(T1)
Figure 4: Empirical rejection probabilities of the sequential tests for a change in the variance
matrix based on the statistics D̂ (solid line), P̂ (dashed line), Q̂ (dotted line). The initial and
total sample size are m = 200 and m(T + 1) = 400, respectively, and the change occurs at
observation 300. The level is α = 0.05.
where {εj }j∈Z = {(εj,1 , εj,2 )> }j∈Z denotes a sequence of centered and independent two-dimensional
Gaussian distributed random variables and the involved matrices are defined as
0.2 0.1
0.3 0.1
0.1 0.05
A1 =
, A2 =
, A3 =
.
(5.3)
0.1 0.2
0.1 0.3
0.05 0.1
For the alternative, we proceed similarly as in Aue et al. (2009a) and define
(
I2
if j ≤ m + b m2 c
>
Cov(εj , εj ) = εj · εj =
,
I2 + δ · I2 if j > m + b m2 c
(5.4)
where I2 denotes the two dimensional identity matrix (the case δ = 0 corresponds to the null
hypothesis of no change). For the sake of brevity we will focus on the non-self-normalized
statistics D̂, P̂ and Q̂ here. In Figure 4 we display the empirical power for the three data
generation processes and the threshold function (T1). The results are similar to those presented
in Section 5.1. The test based on the statistic D̂ is more powerful than the tests based on P̂
and Q̂.
5.3
Changes in the correlation
We conclude this paper with a brief comparison of the three methods for the detection of
a change in the correlation, which has been considered in Wied and Galeano (2013). For the
definition of the data generating processes, we use the models (V1) - (V3) introduced in Section
5.2 but with a different process {εj }j∈Z = {(εj,1 , εj,2 )> }j∈Z . In this section {εj }j∈Z is a sequence
of independent two-dimensional Gaussian random variables such that
(
c1 if j ≤ m + b m2 c ,
Cor(εj,1 , εj,2 ) =
c2 if j > m + b m2 c
22
(V1)
(V2)
(V3)
(T1)
Figure 5: Empirical rejection probabilities of the sequential tests for a change in the correlation
based on the statistics D̂ (solid line), P̂ (dashed line) , Q̂ (dotted line). The initial and total
sample size are m = 500 and n = 1000, respectively, and the change occurs at observation 750.
The level is α = 0.05.
and Var(j,1 ) = Var(j,2 ) = 1. We use c1 = 0.3 for the correlation before the change and
consider different values of c2 . For estimation of the long-run variance matrix we use the
estimator proposed in Wied and Galeano (2013) (the explicit formula for the estimator is given
in the appendix of this paper and omitted here for the sake of brevity). Figure 5 now compares
the power of the non-self-normalized methods for the three models defined above. As in the
previous sections the sequential detection scheme based on D̂ yields substantially better results.
Acknowledgments. This work has been supported in part by the Collaborative Research
Center “Statistical modeling of nonlinear dynamic processes” (SFB 823, Teilprojekt A1, C1)
and the Research Training Group ’High-dimensional phenomena in probability - fluctuations
and discontinuity’ (RTG 2131) of the German Research Foundation (DFG). The authors would
like to thank Claudia Kirch and Dominik Wied for some helpful discussions on this subject.
References
Andrews, D. (1991). Heteroskedasticity and autocorrelation consistent covariance matrix estimation.
Econometrica, 59(3):817–58.
Aue, A., Hörmann, S., Horváth, L., Hušková, M., and Steinebach, J. G. (2012). Sequential testing for
the stability of high-frequency portfolio betas. Econometric Theory, 28(4):804837.
Aue, A., Hörmann, S., Horvth, L., and Reimherr, M. (2009a). Break detection in the covariance
structure of multivariate time series models. The Annals of Statistics, 37(6B):4046–4087.
Aue, A. and Horváth, L. (2013). Structural breaks in time series. Journal of Time Series Analysis,
34(1):1–16.
Aue, A., Horváth, L., and Reimherr, M. L. (2009b). Delay times of sequential procedures for multiple
time series regression models. Journal of Econometrics, 149(2):174 – 190.
Chu, C.-S. J., Stinchcombe, M., and White, H. (1996). Monitoring structural change. Econometrica,
64(5):1045–1065.
23
Csörgö, M. and Horváth, L. (1997). Limit theorems in change-point analysis. Wiley series in probability
and statistics. Wiley.
Davis, R. A., Huang, D., and Yao, Y.-C. (1995). Testing for a change in the parameter values and
order of an autoregressive model. The Annals of Statistics, 23(1):282–304.
Dette, H. and Wied, D. (2016). Detecting relevant changes in time series models. Journal of the Royal
Statistical Society: Series B (Statistical Methodology), 78(2):371–394.
Fremdt, S. (2014). Page’s sequential procedure for change-point detection in time series regression.
Statistics: A Journal of Theoretical and Applied Statistics, 48(1):1–28.
Horváth, L., Hušková, M., Kokoszka, P., and Steinebach, J. (2004). Monitoring changes in linear
models. Journal of Statistical Planning and Inference, 126(1):225 – 251.
Jandhyala, V., Fotopoulos, S., MacNeill, I., and Liu, P. (2013). Inference for single and multiple
change-points in time series. Journal of Time Series Analysis, 34(4):423–446.
Jirak, M. (2015). Uniform change point tests in high dimension. The Annals of Statistics, 43(6):2451–
2483.
Kifer, Y. (2017). Functional Erdös-Rényi law of large numbers for nonconventional sums under weak
dependence. Electronic Journal of Probability, 22(23):1–17.
Kirch, C. and Weber, S. (2017). Modified sequential change point procedures based on estimating
functions. Preprint, Otto-von-Guericke-Universitaet Magdeburg.
Lai, T. L. (1995). Sequential changepoint detection in quality control and dynamical systems. Journal
of the Royal Statistical Society. Series B (Methodological), 57(4):613–658.
Moustakides, G. V. (1986). Optimal stopping times for detecting changes in distributions. The Annals
of Statistics, 14(4):1379–1387.
Page, E. S. (1954). Continuous inspection schemes. Biometrika, 41(1/2):100–115.
Pape, K., Wied, D., and Galeano, P. (2016). Monitoring multivariate variance changes. Journal of
Empirical Finance, 39(Part A):54 – 68.
Shao, Q.-M. (1995). On a conjecture of Révész. Proceedings of the American Mathematical Society,
123(2):575–582.
Shao, X. (2015). Self-normalization for time series: A review of recent developments. Journal of the
American Statistical Association, 110(512):1797–1817.
Shao, X. and Zhang, X. (2010). Testing for change points in time series. Journal of the American
Statistical Association, 105(491):1228–1240.
Wied, D. and Galeano, P. (2013). Monitoring correlation change in a sequence of random variables.
Journal of Statistical Planning and Inference, 143(1):186 – 196.
Wu, W. B. and Rosenblatt, M. (2005). Nonlinear system theory: Another look at dependence. Proceedings of the National Academy of Sciences of the United States of America, 102(40):14150–14154.
Wu, W. B. and Zhou, Z. (2011). Gaussian approximations for non-stationary multiple time series.
Statistica Sinica, 21(3):1397–1413.
Zeileis, A. (2004). Econometric computing with hc and hac covariance matrix estimators. Journal of
Statistical Software, 11(10):1–17.
24
A
Technical details
Proof of Theorem 3.3. From (3.4), (3.6) we obtain the representation
bmsc
bmtc
m−3/2 Ũ(bmrc, bmsc, bmtc) = m−3/2 (bmtc − bmsc)(bmsc − bmrc) θ̂bmr+1c − θ̂bmsc+1
bmtc − bmsc
=
m3/2
+
bmsc
X
i=bmrc+1
bmsc − bmrc
IF(Xi , F, θ) −
m3/2
bmtc
X
IF(Xi , F, θ)
t=bmsc+1
(bmtc − bmsc)(bmsc − bmrc)
Rbmrc+1,bmsc − Rbmsc+1,bmtc .
3/2
m
By Assumption 3.1 we have
n bmtc − bmsc
m3/2
bmsc
bmtc
o
bmsc − bmrc X
IF(Xi , F, θ) −
IF(X
,
F,
θ)
i
m3/2
(r,s,t)∈∆3
i=bmrc+1
i=bmsc+1
n
o
D
1/2
=⇒ ΣF
t − s W (s) − W (r) − s − r W (t) − W (s)
(r,s,t)∈∆3
1/2
= ΣF B(s, t) + B(r, s) − B(r, t) (r,s,t)∈∆3 ,
X
where we use the definition of the process B in (3.11) and the fact
sup
(s,t)∈∆2
2
bmtc − bmsc
− (t − s) ≤
= o(1) .
m
m
Finally, Assumption 3.2 yields
(bmtc − bmsc)(bmsc − bmrc)
Rbmrc+1,bmsc − Rbmsc+1,bmtc = op (1) ,
3/2
m
uniformly with respect to (r, s, t) ∈ ∆3 so that Theorem 3.3 follows.
Proof of Corollary 3.5. Define
k−1
Dm (k) = m−3 max |U> (m + j, m + k)Σ−1
F U(m + j, m + k)| .
j=0
(A.1)
Using the fact, that the detection scheme Dm is piecewise constant and the monotonicity of
the threshold function we obtain the representation
Tm
max
k=0
Dm (k)
Dm (bmtc)
= sup
w(k/m) t∈[0,T ]
w(t)
=
m−3 U> (bmsc, bmtc)Σ−1
F U(bmsc, bmtc)
.
w(t − 1)
t∈[1,T +1] s∈[1,t]
sup
sup
By Remark 3.4 it and the continuous mapping theorem we have
Tm
max
k=1
Dm (k) D
B(s, t)> B(s, t)
=⇒ sup sup
,
w(k/m)
w(t − 1)
t∈[1,T +1] s∈[1,t]
25
where the process B is defined in (3.11). The result now follows from Remark 3.4, the fact, that
wα has a lower bound and that Σ̂m is a consistent estimate of the matrix ΣF , which implies
(observing the definition of D̂ in (A.1))
Tm
max
k=1
Dm (k)
T m D̂m (k)
Tm
− max
≤ max Dm (k) − D̂m (k)
k=1
k=1
w(k/m)
w(k/m)
−1
≤ kΣ̂−1
m − ΣF kop
sup
sup |m−3/2 U(bmsc, bmtc)|2
t∈[1,T +1] s∈[1,t]
= oP (1) .
−1
Here k · kop denotes the operator norm and we have used the estimate kΣ̂−1
m − ΣF kop = oP (1),
which is a consequence of the Continuous Mapping Theorem.
Proof of Theorem 3.8. By the definition of the statistic D̂ in (2.9), we obtain
Tm
max
k=0
U> (bmcc, m(T + 1))Σ̂−1
D̂m (k)
m U(bmcc, m(T + 1)
≥ m−3
,
wα (k/m)
wα (T )
(A.2)
where bmcc denotes the (unknown) location of the change. We can apply expansion (3.4) to
X1 , . . . , Xbmcc and Xbmcc+1 , . . . , XbmT c and obtain
bmcc
m(T
+
1)
−
bmcc
m(T +1)
bmcc
m−3/2 U(bmcc, m(T + 1)) =
−
θ̂
θ̂
1
bmcc+1
m3/2
bmcc
m(T + 1) − bmcc X
=
IF(Xi , F (1) , θF (1) )
m3/2
i=1
bmcc
− 3/2
m
m(T +1)
X
IF(Xi , F (2) , θF (2) )
i=bmcc+1
bmcc m(T + 1) − bmcc
(F (1) )
(F (2) )
θF (1) − θF (2) + R1,bmcc − Rbmcc+1,m(T +1) ,
+
m3/2
where θF (`) = θ(F (`) ) (` = 1, 2). Using Assumption 3.6 we obtain the joint convergence of
p
!
!
Pbmcc
(1)
(T
+
1
−
c)
Σ
W
(c)
(1)
1
m(T + 1) − bmcc
IF(X
,
F
,
θ
)
1
F
D
i
F (1)
Pm(T +1) i=1
=⇒ p
(2)
3/2
c
Σ
W
(T
+
1)
−
W
(c)
m
bmcc i=bmcc+1 IF(Xi , F , θF (2) )
2
2
F (2)
and
bmccm(T + 1) (F (1) )
P
(F (2) )
R
−
R
1,bmcc
bmcc+1,m(T +1) −→ 0.
3/2
m
P
As θF (1) 6= θF (2) this directly implies m−3/2 |U(bmcc, m(T + 1))| −→ ∞ , and the assertion
follows from (A.2) and the assumption that Σ̂m is a consistent estimate for ΣF (1) .
26
Proof of Theorem 3.10. Recalling the definition of Ũ and U in (3.6) and (3.7), respectively,
we obtain for the normalizing process V in (3.19) the representation
bmsc
−4
−4
m V(bmsc, bmtc) = m
X
2
2
j (bmsc − j)
θ̂1j
−
bmsc
θ̂j+1
θ̂1j
−
bmsc
θ̂j+1
>
j=1
bmtc
X
−4
+m
2
2
(bmtc − j) (j − bmsc)
j
θ̂bmsc+1
−
bmtc
θ̂j+1
j
θ̂bmsc+1
−
bmtc
θ̂j+1
>
j=bmsc+1
bmsc
−4
=m
X
U(j, bmsc)U> (j, bmsc)
j=1
bmtc
X
−4
+m
= m−3
+ m−3
Ũ(bmsc, j, bmtc)Ũ> (bmsc, j, bmtc)
j=bmsc+1
Z s
U(bmrc, bmsc)U> (bmrc, bmsc)dr
Z 0t
Ũ(bmsc, bmrc, bmtc)Ũ> (bmsc, bmrc, bmtc)dr .
s
By Theorem 3.3 we have
m−3/2 Ũm (bmrc, bmsc, bmtc)
D
(r,s,t)∈∆3
1/2
=⇒ ΣF
B(s, t) + B(r, s) − B(r, t)
(r,s,t)∈∆3
(A.3)
in the space `∞ (∆3 , Rp ), where the process B is defined in (3.11). Consequently, the Continuous
Mapping Theorem yields
1/2
n m−3/2 · U(bmsc, bmtc) o
n
o
ΣF B(s, t)
D
=⇒
,
1/2
1/2
m−4 · V(bmsc, bmtc)
(s,t)∈∆2
(s,t)∈∆2
ΣF N1 (s) + N2 (s, t) ΣF
(A.4)
where N1 , N2 are defined in (3.22). Now the assertion of Theorem 3.10 follows by a further
application of the Continuous Mapping Theorem.
Proof of Theorem 3.11. By definition of the self-normalized statistic D̂SN in (3.20), we obtain
Tm
max
k=0
SN
U> (bmcc, m(T + 1))V−1 (bmcc, m(T + 1))U(bmcc, m(T + 1)
D̂m
(k)
≥m·
,
wα (k/m)
wα (T )
(A.5)
where bmcc denotes the (unknown) location of the change. The discussion in the proof of
Theorem 3.8 shows
m−3/2 U(bmcc, m(T + 1)) −→ ∞ .
P
27
The proof will be completed by inspecting the random variable V−1 (bmcc, m(T + 1)) in the
lower bound in (A.5). Repeating again the arguments from the proof of Theorem 3.3 we can
rewrite
Z c
−4
−3
U(bmrc, bmsc)U> (bmrc, bmsc)dr
m · V(bmcc, m(T + 1)) = m
0
(A.6)
Z T +1
>
−3
Ũ(bmsc, bmrc, bmtc)Ũ (bmsc, bmrc, bmtc)dr .
+m
c
Using Assumption 3.6 and employing the arguments from the proof of Theorem 3.3 we obtain
weak convergence of
{U(bmrc, bmsc)}0≤r≤s≤c
{B (1) (r, s)}0≤r≤s≤c
D
=⇒
,
{Ũ(bmsc, bmrc, bmtc)}c≤s≤r≤t≤c
{B (2) (r, t) + B (2) (s, r) − B (2) (s, t)}c≤s≤r≤t≤T +1
where we use the extra definition
B (`) (s, t) = tW` (s) − sW` (t)
` = 1, 2
and W1 and W2 are defined in Assumption 3.6. By the Continuous Mapping Theorem and the
representation in (A.6) this implies
1/2
1/2
D
1/2
(1)
1/2
(2)
m−4 · V(bmcc, m(T + 1)) =⇒ ΣF (1) N1 (c) ΣF (2) + ΣF (2) N2 (c, T + 1) ΣF (2) ,
(1)
(2)
where the processes N1 and N2 are distributed like N1 and N2 in (3.22) but with respect to
the processes B (1) and B (2) , respectively.
Proof of Proposition 4.2. For the sake of readability, we will give the proof only for the case
d = 2. The arguments presented here can be easily extended to higher dimension. In view of
the representation in (4.13), we may also assume without loss of generality that µ = E[Xt ] = 0.
Part (a) of the proposition is a consequence of the discussion after Corollary 3.5 provided
that Assumptions 3.1 and 3.2 can be established. For this purpose we introduce the notation
2
2
Xt,1
− E[Xt,1
]
Zt := IF v (Xt , F, V ) = Xt,1 Xt,2 − E[Xt,1 Xt,2 ]
2
2
Xt,2
− E[Xt,2
]
and note that the time series {Zt }t∈Z can be represented as a physical system, that is
2
g12 (εt , . . . ) − E[X1,1
]
Zt = g1 (εt , . . . )g2 (εt , . . . ) − E[X1,1 X1,2 ] := G(εt , εt−1 , . . . ) ,
2
g22 (εt , . . . ) − E[X1,2
]
28
(A.7)
where gi denotes the i-th component of the function g in (4.1). The corresponding physical
Z
in (4.3) are given by
dependence coefficients δt,2
q
Z
0 2
0
0 2 2
2
0 2 2
2
)
Xt,2
) ) + (Xt,1 Xt,2 − Xt,1
− (Xt,2
) ) + (Xt,2
δt,2
=
− (Xt,1
(Xt,1
2
0
0
0 2
2
0 2
2
≤ C · kXt,1 − (Xt,1 ) k2 + kXt,2 − (Xt,2 ) k2 + kXt,1 Xt,2 − Xt,1 Xt,2 k2
n
o
2
0 2
2
0 2
0
0
≤ C · max kXt,1
− (Xt,1
) k2 , kXt,2
− (Xt,2
) k2 , kXt,1 Xt,2 − Xt,1
Xt,2
k2 ,
where C > 0 denotes a sufficiently large constant. Now Hölder’s inequality yields (for another
appropriate constant C)
2
0 2
0
0
kXt,1
− (Xt,1
) k2 ≤ kXt,1 + Xt,1
k4 kXt,1 − Xt,1
k4 ≤ C · δt,4 ,
0
0
0
0
0
kXt,1 Xt,2 − Xt,1 Xt,2 k2 ≤ Xt,1 Xt,2 − Xt,2 2 + Xt,2
Xt,1 − Xt,1
2
0
0
0
≤ Xt,1 4 Xt,2 − Xt,2
+ Xt,2
Xt,1 − Xt,1
≤ C · δt,4 .
4
4
4
P
Z
Combining these results gives ∞
t=m δt,2 ≤ C·Θm,4 < ∞ and Theorem 3 from Wu and Rosenblatt
(2005) implies the weak convergence
bmsc
bmsc
p
1 X
1 X
D
√
IF v (Xt , F, V ) = √
Zt =⇒ ΣF W (s) ,
m t=1
m t=1
in the space `∞ ([0, T + 1], R3 ) as m → ∞, where ΣF is the long-run variance matrix defined in
(3.3). Therfore Assumption 3.1 is satisfied.
To finish part (a) it remains to show that Assumption 3.2 holds. Due to (4.15) this is a
consequence of
j
X
1
sup √
Xt,` − E[Xt,` ] = oP (n1/4 )
j
−
i
+
1
1≤i<j≤n
t=i
(A.8)
for ` = 1, 2, 3. Since the arguments are exactly the same, we will only elaborate the case ` = 1.
For this purpose let
Si =
i
X
Xt,1 − E[Xt,1 ] ,
t=1
and note that the left-hand side of (A.8) can be rewritten as
n
o
1
1
−1
max max √ |Sj+k − Sj | = max max max √ (Sj+k − Sj ) , max max √ (Sj+k − Sj ) .
1≤j≤n 1≤k≤n−j
1≤j≤n 1≤k≤n−j
1≤j≤n 1≤k≤n−j
k
k
k
Thus it suffices to show that both terms inside the (outer) maximum are of order op (n1/4 ) . For
the sake of brevity, we will only prove that
max
1
max √ (Sj+k − Sj ) = op (n1/4 )
k
1≤j≤n 1≤k≤n−j
29
(A.9)
and the other term can be treated in the same way. Assertion (A.9) follows obviously from the
two estimates
Sj+k − Sj
√
= op (1) ,
1≤j≤n 1≤k≤(n−j)∧blog2 (n)c
kn1/4
Sj+k − Sj
√
max
max
= op (1) .
1≤j≤n blog2 (n)c≤k≤n−j
kn1/4
max
max
(A.10)
(A.11)
Since the function g is bounded, one directly obtains that there exists a constant C such that
|Xj,1 − E[Xj,1 ]| ≤ C. This gives
√
kC
Sj+k − Sj
√
| max
max 2
| ≤ max
max 2
= o(1)
1≤j≤n 1≤k≤(n−j)∧blog (n)c
1≤j≤n 1≤k≤(n−j)∧blog (n)c n1/4
kn1/4
and so (A.10) is shown. To establish (A.11) we will use Corollary 1 from Wu and Zhou (2011),
which implies, that (on a richer probability space) there exists a process {Ŝi }ni=1 and independent
Gaussian distributed random variables Y1 , . . . , Yn ∼ N (0, Γ(g)) such that
D
(Ŝ1 , . . . , Ŝn ) = (S1 , . . . , Sn )
and the centered Gaussian process
{Ĝi }ni=1 =
i
nX
t=1
Yt
on
i=1
satisfies
max |Ŝi − Ĝi | = OP n1/4 (log(n))3/2 .
1≤i≤n
Therefore we obtain
Ŝj+k − Ŝj
Ĝj+k − Ĝj
√
√
−
max
max
1≤j≤n blog2 (n)c≤k≤n−j
1≤j≤n blog2 (n)c≤k≤n−j
kn1/4
kn1/4
Ŝj+k − Ŝj − Ĝj+k + Ĝj
|Ŝj − Ĝj |
√
≤ max
≤
2
max
= op (1) .
max
1≤j≤n log2 (n)n1/4
1≤j≤n blog2 (n)c≤k≤n−j
kn1/4
max
max
Now Theorem 1 in Shao (1995) gives
lim max
max
n→∞ 1≤j≤n blog2 (n)c≤k≤n−j
Ĝj+k − Ĝj
√
=0
kn1/4
with probability 1, which completes the proof of Part (a).
For a proof of part (b) of Proposition 4.2 let F (1) , ΣF (1) and F (2) , ΣF (2) denote the distribution
30
function and corresponding long-run variances in equation (3.18) before and after the change
point, respectively. Note that h = A · g and consider the time series
(
g(εt , εt−1 , . . . )
if t < bmcc
X̃t =
,
A−1 · h(εt , εt−1 , . . . ) if t ≥ bmcc
which is strictly stationary with distribution function F (1) . Using similar arguments as in the
proof of part (a), one easily verifies that
n 1 bmsc
o
X
p
D
√
IF v (X̃t , F (1) , V )
=⇒ { ΣF (1) W (s)}s∈[0,T +1] .
m t=1
s∈[0,T +1]
(A.12)
Next, observe that there exists a matrix A(v) ∈ R3×3 , such that for all symmetric matrices
M ∈ R2×2 , the following identity holds
vech(A · M · A> ) = A(v) · vech(M ) .
Further, using (4.10) one observes
>
A · IF(X̃t , F (1) , V ) · A> = AX̃t X̃t A> − A · V (F (1) ) · A> = IF(Xt , F (2) , V )
whenever t ≥ bmcc, which yields
A(v) IF v (X̃t , F (1) , V ) = IF v (Xt , F (2) , V ) for t ≥ bmcc .
(A.13)
Similar arguments give
A(v) ΣF (1) (A(v) )> = ΣF (2) .
(A.14)
Now consider the mapping
∞
` ([0, T + 1], R3 ) → `∞ ([0, c], R3 ) × `∞ ([c, T + 1], R3 )
!
{f (s)}s∈[0,c]
ΦA :
,
{f (s)}s∈[0,T +1] 7→ A(v) (f (s) − f (c))
s∈[c,T +1]
then the Continuous Mapping, (A.12) and (A.13) yield
!
!
1 Pbmsc
p
(1)
√
IF(X
,
F
,
V
)
Σ
W
(s)
(1)
t
F
D
t=1
m
s∈[0,c]
s∈[0,c]
(v) p
1 P
=⇒
bmsc
(2)
√
A
Σ
W
(s)
−
W
(c)
(1)
IF(X
,
F
,
V
)
t
F
t=bmcc+1
s∈[c,T +1]
m
s∈[c,T +1]
!
p
Σ
W
(s)
(1)
F
D
s∈[0,c]
= p
,
ΣF (2) W (s) − W (c) s∈[c,T +1]
31
where the identity in distribution follows from the fact that both components are independent
and the identity
p
p
>
A(v) ΣF (1) A(v) ΣF (1) = A(v) ΣF (1) (A(v) )> = ΣF (2) .
For the verification of Assumption 3.6 it suffices to show that both, the phase before and after
the change point satisfy Assumption 3.2. This can be done using similar arguments as in the
proof of part (a) of Proposition 4.2 and the details are omitted.
32
| 10math.ST
|
Single-trial P300 Classification using PCA with LDA, QDA and
Neural Networks
Nand Sharma
∗
arXiv:1712.01977v1 [cs.NE] 6 Dec 2017
Department of Mathematics, Colorado State University, CO, USA
Abstract
The P300 event-related potential (ERP), evoked in scalp-recorded electroencephalography
(EEG) by external stimuli, has proven to be a reliable response for controlling a BCI. The
P300 component of an event related potential is thus widely used in brain-computer interfaces
to translate the subjects’ intent by mere thoughts into commands to control artificial devices.
The main challenge in the classification of P300 trials in electroencephalographic (EEG) data
is the low signal-to-noise ratio (SNR) of the P300 response. To overcome the low SNR of
individual trials, it is common practice to average together many consecutive trials, which
effectively diminishes the random noise. Unfortunately, when more repeated trials are required
for applications such as the P300 speller, the communication rate is greatly reduced. This has
resulted in a need for better methods to improve single-trial classification accuracy of P300
response. In this work, we use Principal Component Analysis (PCA) as a preprocessing method
and use Linear Discriminant Analysis (LDA)and neural networks for classification. The results
show that a combination of PCA with these methods provided as high as 13% accuracy gain for
single-trial classification while using only 3 to 4 principal components.
1
Introduction
Various neurological diseases can disrupt the neuromuscular channels through which the brain communicates with the external world. In certain cases like hemorrhage in the anterior brain stem or
degenerative neuromuscular diseases like amyotrophic lateral scleriosis (ALS), the patients suffer
from a total motor paralysis [5]. This results in a condition known as locked-in syndrome, wherein
the patient is awake and fully aware but cannot communicate with the outside world due to complete paralysis. For such "locked-in" patients, there is a need for an assistive technology that needs
no muscular activity whatsoever.
A brain-computer interface (BCI) is a device that uses brain signals to provide a direct, nonmuscular communication channel between brain and the outside world [32, 31, 29]. The idea underlying BCIs is to measure electric, magnetic, or other physical manifestations of the brain activity
and to translate these into commands for a computer or other devices [21, 15].
For patients with locked-in syndrome,the P300 event-related potential (ERP), evoked in scalprecorded electroencephalography (EEG) by external stimuli, has proven to be a reliable response
for controlling a BCI [9]. In this study we present comparison of some classification methods to
classify an EEG signal based on the presence of P300 component.
∗
[email protected]
1
1.1
BCI and P300
Types of BCIs can be broadly classified into two categories—those that use an external stimulus,
and those that don’t [21]. In the first method, the external stimuli cause changes in neurophysiologic
signals called event-related potentials (ERPs) [15, 22] which are used to identify a user’s response
to the stimuli presented. In the second method, users generate certain detectable patterns of
neurophysiologic signals by concentrating on a specific mental task. For example, imagination of
hand movement can be used to modify activity in the motor cortex [15].
For recording the activity of the brain, the electroencephalogram (EEG) is the method of choice
for a BCI due to their fast responsivity and covariation with cognitive processes [5]. Although
invasive methods that use electrocorticography (ECoG) signals using implanted electrodes are more
accurate, the non-invasive methods are more attractive because of their ease of use by patients; the
non-invasive methods have also been shown to be comparable to implanted electrodes [30] when
used with appropriate machine learning algorithms.
The EEG non-invasive recordings are done from a set of electrodes placed directly on the scalp
using the International 10-20 system (Jasper, 1958) [25]. In this work, the data is obtained through
non-invasive electroencephalographic (EEG) recordings. Moreover, the experiments are run on
EEG data using only a subset of 8 electrodes that have already been found to be meaningful
for P300 classification [30]. A smaller number of electrodes is also better for practical reasons
of lower cost and higher usability for target patients. These 8 electrodes used for this work are
’F3’,’F4’,’C3’,’C4’,’P3’,’P4’,’O1’,’O2’.
A P300 ERP is characterized by a positive peak about 300ms after the stimulus onset [5, 22]
(Figure 1.1). It is elicited when subjects encounter a rarely occurring, but expected, stimulus
among the presented stimuli. If subjects are assigned the task of assigning a category to each of the
stimuli in a series of stimuli of two types, and if one of the two types occurs rarely, a P300 ERP is
seen in the EEG [20] for the rare stimuli. This experimental paradigm based on extensive research
has been called ’oddball’ paradigm [20], the rarely occurring stimuli being the ’oddballs’.
Figure 1.1: P300 ERP. The image is taken from [8].
Farwell and Donchin utilized this characteristic of the P300 to design a BCI in 1988 [20] that
is called a ’P300 Speller’. This first BCI of its kind lets a user type one letter at a time using the
EEG signals captured through an electrode cap, and needs no muscular movement. There has been
a lot of research following their work on improving their setup and many variations of the setup
have emerged. As definitions for P300 speller paradigm vary in literature, it is made clear that for
2
this work, each of these windows of EEG will be referred to as a ’subtrial’ and a set of row (column)
flashings (six in number which contain one target row(column)) will be called a trial [20].
One of the main challenges in P300 classification is the low signal-to-noise ratio (SNR). As the
EEG recorded from scalp contains a lot of noise from ongoing electrical activity in the brain, a P300
is hard to separate from the resulting noisy signal. The problem of low SNR is usually overcome
by averaging together many subsequent trials which cancels out most of the noise, and makes P300
detection possible. But this approach of averaging comes at the cost of reduced communication rate.
Farwell and Donchin reported in their pioneering work that it needed averaging of between 20 and
40 trials to achieve an accuracy of over 80% [20], with a communication rate of about 12 bits per
min or about 2.3 characters per minute. Although their pioneering work did establish the feasibility
of a P300 based speller, the communication rate was painfully slow for a practical use. This lead
to much work in the last few decades which focused on improving the accuracy and communication
speed of a P300-based BCI like P300 speller. So, one of the focus areas of this research has been the
development of algorithms that can reduce the number of trials required to achieve a reliable P300based BCI. But as it is amply demonstrated that averaging of trials stabilizes the P300 amplitude
by removing noise, the challenge is also seen as improving single trial accuracy.
1.2
Related Work
The last few decades have seen a lot of research on P300 classification, with the eventual aim of
achieving single-trial based reliable P300 BCI [21]. But a literature survey of this research does
not show any single P300 classification method to be the state of the art. Krusienski, et al., [18]
report the results of a comparison of different classifier algorithms, which shows that stepwise linear
discriminant analysis (SWLDA) and support vector machines (SVMs) perform well compared to the
other classifiers. In [19] by using SWLDA as the classification method, Krusienski, et al., achieved at
least 60% accuracy for all participants. Three of the five participants performed above 90% accuracy
with averaging about 15 trials. Sellers and Donchin [10] achieved comparable average results for
healthy and ALS patients using SWLDA. Serby, et al., [13] used matched-filtering with independent
component analysis (ICA) to achieve a communication rate of 5.45 symbols/min with an accuracy of
92.1% which averaged roughly 15 trials. When the detection was made in real-time by online testing
with the same six subjects, the average communication rate achieved was 4.5 symbols/min with an
accuracy of 79.5% which averaged roughly 18 trials. On the same lines, a survey of submissions
to BCI Competition II and BCI Competition III shows that several very different approaches like
SVM, ICA, LDA, peak-picking methods were able to achieve 100% accuracy using between 4 and
15 averaged trials [7, 16, 33] on the BCI Competition II data.
This short list shows that there are many different approaches that all work well, and also that
one particular method is not clearly better than others. Then there are other challenges to P300
classification, namely subject-dependence of EEG and even a session-to-session variation in EEG
responses of the same subject. A recent review of the BCI field by Mak, et al., [23] concludes that
a lot more work is still needed to create a truly reliable P300 speller.
In this work, we use Principal Component Analysis (PCA) as a preprocessing method and use
Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA) and neural networks
for classification. PCA has been shown to work well for P300 classification [8, 12]. In [8], the author
compared various blind source separation methods as preprocessing methods for P300 classification,
and it was seen that PCA generally worked better than the other methods like Independent Component Analysis (ICA) and Maximum Noise Fraction (MNF) for P300 classification. This work builds
upon the work done by Cashero [8] and tests if classification methods other that Support Vector
Machines (SVM) will work well in conjunction with PCA for P300 classification.
3
Among the classification methods, the choice of LDA is also driven by success of this method
and its variants like Stepwise LDA (SWLDA) for P300 classification. Although SWLDA has been
shown to work very well for higher accuracy, it is an expensive method with high processing time
and is therefore not considered suitable for an online P300 speller system [21]. Considering that,
we have preferred to consider LDA which is lightweight and provides good performance generally
for P300 classification. QDA is chosen to compare LDA with, and to see whether a linear or a
nonlinear method works better for P300 classification. Our choice of Neural Networks (NN) as a
method of study for P300 classification is primarily driven by our hypothesis that they should be a
good choice for P300 classification. There have been studies employing Neural Networks for P300
classification [12, 26], and many have shown promise. There is very little work done using PCA
and NN as a combination for P300 classification. As NN provide a lot of flexibility in terms of how
the ’derived’ data is created based on flexibility in the number of hidden units, it could be a useful
method for noisy data like EEG, especially if PCA provides a good set of source components. SVM
is another method which has been shown to work well for P300 classification. We don’t include that
in this study due to the fact that a study using PCA with SVM has already been covered in [8].
To keep our focus on the effect of PCA on these classification methods, we have tested only single
trial accuracies.
The paper is laid out as follows. Section 2 develops the mathematical background for all the
algorithms that are used in the experiments for this work. The classification methods, LDA, QDA,
Neural Networks, and the optimization algorithm used for NN and PCA are developed in this
section. Section 3 describes the datasets used and details the steps used in the experiments. Section
4 explains the experiments as they were performed and what was learned along the way, and the
results of the experiments. Section 5 concludes by providing a summary of the findings and identifies
avenues for future work.
2
2.1
Classification and Preprocessing Methods
LDA and QDA
LDA and QDA [14, 6] belong to a class of classification methods that model discriminant functions
for each class and then classify a given data sample to the class with the largest value for its
discriminant function. LDA and QDA model the posterior probabilities for the purpose of defining
these discriminant functions. Representing a data sample by variable X and the class label by
variable C, X ∈ Rd where d is the dimension of data sample X or equivalently represents the number
of features or predictors in each sample X, and C ∈ {1, 2, ..., K} where there are K classes. We
thus need the class posteriors p(C | X) for a given X and C. Suppose fk (x) is the class-conditional
P
density of X in class C = k, and let πk be the prior probability of class k, with K
k=1 πk = 1.
Applying the Bayes theorem gives us
fk (x)πk
p(C = k|X = x) = PK
l=1 fl (x)πl
(1)
While many different models can be used to model the class-conditional density, LDA and QDA
use multivariate Gaussians to model each class density so that
fk (x) =
1
(2π)d/2 |Σ
k
|1/2
T Σ−1 (x−µ )
k
k
e−1/2(x−µk )
(2)
where Σk and µk are Covariance matrix and the mean of the Gaussian
If we assume
PK Ndistribution.
k
that the classes have a common covariance matrix given by Σ = k=1 N Σk , where Nk is the class
4
k samples and N is the total number of samples of all classes, then we get LDA. In comparing two
classes k and l, it is sufficient to look at the log-ratio
ln
p(C = k | X = x)
fk (x)
πk
= ln
+ ln
p(C = l | X = x)
fl (x)
πl
(3)
Plugging in the fk (x) and fl (x) as defined in (2), we get
ln
p(C = k | X = x)
πk
1
= ln
− (µk − µl )T Σ−1 (µk − µl ) + xT Σ−1 (µk − µl )
p(C = l | X = x)
πl
2
(4)
which is an equation linear in x. So, equation (4) gives us a decision rule for LDA—if the log ratio
is positive, the sample x is classified as belonging to class k and to class l if it is negative. The
value being zero implies the decision boundary which is linear in x. It’s obvious that the linear
discriminant functions
1
δk (x) = xT Σ−1 µk − µTk Σ−1 µk + ln πk
(5)
2
are an equivalent description of the decision rule for LDA. So, the class of a new sample x is simply
C(x) = arg maxk δk (x)
Now if we drop the assumption of a common covariance matrix for all classes, then the convenient
cancellation of terms leading to (4) doesn’t happen and we get the quadratic discriminant functions
(QDA),
1
1
δk (x) = − ln |Σk | − (x − µk )T Σ−1
(6)
k (x − µk ) + ln πk
2
2
The decision boundary in this case between each pair of classes k and l is described by a quadratic
equation given by δk (x) = δl (x). Again, the class of a new sample x is obtained as
C(x) = arg maxk δk (x)
The LDA and QDA are also called Generative models, as they make the assumption of Gaussian
distribution of the data, and base the classification of samples on that assumption. The parameters
of the Gaussian distributions are estimated using the training data. Class priors are calculated
based on the number of samples of each class present in the training data, as
πk =
Nk
,
N
where Nk is the class k samples and N is the total number of samples of all classes. The class means
are estimated as the means of the data of a particular class in the training data, so
X
µk =
xi /Nk
ci =k
where ci is the class of sample xi . Similarly Σk is the covariance matrix for each class based on
the data of that particular class in the training data.
5
2.2
Neural Networks
Neural Networks (NN) [14, 6] are primarily employed in machine learning as nonlinear regression and
classification method. While there are many variations and flavors of NN like Recurrent NN, multilayer perceptron, etc., [14, 6, 34], we use a single hidden layer NN for this work. This basic neural
net, sometimes called the single hidden layer back-propagation network, or two layer perceptron is
a two-stage regression or classification model. It consists of an input layer, a hidden layer and an
output layer.
In the case of K-class classification, there are K output units, and each of the K output units
models the probability of class k so that
Yk ∈ [0, 1] ∀ Yk , k = 1, ..., K
Derived features Zm are created from linear combinations of the inputs, followed by a nonlinear
activation function. The output Yk is modeled as a function of linear combinations of the Zm ,
T
Zm = σ(α0m + αm
X), m = 1, ..., M,
βkT Z,
(7)
k = 1, ..., K,
(8)
fk (X) = gk (Y ), k = 1, ..., K
(9)
Yk = β0k +
where Z = (Z1 , Z2 , ..., ZM ), and Y = (Y1 , Y2 , ..., YK ). σ(v) is the activation function and is chosen
to be a sigmoid defined as
σ(v) =
1
1 + e−v
The output function gk (Y ) allows a final transformation of the vector of outputs Y. We choose
the softmax function for this :
e Yk
(10)
gk (Y ) = PK
Yl
l=1 e
which results in a multilogit model, and produces positive estimates that add to one. Treating these
outcomes as probabilities for the corresponding class, we use negative log-likelihood as the objective
function to minimize. This negative log-likelihood objective function is defined as
LL(θ) = −
N X
K
X
Tik logfk (xi )
(11)
i=1 k=1
where θ denotes the complete set of weights of the network, which consist of {α0m , αm , m =
1, ..., M } and {β0k , βk , k = 1, ..., K}, and Tik is the k th component of the target indicator variable
Ti for ith training sample xi . Finally, the corresponding classifier is defined as
C(x) = arg maxk fk (x)
With the sigmoid activation function and log-likelihood error function, the neural network model
works as a linear logistic regression model in the hidden units, and all the parameters are estimated
by maximum likelihood. But by using the nonlinear transformation σ, it becomes a non-linear
model of inputs X. Another interesting aspect of this model is that the number of hidden units
can be varied to adjust the non-linearity of the model. If there are no hidden units, the model
becomes a simple linear logistic regression model over input data. In this work we use both linear
and non-linear versions of the neural network model and call them LR and NLR respectively.
6
The error function LL(θ) can be minimized by variety of approaches. One of the standard
approaches is the gradient descent, called back-propagation in the neural network setting. But a
conjugate gradient method called Scaled Conjugate Gradient [24] has been more successful and
faster for this optimization problem and we use that in this work and we use SCG for this work.
2.3
PCA
Principal Component Analysis, or PCA, [6, 28, 17] (also known as Karhunen-Loeve transform)
is a technique that is widely used in pattern recognition and machine learning for dimensionality
reduction and feature extraction. There are two commonly used derivations of PCA—one that
maximizes variance of data, and the one that minimizes the projection error. We develop the
maximum variance formulation.
Given a set of data samples {xn , n = 1, ..., N }, xn ∈ RD , the goal of PCA is to project the data
onto a space with dimensionality M ≤ D, while maximizing the variance of the projected data. If
v1 is the first direction of projection, the variance of the data projected on v1 is given by
N
1 X T
{v1 xn − v1T x̄}2 = v1T Sv1
N
(12)
n=1
where x̄ =
1
N
PN
n=1 (xn )
is the sample set mean, and S is the data covariance matrix defined by
S=
N
1 X
(xn − x̄)(xn − x̄)T = v1T Sv1
N
(13)
n=1
So, the optimization problem becomes that of maximizing 12, or equivalently
max v1T X T Xv1 , v1T v1 = 1
v1
(14)
which is a constrained optimization problem. Using a Lagrange multiplier λ, we get
max v1T X T Xv1 + λ(1 − v1T v1 )
v1 ,λ
(15)
By setting the derivative of 15 with respect to v1 equal to zero, we get
2Xv1 − 2λv1 = 0
(16)
X T Xv1 = λv1
(17)
which means that v1 must be an eigenvector of X T X with eigenvalue λ, which also turns out to be
a measure of the variance. So, v1 turns out to be a direction of projection that results in maximum
variance in the projected data, and is called the first ’Principal Component’. The subsequent
directions can be found inductively as follows. Given that we have up to vk Principal Components,
the next Principal Component v can be found by solving the following constrained optimization
problem :
max
v,v⊥v1 ,..,vk
v T X T Xv, v T v = 1
7
(18)
And this problem can be solved by creating the Lagrangian. Solving Equation 18 for all eigenvectors
is equivalent to computing the singular value decomposition (SVD) of X [28, 17, 11]:
X = U ΣV T
(19)
where U ∈ RN ×N , V ∈ RD×D and Σ ∈ RN ×D . In this decomposition, the columns of V, the right
singular vectors, are the eigenvectors of X T X [28, 17, 11] which provide us the required ’Principal
Components’. Additionally, these column vectors are ordered by the variance they produce when
data is projected onto them so that the first column of V is the first Principal Component, the second
one the second and so on. In this work, we thus use SVD to derive the Principal Components for
our experiments.
3
Data Acquisition and Representation
This Section describes the datasets used in this study, as also the methods and EEG recording
equipment used. As data representation is an important part of any signal processing method, we
also describe the data representation used in this study. A novel way of using channel-subtrials is
also defined and explained.
3.1
Datasets
There are four subjects in the study. The EEG data for subjects 1 and 2 was recorded by BCI laboratory at Computer Science department at Colorado State University [2]. The data were recorded
using the g.Tec g.GAMMAsys system [4] with a 8-electrode cap with electrodes located at Fz, Cz,
Pz, Oz, P3, P4, O1, O2. This subset of electrodes has been found to be meaningful for P300 classification [30, 27]. Such small subsets of electrodes are also easier to use for a practical and easy-to-use
BCI. Subject 2 was an able-bodied participant in the study, and subject 1 was a subject with C4
complete Spinal cord injury. C4 is a level of Cervical (neck) injury that results in significant loss of
function at the biceps and shoulders. While the data for subject 2 was collected in the laboratory,
the data for subject 1 was recorded at home.
Three sessions of data collection were performed with both subjects 1 and 2. The subjects
count one of three target letters (’b’, ’d’, ’p’) during a session as various other non-target letters
are randomly flashed on a screen. This data collection is done as per ’odd-ball’ paradigm and the
probability of occurrence for the target letter in each session was 0.25. The data were sampled
at 256 Hz, with an inter-stimulus-interval (ISI) of 1 second. One session consisted in recording 20
target and 60 non-target subtrials of 1000ms each of EEG data at 8 electrodes.
Data for subjects 3 and 4 is taken from BCI competition III (dataset II) [1]. These experiments
being based on the P300 speller as proposed by Farwell and Donchin [20], the subjects were presented
with a 6 by 6 matrix of characters. The subject’s task was to focus attention on characters in a
word that was prescribed by the investigator (i.e., one character at a time). All rows and columns
are successively and randomly intensified at a rate of 5.7Hz. The objective in this contest was to
predict the correct character in each of the provided character selection ’epochs’ that consisted of 15
sequences—each sequence consisting of 12 flashings—6 rows and 6 columns. The data was collected
and bandpass filtered from 0.1-60Hz and digitized at 240Hz. After intensification of a row/column,
the matrix was blank for 75ms. Row/column intensifications were block randomized in blocks of
12. The sets of 12 intensifications were repeated 15 times for each character epoch (i.e., any specific
row/column was intensified 15 times and thus there were 180 total intensifications for each character
epoch). Each character epoch was followed by a 2.5 s period, and during this time the matrix was
8
blank. While data for subjects 3 and 4 is recorded at 64 electrode locations, this work uses data
only from 8 locations which is the same set of electrodes used for subjects 1 and 2—Fz, Cz, Pz, Oz,
P3, P4, O1, O2.
3.2
Data Processing and Representation
The original data for all the four subjects comes as a continuous EEG for an entire recording. All
the data was then bandpass-filtered from 0.23 Hz to 30 Hz. The data were then normalized for zero
mean and unit variance. Figures 3.1 and 3.2 show a 4 second window of data before and after the
bandpass filtering. The bandpass filtering was done using Butterworth bandpass filter [3]. This data
is then sliced to separate the target and non-target subtrials for each channel. Each Dataset was
thus reshaped into a matrix with each row representing a channel-subtrial- 256 datapoints as a time
series for subjects 1 (sub1) and 2 (sub2), and 240 for subjects 3(sub3) and 4(sub4). The channelsubtrials used for classification therefore consist of one-second long windows after each stimulus
onset that are extracted from the continuous signal in each data segment.
Figure 3.1: 4s Original Data sub1.
Figure 3.2: 4s Data sub1 after bandpassing.
Our approach for data representation involves considering a time series from each of the eight
channels as a separate subtrial for the purpose of initial classification. What this means is that a
subtrial that in the usual sense consists of eight different time series, one for each channel, is split into
eight different subtrials in our data matrix. This is done with the view that P300 response is present
in each of the eight channels considered although it varies in degree and has some phase difference,
and that our preprocessing method should be able to get common sources to these responses at
different channels. We call these channel subtrials. Once the training of the algorithm is done, and
results for the test set are collected for each channel sub trial, these results are aggregated to arrive
at a classification for the ’overall’ subtrial. This aggregation is done using a voting method.
The P300 data is generally unbalanced in positive and negative examples, as every trial contains
many more negative examples than the positive ones in accordance with the ’odd-ball’ paradigm.
This usually tends to bias a classifier in favor of the negative examples. So, an equal number of
subtrials of target and non-target EEG were used for these experiments (20 for sub1 and sub2, and
30 for sub3 and sub4).
9
4
Experiments and Results
Experiments were designed to test and compare the various approaches selected for this work with
and without PCA. So, the following workflow was followed.
1. Classification performance of the four classifiers—LDA, QDA, LR and NLR—was recorded
on the raw data, i.e., the data that has not been transformed with PCA. In this, no feature
selection is done, so that all the original features are used.
2. Then classification performance is tested after performing PCA on the data. There are two
procedures in this set of experiments—one with forward selection of principal components
across all the components, and one where forward selection is done on selected number of top
components as per the magnitude of the singular values.
4.1
Without PCA
We start experiments by testing our four methods on the raw data, i.e., the data without PCA
transformation for the four subjects. This would serve as a benchmark for the experiments to follow
using the PCA. For these experiments and that follow, the following common approach was followed.
1. The datasets were randomly partitioned into training and test sets in the ratio of 80:20. As we
consider channel-subtrials as separate subtrials during training and classification before voting,
the data was partitioned to ensure we place complete set of 8 channel-subtrials belonging to
a subtrial together either in the training or test sets.
2. The algorithms are trained on the data comprising the channel-subtrials, and tested on the
test channel-subtrials.
3. The accuracies on the channel-subtrials from the 8 channels are aggregated together using the
voting method to decide on the final class of the actual subtrial.
4. The above process is repeated 20 times and the accuracies over the 20 runs are averaged to
obtain the final accuracies.
For NLR, the training set was partitioned to create a validation set. The accuracies on the
validation set were used to choose the number of hidden units. In the pilot studies on all four
subjects, a range of number of hidden units from 2 to the number of features was tested on the
validation set. It was observed that the validation accuracies generally were the best at number
of hidden units equal to the number of features in the dataset. So, throughout the experiments,
number of hidden units for a dataset have been taken as number of features used for classification.
The accuracies obtained on the raw data for the four subjects is shown in Table 1. First thing
to note is that QDA did not work for any of the subjects, the problem being that the covariance
matrices became singular, that gave a runtime error. So, the accuracies for QDA have been left
blank for raw data. The problem of sample covariance matrices being singular occurs when sample
size is less than the number of features, which happens to be the case when using QDA on raw data.
The same problem was not usually encountered in the case of LDA as the averaging of the sample
covariance matrices of the two classes resulted in the average covariance matrix being non-singular.
In the subsequent experiments, a small subset of features was always used, which ensured that this
problem is not encountered.
Analyzing the results, we observe as generally expected in P300 classification that no single
method is the best across all 4 subjects. LR works the best for subjects sub1 and sub4, but NLR
10
also works the best for two subjects sub3 and sub4. It’s only for sub2 that neither of the NN
versions, linear and nonlinear, work well. And it is LDA that works the best for sub2. So, across
subjects, we can not generalize if linear or nonlinear methods are a clear winner. But for sub1, the
performance of the linear version of NN, the LR, is far superior than the other methods. The best
accuracies obtained for all the subjects are near 50% which is the expected accuracy for a random
classifier for single trials [13, 8] except sub1 which gets a very high accuracy for single trials.
Table 1: Accuracies with all four methods (without PCA)
Subjects −→ sub1
sub2
sub3
sub4
Algorithms ↓
LDA
QDA
LR
NLR
4.2
47.33
65.62
58.75
51.25
48.25
42.12
49.58
47.08
50.83
44.58
51.25
51.25
With PCA
The next step in our experiments was to use PCA for feature extraction. We start our experiments
with PCA by getting the Principal Components (PCs) of the training data using SVD. Then we
plot the projection of training data onto first 20 components for each of the four datasets. Looking
at these projections, we choose the top 3 PCs through visual inspection—for sub1 these turned out
to PCs 2,3, and 4. Projection of sub1 data onto a 3-dimensional subspace of these 3 PCs is shown
in Figure 4.1.
After a similar analysis for subjects sub2, sub3 and sub4, we choose the top 3 components for
each of them. Then using the chosen three components, we run our classification experiments. The
results of these experiments are shown in Table 2.
Table 2: Accuracies for PCA-transformed data using only 3 visually chosen PCs
Subjects −→ sub1 sub2 sub3
sub4
Algorithms ↓
PCA+LDA
PCA+QDA
PCA+LR
PCA+NLR
55.0
52.5
41.6
41.6
55.0
46.0
40.8
56.6
49.16
49.16
45.0
49.16
55.00
51.6
50.83
53.33
While this method improved the best accuracy for sub2 and sub4, for sub1 and sub3 the accuracies decreased with almost all the methods. Although the accuracy gains for sub2 and sub4
are impressive, this method is neither scientific nor desirable for a BCI because of its reliance on
human intervention and visual judgment. Moreover, there is a limit to the number of PCs that we
11
Figure 4.1: sub1 data projected on 3-D subspace.
can inspect visually. But, it does show that if we are able to pick the right components, we could
get accuracy gains for some subjects.
To make the process of choosing the PCs thorough and to automate the process, we use the
method of ’forward selection’ (FS) of the PCs in each of the four algorithms. The forward selection
is done using 3-fold cross-validation to choose the PCs that give the best accuracy for the validation
set. A new PC is thus added to the set of PCs at each iteration. Finally, the minimal set that gives
the best accuracy on the validation set is chosen. As this method is very expensive, we pick and
test validation accuracies till only the top 50 components are chosen. This also ensures that data
sample size being less than the dimensionality doesn’t give singular covariance matrices in the case
of QDA. This method thus iterates through all the PCs at each iteration and picks the best PC at
each iteration. The results of this method are shown in table 3.
The results using forward selection of Principal components are not found to be as encouraging
as expected. The challenge lies in selecting the number of top components based on the validation
accuracies. The method chooses the number of components that give the best validation accuracy,
but this doesn’t necessarily work for the test data. The reason for that is the extreme noise in the
12
Table 3: Accuracies for PCA-transformed data with number of PCs chosen using forward selection(Number of Principal Components in braces)
Subjects −→
sub1
sub2
sub3
sub4
Algorithms ↓
PCA+LDA
PCA+QDA
PCA+LR
PCA+NLR
54.37(11)
49.37(18)
42.50(6)
55.00(4)
53.12(14)
55.12(18)
50.46(5)
48.03(5)
49.58(16)
50.83(15)
47.08(3)
51.25(20)
48.75(14)
47.91(12)
50.00(5)
50.83(5)
data due to which the validation set accuracies don’t generalize well to test data. The FS would
pick such components that might be modeling the noise of the validation set which obviously will
throw the classifier off on the test data. Then there is also over-fitting of training data in the case
of non-linear methods as can be seen for NLR. In addition to these problems, there is the matter
of complexity and runtime. The FS is a very expensive method as each iteration of choosing a new
component has to iterate through all the remaining set of components.
Considering the above pitfalls of the FS method, we need to try a different method for selecting
a good subset of PCs. Looking back at the results of using top 3 PCs chosen visually from top 20
components, we try modifying the forward selection algorithm to select the components only from
a set of an empirically chosen number, say n, of top PCs based on the magnitude of their singular
values. This amounts to just considering the top n PCs as they are ordered already by magnitude
of their singular values. As visually chosen 3 components had shown good results for 2 out of 4
subjects, we try n = {5, 10, 15, 20}. After looking at the results thus obtained, it was observed that
using n = 5 works the best. The results of using FS on a restricted set of top 5 PCs is shown in
table 4.
Table 4: Accuracies for PCA-transformed data choosing the best among only top 5 PCs (Number
of Principal Components in braces)
Subjects −→
sub1
sub2
sub3
sub4
Algorithms ↓
PCA+LDA
PCA+QDA
PCA+LR
PCA+NLR
55.0(2)
55.0(2)
41.66(3)
60.0(3)
48.0(3)
52.7(3)
57.5(2)
40.0(3)
45.0(2)
40.83(2)
56.9(4)
63.05(4)
47.5(3)
53.3(1)
46.66(2)
50.83(4)
The results obtained using FS on the restricted set of top 5 PCs show an improvement in accuracy
for all the four subjects when compared with regular FS. The method gives overall best accuracies
for sub2 and sub3. For sub4 also, the accuracy is better than all methods except visual selection of
PCs which can be discounted as that’s not a practical method for an online BCI. So, discounting
the visual selection method, we get best accuracies for three subjects out of four using this method
of restricted FS. In addition, this method comes with greatly reduced data dimensionality. While
FS selected as high as 18 PCs, this method needed no more than 4 components to achieve much
13
better results.
5
5.1
Conclusions
Summary of Results
As the experiments have shown and as is supported by the literature, the success of a method for
P300 classification depends a lot on the subjects. While LR gave exceedingly good results for sub1
without PCA, sub3 had the best accuracy of all the methods with NLR + PCA (with restricted
FS). NLR + PCA (with restricted FS) also gave the second best accuracy for sub1 at 60%.
For sub2, on the other hand, it was the linear version of NN—the LR (with restricted FS)—that
gave the best accuracy. For sub4 alone, the visually chosen 3 PCs gave the best accuracy. But
overall it’s clear that PCA did help in improving the accuracy of classification of single trials in
all subjects but sub1. The best results except in the case of sub1 all came with a much reduced
dimensionality of data—the dimensionality was a maximum of 4 for all these subjects. So, PCA
feature selection not only increased the classification accuracy but also reduced the execution time
of the algorithms by the resulting dimensionality reduction.
Among the wider distinction between linear and nonlinear methods, the results were split. While
sub1 and sub2 got the best accuracies using linear LR, sub3 and sub4 got the best results with
QDA and NLR—both nonlinear methods. The reason for this seems to be different data acquisition
methods used for the two datasets. While sub1 and sub2 data was collected with an ISI of 1000ms,
data for sub3 and sub4 was collected with an ISI of just 175ms. The short ISI resulted in the
overlapping of P300 amplitude of the EEG which happens about 300ms with the new stimulus
induced signal, which also explains why the grand average signal for sub3 and sub4 was not as
similar to a P300 as it was for sub1 and sub2. The overlapping of signals from many stimuli seems
like the reason that makes these two datasets more difficult to be linearly separable. Another
observation is that the method of treating each channel-subtrial as a separate subtrial for the
purpose of classification works well for P300. The hypothesis of PCA being able to extract the
relevant source components from this channel-subtrial dataset is also validated by the results. Also
for sub1 and sub2, the PCA captures the variance across channels into a component (it was the first
PC for sub1). Then the classifier would ignore that component as it won’t give it any significant
discriminatory value for the purpose of classification. Such variance was not that pronounced in the
case of sub3 and sub4.
5.2
Future Work
This work has many avenues for future work. As discussed earlier, as the P300 classification depends
a lot on the subjects, it would be useful to run these same experiments on data from more subjects to
see how well the results generalize across subjects. Then for the same subjects other classification
methods especially variants of LDA, such as Fisher’s linear discriminant (FLD), stepwise linear
discriminant analysis (SWLDA), and regularized Discriminant Analysis, can be tried to see if the
results could be further improved. Another interesting thing to try would be to compare our method
of channel-subtrial based data with other methods of data representation in spatio-temporal and
frequency domains. Also, the methods used in this work can be studied in more depth. For example,
it’s well known that LDA and QDA require a certain minimum number of data samples for optimal
performance. It would be interesting to see if and by how much could the performance further
improve if more number of samples are used for training the classifiers used in this work. For NLR,
we could also try multiple hidden layers to see how that would work for P300 classification.
14
References
[1] Bci competition iii challenge 2004 (http://www.bbci.de/competition/iii/desc_ii.pdf).
[2] Brain-computer interfaces laboratory, computer science department, colorado state university,
usa (http://www.cs.colostate.edu/eeg/main/data/2011-12_bci_csu).
[3] Butterworth filter design, mathworks (http://www.mathworks.com/help/signal/ref/butter.html).
[4] g.tec g.gammasys active electrode system (http://www.gtec.at/products/electrodes-andsensors/g.gammasys-specs-features).
[5] Kubler A, Kotchoubey B, Kaiser J, Wolpaw JR, and Birbaumer N. Brain-computer communication: unlocking the locked in. Psychological Bulletin, American Psychological Association,
vol 127, pp. 358-375, 2001.
[6] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
[7] V. Bostanov. Bci competition 2003-data sets ib and iib: feature extraction from eventrelated
brain potentials with the continuous wavelet transform and the t-value scalogram. IEEE Trans.
Biomed. Eng., vol. 51, pp. 1057-1061, 2004.
[8] Zachary Cashero. Comparison of EEG preprocessing methods to improve the classification of
P300 trials, M.S Thesis. Department of Computer Science, Colorado State University, USA,
2011.
[9] Donchin E, Spencer KM, and Wijesinghe R. The mental prosthesis: Assessing the speed of
a p300-based brain-computer interface. Rehabilitation Engineering, IEEE Transactions on ,
vol.8, no.2, pp. 174-179, 2000.
[10] Sellers EW and Donchin E. A p300-based brain-computer interface: Initial tests by als patients.
Clinical Neurophysiology 117, pp. 538-548, 2006.
[11] Gene H. Golub and Charles F. Van Loan. Matrix computations (4th ed.). Johns Hopkins
University Press Baltimore, MD, USA, 2013.
[12] Mirghasemi H, Fazel-Rezai R, and Shamsollahi MB. Analysis of p300 classifiers in brain computer interface speller. In the Proceedings of Conf Proc IEEE Eng Med Biol Soc. 2006;1:6205-8,
2006.
[13] Serby H, Yom-Tov E, and Inbar GF. An improved p300-based brain-computer interface. IEEE
Trans Neural Syst Rehabil Eng, vol. 13, pp. 89-98, 2005.
[14] Trevor. Hastie, Robert. Tibshirani, and J Jerome H Friedman. The Elements of Statistical
Learning. Springer New York, 2001.
[15] U. Hoffmann, J. Vesin, and T Ebrahimi. Recent advances in brain-computer interfaces. Multimedia Signal Processing, 2007. MMSP 2007. IEEE 9th Workshop on Recent advances in
brain-computer interfaces, pp. 17-17, 2007.
[16] M. Kaper, P. Meinicke, U. Grossekathoefer, T. Lingner, , and H. Ritter. Bci competition 2003data set iib: support vector machines for the p300 speller paradigm. IEEE Transactions on
Biomedical Engineering, vol. 51, pp. 1073-1076, 2004.
15
[17] Michael Kirby. Geometric Data Analysis: An Empirical Approach to Dimensionality Reduction
and the Study of Patterns. John Wiley & Sons, 2000.
[18] D. Krusienski, E. Sellers, F. Cabestaing, S. Bayoudh, D. McFarland, T. Vaughan, and J. Wolpaw. A comparison of classification techniques for the p300 speller. Journal of Neural Engineering, vol. 3, no. 4, pp. 299-305, 2006.
[19] D.J. Krusienski, E.W. Sellers, D.J. McFarland, T.M. Vaughan, and J.R. Wolpaw. Toward
enhanced p300 speller performance. Journal of Neuroscience Methods, vol. 167, pp. 15-21,
2008.
[20] Farwell LA and Donchin E. Talking off the top of your head: Toward a mental prosthesis
utilizing event-related brain potentials. Electroencephalography and Clinical Neurophysiology,
vol. 70, Issue 6, pp 510-523, 1988.
[21] Kun Li, Vanitha Narayan Raju, Ravi Sankar, Yael Arbel, and Emanuel Donchin. Advances
and challenges in signal analysis for single trial p300-bci. Springer-Verlag, Berlin, pp. 88-94,
2011.
[22] Steven J. Luck. An introduction to the event-related potential technique. MIT Press, Cambridge, USA, 2005.
[23] J. N. Mak, Y. Arbel, J. W. Minett, L. M. McCane, B. Yuksel, D. Ryan, D. Thompson,
L. Bianchi, and D. Erdogmus. Optimizing the p300-based brain-computer interface: current
status, limitations and future directions. Journal of Neural Engineering, vol. 8, no. 2, 2011.
[24] Martin F. Moller. A scaled conjugate gradient algorithm for fast supervised learning. Neural
Networks, vol. 6, pp. 525-533, 1993.
[25] Ernst Niedermeyer and Fernando Lopes da Silva. Electroencephalography: Basic Principles,
Clinical Applications, and Related Fields. Lippincott Williams and Wilkins, 2004.
[26] F. Piccione, F. Giorgi, P. Tonin, K. Priftis, S. Giove, S. Silvoni, G. Palmas, and F. Beverina.
P300-based brain computer interface: Reliability and performance in healthy and paralysed
participants. Clinical Neurophysiology 117, pp. 531-537, 2006.
[27] E.W. Sellers, D.J. Krusienski, D.J. McFarland, and J.R. Wolpaw. Towards brain-computer
interfacing. Cambridge, MA: The MIT Press, pp. 31-42, 2007.
[28] Gilbert Strang. Introduction to Linear Algebra, volume 1. Wellesley-Cambridge Press, 2009.
[29] Vaughan TM, Heetderks WJ, Trejo LJ, Rymer WZ, Weinrich M, Moore MM, Kubler A, Dobkin
BH, Birbaumer N, Donchin E, Wolpaw EW, and Wolpaw JR. Brain-computer interface technology: A review of the second international meeting. IEEE Transactions on Rehabilitation
Engineering, vol. 11(2), pp. 94-109, 2003.
[30] J. R. Wolpaw and D. J. McFarland. Control of a two-dimensional movement signal by a noninvasive brain-computer interface in humans. Proceedings of the National Academy of Sciences
of the United States of America, vol. 101(51), pp. 17849-17854, 2004.
[31] Jonathan R. Wolpaw, Niels Birbaumer, William J. Heetderks, Dennis J. McFarland, P. Hunter
Peckham, Gerwin Schalk, Emanuel Donchin, Louis A. Quatrano, Charles J. Robinson, , and
Theresa M. Vaughan. Brain-computer interface technology: A review of the first international
meeting. IEEE Transactions on Rehabilitation Engineering, vol. 8(2), pp. 164-73, 2000.
16
[32] Jonathan R. Wolpaw, Niels Birbaumer, Dennis J. McFarland, Gert Pfurtscheller, and
Theresa M. Vaughan. BrainâĂŞcomputer interfaces for communication and control. Neural
Networks, vol. 6(4), pp. 525–533, 1993.
[33] N. Xu, X. Gao, B. Hong, X. Miao, S. Gao, , and F. Yang. Bci competition 2003-data set
iib: Enhancing p300 wave detection using ica-based subspace projections for bci applications.
IEEE Transactions on Biomedical Engineering, vol. 3, pp. 1067-1072, 2004.
[34] Guoqiang Peter Zhang. Neural networks for classification: a survey. Systems, Man, and
Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, vol.30, no.4, pp. 451462, 2000.
17
| 9cs.NE
|
GPU-Accelerated Parallel Finite-Difference Time-Domain Method for
Electromagnetic Waves Propagation in Unmagnetized Plasma Media
Xi-min Wang, Lang-lang Xiong, Song Liu, Zhi-yun Peng, and Shuang-ying Zhong
Abstract—The finite-difference time-domain (FDTD) method
has been commonly utilized in the numerical solution of
electromagnetic (EM) waves propagation through the plasma
media. However, the FDTD method may bring about a significant
increment in additional run-times consuming for computationally
large and complicated EM problems. Graphics Processing Unit
(GPU) computing based on Compute Unified Device Architecture
(CUDA) has grown in response to increased concern for reduction
of run-times. We represent the CUDA-based FDTD method with
the Runge-Kutta exponential time differencing scheme (RKETD)
for the unmagnetized plasma implemented on GPU. In the paper,
we derive the RKETD-FDTD formulation for the unmagnetized
plasma comprehensively, and describe the detailed flowchart of
CUDA-implemented RKETD-FDTD method on GPU. The
accuracy and acceleration performance of the posed CUDA-based
RKETD-FDTD method implemented on GPU are substantiated by
the numerical experiment that simulates the EM waves traveling
through the unmagnetized plasma slab, compared with merely
CPU-based RKETD-FDTD method. The accuracy is validated by
calculating the reflection and transmission coefficients for
one-dimensional unmagnetized plasma slab. Comparison between
the elapsed times of two methods proves that the GPU-based
RKETD-FDTD method can acquire better application
acceleration performance with sufficient accuracy.
Index Terms—Parallel FDTD, Unmagnetized plasma, Compute
unified device architecture (CUDA), Graphic processing unit
(GPU)
I. INTRODUCTION
Since the finite-difference time-domain (FDTD) method was
initially delivered to numerically resolve the Maxwell’s
equations by Yee in 1966 [1], it has been widely used in the
numerical solution of electromagnetics (EMs) problems. The
FDTD method has obvious advantages over many other
numerical methods as it updates naturally the field values at
every separate cell in discrete time steps by leap-frog
integrations without complex computations [2]. Over the past
decade, the FDTD numerical modeling approach has led to
applications to diverse difficulties, involving modeling of
objects ranging from aerospace and biological systems to
geometric shapes, analysis and design of complicated
microwave circuits and fast time-varying systems as well as
various other engineering applications [3]. Plentiful numerical
schemes related to FDTD formulations used to simulate the EM
waves propagation in the dispersive media are addressed,
including the recursive convolution (RC) method [4,5],
frequency-dependent Z transform method [6,7], direct
integration (DI) method [8, 9], JE convolution (JEC) method
[10], the auxiliary differential equation (ADE) method [11],
Manuscript received; This work is supported by National Nature Science
foundation of China (No. 61261006 & 11563006 &1116501), the State Key
Laboratory of Millimeter Waves Open Research Program (No. K201605), and
it was also supported by the Natural Science Foundation of Jiangxi Province
(No. 20151BAB202024).
Xi-min Wang is with Electrical Engineering, Nanchang University,
Nanchang Jiangxi 330031, China.
Lang-lang Xiong is with School of Sciences, Nanchang University,
Nanchang Jiangxi 330031, China.
Song Liu is with School of Sciences, Nanchang University, Nanchang
Jiangxi 330031, China; and State Key Laboratory of Millimeter Waves,
Nanjing Jiangsu 210016, China. (Email: [email protected])
Zhi-yun Peng is with Electrical Engineering, Nanchang University,
Nanchang Jiangxi 330031, China.
Shuangying Zhong is with School of Sciences, Nanchang University,
Nanchang Jiangxi 330031, China; (Email: [email protected])
(corresponding author Email: [email protected]).
piecewise linear recursive convolution (PLRC) method [12],
piecewise
linear
current density recursive convolution
(PLCDRC) method [13], Runge-Kutta exponential time
differencing (RKETD) [14]. Simulation of EM waves
propagating through plasma media is a unique and fascinating
application built on FDTD formulations for dispersive media,
where the appearing nonlinear phenomena that are not totally
understood can be explicitly refined by numerical simulation.
Furthermore, aforementioned various FDTD scheme for
implementation of GPU-based RKETD-FDTD method with
dispersive media can be applied to the plasmas.
CUDA. Section Ⅳ designs a numerical simulation that is
Although the FDTD schemes above are well-suited to
carried out to prove the accuracy and efficiency of the
numerical simulation, the original FDTD method can bring
GPU-based RKETD-FDTD method for unmagnetized plasma.
about a significant increment in additional run-times consuming
II. RKETD-FDTD FORMULATION
for computationally large and complicated EM issues. However,
FDTD method is naturally a massively parallel algorithm, thus it
can benefit a lot from most advances in parallel computing
The famous Maxwell’s equations in time domain for the
unmagnetized plasma are provided by
techniques to acquire considerable decrease in time spending.
Graphics Processing Units (GPUs) have been industrialized
rapidly in recent years, which are considered as the most
popular hardware accelerator in parallel computation to settle
massively parallel computations with better application
H E J
0 t
(1)
E H
0 t
(2)
dJ J 2 E
0 p
dt
(3)
where, H is the magnetic intensity, E is the electric field, J
acceleration performance obtained.
Since the Compute Unified Device Architecture (CUDA)
is the polarization current density, is the electron collision
was first introduced by NVIDIA as a parallel computing
frequency, p is plasma frequency, 0 and 0 are the
platform and software programming model, GPUs have turned
out to be much more formidable and generalized for developers
to be utilized to tackle general-purpose parallel computations
with high performance. Previous researches about diverse
permittivity and permeability of free space, respectively.
Considering one-dimensional equations, the one-dimensional
component of the equations can be written as
FDTD schemes based on CUDA-enabled GPU have been
undertaken over the past decade [15-20]. What’s more, FDTD
method carried out on multi-GPU clusters has triggered great
interest to further accelerate large-scale computations with
improved speedup performance [21-24].
This paper presents a GPU-based RKETD-FDTD method
with CUDA for the unmagnetized plasma media to acquire
H y
Ex
Jx
0 t
z
H y
Ex
0
z
t
(5)
dJ x
J x 2p Ex
0
dt
(6)
To derive FDTD formulation with RKETD scheme for the
unmagnetized plasma, equation (6) is multiplied by the
t
better acceleration performance, compared with merely
integrating factor e
CPU-based RKETD-FDTD method. Numerical simulation of
letting
the GPU-based RKETD-FDTD method for the unmagnetized
plasma media is undertaken both on CPU and GPU respectively.
The reflection and transmission coefficients through an
unmagnetized plasma layer in one dimension are calculated to
validate the accuracy of the method. Speedup ratios are
and then integrated from
t tn1 tn and t tn
the Maxwell equations for the unmagnetized plasma and derives
the FDTD formulation with RKETD numerical scheme. Section
Ⅲ
illustrates
the
CUDA
programming
model
and
t n to tn 1 ,
. The result is given by
t
J xn1 et J xn et e F (tn )d
0
(7)
Where, F (tn ) 0 p2 Ex (tn ) .
To derive the second-order RKETD method, first define
K et J xn
F (tn , J x )(1 et )
computed to confirm the speedup performance of the method.
This paper is arranged as follows. In Section Ⅱ, we describe
(4)
(8)
Next the approximation is taken to give
F (tn ) F (tn , J x )
F (tn t, K ) F (tn , J x ) o (t )2
t
Combining equation (7) and (9) yields
t
J xn1 et J xn F (tn , J x )et e d
0
(9)
F (tn t , K ) F (tn , J x )
The x component of J at
et
t
t
0
e d
(10)
n 1 time step can be arranged as
J xn1 et J xn
1 e
t
0 p2 Exn
et 1 t
0 p2 Exn1 Exn
2 t
GPU hardware more efficiently.
CUDA programs involve codes running on two different
platforms concurrently: host codes on CPU and device codes on
GPU. There are distinct hardware differences between CPU and
GPU, whose separate memory named host and device memory
(11)
respectively. Hence, the host is the location where the sequential
So equation (11) can be used to update the x component of the
parts of CUDA programs are executed, and the device is where
polarization current density J by utilizing the x component of
the intensively data-parallel parts are executed in the form of
the electric field E at both n and n 1 time steps.
kernels. Kernels are critical and unique to CUDA programs. A
The discretization of equation (4) follows the Yee grid and
kernel is always executed on the device after being invoked on
leapfrog-style algorithm, which is taken to give
the host, explicitly specified as the calculation and memory
Exn1 (k ) Exn (k )
t
H
0 z
1
n
2
y
1
(k ) H
2
1
n
2
y
manipulation for solely a single thread.
1
(k )
2
t
J xn1 (k ) J xn (k )
2 0
As soon as a kernel is launched on the host, a large number of
(12)
1
concurrently
generated,
each
performing
two-level thread hierarchy, which divides threads into blocks
p2 t t
(e 1)
t
2
E (k ) (1
) Exn (k )
(1 et ) J xn (k )
p2 t
2 0
1 2 (e 1 t )
2
n1
x
1
are
calculations of the same pattern in parallel. CUDA reveals a
Substituting equation (11) into equation (12) gives
n
n
t
1
1
[ H 2 (k ) H y 2 (k )]
0 z y
2
2
threads
and group blocks into grids. In the light of the hierarchy
structure, that is, CUDA organizes blocks and grids in three
(13)
dimension, programmers can utilize the hardware resources for
The discretization of equation (5) strongly resembles the FDTD
optimization more straightforwardly. Figure 1 shows an
formulation for vacuum, and the update equation of H is given
illustrative instance of a thread hierarchy structure with a
by
two-dimensional grid containing two-dimensional blocks.
n
n1
1
t
H y 2 (k ) H y 2 (k )
[ Exn (k 1) Exn (k )]
2
2 0 x
1
1
(14)
Effective thread configuration becomes essential to CUDA
programming for the obtainment of better application
acceleration performance.
III. CUDA IMPLEMENTATION
Device(GPU)
Host (CPU)
Grid 0
A. CUDA Programming Model
Parallel programming techniques are compatible with
Kernel 0
Block
(0,0)
Block
(1,0)
Block
(2,0)
Block
(0,1)
Block
(1,1)
Block
(2,1)
computations that could be operated on concurrence of plentiful
Kernel 1
data elements in parallel, with the intention of acquiring better
acceleration performance by mapping data elements to
Block(1,1)
numerous threads concurrently on GPU. During the process, a
Thread
(0,0)
Thread
(1,0)
Thread
(2,0)
Thread
(3,0)
Thread
(4,0)
bridge is built by CUDA programming model between
Thread
(0,1)
Thread
(1,1)
Thread
(2,1)
Thread
(3,1)
Thread
(4,1)
applications and their implementations on NVIDIA GPU.
Thread
(0,2)
Thread
(1,2)
Thread
(2,2)
Thread
(3,2)
Thread
(4,2)
Moreover, CUDA programming model abstracts the computer
Figure 1 (color online) a thread hierarchy structure with a two-dimensional grid
architecture to offer programmers effective perspectives to
containing two-dimensional blocks
manipulate memory and organize threads on CUDA-supported
Moreover, CUDA programming model represents the
GPU, so that they are able to harness the compute capability of
complete memory hierarchy clearly to access and manage
memory
for
optimal
performance.
Several
types
of
two kernels with dotted line border are launched sequentially
programmable device memory are presented, such as local
and executed in parallel on GPU. Shared memory is often used
memory, global memory, shared memory, constant memory and
to efficiently implement the FDTD update kernels for optimum,
texture memory. Figure 2 illustrates the GPU memory hierarchy
compared to merely global memory utilized. The execution
exposed in CUDA memory model. These memory spaces
configuration for each kernel is governed by the complete
slightly differ in the lifetimes, scopes and caching behaviors.
domain and allocated shared memory, for the reason that the
Local memory is allocated for threads in a kernel. Shared
calculating of a Yee’s cell is mapped to a thread on the device.
memory is shared by all threads of the same block, which can be
START
configured by threads of a block with lifetime as block. Global
memory accessible by all threads is expected to be much slower
Allocate CPU (host)
memory
Organize threads for
E-update-kernel
Initialize fields and
PML parameters
Allocate shared
memory for H-fields
than shared memory. Constant and texture memory accessible
by all threads are read-only memory, optimized for diverse
memory practices. Global, constant and texture memory are
continuing with same lifetime as an application [25].
Allocate GPU
(global) memory
Update E-fields with different
update schemes in plasma media
and free space respectively
Copy data from
CPU memory to
GPU memory
(Device) Grid
Block (1, 1)
Update J-fields in
plasma media
Shared Memory
Invoke kernels
Registers
Registers
Thread (0, 0)
Thread (1, 0)
Local
Memory
Local
Memory
？
Host
Organize threads for
H-update-kernel
Copy data back
from GPU memory
to CPU memory
Allocate shared
memory for E-fields
Retrieve results
Global Memory
Constant Memory
Update H-fields in
computational
domain
Deallocate GPU
memory and CPU
memory
Texture Memory
END
Figure 2 (color online) GPU memory hierarchy in CUDA memory model
B. CUDA-implemented RKETD-FDTD method
Figure 3 Flowchart of GPU-based FDTD method with CUDA
Compared with RKETD-FDTD method on CPU, we
IV. NUMERICAL SIMULATION
illustrate the flowchart of GPU-based RKETD-FDTD method
displayed in Figure 3. The host code is implemented to
A. Simulation environment
complete fields and parameters initialization, device memory
As briefly stated before, it’s theoretically predicted that the
allocation and release, and data transfer between host and
GPU-based RKETD-FDTD method can acquire better
device. To efficiently implement GPU-based RKETD-FDTD
acceleration performance compared with the corresponding
method with CUDA, device codes are considered to be divided
simply
into two kernels: H-update-kernel to update the magnetic
RKETD-FDTD method is utilized to simulate EM waves
intensity and E-update-kernel to update the electric field. The
CPU-based
RKETD-FDTD
method,
when
traveling through the unmagnetized plasma media. Nevertheless,
is vacuum. The key parameters of unmagnetized plasma are
the application acceleration performance of the GPU-based
wp 2 28.7 10 rad / s and 20GHz .
RKETD-FDTD method isn’t always as satisfying as estimated
in practical applications, which frequently changes with
problem
complexities,
GPU
hardware
properties
and
programmers’ skills.
9
The source wave exploited in numerical experimentation is the
Gaussian-derivative pulsed plane wave, which is given in time
domain by
(t 5 )2
Ei (t ) (t 5 ) exp
2 2
To evaluate the acceleration performance of GPU-based
RKETD-FDTD method, we prefer CUDA as the parallel
,
(15)
here, 15t .
programming platform to employ the acceleration of the
The execution configurations of the two kernels are arranged
GPU-based RKETD-FDTD method for the unmagnetized
both as 16 blocks in each grid and 256 threads in each block in
plasma media. The numerical experiment is performed on the
case of the computational domain, with CUDA programming
laptop provided with CUDA-supported NVIDIA GPU. The
model constructed.
CPU of the laptop is Intel Core i5 3230M, and GPU NVIDIA
To validate the accuracy of the GPU-based RKETD-FDTD
GeForce GT 650M. GeForce GT 650 is a low-end product
method with CUDA compared with the RKETD-FDTD method
based on graphics processor of NVIDIA Kepler architecture,
merely CPU-based, reflection and transmission coefficients are
developed for desktop applications. Some key specifications of
analyzed with the Fast Fourier Transform (FFT). Figure 4 and 5
GT650 are tabulated in Table 1. The development environment
clarify the magnitudes of reflection coefficient and transmission
is Microsoft Visual Studio 2013 (Community Edition) with
coefficient computed respectively by CUDA-implemented
CUDA toolkit 7.5 assembled, Windows 10 as operating system.
RKETD-FDTD method on GPU and RKETD-FDTD method
Table 1 Specifications of NVIDIA 650M
on CPU with those of the analytical solution. From Figure 4 and
Specification
GeForce GT 650M
Chip
GK107
CUDA cores
384
Processor clock
835MHz
Memory clock
900MHz
Memory size
4096MB
5,
it’s
taken
to
show
that
the
CUDA-implemented
RKETD-FDTD method achieves agreement in reasonable
accuracy with the RKETD-FDTD method and the analytical
solution at higher frequency.
B. Numerical Results
To exhibit the simulation of abovementioned RKETD-FDTD
formulation for the unmagnetized plasma media, we choose
appropriate parameters that are listed as follows. The entire
computational domain is 4096z in z axis, where z is
selected to be z 75 m as size of a cell. A single time step is
taken to be t 125ns . The unmagnetized plasma slab with
the thickness of 9cm occupies the middle 1200 cells, the PML
absorbing boundary conditions [26] are implemented at both
ends of 5 cells to avoid undesired reflections, and the rest space
Reflection coefficient magnitude (dB)
0
-20
-40
-60
FDTD on CPU
FDTD on GPU
Analytical solution
-80
0
10
20
30
40
50
60
70
80
90
100
Frequency (GHz)
Figure 4 (color online) Reflection coefficient magnitude versus frequency
60
Free Space
Plasma Slab
Free Space
40
-20
20
FDTD on CPU
FDTD on GPU
Analytical solution
Time Step = 3000
Ex(V/m)
Transmission coefficient magnitude (dB)
0
-40
0
-20
-40
FDTD on CPU
FDTD on GPU
-60
0
10
20
30
40
50
60
70
80
90
100
-60
Frequency (GHz)
0
1000
2000
3000
4000
FDTD Cells
Figure 5 (color online) Transmission coefficient magnitude versus frequency
Figure 6 through 9 show the electric field magnitude at each
FDTD cell after various time steps to demonstrate a
Figure 7 (color online) Electric field magnitude versus Yee’s cells after 3000
time steps
Gaussian-derivative pulsed plane wave propagating through an
unmagnetized
plasma
slab
with
50GHz
60
Free Space
and
Plasma Slab
Free Space
40
wp 2 50 10 rad / s . Some sharp corners occur at the
9
boundary conditions must be satisfied.
20
Ex(V/m)
border of the plasma slab and vacuum, for the reason that the
Time Step = 4000
0
-20
-40
FDTD on CPU
FDTD on GPU
60
Free Space
Plasma Slab
-60
Free Space
0
1000
3000
4000
Figure 8 (color online) Electric field magnitude versus Yee’s cells after 4000
20
Time Steps = 2000
Ex(V/m)
2000
FDTD Cells
40
time steps
0
60
-20
Free Space
Plasma Slab
Free Space
40
-40
FDTD on CPU
FDTD on GPU
20
0
1000
2000
3000
4000
FDTD Cells
Ex(V/m)
-60
Time Step = 5500
0
-20
Figure 6 (color online) Electric field magnitude versus Yee’s cells after 2000
-40
FDTD on CPU
FDTD on GPU
time steps
-60
0
1000
2000
3000
4000
FDTD Cells
Figure 9 (color online) Electric field magnitude versus Yee’s cells after 5500
time steps
Due to the fact that the application speedup performance may
vary with GPU device properties and developers’ programming
levels, speedup ratio is calculated to evaluate the practical
application speedup performance, which is termed as the ratio
acquire better acceleration performance, when compared with
between the elapsed times of FDTD computation part only on
the previous RKETD-FDTD method simply implemented on
CPU and on GPU in the simulation. Table 2 presents the elapsed
CPU. Numerical experiments of these two methods have been
time of FDTD computation part only on CPU and on GPU and
done respectively on NVIDIA GPU and on CPU. The accuracy
corresponding speedup ratio at diverse time steps with Yee’s
and
cells unchanged.
GPU-implemented RKETD-FDTD method were assessed by
Table 2 Speedup ratio at diverse time steps in the simulation
the numerical simulation of the EM waves traveling through the
acceleration
performance
of
the
represented
CPU Times
GPU Times
Speedup
unmagnetized plasma slab, compared with simply CPU-based
(s)
(s)
Ratios
RKETD-FDTD method. The accuracy was verified by
1000
0.258
0.058
4.45
calculating the reflection and transmission coefficients for
3000
0.770
0.263
2.93
one-dimensional unmagnetized plasma slab. The comparison
5000
1.292
0.481
2.69
between the elapsed times of two methods proved that the posed
7000
1.832
0.689
2.66
GPU-based RKETD-FDTD method implemented with CUDA
9000
2.321
0.887
2.62
can acquire decent acceleration with sufficient accuracy. The
10000
2.579
1.014
2.54
further research will be predicted to obtain more satisfactory
15000
3.906
1.513
2.58
acceleration performance while the numerical results of
20000
5.063
2.048
2.47
GPU-based RKETD-FDTD method are generalized to
Time Steps
two-dimensional and even three-dimensional conditions for
The speedup ratio is slowly decreasing with time step
increasing. The speedup ratios calculated in Table 2 validate the
speedup performance of the GPU-based RKETD-FDTD
method with respect to that only on CPU, by numerically
practical
applications.
We can easily make further improvement in speedup
performance for practical applications when the numerical
results of GPU-based RKETD-FDTD method are generalized
more
sophisticated
CUDA-supported GPUs specialized in scientific computation
such as Tesla can be utilized to acquire better acceleration
performance.
simulating EM waves traveling through the unmagnetized
plasma slab in one direction.
Besides,
REFERENCES
[1] Yee, Kane S. "Numerical solution of initial boundary value problems
involving Maxwell’s equations in isotropic media." IEEE Trans. Antennas
Propag., vol. 14, no. 3, pp. 302-307, 1966.
[2] Taflove, Allen, and Susan C. Hagness. Computational electrodynamics.
Artech house, 2005.
to the two- dimensional and even three-dimensional conditions
[3] Taflove, Allen. "Review of the formulation and applications of the
to get full use of the GPU computing resources, compared with
finite-difference time-domain method for numerical modeling of
the one dimensional condition just to confirm the speedup
electromagnetic wave interactions with arbitrary structures." Wave Motion,
performance. Besides, NVIDIA GeForce GT 650M used in the
vol. 10, no. 6, pp. 547-582, 1988.
[4] Luebbers, Raymond J., and Forrest Hunsberger. "FDTD for N th-order
simulation is one product of the low-end GPUs that are not
developed for scientific computation. More sophisticated
CUDA-supported GPU such as Tesla can be upgraded to obtain
better speedup performance.
dispersive media." IEEE Transactions on Antennas and Propagation,
vol.40, no. 11, pp. 1297-1301, 1992.
[5] Siushansian, Riaz, and Joe LoVetri. "A comparison of numerical techniques
for modeling electromagnetic dispersive media." IEEE Microwave and
Guided Wave Letters, vol. 5, no. 12, pp. 426-428, 1995.
V. CONCLUSION AND PERSPECTIVES
[6] Sullivan, Dennis M. "Frequency-dependent FDTD methods using Z
transforms." IEEE Transactions on Antennas and Propagation, vol. 40, no.
In this letter, we presented a CUDA-based RKETD-FDTD
method for the unmagnetized plasma implemented on GPU to
10, pp. 1223-1230, 1992.
[7] Pereda, José, Ángel Vegas, and Andrés Prieto. "FDTD modeling of wave
[20] Livesey, Matthew, James Francis Stack, Fumie Costen, Takeshi Nanri,
propagation in dispersive media by using the Mobius transformation
Norimasa Nakashima, and Seiji Fujino. "Development of a CUDA
technique." IEEE Transactions on Microwave Theory and Techniques, vol.
Implementation of the 3 D FDTD Method." IEEE Antennas & Propagation
50, no. 7, pp. 1689-1695, 2002.
Magazine, vol. 54, no. 5, pp. 186-195, 2012.
[8] Young, Jeffrey L. "Propagation in linear dispersive media: Finite difference
[21]Ong C, Weldon M, Cyca D, Okoniewski M. “Acceleration of large-scale
time-domain methodologies." IEEE Transactions on Antennas and
FDTD simulations on high performance GPU clusters”. 2009 Antennas &
Propagation, vol. 43, no. 4, pp. 422-426, 1995.
Propagation Society International Symposium, 2009, 1-4.
[9] Young, Jeffrey L. "A higher order FDTD method for EM propagation in a
[22]Shams R, Sadeghi P. “On optimization of finite-difference time-domain
collisionless cold plasma." IEEE Transactions on Antennas and
(FDTD) computation on heterogeneous and GPU clusters”. Journal of
Propagation, vol. 44, no. 9, pp. 1283-1289, 1996.
Parallel and Distributed Computing, vol. 71, no. 4, pp. 584-593, 2011.
[10] Chen, Qing, Makoto Katsurai, and Paul H. Aoyagi. "An FDTD
[23]Kim K H, Park Q H. “Overlapping computation and communication of
formulation for dispersive media using a current density." IEEE
three-dimensional FDTD on a GPU cluster”. Computer Physics
Transactions on Antennas and Propagation, vol. 46, no. 11, pp. 1739-1746,
1998.
Communications, vol. 183, no. 11, pp. 2364-2369, 2012.
[24]Nagaoka, T., and Watanabe S. Accelerating “three-dimensional FDTD
[11] Takayama, Yoshihisa, and Werner Klaus. "Reinterpretation of the
calculations on GPU clusters for electromagnetic field simulation”.
auxiliary differential equation method for FDTD." IEEE Microwave and
International Conference of the IEEE Engineering in Medicine and
Wireless Components Letters, vol. 12, no. 3, pp. 102-104, 2002.
Biology Society, 2012: 5691-4.
[12] Kelley, David F., and Raymond J. Luebbers. "Piecewise linear recursive
convolution for dispersive media using FDTD." IEEE Transactions on
Antennas and Propagation, vol. 44, no. 6, pp. 792-797, 1996.
[13] Shaobin Liu, Naichang Yuan, and Jinjun Mo. "A novel FDTD formulation
for dispersive media." IEEE Microwave and Wireless Components Letters,
vol. 13, no. 5, pp. 187-189, 2003.
[14] Song Liu, Shuangying Zhong, and Shaobin Liu. "Finite-difference
time-domain algorithm for dispersive media based on Runge-Kutta
exponential time differencing method." International Journal of Infrared
and Millimeter Waves, vol. 29, no. 3, pp. 323-328, 2008.
[15] Wang Ximin, Song Liu, Xuan Li, and Shuangying Zhong.
"GPU-accelerated finite-difference time-domain method for dielectric
media based on CUDA." International Journal of RF and Microwave
Computer-Aided Engineering, vol. 26, no. 6, pp. 512-518, 2016.
[16] Stefanski, Tomasz P., and Timothy D. Drysdale. "Acceleration of the 3D
ADI-FDTD method using Graphics Processor Units." In Microwave
Symposium Digest, 2009. MTT'09. IEEE MTT-S International. 2009.
[17] De Donno, Danilo, Alessandra Esposito, Luciano Tarricone, and Luca
Catarinucci. "Introduction to GPU computing and CUDA programming: A
case study on FDTD." IEEE Antennas and Propagation Magazine, vol. 52,
no. 3, pp. 116-122, 2010.
[18] Zunoubi, Mohammad Reza, Jason Payne, and William P. Roach. "CUDA
Implementation of TEz-FDTD Solution of Maxwell's Equations in
Dispersive Media." IEEE Antennas and Wireless Propagation Letters, no.
9, pp. 756-759, 2010.
[19] Lee, Kim Huat, Iftikhar Ahmed, Rick Siow Mong Goh, Eng Huat Khoo, Er
Ping Li, and Terence Gih Guang Hung. "Implementation of the FDTD
method based on Lorentz-Drude dispersive model on GPU for plasmonics
applications." Progress In Electromagnetics Research vol. 116, pp.
441-456, 2011.
[25]Cheng, John, Max Grossman, and Ty McKercher. “Professional CUDA C
Programming”. John Wiley & Sons, 2014.
[26] J.-P. Berenger, “A perfectly matched layer for the absorption of
electromagnetic waves,” J. Comput. Phys., vol. 114, no. 1, pp. 185–200,
1994.
| 5cs.CE
|
MGAN: T RAINING G ENERATIVE A DVERSARIAL N ETS WITH
M ULTIPLE G ENERATORS
arXiv:1708.02556v4 [cs.LG] 27 Oct 2017
Quan Hoang
University of Massachusetts-Amherst
Amherst, MA 01003, USA
[email protected]
Tu Dinh Nguyen, Trung Le, Dinh Phung
PRaDA Centre, Deakin University
Geelong, Australia
{tu.nguyen,trung.l,dinh.phung}@deakin.edu.au
A BSTRACT
We propose in this paper a new approach to train the Generative Adversarial Nets
(GANs) with a mixture of generators to overcome the mode collapsing problem.
The main intuition is to employ multiple generators, instead of using a single one
as in the original GAN. The idea is simple, yet proven to be extremely effective at
covering diverse data modes, easily overcoming the mode collapsing problem and
delivering state-of-the-art results. A minimax formulation was able to establish
among a classifier, a discriminator, and a set of generators in a similar spirit with
GAN. Generators create samples that are intended to come from the same distribution as the training data, whilst the discriminator determines whether samples are
true data or generated by generators, and the classifier specifies which generator a
sample comes from. The distinguishing feature is that internal samples are created
from multiple generators, and then one of them will be randomly selected as final
output similar to the mechanism of a probabilistic mixture model. We term our
method Mixture Generative Adversarial Nets (MGAN). We develop theoretical
analysis to prove that, at the equilibrium, the Jensen-Shannon divergence (JSD)
between the mixture of generators’ distributions and the empirical data distribution is minimal, whilst the JSD among generators’ distributions is maximal, hence
effectively avoiding the mode collapsing problem. By utilizing parameter sharing,
our proposed model adds minimal computational cost to the standard GAN, and
thus can also efficiently scale to large-scale datasets. We conduct extensive experiments on synthetic 2D data and natural image databases (CIFAR-10, STL-10 and
ImageNet) to demonstrate the superior performance of our MGAN in achieving
state-of-the-art Inception scores over latest baselines, generating diverse and appealing recognizable objects at different resolutions, and specializing in capturing
different types of objects by the generators.
1
I NTRODUCTION
Generative Adversarial Nets (GANs) (Goodfellow et al., 2014) are a recent novel class of deep
generative models that are successfully applied to a large variety of applications such as image, video
generation, image inpainting, semantic segmentation, image-to-image translation, and text-to-image
synthesis, to name a few (Goodfellow, 2016). From the game theory metaphor, the model consists of
a discriminator and a generator playing a two-player minimax game, wherein the generator aims to
generate samples that resemble those in the training data whilst the discriminator tries to distinguish
between the two as narrated in (Goodfellow et al., 2014). Training GAN, however, is challenging as
it can be easily trapped into the mode collapsing problem where the generator only concentrates on
producing samples lying on a few modes instead of the whole data space (Goodfellow, 2016).
Many GAN variants have been recently proposed to address this problem. They can be grouped into
two main categories: training either a single generator or many generators. Methods in the former
1
include modifying the discriminator’s objective (Salimans et al., 2016; Metz et al., 2016), modifying
the generator’s objective (Warde-Farley & Bengio, 2016), or employing additional discriminators to
yield more useful gradient signals for the generators (Nguyen et al., 2017; Durugkar et al., 2016).
The common theme in these variants is that generators are shown, at equilibrium, to be able to
recover the data distribution, but convergence remains elusive in practice. Most experiments are
conducted on toy datasets or on narrow-domain datasets such as LSUN (Yu et al., 2015) or CelebA
(Liu et al., 2015). To our knowledge, only Warde-Farley & Bengio (2016) and Nguyen et al. (2017)
perform quantitative evaluation of models trained on much more diverse datasets such as STL-10
(Coates et al., 2011) and ImageNet (Russakovsky et al., 2015).
Given current limitations in the training of single-generator GANs, some very recent attempts have
been made following the multi-generator approach. Tolstikhin et al. (2017) apply boosting techniques to train a mixture of generators by sequentially training and adding new generators to the
mixture. However, sequentially training many generators is computational expensive. Moreover,
this approach is built on the implicit assumption that a single-generator GAN can generate very
good images of some modes, so reweighing the training data and incrementally training new generators will result in a mixture that covers the whole data space. This assumption is not true in practice
since current single-generator GANs trained on diverse datasets such as ImageNet tend to generate
images of unrecognizable objects. Arora et al. (2017) train a mixture of generators and discriminators, and optimize the minimax game with the reward function being the weighted average reward
function between any pair of generator and discriminator. This model is computationally expensive and lacks a mechanism to enforce the divergence among generators. Ghosh et al. (2017) train
many generators by using a multi-class discriminator that, in addition to detecting whether a data
sample is fake, predicts which generator produces the sample. The objective function in this model
punishes generators for generating samples that are detected as fake but does not directly encourage
generators to specialize in generating different types of data.
We propose in this paper a novel approach to train a mixture of generators. Unlike aforementioned
multi-generator GANs, our proposed model simultaneously trains a set of generators with the objective that the mixture of their induced distributions would approximate the data distribution, whilst
encouraging them to specialize in different data modes. The result is a novel adversarial architecture
formulated as a minimax game among three parties: a classifier, a discriminator, and a set of generators. Generators create samples that are intended to come from the same distribution as the training
data, whilst the discriminator determines whether samples are true data or generated by generators,
and the classifier specifies which generator a sample comes from. We term our proposed model
as Mixture Generative Adversarial Nets (MGAN). We provide analysis that our model is optimized
towards minimizing the Jensen-Shannon Divergence (JSD) between the mixture of distributions induced by the generators and the data distribution while maximizing the JSD among generators.
Empirically, our proposed model can be trained efficiently by utilizing parameter sharing among
generators, and between the classifier and the discriminator. In addition, simultaneously training
many generators while enforcing JSD among generators helps each of them focus on some modes
of the data space and learn better. Trained on CIFAR-10, each generator learned to specialize in
generating samples from a different class such as horse, car, ship, dog, bird or airplane. Overall,
the models trained on the CIFAR-10, STL-10 and ImageNet datasets successfully generated diverse,
recognizable objects and achieved state-of-the-art Inception scores (Salimans et al., 2016). The
model trained on the CIFAR-10 even outperformed GANs trained in a semi-supervised fashion
(Salimans et al., 2016; Odena et al., 2016).
In short, our main contributions are: (i) a novel adversarial model to efficiently train a mixture
of generators while enforcing the JSD among the generators; (ii) a theoretical analysis that our
objective function is optimized towards minimizing the JSD between the mixture of all generators’
distributions and the real data distribution, while maximizing the JSD among generators; and (iii)
a comprehensive evaluation on the performance of our method on both synthetic and real-world
large-scale datasets of diverse natural scenes.
2
ĚŝƐĐƌŝŵŝŶĂƚŽƌ
͙
͙
dŝĞĚƉĂƌĂŵĞƚĞƌƐ
௨
dŝĞĚƉĂƌĂŵĞƚĞƌƐ
͙
M
dŝĞĚƉĂƌĂŵĞƚĞƌƐ
ĚŝƐƚŝŶŐƵŝƐŚďĞƚǁĞĞŶ
࢞ ĂŶĚܩ௨ ሺࢠሻ
M
͙
ǁŚŝĐŚŐĞŶĞƌĂƚŽƌ
ǁĂƐƵƐĞĚ͍
ĐůĂƐƐŝĨŝĞƌ
͙
Figure 1: MGAN’s architecture with K generators, a binary discriminator, a multi-class classifier.
2
G ENERATIVE A DVERSARIAL N ETS
Given the discriminator D and generator G, both parameterized via neural networks, training GAN
can be formulated as the following minimax objective function:
min max Ex∼Pdata (x) [log D (x)] + Ez∼Pz [log (1 − D (G (z)))]
(1)
G
D
where x is drawn from data distribution Pdata , z is drawn from a prior distribution Pz . The mapping
G (z) induces a generator distribution Pmodel in data space. GAN alternatively optimizes D and G
using stochastic gradient-based learning. As a result, the optimization order in 1 can be reversed,
causing the minimax formulation to become maximin. G is therefore incentivized to map every z to a
single x that is most likely to be classified as true data, leading to mode collapsing problem. Another
commonly asserted cause of generating less diverse samples in GAN is that, at the optimal point of
D, minimizing G is equivalent to minimizing the JSD between the data and model distributions,
which has been empirically proven to prefer to generate samples around only a few modes whilst
ignoring other modes (Huszár, 2015; Theis et al., 2015).
3
P ROPOSED M IXTURE GAN S
We now present our main contribution of a novel approach that can effectively tackle mode collapse
in GAN. Our idea is to use a mixture of many distributions rather than a single one as in the standard
GAN, to approximate the data distribution, and simultaneously we enlarge the divergence of those
distributions so that they cover different data modes.
To this end, an analogy to a game among K generators G1:K , a discriminator D and a classifier C
can be formulated. Each generator Gk maps z to x = Gk (z), thus inducing a single distribution
PGk ; and K generators altogether induce a mixture over K distributions, namely Pmodel in the data
space. An index u is drawn from a multinomial distribution Mult (π) where π = [π1 , π2 , ..., πK ]
is the coefficients of the mixture; and then the sample Gu (z) is used as the output. Here, we use
a predefined π and fix it instead of learning. The discriminator D aims to distinguish between this
sample and the training samples. The classifier C performs multi-class classification to classify
samples labeled by the indices of their corresponding generators. We term this whole process and
our model the Mixture Generative Adversarial Nets (MGAN).
Fig. 1 illustrates the general architecture of our proposed MGAN, where all components are parameterized by neural networks. Gk (s) tie their parameters together except the input layer, whilst C and
D share parameters except the output layer. This parameter sharing scheme enables the networks
to leverage their common information such as features at low-level layers that are close to the data
layer, hence helps to train model effectively. In addition, it also minimizes the number of parameters
and adds minimal complexity to the standard GAN, thus the whole process is still very efficient.
More formally, D, C and G1:K now play the following multi-player minimax optimization game:
min max J (G1:K , C, D) = Ex∼Pdata [log D (x)] + Ex∼Pmodel [log (1 − D (x))]
G1:K ,C D
( K
)
X
πk Ex∼PGk [log Ck (x)]
(2)
−β
k=1
3
where Ck (x) is the probability that x is generated by Gk and β > 0 is the diversity hyper-parameter.
The first two terms show the interaction between generators and the discriminator as in the standard
GAN. The last term should be recognized as a standard softmax loss for a multi-classification setting, which aims to maximize the entropy for the classifier. This represents the interaction between
generators and the classifier, which encourages each generator to produce data separable from those
produced by other generators. The strength of this interaction is controlled by β. Similar to GAN,
our proposed network can be trained by alternatively updating D, C and G1:K . We refer to Appendix A for the pseudo-code and algorithms for parameter learning for our proposed MGAN.
3.1
T HEORETICAL A NALYSIS
Assuming all C, D and G1:K have enough capacity, we show below that at the equilibrium point
of the minimax problem in Eq. (2), the JSD between the mixture induced by G1:K and the data
distribution is minimal, i.e. pdata = pmodel , and the JSD among K generators is maximal, i.e. two
arbitrary generators almost never produce the same data. In what follows we present our mathematical statement and the sketch of their proofs. We refer to Appendix B for full derivations.
Proposition 1. For fixed generators G1 , G2 , ..., GK and their mixture weights π1 , π2 , ..., πK , the
∗
and D∗ for J (G1:K , C, D) in Eq. (2) are:
optimal solution C ∗ = C1:K
πk pGk (x)
pdata (x)
Ck∗ (x) = PK
and D∗ (x) =
p
(x)
+ pmodel (x)
data
j=1 πj pGj (x)
Proof. It can be seen that the solution Ck∗ is a general case of D∗ when D classifies samples from
two distributions with equal weight of 1/2. We refer the proofs for D∗ to Prop. 1 in (Goodfellow
et al., 2014), and our proof for Ck∗ to Appendix B in this manuscript.
Based on Prop. 1, we further show that at the equilibrium point of the minimax problem in
Eq. (2), the optimal generator G∗ = [G∗1 , ..., G∗K ] induces the generated distribution p∗model (x) =
PK
∗
k=1 πk pGk (x) which is as closest as possible to the true data distribution pdata (x) while maintaining the mixture components pG∗k (x)(s) as furthest as possible to avoid the mode collapse.
Theorem 2. At the equilibrium point of the minimax problem in Eq. (2), the optimal G∗ , D∗ , and
C ∗ satisfy
G∗ = argmin (2 · JSD (Pdata kPmodel ) − β · JSDπ (PG1 , PG2 , ..., PGK ))
(3)
G
πk pG∗k (x)
pdata (x)
and D∗ (x) =
Ck∗ (x) = PK
p
(x)
+ pmodel (x)
∗
data
π
p
(x)
j=1 j Gj
∗
Proof. Substituting C1:K
and D∗ into Eq. (2), we reformulate the objective function for G1:K as
follows:
pmodel (x)
pdata (x)
+ Ex∼Pmodel log
pdata (x) + pmodel (x)
pdata (x) + pmodel (x)
( K
"
#)
X
πk pGk (x)
−β
πk Ex∼PGk log PK
j=1 πj pGj (x)
k=1
#)
( K
"
K
X
X
pGk (x)
=2 · JSD (Pdata kPmodel ) − log 4 − β
πk Ex∼PGk log PK
−β
πk log πk
j=1 πj pGj (x)
k=1
k=1
L (G1:K ) = Ex∼Pdata log
=2 · JSD (Pdata kPmodel ) − β · JSDπ (PG1 , PG2 , ..., PGK ) − log 4 − β
K
X
πk log πk
(4)
k=1
Since the last two terms in Eq. (4) are constant, that concludes our proof.
This theorem shows that progressing towards the equilibrium is equivalently to minimizing
JSD (Pdata kPmodel ) while maximizing JSDπ (PG1 , PG2 , ..., PGK ). In the next theorem, we further clarify the equilibrium point for the specific case wherein the data distribution has the form
4
PK
pdata (x) =
k=1 πk qk (x) where the mixture components qk (x)(s) are well-separated in the
sense that Ex∼Qk [qj (x)] = 0 for j 6= k, i.e., for almost everywhere x, if qk (x) > 0 then
qj (x) = 0, ∀j 6= k.
PK
Theorem 3. If the data distribution has the form: pdata (x) =
k=1 πk qk (x) where the mixture components qk (x)(s) are well-separated, the minimax problem in Eq. (2) or the optimization
problem in Eq. (3) has the following solution:
pG∗k (x) = qk (x) , ∀k = 1, . . . , K and pmodel (x) =
K
X
πk qk (x) = pdata (x)
k=1
, and the corresponding objective value of the optimization problem in Eq. (3) is −βH (π) =
PK
−β k=1 πk log π1k , where H (π) is the Shannon entropy.
Proof. Please refer to our proof in Appendix B of this manuscript.
Thm. 3 explicitly offers the optimal solution for the specific case wherein the real data are generated from a mixture distribution whose components are well-separated. This further reveals
that if the mixture components are well-separated, by setting the number of generators as the
number of mixtures in data and maximizing the divergence between the generated components
pGk (x)(s), we can exactly recover the mixture components qk (x)(s) using the generated components pGk (x)(s), hence strongly supporting our motivation when developing MGAN. In practice, C, D, and G1:K are parameterized by neural networks and are optimized in the parameter
space rather than in the function space. As all generators G1:K share the same objective function, we can efficiently update their weights using the same backpropagation passes. Empirically,
1
, ∀k ∈ {1, ..., K}, which further minimizes the objective value
we set the parameter πk = K
PK
1
−βH (π) = −β k=1 πk log πk w.r.t π in Thm. 3. To simplify the computational graph, we assume that each generator is sampled the same number of times in each minibatch. In addition, we
adopt the non-saturating heuristic proposed in (Goodfellow et al., 2014) to train G1:K by maximizing
log D (Gk (z)) instead of minimizing log D (1 − Gk (z)) .
4
R ELATED W ORK
Recent attempts to address the mode collapse by modifying the discriminator include minibatch
discrimination (Salimans et al., 2016), Unrolled GAN (Metz et al., 2016) and Denoising Feature
Matching (DFM) (Warde-Farley & Bengio, 2016). The idea of minibatch discrimination is to allow the discriminator to detect samples that are noticeably similar to other generated samples. Although this method can generate visually appealing samples, it is computationally expensive, thus
normally used in the last hidden layer of discriminator. Unrolled GAN improves the learning by
unrolling computational graph to include additional optimization steps of the discriminator. It could
effectively reduce the mode collapsing problem, but the unrolling step is expensive, rendering it
unscalable up to large-scale datasets. DFM augments the objective function of generator with one
of a Denoising AutoEncoder (DAE) that minimizes the reconstruction error of activations at the
penultimate layer of the discriminator. The idea is that gradient signals from DAE can guide the
generator towards producing samples whose activations are close to the manifold of real data activations. DFM is surprisingly effective at avoiding mode collapse, but the involvement of a deep DAE
adds considerable computational cost to the model.
An alternative approach is to train additional discriminators. D2GAN (Nguyen et al., 2017) employs
two discriminators to minimize both Kullback-Leibler (KL) and reverse KL divergences, thus placing a fair distribution across the data modes. This method can avoid the mode collapsing problem
to a certain extent, but still could not outperform DFM. Another work uses many discriminators to
boost the learning of generator (Durugkar et al., 2016). The authors state that this method is robust
to mode collapse, but did not provide experimental results to support that claim.
Another direction is to train multiple generators. The so-called MIX+GAN (Arora et al., 2017) is
related to our model in the use of mixture but the idea is very different. Based on min-max theorem
(Neumann, 1928), the MIX+GAN trains a mixture of multiple generators and discriminators with
5
different parameters to play mixed strategies in a min-max game. The total reward of this game is
computed by weighted averaging rewards over all pairs of generator and discriminator. The lack
of parameter sharing renders this method computationally expensive to train. Moreover, there is no
mechanism to enforce the divergence among generators as in ours.
Some attempts have been made to train a mixture of GANs in a similar spirit with boosting algorithms. Wang et al. (2016) propose an additive procedure to incrementally train new GANs on a
subset of the training data that are badly modeled by previous generators. As the discriminator is
expected to classify samples from this subset as real with high confidence, i.e. D (x) is high, the
subset can be chosen to include x where D (x) is larger than a predefined threshold. Tolstikhin
et al. (2017), however, show that this heuristic fails to address the mode collapsing problem. Thus
they propose AdaGAN to introduce a robust reweighing scheme to prepare training data for the next
GAN. AdaGAN and boosting-inspired GANs in general are based on the assumption that a singlegenerator GAN can learn to generate impressive images of some modes such as dogs or cats but fails
to cover other modes such as giraffe. Therefore, removing images of dogs or cats from the training
data and train a next GAN can create a better mixture. This assumption is not true in practice as
current single-generator GANs trained on diverse data sets such as ImageNet (Russakovsky et al.,
2015) tend to generate images of unrecognizable objects.
The most closely related to ours is MAD-GAN (Ghosh et al., 2017) which trains many generators
and uses a multi-class classifier as the discriminator. In this work, two strategies are proposed to address the mode collapse: (i) augmenting generator’s objective function with a user-defined similarity
based function to encourage different generators to generate diverse samples, and (ii) modifying discriminator’s objective functions to push different generators towards different identifiable modes by
separating samples of each generator. Our approach is different in that, rather than modifying the
discriminator, we use an additional classifier that discriminates samples produced by each generator
from those by others under multi-class classification setting. This nicely results in an optimization problem that maximizes the JSD among generators, thus naturally enforcing them to generate
diverse samples and effectively avoiding mode collapse.
5
E XPERIMENTS
In this section, we conduct experiments on both synthetic data and real-world large-scale datasets.
The aim of using synthetic data is to visualize, examine and evaluate the learning behaviors of
our proposed MGAN, whilst using real-world datasets to quantitatively demonstrate its efficacy
and scalability of addressing the mode collapse in a much larger and wider data space. For fair
comparison, we use experimental settings that are identical to previous work, and hence we quote
the results from the latest state-of-the-art GAN-based models to compare with ours.
We use TensorFlow (Abadi et al., 2016) to implement our model and will release the code after
publication. For all experiments, we use: (i) shared parameters among generators in all layers except
for the weights from the input to the first hidden layer; (ii) shared parameters between discriminator
and classifier in all layers except for the weights from the penultimate layer to the output; (iii) Adam
optimizer (Kingma & Ba, 2014) with learning rate of 0.0002 and the first-order momentum of 0.5;
(iv) minibatch size of 64 samples for training discriminators; (v) ReLU activations (Nair & Hinton,
2010) for generators; (vi) Leaky ReLU (Maas et al., 2013) with slope of 0.2 for discriminator and
classifier; and (vii) weights randomly initialized from Gaussian distribution N (0, 0.02I) and zero
biases. We refer to Appendix C for detailed model architectures and additional experimental results.
5.1
S YNTHETIC DATA
In the first experiment, following (Nguyen et al., 2017) we reuse the experimental design proposed
in (Metz et al., 2016) to investigate how well our MGAN can explore and capture multiple data
modes. The training data is sampled from a 2D mixture of 8 isotropic Gaussian distributions with a
covariance matrix of 0.02I and means arranged in a circle of zero centroid and radius of 2.0. Our
purpose of using such small variance is to create low density regions and separate the modes.
We employ 8 generators, each with a simple architecture of an input layer with 256 noise units
drawn from isotropic multivariate Gaussian distribution N (0, I), and two fully connected hidden
6
(a) Symmetric KL divergence.
(b) Wasserstein distance.
(c) Evolution of data (in blue) generated by GAN, UnrolledGAN, D2GAN
and our MGAN from the top row to the bottom, respectively. Data sampled
from the true mixture of 8 Gaussians are red.
Figure 2: The comparison of our MGAN and GAN’s variants on 2D synthetic dataset.
layers with 128 ReLU units each. For the discriminator and classifier, one hidden layer with 128
ReLU units is used. The diversity hyperparameter β is set to 0.125.
Fig. 2c shows the evolution of 512 samples generated by our model and baselines through time. It
can be seen that the regular GAN generates data collapsing into a single mode hovering around the
valid modes of data distribution, thus reflecting the mode collapse in GAN as expected. At the same
time, UnrolledGAN (Metz et al., 2016), D2GAN (Nguyen et al., 2017) and our MGAN distribute
data around all 8 mixture components, and hence demonstrating the abilities to successfully learn
multimodal data in this case. Our proposed model, however, converges much faster than the other
two since it successfully explores and neatly covers all modes at the early step 15K, whilst two
baselines produce samples cycling around till the last steps. At the end, our MGAN captures data
modes more precisely than UnrolledGAN and D2GAN since, in each mode, the UnrolledGAN
generates data that concentrate only on several points around the mode’s centroid, thus seems to
produce fewer samples than ours whose samples fairly spread out the entire mode, but not exceed
the boundary whilst the D2GAN still generates many points scattered between two adjacent modes.
Next we further quantitatively compare the quality of generated data. Since we know the true distribution Pdata in this case, we employ two measures, namely symmetric Kullback-Leibler (KL)
divergence and Wasserstein distance. These measures compute the distance between the normalized
histograms of 10,000 points generated from the model to true Pdata . Figs. 2a and 2b again clearly
demonstrate the superiority of our approach over GAN, UnrolledGAN and D2GAN w.r.t both distances (lower is better); notably the Wasserstein distances from ours and D2GAN’s to the true distribution almost reduce to zero, and at the same time, our symmetric KL metric is significantly better
than that of D2GAN. These figures also show the stability of our MGAN (black curves) and D2GAN
(red curves) during training as they are much less fluctuating compared with GAN (green curves)
and UnrolledGAN (blue curves).
Lastly, we perform experiments with different numbers of generators. The MGAN models with 2, 3,
4 and 10 generators all successfully explore 8 modes but the models with more generators generate
fewer points scattered between adjacent modes. We also examine the behavior of the diversity
coefficient β by training the 4-generator model with different values of β. Without the JSD force
(β = 0), generated samples cluster around one mode. When β = 0.25, the JSD force is weak and
generated data cluster near 4 different modes. When β = 0.75 or 1.0, the JSD force is too strong and
causes the generators to collapse, generating 4 increasingly tight clusters. When β = 0.5, generators
successfully cover all of the 8 modes. Please refer to Appendix C.1 for experimental details.
7
5.2
R EAL - WORLD DATASETS
Next we train our proposed method on real-world databases from natural scenes to investigate its
performance and scalability on much more challenging large-scale image data.
Datasets. We use 3 widely-adopted datasets: CIFAR-10 (Krizhevsky & Hinton, 2009), STL-10
(Coates et al., 2011) and ImageNet (Russakovsky et al., 2015). CIFAR-10 contains 50,000 32×32
training images of 10 classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck.
STL-10, subsampled from ImageNet, is a more diverse dataset than CIFAR-10, containing about
100,000 96×96 images. ImageNet (2012 release) presents the largest and most diverse consisting of
over 1.2 million images from 1,000 classes. In order to facilitate fair comparison with the baselines
in (Warde-Farley & Bengio, 2016; Nguyen et al., 2017), we follow the procedure of (Krizhevsky
et al., 2012) to resize the STL-10 and ImageNet images down to 48×48 and 32×32, respectively.
Evaluation protocols. For quantitative evaluation, we adopt the Inception score proposed in (Salimans et al., 2016), which computes exp (Ex [KL (p (y|x) kp (y))]) where p (y|x) is the conditional
label distribution for the image x estimated by the reference Inception model (Szegedy et al., 2015).
This metric rewards good and varied samples and is found to be well-correlated with human judgment (Salimans et al., 2016). We use the code provided in (Salimans et al., 2016) to compute the
Inception scores for 10 partitions of 50,000 randomly generated samples. For qualitative demonstration of image quality obtained by our proposed model, we show samples generated by the mixture as
well as samples produced by each generator. Samples are randomly drawn rather than cherry-picked.
Model architectures. Our generator and discriminator architectures closely follow the DCGAN’s
design (Radford et al., 2015). The only difference is we apply batch normalization (Ioffe & Szegedy,
2015) to all layers in the networks except for the output layer. Regarding the classifier, we empirically find that our proposed MGAN achieves the best performance (i.e., fast convergence rate and
high inception score) when the classifier shares parameters of all layers with the discriminator except for the output layer. The reason is that this parameter sharing scheme would allow the classifier
and discriminator to leverage their common features and representations learned at every layer, thus
helps to improve and speed up the training progress. When the parameters are not tied, the model
learns slowly and eventually yields lower performance.
During training we observe that the percentage of active neurons chronically declined (see Appendix C.2). One possible cause is that the batch normalization center (offset) is gradually shifted
to the negative range, thus deactivating up to 45% of ReLU units of the generator networks. Our
ad-hoc solution for this problem is to fix the offset at zero for all layers in the generator networks.
The rationale is that for each feature map, the ReLU gates will open for about 50% highest inputs in
a minibatch across all locations and generators, and close for the rest.
We also experiment with other activation functions of generator networks. First we use Leaky ReLU
and obtain similar results with using ReLU. Then we use MaxOut units (Goodfellow et al., 2013)
and achieves good Inception scores but generates unrecognizable samples. Finally, we try SeLU
(Klambauer et al., 2017) but fail to train our model.
Hyperparameters. Three key hyperparameters of our model are: number of generators K, coefficient β controlling the diversity and the minibatch size. We use a minibatch size of [128/K] for
each generator, so that the total number of samples for training all generators is about 128. We
train models with 4 generators and 10 generators corresponding with minibatch sizes of 32 and 12
each, and find that models with 10 generators performs better. For ImageNet, we try an additional
setting with 32 generators and a minibatch size of 4 for each. The batch of 4 samples is too small
for updating sufficient statistics of a batch-norm layer, thus we drop batch-norm in the input layer
of each generator. This 32-generator model, however, does not obtain considerably better results
than the 10-generator one. Therefore in what follows we only report the results of models with 10
generators. For the diversity coefficient β, we observe no significant difference in Inception scores
when varying the value of β but the quality of generated images declines when β is too low or too
high. Generated samples by each generator vary more when β is low, and vary less but become less
realistic when β is high. We find a reasonable range for β to be (0.01, 1.0), and finally set to 0.01
for CIFAR-10, 0.1 for ImageNet and 1.0 for STL-10.
8
Inception results. We now report the Inception scores obtained by our MGAN and baselines in
Tab. 1. It is worthy to note that only models trained in a completely unsupervised manner without
label information are included for fair comparison; and DCGAN’s and D2GAN’s results on STL10 are available only for the models trained on 32×32 resolution. Overall, our proposed model
outperforms the baselines by large margins and achieves state-of-the-art performance on all datasets.
Moreover, we would highlight that our MGAN obtains a score of 8.33 on CIFAR-10 that is even
better than those of models trained with labels such as 8.09 of Improved GAN (Salimans et al.,
2016) and 8.25 of AC-GAN (Odena et al., 2016). In addition, we train our model on the original
96×96 resolution of STL-10 and achieve a score of 9.79±0.08. This suggests the MGAN can be
successfully trained on higher resolution images and achieve the higher Inception score.
Table 1: Inception scores on different datasets. “–” denotes unavailable result.
Model
CIFAR-10
STL-10
ImageNet
Real data
11.24±0.16 26.08±0.26 25.78±0.47
WGAN (Arjovsky et al., 2017)
3.82±0.06
–
–
MIX+WGAN (Arora et al., 2017)
4.04±0.07
–
–
Improved-GAN (Salimans et al., 2016)
4.36±0.04
–
–
ALI (Dumoulin et al., 2016)
5.34±0.05
–
–
BEGAN (Berthelot et al., 2017)
5.62
–
–
MAGAN (Wang et al., 2017)
5.67
–
–
GMAN (Durugkar et al., 2016)
6.00±0.19
–
–
DCGAN (Radford et al., 2015)
6.40±0.05
7.54
7.89
DFM (Warde-Farley & Bengio, 2016)
7.72±0.13
8.51±0.13
9.18±0.13
D2GAN (Nguyen et al., 2017)
7.15±0.07
7.98
8.25
MGAN
8.33±0.10
9.22±0.11
9.32±0.10
Image generation. Next we present samples randomly generated by our proposed model trained
on the 3 datasets for qualitative assessment. Fig. 3a shows CIFAR-10 32×32 images containing a
wide range of objects in such as airplanes, cars, trucks, ships, birds, horses or dogs. Similarly, STL10 48×48 generated images in Fig. 3b include cars, ships, airplanes and many types of animals, but
with wider range of different themes such as sky, underwater, mountain and forest. Images generated
for ImageNet 32×32 are diverse with some recognizable objects such as lady, old man, birds, human
eye, living room, hat, slippers, to name a few. Fig. 4a shows several cherry-picked STL-10 96×96
images, which demonstrate that the MGAN is capable of generating visually appealing images with
complicated details. However, many samples are still incomplete and unrealistic as shown in Fig. 4b,
leaving plenty of room for improvement.
(a) CIFAR-10 32×32.
(b) STL-10 48×48.
(c) ImageNet 32×32.
Figure 3: Images generated by our proposed MGAN trained on natural image datasets. Due to the
space limit, please refer to the appendix for larger plots.
Finally, we investigate samples generated by each generator as well as the evolution of these samples
through numbers of training epochs. Fig. 5 shows images generated by each of the 10 generators in
our MGAN trained on CIFAR-10 at epoch 20, 50, and 250 of training. Samples in each row corre9
(a) Cherry-picked samples.
(b) Incomplete, unrealistic samples.
Figure 4: Images generated by our MGAN trained on the original 96×96 STL10 dataset.
spond to a different generator. Generators start to specialize in generating different types of objects
as early as epoch 20 and become more and more consistent: generator 2 and 3 in flying objects
(birds and airplanes), generator 4 in full pictures of cats and dogs, generator 5 in portraits of cats
and dogs, generator 8 in ships, generator 9 in car and trucks, and generator 10 in horses. Generator
6 seems to generate images of frog or animals in a bush. Generator 7, however, collapses in epoch
250. One possible explanation for this behavior is that images of different object classes tend to
have different themes. Lastly, Wang et al. (2016) noticed one of the causes for non-convergence in
GANs is that the generators and discriminators constantly vary; the generators at two consecutive
epochs of training generate significantly different images. This experiment demonstrates the effect
of the JSD force in preventing generators from moving around the data space.
(a) Epoch #20.
(b) Epoch #50.
(c) Epoch #250.
Figure 5: Images generated by our MGAN trained on CIFAR10 at different epochs. Samples in each
row from the top to the bottom correspond to a different generator.
6
C ONCLUSION
We have presented a novel adversarial model to address the mode collapse in GANs. Our idea is
to approximate data distribution using a mixture of multiple distributions wherein each distribution
captures a subset of data modes separately from those of others. To achieve this goal, we propose a
minimax game of one discriminator, one classifier and many generators to formulate an optimization
problem that minimizes the JSD between Pdata and Pmodel , i.e., a mixture of distributions induced
by the generators, whilst maximizes JSD among such generator distributions. This helps our model
10
generate diverse images to better cover data modes, thus effectively avoids mode collapse. We term
our proposed model Mixture Generative Adversarial Network (MGAN).
The MGAN can be efficiently trained by sharing parameters between its discriminator and classifier, and among its generators, thus our model is scalable to be evaluated on real-world largescale datasets. Comprehensive experiments on synthetic 2D data, CIFAR-10, STL-10 and ImageNet
databases demonstrate the following capabilities of our model: (i) achieving state-of-the-art Inception scores; (ii) generating diverse and appealing recognizable objects at different resolutions; and
(iv) specializing in capturing different types of objects by the generators.
R EFERENCES
Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S
Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale machine
learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467, 2016. 5
Martin Arjovsky, Soumith Chintala, and Léon Bottou.
arXiv:1701.07875, 2017. 1
Wasserstein gan.
arXiv preprint
Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, and Yi Zhang. Generalization and equilibrium
in generative adversarial nets (gans). arXiv preprint arXiv:1703.00573, 2017. 1, 4, 1
David Berthelot, Tom Schumm, and Luke Metz. Began: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017. 1
Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised
feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp. 215–223, 2011. 1, 5.2
Vincent Dumoulin, Ishmael Belghazi, Ben Poole, Alex Lamb, Martin Arjovsky, Olivier Mastropietro, and Aaron Courville. Adversarially learned inference. arXiv preprint arXiv:1606.00704,
2016. 1
Ishan Durugkar, Ian Gemp, and Sridhar Mahadevan. Generative multi-adversarial networks. arXiv
preprint arXiv:1611.01673, 2016. 1, 4, 1
Arnab Ghosh, Viveka Kulharia, Vinay Namboodiri, Philip HS Torr, and Puneet K Dokania. Multiagent diverse generative adversarial networks. arXiv preprint arXiv:1704.02906, 2017. 1, 4
Ian Goodfellow.
Nips 2016 tutorial:
arXiv:1701.00160, 2016. 1, B
Generative adversarial networks.
arXiv preprint
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014. 1, 3.1, 3.1, B
Ian J Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. Maxout
networks. arXiv preprint arXiv:1302.4389, 2013. 5.2
Ferenc Huszár. How (not) to train your generative model: Scheduled sampling, likelihood, adversary? arXiv preprint arXiv:1511.05101, 2015. 2
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,
2015. 5.2
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014. 5
Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing
neural networks. arXiv preprint arXiv:1706.02515, 2017. 5.2
Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.
5.2
11
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,
2012. 5.2
Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.
In Proceedings of the IEEE International Conference on Computer Vision, pp. 3730–3738, 2015.
1
Andrew L Maas, Awni Y Hannun, and Andrew Y Ng. Rectifier nonlinearities improve neural network acoustic models. In Proc. ICML, volume 30, 2013. 5
Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial
networks. arXiv preprint arXiv:1611.02163, 2016. 1, 4, 5.1
Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In
Proceedings of the 27th international conference on machine learning (ICML-10), pp. 807–814,
2010. 5
J v Neumann. Zur theorie der gesellschaftsspiele. Mathematische annalen, 100(1):295–320, 1928.
4
Tu Dinh Nguyen, Trung Le, Hung Vu, and Dinh Phung. Dual discriminator generative adversarial
nets. In Advances in Neural Information Processing Systems 29 (NIPS), pp. accepted, 2017. 1, 4,
5.1, 5.2, 1
Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. arXiv preprint arXiv:1610.09585, 2016. 1, 5.2
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep
convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 5.2, 1
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual
recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. 1, 4,
5.2
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training gans. In Advances in Neural Information Processing Systems,
pp. 2234–2242, 2016. 1, 4, 5.2, 5.2, 1
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In
Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1–9, 2015.
5.2
Lucas Theis, Aäron van den Oord, and Matthias Bethge. A note on the evaluation of generative
models. arXiv preprint arXiv:1511.01844, 2015. 2
Ilya Tolstikhin, Sylvain Gelly, Olivier Bousquet, Carl-Johann Simon-Gabriel, and Bernhard
Schölkopf. Adagan: Boosting generative models. arXiv preprint arXiv:1701.02386, 2017. 1,
4
Ruohan Wang, Antoine Cully, Hyung Jin Chang, and Yiannis Demiris. Magan: Margin adaptation
for generative adversarial networks. arXiv preprint arXiv:1704.03817, 2017. 1
Yaxing Wang, Lichao Zhang, and Joost van de Weijer. Ensembles of generative adversarial networks. arXiv preprint arXiv:1612.00991, 2016. 4, 5.2
David Warde-Farley and Yoshua Bengio. Improving generative adversarial networks with denoising
feature matching. 2016. 1, 4, 5.2, 1
Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun:
Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv
preprint arXiv:1506.03365, 2015. 1
12
A
A PPENDIX : F RAMEWORK
In our proposed method, generators G1 , G2 , ... GK are deep convolutional neural networks parameterized by θ G . These networks share parameters in all layers except for the input layers. The input
layer for generator Gk is parameterized by the mapping fθG ,k (z) that maps the sampled noise z to
the first hidden layer activation h. The shared layers are parameterized by the mapping gθG (h) that
maps the first hidden layer to the generated data. The pseudo-code of sampling from the mixture is
described in Alg. 1. Classifier C and classifier D are also deep convolutional neural networks that
are both parameterized by θ CD . They share parameters in all layers except for the last layer. The
pseudo-code of alternatively learning θ G and θ CD using stochastic gradient descend is described in
Alg. 2.
Algorithm 1 Sampling from MGAN’s mixture of generators.
1: Sample noise z from the prior Pz .
2: Sample a generator index u from Mult (π1 , π2 , ..., πK ) with predefined mixing probability π =
(π1 , π2 , ..., πK ).
3: h = fθG ,u (z)
4: x = gθG (h)
5: Return generated data x and the index u.
Algorithm 2 Alternative training of MGAN using stochastic gradient descent.
1: for number of training iterations do
2:
Sample a minibatch of M data points x(1) , x(2) , ..., x(M) from the data distribution Pdata .
(1)
(2)
(N)
3:
Sample a minibatch of N generated data points x0 , x0 , ..., x0
and N indices
4:
5:
6:
7:
(u1 , u2 , ..., uN ) from the current
mixture.
PN
1
0 (n)
LC = − N n=1 log Cun x
h
i
1 PN
PM
(n)
1
(m)
− N n=1 log 1 − D x0
LD = − M
m=1 log D x
Update classifier C and discriminator D by descending along their gradient:
∇θCD (LC + LD ).
(1)
(2)
(N)
Sample a minibatch of N generated data points x0 , x0 , ..., x0
and N indices
(u1 , u2 , ..., uN ) from thecurrent
mixture.
PN
β PN
1
0 (n)
0 (n)
log
C
−N
x
8:
LG = − N n=1 log D x
u
n
n=1
9:
Update the mixture of generators G by ascending along its gradient: ∇θG LG .
10: end for
B
A PPENDIX : P ROOFS FOR S ECTION 3.1
Proposition 1 (Prop. 1 restated). For fixed generators G1 , G2 , ..., GK and mixture weights
∗
π1 , π2 , ..., πK , the optimal classifier C ∗ = C1:K
and discriminator D∗ for J (G, C, D) are:
πk pGk (x)
Ck∗ (x) = PK
j=1 πj pGj (x)
D∗ (x) =
pdata (x)
pdata (x) + pmodel (x)
Proof. The optimal D∗ was proved in Prop. 1 in (Goodfellow, 2016). This section shows a similar
proof for the optimal C ∗ . Assuming that C ∗ can be optimized in the functional space, we can
calculate the functional derivatives of J (G, C, D)with respect to each Ck (x) for k ∈ {2, ..., K}
13
and set them equal to zero:
!
!
ˆ
K
K
X
X
δJ
δ
= −β
π1 pG1 (x) log 1 −
Ck (x) +
πk pGk (x) log Ck (x) dx
δCk (x)
δCk (x)
k=2
k=2
πk pGk (x) π1 pG1 (x)
−
(5)
= −β
Ck (x)
C1 (x)
Setting
δJ (G,C,D)
δCk (x)
Ck∗ (x) =
to 0 for k ∈ {2, ..., K}, we get:
πk pGk (x)
PK
j=1 πj pGj (x)
π1 pG1 (x)
π2 pG2 (x)
πK pGK (x)
=
= ... =
∗ (x)
C1∗ (x)
C2∗ (x)
CK
PK
results from Eq. (6) due to the fact that k=1 Ck∗ (x) = 1.
(6)
Reformulation of L (G1:K ). Replacing the optimal C ∗ and D∗ into Eq. (2), we can reformulate
the objective function for the generator as follows:
L (G1:K ) = J (G, C ∗ , D∗ )
pdata (x)
pmodel (x)
= Ex∼Pdata log
+ Ex∼Pmodel log
pdata (x) + pmodel (x)
pdata (x) + pmodel (x)
( K
"
#)
X
πk pGk (x)
πk Ex∼PGk log PK
(7)
−β
j=1 πj pGj (x)
k=1
The sum of the first two terms in Eq. (7) was shown in (Goodfellow et al., 2014) to be 2 ·
JSD (Pdata kPmodel ) − log 4. The last term β{∗} of Eq. (7) is related to the JSD for the K distributions:
#
"
K
X
πk pGk (x)
∗=
πk Ex∼PGk log PK
j=1 πj pGj (x)
k=1
K
K
K
K
X
X
X
X
=
πk Ex∼PGk [log pGk (x)] −
πk Ex∼PGk log
πj pGj (x) +
πk log πk
k=1
=−
K
X
k=1
j=1
k=1
k=1
K
K
X
X
πk H (pGk ) + H
πj pGj (x) +
πk log πk
j=1
= JSDπ (PG1 , PG2 , ..., PGK ) +
k=1
K
X
πk log πk
(8)
k=1
where H (P ) is the Shannon entropy for distribution P . Thus, L (G1:K ) can be rewritten as:
L (G1:K ) = − log 4 + 2 · JSD (Pdata kPmodel ) − β · JSDπ (PG1 , PG2 , ..., PGK ) − β
K
X
πk log πk
k=1
PK
Theorem 3 (Thm. 3 restated). If the data distribution has the form: pdata (x) = k=1 πk qk (x)
where the mixture components qk (x)(s) are well-separated, the minimax problem in Eq. (2) or the
optimization problem in Eq. (3) has the following solution:
pG∗k (x) = qk (x) , ∀k = 1, . . . , K and pmodel (x) =
K
X
πk qk (x) = pdata (x)
k=1
, and the corresponding objective value of the optimization problem in Eq. (3) is −βH (π) =
PK
−β k=1 πk log π1k .
14
Proof. We first recap the optimization problem for finding the optimal G∗ :
min (2 · JSD (Pdata kPmodel ) − β · JSDπ (PG1 , PG2 , ..., PGK ))
G
The JSD in Eq. (8) is given by:
JSDπ (PG1 , PG2 , ..., PGK ) =
K
X
k=1
"
#
πk pGk (x)
πk Ex∼PGk log PK
j=1
πj pGj (x)
−
K
X
πk log πk
(9)
k=1
The i-th expectation in Eq. (9) can be derived as follows:
"
#
πk pGk (x)
Ex∼PGk log PK
≤ Ex∼PGk [log 1] ≤ 0
j=1 πj pGj (x)
and the equality occurs if
πk pGk (x)
PK
j=1 πj pGj (x)
= 1 almost everywhere or equivalently for almost every x
except for those in a zero measure set, we have:
pGk (x) > 0 =⇒ pGj (x) = 0, ∀j 6= k
(10)
Therefore, we obtain the following inequality:
JSDπ (PG1 , PG2 , ..., PGK ) ≤ −
K
X
πk log πk =
k=1
K
X
πk log
k=1
1
= H (π)
πk
and the equality occurs if for almost every x except for those in a zero measure set, we have:
∀k : pGk (x) > 0 =⇒ pGj (x) = 0, ∀j 6= k
It follows that
2 · JSD (Pdata kPmodel ) − β · JSDπ (PG1 , PG2 , ..., PGK ) ≥ 0 − βH (π) = −βH (π)
and we peak the minimum if pGk = qk , ∀k since this solution satisfies both
pmodel (x) =
K
X
πk qk (x) = pdata (x)
k=1
and the conditions depicted in Eq. (10). That concludes our proof.
C
C.1
A PPENDIX : A DDITIONAL E XPERIMENTS
S YNTHETIC 2D G AUSSIAN DATA
The true data is sampled from a 2D mixture of 8 Gaussian distributions with a covariance matrix
0.02I and means arranged in a circle of zero centroid and radius 2.0. We use a simple architecture
of 8 generators with two fully connected hidden layers and a classifier and a discriminator with one
shared hidden layer. All hidden layers contain the same number of 128 ReLU units. The input layer
of generators contains 256 noise units sampled from isotropic multivariate Gaussian distribution
N (0, I). We do not use batch normalization in any layer. We refer to Tab. 2 for more specifications
of the network and hyperparameters. “Shared” is short for parameter sharing among generators or
between the classifier and the discriminator. Feature maps of 8/1 in the last layer for C and D means
that two separate fully connected layers are applied to the penultimate layer, one for C that outputs
8 logits and another for D that outputs 1 logit.
The effect of the number of generators on generated samples. Fig. 6 shows samples produced
by MGANs with different numbers of generators trained on synthetic data for 25,000 epochs. The
model with 1 generator behaves similarly to the standard GAN as expected. The models with 2,
3 and 4 generators all successfully cover 8 modes, but the ones with more generators draw fewer
points scattered between adjacent modes. Finally, the model with 10 generators also covers 8 modes
wherein 2 generators share one mode and one generator hovering around another mode.
15
Table 2: Network architecture and hyperparameters for 2D Gaussian data.
Operation Feature maps Nonlinearity
Shared?
G (z) : z ∼ N (0, I) 256
Fully connected 128
ReLU
×
√
Fully connected 128
ReLU
√
Fully connected 2
Linear
C (x) , D (x) 2
√
Fully connected 128
Leaky ReLU
Fully connected 8/1
Softmax/Sigmoid
×
Number of generators 8
Batch size for real data 512
Batch size for each generator 128
Number of iterations 25,000
Leaky ReLU slope 0.2
Learning rate 0.0002
Regularization constants β = 0.125
Optimizer Adam(β1 = 0.5, β2 = 0.999)
Weight, bias initialization N (µ = 0, σ = 0.02I), 0
(a) 1 generator.
(b) 2 generators.
(c) 3 generators.
(d) 4 generators.
(e) 10 generators.
Figure 6: Samples generated by MGAN models trained on synthetic data with 2, 3, 4 and 10 generators. Generated data are in blue and data samples from the 8 Gaussians are in red.
The effect of β on generated samples. To examine the behavior of the diversity coefficient β,
Fig. 7 compares samples produced by our MGAN with 4 generators after 25,000 epochs of training
with different values of β. Without the JSD force (β = 0), generated samples cluster around one
mode. When β = 0.25, generated data clusters near 4 different modes. When β = 0.75 or 1.0,
the JSD force is too strong and causes the generators to collapse, generating 4 increasingly tight
clusters. When β = 0.5, generators successfully cover all of the 8 modes.
C.2
R EAL - WORLD DATASETS
Fixing batch normalization center. During training we observe that the percentage of active
neurons, which we define as ReLU units with positive activation for at least 10% of samples in
the minibatch, chronically declined. Fig. 8a shows the percentage of active neurons in generators
trained on CIFAR-10 declined consistently to 55% in layer 2 and 60% in layer 3. Therefore, the
quality of generated images, after reaching the peak level, started declining. One possible cause is
that the batch normalization center (offset) is gradually shifted to the negative range as shown in the
histogram in Fig. 8b. We also observe the same problem in DCGAN. Our ad-hoc solution for this
problem, i.e., we fix the offset at zero for all layers in the generator networks. The rationale is that for
each feature map, the ReLU gates will open for about 50% highest inputs in a minibatch across all
locations and generators, and close for the rest. Therefore, batch normalization can keep ReLU units
alive even when most of their inputs are otherwise negative, and introduces a form of competition
that encourages generators to “specialize” in different features. This measure significantly improves
performance but does not totally solve the dying ReLUs problem. We find that late in the training,
the input to generators’ ReLU units became more and more right-skewed, causing the ReLU gates
to open less and less often.
16
(a) β = 0
(b) β = 0.25
(c) β = 0.5
(d) β = 0.75
(e) β = 1.0
Figure 7: Samples generated by MGAN models trained on synthetic data with different values of
diversity coefficient β. Generated data are in blue and data samples from the 8 Gaussians are in red.
(a) % of active neurons in layer 2
and 3.
(b) Histogram of batch normalization centers in layer 2 (left) and
3 (right).
Figure 8: Observation of activate neuron rates and batch normalization centers in MGAN’s generators trained on CIFAR-10.
Experiment settings. For the experiments on three large-scale natural scene datasets (CIFAR10, STL-10, ImageNet), we closely followed the network architecture and training procedure of
DCGAN. The specifications of our models trained on CIFAR-10, STL-10 48×48, STL-10 96×96
and ImageNet datasets are described in Tabs. (3, 4, 5, 6), respectively. “BN” is short for batch
normalization and “BN center” is short for whether to learn batch normalization’s center or set it
at zero. “Shared” is short for parameter sharing among generators or between the classifier and the
discriminator. Feature maps of 10/1 in the last layer for C and D means that two separate fully
connected layers are applied to the penultimate layer, one for C that outputs 10 logits and another
for D that outputs 1 logit. Finally, Figs. (9, 10, 11, 12, 13) respectively are the enlarged version of
Figs. (3a, 3b, 3c, 4a, 4b) in the main manuscript.
Table 3: Network architecture and hyperparameters for the CIFAR-10 dataset.
Operation
G (z) : z ∼ Uniform [−1, 1]
Fully connected
Transposed convolution
Transposed convolution
Transposed convolution
C (x) , D (x)
Convolution
Convolution
Convolution
Fully connected
Number of generators
Batch size for real data
Batch size for each generator
Number of iterations
Leaky ReLU slope
Learning rate
Regularization constants
Optimizer
Weight, bias initialization
Kernel
5×5
5×5
5×5
5×5
5×5
5×5
Strides
2×2
2×2
2×2
2×2
2×2
2×2
Feature maps
100
4×4×512
256
128
3
32×32×3
128
256
512
10/1
10
64
12
250
0.2
0.0002
β = 0.01
Adam(β1 = 0.5, β2 = 0.999)
N (µ = 0, σ = 0.01), 0
17
BN?
√
√
√
BN center?
×
×
×
×
×
√
√
√
√
√
√
×
×
Nonlinearity
ReLU
ReLU
ReLU
Tanh
Leaky ReLU
Leaky ReLU
Leaky ReLU
Softmax/Sigmoid
Shared?
×
√
√
√
√
√
√
×
Table 4: Network architecture and hyperparameters for the STL-10 48×48 dataset.
Operation
G (z) : z ∼ Uniform [−1, 1]
Fully connected
Transposed convolution
Transposed convolution
Transposed convolution
Transposed convolution
C (x) , D (x)
Convolution
Convolution
Convolution
Convolution
Fully connected
Number of generators
Batch size for real data
Batch size for each generator
Number of iterations
Leaky ReLU slope
Learning rate
Regularization constants
Optimizer
Weight, bias initialization
Kernel
5×5
5×5
5×5
5×5
5×5
5×5
5×5
5×5
Strides
2×2
2×2
2×2
2×2
2×2
2×2
2×2
2×2
Feature maps
100
4×4×1024
512
256
128
3
48×48×3
128
256
512
1024
10/1
BN?
√
√
√
√
BN center?
×
×
×
×
×
×
√
√
√
√
√
√
√
√
×
×
Nonlinearity
ReLU
ReLU
ReLU
ReLU
Tanh
Leaky ReLU
Leaky ReLU
Leaky ReLU
Leaky ReLU
Softmax/Sigmoid
Shared?
×
√
√
√
√
√
√
√
√
×
10
64
12
250
0.2
0.0002
β = 1.0
Adam(β1 = 0.5, β2 = 0.999)
N (µ = 0, σ = 0.01), 0
Table 5: Network architecture and hyperparameters for the STL96×96 dataset.
Operation
G (z) : z ∼ Uniform [−1, 1]
Fully connected
Transposed convolution
Transposed convolution
Transposed convolution
Transposed convolution
Transposed convolution
C (x) , D (x)
Convolution
Convolution
Convolution
Convolution
Convolution
Fully connected
Number of generators
Batch size for real data
Batch size for each generator
Number of iterations
Leaky ReLU slope
Learning rate
Regularization constants
Optimizer
Weight, bias initialization
Kernel
5×5
5×5
5×5
5×5
5×5
5×5
5×5
5×5
5×5
5×5
Strides
2×2
2×2
2×2
2×2
2×2
2×2
2×2
2×2
2×2
2×2
Feature maps
100
4×4×2046
1024
512
256
128
3
32×32×3
128
256
512
1024
2048
10/1
10
64
12
250
0.2
0.0002
β = 1.0
Adam(β1 = 0.5, β2 = 0.999)
N (µ = 0, σ = 0.01), 0
18
BN?
√
√
√
√
√
BN center?
×
×
×
×
×
×
×
√
√
√
√
√
√
√
√
√
√
×
×
Nonlinearity
ReLU
ReLU
ReLU
ReLU
ReLU
Tanh
Leaky ReLU
Leaky ReLU
Leaky ReLU
Leaky ReLU
Leaky ReLU
Softmax/Sigmoid
Shared?
×
√
√
√
√
√
√
√
√
√
√
×
Table 6: Network architecture and hyperparameters for the ImageNet dataset.
Operation
G (z) : z ∼ Uniform [−1, 1]
Fully connected
Transposed convolution
Transposed convolution
Transposed convolution
C (x) , D (x)
Convolution
Convolution
Convolution
Fully connected
Number of generators
Batch size for real data
Batch size for each generator
Number of iterations
Leaky ReLU slope
Learning rate
Regularization constants
Optimizer
Weight, bias initialization
Kernel
5×5
5×5
5×5
5×5
5×5
5×5
Strides
2×2
2×2
2×2
2×2
2×2
2×2
Feature maps
100
4×4×512
256
128
3
32×32×3
128
256
512
10/1
BN?
√
√
√
BN center?
×
×
×
×
×
√
√
√
√
√
√
×
×
Nonlinearity
ReLU
ReLU
ReLU
Tanh
Leaky ReLU
Leaky ReLU
Leaky ReLU
Softmax/Sigmoid
10
64
12
50
0.2
0.0002
β = 0.1
Adam(β1 = 0.5, β2 = 0.999)
N (µ = 0, σ = 0.01), 0
Figure 9: Images generated by MGAN trained on the CIFAR-10 dataset.
19
Shared?
×
√
√
√
√
√
√
×
Figure 10: Images generated by MGAN trained on the rescaled 48×48 STL-10 dataset.
20
Figure 11: Images generated by MGAN trained on the rescaled 32×32 ImageNet dataset.
21
Figure 12: Cherry-picked samples generated by MGAN trained on the 96×96 STL-10 dataset.
22
Figure 13: Incomplete, unrealistic samples generated by MGAN trained on the 96×96 STL-10
dataset.
23
| 2cs.AI
|
A Kinodynamic Aggressive Trajectory Planner For Narrow Passages
arXiv:1709.05443v1 [cs.SY] 16 Sep 2017
Yaohui Guo1 , Zhaolun Su2 , Dmitry Berenson3 , Ding Zhao4
Abstract— Planning a path for a nonholonomic robot is a
challenging topic in motion planning and it becomes more
difficult when the desired path contains narrow passages. This
kind of scenario can arise, for instance, when quadcopters
search a collapsed building after a natural disaster. Choosing
the quadcopter as the target platform, this paper proposes the
Kinodynamic Aggressive Trajectory (KAT) motion planning
algorithm, which aims to compute aggressive trajectories for
narrow passages under nonholonomic constraints. This type of
maneuvers is necessary because the dynamics of quadcopters
entail that some narrow passages can only be traversed at high
speed.
To find the best path, the KAT uses RRT to determine
a holonomic path first and then adjusts it to satisfy the
nonholonomic constraints. The innovations in this process are:
1) The states of the robot are divided into near-holonomic
set and non-holonomic set, which makes the constraints local
rather than global; 2) For each of the most confined waypoints
in the path, KAT plans forward and backward simultaneously
around the waypoint to find a feasible local trajectory traversing
the narrow passage. Our approach thus transforms a globallyconstrained planning problem into a problem with local constraints, and as a result, the computation becomes tractable. We
evaluate KAT by applying it to a quadcopter flying through two
inclined holes that require aggressive maneuvers in a simulated
environment. The average computation time to successfully find
a solution for passing two 50◦ inclined holes is around 32
seconds.
I. I NTRODUCTION
A robot’s ability to guide itself is the basis for accomplishing higher level tasks, making motion planning a
popular and practical problem in robotics. But due to its
high computational complexity [1], it is still a challenging
problem. The difficulties in planning a path for an informed
robot in a complex environment arise from two principal
concerns. First, the existence of narrow passages makes
sampling method less efficient. Second, some robots have the
nonholonomic property, which makes their attainable region
a local submanifold of the workspace and thus, similarly,
the probability of reaching the goal is low if the robot
samples in the whole workspace. Therefore, planning for a
nonholonomic robot in an environment with narrow passages
could be difficult.
Manuscript created September 14, 2017.(Yaohui Guo and Zhaolun Su are
co-first authors.)
1 Y. Guo is with the Robotic Institute at the University of Michigan, Ann
Arbor [email protected]
2 Z. Su is with the Department of Electrical and Computer Engineering
at the University of Michigan, Ann Arbor [email protected]
3 D. Berenson is with the Department of Electrical and Computer Engineering and the Robotic Institute at the University of Michigan, Ann Arbor
[email protected]
4 D. Zhao is with the Department of Mechanical Engineering and Robotic
Institute at the University of Michigan, Ann Arbor (corresponding
author: [email protected])
Fig. 1: An aggressive trajectory computed by KAT.
Quadcopters are typical examples of such a problem.
Because of its outstanding mobility, the quadcopter has
been used widely in complex and confined environments for
applications such as exploration, inspection, and mapping.
For instance, a quadcopter is useful for exploring the inside
of a collapsed building in a search-and-rescue scenario,
where the environment is usually narrow and complex [2].
For example, the robot may need to pass through a tilted
window without colliding. Unfortunately, a quadcopter can
not maintain a tilted attitude at low speed because all the
forces applied by the propellers can not entirely offset the
gravitational force. Instead, a quadcopter can achieve an
instantaneous tilted state by exploiting its dynamics. Thus,
an aggressive maneuver is required.
In this paper, we present the Kinodynamic Aggressive
Trajectory (KAT) planner for computing a trajectory for a
nonholonomic robot in an environment with narrow passages.
The main idea of KAT is to eliminate the nonholonomic constraints at the near-holonomic states, and plan the trajectory
around narrow points by forward and backward control. This
is intended to let the algorithm focus on the bottlenecks of the
path, thus the global planning can be broken into several local
planning problems. The near-holonomic state assumption
assumes that when the quadcopter is flying steadily at very
low speed, it is able to change its direction of movement
abruptly by a small amount. This assumption allows the
quadcopter to move freely in the near-holonomic state and
adjust its trajectory to fit the aggressive flying path for the
narrow passage.
The paper is organized as follows: section II presents
a general overview of related work mainly in the aspect
of sampling-based planners; section III describes the math
model of the problem; section IV are the details on KAT
algorithm; section V presents the experiments we designed
for KAT and its results; section VII concludes the work and
provides an overview of the future work.
III. P ROBLEM S TATEMENT
We define our object of interest as a time-invariant dynamic system:
ṡ(t) = g(s(t), u(t)),
II. R ELATED W ORK
The Rapidly-exploring Random Trees (RRT) [3], as a
sampling-based motion planning algorithm, has been widely
used for a broad range of robotic systems. For instance,
in [4], LaValle and Kuffner presented the first randomized
approach to systems with kinodynamic constraints. Rather
than planning in the configuration space, their approach plans
the kinodynamic path in the state space considering the
kinodynamic planning as a generalization of holonomic planning. This approach solves the path planning for nonholonomic systems like spacecrafts and hovercrafts, but cannot
guarantee optimality and is not efficient for more complex
systems due to the high dimensionality of the state space.
To improve RRT for finding the optimal path, Karaman
and Frazzoli proposed RRT* [5] for holonomic systems,
which grows the same way as RRT except that the tree
will locally replan to ensure optimality. In [6], Karaman and
Frazzoli proposed an extension of RRT*, which could handle
nonholonomic dynamics systems. This algorithm leverages
the ball-box theorem to find an optimized extending range
for each step while guaranteeing the optimal path. This algorithm works well for nonholonomic planning but is difficult
to implement when the system is complex. In [7], Webb
and Berg introduced the kinodynamic RRT*. Like RRT*
algorithm, it is an asymptotically optimal motion planning
algorithm, using a fixed-final-state-free-final-time controller
to connect any pair of states optimally for systems with
controllable linear dynamics to achieve optimality. However,
this algorithm still needs to sample the state space, thus it
can be time-consuming on systems with high dimensionality.
Moon and Chung [8] presented the kinodynamic planner
Dual-Tree RRT (DR-RRT) for high-speed navigation of
differential drive robot which is composed of a workspace
tree and a state tree. The DT-RRT does not reduce the degree
of freedom directly. Instead, it searches in the workspace
to reduced the search complexity and tries to validate the
path in the state space. However, this algorithm is mainly
suitable for the low DOF kinodynamic system like Dubin’s
car and hard to implement on high DOF systems. Other
approaches like using motion primitives by building a path
set [9], constructing the state lattice [10], are to discretize the
state space in order to reduce the search complexity. These
methods can be used for real-time path planning due to the
high efficiency, but compromise the optimality.
Based on the previous work, searching in the state space
with the high dimensionality is the bottleneck of most
sampling based planners, when dealing with nonholonomic
systems. The KAT algorithm we proposed here aims to reduce this complexity by planning a path in the configuration
space first and then validating it in the state space.
s(0) = s0
where S ⊂ Rns is the state space of the robot; s(t) = st ∈ S
is the state of the robot at time t; U ⊂ Rnu is the input space
of the system; u(t) = ut ∈ U is the input of the system at
time t; g is the nonholonomic constraint of the system, which
will also be referred as the update function of the system.
For convenience, the following notations are used in this
work. The configuration of a robot is the robot’s location and
attitude; the state of a robot consists of the configuration
and the change rate of the configuration. We define the
configuration space of the object as C ⊂ Rnc . If the robot
is free when it is at configuration c, we define c as a
free configuration; otherwise c is a collision configuration.
Define the free configuration space as Cf ree ⊂ C and the
collision space as Ccol ⊂ C. Define function α : S → C
maps a state s to its configuration part c. Define the free
state set as Sf ree = {s | α(s) ∈ Cf ree } ⊂ S, and the
collision state set as Scol = {s | α(s) ∈ Ccol } ⊂ S.
The narrow configuration set Cnar ⊂ Cf ree is the set of
all the configurations in the narrow passages. The narrow
state set is defined as Snar = {s | α(s) ∈ Cnar } ⊂
Sf ree . Define the start configuration set as Cs ⊂ Cf ree
and the goal configuration set as Cg ⊂ Cf ree . Our goal
is to find a dynamically feasible path l : [0, T ] → Sf ree
connecting Cs and Cg while passing through some narrow
configuration cnar,i , i = 1, 2, 3, ..., nnar . This is equivalent
to α(l(0)) ∈ Cs and α(l(T )) ∈ Cg , and there exists
ti ∈ [0, T ], i = 1, 2, 3, ..., n such that α(l(ti )) = cnar,i ∈
{l(t)|t ∈ [0, T ]}, i = 1, 2, 3, ..., nnar . Also, to satisfy the
nonholonomic constraints, there exists valid input u(t) ∈ U,
t ∈ [0, T ] such that
Z T
l(T ) =
g(l(t), u(t))dt + l(0)
0
For numerical computation, we replace the integration with
summation, and get
l(T ) =
T
X
g(s(t), u(t))∆t + l(0)
0
For the quadcopter, we denote its configuration as c =
[p, r]| , where p = [x, y, z]| ∈ R3 is the translation of the
robot and r ∈ SO(3) is the rotation. For computational
convenience, here we use the quaternion r = [qr , qi , qj , qk ]|
instead of the Euler angles to represent the rotation. Therefore, we have c = [x, y, z, qr , qi , qj , qk ]| . The velocity of a
configuration is represented as: vc = [ẋ, ẏ, ż, ωx , ωy , ωz ]| .
Then a state can be expressed as s = [p, r, v, ω]| , where v =
[ẋ, ẏ, ż]| is the translational velocity and ω = [ωx , ωy , ωz ]|
is the angular velocity. Because a quadcopter can respond to
small changes in its velocity and pose almost instantly by a
linear controller when it is still [11], we can assume the quadcopter is not restricted by the nonholonomic constraints when
it is nearly still. Thus here we define the near-holonomic
state set as Sholo∗ = {s|wω ||ω|| + wv ||v|| + wr ||r − r0 || <
, s = [p, r, v, ω]| } where wv , wω and wr are the weights
and r0 = [1, 0, 0, 0]| is the unit quaternion parallel with z
axis. The quadcopter with a near-holonomic state means the
quadcopter could move freely in any direction within the
Sholo∗ . In this case, we can constrain the initial and goal
states in Sholo∗ , so the quadcopter does not have to obey
nonholonomic constraint when leaving the start configuration
and reaching the goal. The nonholonomic constraint is only
effective at the states where the quadcopter needs to conduct
aggressive maneuver.
IV. M ETHOD
The overview of our algorithm appears in Algorithm 1.
The method that we employ consists of four principal parts:
1) RRT planning in holonomic space
2) Sampling narrow configurations with maximum margin in narrow passage
3) Identifying escape velocity for each narrow configuration
4) Controller based dual-direction planning with nonholonomic constraints
The goal of the first planning in holonomic space is to
efficiently gather information regarding the narrow passage.
By testing the robot along the smoothed holonomic path, we
could identify the exact location of the narrow passage and
collect possible poses that would allow the quadcopter to
move through them.
Since the smoothed trajectory will typically hug the obstacle, it is almost impossible for such a trajectory to be used.
We propose implementing the maximum margin sampling
inside the narrow passage to avoid such scenarios. From
the smoothed holonomic path, we will be able to infer
the general configurations where the robot is in a narrow
passage. The algorithm will uniformly sample around the
cluster centers of the narrow points and replace each cluster
center by the configuration with the maximum margin to the
surrounding passage. It is obvious that using such a pose is
more likely to plan a successful path under nonholonomic
constraints.
Then the algorithm will search a velocity, defined as the
escape velocity, to complete the above configuration as a
candidate narrow state on the path. To reduce the risk of
collision and make it easier for the system to recover to a
near-holonomic state, the escape velocity will be the minimum velocity required to pass through the narrow passage.
Next, starting from the narrow state, a dual-direction
controller is employed to find a trajectory through the narrow
passage. The dual-direction controller plans both forward and
backward the dynamics function. If the planner can reach a
near-holonomic point sf ∈ Sholo∗ by forward planning and
a near-holonomic point sb ∈ Sholo∗ by backward planning,
it will return a local path connecting these two states for the
corresponding narrow passage.
Finally, with the local trajectories through each narrow
passage, we can use RRT again to find the paths connecting
the start point and end point of all these trajectories sequentially within Sholo∗ . Thus we have a global path satisfying
the nonholonomic constraints.
Algorithm 1 KAT(sg , ss , Env, robot)
Input: Cg , Cs , Env, robot
Output: lglobal
Initialization: KAT ← Cg , Cs , Env, robot
1: lholo ← RRT(Cg , Cs , Env, robot)
2: Cnar ← NarrowPoints(lholo , Env, robot)
3: for every cnar,i in Cnar do
4:
cnar,i ← MaxMargin(cnar,i , Env, robot)
5:
snar,i ← cnar,i , EscapeVelocity(cnar,i , Env)
6:
llocal,i ← planFB(snar,i , Env)
7: end for
8: lglobal ← RRTConnectLocalPath( all llocal,i )
9: return lglobal
A. Planning in Holonomic Space with White-listed RRT
The planner begins by sampling in the holonomic space
using the RRT algorithm [3]. The purpose of planning in
holonomic space is to gain information about the direction
and possible poses for crossing the narrow passage.
Fig. 2: An example showing the advantage of whitelisting.
In conventional RRT, if the new sample is biased as the
goal configuration, the nearest neighbor will be found from
the entire explored tree structure. However, we found this
algorithm could be inefficient, particularly when a narrow
passage presents. For example, in Figure 2, the samples
in this 2D environment would not be able to form a direct connection from the goal configuration to the nearest
neighbor A. However, it should be able to connect with
node B. We eliminate this kind of scenario by adding an
additional feature called whitelisting on top of the RRT
algorithm. It will keep a list of newly added nodes in the tree
structure and make sure every node will only be tested once.
Every time RRT samples the goal bias, instead of the entire
explored tree structure, the white list will be used to find the
nearest neighbor node for testing the goal connectivity and
the tested node will be deleted from the it. This means node
A in Figure 2 will only be tested its connectivity with the
goal once and yield for other nodes after it fails. Following
this pattern nodes like B will succeed much earlier during
planning.
Algorithm 2 NarrowPoints(path,Env,robot)
Input: path, Env ,robot
Output: Cnar
0
Initialization: Cnar ← ∅, Cnar
←∅
1: for every ci in path do
2:
set nbr as all 4-connected neighbors of ci
3:
if collisionCount(ngb, Env, robot) >= 4 then
0
0
4:
Cnar
← Cnar
∪ {ci }
5:
end if
6: end for
∗
0
← K-centroids(Cnar
)
7: Cnar
∗
8: Cnar ← MaxMarginSampling(Cnar
)
9: return Cnar
(a) RRT planning in holonomic (b) Maximum margin configuration and escape velocity
space
B. Maximum Margin Sampling in Narrow Passage
Since the algorithm plans under delicate conditions, it is
preferable to find a way to go through the narrow passage
while staying as far from the obstacles as possible. In order
to achieve this, each waypoint should be optimized to have
the margin to the nearest obstacle maximized. Unfortunately,
a smoothed holonomic path would tightly pass through
obstacles and leave very little room to work with. The KAT
algorithm resolves this problem with a maximum margin
sampling scheme.
After the RRT planning, KAT has found a collisionfree holonomic trajectory. The next step is to identify the
narrow points on this trajectory. This process is shown in
Algorithm 2. KAT first checks every point on the path and
records those that have more than four 4-connected neighbors
in Ccol . Since there may be many points around one narrow
passage, KAT uses K-centroids clustering to adaptively select
the cluster centers c∗nar,i identified for each narrow passage.
∗
= {c∗nar,i | i = 1, 2, 3, ..., nnar } constitutes the
The set Cnar
hardest part of the trajectory.
For each cluster center c∗nar,i generated from Algorithm 2,
KAT will sample configurations uniformly in the plane
perpendicular to the planned holonomic path lholo , which is
∗
denoted as C⊥
i = {c|c − cnar,i ⊥ lholo }. For the sampled
poses that are not in collision with an obstacle, we will
find the node in them with the lowest objective function
value. This objective function calculates the sum of squared
distances between this collision-free pose and all of the incollision poses. The objective function is formulated as:
X
cnar,i = arg min
(c − ccol )(c − ccol )T
c∈Cf ree ∩C⊥
i ccol ∈Ccol
Since the value of the objective function is not sensitive to
the points far away from the narrow passage, we can simplify
the equation above by only considering the collision points
near each narrow point. For instance, replace the constraint
on ccol from ccol ∈ Ccol to ccol ∈ Ccol ∩ Bδi , where Bδi =
{c | ||c − c∗nar,i || <= δ} and δ is a parameter related to
the scale of the environment and robot. Fig.3(b) shows the
location of this maximum margin sample derived from the
narrow configurations in a passage.
(c) Pose through narrow passage
(d) Final trajectory
Fig. 3: Four different stages of planning.
C. Escape Velocity
In this step, we will complete the narrow point cnar,i in
configuration space to a narrow state snar,i by appending a
translation velocity vt to it while setting the angular velocity
to zero. If vt is the velocity with the minimum norm that can
lead the robot through the narrow passage using the forward
and backward planning algorithm described in the next part,
it is called the escape velocity, denoted by vescape . KAT will
deduce the direction of vescape by taking the weighted mean
of a set of direction vectors, shown in Algorithm 3. Then the
vescape will be fully determined by the forward and backward
planning in the next part.
The algorithm first samples a direction set D = {di |di =
(xi , yi , zi ), |di | = 1, i = 1, 2, 3...n} uniformly distributed on
the unit sphere S2 . By reusing the cluster center c∗nar,i , we
can find a heuristic direction dnar,i for each narrow passage.
This is to first find the point cnnar,i which is the nearest to
c∗nar,i on lholo by
cnnar,i = arg min|c − c∗nar,i |
c∈lholo
and then identify the tangent direction at cnnar,i by
dnar,i =
dlholo
,
dtnar,i
lholo (tnar,i ) = cnnar,i
Next, for each di in D, if hdi , dnar,i i <= 0, it will be removed from D. After constructing D, a length ti is generated
for each di following a normal distribution N (µnar , σnar ),
where σnar and µnar should be selected according to the
property of the environment, and translate the robot from
the narrow point cnar,i by each pair of (ti , di ) to ctnar,i . If
ctnar,i ∈ Ccol , delete (di ) from D. Finally, if D is not empty,
the direction of the escape velocity will be calculated as
P
di ti
∗
vescape,i = Pdi ∈D
di ∈D di ti
Algorithm 3 EscapeVelocity(cnar ,Env,robot)
Input: cnar , Env ,robot
∗
Output: vescape
Initialization: D = {di |di ∈ S2 , i = 1, 2, 3...n}
1: remove all di goes against lholo from D
2: for every di in D do
3:
generate ti ∼ N (µnar , σnar )
4:
cnar,i = translate(cnar , di , ti )
5:
if cnar,i ∈ Ccol then
6:
delete di from D
7:
end if
8: end for
P
P
9: vescape = ( di ∈D di ti )/
di ∈D di ti
∗
10: return vescape
and removed the fastest and slowest. The video of the whole
computational processing is shown in [12]
A. Dynamic model
D. Forward and Backward Planning With Controller
This step finds local dynamically feasible paths through
∗
each narrow passage by exploiting cnar,i and vescape,i
deduced above. If such paths can be found, the bottlenecks of
the planning are solved since the remaining task is to connect
the starts and ends of each local path to form a global path.
The algorithm here will generate snar,i by adding vescape,i
∗
to cnar,i . The direction of vescape,i is determined by vescape,i
and its norm increases each time. For each generated snar,i
the algorithm will use a forward controller Cf : (S, S) → U
and a time-inverse controller Cb : (S, S) → U to stabilize
snar,i respectively. This process should satisfy
st+1 = g(st , ut )∆t + st
(
Cf (st , sstill ), t > tnar,i
ut =
Cb (st , sstill ), t <= tnar,i
for ∀t ∈ [0, Tnar,i ], and ∃tnar,i ∈ [0, Tnar,i ] such that
stnar,i = snar,i . If s0 , sTnar,i ∈ Sholo∗ and st ∈ Sf ree , ∀t ∈
[0, Tnar,i ], the path lnar,i (t) = st , t ∈ [0, Tnar,i ] is a local
dynamically feasible path through the corresponding narrow
passage with both start and end points in Sholo∗ . If the
algorithm can find a local passage llocal,i for every cnar,i
on lholo , the global path lglobal will be constructed from
connecting the start points and end points of these local
paths sequentially. Since the start and end of each llocal,i are
in Sholo∗ , the algorithm will use RRT to plan the path by
sampling in Sholo∗ without restricted by the nonholonomic
constraints.
V. E XPERIMENTS AND R ESULTS
This section describes the dynamics model that is used for
our experiment, results from different sections of the KAT
algorithm as well as the final path generated for different
environment settings.
Our algorithms were implemented in Python with Openrave. All experiments were executed on a laptop with an
Intel(R) Core(TM) i7-5600U at 2.6GHz, 8GB of RAM. Each
experiment ran until a trajectory was found, or 1 minutes
had elapsed. We performed 30 trials for each experiment
Fig. 4: The configuration of the quadcopter.
We exploit the quadcopter dynamics model and controller
developed in [13]. The configuration of the drone model is
shown in Figure 4. The inertia frame with axes e1 ,e2 ,e3 is a
reference frame attached to the ground; the body frame with
axes b1 ,b2 ,b3 is a frame attached to the drone. The dynamics
of the drone is described by the following equations:
ẋ = v,
v̇ = mge3 − f Re3
Ṙ = RΩ̂, J Ω̇ + Ω × JΩ = M
f
f1
1
1
1 1
M1 0 −d 0 d f2
=
M2 d
0 −d 0 f3
M3
−c c −c c
f4
where m ∈ R is the total mass of the quadcopter; J ∈ R3×3
is the inertia matrix with respect to the drone frame; R ∈
SO(3) is the rotation matrix from the body-fixed frame to
the inertial frame; Ω ∈ R3 is the angular velocity in the drone
frame; x ∈ R3 is the position of the center of mass in the
inertial frame; v ∈ R3 is the velocity of the center of mass in
the inertial frame; d ∈ R is the distance from the each axis
of the rotor to the drone center; fi ∈ R is the thrust of the
ith propeller; f ∈ R is the total thrust; τi ∈ R is the torque
applied to the drone by ith rotor along ith axis; M ∈ R3
is the moment vector in the body frame. The controller is
modified from [13]. The forward controller is
Mf = − kΩ eΩ + Ω × JΩ
+ Y (kzv R> ω̂ > − kz R)e3 ,
ff =(−kv ev + mge3 ) · Re3
and the backward controller is
Mb =kΩ eΩ − Ω × JΩ
− Y (kzv R> ω̂ > − kz R)e3 ,
fb =(kv ev + mge3 ) · Re3
where
0
Y = 1
0
1
0
0
At the last step, KAT connects the end and start points of
the local path with the corresponding nearest nodes in the
holonomic path and returns the result, as shown in Figure 3d.
0
0
0
The desired translational velocity and angular velocity are set
to zero; eΩ , ev are the angular velocity error and translational
velocity error; kΩ , kzv , kz , kv are control parameters;Mf , ff ,
Mb , Mb are the total moment and force of forward control
and backward control respectively.
B. Experiment Setup
The experiment setting is shown in Figure 3a. The start
configuration and goal configuration are at the different sides
of the wall; all paths connecting the two sides contain a
narrow passage on the wall, which is an inclined hole and
the quadcopter cannot reach the goal without passing it.
Fig. 5: Local path through the narrow passage.
C. Experiment Result
The holonomic path found by the RRT algorithm is shown
in Figure 3a. Table I shows the comparison of conventional
RRT with white-list RRT on the same environment setting.
By limiting the computation time to 1 minute, the conventional RRT has a slightly lower success rate than modified
RRT. In addition, the white-listed RRT finishes in less time
and with fewer sampled nodes.
TABLE I: Computation time of holonomic sampling
RRT
Modified RRT
Time
Sampled
Nodes
Success
Rate
Path Length
23.3s
17.7s
12197
9648
100%
100%
9.05
9.03
Fig. 6: Local path state space analysis.
This holonomic path identifies a collision-free trajectory
that can connect the goal and start, but violates the dynamics
of the quadcopter. One can easily find that the quadcopter
should fly at a low speed in order to be able to make a sharp
turn near the hole, which conflicts with the need of a high
speed pass for the inclined hole. By sampling and clustering
adjacent points of each node on the path, the algorithm
identifies the narrow passage and refines the configuration
for passing, as illustrated in Figure 3b. From Figure 3b, we
can see the refined point allows the quadcopter to leave a
safe margin from the wall. The next step is to sample forward
and backward to generate a feasible path passing the narrow
passage while making it possible to connect the start and end
nodes with a holonomic path.
The local path built here is shown in Figure 5. Figure 6
provides an analysis of this process. F is the thrust generated
by each propeller; vt is the translational velocity; Z is the
angle between b3 and e3 defined in Figure 4; tnar is the
time when the quadcopter passes the narrow passage. The
control input saturates when t = tnar , because the feedback
error reaches its maximum, which is the difference between
the instantaneous state and the near-holonomic state. This is
very different from planning the path from one side to the
other.
For testing the effectiveness of KAT, we derived a more
complex environment setup. By setting up two obstacle
walls, each with a different window opening angle, we
proved that KAT could connect multiple aggressive trajectories. Figure 1 shows the finished trajectory for passing these
two obstacles. In [12], the video demonstrates the entire
planning process and simulated execution of the planned
trajectory. We also changed the opening angle on both walls
and analyzed how it would affect the results. Table II shows
that as the opening gets steeper, the maximum speed that
KAT has to sample rises. As the opening becomes nearvertical, the maximum speed of the trajectory will be over
10 meters per second.
TABLE II: Average result of 30 trials for dual-obstacle setup
Opening
Rotation (◦ )
Computation Time
(second)
Max Velocity
(m/s)
Success Rate
(%)
0
30
50
85
17.1
23.2
32.5
52.4
0.3
2.1
6.4
15.2
100.0
100.0
89.47
70.0
VI. D ISCUSSION
R EFERENCES
Compared to other sampling-based planners, the advantage
of KAT is using the dual-direction control scheme to generate
local paths around narrow passages, saving a large amount
of computation time from sampling in the high dimensional
state space. This innovation reduces the time complexity
dramatically and can be extended to other similar motion
planning problems, where a particular subset of the problem
poses a much higher challenge than others.
Although the KAT algorithm has proven to be able to successfully generate aggressive flying patterns in a simulated
environment, it could conceivably encounter some difficulties
during implementation in real world environments because
the execution of a trajectory will always subject to drift in
practice. A robust controller might be required to resolve
this issue and make implementation more feasible. Future
work on KAT may include generalizing the concept of nearholonomic set on other robotic systems, and designing an
evaluation method of the success rate of the local aggressive
paths.
[1] J. T. Schwartz, Planning, geometry, and complexity of robot motion,
vol. 4. Intellect Books, 1987.
[2] N. Michael, S. Shen, K. Mohta, Y. Mulgaonkar, V. Kumar, K. Nagatani, Y. Okada, S. Kiribayashi, K. Otake, K. Yoshida, et al.,
“Collaborative mapping of an earthquake-damaged building via ground
and aerial robots,” Journal of Field Robotics, vol. 29, no. 5, pp. 832–
841, 2012.
[3] S. M. LaValle, Planning algorithms. Cambridge university press, 2006.
[4] S. M. LaValle and J. J. Kuffner Jr, “Randomized kinodynamic planning,” The international journal of robotics research, vol. 20, no. 5,
pp. 378–400, 2001.
[5] S. Karaman and E. Frazzoli, “Incremental sampling-based algorithms
for optimal motion planning,” in Proceedings of Robotics: Science and
Systems, (Zaragoza, Spain), June 2010.
[6] S. Karaman and E. Frazzoli, “Sampling-based optimal motion planning
for non-holonomic dynamical systems,” in Robotics and Automation
(ICRA), 2013 IEEE International Conference on, pp. 5041–5047,
IEEE, 2013.
[7] D. J. Webb and J. van den Berg, “Kinodynamic rrt*: Asymptotically
optimal motion planning for robots with linear dynamics,” in Robotics
and Automation (ICRA), 2013 IEEE International Conference on,
pp. 5054–5061, IEEE, 2013.
[8] C. b. Moon and W. Chung, “Kinodynamic planner dual-tree rrt (dtrrt) for two-wheeled mobile robots using the rapidly exploring random
tree,” IEEE Transactions on Industrial Electronics, vol. 62, pp. 1080–
1090, Feb 2015.
[9] R. A. Knepper and M. T. Mason, “Path diversity is only part of the
problem,” in Robotics and Automation (ICRA), 2009 IEEE International Conference on, pp. 3224–3229, IEEE, 2009.
[10] M. Pivtoraiko, R. A. Knepper, and A. Kelly, “Differentially constrained
mobile robot motion planning in state lattices,” Journal of Field
Robotics, vol. 26, no. 3, pp. 308–333, 2009.
[11] D. Brescianini, M. Hehn, and R. D’Andrea, “Quadrocopter pole
acrobatics,” in Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ
International Conference on, pp. 3472–3479, IEEE, 2013.
[12] Y. Guo, X. Zhang, and Z. Su, “Kat algorithm - youtube.”
https://www.youtube.com/watch?v=5R2-adY8F0E&
feature=youtu.be, 2017. (Accessed on 04/22/2017).
[13] G. Loianno, C. Brunner, G. McGrath, and V. Kumar, “Estimation,
control, and planning for aggressive flight with a small quadrotor with
a single camera and imu,” IEEE Robotics and Automation Letters,
vol. 2, no. 2, pp. 404–411, 2017.
VII. C ONCLUSION
In this work, we have proposed the KAT path planning
algorithm for systems with nonholonomic constraints. The
KAT is aimed to solve planning problem where aggressive
maneuver is required to pass the narrow passages. The
algorithm reduces the computation cost significantly by first
identifying the states allow the robot to pass the narrow
passages and then planning the local path exploiting the
dual-direction control scheme. In the simulation, KAT can
efficiently plan a quadcopter through two walls with tilted
holes, showing it is a effective planner for aggressive trajectories.
| 3cs.SY
|
Comparison Training for Computer Chinese Chess
1
Comparison Training for Computer Chinese
Chess
Wen-Jie Tseng1, Jr-Chang Chen2, I-Chen Wu1, Senior Member, IEEE, Tinghan Wei1
Abstract —This paper describes the application of comparison
training (CT) for automatic feature weight tuning, with the final
objective of improving the evaluation functions used in Chinese
chess programs. First, we propose an n-tuple network to extract
features, since n-tuple networks require very little expert
knowledge through its large numbers of features, while simultaneously allowing easy access. Second, we propose a novel evaluation method that incorporates tapered eval into CT. Experiments
show that with the same features and the same Chinese chess
program, the automatically tuned comparison training feature
weights achieved a win rate of 86.58% against the weights that
were hand-tuned. The above trained version was then improved
by adding additional features, most importantly n-tuple features.
This improved version achieved a win rate of 81.65% against the
trained version without additional features.
Index Terms—comparison training, n-tuple networks, machine
learning, Chinese chess
I. INTRODUCTION
C
HINESE chess is one of the most popular board games
worldwide, with an estimated player base of hundreds of
millions of people [28]. It is a two-player zero-sum game.
The state-space complexity and game-tree complexity of Chinese chess are 1048 and 10150 respectively, which are between
those of shogi and chess [1][11][12].
Most strong Chinese chess programs, such as SHIGA and
CHIMO [6][23], commonly use alpha-beta search
[3][11][14][25][26], similar to computer chess. When performing alpha-beta search, it is critical [5] to measure the
strength of positions accurately based on the features of pieces,
locations, mobility, threat and protection, king safety, etc. Position strength is usually evaluated from the weights of designated features. In the past, these features were carefully chosen
and their weights were manually tuned together with experts in
most programs. However, this work becomes difficult and
time-consuming when the number of features grows.
To avoid manually tuning the evaluation functions, two issues need to be taken into consideration during the design of
evaluation functions: define features and automatically tune
these feature weights. For the former, many proposed n-tuple
This work was supported in part by the Ministry of Science and Technology
of
Taiwan
under
contracts
MOST
106-2221-E-009-139-MY2,
106-2221-E-009-140-MY2,
106-2221-E-305-016-MY2
and
106-2221-E-305-017-MY2.
The authors1 are with the Department of Computer Science, National Chiao
Tung University, Hsinchu 30050, Taiwan. (e-mail: [email protected],
[email protected] (correspondent), and [email protected])
The author2 is with the Department of Computer Science and Information
Engineering, National Taipei University, New Taipei City 23741, Taiwan.
(e-mail: [email protected])
networks, which require very little expert knowledge through
its use of large numbers of features while allowing easy access.
It was successfully applied to Othello [15], Connect4 [22] and
2048 [18][27]. For the latter, machine learning methods were
used to tune feature weights to improve the strength of programs [2][19][20][21][24]. One of the successful methods,
called comparison training, was employed in backgammon
[19][20], shogi and chess programs [21][24]. Since Chinese
chess is similar to shogi and chess, it is worth investigating
whether the same technique can be applied to training for
Chinese chess. In contrast, although deep reinforcement
learning [16][17] recently made significant success on Go,
chess and shogi, the required computing power (5000 TPUs as
mentioned in [17]) is too costly for many developers.
This paper includes an n-tuple network with features that
take into consideration the relationship of material combinations and positional information from individual features in
Chinese chess. We then propose a novel evaluation method that
incorporates the so-called tapered eval [8] into comparison
training. Finally, we investigate batch training, which helps
parallelize the process of comparison training.
Our experiments show significant improvements through the
use of comparison training. With the same features, the automatically tuned comparison training feature weights achieved a
win rate of 86.58% against the weights that were hand-tuned.
We then improved by adding more features, most importantly
n-tuple features, into the above trained version. This improved
version achieved a win rate of 81.65% against the trained version without additional features.
The rest of this paper is organized as follows. Section II reviews related work on n-tuple networks and comparison training. Section III describes features used in Chinese chess programs and Section IV proposes a comparison training method.
Section V presents the experimental results. Section VI makes
concluding remarks.
II. BACKGROUND
In this section, we review Chinese chess in Subsection II.A,
the evaluation function using n-tuple networks in Subsection
II.B, the comparison training algorithm in Subsection II.C and
stage-dependent features in Subsection II.D.
A. Chinese Chess
Chinese chess is a two-player zero-sum game played on a
9 × 10 board. Each of two players, called red and black, has
seven types of pieces: one king (K/k), two guards (G/g), two
ministers (M/m), two rooks (R/r), two knights (N/n), two cannons (C/c) and five pawns (P/p). The abbreviations are uppercase for the red pieces and lowercase for the black pieces.
Red plays first, then the two players take turns making one
Comparison Training for Computer Chinese Chess
2
move at a time. The goal is to win by capturing the opponent’s
king. Rules are described in more detail in [28].
B. N-tuple Network
As mentioned in Section I, many researchers chose to use
n-tuple networks since they require very little expert
knowledge while allowing easy access. Evaluation functions
based on n-tuple networks are linear, and can be formulated as
follows.
eval(𝑤, 𝑠) = 𝑤T 𝜑(𝑠),
(1)
where 𝜑(𝑠) is a feature vector that indicates features in a position 𝑠, and 𝑤 is a weight vector corresponding to these features.
In [18], an n-tuple network was defined to be composed of 𝑚
𝑛𝑖 -tuples, where 𝑛𝑖 is the size of the 𝑖-th tuple. The 𝑛𝑖 -tuple is a
set of 𝑐𝑖1 × 𝑐𝑖2 × … × 𝑐𝑖𝑛𝑖 features, each of which is indexed by
𝑣𝑖1 , 𝑣𝑖2 , … , 𝑣𝑖𝑛𝑖 , where 0 ≤ 𝑣𝑖𝑗 < 𝑐𝑖𝑗 for all 𝑗. For example, one
6-tuple covers six designated squares on the Othello board [15]
and includes 36 features, where each square is empty or occupied by a black or white piece. Another example is that one
4-tuple covers four designated squares on the 2048 board [27]
and includes 164 features, since each square has 16 different
kinds of tiles.
For linear evaluation, the output is a linear summation of
feature weights for all occurring features. Thus, for each tuple,
since one and only one feature occurs at a time, the feature
weight can be easily accessed by table lookup. If an n-tuple
network includes 𝑚 different tuples, we need 𝑚 lookups.
C. Comparison Training
Tesauro introduced a learning paradigm called comparison
training for training evaluation functions [19]. He implemented
a neural-net-based backgammon program NEUROGAMMON
[20], which won the gold medal in the first Computer Olympiad
in 1989. He also applied comparison training to tuning a subset
of the weights in Deep Blue's evaluation function [21]. For the
game of shogi, Hoki [13] used comparison training to tune the
evaluation function of his program B ONANZA, which won the
2006 World Computer Shogi Championship. Thereafter, this
algorithm was widely applied to most top shogi programs.
Comparison training is a supervised learning method. Given
a training position 𝑠 and its best child position 𝑠1 , all the other
child positions are compared with 𝑠1 . The best child 𝑠1 is assumed to be the position reached by an expert’s move. The goal
of the learning method is to adjust the evaluation function so
that it tends to choose the move to 𝑠1 . The features involved in
the evaluation function for comparison are extracted for tuning.
Let 𝑤 (𝑡) be the weight vector in the 𝑡-th update. An online
training method is called averaged perceptron [9], described as
follows.
𝑑 (𝑡)
𝑤 (𝑡) = 𝑤 (𝑡−1) + 𝑑 (𝑡) ,
1
= (𝑡) ∑ (𝜑(𝑠1 ) − 𝜑(𝑠𝑖 )),
|𝑆 |
(𝑡)
(2)
𝑠𝑖 ∈𝑆
quantity for the 𝑡-th update. For each iteration, all training
positions are trained once, and at the end of the iteration, the
weight vector 𝑤 ∗ is updated to the average of all weight vectors,
𝑤 (0) to 𝑤 (𝑇) , where 𝑇 is the total number of training positions.
Then, 𝑤 (0) is set to 𝑤 ∗ at the beginning of the next iteration.
Incidentally, 𝑤 ∗ can be thought of as a measure of the training
quality of one iteration by counting the number of correct
moves of test positions. The whole training process stops when
this number decreases.
The research in [21] observed that the feature weights can be
tuned more accurately if the above evaluation values for all 𝑠𝑖
are more accurate, e.g., when the value of 𝑠𝑖 is obtained
through a deeper search. Thus, 𝑑-ply (comparison) training
was proposed by replacing 𝑠𝑖 in Formula (2) by the leaf 𝑙𝑖 on
𝑠𝑖 's principle variation (PV) in the minimax search with depth
𝑑, as shown in Fig. 1.
Fig. 1. An extension of comparison training (𝑑 = 3).
D. Stage-dependent Features
In many game-playing programs, the choice of features and
the feature weights depended on the game stages. This is because the importance of features varies as games progress. In
Othello, thirteen stages were designated according to the
number of pieces on the board [4]. In 2048, three or more stages
were designated according to certain tiles that occur [27]. In
chess evaluation functions, tapered eval was used to calculate
the stages based on the remaining pieces on the board, and
implemented by most chess programs, such as FRUIT, CRAFTY
and STOCKFISH.
In tapered eval, each feature has two weights representing its
weights at the opening and at the endgame. A game is divided
into many stages, and the weight of a feature is calculated by a
linear interpolation of the two weights for each stage. More
specifically, the weight vector 𝑤 in Formula (1) is replaced by
the following.
𝑤 = 𝛼(𝑠)𝑤𝑜 + (1 − 𝛼(𝑠))𝑤𝑒 ,
(3)
where 𝑤𝑜 is the weight vector at the opening, 𝑤𝑒 is that at the
endgame, and the game stage index 𝛼(𝑠), 0 ≤ 𝛼(𝑠) ≤ 1, indicates how close it is to the opening for position 𝑠.
Tapered eval is also well suited for Chinese chess programs.
For example, experts commonly evaluate cannons higher than
knights in the opening, but less in the endgame. Hence, it is
necessary to use different weights for the same feature in each
stage.
(𝑡)
where 𝑆 is the set of child positions of 𝑠 whose evaluation
values are higher than that of 𝑠1 when 𝑤 (𝑡−1) is applied to the
evaluation function, |𝑆 (𝑡) | is the set’s cardinality, and 𝜑(𝑠𝑖 ) is
the feature vector of 𝑠𝑖 . In this paper, 𝑑 (𝑡) is called the update
III. DEFINING FEATURES FOR CHINESE CHESS
This section describes three types of features in our evaluation function for Chinese chess in Subsections III.A, III.B and
III.C. Subsection III.D summarizes those features.
Comparison Training for Computer Chinese Chess
A. Features for Individual Pieces
This type of features consists of the following three feature
sets. The material (abbr. MATL) indicates piece types. The
location (abbr. LOC) indicates the occupancies of pieces on the
board. Symmetric piece locations share the same feature since
the importance of symmetric locations should be the same. E.g.,
with symmetry, there are only 50 LOC features for knights. The
mobility (abbr. MOB) indicates the number of places that pieces
can be moved to without being captured. E.g., in Fig. 2, the
mobility of piece R at location f4 is counted as seven since R
can only move to d4, e4, f2, f3, f6, f8 and h4 safely.
3
attack (e.g., by r). One 2-tuple of 50 × 90 can be used to represent the location relation of LOC2 for two knights when the
left-right symmetry of piece locations is considered. Thus, one
2-tuple of LOC2 is used for each combination of two pieces.
D. Summary of Features
Table I lists the numbers of features for each feature set 1.
Table I. Feature sets for Chinese chess.
Feature set
MATL LOC MOB AKA
# of features
6
194
26
5
SPC COP MATL2 LOC2
5
32
462
119,960
IV. COMPARISON TRAINING
In Subsection IV.A, we investigate comparison training for
feature weights based on tapered eval. In Subsection IV.B, we
present a batch training for comparison training. In Subsection
IV.C, we discuss the issue for initialization.
Fig. 2. A position for feature explanation.
B. Features for King Safety
This type of features indicates the threats to the king. Two
effective threats are attacking the king’s adjacency (abbr. AKA)
and safe potential check (abbr. SPC). When the king’s adjacent
locations are under attack, the mobility of the king is reduced so
that it increases the chance for the opponent to checkmate. The
AKA value is the summation of the number of such attacks. For
example, in Fig. 2, the AKA value to k is four, with contributions from one R, two from N, and one from C.
The SPC value is the summation of the number of moves,
made by opponent pieces 𝑝, which can check the king in the
next ply while 𝑝 will not be captured immediately after the
move. For example, in Fig. 2, the SPC value to K is three, where
one is from r, one from n, and one from c. Each AKA or SPC
value is given a weight individually and is seen as a feature.
C. Features for Piece Relation
This type of features consists of three feature sets. The first
feature set is chasing the opponent’s pieces (abbr. COP), indicating that piece 𝑝 chases opponent's piece 𝑞 if 𝑞 is immediately under 𝑝’s attack.
The second feature set, MATL2, identifies the material relations between the numbers of pieces of two piece types. For
example, it is well known in the Chinese chess community that
it is favorable towards the opponent if a player does not have
both guards and the opponent has at least one knight. Since
there are at most two guards and two knights, we can use one
2-tuple of (2 + 1) × (2 + 1) features to represent the material
relation of MATL2 for guards and knights. Each combination
of two piece types composes a 2-tuple of MATL2.
The third feature set, LOC2, extracts the locational relation
of two pieces, which may include attack or defense strategies.
For example, in Fig. 2, both pieces n at c7 and p at c6 are not in
a good shape since p blocks the move of n. Another example is
that two Ns protect each other, preventing from opponent’s
A. Comparison Training for Tapered Eval
For tapered eval, 𝑤 does not physically exist and is calculated from 𝑤𝑜 and 𝑤𝑒 for a position 𝑠 based on the proportion,
𝛼(𝑠) and 1 − 𝛼(𝑠), as in Formula (3). Thus, intuitively, the
update quantity for updating 𝑤𝑜 and 𝑤𝑒 should be also proportional to 𝛼(𝑠) and 1 − 𝛼(𝑠) respectively, as follows.
𝑤𝑒
= 𝑤𝑜 (𝑡−1) + 𝛼(𝑠)𝑑 (𝑡)
= 𝑤𝑒 (𝑡−1) + (1 − 𝛼(𝑠))𝑑 (𝑡)
𝑤𝑜
(𝑡)
(𝑡)
(4)
(5)
Unfortunately, when 0 < 𝛼(𝑠) < 1, the update quantity actually updated is less than what it should be, making the training imbalanced. For example, if 𝛼(𝑠) = 0.5, then
(𝑡)
𝑤 (𝑡) = 0.5𝑤𝑜 + 0.5𝑤𝑒 (𝑡)
= 0.5𝑤𝑜 (𝑡−1) + 0.5𝑤𝑒 (𝑡−1) + 0.5𝑑 (𝑡)
= 𝑤 (𝑡−1) + 0.5𝑑 (𝑡) ,
(6)
where only 0.5𝑑 (𝑡) is updated. Therefore, we propose new
formulas to make the training balanced as follows.
𝑤𝑜 (𝑡) = 𝑤𝑜 (𝑡−1) +
𝑤𝑒 (𝑡) = 𝑤𝑒 (𝑡−1) +
𝛼(𝑠)
𝛼(𝑠)2
2
𝑑 (𝑡)
(7)
2
𝑑 (𝑡)
(8)
+ (1 − 𝛼(𝑠))
1 − 𝛼(𝑠)
𝛼(𝑠)2 + (1 − 𝛼(𝑠))
The update quantity is proved to be equivalent to that in Formula (2) as follows.
𝑤 (𝑡) = 𝛼(𝑠)𝑤𝑜
(𝑡)
= (𝛼(𝑠)𝑤𝑜
+ (1 − 𝛼(𝑠))𝑤𝑒 (𝑡)
(𝑡−1)
+ (1 − 𝛼(𝑠))𝑤𝑒 (𝑡−1) )
2
+
𝛼(𝑠)2 + (1 − 𝛼(𝑠))
(𝑡)
2𝑑
𝛼(𝑠)2 + (1 − 𝛼(𝑠))
(9)
= 𝑤 (𝑡−1) + 𝑑 (𝑡)
B. Batch Training and Parallelization
In this subsection, we investigate batch training for comparison training. Batch training was commonly used to help
improve the training quality of machine learning. Since batch
1
For MOB, we only consider the two strong piece types, R and N. For AKA
and SPC, their possible values are 1 to 4 and ≥ 5. For LOC2, the number is
calculated by removing symmetry and illegal locations.
Comparison Training for Computer Chinese Chess
4
training also supports multiple threads, it helps speed up the
training process.
The most time-consuming part during training is searching
the training positions for the leaf nodes on PVs, such as 𝑙1 and
𝑙2 in Fig. 1. Once search is complete, some leaf positions are
selected into 𝑆 (𝑡) of Formula (2) as described in Subsection
II.C, and the new 𝑤 (𝑡) is updated based on the values of 𝑤 (𝑡−1) .
For batch training, we update the weight vector once after
searching a batch of 𝑁 training positions. Namely, Formula (2)
is changed to the following.
𝑡
𝑤
(𝑡)
=𝑤
(𝑡−𝑁)
+
∑
𝑑 (𝑖) ,
(10)
𝑖=𝑡−𝑁+1
(𝑖)
where 𝑑 is calculated based on the weight 𝑤 (𝑡−𝑁) . Let 𝑇 be
the total number of training positions. Then, only ⌈𝑇/𝑁⌉ updates are needed in one iteration.
The above batch training provides a way for parallelism,
using multiple threads to search 𝑁 positions in one batch.
Namely, each thread grabs one position to search whenever
idling. The computation of averaged weight vector 𝑤 ∗ remains
unchanged.
C. Weight Initialization
Before training, the weights of MATL are initialized as in
Table II. Other feature weights are initialized to zero.
Table II. Initial weights of training for MATL.
Guard
Minister
Rook
Knight
350
350
2000
950
Cannon
950
Pawn
300
V. EXPERIMENTS
In our experiments, the training data collected from [10] include 63,340 game records of expert players whose Elo ratings
exceed 2000. From these game records, 1.4 million positions
were randomly selected, one million for training and the rest for
testing.
For benchmarking, 1071 opening positions were selected
based on the frequencies played by experts, similar to [7].
Considering from the perspective of both players, a total of
2142 games were played in each experiment. Game results
were judged according to the Asian rules [28], and a game with
more than 400 plies was judged as a draw.
We list all the versions done in our experiments in Subsection V.A, and describe the effect of our training methods including tapered eval in Subsection V.B. The experiment and
comparison of all versions are described in Subsection V.C.
Subsection V.D shows experiments for batch training, and
Subsection V.E for 1-, 2-, 3- and 4-ply training.
A. List of Versions for Evaluation Functions
We used fourteen evaluation versions in experiments, based
on the feature sets in Table I. These versions are listed in Table
III. The version EVAL0 consists of features in the three feature
sets, MATL, LOC and MOB. EVAL1 includes the feature sets
used by EVAL0 plus the feature set AKA, and similarly,
EVAL2 to EVAL7 include extra feature sets as shown in Table
III. EVAL7 includes all the three feature sets, AKA, SPC and
COP.
EVAL8 to EVAL13 contain 2-tuple feature sets, MATL2
and/or LOC2, in addition to EVAL0 and EVAL7. The number
of features used in each version is also listed in the third column
of the table.
Table III. Features of evaluation functions.
Features sets
Evaluation versions
EVAL0
MATL+LOC+MOB
EVAL1
EVAL0+AKA
EVAL2
EVAL0+SPC
EVAL3
EVAL0+COP
EVAL4
EVAL0+AKA+SPC
EVAL5
EVAL0+AKA+COP
EVAL6
EVAL0+SPC+COP
EVAL7
EVAL0+AKA+SPC+COP
EVAL8
EVAL0+MATL2
EVAL9
EVAL7+MATL2
EVAL10
EVAL0+LOC2
EVAL11
EVAL7+LOC2
EVAL12
EVAL0+MATL2+LOC2
EVAL13
EVAL7+MATL2+LOC2
# of features
226
231
231
258
236
263
263
268
688
730
120,186
120,228
120,648
120,690
B. Training and Tapered Eval
This subsection describes the effect of our training methods
including tapered eval. The comparison training method described in Section IV.A is incorporated into the Chinese chess
program, CHIMO, which won the second place in Computer
Olympiad 2015. In the rest of this subsection, the original version (without training) is called the hand-tuned version (since
all the feature weights are tuned manually), while the incorporated version is called the trained version. For simplicity of
analysis, we only consider the features in EVAL0. Each move
takes 0.4 seconds on an Intel® Core™ i5-4690 processor,
which is about four hundred thousand nodes per move. In the
rest of this section, the time setting for each move is the same.
Table IV presents the performance comparisons of the
hand-turned and trained versions, with and without tapered eval
(marked with stars in the table when tapered eval is used). Note
that the trained version is based on 3-ply training. From the
table, the trained version clearly outperforms the hand-tuned
with and without tapered eval by win rates of 86.58% and
82.94% respectively. This shows that comparison training does
help improve strength significantly. Moreover, both trained and
hand-tuned versions using tapered eval also outperform those
without it by win rates of 62.75% and 53.76% respectively.
Table IV. Experiment results for comparison training and tapered eval. The
version marked with a star uses tapered eval.
Players
Win rate
trained vs. hand-tuned
82.94%
trained* vs. hand-tuned*
86.58%
hand-tuned* vs. hand-tuned
53.76%
trained* vs. trained
62.75%
In the rest of the experiments, all versions in Table III use
tapered eval and 3-ply training, unless specified explicitly.
C. Comparison for Using Different Feature Sets
This subsection compares all the versions listed in Table III
against EVAL0. Fig. 3 shows the win rates of
EVAL1-EVAL13 versions playing against EVAL0.
Comparison Training for Computer Chinese Chess
5
Batch training also helps significantly on training speedup.
Fig. 5 shows that for all versions EVAL0-EVAL13 the training
speedups on 4-core CPUs are 2.87-3.59 times as fast as those on
single-core CPUs for batch size 50, 3.37-3.68 times for 100,
and 3.57-3.7 times for 200.
Fig. 3. Win rates of all versions against EVAL0.
In general, as in Fig. 3, the more feature sets are added, the
higher the win rates are. The only exception is the case that
adds AKA from EVAL3 to EVAL5. Consider the three
non-tuple feature sets AKA, SPC and COP. From
EVAL1-EVAL3, SPC improves more than AKA and COP.
EVAL7 performs better than EVAL1-EVAL6.
Consider the two 2-tuple feature sets, MATL2 and LOC2.
Clearly, all versions including LOC2, EVAL10-EVAL13,
significantly outperform others. In contrast, the versions including MATL2 do not. EVAL13 that includes all feature sets
reaches a win rate of 81.65%, the best among all the versions.
D. Batch Training and Parallelization
First, we analyze the quality of batch training with three
batch sizes of 50, 100 and 200 for comparison training. Fig. 4
shows the win rates of the versions trained with three batch
sizes against those without batch training. The results show that
the versions with batch training perform slightly better for
EVAL0-EVAL9 and roughly equally for EVAL10-EVAL13.
For EVAL13, the win rates are 50.56%, 50.21% and 49.53%
with batch sizes 50, 100 and 200 respectively. Hence, in general,
batch training does not improve playing strength significantly
in this paper.
Fig. 4. Win rates of all versions with batch training against those without batch
training.
Fig. 5. Speedups.
E. Comparison from 1-ply to 4-ply Training
This subsection analyzes the strength of 1-, 2-, 3- and 4-ply
training. From Fig. 6, all versions trained with 2-, 3- and 4-ply
generally outperform those with 1-, 2- and 3-ply training respectively, except for EVAL10. Namely, for 4-ply vs. 3-ply,
EVAL5 improves most by a win rate of 55.56%, while
EVAL10 shows no improvement, with a win rate of 49.37%.
For 𝑑-ply training where 𝑑 ≥ 5, we tried one case for EVAL13
(performing best among the above versions) with 5-ply training.
5-ply training in this case shows no improvement over 4-ply
training, with a win rate of only 49.42%. Other experiments for
𝑑 ≥ 5 were omitted due to time constraints.
In our experiments, 3-ply training is sufficient since for
EVAL10-EVAL13, 4-ply and 3-ply training performed about
the same. Although for EVAL0-EVAL9, 4-ply training outperforms 3-ply training, EVAL10-EVAL13 are the stronger
versions that are used when competing.
Fig. 6. Win rates of 𝑑-ply training, 1 ≤ 𝑑 ≤ 4.
VI. CONCLUSIONS
This paper designs comparison training for computer Chinese chess. First, we propose to extract large numbers of features by leveraging n-tuple networks that require very little
expert knowledge while allowing easy access. Second, we also
propose a novel method to incorporate tapered eval into comparison training,
In our experiments, EVAL0 with 3-ply training outperforms
our original hand-tuned version by a win rate of 86.58%. With
3-ply training, the version EVAL13, including a 2-tuple network (consisting of LOC2 and MATL2), outperforms EVAL0
by a win rate of 81.65%. Moreover, EVAL13 with 4-ply
training performs better than 3-ply training by a win rate of
51.49%. The above shows efficiency and effectiveness to use
comparison training to tune feature weights. In our experiments,
batch training does not improve playing strength significantly.
However, it speeds up the training process by a factor of
2.87-3.7 with four threads. We incorporated the above into the
Chinese chess program CHIMO and won gold in Computer
Olympiad 2017 [23].
Our results also show that all versions including LOC2 perform much better. This justifies that, with the help of a 2-tuple
network, LOC2 is able to extract useful features with very little
expert knowledge. This result also shows the potential for locational relations among three or more pieces. However, add-
Comparison Training for Computer Chinese Chess
ing just one more piece increases the number of features dramatically. Note that the number of features for LOC2 is
119,960. This makes it difficult to complete training in a reasonable amount of time. We leave it as an open problem for
further investigation.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
Allis, L.V., Searching for Solutions in Games and Artificial Intelligence,
Ph.D. Thesis, University of Limburg, Maastricht, The Netherlands,
1994.
Baxter, J., Tridgell, A. and Weaver, L., Learning to Play Chess Using
Temporal Differences, Machine Learning 40(3), 243-263, 2000.
Beal, D.F., A Generalised Quiescence Search Algorithm, Artificial
Intelligence 43(1), 85-98, 1990.
Buro, M., Experiments with Multi-ProbCut and a New High-Quality
Evaluation Function for Othello, Technical Report 96, NEC Research
Institute, 1997.
Campbell, M., A. Joseph Hoane Jr., and Hsu, F.-H., Deep Blue, Artificial Intelligence 134(1-2), 47-83, 2002.
Chen, J.-C., Yen, S.-J., and Chen, T.-C., Shiga Wins Chinese Chess
Tournament, ICGA Journal 36(3), 173-174, 2013.
Chen, J.-C., Wu, I-C., Tseng, W.-J., Lin, B.-H., Chang, C.-H.,
Job-Level Alpha-Beta Search, IEEE Transactions on Computational
Intelligence and AI in Games 7(1), 28-38, 2015.
Chess Programming Wiki, Taper Eval, [Online], Available:
https://chessprogramming.wikispaces.com/Tapered+Eval
Collins, M., Discriminative Training Methods for Hidden Markov
Models: Theory and Experiments with Perceptron Algorithms, In
EMNLP’02, pp. 1-8, 2002.
Dong Ping Xiang Qi, A Chinese chess website (in Chinese) [Online].
Available: http://www.dpxq.com/.
Hsu, S.-C., Introduction to Computer Chess and Computer Chinese
Chess, Journal of Computer 2(2), 1-8, 1990. (in Chinese)
Iida, H., Sakuta, M., and Rollason, J., Computer Shogi, Artificial Intelligence 134, 121-144, 2002.
Kaneko, T. and Hoki, K., Analysis of Evaluation-Function Learning by
Comparison of Sibling Nodes, In Advances in Computer Games 13,
LNCS 7168, 158-169, 2012.
Knuth, D.E. and Moore, R.W., An Analysis of Alpha-Beta Pruning,
Artificial Intelligence 6(4), 293-326, 1975.
Lucas, S. M., Learning to Play Othello with N-tuple Systems, Australian Journal of Intelligent Information Processing 4, 1-20, 2007.
Silver, D. et al., Mastering the Game of Go with Deep Neural Networks
and Tree Search, Nature 529, 484-489, 2016.
Silver, D. et al., Mastering Chess and Shogi by Self-Play with a General
Reinforcement Learning Algorithm, arXiv:1712.01815, 2017
Szubert, M. and Jaśkowski, W., Temporal Difference Learning of
N-tuple Networks for the Game 2048, In 2014 IEEE Conference on
Computational Intelligence and Games (CIG), pp. 1-8, 2014.
Tesauro, G., Connectionist Learning of Expert Preferences by Comparison Training. Advances in Neural Information Processing Systems
1, 99-106, Morgan Kaufmann, 1989.
Tesauro, G., Neurogammon: a Neural Network Backgammon Program,
IJCNN Proceedings III, 33-39, 1990.
Tesauro, G., Comparison Training of Chess Evaluation Functions. In:
Machines that learn to play games, pp. 117-130, Nova Science Publishers, Inc., 2001.
Thill , M., Koch, P. and Konen, W., Reinforcement Learning with
N-tuples on the Game Connect-4, In Proceedings of the 12th International Conference on Parallel Problem Solving from Nature - Volume
Part I (PPSN’12), pp. 184-194, 2012.
Tseng, W.-J., Chen, J.-C. and Wu, I-C., Chimo Wins Chinese Chess
Tournament, ICGA Journal, to appear.
Ura, A., Miwa, M., Tsuruoka, Y., and Chikayama, T., Comparison
Training of Shogi Evaluation Functions with Self-Generated Training
Positions and Moves, CG 2013, 2013.
Ye, C. and Marsland, T.A., Selective Extensions in Game-Tree Search,
Heuristic Programming in Artificial Intelligence 3, pp. 112-122. Ellis
Horwood, Chichester, UK, 1992.
Ye, C. and Marsland, T.A., Experiments in Forward Pruning with
Limited Extensions, ICCA Journal 15(2), 55-66, 1992.
Yeh, K.-H., Wu, I-C., Hsueh, C.-H., Chang, C.-C., Liang, C.-C. and
6
[28]
Chiang, H., Multi-Stage Temporal Difference Learning for 2048-like
Games, IEEE Transactions on Computational Intelligence and AI in
Games, 9(4), 369-380, 2017.
Yen, S.-J., Chen, J.-C., Yang, T.-N. and Hsu, S.-C., Computer Chinese
Chess, ICGA Journal 27(1), 3-18, 2004.
Wen-Jie Tseng received his B.S. in Applied Mathematics from National Chung
Hsing University and M.S. in Computer
Science from National Chiao Tung University in 2006 and 2008, respectively. He
is currently a Ph.D. candidate in the Department of Computer Science, National
Chiao Tung University. His research interests include artificial intelligence and
computer games. He is the leader of the team developing the
Chinese chess program, named CHIMO, which won the gold
medal in Computer Olympiad 2017.
Jr-Chang Chen is an associate professor
of the Department of Computer Science
and Information Engineering at National
Taipei University. He received his B.S.,
M.S. and Ph.D. degrees in Computer Science and Information Engineering from
National Taiwan University in 1996, 1998,
and 2005 respectively. He served as the
Secretary General of Taiwanese Association for Artificial Intelligence in 2015 and 2016. Dr. Chen’s
research interests include artificial intelligence and computer
games. He is the co-author of the two Chinese chess programs
named ELP and CHIMO, and the Chinese dark chess program
named YAHARI, which have won gold and silver medals in the
Computer Olympiad tournaments.
I-Chen Wu (M’05-SM’15) is with the
Department of Computer Science, at National Chiao Tung University. He received
his B.S. in Electronic Engineering from
National Taiwan University (NTU), M.S.
in Computer Science from NTU, and Ph.D.
in Computer Science from Carnegie-Mellon University, in 1982, 1984 and
1993, respectively. He serves in the editorial boards of the IEEE Transactions on Computational Intelligence and AI in Games, ICGA Journal and Journal of Experimental & Theoretical Artificial Intelligence. He also served
as the presi-dent of the Taiwanese Association of Artificial
Intelligence in 2015 and 2016. His research interests include
artificial intelligence, machine learning, computer games, and
volunteer computing.
Dr. Wu introduced the new game, Connect6, a kind of
six-in-a-row game. Since then, Connect6 has become a
tour-nament item in Computer Olympiad. He led a team develop-ing various game playing programs, including CGI for
Go and CHIMO for Chinese chess, winning over 30 gold
medals in international tournaments, including Computer
Olympiad. He wrote over 120 papers, and served as chairs and
committee in over 30 academic conferences and organizations,
including the conference chair of IEEE CIG conference 2015.
Comparison Training for Computer Chinese Chess
Tinghan Wei received his B.AS. in Electrical Engineering from University of
British Columbia in 2008. He is currently a
Ph.D. candidate in the Department of
Computer Science, National Chiao Tung
University. His research interests include
artificial intelligence, machine learning
and computer games.
7
| 2cs.AI
|
Parallel Local Search: Experiments with a
PGAS-based programming model
Rui Machado1,2 , Salvador Abreu2 , and Daniel Diaz3
arXiv:1301.7699v2 [cs.PL] 10 May 2013
1
2
Fraunhofer ITWM, Kaiserslautern, Germany
[email protected]
Universidade de Évora and CENTRIA, Portugal
[email protected]
3
University of Paris 1-Sorbonne, France
[email protected]
Abstract. Local search is a successful approach for solving combinatorial optimization and constraint satisfaction problems. With the progressing move toward multi and many-core systems, GPUs and the quest
for Exascale systems, parallelism has become mainstream as the number
of cores continues to increase. New programming models are required
and need to be better understood as well as data structures and algorithms. Such is the case for local search algorithms when run on hundreds
or thousands of processing units. In this paper, we discuss some experiments we have been doing with Adaptive Search and present a new parallel version of it based on GPI, a recent API and programming model
for the development of scalable parallel applications. Our experiments
on different problems show interesting speedups and, more importantly,
a deeper interpretation of the parallelization of Local Search methods.
Keywords: Parallel Local Search, GPI, Adaptive Search, Constraint
Programming
1
Introduction
Systematic and complete search algorithms impose a limitation on the problem
size they are able to solve due to the exponential increase in processing time and
memory requirements. For this reason, heuristics-based search algorithms are
used (and necessary) for larger problem sizes. Instead of exploring the complete
search space, heuristics are used to guide the search to portions of the search
space where solutions might be found. Local Search and Meta-heuristics are an
interesting paradigm for combinatorial search and have been shown very effective
for solving real-life problems [9,8]. But despite the effectiveness of local search
methods, for really large problem instances, the running time required might still
be substantial. One way to cope with this problem is by introducing parallelism.
The current trend we are facing is an inevitable paradigm shift towards multicore technologies where parallelism is now omnipresent. In recent systems parallelism spreads over several systems levels and heterogeneity is growing on the
node as well as on the chip level. Data must be maintained across a hierarchy
of memory levels and most applications and algorithms are not yet ready to
take full advantage of available capabilities. There is a demand for programming
models with a flexible threads model and asynchronous communication to cope
with this gap.
PGAS (Partitioned Global Address Space) programming models have been
discussed as an alternative to MPI [12] for some time. The PGAS approach offers
the developer an abstract shared address space which simplifies the programming
task and at the same time facilitates data-locality, thread-based programming
and asynchronous communication. GPI is a PGAS API that follows this philosophy and delivers the full performance of RDMA-enabled4 networks directly to
the application without interrupting the CPU.
In this paper we aim at bringing together both the need for parallelism to
solve large problem instances with Local Search and its availability in current
systems. We implemented a new parallel version of the Adaptive Search algorithm based on GPI that goes beyond the simple independent multiple-walk.
Our new design shows interesting speedup gains on benchmarks with scalability
problems and more importantly, a deeper interpretation on the parallelization
of Adaptive Search in particular and Local Search methods in general, based on
some characteristics of the benchmarks.
The rest of the paper is organized as follows: in section 2 we present GPI
and its programming model, hightlighting some its major features. Section 3
provides some background on the Adaptive Search algorithm and section 4 focuses on its parallelization. In section 5, we detail our parallelization strategy
based on GPI and in section 6 we show the obtained results and compare it to
the previous implementation. Section 7 examines and interprets our experimental findings, correlating them with the characteristics of the problems. Finally,
section 8 presents a short conclusion and perspectives of future work.
2
GPI
GPI (Global address space Programming Interface) is a PGAS API for parallel
applications running on clusters. The thin communication layer delivers the full
performance of RDMA-enabled networks directly to the application without
interrupting the CPU. Fig. 1 depicts the architecture of GPI.
The local memory is the internal memory available only to the node and allocated through typical allocators (e.g. malloc). This memory cannot be accessed
by other nodes. The global memory is the partitioned global shared memory
available to other nodes and where data shared by all nodes should be placed.
The DMA interconnect connects all nodes and is through this interconnect that
GPI operations are issued. At the node level, the Manycore Threading Package
(MCTP) is used to take advantage of all cores present on the system and make
use of the GPI functionality and global memory. The MCTP was developed to
4
RDMA - Remote Direct Memory Access.
The Global Programming Interface (GPI)
Local Memory
Local Memory
Local Memory
Global Memory
Global Memory
Global Memory
DMA interconnect
MCTP threads
MCTP threads
MCTP threads
Node 1
Node 2
Node n
Fig. 1. GPI
help programmers take better advantage of new architectures and ease the development of multi-threaded applications. The MCTP is a threading package
based on thread pools that abstracts the native threads of the platform.
GPI is constituted by a pair of components: the GPI daemon and the GPI
library. The GPI daemon runs on all nodes of the cluster, waiting for requests to
start applications and the library holds the functionality available for a program
to use: read/write global data, passive communication, global atomic counters,
collective operations. The two components are described in more detail in our
previous contribution [10].5
The GPI core functionality can be summarized as follows:
–
–
–
–
–
–
–
read and write global data
passive communication
send and receive messages
commands
global atomic counters and spinlocks
barriers
collective operations
In the context of this work, the most important functionality is the read/write
of global data.
Two operations exist to read and write from global memory independent
of whether it is a local or remote location. One important point is that those
operations are one-sided that is, only the peer that issues such operation takes
part in it. This is different from a two-sided scheme (message passing) where the
5
GPI was previously known as Fraunhofer Virtual Machine (FVM).
peer that sends (sender ) has a corresponding peer (receiver ) that needs to issue
a receive operation. Moreover, this functionality is non-blocking and completely
off-loaded to the interconnect, allowing the program to continue its execution
and hence take better advantage of CPU cycles. The data movement does not
require any intermediate buffers and protocols to maintain those buffers. If the
application needs to make sure the data was transferred (read or write), it needs
to call a wait operation that blocks until the transfer is finished and asserting
that the data is usable.
3
Adaptive Search
Local Search is based on the simple idea of “searching” by iteratively moving
from one solution to one of its neighbours. The neighborhood of a solution is
the set of solutions that can be obtained by applying a move. A move is a local
change (hence the name Local Search).
The mechanism used to select a neighbour and thus the definition of what
constitutes a neighbourhood is the main issue that differentiates between different local search methods. In general, it is problem dependent and is related to
the definition of the objective function.
The Adaptive Search method [4] is one of many different local search methods
and has proved to be very efficient in the types of problems where it was tested.
It is a generic, domain-independent constraint-based local search method.
This meta-heuristic takes advantage of the structure of the problem in terms
of constraints and variables and can guide the search more precisely than a single
global cost function to optimize, such as for instance the number of violated
constraints. The algorithm also uses an short-term adaptive memory in the spirit
of Tabu Search in order to prevent stagnation in local minima and loops. This
method is generic, can be applied to a large class of constraints (e.g. linear and
non-linear arithmetic constraints, symbolic constraints, etc) and naturally copes
with over-constrained problems.
The input of the method is a problem in CSP format, that is, a set of variables
with their (finite) domains of possible values and a set of constraints over these
variables. For each constraint, an “error function” needs to be defined; it will
give, for each tuple of variable values, an indication of how much the constraint is
violated. For instance, the error function associated with an arithmetic constraint
|X − Y | < c, for a given constant c ≥ 0, can be max(0, |X − Y | − c).
Adaptive Search relies on iterative repair, based on variable and constraint
error information, seeking to reduce the error on the worst variable so far. The
basic idea is to compute the error function for each constraint, then combine for
each variable the errors of all constraints in which it appears, thereby projecting
constraint errors onto the relevant variables. Finally, the variable with the highest
error will be taken and its value will be modified. In this second step, the well
known min-conflict heuristic is used to select the value in the variable domain
which is the most promising, that is, the value for which the total error in the next
configuration is minimal. In order to prevent being trapped in local minima, the
Adaptive Search method also includes a short-term memory mechanism to store
variables to avoid (variables can be marked Tabu and “frozen” for a number
of iterations). It also integrates reset transitions to escape stagnation around
local minima. A (partial) reset consists in assigning fresh random values to some
variables (also randomly chosen). A reset is guided by the number of variables
being marked Tabu. As in any local search method, it is also possible to restart
from scratch when the number of iterations reaches a given limit.
4
Parallel Adaptive Search
When parallelizing an algorithm one aims at identifying hotspots and sources of
parallelism. As with most of meta-heuristics, in Adaptive Search these sources
of parallelism are essentially: (1) the inner loop of the algorithm i.e., computing
and combining the errors of variables and selecting the variable with highest
error and (2) the search space of the problem.
The problem with exploiting the inner loop of the algorithm is its granularity: it is
too fine-grained. The overhead associated with synchronization and dispatching
of tasks comes at a too high cost.
The other main source of parallelism is the search space (domain) of the
problem itself. Theoretically, this domain could be decomposed in several disjunct partitions to be explored in parallel and without dependencies. However,
in practice, several issues arise with this. Each partition is in general still too
large for a sequential execution and more importantly, not the whole search
space is equally valid and the exploration should avoid areas of it that lead to
poor solutions. Moreover, it is hard and expensive to control and maintain the
search conducted in the different partitions since a Local Search algorithm only
has a local view of the search space. One example is the class of problems that
have the best solutions clustered in a certain ’zone’ of the search space. In this
case, the algorithm should converge to that zone but in case of parallel execution
avoid too much redundant work.
The Adaptive Search method has already been subject to some research on its
parallel behaviour. Previous work on parallel implementations of the Adaptive
Search algorithm have mostly focused on independent multiple-walks, requiring
no commmunication neither shared memory between processing units.
In [5], the authors present a parallel implementation of the Adaptive Search
algorithm for the Cell/BE, a heterogenous multicore architecture. The system
includes 16 processors (the SPEs) where each one starts with a different random
initial solution. The PPE acts as the master processor, waiting for the message
of a found solution. For such number of processing units, the results were very
promising, achieving for some problems linear speed-up.
Further work with Parallel Adaptive Search continued to follow the same
approach with no communication between workers but more interestingly, concentrating on cluster systems with a larger number of cores.
In [2], the authors experiment and investigate the performance of a multiple
independent-walk on a system with up to 256 cores. The parallelization was done
with MPI and involves the introduction of a “communication step” which tests
if termination was detected (a solution was found) and terminates the execution
properly.
The presented performance results are relatively modest in terms of parallel
efficiency and still far away for the ideal speed-up which contrasts with the results
obtained at a smaller scale (ie. up to 16 cores) in previous work. This points out
the need for better alternative strategies in order to better exploit large-scale
parallelism.
Since that the independent multiple-walk approach still leaves space for improvement in terms of parallel efficiency and scalability for some problems, new
ways to take take full advantage of parallel systems must be found.
In [1], the authors experiment with more complex strategies, where processes
exchange messages resembling branch-and-bound methods where the bound is
exchanged between all participants. In their work, two alternatives are attempted:
exchanging the cost of the current solution of each process and the current cost
plus the number of iterations needed to achieve that cost. Unfortunately, both
approaches do not achieve better results than an independent multiple-walk.
5
Adaptive Search with GPI
Previous work with parallel Adaptive Search provides some groundwork to build
upon and has showed that some benchmarks exhibit scalability problems when
run on a large number of cores.
GPI seems, à priori, an interesting match to the problem of parallelization. Local
search methods work with local information, trying to progress and converge to
solutions in a global search space, requiring low global information. However, in
a parallel setting, communication and cooperation are crucial and in this case,
required to overcome the low parallel efficiency in some problems. The communication with GPI is based on one-sided primitives that might benefit the local
view on a global search space, allowing threads to cooperate asynchronously.
Moreover, communication is very efficient as GPI exploits the full performance
of the interconnect. Hence, we continue to explore ways to further improve the
parallelization of the Adaptive Search algorithm, exploiting GPI and its programming model, with the objective of getting some further benefits. But more
importantly, to find mechanisms, concepts or limitations that are general.
In general, we can define the following objectives:
– further investigate and understand the behavior of parallel Adaptive Search
on different problems.
– investigate the possibilites given by GPI and devise more complex mechanisms for the parallel execution of Adaptive Search, improving its performance
– identify the, possibly new, problems generated by the previous point.
The new parallel version of Adaptive Search based on GPI includes two variants which we name TDO (Termination Detection Only) and PoC (Propagation
of Configuration).
The TDO variant implements the simple independent multiple-walk and
serves mostly has our basis for comparison. First, with the existing MPI version, making sure that the implementation is correct and the performance is as
expected. Second, to allow us to measure the improvement (if present) obtained
with the more complex PoC variant. The PoC variant introduces more communication and sharing between working threads, by means of GPI primitives and
threaded model.
The next sections present the two different variants in more detail.
5.1
Termination Detection Only
The variant with Termination Detection Only (TDO) is rather straightforward
and implements the idea of an independent multiple-walk: all available cores
execute the sequential version of the Adaptive Search algorithm.
We name this variant as Termination Detection Only since it subsumes itself to a termination detection problem i.e., detecting the termination of a distributed computation. Termination Detection is itself a subject of much research
and several algorithms have been and continue to be proposed( [6,11,13]).
In the case of the Parallel Adaptive Search method, we are interested in
detecting termination as soon as one of the participating threads has found a solution, instead of waiting for all threads to finish as some of them can potentially
require too many steps in order to find a solution (it is enough to be trapped in
a ’zone’ of the search space with no possible solutions).
The implementation of this variant is simple as it only involves the implementation of a mechanism of triggering and detecting termination.
The GPI implementation follows a similar line of the previous work with
MPI. Whenever a thread finds a solution, it triggers termination by writing to
its peers that it has found a solution. Thus, the time of the parallel execution is
the time taken by this fastest thread.
Other threads must detect termination. This is only possible by introducing
a communication step inside the internal loop of the Adaptive Search algorithm.
This is required since there is no other way for a GPI instance to react on an
remote event (i.e., termination) other than with communication. In this communication step, a check for termination is done on a particular memory address
that is written on termination emmission as described above. The communication step introduces some overhead that needs to be minimized. Thus the
communication step is only executed every k iterations.
5.2
Propagation of configuration
The experiments in previous work and with the TDO variant have found that
the simple approach to parallelization, namely, the independent multiple-walk,
proves itself insufficient in obtaining parallel efficiency on some problems specially when experimenting with a large number of cores. Moreover, exchanging
some simple information such as the cost leads to no improvement.
Hence, we aim at communicating more and more meaningful information, introducing cooperation. By cooperation we mean mechanisms that allow threads
to share information about their state and thus benefit from the collective search.
Also, we want to exploit the potential and benefits of GPI and its programming
model (one-sided communication, no wait for communication, global access to
data, threaded model, etc.).
One of the most powerful aspects of Local Search is its simplicity. And due
to this simplicity, it is hard to extract what could be considered as meaningful information to be shared and communicated. One logical candidate not yet
tried is the whole current solution or configuration. Because the term solution
is sometimes misleading, we refer to the current solution as a configuration. The
final solution represents the solution when the algorithm stops.
The used implementation of the Adaptive Search method deals only with
permutation problems and thus, a configuration is the permutation vector of the
problems’ variables.
Similarly to other approaches to the parallelization of local search methods
which introduce cooperation, several important questions arise, namely:
1.
2.
3.
4.
Who does the communication?
When to do the communication?
How to do the communication?
What to communicate?
Answering most of these questions requires carrying out actual experiments since
the best and correct answer it is not, in our opinion, foreseeable.
Our approach, which we call Propagation of Configuration (PoC), aims at answering these questions and give a better understanding of how cooperation can
help with increasing the scalability of Local Search in general and the Adaptive
Search method in particular.
Who does the communication?
Answering the question of who does the communication involves deciding whether
a single thread or all threads actually perform communication. Note that by communication we mean that, in a distributed setting, messages between nodes are
exchanged. In a single node and given the GPI programming model, we can
benefit from the threaded-model and shared memory. Notwithstanding the best
option for this, it is clear that all threads must benefit from it.
There are potential advantages and disvantages with both options. If all
threads perform communication, any shared resources must be protected by a
mutual exclusion mechanism, which might suffer from high contention. Moreover,
when all threads perform communication a lot more pressure on the interconnect
follows, increasing the parallel overhead and with possibly a lot of redundant
communication happening (the same configuration being communicated several
times). But, on the other hand, there will be a rapid progress towards the best
promising neighborhood, intensifying the search. Of course, this can be positive
but can also become dangerous since most of threads might get trapped in a local
minimum or poor quality neighborhood. A good trade-off between intensification
and diversification must be achieved.
If a single master thread communicates, the effects are potentially the opposite: less intensification but also less contention, less pressure on the interconnect
and less redundant work.
Preliminary tests made clear that the best option is the one with a single
communicating thread since it reduces the parallel overhead. Plus, with GPI, all
threads in a single node benefit immediately from the results obtained by the
master thread without any exchange of messages.
When to do the communication?
The first possible answer to this question is to follow the same strategy as with
the Termination Detection Only variant: introduce a communication step and
perform communication every k iteration. The value of k is fundamental on
how well this option might perform. With a low value (e.g., k = 10), a strong
intensification of the search is achieved but with the danger that threads might
give up too soon on a promising neighborhood. With a high value of k, we avoid
that danger but less intensification will be achieved since less information will
be propagated.
The other option is to not interrupt the normal flow of the algorithm for
communication, letting the search progress normally and independently until a
local minimum is achieved. Only at this point the configuration is propagated
and possibly used. One danger however is if threads don’t hit local minima
that often, the propagation of configuration won’t progress and some threads
might never see an up-to-date configuration, achieving less intensification. A
solution to this problem is to still have communication every k iteration, where
threads simply keep the communication progressing but only use the propagated
information when they are in trouble i.e., hit a local minimum. However, this
option increases the overhead by adding the extra communication step in some
iterations.
In principle the second option might seem more promising as no disturbance
is caused when the algorithm is progressing positively. But the forementioned
danger that the propagation of configurations won’t progress can have the consequence that there won’t be a benefit from the communication scheme when
compared to the simple TDO variant. We performed some tests on a problem
with low number of local minima (Magic Square) and in fact, this is what happens.
Based on this reasoning, our chosen option to when to communicate is to
have a communication step. Moreover, we still need to detect termination thus
a communication step must be present, even if with a much lower influence in
terms of overhead. Our PoC variant combines termination detection and the
propagation of configurations in a single step that happens every k iterations
and we focus on finding an optimal value for k.
How to do the communication?
With this question, we consider a single alternative. Since we aim at large scale
executions, we need an efficient approach. Communication is done in a tree-based
topology, in which each node only communicates with its parent and children (if
any). Currently, a binary tree is used but this can be parametrized at initialization. At each communication step, the propagation of the configuration is done
either up (to parent) or down (to the children) the tree. This only happens if
a configuration was propagated from the children (in case of the up direction)
or from the parent (down direction). The propagation of the communication
behaves then like a wave, up and down the tree, with possibly different configurations being propagated at different points of the tree and contributing to some
diversification.
Communication is performed by using GPI one-sided primitives. A thread
posts a write operation and returns immediately to work. The configuration
to be propagated will be directly written to the memory of the remote node
asynchronously, without any acknowledgement of it and overlapped with the
algorithm’s computation. The remote node on the other hand, on its communication step, checks if a valid configuration was written to its memory, decides
how to act on it and propagates its decision further.
We consider this single alternative since it gives us a good balance between
intensification and diversification and because having a tree-based topology provides an efficient pattern to achieve communication scalability. The final objective is to have a communication step with low overhead and here GPI provides
us with mechanisms to do so.
What to communicate?
The Adaptive Search method (as many other Local Search methods) is very
simple and includes very few elements that can be communicated.
The proposed option already mentioned before, is to communicate a full
configuration. To this, we only add the cost of the configuration as it is the
metric to evaluate the configuration. Plus, computing the cost everytime we
communicate a configuration is a source of extra overhead specially if a problem
has a large number of variables.
Still, the question remains of which configuration to communicate. In our
design the best configuration i.e., the configuration with better cost is communicated. At a communication step, a thread decides to propagate its own current
configuration or the propagated configuration(s).
Communicating configurations can be of advantage because it includes implicitly more information about the state of the search since it, in a sense,
provides a semi-exact positioning within the whole search space. As the best
configurations are being propagated, other threads that are currently on poorer
neighborhoods might benefit from moving to the best ones. With the stochastic
behavior of Adaptive Search and enough diversification, the whole search procedure can be performed on the best neighborhoods and possibly, converge faster
to good solutions.
6
Experimental results
In this section we present the obtained results using different problems.
– all-interval: the All Interval Series problem (prob007 in CSPLib [7]),
– costas-array: the Costas Array problem,
– magic-square: the Magic Square problem (prob019 in CSPLib).
The experiments were conducted on a cluster system where each node includes
a dual Intel Xeon 5148LV (“Woodcrest”) (i.e., 4 CPUs per node) with 8 GB
of RAM. The full system is composed of 620 cores connected with Infiniband
(DDR). Since we aim at large scale, we performed our experiments on the system
using up to 256 cores on some problems and 512 cores on others. This difference
is due to the fact that the system is largely used and is hard to get access to the
full system.
Note that Adaptive Search, as many other Local Search methods, has a
stochastic behavior to achieve diversity on the search. To benchmark such behavior, several executions must be done and averaged. In our experiments we
ran each problem 100 times in order to obtain meaningful results.
We compare both GPI variants (TDO and PoC) with the MPI implementation from previous work, which serves as our basis for comparison.
Fig. 2 depicts the obtained results for the Costas Array problem (CAP) with
n=20.
Costas Array 20
256
GPI
MPI
Speed-up
128
64
32
16
16
32
64
No. of cores
128
256
Fig. 2. Costas Array (n=20) on 256 cores (64 nodes)
As already observed in previous work [3], the CAP shows an almost optimal
scalability using an independent multiple-walk with no cooperation. We can
observe that our implementation obtains similar, although slight better, results.
This is the expected result since both approaches (TDO and MPI) are equivalent
and a confirmation that our implementation performs as expected.
Although we aspired at obtaining even better results with the PoC variant
(possibly super linear) for this problem, our experiments showed that this variant
performs much worse than the simple TDO variant and thus we only present
the speedup obtained with GPI using the TDO variant.
The Fig. 3 depicts the obtained results for the Magic Square problem up to
512 cores.
Magic Square 200
100
GPI
MPI
PropConf (k=1000)
PropConf (k=10)
90
80
Speed-up
70
60
50
40
30
20
10
16
32
64
128
No. of cores
256
512
Fig. 3. Magic Square 200 on 512 cores (128 nodes)
For this problem we present the speedup obtained with the TDO and PoC
variants and compare it with the MPI version. The GPI TDO variant presents
again, as expected, results similar to the MPI version.
The Magic Squares benchmark is one of the problems that results in disappointing scalability when using the simple independent multiple-walk and therefore a major target for improvement with more sophisticated approaches. Indeed,
for this problem, our PoC variant improves the performance and scales better
as we increase the number of cores used.
We wanted to answer the question of when to do communication: as we mentioned, in our preliminary experiments it turned out that the best approach is
to have a communication step every k iterations where the value of k is decisive. Surprinsingly, for this benchmark, a lower value of k (k =10 in contrast to
k =1000) improves scalability by a factor of 2, achieving a speedup of 97 with
512 cores. Still a low parallel efficiency but a large improvement over the other
options and variants.
The obtained results for the last problem, the All Interval series (n=400), is
shown in Fig. 4.
The All Interval Series benchmark is also one of the problems where good
scalability was hard to reach when using a large number of cores. In Fig. 4 it is
possible to observe this fact, where both the MPI and GPI TDO versions reach
a modest speedup factor of 20 and 25, respectively (with 256 cores). Our PoC
All Interval
30
GPI
MPI
PropConf(k=1000)
25
Speed-up
20
15
10
5
0
16
32
64
No. of cores
128
256
Fig. 4. All Interval 400 on 256 cores (64 nodes)
variant however, performs much worse than the TDO variant at a low number
of cores but it improves as we increase the number of cores, hinting that this
variant can be of advantage if we increase the number of cores and the problem
size. In Fig. 4 we only depict the obtained results for the PoC variant with k
= 1000 since, for this benchmark, it is the best value. In contrast to the Magic
Squares benchmark, a lower value of k results in a much worse performance.
7
Discussion
The experimental results presented large differences in how the different problems benefit from parallelism and the implemented variants. One of our main
objectives is to investigate and understand why this happens.
In order to be able to draw some conclusions on our experiments, it is important to characterize the chosen problems from different perspectives. We characterize the problems using different information such as the number of iterations
and local minima. This characterization will give us a basis to better understand
the problems at hand and possibly explain our results.
The Table 1 presents the obtained values for acquired information when
running some instances of the previously presented problems. This information
is the following:
Problem The problem instance.
Iterations The number of iterations required to find a solution.
Local Minima The number of local minima found.
Resets The number of partial resets performed (not full restart).
Same var / Iteration The number of times that existed more than one candidate variable (highest error value) to be selected.
Problem
Iterations Local Minima
Magic Square 200 413900.505
Resets
Same var/Iteration
25864.75
3.01
23.36
204024.89
Costas 18
389932.263
204024.89
1.00
Costas 19
3364807.772 1714299.50 1714299.50
0.99
All Interval 200
11229.220
495.27
495.27
5.97
All Interval 400 41122.406
1422.15
1422.15
9.19
Table 1. Information collected for different problems instances.
This information allows us to better understand how does the Adaptive Search
algorithm progress towards a solution, the neighborhood structure and extract
further information (e.g., number of local minima per iteration).
From Table 1 we can see that the different problems exhibit a quite different
behavior. The Magic Square problem performs a low number of partial resets
when compared to the total number of iterations or to the number of identified
local minima. On the other hand, it is the problem where the number of candidate variables per iterations (Same var/Iteration) is high, meaning that at each
iteration there are several possible moves towards the next configuration.
The Costas Array problem exhibits a completely different behavior. In this
problem, the number of local minima identified is very large (almost every second
iteration finds a local minimum) and the number of partial resets is very high,
coincident with the number of local minima i.e., at each local minimum found,
a partial reset is performed. Also the number of possible moves at each iteration
is close to 1.
The All Interval problem is yet another kind problem. Here, the number of
resets is as with the CAP equal to the number of local minima but these happen
much less often. The number of possible variable choices or moves is higher than
1, meaning that some diversification could be achieved.
If we relate this characterization of problems with the obtained experimental
results, some conclusions can be conjectured in order to better understand the
parallelization of such algorithm or, more concretely, how much can it benefit
from a communication scheme such as the one we designed.
We argue that one critical aspect is the neighborhood of a configuration or the
set of possible moves, which define transitions between configurations. Since we
are propagating configurations we can look at our problems at hand according to
this aspect. If a problem has a dense neighborhood or, in other words, the set of
possible moves at each transition is (much) larger than one, each of these moves
can be explored in parallel. Thus, when a promising configuration is propagated
and several moves are possible and explored in parallel, the probability that one
of these moves leads to a faster path towards an optimal solution increases.
Another important aspect is the number of local minima and resets and how
both relate. A problem that finds a large number of local minima before encoun-
tering an optimal solution benefits less from processing a configuration which
seems promising. This configuration is heuristically promising but in reality this
information is less meaningful than it should. Similarly, a problem with a high
number of partial resets suffers from the same problem.
Looking back at our experimental results with the different problems, we can
better understand a) the difference in scalability and b) the improvement factor
brought by the PoC variant to some problems.
In the Magic Square problem, each configuration has a dense neighborhood
and benefits from the parallel exploration of different moves. Thus, the PoC variant improves the performance and scalability of the algorithm. When a working thread adopts a propagated configuration, it will define its own path from
that configuration and differently from one other thread that receives that same
promising configuration. Moreover, this problem has a low number of local minima and resets meaning that paths from one (initial) configuration towards an
optimal solution are a series of transitions from neighbor configurations.
The Costas Array Problem exhibits optimal scalability with the independent
multiple-walk MPI version or with our TDO variant and this is already per
se satisfactory. On the other hand, it performs worse with the PoC variant:
propagating a configuration is only a source of parallel overhead and will limit
the search allowing less diversification. A propagated configuration will allow,
on average, a single move and two threads taking the same configuration results
in redundant work which is also probably unfruitful since the CAP is one of
the problems with a high number of local minima and reset. This also explains
the good scalability using the TDO variant, where increasing the number of
cores allows covering more of the total search space together with the fact that
solutions for this problem are well spread over it.
Finally, the All Interval Series problem shows a mixed behavior. Similarly
to the CAP, the larger number of local minima found and same number of resets point to the same problem. There is less benefit from taking a propagated
configuration since its meaningfulness is low. The PoC variant only introduces
unnecessary overhead and this could explain the much worse performance at a
lower number of cores. On the other hand, and similarly to the Magic Square
benchmark, there is more than one possible move, on average i.e., some diversification can be achieved. With a large enough number of cores, the parallel
overhead can be amortized by the gain obtained with this diversification. This
could be the reason for the steeper curve for the PoC variant on Fig. 4. Of course,
with further experiments we will be able to understand this better.
In summary, problems where configurations have a denser neighborhood benefit from a cooperation scheme such as the PoC variant where the full configuration is communicated. Contrarily, problems that follow a trajectory with a single
move possible won’t benefit from a communication scheme that propagates the
best current configuration. Also, if a large number of local minima is found and
partial resets are required in the same number, the expectation for improvement
in performance is zero.
8
Conclusion
In this paper we presented our work on the parallel implementation of the Adaptive Search method using a different programming model. GPI is an API designed
for high-performance and scalable parallel applications. We aimed at investigating and understanding the behavior of Adaptive Search in a parallel setting,
focusing on different problems particularly those that, in previous work, showed
scalability problems when targeting a large number of cores. GPI and its programming model allowed us to design a new communication and parallelization
scheme which in our experimental evalution allowed a gain of a factor of 2 in
terms of speedup for some problems. More importantly, it provided deeper insight and understanding on the parallelization of Local Search methods given
different problems with disparate characteristics such as the neighborhood of a
configuration, the number of local minima and partial resets.
In the future, we intend to examine our design and conclusions with other
larger problems and experiment with complexer parallelization schemes. One
possible direction is instead of using promising information (configurations, cost,
statistics) directly, act on the complement of it, avoiding redundant work and
cover as much as possible from the search space since this is the main source of
parallelism.
One of our potential final goals is the design of a new Local Search algorithm
more amenable to parallelization that builds upon these experiences.
Acknowledgements This work was partly supported by Fundação para a
Ciência e Tecnologia under grant PTDC/EIA-EIA/100897/2008 (HORUS).
References
1. Yves Caniou and Philippe Codognet. Communication in parallel algorithms for
constraint-based local search. In IPDPS Workshops, pages 1961–1970, 2011.
2. Yves Caniou, Philippe Codognet, Daniel Diaz, and Salvador Abreu. Experiments
in parallel constraint-based local search. In Peter Merz and Jin-Kao Hao, editors, EvoCOP, volume 6622 of Lecture Notes in Computer Science, pages 96–107.
Springer, 2011.
3. Yves Caniou, Daniel Diaz, Florian Richoux, Philippe Codognet, and Salvador
Abreu. Performance analysis of parallel constraint-based local search. In Proceedings of the 17th ACM SIGPLAN symposium on Principles and Practice of
Parallel Programming, PPoPP ’12, pages 337–338, New York, NY, USA, 2012.
ACM.
4. P. Codognet and D. Diaz. Yet another local search method for constraint solving.
Stochastic Algorithms: Foundations and Applications, pages 342–344, 2001.
5. Daniel Diaz, Salvador Abreu, and Philippe Codognet. Targeting the cell broadband
engine for constraint-based local search. Concurrency and Computation: Practice
and Experience, 24(6):647–660, 2012.
6. Edsger W. Dijkstra, W. H. J. Feijen, and A. J. M. van Gasteren. Derivation of a
termination detection algorithm for distributed computations. Inf. Process. Lett.,
16(5):217–219, 1983.
7. Ian P. Gent and Toby Walsh. Csplib: A benchmark library for constraints. In CP,
pages 480–481, 1999. http://www.csplib.org.
8. T. Gonzalez, editor. Handbook of Approximation Algorithms and Metaheuristics.
Chapman and Hall / CRC, 2007.
9. T. Ibaraki, K. Nonobe, and M. Yagiura, editors. Metaheuristics: Progress as Real
Problem Solvers. Springer Verlag, 2005.
10. Rui Machado and Carsten Lojewski. The fraunhofer virtual machine: a communication library and runtime system based on the rdma model. In Computer
Science-Research and Development, volume 23(3), pages 125–132, 2009.
11. Friedemann Mattern. Algorithms for distributed termination detection. Distributed
Computing, 2:161–175, 1987. 10.1007/BF01782776.
12. MPI Forum. MPI: A Message-Passing Interface Standard. Version 2.2, September
4th 2009. available at: http://www.mpi-forum.org (Dec. 2009).
13. Vijay A. Saraswat, Prabhanjan Kambadur, Sreedhar B. Kodali, David Grove, and
Sriram Krishnamoorthy. Lifeline-based global load balancing. In PPOPP, pages
201–212, 2011.
| 6cs.PL
|
A Bernoulli-Gaussian Physical Watermark for Detecting Integrity
Attacks in Control Systems
arXiv:1710.01105v2 [cs.SY] 8 Oct 2017
Sean Weerakkody
Omur Ozel
Abstract— We examine the merit of Bernoulli packet drops in
actively detecting integrity attacks on control systems. The aim
is to detect an adversary who delivers fake sensor measurements
to a system operator in order to conceal their effect on the
plant. Physical watermarks, or noisy additive Gaussian inputs,
have been previously used to detect several classes of integrity
attacks in control systems. In this paper, we consider the
analysis and design of Gaussian physical watermarks in the
presence of packet drops at the control input. On one hand,
this enables analysis in a more general network setting. On
the other hand, we observe that in certain cases, Bernoulli
packet drops can improve detection performance relative to a
purely Gaussian watermark. This motivates the joint design
of a Bernoulli-Gaussian watermark which incorporates both
an additive Gaussian input and a Bernoulli drop process. We
characterize the effect of such a watermark on system performance as well as attack detectability in two separate design
scenarios. Here, we consider a correlation detector for attack
recognition. We then propose efficiently solvable optimization
problems to intelligently select parameters of the Gaussian
input and the Bernoulli drop process while addressing security
and performance trade-offs. Finally, we provide numerical
results which illustrate that a watermark with packet drops
can indeed outperform a Gaussian watermark.
I. I NTRODUCTION
The security of cyber-physical systems (CPS) has become
a critical issue [1]. Since CPS such as the smart grid, waste
management systems, water distribution systems, transportation systems, and smart buildings are linked to critical
infrastructures, it is imperative that they operate securely. Unfortunately, attacks have occurred against CPS. This includes
Stuxnet [2], which targeted uranium enrichment facilities
in Iran, the Maroochy Shire incident [3], an attack by a
malicious insider on a sewage management system, and the
Ukraine power attack [4], a hack resulting in widespread
blackouts in Ukraine. The threat does not appear to be over
as the growing connectivity and heterogeneity of our system
architectures provide new attack surfaces for adversaries.
We focus on detecting integrity attacks in control systems
in the presence of packet drops at the control input. In an
integrity attack, an adversary modifies inputs and sensor
measurements in a control system. The goal of such an
S Weerakkody, O. Ozel, and B. Sinopoli are with the Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA, USA 15213. Email: {sweerakk,
oozel}@andrew.cmu.edu, [email protected]
S. Weerakkody is supported in part by the Department of Defense (DoD)
through the National Defense Science & Engineering Graduate Fellowship
(NDSEG) Program. The work by S. Weerakkody, O. Ozel, and B. Sinopoli
is supported in part by the Department of Energy under Award Number
de-oe0000779 and by the National Science Foundation under award number
CCF 1646526.
Bruno Sinopoli
attacker may be to achieve some economic benefit or cause
physical damage to a system. An attacker can potentially
maximize his impact by hiding his presence from the operator. Remaining stealthy allows an attacker to affect the system for long periods of time without defender interference.
The adversary can avoid detection by intelligently modifying
the sensor measurements to fool detectors. One example
is a replay attack, as used in Stuxnet, where an attacker
replaces true outputs with a previously recorded sequence
of measurements [5].
We consider the use of physical watermarking to detect
integrity attacks. A physical watermark is a noisy Gaussian signal added on top of an optimal control input to
authenticate a system’s dynamics. Physical watermarking is
a method of active detection, where a defender alters his
strategy to recognize attacks. These methods are necessary
when standard fault detection methods provably fail [6], [7].
Recent work has investigated physical watermarking. In [5],
[8]–[10], the design of watermarks against replay attacks
was examined. Additionally, [11] and [12] design asymptotic
detectors in systems implementing physical watermarking
to ensure zero additive distortion power is introduced into
sensor measurements. Additionally, in a scalar setting, [13]
demonstrates the optimality of Gaussian watermarks against
Gaussian attackers and vice versa. [14] evaluates the use of
non-stationary watermarks to hamper system identification.
Finally, [15] considers watermarks to thwart adversaries who
have access to a subset of inputs and model knowledge.
However, prior work fails to consider the scenario where
there exists packet drops in the network. In this paper, we
generalize the design of the Gaussian physical watermark
by incorporating Bernoulli drops at the control inputs. This
enables the operator to account for imperfect networks when
designing a Gaussian watermark for secure detection. We
also argue that using Bernoulli drops together with a Gaussian watermark can improve detection. This motivates the
analysis and design of a joint Bernoulli-Gaussian watermark.
In our preliminary work [16], we proposed using packet
drops in a setting without Gaussian watermarks to detect
replay attacks. This article extends these results by providing
a rigorous mathematical setting to jointly design parameters
of both a Gaussian watermark and Bernoulli drop process.
We investigate two types of watermark design: 1) a watermark with an independent and identically distributed (IID)
Gaussian additive input multiplied by a Markovian Bernoulli
drop process at the control input and 2) a watermark with
a stationary Gaussian additive input generated by a hidden
Markov model (HMM) multiplied by an IID Bernoulli drop
process at the control input. We incorporate a correlation
detector [17], [8] to recognize integrity attacks and characterize adversarial scenarios where the Bernoulli-Gaussian
watermark is provably effective. Next, we provide efficiently
solvable optimization problems to design parameters of the
Gaussian input and the Bernoulli drop process. Simulation
results illustrate scenarios where packet drops improve detection performance relative to a purely Gaussian watermark.
Plant(A,B)
Sensors ( C )
uk,c
yk
m|b
uk
LQG
x̂k
Estimator
Controller
∆uk
II. S YSTEM M ODEL
Gaussian Input
We consider a discrete time LTI control system as follows
ηk
xk+1 = Axk + Buk,c + wk , yk = Cxk + vk .
(1)
xk ∈ Rn is the state vector at time k. A set of m sensor
measurements yk ∈ Rm is delivered to a supervisory control
and data acquisition (SCADA) system at time k in order to
perform remote estimation and compute an intended control
input uk ∈ Rp . A set of p control inputs uk,c ∈ Rp actuate
the system. We differentiate uk,c , the control input applied
to the system, versus uk , the input computed by a SCADA
operator. We assume wk ∼ N (0, Q) is IID process noise
and vk ∼ N (0, R) is IID measurement noise (independent
of {wk }), where Q ≻ 0, R ≻ 0. A Kalman filter performs
state estimation as follows.
x̂k+1|k = Ax̂k|k + Buk,c , x̂k|k = x̂k|k−1 + Kzk ,
T
T
−1
K = P C (CP C + R)
, zk = yk − C x̂k|k−1 ,
P = AP AT + Q − AP C T (CP C T + R)−1 CP AT .
(2)
(3)
(4)
The defender minimizes a cost function J:
" N
#
X
1
T
T
E
xk W xk + uk,c U uk,c , (5)
J = lim
N →∞ 2N + 1
k=−N
1
where W ≻ 0 and U ≻ 0. We assume (A, B) and (A, Q 2 )
1
are controllable and (A, C) and (A, W 2 ) are observable.
A. Control in Uncertain Networks
As shown in Fig. 1, the control input uk may be dropped
as it is sent from the SCADA system to the plant. Here,
uk,c = ηk uk ,
(6)
where ηk ∈ {0, 1} is a Bernoulli random variable. The
control input uk may be dropped due to network imperfections. In this case, we assume the operator receives an acknowledgement (ACK), which specifies if uk was delivered.
Alternatively, the input uk may be intentionally dropped as
a means to watermark the system, enabling the detection of
integrity attacks that fail to preserve the effect of the drop
process. This strategy was initially investigated in [16]. We
consider both IID and Markovian drop processes.
1) IID Bernoulli Process: First, we assume {ηk } is
an IID Bernoulli process where P (ηk = 1) = 1 −
pd . LQG control with IID Bernoulli packet losses was
studied in [18]. Consider the information set Fk ,
{y−∞:k , η−∞:k−1 , u−∞:k−1 }. We suppose pd is chosen (or
given) so that the system (1) can have finite cost J. The
Failure Detector
Packet Drop
Sequence
Fig. 1.
System model
optimal control strategy at time k given Fk is as follows
[16], [18]:
ubk = Lk x̂k|k , Lk = −(B T Sk+1 B + U )−1 B T Sk+1 A,
Sk = AT Sk+1 A + W + (1 − pd )AT Sk+1 BLk .
(7)
As we expect the system has been running for a long time,
both Lk and Sk have converged to fixed point values so that
ubk = L(b) x̂k|k , L(b) = −(B T S(b) B + U )−1 B T S(b) A,
S(b) = AT S(b) A + W + (1 − pd )AT S(b) BL(b) .
(8)
J = J(b) for this strategy where J(b) is
J(b) = tr S(b) Q + (AT S(b) A + W − S(b) )(P − KCP ) .
(9)
2) Markovian Bernoulli Process: In this setup, we assume
there are Markovian packet losses [19] at the input where
P (ηk+1 = 0|ηk = 0)P (ηk+1 = 1|ηk = 0)
ᾱ α
=
P (ηk+1 = 0|ηk = 1)P (ηk+1 = 1|ηk = 1)
β β̄
(10)
and ᾱ , 1 − α, β̄ , 1 − β. Here, we assume 0 < α ≤ 1,
0 < β ≤ 1 so that ηk is irreducible. Moreover, we assume
ηk is stationary, which can be obtained by letting its initial
β
. Finally, we assume that
distribution be P (η−∞ = 0) = α+β
α and β are selected (or given) so that the system (1) can
have finite cost J. The optimal strategy at time k given Fk
is
T
−1 T
um
B R(m) A,
k = L(m) x̂k|k , L(m) = −(B R(m) B + U )
Rm = AT (βS(m) + β̄R(m) )A + W + β̄AT R(m) BL(m) ,
Sm = AT (ᾱS(m) + αR(m) )A + W + αAT R(m) BL(m) ,
where L(m) , R(m) , S(m) are parameters which converged to
their steady state values. The resulting cost of control is
tr(βS(m) Q + αR(m) Q)
(11)
α+β
tr((AT (ᾱS(m) + αR(m) )A + W − S(m) )(P − KCP ))
.
+
α+β
J = J(m) =
Remark 1: The prior strategies are optimal when the defender only has knowledge of the observed drop sequence
η−∞:k−1 . However, if the drop sequence is intentionally introduced using a pseudo random number generator (PRNG),
the defender knows future values of ηk . The design of a
controller that uses this information is left for future work.
B. Joint Bernoulli-Gaussian Physical Watermarking
To account for adversarial behavior, we consider additive
Gaussian physical watermarks ∆uk . First introduced in [5]
to detect replay attacks, an additive Gaussian watermark can
be leveraged to verify the freshness of outputs. We aim to
intelligently combine the Gaussian watermarks considered
in [8] and [9] with a Bernoulli drop process at the input.
Such a design accomplishes two goals: 1) to expand the
analysis of physical watermarking to a more realistic network
setting with packet drops and 2) to potentially improve
performance by considering a more general joint BernoulliGaussian watermark.
We consider two main joint designs.
Watermark 1: IID Gaussian Input + Markovian Drops
uk,c = ηk (um
k + ∆uk ).
(12)
{ηk } is a Markovian Bernoulli process and ∆uk ∼ N (0, Q)
is an IID Gaussian watermark [5]. We assume ∆uk is
independent of other stochastic processes in the system.
Watermark 2: Stationary Gaussian Input + IID Drops
uk,c = ηk (ubk + ∆uk ).
(13)
In this case, {ηk } is an IID Bernoulli process. The Gaussian
input ∆uk is assumed to be a stationary process generated
by a hidden Markov model (HMM) as considered in [9].
ζk+1 = Aω ζk + ψk ,
∆uk = Ch ζk .
(14)
ζk is the hidden state of the HMM, Aω has spectral radius
ρ(Aω ) ≤ ρ̄ ≤ 1, and ψk ∼ N (0, Ψ) is IID Gaussian noise.
For stationarity, Cov(ζ0 ) = Aω Cov(ζ0 )ATω + Ψ. ∆uk is
independent of other stochastic processes in the system.
Remark 2: Here, ρ̄, the maximum allowable spectral radius, is a design parameter for the defender. We observe a
larger ρ̄ improves expected detection performance. However,
a larger ρ̄ means a larger correlation between watermarks and
this could facilitate the prediction of future watermarks if the
attacker guesses an initial Gaussian input ∆uk .
III. ATTACK M ODEL
In this section we describe a model of our adversary in
terms of knowledge, capabilities, and potential strategies.
A. Attacker Capabilities
Without loss of generality, we assume an attack begins at
time k = 0. We make the following assumptions.
1) The attacker can modify all measurements yk , k ≥ 0.
The falsified outputs at time k are denoted by ykv .
2) The attacker inserts an input B a uak into the system.
3) The attacker is unable to read the true control inputs
uk,c . As a result, he is unaware of the drop sequence
{ηk } and the Gaussian watermark {∆uk }.
The system under attack is given by
xk+1 = Axk + Buk,c + B a uak + wk ,
v
x̂k+1|k+1 = (I − KC)(Ax̂k|k + Buk,c ) + Kyk+1
.
(15)
(16)
Remark 3: Attackers can inject B a uak by appropriating
the defender’s actuators or inserting their own. The attacker
could possibly modify inputs without being able to read them
if the inputs are encrypted. Alternatively, the attacker can
cause damage even if B a uak = 0. For example, the attacker
can destabilize the plant if A is open loop unstable.
B. Attack Strategy
The attacker generates ykv through a virtual system:
xvk+1 = Axvk + ηkv B(Lm|b x̂vk|k + ∆uvk ) + wkv ,
x̂vk+1|k+1
= (I − KC)(A + ηkv BLm|b )x̂vk|k
+ ηkv (I − KC)B∆uvk ,
ykv = Cxvk + vkv .
+
v
Kyk+1
(17)
(18)
(19)
In the case of Watermark 1, Lm|b = L(m) , ηkv follows
a Markovian process (10) with parameters α and β and
∆uvk ∼ N (0, Q) is an IID Gaussian process. In the case of
Watermark 2, Lm|b = L(b) , ηkv is an IID Bernoulli process
with drop probability pd and ∆uvk is a stationary Gaussian
process which satisfies (14). Additionally, vkv ∼ N (0, R)
and wkv ∼ N (0, Q) are IID processes. Finally, we assume
the stochastic processes {ηkv , ∆uvk , wkv , vkv } are independent
of the real system’s stochastic parameters {ηk , ∆uk , wk , vk }.
The previous attack strategy can be generated (approximately) by a replay attack where the attacker records a
long sequence of outputs y−T ′ :−T ′ +T and, starting at time 0,
replaces yk with ykv = yk−T ′ for 0 ≤ k ≤ T . Attackers who
do not have precise knowledge of the model may engage in
replay attacks, which only require access to the outputs [5],
[8], [9]. Alternatively, this attack strategy can be constructed
by an adversary who is familiar with the model, for instance a
malicious insider. In this case, the attacker simulates a virtual
copy of the system dynamics to fool a bad data detector. It
was previously shown [9] that if pd = 0 and there is no
Gaussian watermark, the given strategies are asymptotically
stealthy when A , (A + BL(b) )(I − KC) is Schur stable.
A model aware attacker could also potentially pursue an
additive attack, for instance a false data injection attack [20]
or a zero dynamics attack [21], [22]. In these attacks, the
adversary injects an additive bias into the system which
preserves the watermark and allows the attacker to remain
stealthy. However, there are scenarios where additive attacks
on sensor measurements are not feasible. As an example,
suppose the defender uses public key cryptography, where
a public key is used to encrypt the measurements while a
private key is used to decrypt the associated cipher text. An
attacker could send his own virtual measurements encrypted
with the public key. However, such an attack could not
leverage information in the true measurement as that would
require access to the defender’s private key to learn yk . In this
case, additive attacks constructed by replacing a true output
packet with a virtual packet would be infeasible. By assumption, an additive networked-based attack on the defender’s
control input is also impossible because the adversary is
unable to read the defender’s input.
We argue that alternative attack strategies which manipulate all sensors yk in a setting with public key cryptography
also fail due to the fact that the resulting attack sequence
{ykv } is independent of the watermarks {∆uk , ηk }. Specifically, an attacker who is unable to read the inputs or outputs
will have no information about the watermarks. As a result,
the outputs he can construct will fail to fool the correlation
detector, which we propose in the next section.
IV. A C ORRELATION D ETECTOR
We consider a correlation detector, proposed in [8]. The
defender computes a virtual output yk′ , which explicitly
characterizes the effect of watermarks on yk .
x′k+1 = Ax′k + ηk B(Lm|b x̂′k|k + ∆uk ),
x̂′k+1|k+1
= (I − KC)(A +
yk′ = Cx′k , (20)
ηk BLm|b )x̂′k|k
+
′
Kyk+1
(21)
µ. Note that κ corresponds to the time index of the event
triggered correlation detector and increases at instants when a
new detection statistic is computed. Identifying attacks on an
individual sensor i can be done by focusing on the correlation
between individual measurements. An appropriate statistic gκi
′
would be ytiκ ytiκ where ytiκ is the ith entry of ytκ .
Remark 4: A detector with an adaptive threshold could
address issues of small yk′ . However, such a detector is more
prone to misses, mistaking an attack for noise. Incorporation
and analysis of such a detector is left for future work.
Remark 5: An adversary that can not read {uk }, {yk } can
not take advantage of instances when detection does not
occur, because such instances are entirely dependent on the
realization of previous watermarks. An attacker who is forced
to act independently of the real time watermarking sequence
cannot determine if a detection has been triggered.
We now verify that the expected correlation is 0, if the
outputs ykv are generated independently of the watermarks.
Theorem 6: If ykv and {∆uk , ηk } are independent, then
h
i
E ykv T yk′ kyk′ k22 ≥ µ = 0.
Proof: Observe that yk′ can be written as a linear
function of the Gaussian watermarks ∆uk so that
+ ηk (I − KC)B∆uk ,
yk′ =
where with some abuse of notation x′−∞ = 0, x̂′−∞|−∞ = 0.
We can simplify (20) and (21) to obtain
x′k+1 = (A + ηk BLm|b )x′k + ηk B∆uk ,
yk′ = Cx′k . (22)
This virtual process created by the defender is driven entirely by the sequence of Bernoulli-Gaussian watermarks
{∆uk , ηk }. Thus, if we were to multiply the true outputs
yk with the defender’s virtual outputs yk′ we would expect
a positive correlation. However, if an attacker introduces
measurements ykv , which are driven by an independent sequence of watermarks, the expected correlation drops to 0.
This motivates consideration of the detection statistic ykT yk′ ,
where a large statistic is indicative of normal behavior while
a small statistic indicates malicious behavior. Observe due to
the random real time selection of watermarks, kyk′ k2 may be
close to 0, impacting detector performance since the correlation will likely also approach 0 even under normal operation.
As a result, we propose an event triggered detector:
If kyk′ k22 ≥ µ
Perform Detection
κ = κ + 1, tκ = k
κ
X
H0
gj ≷ τ, gκ = ytTκ yt′ κ .
j=κ−W+1
(23)
H1
The null hypothesis H0 is that the system is operating
without malicious behavior while the alternative hypothesis
H1 is that the system is under attack. W is the size of the
detector’s window. A detection event is triggered if kyk′ k22 is
greater than some user defined threshold µ, preventing false
alarms from being raised when yk′ is small, while sacrificing
time to detection. This tradeoff can be addressed by tuning
k−1
X
Gj (ηj:k−1 )∆uj ,
(24)
j=−∞
where Gj is some linear gain, determined by the sequence
of Bernoulli drops ηj:k−1 . Thus, we have
k−1
X
T
T
Gj (ηj:k−1 )∆uj kyk′ k22 ≥ µ
E[ykv yk′ ] = E ykv
j=−∞
=
k−1
X
j=−∞
h
i
T
E [ykv ] E Gj (ηj:k−1 )∆uj kyk′ k22 ≥ µ = 0.
The proposed detector can often differentiate between
faulty and malicious scenarios. During a fault, we expect
to see the effect of the embedded watermarks in the output
and it could be measured through correlation. Alternatively,
residue based detectors such as a χ2 detector (gκ =
−ztTκ (CP C T + R)−1 ztκ ), which measures the difference
between measured and expected behavior, will likely raise an
alarm during faulty behavior and malicious behavior. Both
detectors can be used in tandem. A residue based detector
can raise alarms in the case of faulty or malicious behavior,
while a correlation detector can distinguish these events. In
this article, we focus on the correlation detector.
V. T HE F IRST WATERMARK D ESIGN
We consider the design of a watermark consisting of an
IID Gaussian input and Markovian drops. This requires the
evaluation of a detection and performance trade-off. We wish
to maximize the correlation of yk and yk′ to distinguish the
system under attack from normal operation. However, we
also need to ensure the system meets an adequate level of
¯ starting at k = 0.
performance as characterized by the cost J,
"N −1
#
X
1
T
T
¯
E
(25)
xk W xk + uk,c U uk,c
J = lim
N →∞ N
k=0
As such, we design the parameters α, β, Q by solving the
following optimization problem
maximize
α,β,Q
lim E[ykT yk′ |H0 ]
k→∞
(26)
subject to J¯ ≤ δ, 0 < α, β ≤ 1.
To begin with, we use [19, Theorem 3] to analytically
compute the cost J¯ as follows.
Theorem 7: Suppose α and β are chosen so that the
system has finite cost J(m) in the absence of a Gaussian
watermark. The LQG cost J¯ of the control system (1) with
IID Gaussian and Markovian watermark (12) is:
α
J¯ = J(m) (α, β) +
tr (B T R(m) B + U )Q . (27)
α+β
Proof: Consider
the
in a finite
hori
T cost to go
PN
T
zon, Vk (xk ) ,
j=k E xj W xj + uj,c U uj,c |Fk , and let
uN,c = 0. Similar to, [19], it can be shown that
(
E[xTk Sk xk |Fk ] + ck (ηk−1 = 0)
Vk (xk ) =
,
(28)
E[xTk Rk xk |Fk ] + dk (ηk−1 = 1)
where cN = dN = 0, RN , SN = W, P̄ = P − KCP and
F = (A + BL(m) ),
Rk = W + βAT Sk+1 A + β̄F T Rk+1 F + β̄LT(m) U L(m) ,
Sk = W + ᾱAT Sk+1 A + αF T Rk+1 F + αLT(m) U L(m) ,
T
T
ck = −αtr((F Rk+1 F − A Rk+1 A +
LT(m) U L(m) )(P̄ ))
T
+ α[tr(Rk+1 Q) + dk+1 + tr((B Rk+1 B + U )Q)]
+ ᾱ[tr(Sk+1 Q) + ck+1 ],
(29)
dk = −β̄tr((F T Rk+1 F − AT Rk+1 A + LT(m) U L(m) )(P̄ ))
T
+ β̄[tr(Rk+1 Q) + dk+1 + tr((B Rk+1 B + U )Q)]
+ β[tr(Sk+1 Q) + ck+1 ].
(30)
i
hP
N
T
T
Let J¯N = E
k=0 xk W xk + uk,c U uk,c = E[V0 (x0 )]. We
find that
J¯N = P (η−1 = 0) E[xT0 S0 x0 |η−1 = 0] + c0
+ P (η−1 = 1) E[xT0 R0 x0 |η−1 = 1] + d0 .
Leveraging the fact that {ηk } is stationary with P (ηk = 0) =
β
α+β as well as (29) and (30), we obtain
J¯N =
N −1
1 X
− αtr((F T Rk+1 F − AT Rk+1 A
α+β
k=0
+ LT(m) U L(m) )(P̄ )) + tr((βSk+1 + αRk+1 )Q)
+ αtr((B T Rk+1 B + U )Q)
+
0
βE[xT0 S0 x0 |η−1
]
1
αE[xT0 R0 x0 |η−1
]
+
α+β
,
j
where η−1
refers to the condition η−1 = j. It can be shown
(in a similar manner to the proof of Theorem 8) that the last
term is bounded. Note J¯ = limN →∞ N1 J¯N −1 . Moreover,
from [19][Theorem 3, Lemma 4], {Sk }, {Rk } converge to
S(m) , R(m) , respectively. This proves the desired result.
We now compute the expected correlation without attacks.
Theorem 8: Suppose α and β are chosen so the resulting
system has finite cost J(m) [19][Theorem 3] in the absence
of a Gaussian watermark. Then, for the control system (1)
with IID Gaussian and Markovian watermark (12), we have
lim E[ykT yk′ |H0 ] =
k→∞
tr(C(αX1 + βX0 )C T )
,
α+β
(31)
where
X0 = A(ᾱX0 + αX1 )AT ,
(32)
T
X1 = (A + BL(m) )(βX0 + β̄X1 )(A + BL(m) ) + BQB T
Proof: We begin with the Lemma below.
Lemma 9: ∀ M ∈ R2n×n , limk→∞ Lk0 (M ) = 0 where,
X
A(ᾱX + αY )AT
=
.
L0
Y
(A + BL(m) )(βX + β̄Y )(A + BL(m) )T
The proof is in the appendix. The closed loop dynamics are
xk+1 = (A + ηk BL(m) )xk − ηk BL(m) ek + wk + ηk B∆uk
ek+1 = (A − KCA)ek + (I − KC)wk − Kvk+1 ,
where ek = xk − x̂k|k . From (22), when ηk = 1. we obtain
E[x′k+1 xTk+1 |ηk = 1]
= (A + BL(m) )E[x′k xTk |ηk = 1](A + BL(m) )T −
(A + BL(m) )(E[x′k eTk |ηk = 1]LT(m) B T − E[x′k wkT |ηk = 1])
+ (A + BL(m) )E[x′k ∆uTk |ηk = 1]B T
+ BE[∆uk xTk |ηk = 1](A + BL(m) )T
+ B E[∆uk wkT |ηk = 1] + E[∆uk ∆uTk |ηk = 1]B T
− BE[∆uk eTk |ηk = 1](BL(m) )T ,
where we implicitly condition on H0 . x′k is independent of
∆uk , wk , ek and ∆uk is independent of xk , wk , ek . Thus,
E[x′k+1 xTk+1 |ηk = 1]
= (A +
(33)
BL(m) )E[x′k xTk |ηk
T
= 1](A + BL(m) ) + BQB T .
Next, since the Markov process is stationary and xk , x′k and
ηk are conditionally independent given ηk−1 , we observe
E[x′k xTk |ηk = 1]
(34)
= P (ηk−1 = 1|ηk = 1)E[x′k xTk |ηk = 1, ηk−1 = 1]
+ P (ηk−1 = 0|ηk = 1)E[x′k xTk |ηk = 1, ηk−1 = 0],
= β̄E[x′k xTk |ηk−1 = 1] + βE[x′k xTk |ηk−1 = 0].
It can be similarly shown that
E[x′k+1 xTk+1 |ηk = 0] = AE[x′k xTk |ηk = 0]AT .
E[x′k xTk |ηk = 0]
= αE[x′k xTk |ηk−1
= 1] + ᾱE[x′k xTk |ηk−1 = 0].
(35)
(36)
Letting Xk,j = E[x′k xTk |ηk−1 = j] we have
Xk,0
0
Xk+1,0
+
.
= L0
BQB T
Xk,1
Xk+1,1
(37)
Since L0 is stable, limk→∞ E[x′k xTk |ηk−1 = 0] and
limk→∞ E[x′k xTk |ηk−1 = 1] are obtained by solving a fixed
point equation which has a unique solution X0 and X1 . (32)
immediately follows from (37). Next, we find that
lim E[x′k xTk ] = P (ηk−1 = 1)X1 + P (ηk−1 = 0)X0 , (38)
Rather than optimizing over the parameters of the HMM,
we instead optimize over the autocovariance functions
Γ(d) , E[∆uk ∆uTk+d ]. For tractable analysis we replace the
constraint ρ(Aω ) ≤ ρ̄ with the following related assumption.
Assumption 1: Let Γ(d) be an autocovariance function for a
Gaussian process generated by an HMM (Aω , Ch , Ψ). Then
(Aω , Ch , Ψ, ρ̄) is feasible only if Γ̃(d) , ρ̄−|d| Γ(d) is a
autocovariance function of a stationary Gaussian process.
Γ̃(d) can be potentially realized by an alternate HMM
k→∞
αX1 + βX0
=
.
α+β
Finally, we observe that
E[ykT yk′ ]
= tr
E[(yk′ ykT )]
= tr
CE[x′k xTk ]C T
.
(39)
Thus, the watermark design problem (26) is given by
maximize
α,β,Q
subject to
tr(C(αX1 + βX0 )C T )
α+β
0
X0
X0
+
,
= L0
BQB T
X1
X1
J(m) (α, β) + tr((B T R(m) B + U )Q) ≤ δ,
0 < α, β ≤ 1.
(40)
For fixed α and β, the problem is an efficiently solvable
semidefinite program. However, to optimize over α and β,
we have to solve multiple instances of the problem over a
finite 2 dimensional space. Ideally a designer will sample the
space sufficiently. Note, not all (α, β) in (0, 1] × (0, 1] are
feasible as some selections of α and β lead to unbounded
cost. Likewise, there may be naturally occurring drops which
constrain α and β. For instance, if we add an artificial
Markovian drop process on top of a naturally occurring
IID drop process with drop probability pd , we know that
α ≤ (1 − pd ), β̄ ≤ (1 − pd ).
Remark 10: The optimal design of Watermark 1 requires
solving multiple instances of a convex optimization problem
with parameters varying over a bounded 2 dimensional space.
This will also be true for Watermark 2. A formulation that
considers a stationary Gaussian input with a Markovian drop
process is nontrivial. Even if analysis can be performed, optimal design will likely require searching over 3 dimensions.
This more complicated case is left for future work.
VI. T HE S ECOND WATERMARK D ESIGN
We now investigate a watermark consisting of stationary
Gaussian noise generated by a HMM (14) and an IID
Bernoulli drop process at the control input with drop probability equal to pd . Again, we design a watermark to address
a performance and security trade-off. We wish to solve:
maximize
lim E[ykT yk′ |H0 ]
pd ,Aω ,Ch ,Ψ
k→∞
subject to
J¯ ≤ δ, ρ(Aω ) ≤ ρ̄,
0 ≤ pd ≤ 1.
ζ̃k+1 = (Aω /ρ̄)ζ̃k + ψ̃k , ∆ũk = Ch ζ̃k ,
(42)
Cov(ζ̃0 ) = Aω Cov(ζ̃0 )ATω + Ψ,
(43)
ψ̃k ∼ N (0, Cov(ζ̃0 ) − Aω Cov(ζ̃0 )ATω /ρ̄2 ).
(44)
Note, that if ρ(Aω ) > ρ̄, (42) can not be a stationary
process. This HMM can be realized if and only if Cov(ζ̃0 ) −
Aω Cov(ζ̃0 )ATω /ρ̄2 is positive semidefinite. Intuitively, if
ρ(Aω ) is marginally less than ρ̄, there is a larger chance
that Cov(ζ̃0 ) − Aω Cov(ζ̃0 )ATω /ρ̄2 is positive semidefinite.
Remark 11: When ρ̄ = 1, Assumption 1, introduces no
relaxation. In fact, the resulting formulation optimizes all
stationary Gaussian processes in general. However, in the
case ρ̄ = 1, we will prove that the resulting Gaussian process
{∆uk } is entirely deterministic except for the initial watermark. A lower parameter ρ̄ reduces average performance, but
prevents an attacker who learns or guesses the current hidden
state from adequately predicting future watermarks.
We arrive at a relaxed formulation to (41) below.
Theorem 12: Consider the control system (1) with IID
Bernoulli and stationary Gaussian watermark (14). Suppose pd is chosen so that the system has finite cost J(b)
[19][Theorem 3] in the absence of a Gaussian watermark. An
equivalent formulation to (41) after replacing the constraint
ρ(Aω ) ≤ ρ̄ with Assumption 1 is given by
maximize
tr(CF2 (ω, H, pd )C T )
subject to
J(b) (pd ) + F1 (ω, H, pd ) ≤ δ,
ω,H,pd
H ∈ Cp×p , H 0.
where
F2 (ω, H, pd ) = 2Re 2sym L1 (M2 HB T ) + L1 (BHB T )
F1 (ω, H, pd ) = tr(U Θ) + tr((W + p̄d LT(b) U L(b) )F2 ),
Θ(ω, H, pd ) = 2Re (2sym [p̄d M1 H] + p̄d H) ,
−1
M2 = p̄d ρ̄s(A + BL(b) ) I − sρ̄(A + p̄d BL(b) )
B,
−1
M1 = p̄d ρ̄sL(b) I − sρ̄(A + p̄d BL(b) )
B,
L1 (X) = p̄d (A + BL(b) )L1 (X)(A + BL(b) )T + X
+ pd AL1 (X)AT ,
sym(X) =
(41)
(45)
0 ≤ pd ≤ 1, 0 ≤ ω ≤ 0.5,
X + XT
,
2
s = exp(2πjω),
p̄d = 1 − pd .
There is also an optimal solution (H∗ , ω∗ , pd ∗ ) such that
H∗ = hhH where hH denotes the conjugate transpose of
In this section, we illustrate the performance of the proposed watermarking designs through extensive numerical
results. We tested our watermark designs in various randomly
generated systems and, unless otherwise stated, averaged
results over 1500 trials. Replay attacks are considered.
In Fig. 2, we utilize Watermark 1, which has a Markovian drop process defined by parameters (α, β) and an IID
Gaussian watermark. The watermark is tested on a randomly
generated open loop stable system with 5 states, 4 inputs,
and 2 outputs. We plot the receiver operating characteristic
(ROC) curve for both the proposed correlation detector and a
χ2 detector. The χ2 detector serves as a benchmark, having
been previously used for attack detection [5], [8], [10], [16]
in watermarked systems. The threshold µ is chosen to be
a constant multiple of limk→∞ E[ykT yk′ ]. The ROC curves
are collected at multiple different costs ∆J = 1.05J ∗ ,
∆J = 0.45J ∗ and ∆J = 0.15J ∗ . Here, ∆J represent the
increase in the cost J¯ relative to optimal cost J ∗ without
drops or a Gaussian watermark. We compare a system with
drops (α = 0.69, β = 0.9) to a system without drops
(α = 1, β = 0). The proposed detector outperforms the
χ2 detector in all cases and packet drops improve the ROC
curve for both detectors. The improvement appears to be
higher for moderately valued ∆J before saturating. In Fig.
0.8
0.7
0.7
ROC Curve
∆ J= 0.15 J * , α=0.69, β=0.9
Probability of Detection β
Probability of Detection β
ROC Curve
0.8
0.6
0.5
0.4
0.3
∆ J = 0.15 J *, α=0.69, β=0.9
∆ J = 0.15 J *, α=1, β=0
0.2
∆ J = 0.45 J *, α=0.69, β=0.9
∆ J = 0.45 J *, α=1, β=0
∆ J= 0.45 J * , α=0.69, β=0.9
∆ J= 0.45 J * , α=1, β=0
∆ J= 1.05 J * , α=0.69, β=0.9
∆ J= 1.05 J * , α=1, β=0
0.5
0.4
0.3
0.2
∆ J = 1.05 J *, α=0.69, β=0.9
0.1
∆ J= 0.15 J * , α=1, β=0
0.6
0.1
*
∆ J = 1.05 J , α=1, β=0
0
0
0.05
0.1
0.15
0
0.2
0
Probability of False Alarm α
0.05
0.1
0.15
0.2
Probability of False Alarm α
(a) Correlation Detector
(b)
χ2
Detector
Fig. 2. Detection probability versus false alarm rate for χ2 and correlation
detectors for a system using Watermark 1.
∆
∆
∆
∆
∆
∆
60
50
40
J = 0.15,
J = 0.15,
J = 0.45,
J = 0.45,
J = 1.05,
J = 1.05,
Expected Time to Detection
Expected Time to Detection
Expected Time to Detection
Expected Time to Detection
70
α=0.69, β=0.9
α=1, β=0
α=0.69, β=0.9
α=1, β=0
α=0.69, β=0.9
α=1, β=0
30
20
10
0
0
0.05
0.1
0.15
250
∆ J = 0.15 J *, α = 0.69, β=0.9
*
∆ J = 0.15 J , α = 1, β=0
200
∆ J = 0.45 J *, α = 0.69, β=0.9
∆ J = 0.45 J *, α = 1, β=0
150
*
∆ J = 1.05 J , α=0.69, β=0.9
∆ J = 1.05 J *, α=1, β=0
100
50
0
0.2
0
Probability of False Alarm α
0.05
0.1
0.15
0.2
Probability of False Alarm α
(a) Correlation Detector
χ2
(b)
Fig. 3. Expected time to detection for
system using Watermark 1.
χ2
Detector
and correlation detectors for a
In Fig. 4, we introduce Watermark 2, which has IID
drops (with probability of drop pd ) and a stationary Gaussian
watermark. The watermark is added to a randomly generated
open loop stable system with 6 states, 5 inputs, and 5 outputs.
We plot ROC curves generated by both the correlation
detector and χ2 detector for a system with drops (pd = 0.6)
and a system without drops (pd = 0), at various costs of
control ∆J = 0.95J ∗ , ∆J = 0.45J ∗ and ∆J = 0.15J ∗ .
Time to detection plots are provided in Fig. 5. The results
and patterns observed here are similar to the results seen in
the system with Watermark 1.
ROC Curve
0.9
ROC Curve
∆ J=0.95 J *, p d =0.6
0.8
0.8
∆ J=0.95 J *, p d =0
0.7
0.7
Probability of Detection β
VII. S IMULATIONS
3, we plot the expected time to detection for both detectors
in a system with Watermark 1. The packet drop process
introduces an additional delay in the time to detection though
this additional time is less significant as ∆J is increased.
Probability of Detection β
h ∈ Cp . Letting Re and Im be the real and imaginary parts
of a matrix/vector, respectively, an optimal Aω , Ch , Ψ is
cos(2πω∗ ) − sin(2πω∗ )
Aω = ρ̄
,
sin(2πω∗ ) cos(2πω∗ )
√
(46)
Ch = 2 Re(h) Im(h) , Ψ = (1 − ρ̄2 )I.
The proof is similar in nature to the proof of Theorem
6 in [9]. A sketch is found in the appendix. For fixed
pd and ω, the proposed problem is an efficiently solvable
semidefinite program. To approximate a global maximum,
we solve the problem repeatedly over the space 0 ≤ ω ≤ 0.5
and 0 ≤ pd ≤ 1. For sufficiently large pd , the cost J¯
becomes infinite in open loop unstable systems [18], limiting
the feasible space. We can account for natural packet drops
in the system as before. For instance, if the input is dropped
naturally with probability p′d , we have p′d ≤ pd ≤ 1.
Remark 13: An optimal watermark for a given pd 6= pd ∗
may have better detection performance than the globally optimal watermark. Future work aims to use objective functions
that better highlight the relative performance of watermarks.
Remark 14: While packet drops at the sensor measurements are not modeled in this paper, our framework could
be extended to address this behavior without significantly
changing the formulations of the proposed optimization
problems. The main effect of packet drops at the sensor
side is a time varying Kalman gain. The objective function
and increase in cost J¯ due to the Gaussian portion of the
watermark are not affected by time variations in the Kalman
gain in both watermarking settings. Both J(m) and J(b) can
be empirically evaluated for fixed (α, β) and pd , respectively,
to account for packet drops at the sensor measurements.
0.6
0.5
0.4
*
∆ J = 0.95 J , p d =0.6
∆ J = 0.95 J *, p =0
0.3
d
*
∆ J = 0.45 J , p d =0.6
0.2
∆ J = 0.45 J *, p d =0
*
∆ J=0.45 J , p d =0.6
∆ J=0.45 J *, p =0
0.6
d
∆ J=0.15 J *, p d =0.6
*
0.5
∆ J=0.15 J , p d =0
0.4
0.3
0.2
*
∆ J = 0.15 J , p d =0.6
0.1
0
0.1
∆ J = 0.15 J *, p d =0
0
0.05
0.1
0.15
Probability of False Alarm α
(a) Correlation Detector
0.2
0
0
0.05
0.1
0.15
0.2
Probability of False Alarm α
(b)
χ2
Fig. 4. Detection probability versus false alarm rate for
detectors for a system using Watermark 2.
Detector
χ2
and correlation
Expected Time to Detection
60
Expected Time to Detection
90
∆ J= 0.15 J *, p =0.6
d
Expected Time to Detection
Expected Time to Detection
80
50
∆ J = 0.15 J *, p =0.6
d
40
*
∆ J = 0.15 J , p d =0
∆ J = 0.45 J *, p =0.6
d
30
*
∆ J = 0.45 J , p d =0
∆ J = 0.95 J *, p =0.6
d
20
*
∆ J = 0.95 J , p d =0
10
*
∆ J = 0.15 J , p =0
d
∆ J = 0.45 J *, p =0.6
70
d
*
∆ J = 0.45 J , p =0
60
d
*
∆ J = 0.6 J , p =0.6
d
50
*
∆ J = 0.6 J , p =0
d
40
30
20
10
0
0
0.05
0.1
0.15
0
0.2
0
Probability of False Alarm α
0.05
0.1
0.15
0.2
Probability of False Alarm α
(b) χ2 Detector
(a) Correlation Detector
Fig. 5. Expected time to detection for
system using Watermark 2.
χ2
and correlation detectors for a
R EFERENCES
Correlation Detector Statistic vs Time
0.1
0.09
0.08
0.07
0.06
0.05
0
200
400
600
800
Chi-Squared Detector Statistic vs Time
40
No Fault
Fault
0.11
Average Residue Detector Statistic
Average Correlation Detector Statistic
In Figs. 6 and 7, we plot χ2 detector and correlation
detector statistics (averaged over 500 trials) during a fault
in the system. The fault introduced (at time 210) is a
constant additive bias added to a subset of sensors (i.e. due
to disturbances/sensor drift). While the χ2 detector raises an
alarm, the correlation detector does not since the watermark
is preserved in the system. This motivates the use of both the
correlation and χ2 detector to distinguish faults from attacks.
If both detectors raise an alarm, indicating the watermark is
absent in the outputs, we consider a likely attack scenario.
If only the χ2 detector raises an alarm, we expect that the
watermark is preserved while the dynamics are inconsistent
with modeling. As such, we anticipate a fault.
0.12
35
30
20
15
10
1000
No Fault
Fault
25
0
200
400
Time
600
800
1000
Time
(a) Correlation Detector
(b)
χ2
Detector
χ2
Correlation Detector Statistic vs Time
170
160
150
140
130
120
110
100
0
200
400
600
800
Chi-Squared Detector Statistic vs Time
28
No Fault
Fault
180
1000
Average Residue Detector Statistic
Average Correlation Detector Statistic
Fig. 6. Average correlation detector and
detector statistics under a fault
at the sensor output for a system using Watermark 1.
190
termarks. We argued that packet drops can be beneficial for
detection and consequently considered the design of a joint
Bernoulli-Gaussian watermark to detect integrity attacks.
We proposed two main watermark designs in conjunction
with a correlation detector and provided efficiently solvable
optimization problems to address the trade-off between detection and control performances. In future work we aim to
generalize our watermarking approach to allow us to drop
either the entire control input or the Gaussian portion of the
watermark. We also hope to conduct testing in real systems.
26
24
22
No Fault
Fault
20
18
16
14
0
Time
200
400
600
800
1000
Time
(b) χ2 Detector
(a) Correlation Detector
χ2
Fig. 7. Average correlation detector and
detector statistics under a fault
at the sensor output for a system using Watermark 2.
VIII. C ONCLUSION AND F UTURE W ORK
In this paper, we showed how to incorporate Bernoulli
packet drops at the control input in the design of physical wa-
[1] A. A. Cárdenas, S. Amin, and S. S. Sastry, “Secure Control: Towards
Survivable Cyber-Physical Systems,” in Distributed Computing Systems Workshops, 2008. ICDCS ’08. 28th International Conference on
DOI - 10.1109/ICDCS.Workshops.2008.40. IEEE, 2008, pp. 495–500.
[2] T. M. Chen, “Stuxnet, the real start of cyber warfare? [editor’s note],”
IEEE Network, vol. 24, no. 6, pp. 2–3, 2010.
[3] J. Slay and M. Miller, “Lessons learned from the Maroochy water
breach,” in Critical Infrastructure Protection. Springer US, 2008,
pp. 73–82.
[4] G. Liang, S. R. Weller, J. Zhao, F. Luo, and Z. Y. Dong, “The 2015
ukraine blackout: Implications for false data injection attacks,” IEEE
Transactions on Power Systems, vol. 32, no. 4, pp. 3317–3318, 2017.
[5] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in
47th Annual Allerton Conference on Communication, Control, and
Computing, Sept 2009, pp. 911–918.
[6] C.-Z. Bai, F. Pasqualetti, and V. Gupta, “Security in stochastic control
systems: Fundamental limitations and performance bounds,” in American Control Conference (ACC), 2015. IEEE, 2015, pp. 195–200.
[7] S. Weerakkody, B. Sinopoli, S. Kar, and A. Datta, “Information flow
for security in control systems,” in 55th IEEE Conference on Decision
and Control (CDC). IEEE, 2016, pp. 5065–5072.
[8] Y. Mo, R. Chabukswar, and B. Sinopoli, “Detecting integrity attacks on
SCADA systems,” IEEE Transactions on Control Systems Technology,
vol. 22, no. 4, pp. 1396–1407, 2014.
[9] Y. Mo, S. Weerakkody, and B. Sinopoli, “Physical authentication
of control systems: designing watermarked control inputs to detect
counterfeit sensor outputs,” IEEE Control Systems Magazine, vol. 35,
no. 1, pp. 93 – 109, 2015.
[10] F. Miao, M. Pajic, and G. J. Pappas, “Stochastic game approach for
replay attack detection,” in 52nd IEEE Conference on Decision and
Control (CDC). IEEE, 2013, pp. 1854–1859.
[11] B. Satchidanandan and P. Kumar, “Dynamic watermarking: Active
defense of networked cyber–physical systems,” Proceedings of the
IEEE, vol. 105, no. 2, pp. 219–240, 2017.
[12] P. Hespanhol, M. Porter, R. Vasudevan, and A. Aswani, “Dynamic watermarking for general LTI systems,” arXiv preprint arXiv:1703.07760,
2017.
[13] M. Hosseini, T. Tanaka, and V. Gupta, “Designing optimal watermark
signal for a stealthy attacker,” in 2016 European Control Conference
(ECC). IEEE, 2016, pp. 2258–2262.
[14] J. Rubio-Hernan, L. De Cicco, and J. Garcia-Alfaro, “On the use of
watermark-based schemes to detect cyber-physical attacks,” EURASIP
Journal on Information Security, vol. 2017, no. 1, 2017.
[15] S. Weerakkody, Y. Mo, and B. Sinopoli, “Detecting integrity attacks
on control systems using robust physical watermarking,” in 53rd IEEE
Conference on Decision and Control (CDC), Los Angeles, California,
2014, pp. 3757–3764.
[16] O. Ozel, S. Weerakkody, and B. Sinopoli, “Physical watermarking
for securing cyber-physical systems via packet drop injections,” in To
appear, 8th IEEE International Conference on Smart Grid Communications, 2017.
[17] R. Chabukswar, Y. Mo, and B. Sinopoli, “Detecting integrity attacks
on SCADA systems,” in 18th IFAC World Congress, Milan, Italy, Aug
2011, pp. 11 239–11 244.
[18] L. Schenato, B. Sinopoli, M. Franceschetti, K. Poolla, and S. S.
Sastry, “Foundations of control and estimation over lossy networks,”
Proceedings of the IEEE, vol. 95, no. 1, pp. 163–187, 2007.
[19] Y. Mo, E. Garone, and B. Sinopoli, “LQG control with Markovian
packet loss,” in Control Conference (ECC), 2013 European. IEEE,
2013, pp. 2380–2385.
[20] Y. Mo and B. Sinopoli, “False data injection attacks in cyber physical
systems,” in First Workshop on Secure Control Systems, Stockholm,
Sweden, April 2010.
[21] F. Pasqualetti, F. Dorfler, and F. Bullo, “Attack detection and identification in cyber-physical systems,” IEEE Transactions on Automatic
Control, vol. 58, no. 11, pp. 2715–2729, 2013.
[22] A. Teixeira, D. Perez, H. Sandberg, and K. H. Johannson, “Attack
models and scenarios for networked control systems,” in Proceedings
of the 1st international conference on High Confidence Networked
Systems, Beijing, China, 2012, pp. 55–64.
[23] T. Chonavel and J. Ormrod, Statistical Signal Processing: Modelling
and Estimation, ser. Advanced Textbooks in Control and Signal
Processing. SPRINGER VERLAG GMBH, 2002.
[24] P. Delsarte, Y. Genin, and Y. Kamp, “Orthogonal polynomial matrices
on the unit circle,” IEEE Transactions on Circuits and Systems, vol. 25,
no. 3, pp. 149–160, 1978.
IX. A PPENDIX : P ROOF
OF
L EMMA 9
Then, we have
i j′
lim E[x̄ik x̄j′
k |ηk−1 = 0] = 0, lim E[x̄k x̄k |ηk−1 = 1] = 0,
k→∞
lim E[x̄ik x̄j′
k ] = 0,
k→∞
Sk+1 = g1 (Sk , Rk ), Rk+1 = g2 (Sk , Rk ), R0 = S0 = ǫI.
(50)
We observe that g1 (X, Y ) and g2 (X, Y ) are monotonically
increasing functions in (X, Y ). Because S1 ≥ S0 and R1 ≥
R0 , we see that {Sk } and {Rk } are monotonically increasing
in the semidefinite sense. Now consider the sequence
S̄k+1 = h1 (S̄k , R̄k ), R̄k+1 = h2 (S̄k , R̄k ), R̄0 = S̄0 = ǫI,
(51)
where we define
h1 (X, Y ) , W + αLT(m) U L(m) + ᾱAT XA + αF T Y F,
h2 (X, Y ) , W + β̄LT(m) U L(m) + βAT XA + β̄F T Y F.
Proof: We begin with the following Lemma.
Lemma 15: Assume {ηk } is a stationary Markovian drop
process. Suppose α > 0 and β > 0 are chosen so that the
system has finite cost J(m) [19][Theorem 3] in the absence of
a Gaussian watermark. Consider x̄k+1 = (A + ηk BL(m) )x̄k
and x̄′k+1 = (A + ηk BL(m) )x̄′k ,
(
( ′
x00 η−1 = 0
x00 η−1 = 0
′
x̄0 =
,
x̄
=
. (47)
′
0
x10 η−1 = 1
x10 η−1 = 1
k→∞
We claim that limN →∞ JN exists. Consider the sequence
lim E[(x̄ik )2 ] = 0,
k→∞
Again, we observe that h1 (X, Y ) and h2 (X, Y ) are monotonically increasing in (X, Y ). Because S̄1 ≥ S̄0 and R̄1 ≥
R̄0 , it can be seen that {S̄k } and {R̄k } are monotonically increasing in the semidefinite sense. Moreover, due to Lemma
4 in [19], {S̄k } and {R̄k } converge. We observe that if
X ≤ X̄ and Y ≤ Ȳ , then g1 (X, Y ) ≤ h1 (X̄, Ȳ ) and
g2 (X, Y ) ≤ h2 (X̄, Ȳ ). Since R0 = R̄0 and S0 = S̄0 , it
can be seen that Sk ≤ S̄k and Rk ≤ R̄k for all k.
As a result, {Sk } and {Rk } are bounded above by a
monotonically increasing, convergent sequence, which in
turn means that {Sk } and {Rk } are bounded. From the
monotone convergence theorem {Sk } and {Rk } converge
to some S ∗ and R∗ . It is immediately seen that
T
lim JN
x̄ik
where i, j ∈ {1, · · · , n} and
is the ith element of x̄k and
′
x̄j′
is
the
jth
element
of
x̄
.
k P
k
N
T
Proof: Define JN ,
k=0 ǫE[x̄k x̄k ] where ǫ > 0
is chosen so ǫI ≤ WP
. Moreover the cost to go function is
N
defined as V̄j (x̄j ) , k=j ǫE[x̄Tk x̄k |η−1:j−1 ]. We show that
(
E[x̄Tk S̄k x̄k |η−1:k−1 ] (ηk−1 = 0)
V̄k (x̄k ) =
,
(48)
E[x̄Tk R̄k x̄k |η−1:k−1 ] (ηk−1 = 1)
where
N →∞
lim E[(x̄ik )2 ] = 0.
2
By symmetry this also implies that limk→∞ E[(x̄j′
k ) ] = 0.
By Cauchy Schwartz, we see
2
Since E[(x̄ik )2 ]E[(x̄j′
k ) ] converges
i j′ 2
(E[x̄k x̄k ]) converges to 0 and thus
T
g2 (X, Y ) , ǫI + βA XA + β̄F Y F,
and F = A + BL(m) , S̄N = ǫI, R̄N = ǫI. The proof is
by induction. (48) holds for k = N . Assume it holds for
k = t + 1. We next show (48) holds for k = t. Conditioned
on ηt−1 = 0, we have
V̄t (x̄t ) = E[ǫx̄Tt x̄t + V̄t+1 (x̄t+1 )|η−1:t−1 ],
= E[ǫx̄Tt xt + x̄Tt (ᾱAT S̄t+1 A + αF T R̄t+1 F )x̄t |η−1:t−1 ],
=
E[x̄Tt g1 (S̄t+1 , R̄t+1 )x̄t |η−1:t−1 ]
=
E[x̄Tt S̄t x̄t |η−1:t−1 ].
The case when ηt−1 = 1 is similar. Thus,
JN = E[V̄0 (x̄0 )] =
T
βx00 S̄0 x00
T
αx10 R̄0 x10
+
α+β
(49)
to
0,
(54)
we
lim E[x̄ik x̄j′
k ] = 0.
k→∞
know
(55)
Next, we observe that
0≤
min(α, β)
E[(x̄ik )2 |ηk−1 = l] ≤ E[(x̄ik )2 ],
α+β
(56)
where l ∈ {0, 1}. As α, β > 0 by assumption, we know
limk→∞ E[(x̄ik )2 |ηk−1 = l] = 0. Using the Cauchy Schwartz
inequality in a similar manner as before, we see
i j′
lim E[x̄ik x̄j′
k |ηk−1 = 1] = 0, lim E[x̄k x̄k |ηk−1 = 0] = 0.
k→∞
(57)
k→∞
.
(53)
k→∞
j′ 2
2
i 2
0 ≤ (E[x̄ik x̄j′
k ]) ≤ E[(x̄k ) ]E[(x̄k ) ].
g1 (X, Y ) , ǫI + ᾱAT XA + αF T Y F,
(52)
P
T
Note that JN = N
k=0 ǫE[x̄k x̄k ]. Since JN converges to a
T
finite constant, limk→∞ E[x̄k x̄k ] = 0. Since 0 ≤ E[(x̄ik )2 ] ≤
E[x̄Tk x̄k ], we immediately obtain
S̄k = g1 (S̄k+1 , R̄k+1 ), R̄k = g2 (S̄k+1 , R̄k+1 ),
T
T
βx00 S ∗ x00 + αx10 R∗ x10
=
.
α+β
We are now ready to prove the desired result. To this end,
we first observe that
E[x̄′k+1 x̄Tk+1 |ηk = 0]
E[x̄′k x̄Tk |ηk−1 = 0]
.
= L0
E[x̄′k x̄Tk |ηk−1 = 1]
E[x̄′k+1 x̄Tk+1 |ηk = 1]
(58)
As a result,
0′ 0 T
E[x̄′k x̄Tk |ηk−1 = 0]
k x0 x0
.
(59)
= L0
′
E[x̄′k x̄Tk |ηk−1 = 1]
x10 x10 T
Leveraging (57), we see that limk→∞ E[x̄′k x̄Tk |ηk−1 = l] = 0
for l ∈ {0, 1}. Consequently, we have
0′ 0 T
k x0 x0
= 0.
(60)
lim L0
′
k→∞
x10 x10 T
′
′
Note that x00 , x00 , x10 , and x10 can be chosen so that Lk0 is
applied to an arbitrary canonical basis vector in R2n×n . Thus,
for all M ∈ R2n×n , limk→∞ Lk0 (M ) = 0. Thus, L0 is stable.
X. A PPENDIX : P ROOF
OF
T HEOREM 12
(61)
Noting that the IID drop case is a special instance
of Markovian drops, we know from Lemma 15 that
limk→∞ E[(x̄ik )2 ] = 0. Using the fact that E[(x̄ik )2 ] ≥
(E[x̄ik ])2 ≥ 0, we have limk→∞ E[x̄k ] = 0. As a result,
for all x0,∗ ∈ Rn
lim (A + p̄d BL(b) )k x0,∗ = 0.
k→∞
(62)
Thus, (A + p̄d BL(b) ) is Schur stable. Next, we note that
E[x̄′k+1 x̄Tk+1 ]
=
L1 (E[x̄′k x̄Tk ]).
for fixed t ≥ 0. It can be seen that
xt = l1,{ηk } (w−∞:t−1 , v−∞:t−1 ) + γt (∆u−∞:t−1 ),
i=1−t
j=1−i
where l1 and l2 are linear functions of the process and sensor
noise for fixed realizations of the drop process ηk . Since
{wk } and {vk } are independent of ∆uk , we observe that
N
−1
X
1
tr(W Cov(γt ))+
lim
J¯ = J(b) (pd ) +
N N →∞ t=0
!
j=1−i
We see that E[Zt+1 ] is equal to
E[(A + ηt BL(b) )Zt (A + ηt BL(b) )T + ηt2 B∆ut ∆uTt B T ].
Since ηt is independent of Zt , we have
E[Zt+1 ] = L1 (E[Zt ]) + p̄d BΓ(0)B T .
Since L1 is stable and the system has been running since
k = −∞, E[Zt ] is the unique solution of the following fixed
point equation.
E[Zt ] = L1 (E[Zt ]) + p̄d BΓ(0)B T = L1 (BΓ(0)B T ).
In addition, let
Ytd ,
(64)
k→∞
∞
X
d T
ξi,t
γi,t ,
i=1−t
t−1
Y
j=1−i−d
(A + ηj BL(b) ) η−i−d B∆u−i−d .
By similar reasoning, we find that E[Ytd ] equals
L1 p̄d (A + BL(b) )(A + p̄d BL(b) )d−1 BΓ(d)B T .
We argue
x′0,∗
and x0,∗ can be chosen so that Lk1 is applied
Note that
to an arbitrary canonical basis vector in Rn×n . Thus, for all
M ∈ Rn×n , limk→∞ Lk1 (M ) = 0. Thus, L1 is stable.
We now proceed to the main proof. We obtain an equivalent
realization to (41) by using autocovariance functions Γ(d) ,
E[∆uk ∆uk+d ].
(67)
Step 1a: Calculate
Cov(γt ):
P
T
Define Zt , ∞
i=1−t γi,t γi,t where
t−1
Y
(A + ηj BL(b) ) η−i B∆u−i .
γi,t ,
d
ξi,t
,
Leveraging (55), we note limk→∞ E[x̄′k x̄Tk ] = 0 and this
implies
lim Lk1 (x′0,∗ xT0,∗ ) = 0.
(65)
(66)
ut,c = l2,{ηk } (w−∞:t−1 , v−∞:t ) + ηt L(b) γt (∆u−∞:t−1 )
+ ηt ∆ut ,
∞ t−1
Y
X
(A + ηj BL(b) ) η−i B∆u−i ,
γt (∆u−∞:t−1 ) =
(63)
As a result,
E[x̄′k x̄Tk ] = Lk1 (x′0,∗ xT0,∗ ).
E[xTt W xt + uTt,c U ut,c ] = tr(W Cov(xt )) + tr(U Cov(ut,c )),
tr(U Cov(ηt [L(b) γt + ∆ut ])) .
Proof: We begin with the following Lemma.
Lemma 16: Suppose pd is chosen so the system with
IID drops has finite cost J(b) [19][Theorem 3]. Then the
matrix (A+ p̄d BL(b) ) is Schur stable. Moreover, the operator
L1 (X) , p̄d (A + BL(b) )X(A + BL(b) )T + pd AXAT is stable. Specifically, ∀M ∈ Rn×n , we have limk→∞ Lk1 (M ) = 0
Proof: Consider the systems x̄k+1 = (A+ηk BL(b) )x̄k ,
and x̄′k+1 = (A + ηk BL(b) )x̄′k . where ηk is an IID drop
process with drop probability pd and x̄0 = x0,∗ , x̄′0 = x′0,∗ .
Observe that
E[x̄k ] = (A + p̄d BL(b) )k x0,∗ .
Step 1: Calculate J¯ in terms of Γ(d):
Let us first compute
"
Cov(γt ) = E Zt +
=2
∞
X
d=1
∞
X
d=1
Ytd
+
(Ytd )T
#
sym[Y∗d ] + L1 (BΓ(0)B T ),
(68)
that Γ̃(d) is real. Moreover, from (68), we see that
where
Y∗d = L1 p̄d (A + BL(b) )(A + p̄d BL(b) )d−1 BΓ(d)B
Step 1b: Calculate Cov(ηt [L(b) γt + ∆ut ]):
We argue that
E[ηt2 L(b) γt ∆uTt ] = p̄2d L(b)
∞
X
T
.
(A + p̄d BL(b) )d BΓ(d + 1).
d=0
Therefore, we obtain
Cov(ηt [L(b) γt + ∆u1 ]) = p̄d Γ(0) + L(b) Cov(γ)LT(b)
!
∞
X
(A + p̄d BL(b) )d BΓ(d + 1) , (69)
+ 2sym p̄2d L(b)
d=0
where Cov(γ) , Cov(γt ) is given in (68). Note, Cov(γt ) is
constant in t. Substituting (69) into (67), we have J¯ =
J(b) (pd ) + tr(p̄d U Γ(0)) + tr((W + p̄d LT(b) U L(b) )Cov(γ))
!
∞
X
2
d
+ tr 2U sym p̄d L(b)
(A + p̄d BL(b) ) BΓ(d + 1)
.
d=0
Step 2: Calculate E[ykT yk′ |H0 ] in terms of Γ(d):
Recall from the proof of Theorem 8 and (39)
E[ykT yk′ |H0 ] = tr CE[x′k xTk ]C T .
We observe that x′k = γk . Thus, from (66), we assert
(70)
lim E[ykT yk′ |H0 ] = tr CCov(γ)C T .
k→∞
Step 3: Convert to Frequency Domain:
Optimizing over the autocovariance functions is intractable
as there are infinitely many optimization variables. In this
case, as in the work [9], we will leverage Bochner’s theorem
in [23, p.64] (see also [24]). This theorem provides a frequency domain representation of an autocovariance function
of a stationary process:
Theorem 17 (Bochner’s theorem): Γ(d) is an autocovariance function of a stationary Gaussian process {∆uk } if and
only if there exists a unique positive Hermitian measure ν
of size p × p satisfying
Z 0.5
Γ(d) =
exp(2πjdω)dν(ω).
−0.5
Note that a positive Hermitian measure ν takes a Borel set
in [−0.5 0.5] and outputs a positive semidefinite Hermitian
matrix in Cp×p . We choose to optimize over Γ̃(d), which
has bijective relationship with Γ(d). By assumption Γ̃(d) is
an autocovariance function of a stationary Gaussian process.
As a result, we can use Bochner’s theorem to rewrite Γ̃(d)
in terms of a Riemann sum. Specifically,
#
" q
X
exp(2πjdωi )ν̃(Ii ) ,
(71)
Γ̃(d) = lim 2Re
σ→0
i=1
∪qi=1
where Ii ∩ Ij = ∅,
Ii = [0, 0.5], ωi ∈ Ii and σ is
the maximum length of Ii . Here, we also leverage the fact
Cov(γ)
= lim
σ→0
∞
X
q
h
h
X
2Re 2sym L1 p̄d exp(2πjωi )ρ̄(A + BL(b) )
i=1
(ρ̄ exp(2πjωi )(A + p̄d BL(b) ))d−1 B ν̃(Ii )B T
d=1
ii
h
,
+ L1 B ν̃(Ii )B T
= lim
σ→0
q
h
h
X
2Re 2sym L1 p̄d exp(2πjωi )ρ̄(A + BL(b) )
i=1
(I − ρ̄ exp(2πjωi )(A + p̄d BL(b) ))−1 B ν̃(Ii )B T
ii
h
,
+ L1 B ν̃(Ii )B T
= lim
σ→0
i
q
X
i
F2 (ωi , ν̃(Ii ), pd ).
(72)
i=1
The inverse is well defined since we showed (A + p̄d BL(b) )
is Schur stable. By similar reasoning it can be shown that
q
X
F1 (ωi , ν̃(Ii ), pd ).
(73)
J¯ = J(b) (pd ) + lim
σ→0
i=1
Replacing ρ(Aω ) ≤ ρ̄ with Assumption 1 in problem (41),
we arrive at the following equivalent formulation:
q
X
tr(CF2 (ωi , ν̃(Ii ), pd )C T )
maximize lim
σ→0
ν̃(Ii ),pd
subject to
i=1
J(b) (pd ) + lim
σ→0
0 ≤ pd ≤ 1.
q
X
i=1
F1 (ωi , ν̃(Ii ), pd ) ≤ δ,
(74)
Step 4: Demonstrate Equivalence:
The rest of the result follows from Steps 3 and 4 in the
proof of Theorem 6 in [9] when pd < 1. In particular, we can
leverage the linearity of F2 and F1 in H for fixed pd < 1 and
ω to show that the optimal value of (45) is an upper bound
on the optimal value for problem (74). Then, we show that
for Borel set Sb ⊂ [−0.5, 0.5], the measure
ν̃(Sb ) = Iω∗ ∈Sb H∗ + I−ω∗ ∈Sb conj(H∗ ),
(75)
where I is the indicator function and conj refers to the
complex conjugate, achieves this upper bound. The resulting
autocovariance function is
Γ(d) = 2ρ̄|d|Re(exp(2πjdω∗ )H∗ ),
(76)
and can be generated by the HMM (46) if there exists an optimal H∗ , which has rank 1. Theorem 7 of [8] demonstrates the
existence of such a solution, while the associated proof shows
how such a solution can be constructed from an optimal H∗
with rank greater than 1. When pd = 1, F2 and F1 are
identically 0, establishing the equivalence of (45) and (74)
in this scenario. Note also in this case, (if J(b) (pd ) ≤ δ) any
stationary Gaussian process in the feasible region is optimal
since the resulting additive input is immediately dropped.
| 3cs.SY
|
Under review at Computer Vision and Image Understanding Journal.
Taking Visual Motion Prediction To New Heightfields
Sébastien Ehrhardt1
1
Aron Monszpart2
2
Dept. of Engineering Science,
University of Oxford
{hyenal, vedaldi}@robots.ox.ac.uk
arXiv:1712.09448v1 [cs.CV] 22 Dec 2017
Niloy Mitra2
Andrea Vedaldi1
Dept. of Computer Science,
University College London
{a.monszpart, n.mitra}@cs.ucl.ac.uk
Abstract
to the description of a physical scenario is anything but
trivial. First, the scenario needs to be abstracted (e.g., by
segmenting the scene into rigid objects, deciding which equations to apply, and estimating physical parameters such as
mass, linear and angular velocity, etc.). Then, prediction still
requires the numerical integration of complex systems of
equations. It is unlikely that this is the process of mental
modeling followed by natural intelligences.
While the basic laws of Newtonian mechanics are well
understood, explaining a physical scenario still requires
manually modeling the problem with suitable equations and
estimating the associated parameters. In order to be able
to leverage the approximation capabilities of artificial intelligence techniques in such physics related contexts, researchers have handcrafted the relevant states, and then
used neural networks to learn the state transitions using
simulation runs as training data. Unfortunately, such approaches are unsuited for modeling complex real-world scenarios, where manually authoring relevant state spaces tend
to be tedious and challenging. In this work, we investigate
if neural networks can implicitly learn physical states of
real-world mechanical processes only based on visual data
while internally modeling non-homogeneous environment
and in the process enable long-term physical extrapolation.
We develop a recurrent neural network architecture for this
task and also characterize resultant uncertainties in the form
of evolving variance estimates. We evaluate our setup to
extrapolate motion of rolling ball(s) on bowls of varying
shape and orientation, and on arbitrary heightfields using
only images as input. We report significant improvements
over existing image-based methods both in terms of accuracy of predictions and complexity of scenarios; and report
competitive performance with approaches that, unlike us,
assume access to internal physical states.
In an effort to develop models of physics that are more
suitable for artificial intelligence, in this work, we ask
whether a representation of the physical state of a mechanical system can be learned implicitly by a neural network,
and whether this can be used to perform accurate predictions
efficiently (i.e., extrapolating to predict future events). To
this end, we propose a new learnable representation with
several important properties. First, the representation is not
handcrafted, but rather automatically induced from visual
observations. Second, the representation is distributed and
can model physical interactions of objects with complex surrounding, such as irregularly-shaped ground. Third, despite
its distributed nature, the representation can model a number of interacting discrete objects such as colliding balls,
without the need of ad-hoc components such as collision
detection subnetworks. Fourth, since physical predictions
integrate errors over time and are thus inherently ambiguous,
the representation produces robust probabilistic predictions
which model such ambiguity explicitly. Finally, through extensive evaluation, we show that the representation performs
well for both extrapolation and interpolation of mechanical
phenomena.
1. Introduction
Our paper is not the first that looks at learning to predict mechanical phenomena using deep networks, but, to
the best of our knowledge, it is the first to achieve most of
the benefits listed above, particularly when these are combined. For example, the recent Neural Physics Engine (NPE)
of [Chang et al., 2017] uses a neural network to learn the
state transition function of mechanical systems. Differently
from our approach, their state is handcrafted and includes
physical parameters such as positions, velocities, and masses
of rigid bodies. While NPE works well, it still requires to
Animals can make remarkably accurate and fast predictions of physical phenomena in order to perform activities
such as navigate, prey, or burrow. However, the nature of the
mental models used to perform such predictions remains unclear and is still actively researched [Hamrick et al., 2016].
In contrast, science has developed an excellent formal
understanding of physics; for example, mechanics is nearly
perfectly described by Newtonian physics. However, while
the constituent laws are simple and accurate, applying them
1
Under review at Computer Vision and Image Understanding Journal.
abstract the physical system manually, by identifying the
objects and their physical parameters, and by explicitly integrating such parameters. In contrast, our abstractions are
entirely induced from external observations of object motions. Hence, our system implicitly discovers any hidden
variable or state required to perform tasks such as long-term
physical extrapolation in an optimal manner. Furthermore,
the integration of physical parameters over time is also implicit and performed by a recurrent neural network architecture. This is needed since the nature of the internal state
is undetermined; it also has a major practical benefit as, as
we show empirically, the system can be trained to not only
extrapolate physical trajectories, but also to interpolate them.
Remarkably, interpolation is still obtained by computing the
trajectory in a feed-forward manner, from the first to the last
time step, using the recurrent model.
Another significant difference with NPE is in the fact that
our system uses visual observations to perform its predictions. In this sense, the work closest to ours is the Visual
Interaction Networks (VIN) of [Watters et al., 2017], which
also use visual input for prediction. However, our system
is significantly more advanced as it can model the interaction of objects with complex and irregular terrain. We show
empirically that VIN is not very competitive in our more
complex experimental setting.
There are also several aspects that we address for the
first time in this paper. Empirically, we push our model
by considering scenarios beyond the ‘flat’ ones tackled by
most recent papers, such as objects sliding and colliding on
planes, and look for the first time at the case of ball(s) rolling
on non-trivial 3D shapes (e.g., bowls of varying shape and
orientation, or terrains modeled as arbitrary heightfields),
where both linear and angular momenta are tightly coupled.
Furthermore, since physical extrapolation is inherently ambiguous, we allow the model to explicitly estimate its prediction uncertainty by estimating the variance of a Gaussian
observation model. We show that this modification further
improves the quality of long-term predictions.
The rest of the paper is organized as follows. The relation
of our work to the literature is discussed in section 2. The
detailed structure of the proposed neural networks is given
and motivated in section 3. These networks are tested on a
large dataset of simulated physical experiments described
in section 4 and extensively evaluated and contrasted against
related works in section 5. We conclude by discussing current limitations and directions for future investigation in
section 6.
via empirical and visual observation of the motion of objects.
This research is thus related to a number of recent works in
various machine learning sub-areas, discussed next.
Learning intuitive physics [Battaglia et al., 2013] are
one of the first to consider ‘intuitive’ physical reasoning;
their aim is to answer simple qualitative questions related to
rigid body processes, such as determining whether a certain
tower of blocks is likely to fall or not. They approach the
problem by using a sophisticated physics engine that incorporates all required knowledge about Newtonian physics
a-priori. More recently, [Mottaghi et al., 2016] used static
images and a graphics rendering engine (Blender) to predict
motion and forces from a single RGB image. Motivated
by the recent success of deep learning for image analysis
(e.g., [Krizhevsky et al., 2012]), they trained a convolutional
neural network to predict such quantities and used it to
produce a “most likely motion,” rendering it using a traditional computer graphics pipeline. With a similar motivation,
[Lerer et al., 2016] and [Li et al., 2017] also applied deep
networks to predict the stability of towers of blocks purely
from images. These approaches demonstrated that such networks can not only predict instability, but also pinpoint the
source of such instability, if any. Other approaches such as
[Agrawal et al., 2016] or [Denil et al., 2016] have attempted
to learn intuitive physics of objects through manipulation;
however, their models did not aim to capture the underlying
dynamics of the systems.
Learning physics The work by [Wu et al., 2015] and its
extension [Wu et al., 2016] propose methods to learn physical properties of scenes and objects. [Wu et al., 2015] use
an MCMC-sampling based approach that assumes complete knowledge of the physical equations necessary to estimate physical parameters. In [Wu et al., 2016], a deep
learning based approach was used instead of MCMC, albeit still explicitly encoding physics in a simulator. Physical laws were also explicitly incorporated in the model
by [Stewart and Ermon, 2017] to predict the movement of a
pillow from unlabelled data. Their method was, however,
only applied to a fixed small number of future frames.
The research performed by [Battaglia et al., 2016]
and [Chang et al., 2017] focused on dynamics and attempted
to partially substitute the physics engine with a neural network that captures a selection of relevant physical laws. Both
approaches were able to use such networks to accurately predict updates for the physical state of the world. Although
results are plausible and promising, [Chang et al., 2017] suggest that long-term predictions remain difficult. Furthermore,
in both approaches, their neural networks only predict instantaneous updates of physical parameters that are then
2. Related Work
We address the problem of training deep neural networks
that can perform long-term predictions of mechanical phenomena while learning the required physical laws implicitly,
2
Under review at Computer Vision and Image Understanding Journal.
Input images t = 0 . . . 3
explicitly integrated. In contrast, in this work propagation is
implicit and applies a recurrent neural network architecture
to an implicit representation of the world.
Closer to our approach, [Fragkiadaki et al., 2016] and
[Watters et al., 2017] attempted to learn an internal representation of the physical world from images. In addition
to observing images, it is also possible to generate them
as [Fragkiadaki et al., 2016] learn to perform long-term extrapolation more successfully. Image generation can be
seen as a constraint that avoids the over time degeneration of the internal representation of dynamics. The work
of [Watters et al., 2017] extends the Interaction Network by
[Battaglia et al., 2016] to propagate an implicit representation of the dynamics of objects, obtaining a Visual Interaction Network (VIN). While their approach is the closest to ours, it has various limitations including not modeling the interaction with complex environments and the
relatively small size of the input images. The Predictron
by [Silver et al., 2016] also propagates a tensor state, but
suffers from the same drawbacks. [Ehrhardt et al., 2017b]
showed, how long-term extrapolation models can be trained
for one object moving on smooth analytic surfaces, such as
ellipsoids.
ENCODER
φenc
3 × 3 × Nf
TRANSITION
DECODER
Approximating physics for plausible simulation Several authors focused on learning to perform plausible physical predictions, for example to generate realistic future frames in a video [Tompson et al., 2016,
Ladický et al., 2015], or to infer rigid body collision parameters from monocular videos [Monszpart et al., 2016]. In
these approaches, physics-based losses are used to learn
plausible yet not necessarily accurate results, which may
be appropriate for tasks such as rendering and animation.
[Battaglia et al., 2016] also use a loss that captures the concept of energy conservation. The latter can be seen as a
way to incorporate knowledge about physics a-priori into
the network, which differs from our goal of learning any
required physical knowledge from empirical observations.
DispNet / PosNet
φtrans
φdec
ProbNet
Figure 1. Overview of our proposed pipeline. The first four images of a sequence first pass through a partially pre-trained feature
encoder network to build the concept of physical state. It then
recursively passes through a transition layer to produce long-term
predictions about the future states of the objects. It is then decoded
to produce state estimates. While our DispNet and PosNet models
are trained to regress the next states, the ProbNet model trained
with the log-likelihood loss is also able to handle the notion of
uncertainty thanks to its extended state space. Note here that only
one object is considered, extension for multiple objects is discussed
in section 3.4.
a simple deep recurrent architecture. Our work builds on
their success, and propagates a tensor-based state representation instead of a vector-based one. Using spatial convolutional operators allows for knowledge to be stored and
propagated locally w. r. t. the object locations in the images.
Learning dynamics Physical extrapolation can be performed without integrating physical equations explicitly. For example, [Hochreiter and Schmidhuber, 1997]
used LSTMs to make accurate long-term predictions
in human pose estimation [R. Villegas, 2017] and in
simulated environments [Oh et al., 2015, Chiappa, 2017].
Propagation can also be done using simpler convolutional operators; [Xue et al., 2016], in particular, used
these to generate possible future frames given a single static image and [De Brabandere et al., 2016] applied
it to the moving MNIST dataset for long-term prediction. The work by [Ondruska and Posner, 2016] and
[Ehrhardt et al., 2017a] also showed that an internal representation of dynamics can be propagated through time using
3. Method
In this section, we propose a novel neural network model
to make predictions about the evolution of a mechanical system from visual observations of its initial state. In particular,
this network, summarized in Fig. 1, can predict the motion
of one or more rolling objects accounting for variations in
the 3D geometry of the environment.
Formally, let yt be a vector of physical quantities that
we would like to predict at time t, such as the position of
one or more objects. Physical systems satisfy a Markov
3
Under review at Computer Vision and Image Understanding Journal.
condition, in the sense that there exists a state vector ht such
that (i) measurements yt = g(ht ) are a function of the state
and (ii) the state at the next time step ht+1 = f (ht ) depends
only on the current value of the state ht . Uncertainty in
the model can be encoded by means of observation p(yt |ht )
and transition p(ht+1 |ht ) probabilities, resulting in a hidden
Markov model.
State-only methods, such as the Neural Physics Engine
(NPE) by [Chang et al., 2017] start from an handcrafted definition of the state ht . For instance, in order to model a
scenario with two balls colliding, one may choose ht to contain the position and velocity of each ball. In this case, the
observation function g may be as simple as extracting the
position components from the state vector. It is then possible to use a neural network φ to approximate the transition
function f . In particular, [Chang et al., 2017] suggest that
it is often easier for a network to predict a rate of change
∆t = φ(ht ) for some of the physical parameters (e.g., the
balls’ velocities), which can be used to update the state using
a hand-crafted integrator ht+1 = f˜(ht , ∆t ).
While this approach works well, there are several limitations. First, even if the transition function is learned, the
state ht is defined by hand. Even in the simple case of the
colliding balls, the choice of state is ambiguous; for example, one could include in the state not only the position and
velocity of the balls, but also their radius, mass, elasticity,
friction coefficients, etc. Learning the state as well has the
significant benefit of making such choices automatic. Second, training a transition function requires knowledge of the
state values, which may be difficult to obtain except in the
case of simulated data. Third, in order to use such a system
to perform predictions, one must know the initial value of
the state h0 of the system, whereas in many applications one
would like to start instead from sensory inputs xt such as
images [Fragkiadaki et al., 2016].
We propose here an approach to address these difficulties. We assume that the state ht is a hidden variable, to
be determined as part of the learning process. Since the ht
cannot be observed, the transition function ht+1 = f (ht )
cannot be learned directly as in the NPE. Instead, state and
transitions must be inferred jointly as a good explanation
of the observed physical measurements yt . Any integrator
involved in the computation of the transition function is
implicitly moved inside the network, which is a recurrent
neural network architecture. In our experiments (section 5),
we show that the added flexibility of learning an internal
state representation and its evolution automatically allows
the system to scale well to the complexity of the physical
scenario.
Since the evolution of the state ht cannot be learned by
observing measurements yt in isolation, the system is supervised using sequences y[0,T ) = (y0 , . . . , yT −1 ) of observations. This is analogous to a Hidden Markov Model (HMM),
which is often learned by maximizing the likelihood of the
observation sequences after marginalizing the hidden state.1
As an alternative learning formulation, we propose instead to
consider the problem of long-term predictions starting from
an initial set of observations. Not only this is more directly
related to applications, but it has the important benefit that
predictions can be performed equally well from initial observations of the physical quantities yt or of some other sensor
reading xt , such as images.
Our system is thus based on learning three modules:
(i) an encoder function that estimates the state ht =
φenc (x(t−T0 ,t] ) from the T0 most recent sensor readings
(alternatively ht = φenc (y(t−T0 ,t] ) can use the T0 most
recent physical observations); (ii) a transition function
ht+1 = φtrans (ht ) that evolves the state through time; and
(iii) a decoder function that maps the state ht to a physical
observation yt = φdec (ht ), and in some case an uncertainty
associated. The rest of the section discusses the three modules, encoder, transition, and decoder maps, as well as the
loss function used for training. Further technical details can
be found in section 5.
3.1. Encoder Map: from images to state
The goal of the encoder map is to take T0 consecutive
video frames observing the initial part of the object motion and to produce an estimate h0 = φenc (x(−T0 ,0] ) of
the initial state of the physical system. In order to build
this encoder, we follow [Fragkiadaki et al., 2016] and concatenate the RGB channels of the T0 images in a single
Hi × Wi × 3T0 tensor. The latter is passed to a convolutional neural network φenc outputting a feature tensor
s0 ∈ RH×W ×C , used as internal representation of the system’s state. Note that this representation is spatially distributed and differs from the concentrated vector representation of the VIN of [Watters et al., 2017]. In the experiments,
we will show the advantage of using a tensorial representation in modeling complex environments. We also augment
our tensor representation with a state vector pt ∈ Rn , so
that the state is the pair ht = (st , pt ). In deterministic cases,
n = 2 and pt is the 2D projection of the object’s location on
the image plane. For multiple objects (see section 3.4) this
state is computed for each object independently.
1 Formally, a Markov model is given by p(y
[0,T )] , h[0,T ) ) =
Q −2
p(h0 )p(y0 |h0 ) T
p(h
|h
)p(y
|h
);
traditionally,
p can be
t+1 t
t+1 t+1
t=0
learned as the maximizer of the log-likelihood maxp Ey [log Eh [p(y, h)]],
where we dropped the subscripts for compactness. Learning to interpolate/extrapolate can be done by considering subsets ȳ ⊂ y of the measurements as given and optimizing the likelihood of the conditional probability
maxp Ey [log Eh [p(y, h|ȳ)]].
4
Under review at Computer Vision and Image Understanding Journal.
3.2. Transition Map: evolving the state
where σλ,α (z) = λ/(1 + exp(−z)) + α. In the following,
we will refer to this method as ProbNet, whereas the other
method estimated displacement without uncertainty will be
referred to as DispNet. Table 1 summarize the different
methods and their specificity.
The state ht is evolved through time by learning the transition function φtrans : ht 7→ ht+1 . Since the initial state h0
is obtained from the encoder map, the state at time t can be
written as, ht = φttrans (φenc (x(−T0 ,0] )).
More in detail, the distributed state component st is updated by using a convolutional network st+1 = φs (st ).
The concentrated component pt is updated incrementally
as pt+1 = pt + φp (st ), where φp (st ) is estimated using a
single layer perceptron regressor from the distributed representation. Combined, the state update can be written as,
3.4. Extension to multiple objects
We now consider how the model described above can
be extended to handle multiple interacting objects. This is
more challenging as it requires to handle complex object
interactions such as collisions.
In order to do so, for each object oi , i = 1, . . . , Nobjects
we consider a separate copy of the distributed state tensor
oN
sot i (hence the overall state is st = (sot 1 , . . . , st objects )). The
encoder network φenc is thus modified to output a H × W ×
Nobjects C tensor. It is then split along the third dimension to
produce H × W × C tensor for each of the Nobjects . We
order objects w. r. t. their color so that each feature is always
responsible for the same object identified by its color. We
recall here that this extension studies the ability of handling
collisions of our model without any explicit module. We aim
in the future to build more object agnostic representation.
The input of the transition module is also modified to take
into account the interaction between objects. Focusing on an
o
object of with state st f , the update is written as
(st+1 , pt+1 ) = φtrans (st , pt ) = (φs (st ), pt + φp (st )).
Inspired by the work of [Watters et al., 2017], we also consider an alternative architecture where pt is estimated directly
from st rather than incrementally. In order to do so, the location x and y of each pixel is appended as feature channels
C+1 and C+2 of the distributed state tensor st , obtaining an
augmented tensor augxy (st ). Then the object’s position pt
is estimated by a two-layer perceptron pt = φp (augxy (st )).
3.3. Decoder Map: from state to probabilistic predictions
For deterministic models, the projected object position
pt is part of the neural network state, the decoder map ŷt =
φdec (st , pt ) = pt simply extracts and returns that part of the
state. Training optimizes the average L2 distance
between
PT −1
ground truth yt and predicted ŷt positions T1 t=0 kŷt −
yt k2 .
In addition to this simple scheme, we also consider a more
robust variant based on probabilistic predictions. In fact, the
extrapolation error accumulates and increases over time, and
the L2 -based loss may be dominated by outliers, unbalancing learning. Hence, we modify the model to explicitly and
dynamically express its own prediction uncertainty by outputting the mean and variance (µt , Σt ) of a bivariate Gaus2
sian observation model. The L
is thus replaced with the
PTloss
−1
1
negative log likelihood − T t=0 log N (yt ; µt , Σt ) under
this model.
In order to estimate the Gaussian parameters µt and Σt ,
we extend the state component pt = (µt , λ1,t , λ2,t , θt ) to
include both the mean as well as the eigenvalues and rotation
of the covariance matrix Σt = R(θt )| diag(λ1,t , λ2,t )R(θt ).
In order to ensure numerical stability, eigenvalues are constrained to be in the range [0.01, 100] by setting them as the
output of a scaled and translated sigmoid λi,t = σλ,α (βi,t ),
o
o
f
st+1
= φs st f ,
X
sot i
i6=f
where the second argument is the sum of the state subtensors
for all other objects. Since the function φs is the same for
all objects of , this ensures that object interactions are symmetric and commutative. Note that, as opposed to methods
such as [Chang et al., 2017], no explicit collision detection
module is implemented here. Instead, handling collisions is
left to the discretion of the network.
s2
t
s2
t+1
s1
t
s3
t
Σ
φs
Table 1. Neural network variants.
Name
DispNet
ProbNet
PosNet
pt regression
incremental
incremental
direct
pt+1
pt + φp (st )
pt + φp (st )
φp (st )
output and loss
deterministic
probabilistic
deterministic
Figure 2. Multiple object extension. For each object (here object
2) we concatenate the state of this object with the addition of the
other objects features. We then give this tensor to the module φs to
obtain our new state s2t+1
5
Under review at Computer Vision and Image Understanding Journal.
With this modification, the transition subnetwork is illustrated in Fig. 2. The rest of the pipeline is essentially the
same as before and is applied independently to each object.
The same network parameters are used for each application
of a module regardless of the specific object.
texture attached to its surface, so it appears as a painted object. The texture is used to observe the object rotation. We
study the impact of being able to visually observe rotation
by re-rendering the single ball experiments with a uniform
white color (see Table 3). In the multi-object experiments,
instead, each ball has a constant, distinctive diffuse color (intensity 0.8) with Phong specular component (intensity 0.5).
We initially position the ball at angles (θ, φ) with respect to
the the bowl center, where the elevation θ is uniformly sampled in the range θ ∈ U [−9π/10, −π/2] and the azimuth
φ ∈ U [−π, π]. The minimum elevation is set to −9π/10
to avoid starting the ball at the bottom of the bowl. Due to
friction, at the end of each experiment the ball rests at the
bottom of the bowl.
The initial orientation of the ball (relevant for the multicolored texture) is obtained by uniformly sampling its
xyz Euler angles in [−π, π]. The ball’s initial velocity
v is obtained by first sampling vx , vy uniformly in the
range U [5, 10], assigning each of vx , vy a random sign
(∼ 2B (0.5)−1), and then by projecting the vector (vx , vy , 0)
to be tangential to the bowl’s surface. In the multi-object
‘Ellipsoidal bowl’ scenario, in order to achieve more interesting motion patterns, the magnitude of the initial velocities
is set uniformly in the range U [10, 15]; if, after simulation,
a ball leaves the bowl due to a collision or excessive initial
velocity, the scene is discarded. Sequences are recorded
until all objects stop moving. Short sequences (less than 250
frames) are discarded as well. The average angular velocity
computed over all ‘Bowl’ scenes was 5.94 radian/s.
4. Experimental Setup
In the experiments (Fig. 3), we consider two physical scenarios: spheres rolling on a 3D surface, which can be either
a semi-ellipsoid with random parameters or a continuous
randomized heightfield. When the semi-ellipsoid is isotropic
(i.e. a hemisphere) we refer to it as ‘Hemispherical bowl’,
and in the more general case as ‘Ellipsoidal bowl’ (see Table 2), whereas the heightfield scenario is referred to as
‘Heightfield.’
4.1. Hemispherical bowl and Ellipsoidal bowl scenarios
The symbol p = (px , py , pz ) ∈ R3 denotes a point in 3D
space or a vector (direction). The camera center is placed at
location (0, 0, cz ), cz > 0 and looks downward along vector
(0, 0, −1) using orthographic projection, such that the point
(px , py , pz ) projects to pixel (px , py ) in the image.
(0, 1, 1)
z
y
(0, 0, cz )
(0, 0, 0)
(a, 0, 1)
Ellipse
x
(a)
(b)
(c)
(a)
Figure 3. Problem setup. We consider the problem of understanding and extrapolating mechanical phenomena with recurrent deep
networks. (a) Experimental setup: an orthographic camera looks at
a ball rolling in a 3D bowl. (b) Example of a 3D trajectory in the
3D bowl simulated using Blender 2.78’s OpenGL renderer. (c) An
example of a rendered frame in the ‘Ellipsoidal bowl’ experiment
that is fed to our model as input.
Ellipsebowl
Ellipsoidal
Heightfield
1.00
0.86
1.00
0.71
0.86
(b)
Heightfield
Heightfield
1.00
0.88
1.00
0.75
0.88
(c)
camera
0.62
0.75
0.50
0.62
0.57
0.71
0.43
0.57
0.38
0.50
0.25
0.38
0.29
0.43
0.14
0.29
0.12
0.25
0.00
0.12
0.00
0.14
Network input and contours 0.00Network input and contours 0.00
Thus, the ‘Ellipsoidal bowl’ is the bottom half of an ellipsoid of equation x2 /a2 + y 2 + (z − 1)2 = 1 with its
axes aligned to the xyz axes and its lowest point corresponding to the origin. For the ‘Ellipsoidal bowl’ scenario, the
ellipsoid shape is further varied by sampling a ∈ U [0.5, 1]
for the (a = 1 for the ‘Hemispherical bowl’ scenario) and
by rotating the resulting shape randomly around the z-axis.
Both ‘Hemispherical bowl’ and ‘Ellipsoidal bowl’ are rendered by mapping a checker board pattern to their 3D surface
(to make it visible to the network).
The rolling object is a ball of radius ρ ∈ {0.04, 0.225}.
The ball’s center of mass at time t is denoted as qt =
(qxt , qyt , qzt ), which, due to the orthographic projection, is
imaged at pixel (qxt , qyt ). The ball has a fixed multi-color
Network input and contours
Network input and contours
Network input and contours
Network input and contours
Heightfield simulation setup
Figure 4. Experimental setups. (a) ‘Ellipsoidal bowl’ experiment
setup, depth map on the top, network input with isocontours at
the bottom. We create the dataset by varying the ellipsoid’s main
axis ratio and orientation, and the starting position and velocity
of the balls. (b-c) ‘Heightfield’ rendering setup. Each sequence
is generated using a random translation and rotation of the fixed
heightfield geometry. Walls ensure the automatically generated
sequences are long enough. A randomly positioned area light
presents additional generalization challenges to the network.
6
Under review at Computer Vision and Image Understanding Journal.
Note that, while some physical parameters of the ball’s
α
state are included in the observation vector y[−T
, these
0 ,T )
are not part of the state h of the neural network, which is
inferred automatically. The network itself is tasked with
predicting part of these measurements, but their meaning is
not hardcoded.
is visible under the static camera. In order to generate motion sequences of enough length for long-term extrapolation,
we also surround the camera frustum with perfectly elastic
walls (see Fig. 4c). The noise texture has a scale parameter,
which we vary between 0.7 (fairly planar) and 0.2 resulting in high curvature surfaces that have holes comparable
with the ball diameter. We set the surface elasticity to 0 in
order to encourage the balls to roll and not bounce. The
initial placement of the ball, similarly to the bowl case, is
drawn from a 2D uniform distribution. Then, we use sphere
tracing to push the ball onto the surface from the camera
plane. We add a small random initial velocity (U [2, 4]), and
similarly to the ‘Hemispherical bowl’ case, we project the
initial velocity onto the local surface normal. The average
angular velocity computed over all ‘Heightfield’ scenes was
2.8 radian/s. The surface is lit with a small (0.1 × 0.1) area
light from a random location. We draw the 2D position of
the light as x, y ∼ (2B (0.5) − 1) (U [1, 1.5] × U [1, 1.5]),
with a fixed camera height z = 2.
Simulation details
For efficiency, we extract multiple sub-sequences xα
[−T0 ,T )
form a single longer simulation (training, test, and validation
sets are however completely independent). The simulator
runs at 120fps for accuracy, but the data is subsampled to
40fps. We use Blender 2.78’s OpenGL renderer and the
Blender Game Engine (relying on Bullet 2 as physics engine). The ball is a textured sphere with unit mass. The
simulation parameters were set as: max physics steps = 5,
physics substeps = 3, max logic steps = 5, fps = 120. Rendering used white environment lighting (energy = 0.7) and
no other light source in the ‘Hemispherical bowl’ case, environment energy = 0.2, and a spotlight at the location of the
camera in the ‘Ellipsoidal bowl’ case. We used 70% the data
for training, 15% for validation, and 15% for test, 12500 sequences in the ‘Hemispherical bowl’/‘Ellipsoidal bowl’ experiments and 6400 in the ‘Heightfield’ case. During training, we start observation at a random time while it is fixed
for test. The output images were stored as 256 × 256 color
JPEG files. For multiple objects in the ellipsoid experiment,
we set the elasticity parameter of the balls to 0.7 in order to
get a couple of collisions before they settle in the middle of
the scene.
5. Results and Discussions
5.1. Baselines
(i) Least squares fit We compare the performance of our
methods to two simple least squares baselines: Linear and
Quadratic. In both cases, we fit least squares polynomials
to the screen-space coordinates of the first T = 10 frames,
which are not computed but given as inputs. The polynomials
are of first and second degree(s), respectively. Note, that
being able to observe the first 10 frames is a large advantage
compared to the networks, which only see the first T0 = 4
frames.
4.2. Heightfield scenario
(ii) NPE The NPE method and its variants were trained
using available online code. We used the same training procedure as reported in [Chang et al., 2017]. Additionally, we
added angular velocities as input and regressed type of parameter. In the case of the Ellipsoidal bowl, both scaling and
bowl rotation angle are also given as input to the networks.
In this case NPE’s method carries forward the estimated
states via the network.
An important part of our experiments involve randomly
generated continuous heightfields. Long-term motion prediction on random heightfields represent a tougher challenge,
since solely observing the motion of the object at the beginning of the sequence does not contain enough information
for successful mechanical predictions. In contrast to the
‘Ellipsoidal bowl’ cases, where the 2D shape that the container occupies in the image is theoretically enough to infer
the analytical shape of the local surface at any future 3D
point of interest, in the ‘Heightfield’ case the illumination
conditions of the surface have to be parsed. Furthermore, a
more elaborate understanding about the interaction between
surface and 3D rolling motion has to be developed.
Similar to the ellipsoid cases, we generate randomized
sequences of a ball rolling on a random (heightfield) surface. We approximate random heightfields by generating a
large (8 × 8) Improved Perlin noise texture and applying
it as a displacement map to a highly tessellated plane. For
each scene, we uniform randomly rotate and translate the
plane so that a different part (2.5 × 2.5) of the heightfield
(iii) VIN and IN From State (IFS) Finally, we used
VIN network and its state variant IN From State from
[Watters et al., 2017]. IFS is essentially a version of VIN
where the propagation mechanism is the same but the first
state vector is not deduced from visual observation but given
as ground truth position and velocity as in the NPE. The VIN
network uses downscaled 32 × 32 images. Both networks
use training procedures as reported in [Watters et al., 2017]
with the exceptions that for IFS the learning rate was updated
using our method (see section 5.2) and we rely on the first 4
states and 16 rolled out steps. As with NPE, angular velocity
7
Under review at Computer Vision and Image Understanding Journal.
Figure 5. Errors in bowls. Pixel errors and angular velocity RMSE in radian/s (first two columns of Table 2). Our method performs
comparably to state based methods, which use ground truth state information for initialization compared to ours, which operates with visual
input. Hatched denotes non-visual input (i.e. direct access to physical states).
was also added to IFS input and regressed parameters. Scaling and rotation angle of the bowl were also given as input
to the network in Ellipsoidal bowl experiment. Note that
VIN and our models work with images as direct observation
of the world rather than perfect states, which represents a
much more difficult problem whilst yielding a more general applicability. Physical properties are then deduced from
the observations and integrated through our Markov model.
Thus, these methods do not need a simulator to estimate
parameters of the physical worlds (such as scaling and rotation angle) and can be trained on changing environments
without requiring additional external measurements of the
underlying 3D spaces.
ments of the loss have been found after 100 consecutive
epochs. Training is halted when the loss has not decreased
after 200 successive epochs; 2,000 epochs were found to be
usually sufficient for convergence. In every case the loss is
the sum of the L2 angular velocity loss and either L2 position errors (PosNet, DispNet) or likelihood loss (ProbNet)
(see section 3.3). We omit the angular loss, when angular
velocity is not regressed (labelled as “* w/o ang. vel.” in the
tables).
Since during the initial phases of training the network is
very uncertain, the model using the Gaussian log-likelihood
loss was found to get stuck on solutions with very high
variance
Σ(t). To address this, we added a regularizer
P
λ t det Σ(t) to the loss, with λ = 0.01.
In all our experiments we used Tensorflow
([Abadi et al., 2015]) r1.3 on a single NVIDIA Titan
X GPU.
5.2. Results
Implementation details The encoder network φenc is obtained by taking the ImageNet-pretrained VGG16 network
of [Simonyan and Zisserman, 2015] and retaining the layers up to conv5 (for an input image of size (Hi , Wi ) =
(128, 128, 3) this results in a (8, 8, Nf = 512) state tensor
st ). In the 3 balls experiments, we replaced the last conv5
layer with a convolutional layer of output 256 × 3 channels.
Object features are thus obtained by splitting this last tensor
along the channel dimension into (8, 8, Nf = 256) state tensor per object. The filter weights of all layers except conv1
are retained for fine-tuning on our problem. The conv1
is reinitialized as filters must operate on images with 3T0
channels. The transition network φs (st ) uses a simple chain
of two convolution layers with 256 and Nf filters respectively, of size 3 × 3, stride 1, and padding 1 interleaved by
a ReLU layer. Network weights are initialized by sampling
from a Gaussian distribution. Additionally, angular velocity
is always regressed from the state st using a single layer
perceptron.
Training uses a batch size of 50 using the first Ttrain positions and angular velocity (or only position when explicitly mentioned) of each video sequence using RMSProp by
[Tieleman and Hinton, 2012]. We start with a learning rate
of 10−4 and decrease it by a factor of 10 when no improve-
5.2.1
Extrapolation
(i) Experiments using a single ball Table 2 compares the
baseline predictors and the eight networks on the task of
long term prediction of the object trajectory. All methods observed only the first T0 = 4 inputs (either object
states or simply image frames) except for the linear and
quadratic baselines, and aimed to extrapolate the trajectory
to Tgen = 40 time steps. Predictions are “long term” relative to the number of inputs T0 Tgen . Note also that
during training networks only observe sequences of up to
Ttrain ≤ Tgen frames; hence, the challenge is not only to
extrapolate physics, but to generalize beyond extrapolations
observed during training.
Table 2 reports the average errors at time Ttrain = 20
and Tgen = 40 for the different estimated parameters. Our
methods outperform state-only approaches for predictions
of up-to Ttrain steps. For example, PosNet has a pixel error
of 1.0/1.2/6.8 in the Hemispherical/Ellipsoidal/Heightfield
scenarios vs 3.3/2.7/10.9 of NPE and 1.6/3.1/8.7 of IFS. This
is non-trivial as our networks know nothing about physical
8
Under review at Computer Vision and Image Understanding Journal.
Table 2. Long term predictions. All of our models (below thick line) observed the T0 = 4 first frames as input. All networks have been
trained to predict the Ttrain = 20 first positions, except for the NPEs which were given T0 = 4 states as input and train to predict state at time
T0 + 1. We report here results for time Ttrain = 20 and Tgen = 40. Unless noted, reported models are trained to predict position and angular
velocity. For each time we report on the left average pixel error and root squared L2 angular velocity loss on the right. Perplexity (loge
values shown in the table) is defined as 2−E[log2 (p(x))] where p is the estimated posterior distribution. This value is shown in bracket.
Method
State
Linear
Quadratic
NPE w/o ang. vel.
NPE
IFS w/o ang. vel.
IFS
VIN w/o ang. vel.
PosNet w/o ang. vel.
PosNet
DispNet w/o ang. vel.
DispNet
ProbNet w/o ang. vel.
GT
GT
GT
GT
GT
GT
Visual
Visual
Visual
Visual
Visual
Visual
ProbNet
Visual
Hemispherical bowl
Errors (Perplexity)
Ttrain
Tgen
pixel
39.2
164.3
2.6
3.3
1.3
1.6
40.4
1.0
1.0
3.0
3.5
2.9
(4.5)
3.4
(4.7)
ang. vel.
7.5
18.4
–
0.8
–
0.3
–
–
0.4
–
1.2
–
1.2
pixel
127.5
120.1
6.0
9.6
2.9
2.2
37.8
18.1
13.8
29.7
15.9
24.2
(21.9)
15.3
( 9.2)
Ellipsoidal bowl
Errors (Perplexity)
Ttrain
Tgen
ang. vel.
17.9
861.2
–
1.7
–
0.4
–
–
3.0
–
4.3
–
3.4
pixel
61.9
11.7
3.2
2.7
3.3
3.1
24.0
1.6
1.2
2.5
2.1
2.9
(32.1)
4.0
(4.5)
ang. vel.
23.3
14.8
–
1.4
–
1.0
–
–
0.5
–
1.0
–
1.8
pixel
20.1
93.1
6.1
7.6
8.9
6.9
30.2
24.4
11.8
20.6
16.1
21.8
(54.0)
16.7
(9.3)
Heightfield
Errors (Perplexity)
Ttrain
Tgen
ang. vel.
80.0
70.6
–
2.9
–
1.4
–
–
3.0
–
4.4
–
3.8
pixel
21.3
26.7
12.0
10.9
10.4
8.7
42.6
7.2
6.8
7.7
7.2
6.4
(9.5 )
6.8
(10.8)
ang. vel.
9.4
27.4
–
3.7
–
2.5
–
–
2.1
–
2.0
–
2.1
pixel
61.9
126.0
38.5
32.9
27.6
26.1
42.7
24.6
23.2
25.8
21.6
22.5
(12.7)
20.5
(12.3 )
ang. vel.
19.3
122.2
–
4.6
–
2.8
–
–
4.2
–
3.3
–
2.7
laws a-priori, and observe the world through images rather
than being given the initial ground-truth state values. On the
other hand, our methods can, through images, better observe
and hence model the underlying environments. The gap in
the heightfield results, in particular, shows the value in observing the environments in this manner as we constantly
out-perform state-only methods. Our methods also shown to
make significantly better predictions compared to the other
visual competitor (VIN) which failed to be able to model
interactions between the object and its environment and performed poorly even on training regimes (40.4/24.0/42.6 in
the Hemispherical/Ellipsoidal/Heightfield scenarios respectively).
An interesting question is whether the model learns or
not to measure angular velocity from images, or whether
predicting this quantity during simply induces a better internal understanding of physics. To tease this effect out, we
prevent the network from observing the ball spin by removing the texture on the ball. Table 3 shows that this results
approximately in the same accuracy as the textured cases,
indicating that angular velocity is not estimated visually.
Finally, introducing the probability-based loss in DispNet
results in the ProbNet network. As shown in Table 2, This
change significantly outperforms the deterministic DispNet
results in most cases.
All methods can perform arbitrary long predictions. Our
networks, which are only trained to predict the first Ttrain positions, are still competitive with state-only methods (which
only predict a transition function and hence implicitly generalize to arbitrarily lengths) even when predictions are generalized to Tgen steps. In particular, while performances
around Tgen deteriorates, PosNet provides very promising results, reaching nearly state-only models performances on the
‘Ellipsoidal bowl’ experiments (11.8 pixel prediction error
vs 6.1 of NPE).
(ii) Experiments using multiple balls We also trained our
models with two and three balls in the ‘Ellipsoidal bowl’ environment to study the ability of our models to handle object
interactions without explicit collision modules. The aforementioned training setups are maintained in these experiments. Quantitatively, Table 4 shows that our models were
able to get competitive results w. r. t. state-only methods
Table 3. Impact of ball texturing on prediction. We compare the
impact of ball texturing on predictions. Table layout and measures
are same as Table 2. Results show that ball texture is rather ignored
to make predictions.
We also note that learning to regress angular velocity
generally improve the ability of our models to predict position, in particular when generalizing to Tgen steps. For
example, PosNet in the Ellipsoidal bowl reduces its position
error from 24.4 to 11.8 at Tgen when it is required to predict
angular velocity during training. For further comparisons,
see the similarly colored, adjacent bars in Fig. 5 (left) and
Fig. 7 (left)). This is remarkable as angular velocity as such
remains very challenging to predict.
Method
PosNet w/o ang. vel.
PosNet
DispNet w/o ang. vel.
DispNet
ProbNet w/o ang. vel.
ProbNet
9
Ellipsoidal bowl
Errors (Perplexity)
Ttrain
Tgen
pixel
1.6
1.2
2.5
2.1
2.9
(32.1)
4.0
(4.5)
ang. vel.
–
0.5
–
1.0
–
1.8
pixel
24.4
11.8
20.6
16.1
21.8
(54.0)
16.7
(9.3)
Ellipsoidal bowl (no ball texture)
Errors (Perplexity)
Ttrain
Tgen
ang. vel.
–
3.0
–
4.4
–
3.8
pixel
1.6
1.1
1.7
1.6
3.1
(5.0)
4.3
(4.5)
ang. vel.
–
0.6
–
1.0
–
1.3
pixel
23.7
12.7
26.3
16.2
24.0
(12.7)
15.0
(8.2)
ang. vel.
–
3.5
–
3.8
–
3.5
Under review at Computer Vision and Image Understanding Journal.
5.3. Discussion
containing explicit collision modules, e.g., NPE. Probabilistic model shows an increase in uncertainty at Ttrain , which
reveals that the task to solve were harder due to the chaotic
nature of the system. In addition, angular velocity seems to
be very challenging to estimate in this case. Qualitatively,
Section 5.2.1 shows that collisions are well handled by our
model despite not being explicitly encoded.
In addition to the various results we presented, we discuss
our conclusions regarding the main sources of prediction
error in the conducted experiments.
Does training for longer horizons help? Training for longer
horizons Ttrain = 40 in Table 5 compared to Ttrain = 20
in Table 2 results in better position estimates as expected.
When a single end state is also observed (interpolation) the
model manages to infer plausible trajectories even though
the initial and final states are far apart in time.
This motivates us to design more structured representations in the future, which would generalize even better
outside the supervised time spans (see Table 6).
Table 4. Multiple balls experiment.
We extend the
‘Ellipsoidal bowl’ setup adding more balls. We show that in this
case our networks get comparable performances to state-only methods. Table layout and measures are the same as Table 2 except that
Ttrain = 15 and Tgen = 30.
Method
States
NPE
IFS
GT
GT
PosNet
DispNet
ProbNet
Visual
Visual
Visual
5.2.2
Ellipsoidal bowl 2 balls
Errors (Perplexity)
Ttrain
Tgen
pixel
5.3
4.1
ang. vel.
1.5
1.3
pixel
13.4
9.6
4.2
3.9
5.3
(7.0)
2.4
2.4
2.5
11.7
18.5
19.8
(14.0)
Ellipsoidal bowl 3 balls
Errors (Perplexity)
Ttrain
Tgen
ang. vel.
2.0
1.5
2.8
4.3
3.6
pixel
5.0
4.3
ang. vel.
1.6
1.5
pixel
13.3
10.0
5.7
6.2
6.5
(7.5)
4.0
3.9
3.9
15.6
17.4
17.1
(12.6)
ang. vel.
2.0
1.6
Can the models handle collisions of multiple objects?
Adding additional objects to our scenes has appeared to
be a challenging task for our models. Training for longer
time horizons did not decrease the errors significantly in
these cases, which shows that collisions remain difficult
to estimate. Promisingly, InterpNet manages to improve
performance similarly to the earlier cases, the remaining
ambiguity in the middle of the sequences matches the
ratios of single object examples (ErrorT =10 /ErrorT =20 :
1.0/1.6 ' 3.2/4.5 ' 3.3/4.5 in Table 5 middle columns).
4.5
4.6
4.1
Interpolation
So far, we have consider the problem of extrapolating trajectories without any information on the possible final state
of the system. We aim here to study the impact of injecting
such knowledge in our networks.
In order to do so, in this experiment we concatenate to
the first T0 = 4 input frames the last observed frame at time
Tfinal and give the resulting stack as input to the encoder
network h0 = φenc (x(−T0 ,0] , xTfinal ) to estimate the first state
h0 . In this setting, the model performs “interpolation” as
it sees images at the beginning as well as the end of the
sequence. The rest of the model works as before with the
exception that the first state h0 is decoded in a prediction
(y0 , yTfinal ) = φdec (h0 ) of both the first and the last position
yTfinal ; in this manner, the loss encourages state h0 to encode
information about the last observed frame xTfinal .
Table 5 indicates that the ability of observing an image
of the final state enables our models to provide far better
estimations. Even in the more complex scenarios with 2
and 3 balls and the heightfield experiments, the errors are
significantly lower than for extrapolation. As expected, for
InterpNet the highest errors are always found in the middle
of the estimate as these points are less predictable from the
available information; by contrast, for DispNet the highest
errors are at the end.
Still, we note that harder scenarios result in larger errors
even for interpolation, and particularly for colliding balls
due to the chaotic nature of this dynamics. This also shows
the current limitation of our system in modeling collisions
and complex variable environments.
Does regression of angular velocity help? Almost all models
benefit from the additional supervision signal coming from
the loss on angular velocity, as shown in Fig. 5(left) and
Fig. 7 (left). The objects’ texture at these resolutions is
difficult to interpret, and the connection between pixel
color and rotation around axis is highly non-linear, which
encourages us to look for a different representation of
rotation in the future to improve our angular prediction
errors.
Are changing environments more difficult? The characteristics of the environment also appear to strongly contribute
to the final estimation errors. When only following one
ball we notice that for simple shapes where the environment parameters can vary along at most 3 dimensions (in the
‘Hemispherical bowl’ and ‘Ellipsoidal bowl’ cases), the system can obtain nearly perfect estimates in the interpolation
experiments. However in the ‘Heightfield’ scenes interaction
with the environment is much more difficult to estimate and
the maximum errors are larger, even for InterpNet the errors
remains substantial.
6. Conclusions
In this paper, we studied the possibility of abstracting
knowledge of physics using a single neural network with a
recurrent architecture to model long term predictions with a
10
Under review at Computer Vision and Image Understanding Journal.
NPE
(a)
30
3030 2020
302020
1 10
10
IFS
30
20
30 30
20
30
20 20
10
10
1 10
1
10
10
10
10
1
20
20
30
30
DispNet
30
20
30
20 20
30
3020
10
10
1 10
10
1
10
10
1
20 30
20
30
PosNet
30
20
30
20
30 20
30 20
1 10
10
1
10
10
1
20
20 30
30
10
10
ProbNet
30
3030 20
20 20
20
30
1 10
1
10
10
10
11
11
20
30
20
30
10
10
1
10
2030
20
30
11
(b)
30
30
30
20
20
10
20
10
1 30
10
10
30
20
3020 1
10
10
1
(c)
2020
1
30
30
30
20 30 30
20
10 20
13020 10
10
10
30
20
1
30 20
10
10
1
10
10
30
30
201
10
10
20
30
10
30
20
20
20
20
1
30
30
10
10
1
30
201
30
10
10
20
10
30
10
302020
30
30 20
20
20 3030
20
10 10
1
10
3030
20
11
20
10
1
20
20
1
30
30
10
10
1
30
102011
10
30
20
10
10
30
20
20
30
30
30
20
3020
20
10 20
130 10
10
10
30
20
1
3020 1
10
10
11
2020
1
10
20
20
1
30
1
30
1
10201
10
30 20
30
10
30
10
20
20
30
10
10
1
30
102011
10
30 30
30 2030
10
10
30
20
20
30
30
2030 20
20
0 10
10230
10
11
10
30
30
20
1
20
10
10
1
1
1
(d)
30
40
2040
30
2010
10
(e)
(f)
40
30
1
40
30
20
20
1030
40
1
40
20
40
30
20
10
10
1
40
40
30
30
20
20
10
1
40
30 20
2010
30
40
1
NPE
40
IFS
40
30
20
10
20 10
11
30
40
11
40
4030
2030
20
10
11
40
30
20
30 20
40
30
20
3020
10
10
40
10
1
40 40
3030
20
20
10
1
20
30 30
20
40
PosNet
40
30 3040
2020
10
10
10
10
1
DispNet
1
40
30
40
30
20
20
10
1
40 40
30
30
20
20
40
10
10
1
ProbNet
40
30
30 2020
10
10
1
GT
Figure 6. Ellipsoidal bowl and Heightfield extrapolations. (a-c) Example scene from the 3 balls in the ‘Ellipsoidal bowl’ experiment.
Extrapolation on multiple objects generalises well to 3 objects. Note how in (b) the collision of the red and green ball is predicted by our
networks, solely by seeing the first 4 frames of the sequence. Remember, NPE and IFS start with the ground truth knowledge of the physical
state of the objects. (d-f) Our models, taking only 4 images as input, have learned to parse the illumination of a quickly changing heightfield
surface and use it to predict the long-term (up to 10x the length of initial observation) motion of an object. (d) For homogeneously lit flat
regions, it is difficult to make decisions, indicated by ProbNet’s large uncertainty estimates. (e) IFS, DispNet and PosNet correctly interpret
the ball’s initial angular velocity to predict the future path. ProbNet demonstrates the power of anisotropic uncertainty estimation (c, f). It is
more certain in the direction of motion than orthogonal to it. Note, that NPE and IFS were given the ground truth object positions for the
first four frames, and do not have the capability to take images as input.
11
Under review at Computer Vision and Image Understanding Journal.
Figure 7. Errors on Heightfields. Position errors (left) and angular velocity error (right) for trained (Ttrain = 20) and untrained (Tgen = 40)
generalization on increasing difficulty heightfields (’Mean’ is reported in the right column of Table 2). Note, how angular velocity estimation
helps position accuracy. Hatches denote non-visual methods.
Table 5. Extrapolation vs Interpolation. We constructed InterpNet as an extension of DispNet, where in addition to the concatenation
of the first T0 = 4 frames, also the last frame at Tf inal is provided to the model as inputs. All networks have been trained to predict the
Ttrain := Tf inal positions. As expected, InterpNet learned to predict the positions at Tf inal by relying on the features extracted from the last
input image. We report the pixel errors at different times along the sequences. Tf inal is the last value shown for every experiment.
Method
DispNet
InterpNet
Hemispherical bowl
pixel error, Ttrain = 40
T=10 20
30
40
2.2
3.6
3.9
5.0
1.4
1.8
1.6
1.0
Ellipsoidal bowl 1 ball
pixel error, Ttrain = 40
T=10 20
30
40
1.4
2.4
2.7
3.0
1.0
1.6
1.3
0.6
Ellipsoidal bowl 2 balls
pixel error, Ttrain = 30
T=10 20
30
2.8
5.8
8.7
3.2
4.5
3.1
Table 6. Length of supervision. The maximum position error of
DispNet decreases when we add more supervision during training.
Dataset
‘Hemispherical bowl’
‘Ellipsoidal bowl’
‘Heightfield’
Extrapolation
Ttrain = 20 Ttrain = 40
15.9
5.0
16.1
3.0
21.6
17.9
Ellipsoidal bowl 3 balls
pixel error, Ttrain = 30
T=10 20
30
3.2
8.1
12.0
3.3
4.5
2.1
Heightfield
pixel error, Ttrain = 40
T=10 20
30
40
3.6
7.9 12.9 17.9
2.5
5.2
5.1
1.6
resentation of the system state which, at the same time, is
still able to model concentrated object interactions such as
collisions.
Interpolation
Ttrain = 40
1.8
1.6
5.2
While keeping the same architecture, we further demonstrate that it is possible to remove ambiguity by showing
the network an image of the final state of the system, performing interpolation. However, in this case the internal
state propagation mechanism is still limited by its ability to
make accurate long term predictions outside temporal spans
observed during training.
changing environment. We compared our model to various
baselines on the non-trivial motion of ball(s) rolling on a
surfaces with different possible shapes (e.g. ellipsoidal bowls
or randomized heightfields). As opposed to some concurrent approaches, we do not integrate physical quantities but
implicitly encode the states in a feature vector that we can
propagate through time.
In the future, we aim at increasing the robustness and
generalization capabilities of our models by enforcing more
explicitly temporal and spatial invariance (as physical laws
are constant and homogeneous). Another important step will
be to test our framework on real video footage to assess the
ability of the predictions to generalize to situations affected
by real-world nuisance factors.
Our experiments on synthetic simulations indicate that
our networks can predict mechanical phenomena more accurately than networks that build on hand-crafted physicallygrounded representations of the system state. This means
that our approach can both infer automatically an internal
representation of these phenomena and work with visual inputs in order to initialize such a representation and use it for
extrapolation. Our models can also estimate a distribution
over physical measurements such as position to account for
uncertainty in the predictions.
References
[Abadi et al., 2015] Abadi et al. (2015). TensorFlow: Large-scale
machine learning on heterogeneous systems. Software available
from tensorflow.org.
[Agrawal et al., 2016] Agrawal, P., Nair, A. V., Abbeel, P., Malik,
J., and Levine, S. (2016). Learning to Poke by Poking: Experiential Learning of Intuitive Physics. In Proc. NIPS, pages
5074–5082.
We also demonstrated another significant difference with
existing networks, namely the ability to account for complex
variable environments. The latter leverage a distributed rep12
Under review at Computer Vision and Image Understanding Journal.
[Battaglia et al., 2016] Battaglia, P., Pascanu, R., Lai, M.,
Rezende, D. J., et al. (2016). Interaction networks for learning about objects, relations and physics. In Proc. NIPS, pages
4502–4510.
[Mottaghi et al., 2016] Mottaghi, R., Bagherinezhad, H., Rastegari, M., and Farhadi, A. (2016). Newtonian scene understanding: Unfolding the dynamics of objects in static images. In
IEEE CVPR.
[Battaglia et al., 2013] Battaglia, P. W., Hamrick, J. B., and Tenenbaum, J. B. (2013). Simulation as an engine of physical scene
understanding. PNAS, 110(45):18327–18332.
[Oh et al., 2015] Oh, J., Guo, X., Lee, H., Lewis, R. L., and Singh,
S. (2015). Action-conditional video prediction using deep networks in atari games. In Advances in Neural Information Processing Systems, pages 2863–2871.
[Chang et al., 2017] Chang, M. B., Ullman, T., Torralba, A., and
Tenenbaum, J. B. (2017). A compositional object-based approach to learning physical dynamics. In Proc. ICLR.
[Ondruska and Posner, 2016] Ondruska, P. and Posner, I. (2016).
Deep tracking: Seeing beyond seeing using recurrent neural
networks. In Proc. AAAI.
[Chiappa, 2017] Chiappa, Racaniere, W. M. (2017). Recurrent
environment simulators. In Proc. ICLR.
[R. Villegas, 2017] R. Villegas, J. Yang, Y. Z. S. S. X. L. H. L.
(2017). Learning to generate long-term future via hierarchical
prediction. In Proc. ICML.
[De Brabandere et al., 2016] De Brabandere, B., Jia, X., Tuytelaars, T., and Van Gool, L. (2016). Dynamic filter networks. In
Proc. NIPS.
[Silver et al., 2016] Silver, D., van Hasselt, H., Hessel, M., Schaul,
T., Guez, A., Harley, T., Dulac-Arnold, G., Reichert, D., Rabinowitz, N., Barreto, A., and Degris, T. (2016). The predictron:
End-to-end learning and planning. CoRR, abs/1612.08810.
[Denil et al., 2016] Denil, M., Agrawal, P., Kulkarni, T. D., Erez,
T., Battaglia, P., and de Freitas, N. (2016). Learning to perform
physics experiments via deep reinforcement learning. Deep
Reinforcement Learning Workshop, NIPS.
[Simonyan and Zisserman, 2015] Simonyan, K. and Zisserman, A.
(2015). Very deep convolutional networks for large-scale image
recognition. In International Conference on Learning Representations.
[Ehrhardt et al., 2017a] Ehrhardt, S., Monszpart, A., J. Mitra, N.,
and Vedaldi, A. (2017a). Learning A Physical Long-term Predictor. arXiv e-prints arXiv:1703.00247.
[Stewart and Ermon, 2017] Stewart, R. and Ermon, S. (2017).
Label-free supervision of neural networks with physics and
domain knowledge. In AAAI, pages 2576–2582.
[Ehrhardt et al., 2017b] Ehrhardt, S., Monszpart, A., Vedaldi, A.,
and J. Mitra, N. (2017b). Learning to Represent Mechanics
via Long-term Extrapolation and Interpolation. arXiv preprint
arXiv:1706.02179.
[Tieleman and Hinton, 2012] Tieleman, T. and Hinton, G. (2012).
Lecture 6.5—RMSProp: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for
Machine Learning.
[Fragkiadaki et al., 2016] Fragkiadaki, K., Agrawal, P., Levine, S.,
and Malik, J. (2016). Learning visual predictive models of
physics for playing billiards.
[Tompson et al., 2016] Tompson, J., Schlachter, K., Sprechmann,
P., and Perlin, K. (2016). Accelerating Eulerian Fluid Simulation
With Convolutional Networks. ArXiv e-print arXiv:1607.03597.
[Hamrick et al., 2016] Hamrick, J. B., Battaglia, P. W., Griffiths,
T. L., and Tenenbaum, J. B. (2016). Inferring mass in complex
scenes by mental simulation. Cognition, 157:61–76.
[Watters et al., 2017] Watters, N., Zoran, D., Weber, T., Battaglia,
P., Pascanu, R., and Tacchetti, A. (2017). Visual interaction
networks: Learning a physics simulator from video. In Guyon,
I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R., editors, Advances in Neural Information Processing Systems 30, pages 4542–4550. Curran
Associates, Inc.
[Hochreiter and Schmidhuber, 1997] Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory. Neural Comput.,
9(8):1735–1780.
[Krizhevsky et al., 2012] Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Proc. NIPS, pages 1097–1105.
[Wu et al., 2016] Wu, J., Lim, J. J., Zhang, H., Tenenbaum, J. B.,
and Freeman, W. T. (2016). Physics 101: Learning physical
object properties from unlabeled videos. In Proc. BMVC.
[Ladický et al., 2015] Ladický, Jeong, SoHyeon, Solenthaler, B.,
Pollefeys, M., Gross, M., et al. (2015). Data-driven fluid simulations using regression forests. ACM Trans. on Graphics (TOG),
34(6):199.
[Wu et al., 2015] Wu, J., Yildirim, I., Lim, J. J., Freeman, B., and
Tenenbaum, J. (2015). Galileo: Perceiving physical object
properties by integrating a physics engine with deep learning.
In Proc. NIPS, pages 127–135.
[Lerer et al., 2016] Lerer, A., Gross, S., and Fergus, R. (2016).
Learning physical intuition of block towers by example. In Proceedings of the 33rd International Conference on International
Conference on Machine Learning - Volume 48, pages 430–438.
[Xue et al., 2016] Xue, T., Wu, J., Bouman, K. L., and Freeman,
W. T. (2016). Visual dynamics: Probabilistic future frame synthesis via cross convolutional networks. In Proc. NIPS.
[Li et al., 2017] Li, W., Leonardis, A., and Fritz, M. (2017). Visual
stability prediction and its application to manipulation. AAAI.
[Monszpart et al., 2016] Monszpart, A., Thuerey, N., and Mitra, N.
(2016). SMASH: Physics-guided Reconstruction of Collisions
from Videos. ACM Trans. on Graphics (TOG).
13
| 1cs.CV
|
Computing Hitting Set Kernels By AC0 -Circuits
Max Bannach
Till Tantau
arXiv:1801.00716v1 [cs.CC] 2 Jan 2018
Institute for Theoretical Computer Science,
Universität zu Lübeck
Lübeck, Germany
{bannach,tantau}@tcs.uni-luebeck.de
Abstract
Given a hypergraph H = (V, E), what is the smallest subset X ⊆ V such that e ∩ X 6= ∅
holds for all e ∈ E? This problem, known as the hitting set problem, is a basic problem
in parameterized complexity theory. There are well-known kernelization algorithms for it,
which get a hypergraph H and a number k as input and output a hypergraph H ′ such that
(1) H has a hitting set of size k if, and only if, H ′ has such a hitting set and (2) the size of
H ′ depends only on k and on the maximum cardinality d of edges in H. The algorithms run
in polynomial time, but are highly sequential. Recently, it has been shown that one of them
can be parallelized to a certain degree: one can compute hitting set kernels in parallel time
O(d) – but it was conjectured that this is the best parallel algorithm possible. We refute this
conjecture and show how hitting set kernels can be computed in constant parallel time. For
our proof, we introduce a new, generalized notion of hypergraph sunflowers and show how
iterated applications of the color coding technique can sometimes be collapsed into a single
application.
1
Introduction
The hitting set problem is the following combinatorial problem: Given a hypergraph H = (V, E)
as input, consisting of a set V of vertices and a set E of hyperedges with e ⊆ V for all e ∈ E, find a
set X ⊆ V of minimum size that “hits” all hyperedges e ∈ E, that is, e ∩ X 6= ∅. Many problems
reduce to the hitting set problem, including the vertex cover problem (it is exactly the special
case where all edges have size |e| = 2) and the dominating set problem (a dominating set of a
graph is exactly a hitting set of the hypergraph whose hyperedges are the closed neighborhoods
of the graph’s vertices). The computational complexity of the hitting set problem is thus of
interest both in classical complexity theory and in parameterized complexity theory.
The first result on the parameterized complexity of the hitting set problem was an efficient kernelization algorithm for this problem restricted to edges of cardinality three [16]. This
was later improved to a kernelization for the d-uniform version (all hyperedges have size exactly d) [15], which is based on the so-called Sunflower Lemma [13]. We will later have a closer
look at this algorithm; at this point let us just summarize its main idea by “repeatedly find
sunflowers and replace them by their cores until there are no more sunflowers.” The Sunflower
Lemma tells us that this algorithm will stop only when the input graph has been reduced to a
kernel. The just-sketched kernelization algorithm is highly sequential, but Chen et al. [11] have
recently shown that it can be parallelized: Instead of reducing sunflowers one-at-a-time, one can
replace all sunflowers in a hypergraph by their cores simultaneously in constant parallel time.
This process only needs to be repeated d(H) = maxe∈E |e| times, leading to a parallel algorithm
running in time O(d(H)). However, there were good reasons to believe that this algorithm is
essentially the best possible (we will later discuss them) and Chen et al. conjectured that the
1
hitting set problem does not admit a kernelization algorithm running in constant parallel time
(that is, in time completely independent of the input graph).
Our Contributions. In the present paper we refute the conjecture of Chen et al. and show
that there is a constant parallel time kernelization algorithm for the hitting set problem:
Problem 1.1. pk,d -hitting-set
Instance: A hypergraph H = (V, E) and a number k ∈ N.
Parameter: k + d(H)
Question: Does H have a hitting set X with |X| ≤ k?
Theorem 1.2 (Main Theorem). There is a dlogtime-uniform AC0 -circuit family that maps
every hypergraph H = (V, E) and number k to a new hypergraph H ′ = (V, E ′ ) that has the same
size-k hitting sets as H, has d(H ′ ) ≤ d(H), and has |E ′ | ≤ f (k, d(H)) for some fixed computable
function f .
Let us stress at this point that the AC0 -family from the theorem really has a size that is
polynomial in the input length (no exponential or even worse dependency on the parameters) and
has a depth that is completely independent of the input. The hypergraph H ′ has the same vertex
set V as H – a feature shared by all hypergraphs considered in this paper that simplifies the
presentation. However, since V is still “large,” the circuit is not quite a kernelization
algorithm.
S
Fortunately, this is easy to fix by replacing the vertex set of H ′ by V ′ = e∈E ′ e, yielding the
following corollary:
Corollary 1.3 (Constant-Time Kernelization). There is a dlogtime-uniform AC0 -circuit family that computes a kernel for every instance for pk,d -hitting-set.
The theorem and corollary imply that all problems that can be reduced to pk,d -hitting-set
via a parameter-preserving AC0 -reduction admit a kernelization computable by an AC0 -circuit
family. This includes pk -vertex-cover, which is just pk,d -hitting-set with d fixed at 2; pk triangle-removal, where the objective is to remove at most k vertices from an undirected
graph so that no triangles remain; and also pk,deg -dominating-set, where we must find a
dominating set of size at most k in an undirected graph and we parametrize by k and the
maximum degree of the vertices.
Our proof of the main theorem requires the development of two new ideas, which we believe
may also be useful in other situations. The above-mentioned parallel kernelization algorithm
for the hitting set problem with runtime O(d(H)) essentially does the following: “Repeat d(H)
times: replace all sunflowers of size k + 1 by their cores” and the difficult task in each of the
d(H) iterations is to find the sunflowers. It turns out that this can be done in constant parallel
time using the color coding technique [2] and it has been shown in [3] and again in [11] that this
technique can be implemented in constant time. Our first idea for turning the circuits depth
from O(d) into O(1) is to collapse the color codings from the d rounds into a single application of
the color coding technique: Instead of applying color coding in each round to filter and describe
“objects,” we would like to apply one global application of color coding that already contains
the internal colorings and does away with the intermediate objects.
Unfortunately, there does not appear to be a simple (or any) way of actually collapsing the
colorings used when we “replace all sunflowers by their cores”: The coloring coding technique is
good at imposing requirements of the form “these objects must be disjoint,” but cannot impose
requirements of the form “these objects must be the same.” For this reason, as our second new
idea, we develop a generalization of the notion of a sunflower (which we dub “pseudo-sunflowers”)
that is tailored to the collapsing of color coding.
2
Related Work. The sequential kernelization algorithm for the hitting set problem based
on the Sunflower Lemma has been known for a longer time [15], but there have been recent
improvements that bring down the runtim to linear time [17]. A parallel version has recently
been studied by Chen et al. [11] and they show how kernels for pk,d -hitting-set can be computed
by circuits of depth O(d(H)). Chen et al. also conjecture that the circuit depth of O(d(H)) is
unavoidable (which we refute).
The results of this paper fit into the larger, fledgling field of parallel parameterized complexity
theory, which has already been studied both from a practical [1] and a theoretical point of
view [8]. First results go back to research on parameterized logarithmic space [7, 10, 14], since it
is known from classical complexity theory that problems that are solvable with such a resource
bound can also be parallelized. A more structured analysis of parameterized space and circuit
classes was later made by Elberfeld et. al [12], which addresses parallelization more directly.
Current research on parameterized parallelization – including this paper – focuses on constanttime computations, that is, on a parameterized analogue of AC0 [9, 11, 3, 4]. We remark that
many previous results (including several of the authors) boil down to showing that instead
of using a known reduction rule many times sequentially, one can simply apply it in parallel
“everywhere,” but “only once.” In contrast, the kernelization algorithm developed in the present
paper had no previous counterpart in the sequential setting.
Organization of This Paper. After a short section on preliminaries, in Section 3 we review
known kernelization algorithms for the hitting set problem – both the sequential ones and the
parallel one. In Section 4 we discuss the obstacles that must be surmounted to turn the known
parallel algorithm into one that needs only constant time. Towards this aim, we introduce
the notions of pseudo-cores and pseudo-sunflowers as replacements for the cores and sunflowers
used in the known algorithms. In Section 5 we then argue that these pseudo-sunflowers can be
computed in constant time by “collapsing” multiple rounds of color coding into a single round.
Full proofs can be found in the full version of the paper [5].
2
Preliminaries
A hypergraph is a pair H = (V, E) such that for all hyperedges e ∈ E we have e ⊆ V . We write
V (H) = V and E(H) = E for the vertex and hyperedge sets of H. Let d(H) = maxe∈E |e|.
Throughout this paper, all hypergraphs will always have the same vertex set V , which is the
input vertex set. For this reason, in slight abuse of notation, for two hypergraphs H1 = (V, E1 )
and H2 = (V, E2 ) we also write H1 ⊆ H2 for E(H1 ) ⊆ E(H2 ) and H1 ∪H2 for (V, E(H1 )∪E(H2 )).
Concerning circuit classes and parallel computations, we will only need the notion of ACcircuit families, which are sequences C = (C0 , C1 , C2 , . . . ) of Boolean circuits where each Ci is a
directed acyclic graph whose vertices are gates such that there are i input gates, the inner gates
are ∧-gates or ∨-gates with unbounded fan-in, or ¬-gates; and the number of output gates is
either 1 (for decision problems) or depends on the number of input gates (for circuits computing
a function). The size function S maps circuits to their size (number of gates) and the depth
function D maps them to their depth (longest path from input gates to output gates). When
D(Cn ) ∈ O(1) and S(Cn ) ∈ nO(1) hold, we call C an AC0 -circuit family. Concerning circuit
uniformity, all circuit families in this paper will be dlogtime uniform, which is the strongest
notion of uniformity commonly considered [6]: there is a dtm that on input of bin(i)# bin(n),
where bin(x) is the binary encoding of x, outputs the ith bit of a suitable encoding of Cn in at
most O(log n) steps.
Even though this paper is about a parallel kernelization algorithm, we will need only little
from the machinery of parallel parameterized complexity theory. We do need the following
notions: A parameterized problem is a pair (Q, κ) where Q ⊆ Σ∗ is a language and κ is a function
3
κ : Σ∗ → N that is computable by a dlogtime-uniform AC0 -circuit family. When we write down
a parameterized problem such as pk,d -hitting-set, the indices of “p” (for “parameterized”)
indicate which parameter function κ we mean. A kernelization for a parameterized problem
(Q, κ) is a function K that maps every instance x ∈ Σ∗ to a new instance K(x) ∈ Σ∗ such that
for all x ∈ Σ∗ we have (1) x ∈ Q ⇐⇒ K(x) ∈ Q and (2) |K(x)| ≤ f (κ(x)) for some fixed
computable function f .
A parameterized problem (Q, κ) lies in FPT if x ∈ Q can be decided by a sequential algorithm
running in time f (κ(x)) · |x|O(1) for a computable function f . The AC0 -analogue of FPT is the
class para-AC0 . It contains all problems (Q, κ) for which there is a circuit family (Cn,k )n,k∈N
such that for all inputs x we have C|x|,κ(x)(x) = 1 if, and only if, x ∈ Q, and D(Cn,k ) ∈ O(1) and
S(Cn,k ) ∈ f (k) · nO(1) . It is well-known that (Q, κ) ∈ FPT holds if, and only if, Q is decidable
and there is a kernelization for (Q, κ) that is computable in polynomial time. The same proof
as for the polynomial-time case also shows that we have (Q, κ) ∈ para-AC0 if, and only if, Q is
decidable and (Q, κ) has a kernelization that can be computed by an AC0 -circuit family. (We
stress once more that this means that the kernelization is a normal AC0 -circuit family, having
size S(Cn ) ∈ nO(1) .)
We will use the color coding technique a lot. First introduced in [2], it has recently been
shown to work in the context of constant time computations [3, 11]. The key observation
underlying this technique is the following: Suppose we are given a set of n elements and suppose
you have k special elements x1 , . . . , xk together with some specific colors c1 , . . . , ck for them
“in mind”. Then we can compute a set Λ of “candidate colorings” of all elements of the set
such that at least one λ ∈ Λ colors each “in mind” vertex xi with the “desired” color ci , that is
λ(xi ) = ci . Formally, the following holds (the original version of this lemma due to Alon et. al [2]
is equivalent to the statement below – only without any depth guarantees):
Fact 2.1 (Color Coding Lemma, [3]). There is a dlogtime-uniform family (Cn,k,c )n,k,c∈N of
without inputs such that each Cn,k,c
AC-circuits
1. outputs a set Λ of functions λ : {1, . . . , n} → {1, . . . , c} (coded as a sequence of function
tables) with the property that for any k mutually distinct x1 , . . . , xk ∈ {1, . . . , n} and any
c1 , . . . , ck ∈ {1, . . . , c} there is a function λ ∈ Λ with λ(xi ) = ci for all i ∈ {1, . . . , k},
2. has constant depth (independent of n, k, or c), and
2
3. has size at most O(log c · ck · k4 · n log2 n).
3
3.1
Known Kernelization Algorithms for the Hitting Set Problem
Known Sequential Kernelization Algorithms
The knwon sequential kernelization algorithms for pk,d -hitting-set are based on the so-called
Sunflower Lemma. The perhaps simplest application of this lemma is to repeatedly collapses
sufficiently large sunflowers to their cores until there are no longer any large sunflowers in the
graph and, then, the Sunflower Lemma tells us that the graph “cannot be very large.” In detail,
the definitions and algorithm are as follows:
Definition 3.1 (Sunflower). A sunflower S with core C is a set of proper supersets of C such
that for any two distinct p, q ∈ S we have p ∩ q = C. The elements of a sunflower are called
petals. A sunflower in a hypergraph is a sunflower whose petals are hyperedges of the hypergraph.
Fact 3.2 (Sunflower Lemma [13]). Every hypergraph H with more than kd(H) · d(H)! hyperedges
contains a sunflower of size k + 1.
4
The importance of the Sunflower Lemma for the hitting set problem lies in the following
observation: Suppose a hypergraph H contains a sunflower S of size at least k + 1. Then H
has a size-k hitting set if, and only if, the hypergraph obtained from H by removing all petals
of the sunflower and adding its core has such a hitting set (we cannot hit the k + 1 petals in
the sunflower using only k vertices without using at least one vertex of the core; thus, we hit
all petals if, and only if, we hit the core). In other words, replacing a sunflower of size k + 1 by
its core is a reduction rule for the hitting set problem; and if we can no longer apply this rule,
the Sunflower Lemma tells us that the hypergraph’s size is bounded by a function that depends
only on k and d(H) – in other words, it is a kernel.
The just-described kernelization algorithm is simple, but “very sequential.” It is, however,
not too difficult to turn it into a more parallel algorithm – at least, as long as d(H) is fixed. This
was first noted by Chen et al. [11] and we explain the ideas behind their proof below, rephrased
for the purposes of the present paper.
A better sequential kernelization algorithm has recently [17] been proposed (it runs in time
O(2d(H) |E|), which is linear from a parameterized point of view) – but the algorithm is arguably
“even more sequential” and does not lend itself to easy parallelization.
3.2
Known Parallel Kernelization Algorithm
The first step towards a parallel kernelization is the observation that we can compute many
cores in parallel. Given a hypergraph H = (V, E) and a number k, let a k-core in H be a core C
of a sunflower in H with more than k petals. Let k -cores(H) = V, {C | C is a k-core in H} .
While in the sequential algorithm we always replace one sunflower by its core, we now replace
all sunflowers by their cores. This leaves behind some hyperedges, but the Sunflower Lemma
will show that their number is “small.” Unfortunately, the set of cores itself may still be large
and we need to apply the replace-all-sunflowers-by-cores operation repeatedly. This process does
stop after at most d(H) rounds since the size of the cores decreases by 1 in each round and,
hence, after d(H) rounds it has shrunk to 0.
H = H0
f
c
g
h
u
v
w
r
s
o
i
d
u
v
w
t
r
s
t
p
q
o
p
q
j
k
i
j
k
H1
l
m n
f
e
c
b
b
a
a
g
h
d
l
m n
e
Figure 1: Visualization of a hypergraph H0 and of its 2-cores H1 = 2-cores(H0 ). Vertices are
drawn as rectangles, while the ten hyperedges of H0 are drawn as lines: they contain all vertices
that they touch. For instance, the leftmost line starting in the vertex a in H0 visualizes the
hyperedge {a, b, c, f, u, v, w} and the rightmost line visualizes the hyperedge {a, b, e, n}. The
hypergraph H0 contains three sunflowers of size 3, visualized by the red, blue, and green lines,
respectively. Their cores are the hyperedges shown in H1 . These cores, in turn, form a sunflower
in H1 with core {a, b}, but note that {a, b} is not a 2-core of H0 . It is the only hyperedge of H2 .
Let us now formalize these ideas a bit: Let H0 = H and let Hi+1 = k -cores(Hi ). Then H0
is the original hypergraph; H1 is the set of its k-cores; H2 is the set of H1 ’s k-cores and thus
the set of “cores of cores” of H; next H3 is the set of “cores of cores of cores” of H; and so on,
5
see Figure 1 for an example. In a sense, each Hi is nested into the previous hypergraph, leading
to a whole sequence resembling a matryoshka doll. Below, we define a matryoshka sequence as
a sequence that has this “nested in some sense” property and then show in Lemma 3.4 that
(H0 , H1 , . . . ) is, indeed, such a matryoshka sequence:
Definition 3.3 (Matryoshka Sequence). A matryoshka sequence for a hypergraph H = (V, E)
and a number k is a sequence (M0 , M1 , . . . , Md(H) ) of hypergraphs, all of which have the same
vertex set V , with the following properties for all i ∈ {0, . . . , d(H)}:
1. M0 = H,
2. d(Mi ) ≤ d(H) − i,
3. k -cores(Mi ) ⊆ Mi+1 , and
4. every size-k hitting set of H is also a hitting set of Mi .
Lemma 3.4 (Cores of Cores Form a Matryoshka Sequence). For every hypergraph H and number k, the sequence (H0 , . . . , Hd(H) ) is a matryoshka sequence for H and k.
Proof. The first property of a matryoshka sequence is true by definition. The second property
follows since each time we apply the operator k -cores to a hypergraph, the maximum size of the
hyperedges decreases by at least 1 (cores are smaller than the largest petals of their sunflowers).
Since we start at d(H), we get d(Hi ) ≤ d(H) − i. For the third property we actually even have
equality here by definition. The last property is proved by induction on i. The case i = 0
is trivial, so consider a hitting set X of size k for Hi−1 and consider a k-core C of Hi−1 . By
definition there must be a sunflower {e1 , . . . , ek+1 } ⊆ E(Hi−1 ) with core C. If X did not hit C
(that is, if X ∩ C = ∅), then the size-k set X would have to hit all of the k + 1 pairwise disjoint
sets ei − C for i ∈ {1, . . . , k + 1}, which is impossible.
Recall that the idea behind the parallel computation of a kernel for the hitting set problem
is to repeatedly remove all sunflowers from H, each time perhaps leaving a manageable number
of hyperedges – and after d rounds, no hyperedges will remain. We use the following notation
for the “removal” operation: For two hypergraphs H = (V, E) and H ′ = (V, E ′ ) let H ⊖ H ′ =
V, { e ∈ E | ∀e′ ∈ E ′ : e′ 6⊆ e } , that is, we remove all hyperedges from H that contain a
hyperedge of H ′ . Thus, H ⊖ H1 is the set of all hyperedges in H that are not involved in any
sunflower of size at least k + 1 since we remove all edges that contain a core.
The following theorem shows that the repeated removing operation only leaves behind a
“small” number of hyperedges. We formulate the theorem for arbitrary matryoshka sequences
(we will need this later on), but it is best to think of the Mi as the sets Hi .
Theorem 3.5 (Kernel Theorem). Let (M0 , . . . , Md(H) ) be a matryoshka sequence for H and k.
Let K = (M0 ⊖ M1 ) ∪ (M1 ⊖ M2 ) ∪ (M2 ⊖ M3 ) ∪ · · · ∪ (Md(H)−1 ⊖ Md(H) ) ∪ Md(H) .
1. Then K has at most
Pd(H)
i=0
ki i! hyperedges and
2. H and K have the same size-k hitting sets.
Proof. For the first item, fix an i and consider Mi ⊖ Mi+1 . We claim that these “remaining
hyperedges” cannot contain a sunflower {e1 , . . . , ek+1 }: If it did, by the third property of a
matryoshka sequence the sunflower’s core would be an element of Mi+1 and, thus, none of the
ei would be in Mi ⊖ Mi+1 . By the Sunflower Lemma and the fact that d(Mi ⊖ Mi+1 ) ≤ d(Mi ) ≤
d(H) − i by the second property of a matryoshka sequence, we get that Mi ⊖ Mi+1 cannot have
more than kd(H)−i · (d(H) − i)! hyperedges. This means that the union K cannot have more
hyperedges than the sum of these numbers for i ∈ {1, . . . , d(H)} plus the number of hyperedges
6
in Md(H) . However, by the second property we have d(Md(H) ) ≤ 0 and, thus, this hypergraph
can contain at most one edge (the empty edge). We account for this single edge by the value
Pd(H)
k0 0! = 1 for i = 0 in the sum i=0 ki i!.
For the second item, we make a simple observation: Let A, B, and C be hypergraphs (all
with the same vertex set V as always) such that every size-k hitting set of A ∪ B is also a hitting
set of C. We claim that
A ∪ B and A ∪ (B ⊖ C) ∪ C have the same size-k hitting sets.
(∗)
To see this, first consider a size-k hitting set X of A ∪ B. Trivially, X is also a hitting set of
A ∪ (B ⊖ C) ⊆ A ∪ B and X is also hitting set of C by assumption. Now, second, consider a
size-k hitting set X of A ∪ (B ⊖ C) ∪ C. Trivially, X hits all of A as well as all hyperedges in B
that are in B ⊖ C, so consider an edge e ∈ B − (B ⊖ C). By definition, this means that e must
be a superset of some e′ ∈ E(C) and X was a hitting set for C and, thus, hits e′ and therefore
also e.
Let us now prove the second item by proving by induction on i that H and
(M0 ⊖ M1 ) ∪ (M1 ⊖ M2 ) ∪ (M2 ⊖ M3 ) ∪ · · · ∪ (Mi−1 ⊖ Mi ) ∪ Mi .
have the same size-k hitting sets. The base case i = 0 is true by the first property of a matryoshka
sequence. For the inductive step from i to i + 1, let A = (M0 ⊖ M1 ) ∪ · · · ∪ (Mi−1 ⊖ Mi ) and
B = Mi and C = Mi+1 . By the inductive assumption, H and A ∪ B have the same size-k hitting
sets. The fourth property of a matryoshka sequence now implies that every size-k hitting set of
A ∪ B is also a hitting set of C. By (∗) we get that A ∪ B has the same size-k hitting sets as
A ∪ (B ⊖ C) ∪ C. Thus, H and A ∪ (B ⊖ C) ∪ C have the same size-k hitting sets, which was
exactly the inductive claim for i + 1.
Instantiating the theorem with (H0 , . . . , Hd(H) ) tells us that, if we can compute the elements
of K = (H0 ⊖ H1 ) ∪ · · · ∪ (Hd(H)−1 ⊖ Hd(H) ) ∪ Hd(H) in parallel, we can compute a kernel for the
hitting set problem in parallel. Clearly, “computing K” essentially boils down to “computing
the Hi ” in parallel. Thus, the real question, which we address next, is how quickly and easily
we can compute the hypergraphs Hi .
At this point, we briefly need to address some technical issues concerning the coding of
hypergraphs. For our purposes, it is largely a matter of taste how the input hypergraph H0 is
encoded, but the encoding of the later graphs Hi becomes important in the context of parallel
constant-time computations. We consider H = (V, E) fixed and encoded using, for instance,
an incidence matrix (having |V | columns and |E| rows). We encode a refinement of H, that
is, a hypergraph H ′ = (V, E ′ ) with the property that each e′ ∈ E ′ is a subset of some e ∈ E,
using a matrix of 2d(H) columns and |E| rows. There is a column for each of the at most 2d(H)
possible subsets of an edge e ∈ E and the entry at the column for a given row is 1 if this
subset is an element of E ′ ; otherwise it is 0. Let us call this the refinement matrix enconding of
hypergraph H ′ (with respect to the fixed input hypergraph H).
Lemma 3.6 (Computing Cores in Constant Depth). For each d and i there is a dlogtimeuniform family of AC-circuits that
1. on input of the incidence matrix of a hypergraph H with d(H) ≤ d, a number k, and the
refinement matrix encoding of the hypergraph Hi ,
2. outputs the refinement matrix encoding of Hi+1 ,
3. has constant depth, and
4. has size f (k, d) · |V |O(1) |E|O(1) where f is some computable function.
7
Proof. By the definition of refinement matrix encodings, it suffices to show how we can decide
for a single set e′ ⊆ e ∈ E whether it is a hyperedge of Hi+1 or not using a circuit of constant
depth and size f (k, d) · |V |O(1) |E|O(1) – it then follows that we can do this for all possible e′
in parallel without increasing the circuit depth at all and without increasing the circuit size by
more than 2d |E|.
By definition, e′ is a hyperedge of Hi+1 if it is a k-core of Hi = (V, Ei ). Thus, “all” we
need to test is whether there are petals p1 , . . . , pk+1 ∈ Ei that form a sunflower in Hi with
core e′ . This question can be answered very quickly in parallel using color coding as follows:
The vertices from the Color Coding Lemma (Fact 2.1) that we “have in mind” are the vertices
in the petals and the color “we have in mind for them” is i for all vertices in pi − e′ . Formally,
we use Fact 2.1 to obtain a set Λ of colorings λ : V → {1, . . . , k + 1} and check whether for one
of these colorings for each j ∈ {1, . . . , k + 1} there exists a hyperedge pj ∈ Ei with the properties
that (1) pj ⊇ e′ and (2) all vertices in pj − e′ have the color j. Clearly, such a coloring together
with the hyperedges p1 to pk+1 exists if, and only if, e′ is the core of a sunflower of size k + 1
in Hi . Even more importantly, Fact 2.1 provides us with such a coloring λ – if it exists – via a
circuit of constant depth and size at most some polynomial in the number |V | of vertices times a
computable function f (k, d) depending on the number k + 1 of colors and the maximum number
d(k + 1) of vertices in the sunflower (for which we “had colors in mind”).
The lemma tells us that once we have computed some Hi , we can compute the next Hi+1
using only constant additional depth and using f (k, d) · |V |O(1) |E|O(1) additional size. Since
Hi ⊖ Hi+1 can easily be computed from Hi and Hi+1 in constant depth, we get:
Theorem 3.7 (Depth-O(d) Kernelization Algorithm, [11]). For each d there is a dlogtimeuniform family of AC-circuits that
1. on input of a hypergraph H with d(H) ≤ d and a number k
2. outputs a hypergraph K having the same size-k hitting sets as H and having at most
Pd(H) i
i=0 k i! hyperedges,
3. has depth O(d),
4. and has size f (k, d) · |V |O(1) |E|O(1) where f is some computable function.
4
Pseudo-Cores and Pseudo-Sunflowers
The parallel kernelization algorithm described in the previous section has a depth that is linear
in the parameter d, the maximum size of any hyperedge in the input hypergraph. The reason
for this linear dependency was that, while we managed to reduce not just one but all sunflowers
in the hypergraph to their cores in parallel, we had to repeat this “reduce to core” procedure d
times – and each round adds a constant number of layers to the circuit.
It is not obvious how this build-up of layers can be avoided. In the following, we first explain
why there are good reasons to believe that the computation of the hypergraphs Hi necessitates
deeper and deeper circuits. Following this discussion, we explain our proposal for side-stepping
these difficulties: we replace the hypergraphs Hi by new hypergraphs Hi′ that are easier to
compute but still form a matryoshka sequence and – hence – can serve as a replacement for the
Hi in the Kernel Theorem, Theorem 3.5.
The Difficulty: Cores of Cores Are Hard to Compute There are several reasons to
believe that one cannot compute kernels for the hitting set problem in constant depth using the
repeated sunflower-reduction-procedure. A first idea for reaching a constant depth is to apply
8
the reduction procedure only a constant number of times (instead of d times). Indeed, it is
not immediately clear that a “core of cores” is not already a core in the first round – so do
we actually need more than one round? Unfortunately, the answer is “yes, we do”: Figure 1
shows an example where {a, b} is a 2-core of the 2-cores, but it is not a 2-core of the original
hypergraph. For a more complex example, where d rounds are needed to arrive at a constant
size kernel, consider the trees Tdℓ (defined in detail later on) that are perfectly balanced trees of
depth d with ℓ + 1 children per node for a number ℓ ≥ k – and now consider the hypergraph
H d that has one hyperedge for each leaf of Tdℓ and this hyperedge contains all the nodes on the
path from the leaf to the root r. Now, for i > 0 we have k -cores(H i ) = H i−1 and the latter
hypergraphs all have a size of at least the arbitrarily large ℓ for i > 1. Thus, we need to apply
the “core of cores” procedure at least d − 1 times before arriving at a hypergraph whose size
depends only on the parameter.
A second, more promising idea is the observation that it might be possible to somehow
“collapse” two (and then, hopefully, all) applications of the sunflower-reduction-procedure “into
a single application.” Unfortunately, we also run into a problem here, namely in the “collapsed
color coding process.” In essence, color coding is great at ensuring that certain vertex sets
are disjoint (namely those vertex sets that receive different colors), but fails at enforcing that
the same vertices are used in different hyperedges – which is exactly what is needed when the
definition of some Hi refers to Hi−1 , which in turn refers to some Hi−2 .
These problems with avoiding the build-up of additional layers with rising d have led Chen
et al. [11] to the conjecture that the build-up is unavoidable and that all parallel kernelization
algorithms for pk,d -hitting-set have a runtime that is linear in d. We agree with Chen et al. in
their assessment that the computation of the Hi presumably necessitates a linear circuit depth
– but, nevertheless, we will refute their conjecture in the following.
The Solution: Pseudo-Cores As a Replacement For Cores Our idea is not to compute
the sets Hi (we do not see how this can be done in constant time), but to compute hypergraphs
Hi′ with rather similar properties (formally, they will form matryoshka sequences as well) that
we can compute in constant time for all d and i. We introduce a new notion of k-pseudo-cores
of level i and Hi′ will be the hypergraph whose edges are the k-pseudo-cores of level i. Crucially,
the definition of Hi′ (only) refers directly to the original input graph H and its hyperedges can be
obtained from H directly using color coding. At the same time, the Hi′ will form a matryoshka
′
sequence and, hence, just as for the Hi , the core of any sunflower of Hi−1
must already be
′
present in Hi .
The definition of pseudo-cores is somewhat technical. We will, however, show that all cores
are pseudo-cores of level 1, cores of cores are pseudo-cores of level 2, and so on. The reverse
implication does not hold (for instance, pseudo-cores of level 2 need not be cores of cores). For
a “level” L and a number k, let TLk denote the rooted tree in which all leafs are at the same
depth L and all inner nodes have exactly k + 1 children. The root of TLk will always be called r
in the following. Thus, T1k is just a star consisting of r and its k + 1 children, while in T2k each
of the k + 1 children of r has k + 1 new children, leading to (k + 1)2 leafs in total. For each
l ∈ leafs(TLk ) = { l | l is a leaf of TLk } there is a unique path (l0 , l1 , . . . , lL ) from l0 = r to lL = l.
An example for the following definition is shown in Figure 2.
Definition 4.1 (Pseudo-Sunflowers and Pseudo-Cores). Let H = (V, E) be a hypergraph and let
L and k be fixed. A set C ⊆ V is called a k-pseudo-core of level L in H if there exists a mapping
S : leafs(TLk ) × {0, 1, . . . , L} → 2V , called a TLk -pseudo-sunflower for H with pseudo-core C, such
that for all l, m ∈ leafs(TLk ) with l 6= m we have:
1. S(l, 0) = C.
2. S(l, 0) ∪ S(l, 1) ∪ · · · ∪ S(l, L) ∈ E and let us write S(l) for this hyperedge.
9
3. S(l, i) ∩ S(l, j) = ∅ for 0 ≤ i < j ≤ L, but S(l, i) 6= ∅ for i ∈ {1, . . . , L}.
4. Let z ∈ {1, . . . , L} be the smallest number such that lz 6= mz , that is, z is the depth
where the path from r to l and the path from r to m diverge for the first time. Then
S(l, z) ∩ S(m, z) = ∅ must hold.
u
H
r
f
g
h
v
s
w
t
o
p
q
i
j
k
T22
l
m n
r
c
d
l
c1
c2
e
b
c3
a
S(l, 0)
S(l, 1)
S(l, 2)
{a, b}
{a, b}
{a, b}
{a, b}
{a, b}
{a, b}
{a, b}
{a, b}
{a, b}
{c, f }
{c, g}
{c, h, o}
{d, i}
{d, j}
{d}
{e}
{e}
{e}
{u, v, w}
{r, s, t, m}
{p, q, l, e}
{o, r, u}
{p, s, v}
{k, q, t, w}
{l}
{m}
{n}
Figure 2: A T22 -pseudo-sunflower S for the level 2 pseudo-core {a, b} in the hypergraph H. The
four properties of pseudo-sunflowers hold: In “column S(l, 0)” we always have the pseudo-core,
the union of each row is a hyperedge, the sets in a row form a partition of this hyperedge,
and – most importantly – we have the disjointness property at each “branch” of the tree. This
property requires that for column S(l, 1) the sets of all red vertices, of all blue vertices, and of
all green vertices are pairwise disjoint; whereas for column S(l, 2) it requires that the three red
sets are pairwise disjoint, likewise for the three blue sets, and the three green sets. However, it
is permissible (and the case) that a red vertex in the third column is the same as green vertex
in the third or the second column.
′
Definition 4.2. For a hypergraph H
= (V, E) ′ and numbers k and i ≥ 1 let Hi = V, {C |
C is a k-pseudo-core of level i of H} and let H0 = H.
To get some intuition, let us have a closer look at H1′ . As the following lemma shows, pseudocores and cores are still very closely related at this first level – while for larger levels, we no
longer have Hi = Hi′ , but only Hi ⊆ Hi′ .
Lemma 4.3. Let H be a hypergraph and k a number. Then H1 = H1′ .
Proof. Consider a k-pseudo-core of H of level 1. The tree T1k consists of a root r with leafs l1 to
lk+1 . Consider a T1k -pseudo-sunflower S and let us fix some leaf l of T1k . The pseudo-sunflower
must map (l, 0) to the pseudo-core C and (l, 1) to a set of vertices that is disjoint from C. This
means that S(l) = S(l, 0) ∪ S(l, 1) = C ∪ S(l, 1) ∈ E is a hyperedge in H that contains the
(pseudo)core C. Furthermore, for any two different leafs l and m we have S(l, 1) ∩ S(m, 1) = ∅
or, equivalently, S(l) ∩ S(m) = C. This shows that {S(l1 ), . . . , S(lk+1 )} is a sunflower with
core C. For the other direction, given a sunflower {e1 , . . . , ek+1 } of size k + 1 in H with core C,
the T1k -pseudo-sunflower S given by S(li , 0) = C and S(li , 1) = ei − C, where the li are the k + 1
leafs of T1k , witnesses that C is also k-pseudo-core of level 1 of H.
5
The Constant-Depth Kernelization
We show that hitting set kernels can be computed in constant depth in two steps:
′
) is a matryoshka sequence.
1. We show that (H0′ , . . . , Hd(H)
10
2. We show that all Hi′ can be computed by a constant depth circuit whose depth is independent of both k and d(H).
By the Kernel Theorem, Theorem 3.5, taken together, these two items yield the desired kernelization algorithm.
Step 1: Pseudo-Cores Form Matryoshka Sequences. Our first aim is to show the following theorem, which is an analogue of Lemma 3.4 for pseudo-cores:
′
) from DefTheorem 5.1. For every hypergraph H and number k, the sequence (H0′ , . . . , Hd(H)
inition 4.2 is a matryoshka sequence for H and k.
The proof consists of four lemmas, one for each of four properties of a matryoshka sequence:
Lemma 5.2. H0′ = H.
Proof. By definition.
Lemma 5.3. d(HL′ ) ≤ d(H) − L holds for all L ∈ {0, . . . , d(H)}.
˙
˙ · · ∪S(l,
˙
Proof. For every leaf l we have S(l) = S(l, 0)∪S(l,
1)∪·
L) and all S(l, i) for i ∈ {1, . . . , L}
are non-empty sets. This implies that |S(l, 0)| ≤ |S(l)| − L ≤ d(H) − L.
′
Lemma 5.4. k -cores(HL′ ) ⊆ HL+1
holds for all L ∈ {0, . . . , d(H)}.
Proof. We show the claim by induction on L. The base case L = 0 was already handled in
Lemma 4.3. For larger L, let e be a k-core of HL′ . Then there is a sunflower {e1 , . . . , ek+1 } ⊆
E(HL′ ) with core e and |e| ≤ d(H) − L − 1 since all ei have the property |ei | ≤ d(H) − L by
definition and since a core is always smaller than its largest petal. For each j ∈ {1, . . . , k + 1}
there must now be a TLk -pseudo-sunflower Sj with pseudo-core ej . From these k + 1 different
k -pseudo-sunflower S whose pseudo-core is e as follows:
pseudo-sunflowers, we construct a TL+1
k
k , let l1 = c be
In the tree TL+1 , let c1 to ck+1 be the children of the root r. For a leaf l of TL+1
j
the child of r on the path from r to l and let us view l also as a leaf in the tree TLk rooted at cj .
We define S as follows:
1. S(l, 0) = e.
2. S(l, 1) = ej − e.
3. S(l, i) = Sj (l, i − 1) for i ∈ {2, . . . , L + 1}.
It remains to show that the mapping S defined in this way satisfies the four properties of a
k :
pseudo-sunflower. Consider any two leafs l and m of TL+1
1. By definition, S(l, 0) = e and, thus, e is the pseudo-core of S.
2. S(l, 0)∪S(l, 1)∪· · ·∪S(l, L+1) = e∪(ej −e)∪Sj (l, 1)∪· · ·∪Sl (l, L) = e∪(ej −e)∪(Sj (l)−ej ).
We know that Sj (l) ∈ E holds (since Sj is a pseudo-sunflower), that ej ⊆ Sj (l) holds (since
ej is the pseudo-core of Sj ), and that e ⊆ ej holds (since e is the core of a sunflower that
has ej as one of its petals). This implies S(l) = Sj (l) and the latter is an element of E.
3. Clearly, S(l, 0) = e and S(l, 1) = ej − e are disjoint and S(l, 1) has size at least 1 since ej is
not a subset of e. The other S(l, i) are also disjoint from one another since the Sj (l, i − 1)
are disjoint among one another, and they are also disjoint from S(l, 0) and S(l, 1) (since
all Sj (l, i − 1) are disjoint from Sj (l, 0) = ej ).
11
4. Finally, let z be the “divergence depth” of l and m, that is, the smallest number with
lz 6= mz . For z = 1, we have S(l, 1)∩S(m, 1) = ∅ since S(l, 1) = ej −e and S(m, 1) = ej ′ −e
for some indices j and j ′ and since the intersection of the two sunflower petals ej and ej ′
is exactly e. For z > 1, the leafs l and m lie in the same tree TLk rooted at some child cj
of r and we have S(l, z) ∩ S(m, z) = Sj (l, z − 1) ∩ Sj (m, z − 1) and the latter intersection
is empty since Sj is a pseudo-sunflower.
Lemma 5.5. Every size-k hitting set of H is also a size-k hitting set of HL′ for all L ∈
{0, . . . , d(H)}.
Proof. We must show that every hitting set X of H with |X| ≤ k is also a hitting set of HL′ .
That is, we must show that every k-pseudo-core C of level L contains at least one element of X.
Let S be a TLk -pseudo-sunflower with pseudo-core C.
The following definition will be crucial in the following: We say that X hits a node n of TLk
if there is a leaf l of TLk such that n = lD (that is, n is the Dth node on the path l0 , l1 , . . . , lL
from the root r = l0 to l = lL ) and X ∩ (S(l, 0) ∪ S(l, 1) ∪ · · · ∪ S(l, D)) 6= ∅.
First, observe that X hits every leaf of TLk since, for them, D = L and S(l, 0) ∪ · · · ∪
. . . S(l, L) = S(l) is a hyperedge of H and, hence, gets hit by X.
Second, we claim that if X hits all children c1 , . . . , ck+1 of a node n of TLk , then X also hits n.
Let n be at depth D, so the children are at depth D + 1. By definition of “being hit by X,” for
each child ci of n there must be a leaf li such that
X ∩ (S(li , 0) ∪ · · · ∪ S(li , D + 1)) 6= ∅.
(1)
We claim that at least one of the li also witnesses that X hits n. Otherwise, since all li are in
the subtree rooted at n, for all of them we would have
X ∩ (S(li , 0) ∪ · · · ∪ S(li , D)) = ∅.
(2)
From (1) and (2) we immediately get that
X ∩ S(li , D + 1) 6= ∅ must hold for all i ∈ {1, . . . , k + 1}.
(3)
Now, for any two different leafs li and lj consider the two paths from the root to them. These
paths will be identical exactly up to the node n and will then split into a path via the child
ci and a path via the child cj . Now, in this situation the fourth property of pseudo-sunflowers
tells us that S(li , D + 1) ∩ S(lj , D + 1) = ∅ must hold. In other words, the k + 1 many sets
S(li , D + 1) in equation (3) are pairwise disjoint. However, this means that the size-k set X
cannot contain one element of each of them. Thus, our assumption that X does not hit n has
lead us to a contradiction.
Third, we claim that X hits the root of TLk . This follows easily from the first two claims
since X hits all leafs of TLk and whenever it hits all children of a node, it also hits the node.
Clearly, this implies that X hits all nodes, including the root.
Now, we are done since “X hits the root” means that X ∩ S(l, 0) 6= ∅ holds for at least one
leaf l and S(l, 0) = C. Thus, X ∩ C 6= ∅, which was the claim.
Step 2: Pseudo-Cores Can Be Computed in Constant Depth. Theorem 5.1 states that
the hypergraphs Hi′ form a matryoshka sequence and, thus, the Kernel Theorem tells us that
the following hypergraph is a kernel for the hitting set problem: K = (H0′ ⊖ H1′ ) ∪ (H1′ ⊖ H2′ ) ∪
′
′
′
· · · ∪ (Hd(H)−1
⊖ Hd(H)
) ∪ Hd(H)
. Of course, the whole effort that went into the definition of the
′
Hi and the proof of the matryoshka properties would be for nothing, if the Hi′ were not easier
to compute than the Hi .
12
This is exactly what we claim in the following theorem and prove in the rest of this paper: It
is an analogue of Lemma 3.6 for pseudo-cores. The crucial difference in the formulation is that,
′
now, we no longer get Hi−1
as input when we compute Hi′ , but rather we compute Hi′ “directly”
from the original graph H.
Theorem 5.6 (Computing Pseudo-Cores in Constant Depth). There is a dlogtime-uniform
family of AC-circuits that
1. on input of the incidence matrix of a hypergraph H = (V, E) and numbers k and L,
2. outputs the refinement matrix encoding of HL′ ,
3. has constant depth (in particular, it is independent of |V |, |E|, d(H), k, and L), and
4. has size f (k, d(H)) · |V |O(1) |E|O(1) where f is some computable function.
To compute the encoding of HL′ , we can consider all candidate pseudo-cores in parallel.
Thus, proving the theorem boils down to deciding for a subset C ⊆ V whether there exists a
TLk -pseudo-sunflower S of H whose pseudo-core is C. Of course, we wish to use color coding for
this and our definition of pseudo-cores and pseudo-sunflowers was carefully crafted so that it
includes only requirements of the form “these parts of these hyperedges must be disjoint” (and
not – as is necessary for describing cores of cores – statements like “these hyperedges must share
the vertices that form petals”). Unfortunately, while we no longer need to ensure that certain
parts of different hyperedges are identical, we must be careful that we do not inadvertently
forbid vertices to be the same across hyperedges when we “do not care whether they are the
same”:
Example 5.7. Suppose we wish to find two disjoint hyperedges e1 = {v1 , v2 , v3 } and e2 =
{v4 , v5 , v6 } in a hypergraph H plus another hyperedge e3 = {x, y} such that x ∈
/ e1 ∪ e2 , but do
not care whether y ∈ e1 ∪ e2 holds or not. We can easily enforce the disjointness properties by
coloring v1 to v6 using colors 1 to 6 and x using color 7. However, how should we color y for
which we do not care about disjointness (at least with respect to e1 and e2 )? Fixing any of the
colors 1 to 3 for y or any of the colors 4 to 6 (or, for that matter, any other color) would be
wrong, since this would enforce either y ∈
/ e2 or y ∈
/ e1 (or both).
Fortunately, there is a way out of the dilemma: we consider all feasible colors y could get
in parallel. To formalize this “trick”, we define a technical problem in which an undirected
graph G is used to specify which vertices in hyperedges of a hypergraph H should be different.
As is customary, a proper coloring of an undirected graph G = (U, F ) is a mapping c : U → C
to some set C of colors with c(u) 6= c(v) for all {u, v} ∈ F . Let us write f [X] = {f (x) | x ∈ X}
for the image of a set X under a function f . For an example instance see Figure 3.
Problem 5.8. pG -restricted-coloring
Instance: A hypergraph H = (V, E) and an undirected graph G = (U, F ) together with a partition U = U1 ∪˙ · · · ∪˙ Um of U .
Parameter: |G|
Question: Is there a proper coloring c : U → V of G such that c[Ui ] ∈ E holds for all i ∈
{1, . . . , m}?
Lemma 5.9. The problem pG -restricted-coloring can be solved by a dlogtime-uniform
family of AC-circuits of constant depth and size f (|G|)|V |O(1) |E|O(1) for some computable function f .
13
Proper coloring
H′
f
c
g
h
u
v
w
r
s
t
o
p
q
i
j
k
d
G
l
m n
e
c f f f
u v ww
c g g g
r s tm
c h o o
p q l e
d i i i
o r uu
d j j j
p s v v
d d d d
k q tw
e e e e
l l l l
e e e e
mmmm
e e e e
nnnn
Figure 3: An instance of pG -restricted-coloring consisting of a hypergraph H ′ and a graph G
(a thick edge connecting two areas with dashed borders indicates that there is an edge between
each vertex of the first area and each vertex of the second area; thus, in the example, each
thick edge corresponds to 12 · 12 = 144 edges). This instance is the one resulting from the
reduction described in the proof of Theorem 5.6 for L = 2, the hypergraph H from Figure 1,
and the core {a, b} (except that we use only four vertices in G per set S(l, i) instead of d = 9). A
proper coloring is shown right (the table indicates the values c(u) ∈ V (H ′ ) for the corresponding
vertices u of G).
Proof. Let G = (U, F ) and H = (V, E) be given as input. We assume that |U | ≤ |V | holds
since, otherwise, the number of possible mappings c that must be checked only depends on the
parameter |G| and, thus, they can easily be checked in parallel in constant depth (any function
can be computed in constant depth when circuit size is not an issue).
The objective is, of course, to use color coding to find the mapping c. Towards this aim, we
search for a mapping d : V → {1, . . . , |U |} with the following two Properties 1 and 2:
1. There is a proper coloring c′ : U → {1, . . . , |U |} of G such that
2. for each i ∈ {1, . . . , m} there is a hyperedge ei ∈ E with |d[ei ]| = |ei | and d[ei ] = c′ [Ui ].
The idea behind this search is the following: The graph G imposes restrictions of the form
“for the edge {u, v} the vertices c(u) and c(v) must be different.” In principle, such restrictions
are easy to enforce using color coding: use one color for c(u) and a different color for c(v).
However, as explained in Example 5.7, we may run into a problem when there is no edge
between two vertices x and y since, then, we may not rule out that c(x) = c(y) holds – which we
would rule out when two different colors are used. The solution offered following the example is
to try out all possible ways in which we may assign colors to vertices for which we “actually do
not care” about their colors. These “possible ways” are modeled by the proper coloring c′ from
above.
In detail, recall the situation of Example 5.7 where we searched for two disjoint hyperedges
e1 = {v1 , v2 , v3 } and e2 = {v4 , v5 , v6 } and a hyperedge e3 = {x, y} with x ∈
/ e1 ∪ e2 . This search
can be modeled by a graph G whose vertex set U can be partitioned into U1 = {u1 , u2 , u3 },
U2 = {u4 , u5 , u6 }, and U3 = {u7 , u8 } and where the edge set F is a clique on U1 ∪ U2 (to ensure
that the hyperedges c[U1 ] and c[U2 ] are disjoint and have size 3) and there are edges between u7
and all elements of U1 ∪ U2 (to ensure that c(u7 ) ∈
/ c[U1 ] ∪ c[U2 ]). A proper coloring c : U → V
might now map c(ui ) = vi for i ∈ {1, . . . , 6} and c(u7 ) = x and c(u8 ) = y. All vertices in
{v1 , . . . , v6 , x} must be distinct, but y must not necessarily be distinct from them – in fact, it
could be any of them. In this situation, the different possible values of y give rise to different c′
14
and d (note that in all of these examples, c′ is a proper coloring of G and that |d[ei ]| = |ei | and
d[ei ] = c′ [Ui ] hold for i ∈ {1, 2, 3}):
Example 5.10. Suppose that y = v1 holds, that is, e1 ∩e3 = {v1 } and e2 ∩e3 = ∅. This situation
is modeled by the following functions c′ and d: c′ (ui ) = i for i ∈ {1, . . . , 7} and c′ (u8 ) = 1; and
d(vi ) = i for i ∈ {1, . . . , 6} and d(x) = 7 and d(v) can be arbitrary for v ∈
/ {v1 , . . . , v6 , x} (note
that we do not need to define d(y) since d(v1 ) is already defined and y = v1 ).
Example 5.11. Suppose that y = v5 holds, that is, e1 ∩ e3 = ∅ and e2 ∩ e3 = {v5 }. Here, we
can use almost identical functions c′ and d as in the previous example, except that c′ (u8 ) = 5.
Example 5.12. Suppose that y ∈
/ {v1 , . . . , v6 , x}. Then we use c(u8 ) = 8 and d(y) = 8.
Let us now formally argue that the search for d can be performed using color coding: First,
observe that the test “there is a proper coloring c′ ” can be performed in parallel by testing all
possible colorings of G (their number depends only on |G|). Second, given a mapping d : V →
{1, . . . , |U |}, we can determine the existence of hyperedges ei ∈ E with both |d[ei ]| = |ei | and
d[ei ] = c′ [Ui ] in constant depth. Third, if d : V → {1, . . . , |U
S |} has Properties 1 and 2, so does
anySother d′ as long as it is identical to d on the vertices of m
i=1 ei . Since the number of vertices
m
in i=1 ei is at S
most |U |, which depends only on the parameter |G|, instantiating Fact 2.1 with
′
{x1 , . . . , xk } = m
i=1 ei tells us that we can find one such d in constant depth.
It remains to argue that the following two statements are equivalent:
• There is a mapping d : V → {1, . . . , |U |} with the Properties 1 and 2.
• There is a mapping c : U → V such that is a proper coloring of G and c[Ui ] ∈ E holds for
all i ∈ {1, . . . , m}.
For the first direction, let d : V → {1, . . . , |U |} be a mapping and c′ : U → {1, . . . , |U |}
a proper coloring of G such that for each i ∈ {1, . . . , m} there is a hyperedge ei ∈ E with
|d[ei ]| = |ei | and d[ei ] = c′ [Ui ]. Define c : U → V as follows: We know that each u ∈ U lies in a
unique set Ui , and since d[ei ] = c′ [Ui ] for the hyperedge ei ∈ E, there must be an element vu ∈ ei
with d(vu ) = c′ (u). Since |d[ei ]| = |ei |, the element vu ∈ ei must be unique with this property.
Define c(u) = vu and observe that d ◦ c = c′ holds. (As an example, consider the functions c′
and d from Example 5.11 and u5 ∈ U and u8 ∈ U . For u5 , we have u5 ∈ U2 = {u4 , u5 , u6 }
and d[e2 ] = c′ [U2 ] = {4, 5, 6}. The element v5 has the property d(v5 ) = 5 and, indeed, it is the
only element with this property. Thus c(u5 ) = v5 . For u8 , we have u8 ∈ U3 = {u7 , u8 } and
d[e3 ] = c′ [U3 ] = {7, 5}. Again, the element v5 has the property d(v5 ) = 5 and c(u8 ) = v5 .)
The just-defined mapping c is a proper coloring of G since for every edge {u, v} ∈ F we
have d(c(u)) = c′ (u) 6= c′ (v) = d(c(v)), which implies c(u) 6= c(v). To prove c[Ui ] ∈ E for
i ∈ {1, . . . , m}, fix some i. Since d ◦ c = c′ , we also have d[c[Ui ]] = c′ [Ui ] and this equals d[ei ] by
assumption. By construction of c, we clearly also have c[Ui ] ⊆ ei . Finally, since |d[ei ]| = |ei |, we
know that d is injective on ei . Now, from d[c[Ui ]] = d[ei ] and c[Ui ] ⊆ ei and d being injective
on ei , we conclude that c[Ui ] = ei ∈ E must hold.
For the second direction, let c : U → V be a proper coloring of G with c[Ui ] ∈ E for all
i ∈ {1, . . . , m}. We need to construct a mapping d : V → {1, . . . , |U |} and a proper coloring
c′ : U → {1, . . . , |U |} with the property that for each i ∈ {1, . . . , m} there is a hyperedge ei ∈ E
with |d[ei ]| = |ei | and d[ei ] = c′ [Ui ].
For the definition of d, let b : c[U ] → {1, . . . , |c[U ]|} be a bijection. Define d as follows:
(
b(v) for v ∈ c[U ] and
d(v) =
1
otherwise.
15
Let c′ = d ◦ c. First, c′ is a proper coloring of G since c is a proper coloring and d restricted to
c[U ] is a bijection. Now consider an i ∈ {1, . . . , m}. Then ei = c[Ui ] is a hyperedge in E. We
have |d[ei ]| = |ei | since d restricted to ei ⊆ c[U ] is a bijection. We have d[ei ] = d[c[Ui ]] = c′ [Ui ]
by definition of c′ .
We are now ready to prove Theorem 5.6:
Proof of Theorem 5.6. Recall that in order to prove the claim, we must solve the following
problem in constant depth: Given H = (V, E), k, L, and a subset C ⊆ e ∈ E, check whether
there exists a TLk -pseudo-sunflower S of H whose pseudo-core is exactly C. We must now
show how the existence of the pseudo-sunflower can be checked using the technical problem pG restricted-coloring.
The input for the restricted coloring problem will consist of a special graph G that encodes
the different disjointedness properties of pseudo-sunflowers
using edges and will consist of the
hypergraph H ′ = V, { e − C | e ⊇ C, e ∈ E } . In other words, we restrict H to those edges that
contain the alleged core C (other edges cannot be part of the sought pseudo-sunflower anyway)
and we remove the core from the edge since they all contain it.
Let us now define the graph G = (U, F ). The objective of this definition is, of course, that
there is a TLk -pseudo-sunflower S with core C if, and only if, there a proper coloring c : U → V
of G such that c[Ui ] ∈ E(H ′ ) for all i ∈ {1, . . . , m}.
Pseudo-sunflowers are mappings from leafs(TkL ) × {0, . . . , L} to subsets of V such that for
each leaf l the union S(l, 0) ∪ · · · ∪ S(l, L) is a hyperedge in E. In our case, we must have
S(l, 1) ∪ · · · ∪ S(l, L) ∈ E(H ′ ) since S(l, 0) = C and we removed the fixed core C already from
the hyperedges of H ′ . In G, we will have one set Ul for each leaf l of TLk : The vertices that
will be assigned to the elements of Ul by the coloring c should then form exactly the hyperedge
S(l, 1) ∪ · · · ∪ S(l, L).
If we knew that each S(l, i) had size exactly 1, we could set U = leafs(TkL ) × {1, . . . , L}: For
each leaf l the coloring c would need to pick L vertices which, together, make up the hyperedge
S(l) − C of H ′ . To ensure that S(l, i) and S(l, j) are disjoint for i 6= j, we would make each
{l} × {1, . . . , L} a clique in G. However, the sets S(l, i) can have different sizes. For this reason,
we do not use a single vertex in G for each S(l, i), but d different vertices (actually, d − L + 1
vertices would suffice): The different elements of S(l, i) can be represented by different vertices
– and if |S(l, i)| < d, the coloring c can map the superfluous vertices to any of the vertices of
S(l, i).
We set U = leafs(TkL ) × {1, . . . , L} × {1, . . . , d} and define the partition of U by Ul =
{l} × {1, . . . , L} × {1, . . . , d} for each l ∈ leafs(TLk ). It remains to explain how we put edges
into G such that the colorings of G induce pseudo-sunflowers. The following edges are present
in G to ensure the four properties from Definition 4.1:
1. Nothing needs to be done to ensure the first property (S(l, 0) = C) since H ′ only contains
hyperedges that used to contain C.
2. Nothing needs to be done to ensure the second property (S(l) ∈ E) since c[Ul ] ∈ E(H ′ )
will ensure that S(l) − C ∈ E(H ′ ) holds and, thus, S(l) ∈ E(H).
3. To ensure the third property (S(l, i) ∩ S(l, j) = ∅ for i 6= j), for each l ∈ leafs(TLk ) and
every i 6= j and all x, y ∈ {1, . . . , d} let {(l, i, x), (l, j, y)} be an element of F , that is, let it
be an edge of G.
4. To ensure the fourth property (S(l, z) ∩ S(m, z) = ∅ must hold when lz and mz have the
same parent), for each l, m ∈ leafs(TLk ) and the smallest number z with lz 6= mz and all
x, y ∈ {1, . . . , d} let {(l, z, x), (m, z, y)} be an element of F .
16
With this definition, we claim that C is a k-pseudo-core of level L of H if, and only if, (H ′ , G)
is a element of pG -restricted-coloring. If we can show this, we are done by Lemma 5.9.
We need to prove two directions. First, let a TLK -pseudo-sunflower S of H with core C be
given. We must argue that there is a proper coloring c : U → V of G with c[Ul ] ∈ E(H ′ ) for all
leafs of TLk . This coloring is the following: Consider all leafs l and all numbers i ∈ {1, . . . , L}. For
each pair, the set S(l, i) consist of some vertices v1 , . . . , vp ∈ V for some p = |S(l, i)| ∈ {1, . . . , d}.
We set c(l, i, x) = vx for x ∈ {1, . . . , p} and c(l, i, x) = vp (or any other element of S(l, i), it does
not matter) for x ∈ {p + 1, . . . , d}.
With this definition, we clearly have c[Ul ] = S(l, 1)∪· · ·∪S(l, L) and since S(l) ∈ E, the latter
is an element of E(H ′ ). Furthermore, c is a proper coloring: For all edges {(l, i, x), (l, j, y)} ∈ F
we know that the colors c(l, i, x) and c(l, j, y) are different since c(l, i, x) ∈ S(l, i) and c(l, j, y) ∈
S(l, j) and S(l, i) ∩ S(l, j) = ∅. Next, for the edges of the form {(l, z, x), (m, z, y)} ∈ F we also
have that c(l, z, x) and c(m, z, y) are different since S(l, z) and S(m, z) are disjoint.
For the other direction, let a coloring c be given. Define a mapping S from leafs(TLk ) ×
{0, . . . , L} to subsets of V as follows: For all l ∈ leafs(Tlk ) let S(l, 0) = C and for i ∈ {1, . . . , L}
let S(l, i) = {c(l, i, 1), c(l, i, 2), . . . , c(l, i, d)}.
To see that S has the properties of a pseudo-sunflower, consider the four properties. The
first property is clearly true by definition. The second follows from c[Ui ] ∈ E(H ′ ) and, hence
C ∪c[Ui ] ∈ E(H). The third item follows from the following fact: For any two vertices vx ∈ S(l, i)
and vy ∈ S(l, j) for i 6= j, there is an edge between (l, i, x) and (l, j, y) in G and, thus, vx 6= vy .
This shows that S(l, i) ∩ S(l, j) = ∅ must hold; and note that, clearly, S(l, i) 6= ∅ always holds.
For the fourth item, we have S(l, z) ∩ S(m, z) = ∅ since for all vx ∈ S(l, z) and vy ∈ S(m, z)
there is an edge between (l, z, x) and (m, z, y) in G.
Theorem 5.6 now implies Theorem 1.2 by simple standard arguments:
Proof of Theorem 1.2. The only difference between the above claim and the claim of Theorem 5.6 (apart from the exact formulation) is that Theorem 1.2 requires the AC0 -circuit family
to have size |V |c |E|c for some constant c, while Theorem 5.6 allows it to have size f (k, d)|V |c |E|c .
To reduce the size, on input (H, k), a kernelization algorithm for Theorem 1.2 first checks whether
we have f (k, d) > |V |c |E|c and, if so, just outputs (H, k); otherwise it runs the kernelization
algorithm from Theorem 5.6, which needs size f (k, d)|V |c |E|c ≤ |V |2c |E|2c .
6
Conclusion
The results of this paper can be summarized as pk,d -hitting-set ∈ para-AC0 or, equivalently,
that kernels for the hitting set problem parameterized by k and d can be computed by a single
AC0 -circuit family. This result refutes a conjecture of Chen et al. [11]. The proof introduced a
new technique: Iterated applications of color coding can sometimes be “collapsed” into a single
application. This collapsing is not always straightforward (as the present paper showed) and
additional technical machinery may be needed to make it work.
The proof of our main result would be much simpler if the number of k-cores of a hypergraph
depended only on the parameters k and d (since, then, only one round would be needed in the
parallel algorithm). While we gave examples that refute this hope, it may be possible to tweak
the idea a bit: We can compute in constant parallel time the set of all inclusion-minimal
k-cores of a hypergraph. We believe that we can prove that the number of these inclusionminimal k-cores depends only on k and d (unfortunately, we need rather involved and technical
combinatorics and the dependence on k and d seems to be “quite bad”). Nevertheless, if this
is the case, we get a different proof that pk,d -hitting-set has an AC0 -kernelization, where the
complexity of proving correctness is shifted away from the algorithm (which gets much simpler)
towards the underlying graph theory and combinatorics.
17
References
[1] F. N. Abu-Khzam, M. A. Langston, P. Shanbhag, and C. T. Symons.
able parallel algorithms for fpt problems.
Algorithmica, 45(3):269–284,
doi:10.1007/s00453-006-1214-1.
Scal2006.
[2] Noga Alon, Raphael Yuster, and Uri Zwick. Color-coding. Journal of the ACM, 42(4):844–
856, 1995. doi:10.1145/210332.210337.
[3] Max Bannach, Christoph Stockhusen, and Till Tantau. Fast parallel fixed-parameter
algorithms via color coding.
In Proceedings of the Tenth International Symposium on Parameterized and Exact Computation (IPEC 2015), pages 224–235, 2015.
doi:10.4230/LIPIcs.IPEC.2015.224.
[4] Max Bannach and Till Tantau. Parallel multivariate meta-theorems. In Proceedings of the
Eleventh International Symposium on Parameterized and Exact Computation (IPEC 2016),
pages 4:1–4:17, 2016. doi:10.4230/LIPIcs.IPEC.2016.4.
[5] Max Bannach and Till Tantau. Computing hitting set kernels by AC0 -circuits. Technical
Report arxiv:. [cs.], ArXiv e-prints, 2018. URL: http://arxiv.org/abs/.
[6] David A. Mix Barrington, Neil Immerman, and Howard Straubing. On uniformity within
NC1 . In Proceedings of the Third Annual Structure in Complexity Theory Conference, pages
47–59, 1988. doi:10.1109/SCT.1988.5262.
[7] L. Cai, J. Chen, R. G. Downey, and M. R. Fellows.
Advice classes of parameterized tractability.
Annals of Pure and Appied Logic, 84(1):119–138, 1997.
doi:10.1016/S0168-0072(95)00020-8.
[8] Marco Cesati and Miriam Di Ianni. Parameterized parallel complexity. In Proceedings of
the Fourth International Euro-Par Conference, volume 1470 of Lecture Notes in Computer
Science, pages 892–896. Springer, 1998. doi:10.1007/BFb0057945.
[9] Y. Chen and J. Flum. Some lower bounds in parameterized AC0 . In Proceedings of the
41st International Symposium on Mathematical Foundations of Computer Science (MFCS
2016), pages 27:1–27:14, 2016. doi:10.4230/LIPIcs.MFCS.2016.27.
[10] Y. Chen, J. Flum, and M. Grohe.
Bounded nondeterminism and alternation in
parameterized complexity theory. In Proceedings of the 18th IEEE Conference on
Computational Complexity (CCC 2003), pages 13–29. IEEE Computer Society, 2003.
doi:10.1109/CCC.2003.1214407.
[11] Yijia Chen, Jörg Flum, and Xuangui Huang. Slicewise definability in first-order logic
with bounded quantifier rank. In 26th EACSL Annual Conference on Computer Science Logic, CSL 2017, August 20-24, 2017, Stockholm, Sweden, pages 19:1–19:16, 2017.
doi:10.4230/LIPIcs.CSL.2017.19.
[12] M. Elberfeld, C. Stockhusen, and T. Tantau.
On the Space Complexity of Parameterized Problems: Classes and Completness. Algorithmica, 71(3):661–701, 2014.
doi:10.1007/s00453-014-9944-y.
[13] P. Erdős and R. Rado. Intersection theorems for systems of sets. Journal of the London
Mathematical Society, 1(1):85–90, 1960.
18
[14] J. Flum and M. Grohe. Describing parameterized complexity classes. In Proceedings
of the 19th Annual Symposium on Theoretical Aspects of Computer Science (STACS
2002), volume 2285 of Lecture Notes in Computer Science, pages 359–371. Springer, 2002.
doi:10.1007/3-540-45841-7_29.
[15] J. Flum and M. Grohe. Parameterized Complexity Theory. Texts in Theoretical Computer
Science. Springer, 2006. doi:10.1007/3-540-29953-X.
[16] Rolf Niedermeier and Peter Rossmanith.
An efficient fixed-parameter algorithm for 3-hitting set.
Journal of Discrete Algorithms, 1(1):89–102, 2003.
doi:10.1016/S1570-8667(03)00009-1.
[17] René van Bevern. Towards optimal and expressive kernelization for d-hitting set. Algorithmica, 70(1):129–147, September 2014. doi:10.1007/s00453-013-9774-3.
19
| 8cs.DS
|
An OvS-MultiObjective Algorithm Approach for
Lane Reversal Problem
E. G. Baquela
A.C. Olivera
GISOI - FRSN
Universidad Tecnológica Nacional
San Nicolás de los Arroyos, Argentine
[email protected]
DCEyN,UACO
Universidad Nacional de la Patagonia Austral
Santa Cruz, Argentine
[email protected]
Abstract- The lane reversal has proven to be a useful method
to mitigate traffic congestion during rush hour or in case of
specific events that affect high traffic volumes. In this work we
propose a methodology that is placed within optimization via
Simulation, by means of which a multi-objective genetic
algorithm and simulations of traffic are used to determine the
configuration of ideal lane reversal.
Keywords—optimization;
algorithms; lane reversal
simulation;
traffic;
genetic
INTRODUCTION
Despite the vast amount of investigations about this topic,
the selection of lanes to be reversed is carried out by means of
a mathematical programming model, locally to the routes
under study, without taking into account either the effect over
the whole traffic system or the dynamic behavior of such
systems (for example, spontaneous blockages due to relative
differences of speed). Reference [7] uses formalism based on
the Optimization via Simulation (OvS) to deal with the related
problem of traffic contraflow, using Genetic Algorithms in the
microscopic simulation and optimization cycle for the
evaluation of solutions, but the local environment of the
problem under study is still being restricted.
The Lane Reversal has proven to be a very effective
method to solve traffic congestion problems when there
appear what bibliography calls traffic waves (temporary
changes in the dynamics of vehicle traffic that unbalance the
flow of traffic in a network) [1]. These traffic waves are
generated, in most cases, by the dynamics of traffic during
rush hour, by the flow generated after an event (a show, for
example) or by emergencies [2].
In this work we propose an optimization model based on
OvS that uses a Multi-Objective Genetic Algorithm and
microscopic simulations of traffic to determine which lanes
are going to be reversed, evaluating the effect of each
alternative over the whole system and taking into account their
dynamic effects.
There exists a significant amount of works in which this
topic is dealt with. Reference [3] analyzes possible
applications of lane reversal and [4] formalizes the problem of
selection of lanes to be reversed with a nonlinear
programming model. Reference [2] analyzes the status of 53
implementations distributed worldwide. Reference [5] studies
the characteristics of traffic systems submitted to lane
reversal; [6] analyzes the impact of lane reversal according to
the characteristics of traffic routes and of the vehicle flow
pattern. On the other hand, [1] analyzes the bibliography that
is available and, detecting that investigations deal with lane
reversal in isolation regarding the rest of traffic policies,
proposes a multi-objective mixed linear programming model
that integrates lane reversal and traffic lights coordination, in
order to optimize the system comprehensively.
Given a traffic network T, with a distribution of the traffic
flow F (traffic flow in normal conditions) associated to a
number of I vehicles, the direction of lanes and roads that
belong to set C is intended to be modified in such a way as to
absorb the variation in the traffic flow distribution linked to a
specific pattern of circulation P, mitigating the impact in the
performance of the traffic system.
DEFINITION OF THE PROBLEM
Our decision variables (called X) are associated with the
decision of reversing or not the traffic flow in a particular
lane. Therefore, we have N binary variables, being N the
number of lanes that are available for reversal. For coding
purposes, we consider that value “1” represents directionality
reversal and “0” represents no reversal.
The definition of objective function admits many points of
view according to the nature of the disturbance. On one hand,
we want the traffic wave impact to be as small as possible
over the vehicle flow. For this reason, we choose to maximize
the average speed of the vehicles that circulate by the system
(1). On the other hand, the lane reversal involves some side
effects. Drivers may initially be out of the habit of driving by a
lane that has changed its directionality, with the subsequent
difficulty to rearrange their road map. Therefore, it is desirable
to have as few changes in the network as possible. So, we will
mitigate the number of changes to the system as well (2). Then
we have a multi-objective nonlinear programming problem
with binary decision variables, where each one represents the
decision of reversing traffic directionality in a particular lane.
Max Z1 = Σ(Speedi)/ I
(1)
Min Z2 = ΣX
(2)
RESTRICTIONS TO THE SYSTEM
The main restriction to the selection of lanes to be reversed
consists in defining in which lanes it is possible to change
directionality. This is defined by choosing which lanes are
going to be represented by decision variables and which ones
are not. Restrictions related to lanes capacity, interconnection
of them and other network characteristics are defined
implicitly in the simulation model, being left for explicit
definition only the restrictions related to lane reversal policies.
For example, lanes in which directionality reversal are
mutually exclusive, lanes with direction dependency, etc. Due
to the adopted formalism, these restrictions are modeled as
binary restrictions and they do not represent great obstacles.
between each chromosome and each variation of the original
structure of the network.
The MOGA scheme used is an adaptation of [8]. In our
research, elitism was not used; crossing in a point was
generated by binary competence and a simple mutation
scheme. As regards crossing, crossing schemes of 10, 25 and
50% of the chromosomes were tested.
The mutation that was used was of 0,5, 0,1 and 0,15 in the
chromosome level. The bit to be mutated was selected with a
uniform probability.
SYSTEM SIMULATION
The evaluation of the target function Z1 is carried out by
means of traffic microscopic simulations. In a microscopic
simulation, the individual behavior of each vehicle is
simulated, arising the global dynamics of the system from the
interactions among them. These interactions are in general
simple, since vehicles are modeled by means of a pattern of
the type "Car-Following", in which the speed of a vehicle
depends on its own speed and on the immediately preceding
car's speed. In Fig. 01, a traffic simulation is shown.
With the aim of evaluating the target function Z1, it is
necessary to configure the simulation model according to the
characteristics of the solution to be evaluated. Firstly, it is
necessary to re-structure traffic network connections so that
they can be representative of the new distribution of
directions. Then, starting from a fixed traffic flow definition
(i.e. a number of vehicles that start in an origin i towards a
destination j), it is necessary to translate such flows into an
individual route for each vehicle. Afterwards, the simulation is
carried out during a representative time horizon to the traffic
wave that was studied, giving as a result the total time of the
travel for each vehicle. Finally, the average travel time is
calculated as the arithmetical average of the previous values.
The simulation admits the definition of various additional
parameters, being the degree of random variation of vehicles'
speed the most important one. This parameter has to be
calculated on the basis of samples of the system under
analysis. However, it has the effect of allowing us to simulate
traffic congestion in a realistic way, not because of saturation
of the network's capacity but due to speed variations among
vehicles.
Fig. 01: Traffic Simulation (cross of two avenues)
MODELING BY MULTI-OBJECTIVE GENETIC
ALGORITHMS
For this problem, the chosen modeling and resolution
algorithm was a multi-objective genetic algorithm (MOGA).
MOGAs have proven to be very useful to optimize multiobjective nonlinear systems, generating solutions that adjust to
Pareto boundary of the two objectives. Besides, the use of
binary genetic algorithms is presented as a natural way to
model the decision variables of this problem, owing to the fact
that each bit of a particular chromosome represents the
reversal or not of the lane related to it. Then, given a
predefined traffic network, there exists a correlation 1:1
When simulated all traffic network, the target function is
not limited to evaluate the environment of vehicles directly
affected by the traffic wave but the impact over the total
number of vehicles that circulate by the system.
COMPLETE OPTIMIZATION PROCESS
So, the complete process of OvS can be summarized into
the following steps [8]:
1) Generation of the initial population of solutions to be
evaluated.
2) Evaluation of each solution.
2.1) Generation of a traffic network for each scenario (or
solution).
2.2) Calculation of individual routes of each vehicle for
each scenario.
2.3) Simulation of each scenario.
3) Order of solutions for the criterion of non-dominant
solutions.
4) Selection of two parents for crossover using the binary
tournament selection method.
5) Crossover
Acknowledgment
The work of E.G. Baquela was supported by the
Universidad Tecnológica Nacional, under PID 816.
The work of A. C. Olivera was supported by the Consejo
Nacional de Investigaciones Científicas y Técnicas of
Argentina and the ANPCyT under Contract PICT-2011-0639.
References
6) Generation of mutation with probability M for each
offspring.
7) Incorporation of the offspring to the new population,
repetition of 4 to 6 until generating a new population with the
same size as the initial one.
8) Repetition of steps 2 to 7.
In an attempt not to waste computer time, we keep a record
of the solutions that were evaluated in order to avoid repeating
unnecessary simulations.
EXPERIMENTS AND RESULTS
In order to test the methodology, rectangular traffic
networks were used, with pre-generated traffic flows and
traffic waves randomly generated. In Fig. 02 a part of traffic
net is shown.
Fig. 03 shows the Pareto frontier for one example,
sampling the best solution in different generations. It can be
seen that an increase in the amount of change generates an
interesting restructuring of the network traffic, which
increases average speeds in the system.
In Fig. 04 the evolution of the average speed is displayed
when the number of changes is restricted to a fixed value. A
large number of generations needed for the algorithm to detect
a set of lanes directly related to the traffic wave. Selecting
initial solutions invest close lanes to traffic flow generated by
the waves; it is possible to accelerate the process.
CONCLUSIONS
The developed lane reversal optimization model has shown
satisfactory results on the tests that were carried out, and it
proved to be a useful alternative close to reality to evaluate
lane reversal policies.
Moreover, with reference to the lane reversal problem
itself, the apparently opposite objectives of mitigating the
impact over average speed and generating as few changes as
possible are not necessarily opposite objectives.
Fig. 02: Traffic Net (two nodes and one of possible routes between them)
J. Zhao, W. Ma, Y. Liu, and X. Yang, “Integrated design and operation of
urban arterials with reversible lanes,” Transportmetrica B: Transport
Dynamics, vol. 0, pp. 1-21, May 2014.
R. Kishore Kamalanathsharma, “Reversible Lanes: State of Implementation
on a Global Level.”, Annual ATES Student Paper Competition, March 2011.
Z. Bede, G. Szabó, and T. Péter, “Optimalization of road traffic with the
applied of reversible direction lanes,” Transportation Engineering , vol.
38-1, pp. 3-8, 2010.
Z. Bede, and T. Péter, “The mathematical modeling of
reversible lane system,” Transportation Engineering , vol. 39-1, pp. 710, 2011.
L. Lambert, and B. Wolshon, “Characterization and comparison of traffic flow
on reversible roadways,” Journal of Advanced Transportation , vol. 44-2, pp.
113-122, April 2010.
W. Cheng, X. Li, and S. Liu, “Research on the traffic control and organization
coordinate applications of reversible lanes,” 3rd International Conference on
Awareness Science and Technology (iCAST) , pp. 72-75, September 2011.
A. Karoonsoontawong, “Genetic Algorithms for dynamic contraflow
problem,” Journal of Society for Transportation and Traffic Studies, vol.1-2,
pp. 1-17, June 2010.
H. W. Ding, L. Benyoucef, and X.Xie,
“Stochastic multi-objective
production-distribution network design using simulation-based optimization,”
International Journal of Production Research, vol.47-2, pp. 479-505, 2009.
Fig. 03: Pareto Frontier from generations 15, 30 and 50.
| 9cs.NE
|
1
MIMO Transmit Beampattern Matching Under
Waveform Constraints
arXiv:1802.06957v1 [math.OC] 20 Feb 2018
Ziping Zhao, Student Member, IEEE, and Daniel P. Palomar, Fellow, IEEE
Abstract
In this paper, the multiple-input multiple-output (MIMO) transmit beampattern matching problem
is considered. The problem is formulated to approximate a desired transmit beampattern (i.e., an energy
distribution in space and frequency) and to minimize the cross-correlation of signals reflected back to the
array by considering different practical waveform constraints at the same time. Due to the nonconvexity
of the objective function and the waveform constraints, the optimization problem is highly nonconvex.
An efficient one-step method is proposed to solve this problem based on the majorization-minimization
(MM) method. The performance of the proposed algorithms compared to the state-of-art algorithms is
shown through numerical simulations.
Index Terms
MIMO, waveform diversity, beampattern design, waveform constraints, nonconvex optimization.
I. I NTRODUCTION
Multiple-input multiple-output (MIMO) systems [1] have the capacity to transmit independent
probing signal or waveforms from each transmit antenna. Such waveform diversity feature leads
to many desirable properties for MIMO systems. For example, a modern MIMO radar has many
appealing features, like higher spatial resolution, superior moving target detection and better
parameter identifiability, compared to the classical phased-array radar [2]–[4].
The MIMO transmit beampattern matching problem is critically important in many fields,
like in defense systems, communication systems, and biomedical applications. This problem is
This work was supported by the Hong Kong RGC 16206315 research grant. The author Z. Zhao is supported by the Hong
Kong PhD Fellowship Scheme (HKPFS).
The authors are with the Department of Electronic and Computer Engineering, The Hong Kong University of Science and
Technology (HKUST), Clear Water Bay, Kowloon, Hong Kong (e-mail: [email protected]; [email protected]).
2
concerned with designing the probing waveforms to approximate a desired antenna array transmit
beampattern (i.e., an energy distribution in space and frequency) and also to minimize the the
cross-correlation of the signals reflected back from various targets of interest by considering some
practical waveform constraints. The MIMO transmit beampattern matching problem appears to be
difficult from an optimization point of view because the existence of the fourth-order nonconvex
objective function and the possibly nonconvex waveform constraints which are used to represent
desirable properties and/or enforced from an hardware implementation perspective [5].
In [6], the MIMO transmit beampattern matching problem was formulated to minimize the
difference between the designed beampattern and the desired one. The formulation in [6] was
modified in [7], [8] by introducing the cross-correlation between the signals. And in [8], the
authors proposed to design the waveform covariance matrix to match the desired beampattern
through semidefinite programming. A closed-form waveform covariance matrix design method
was also proposed based on discrete Fourier transform (DFT) coefficients and Toeplitz matrices
in [9], [10]. But such kind of methods can perform badly for small number of antennas. After
the waveform covariance matrix is obtained, other methods should be applied to synthesize a
desired waveform from its covariance matrix. For example, a cyclic algorithm was proposed in
[11] to synthesize a constant modulus waveform from its covariance matrix. These methods are
usually called two-steps methods. In practice, they could become inefficient and suboptimal if
more waveform constraints are considered.
In [12], it was found that directly designing the waveform to match the desired beampattern
can give a better performance, which is referred to as the one-step method. But the method
in [12] is tailored to the constant modulus constraint and can be slow in convergence. In [13],
the problem was solved based on the alternating direction method of multipliers (ADMM) [14].
However, again the proposed algorithm is only designed for dealing with unimodulus constraint.
The majorization-minimization (MM) method [15], [16] has shown its great efficiency in deriving fast and convergent algorithms to solve nonconvex problems in many different applications
[17], [18]. In this paper, we propose a one-step method to directly solve the MIMO transmit
beampattern matching problem based on the MM method by considering different waveform
constraints. The performance of our algorithms compared to the existing algorithms is verified
through numerical simulations.
3
)*+*)*,-*
./)*-0/1,
!
" &'( !
&
'
"#$%#
"#$%#
"#
(
!!!!!!
"#$%#
$
%
Fig. 1. MIMO transceiver with M antennas and θ is the spacial direction of interest.
II. MIMO T RANSMIT B EAMPATTERN M ATCHING P ROBLEM F ORMULATION
A colocated MIMO radar [19] with M transmit antennas in a uniform linear array (ULA),
as shown in Fig. 1, is considered. Each transmit antenna can emit a different waveform xm (n)
with m = 1, 2, . . . , M, n = 1, 2, . . . , N, where N is the number of samples. Let x (n) =
h
iT
h
x1 (n) , x2 (n) , . . . , xM (n) be the nth sample of the M transmit waveforms and x = xT (1) ,
iT
xT (2) , . . . , xT (N) denote the waveform vector.
The signal at a target location with angle θ (θ ∈ Θ, which is the angle set) is represented by
M
X
e−jπ(m−1) sin θ xm (n) = aT (θ) x (n) , n = 1, . . . , N,
m=1
T
where a (θ) is the transmit steering vector written as a (θ) = 1, e−jπ sin θ , . . . , e−jπ(M −1) sin θ .
Then, the power for the probing signal x at location θ which is named the transmit beampattern
can be written as follows:
P (θ, x)
=
N
X
aT (θ) x (n)
n=1
=
∗
H
IN ⊗ aT (θ) x
=xH
aT (θ) x (n)
IN ⊗ aT (θ) x
IN ⊗ a∗ (θ) aT (θ) x = xH A (θ) x,
where A (θ) = IN ⊗ a∗ (θ) aT (θ).
4
Suppose there are K targets of interest, and then the spatial cross-correlation sidelobes (crosscorrelation beampattern) between the probing signals at locations θi and θj (i 6= j, i, j = 1, . . . , K
and θi , θj ∈ Θ) is given by
Pcc (θi , θj , x)
=
N
X
aT (θi ) x (n)
n=1
=
∗
aT (θj ) x (n)
H
IN ⊗ aT (θi ) x
=xH
IN ⊗ aT (θj ) x
IN ⊗ a∗ (θi ) aT (θj ) x = xH A (θi , θj ) x,
where A (θi , θj ) = IN ⊗ a∗ (θi ) aT (θj ).
The objective of the transmit beampattern matching problem is as follows: i) to match a
desired transmit beampattern denoted as p (θ), which can be formulated as follows1:
X
J (α, x) =
ω (θ) |αp (θ) − P (θ, x)|2 ,
(1)
θ∈Θ
where ω (θ) ≥ 0 is the weight for the direction θ; and ii) to minimize the cross-correlation
between the probing signals at a number of given target locations due to the fact that the statistical
performance of adaptive MIMO radar techniques rely on the cross-correlation beampattern, which
is given as
E (x) =
X
|Pcc (θi , θj , x)|2 .
(2)
θi ,θj ∈Θ, i6=j
Then, by considering J (α, x) and E (x), the MIMO transmit beampattern matching problem is
formulated as follows:
minimize
α,x
f (α, x) , J (α, x) + ωcc E (x)
(3)
subject to x ∈ X , X0 ∩ (∩i Xi ) ,
where ωcc controls the sidelobe term, X generally denotes the waveform constraint, and X0 =
x ∈ CM N | kxk22 = c2e representing the total transmit energy (power) constraint. We are
also interested in other practical waveform constraints:
i) Constant modulus constraint is to prevent the non-linearity distortion of the power amplio
n
ce
√
fier to maximize the efficiency of the transmitter, which is given by X1 = x | |x (l)| = cd = M N
for l = 1, . . . , MN;
1
Variable α is introduced since p (θ) is typically given in a “normalized form” and we want to approximate a scaled version
of p (θ), not p (θ) itself.
5
ii) Peak-to-Average Ratio (PAR) constraint is the ratio of the peak signal power to its
average power (PAR (x) =
max|x(l)|2
kxk22 /M N
with 1 ≤ PAR (x) ≤ MN). The PAR (x) is constrained to
a small threshold, so that the analog-to-digital and digital-to-analog converters can have lower
dynamic range, and fewer linear power amplifiers are needed. Since X0 , the PAR constraint is
o
n
ce
≤
c
≤
c
for l = 1, . . . , MN;
X2 = x | |x (l)| ≤ cp , √M
p
e
N
iii) Similarity constraint is to allow the designed waveforms to lie in the neighborhood
of a reference one which already can attain a good performance [20], which is denoted as
o
n
X3 = x | |x − xref | ≤ cǫ , 0 ≤ cǫ ≤ √M2 N .
Problem (3) is a constrained nonconvex problem due to the nonconvex objective and con-
straints. We are trying to solve it by using efficient nonconvex optimization methods.
III. P ROBLEM S OLVING
VIA
T HE MM M ETHOD
A. The Majorization-Minimization (MM) Method
The MM method [15], [16], [21] is a generalization of the well-known EM method. For an
optimization problem given by
minimize
f (x)
x
subject to x ∈ X ,
instead of dealing with this problem directly which could be difficult, the MM-based algorithm
solves a series of simpler subproblems with surrogate functions that majorize f (x) over X . More
specifically, starting from an initial point x(0) , it produces a sequence x(k) by the following
update rule:
x(k) ∈ arg min f x, x(k−1) ,
x∈X
where the surrogate majorizing function f x, x(k)
f x(k) , x(k) = f x(k) ,
f x, x(k) ≥ f (x) ,
′
f x(k) , x(k) ; d = f ′ x(k) ; d ,
satisfies
∀x(k) ∈ X ,
∀x, x(k) ∈ X ,
∀d, s.t. x(k) + d ∈ X .
The objective function value is monotonically nonincreasing at each iteration. To use the MM
method, the key step is to find a majorizing function to make the subproblem easy to solve,
which will be discussed in the following subsections.
6
B. Majorization Steps For The Beampattern Matching Term J (α, x)
In this section, we discuss the majorization steps, i.e., how to construct a good majorizing
function for the beampattern matching term J (α, x) in (1). First, we have
X
J (α, x) =
ω (θ) |αp (θ) − P (θ, x)|2
θ∈Θ
=α2
X
ω (θ) p2 (θ) − 2α
θ∈Θ
X
ω (θ) p (θ) P (θ, x) +
θ∈Θ
X
ω (θ) (P (θ, x))2 ,
θ∈Θ
which is a quadratic function in variable α. Then, it follows that the minimum of J (α, x) is
attained when
α (x) =
X
ω (θ) p (θ) P (θ, x) /
θ∈Θ
X
ω (θ) p2 (θ) .
θ∈Θ
Substituting α (x) back into J (α, x) and considering
H
P (θ, x) = Tr xxH A (θ) = vec xxH vec (A (θ)) ,
we get
J (x) =
X
θ∈Θ
2
H H
ω (θ) vec xx
vec (A (θ)) −
X
ω (θ) p2 (θ)
θ∈Θ
!−1
X
2
H H
× vec xx
vec
ω (θ) p (θ) A (θ)
where
θ∈Θ
H
=vec xxH HJ vec xxH ,
HJ =
X
ω (θ) vec (A (θ)) vec (A (θ))H −
θ∈Θ
× vec
X
θ∈Θ
X
θ∈Θ
ω (θ) p2 (θ)
−1
X
H
ω (θ) p (θ) A (θ) vec
ω (θ) p (θ) A (θ) ,
θ∈Θ
and it is easy to see that J (x) is a quartic function in x. Next, we introduce a useful lemma.
Lemma 1. Let A ∈ HK and B ∈ HK such that B A. At any point x0 ∈ CK , the quadratic
function xT Ax is majorized by xH Bx + 2Re xH (A − B) x0 + xH
0 (B − A) x0 .
Proof: Notice that (x − x0 )H (B − A) (x − x0 ) ≥ 0.
Based on Lemma 1, we can choose ψJ,1 ≥ λmax (HJ ), and because ψJ,1 I HJ , at iterate
x(t) we have
H
J (x) ≤ψJ,1 vec xxH vec xxH
H
+ 2Re vec xxH (HJ − ψJ,1 I) vec x(t) x(t)H
H
+ vec x(t) x(t)H (ψJ,1 I − HJ ) vec x(t) x(t)H ,
7
where since vec xxH
H
vec xxH = kxk42 = c4e , the first term is just a constant. Then after
ignoring the constant terms, we get the following majorizing function for J (x):
H
J 1 x, x(t) ≃ 2Re vec xxH (HJ − ψJ,1 I) vec x(t) x(t)H ,
where “≃” stands for “equivalence” up to additive constants. Substituting HJ back into function
J 1 x, x(t) and dropping the constants, we have
J 1 x, x(t) ≃ 2xH MJ − ψJ,1 x(t) x(t)H x,
(4)
P
where MJ = θ∈Θ ω (θ) P θ, x(t) − p (θ) α x(t) A (θ). It is easy to see that after majoriza
tion, the majorizing function J 1 x, x(t) becomes quadratic in x rather than quartic in J (x).
However, using this function as the objective to solve is still hard due to the waveform constraint
X .2 So we propose to majorize J 1 x, x(t) again to simplify the problem to solve in each
iteration. Thus, we can consider choosing ψJ,2 ≥ λmax (MJ ) ≥ λmax MJ − ψJ,1 x(t) x(t)H for
majorization, where we can have the following useful property.
Lemma 2. [22], [23] Define
X
B=
ω (θ) P θ, x(t) − p (θ) α x(t) a∗ (θ) aT (θ)
θ∈Θ
=
b0
b1
..
.
b∗1
···
..
.
b0
.. ..
.
.
bM −1 . . .
b1
b∗M −1
..
.
b∗1
b0
,
T
which is Hermitian Toeplitz, F as a 2M×2M FFT matrix, and b = b0 , b1 , . . . , bM −1 , 0, b∗M −1 , , . . . , b∗1 .
Then, we have MJ = IN ⊗ B, λmax (MJ ) = λmax (B), and
1
max µ2i + max µ2i−1 ,
λmax (B) ≤ λµ =
1≤i≤M
2 1≤i≤M
where µ = Fb, which is the discrete Fourier transform for b.
Lemma 2 provides an easy way for the computation of ψJ,2 . Based on Lemma 1 and using
ψJ,2 = λµ , the majorizing function J 1 x, x(t) can be further majorized as
J 1 x, x(t) ≤2ψJ,2 xH x + 4Re xH MJ − ψJ,1 x(t) x(t)H − ψJ,2 I x(t)
+ 2x(t)H ψJ,2 I − MJ + ψJ,1 x(t) x(t)H x(t) ,
2
It is a NP-hard unimodular quadratic program even only considering X1 .
8
where since kxk22 = c2e , the first term is a constant. Then by ignoring the constant terms, the
objective becomes a linear majorizing function at iterate x(t) as follows:
J 2 x, x(t) ≃ −4Re xH yJ ,
(5)
where yJ = − (MJ − c2e ψJ,1 I − ψJ,2 I) x(t) .
C. Majorization Steps For The Sidelobe Term E (x)
To deal with the sidelobe term E (x) in (2), the majorization steps are similar to J (x). First,
we have
E (x) =
X
|Pcc (θi , θj , x)|2
θi ,θj ∈Θ, i6=j
=vec xxH
HE vec xxH ,
vec (A (θi , θj )) vec (A (θi , θj ))H . Then, based on Lemma 1, by choos
≥ λmax (HE ) and ψE,2 ≥ λmax ME − ψE,1 x(t) x(t)H , we can get the majorizing
where HE =
ing ψE,1
H
P
θi ,θj ∈Θ, i6=j
functions at iterate x(t) written as follows:
E 1 x, x(t) ≃2xH ME − ψE,1x(t) x(t)H x
(6)
≤E 2 x, x(t)
≃ − 4Re xH yE ,
P
where ME = θi ,θj ∈Θ, i6=j Pcc θj , θi , x(t) A (θi , θj ) and yE = − (ME − c2e ψE,1 I − ψE,2 I) x(t) .
D. Solving The Majorized Subproblem in MM
By combing the two majorizing functions J 2 x, x(t) and E 2 x, x(t) , the overall majorizing
function at iterate x(k) for the objective f (x) is given as follows:
f (x) ≤f x, x(t)
=J 2 x, x(t) + ωcc E 2 x, x(t)
≃ − 4Re xH yJ − 4ωcc Re xH yE
= − Re xH y ,
where
y = − 4 MJ + ωcc ME − c2e (ψJ,1 + ωcc ψE,1 ) I
− (ψJ,2 + ωcc ψE,2 ) I) x(t) .
9
Finally, by majorizing the objective function in (3) using the MM method, the subproblem
we need to solve at each iteration is given as follows:
minimizex f x, x(t) ≃ −Re xH y
(7)
subject to x ∈ X .
For problem (7), as to different interested waveform constraints, closed-form optimal solutions
x⋆ can be derived, which are summarized in the following lemma.
Lemma 3. i) For fixed energy constraint (i.e., X = X0 ), x⋆ = ce y/ kyk2 ; ii) for constant
modulus constraint (i.e., X = X1 ), x⋆ = cd ej arg(y) ;3 iii) for fixed energy with PAR constraint
(i.e., X = X0 ∩ X2 ), the solution x⋆ can be found in [24, Alg. 2]; iv) for constant modulus with
similarity constraint (i.e., X = X1 ∩ X3 ), the solution x⋆ can be found in [25].
E. The MM-Based Beampattern Matching Algorithm
Based on the MM method, in order to solve the original problem (3), we just need to iteratively
solve the subproblem (7) with a closed-form solution update in Lemma 3 at each iteration. The
overall algorithm is summarized as follows.
Input: a (θ), p (θ), x(0) and t = 0.
Repeat
1. Compute MJ , ME , ψJ,1 , ψE,1 , ψJ,2 , ψE,2 and y;
2. Update x(t) in a closed-form according to Lemma 3;
3. t = t + 1;
Until x and f (x) satisfy a termination criterion.
Output: α, x.
IV. N UMERICAL S IMULATIONS
The performance of the proposed algorithm for MIMO transmit beampattern matching is
evaluated by numerical simulations. A colocated MIMO radar system is considered with a ULA
comprising M = 10 antennas with half-wavelength spacing between adjacent antennas. Without
loss of generality, the total transmit power is set to c2e = 1. Each transmit pulse has N = 32
samples. The range of angle is Θ = (−90◦ , 90◦ ) with spacing 1◦ under which the weight
3
The operation arg (y) is applied element-wise for y.
10
ω (θ) = 1 for θ ∈ Θ, and ωcc = 0, which is the same setting as [13]. We consider a desired
beampattern with three targets or mainlobes (K = 3) at θ1 = −40◦ , θ2 = 0◦ , θ3 = 40◦ , and each
width of them is △θ = 20◦ . The desired beampattern is
1, θ ∈ [θk − △θ/2, θk + △θ/2] , k = 1, 2, K
p (θ) =
0, otherwise.
We compare the convergence property over iterations of the objective function for the beam-
pattern matching problem under unimodulus waveform constraint by using the proposed MMbased algorithm (denoted as MM-based algorithm (prop.)) and the ADMM-based algorithm in
[13] (denoted as ADMM-based algorithm) , which is shown in Fig. 2.
160
MM-based algorithm (prop.)
ADMM-based algorithm
140
Objective Function Value
120
100
80
60
40
20
0
0
10
10
1
10
2
Iterations
Fig. 2. Convergence comparison for objective function value.
As shown in Fig. 2, the MM-based algorithm can have a monotonic convergence property.
And it can converge within 20 iterations which is faster than the benchmark algorithm.
Then, we also compare the matching performance of the designed beampatterns in terms of
the mean-squared error (MSE) defined as
MSE (P (θ, x)) = E
"
X
θ∈Θ
ω (θ) |αp (θ) − P (θ, x)|
2
#
.
In Fig. 3, we show the simulation results for MSE (P (θ, x)) by using different design methods.
11
2.4
desired beampattern
MM-based algorithm (prop.) MSE=24.6264
ADMM-based algorithm MSE=24.8211
2.2
2
1.8
Beampattern
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
-90 -80 -70 -60 -50 -40 -30 -20 -10
0
10
20
30
40
50
60
70
80
90
θ (degree)
Fig. 3. Transmit beampattern design with 3 targets .
From Fig. 3, we can see that compared to the benchmark, our proposed algorithm can have
a tighter matching performance and can obtain a lower MSE. Based on these, the proposed
algorithm is validated.
V. C ONCLUSIONS
This paper has considered the MIMO transmit beampattern matching problem. Efficient algorithms have been proposed based on the MM method. Numerical simulations show that the
proposed algorithms are efficient in solving the beampattern matching problem and can obtain
a better performance compared to the the state-of-art method.
R EFERENCES
[1] J. Li and P. Stoica, MIMO Radar Signal Processing.
Wiley, 2008.
[2] D. Bliss and K. Forsythe, “Multiple-input multiple-output (MIMO) radar and imaging: degrees of freedom and resolution,”
in Proc. of the 37th Asilomar Conf. on Signals, Systems and Computers., vol. 1.
IEEE, 2003, pp. 54–59.
[3] E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, and R. Valenzuela, “MIMO radar: An idea whose time has
come,” in Proc. of the 2004 IEEE Radar Conf.. IEEE, 2004, pp. 71–78.
[4] K. Forsythe, D. Bliss, and G. Fawcett, “Multiple-input multiple-output MIMO radar: Performance issues,” in Proc. of the
38th Asilomar Conf. on Signals, Systems and Computers., vol. 1. IEEE, 2004, pp. 310–315.
[5] M. Skolnik, “Radar handbook,” New York: McGraw-Hill, 1990.
[6] D. R. Fuhrmann and G. San Antonio, “Transmit beamforming for MIMO radar systems using partial signal correlation,”
in Proc. of the 38th Asilomar Conf. on Signals, Systems and Computers., vol. 1.
IEEE, 2004, pp. 295–299.
[7] ——, “Transmit beamforming for MIMO radar systems using signal cross-correlation,” IEEE Trans. Aerosp. Electron.
Syst., vol. 44, no. 1, 2008.
[8] P. Stoica, J. Li, and Y. Xie, “On probing signal design for MIMO radar,” IEEE Trans. Signal Process., vol. 55, no. 8, pp.
4151–4161, 2007.
12
[9] J. Lipor, S. Ahmed, and M.-S. Alouini, “Fourier-based transmit beampattern design using MIMO radar,” IEEE Trans.
Signal Process., vol. 62, no. 9, pp. 2226–2235, 2014.
[10] T. Bouchoucha, S. Ahmed, T. Al-Naffouri, and M.-S. Alouini, “DFT-based closed-form covariance matrix and direct
waveforms design for MIMO radar to achieve desired beampatterns,” IEEE Trans. Signal Process., vol. 65, no. 8, pp.
2104–2113, 2017.
[11] P. Stoica, J. Li, and X. Zhu, “Waveform synthesis for diversity-based transmit beampattern design,” IEEE Trans. Signal
Process., vol. 56, no. 6, pp. 2593–2598, 2008.
[12] Y.-C. Wang, X. Wang, H. Liu, and Z.-Q. Luo, “On the design of constant modulus probing signals for MIMO radar,”
IEEE Trans. Signal Process., vol. 60, no. 8, pp. 4432–4438, 2012.
[13] Z. Cheng, Z. He, S. Zhang, and J. Li, “Constant modulus waveform design for MIMO radar transmit beampattern,” IEEE
Trans. Signal Process., vol. 65, no. 18, pp. 4912–4923, 2017.
[14] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating
direction method of multipliers,” Found. Trends Mach. Learn., vol. 3, no. 1, pp. 1–122, 2011.
[15] D. R. Hunter and K. Lange, “A tutorial on MM algorithms,” Amer. Statist., vol. 58, no. 1, pp. 30–37, 2004.
[16] Y. Sun, P. Babu, and D. P. Palomar, “Majorization-minimization algorithms in signal processing, communications, and
machine learning,” IEEE Trans. Signal Process., vol. 65, no. 3, pp. 794–816, 2016.
[17] J. Song, P. Babu, and D. P. Palomar, “Optimization methods for designing sequences with low autocorrelation sidelobes,”
IEEE Trans. Signal Process., vol. 63, no. 15, pp. 3998–4009, 2015.
[18] Z. Zhao and D. P. Palomar, “Mean-reverting portfolio with budget constraint,” IEEE Trans. on Signal Process., vol. PP,
no. 99, p. 1, 2018.
[19] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal Process. Mag., vol. 24, no. 5, pp. 106–114, 2007.
[20] J. Li, J. R. Guerci, and L. Xu, “Signal waveform’s optimal-under-restriction design for active sensing,” IEEE Signal
Process. Lett., vol. 13, no. 9, pp. 565–568, Sep. 2006.
[21] M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence analysis of block successive minimization methods for
nonsmooth optimization,” SIAM J. Optim., vol. 23, no. 2, pp. 1126–1153, 2013.
[22] P. Jorge and S. Ferreira, “Localization of the eigenvalues of Toeplitz matrices using additive decomposition, embedding
in circulants, and the fourier transform,” in Proc. Symp. Syst. Identif., vol. 3, 1994, pp. 271–275.
[23] R. A. Horn and C. R. Johnson, Matrix analysis.
Cambridge Univ. Press, 1990.
[24] J. A. Tropp, I. S. Dhillon, R. W. Heath, and T. Strohmer, “Designing structured tight frames via an alternating projection
method,” IEEE Trans. Inf. Theory, vol. 51, no. 1, pp. 188–209, 2005.
[25] L. Zhao and D. P. Palomar, “Maximin joint optimization of transmitting code and receiving filter in radar and
communications,” IEEE Trans. Signal Process., vol. 65, no. 4, pp. 850–863, 2017.
| 7cs.IT
|
RELATING F-SIGNATURE AND F-SPLITTING RATIO OF PAIRS
USING LEFT-DERIVATIVES
arXiv:1206.4293v4 [math.AC] 27 Nov 2012
ERIC CANTON
Abstract. We first relate an approximate nth -order left derivative of s(R, f t ) at the F-pure threshold c
to the F-splitting ratio rF (R, f c ). Next, we apply the methods developed by Monsky and Teixeira in their
investigation of syzygy gaps and p-fractals to obtain uniform convergence of the F-signature when f is
a product of distinct linear polynomials in two variables. Finally, we explicitly compute the F-signature
function for several examples using Macaulay2 code outlined in the last section of this paper.
1. Introduction
Let (R, m, k) be a local ring containing a field k, and assume char(R) = p > 0. For simplicity we assume
that k = k p (k is perfect); this is always the case when our field is finite. For an ideal I ⊆ R we denote by
e
Pn
e
I [p ] = { 1 ci rip : ci ∈ R and ri ∈ I} the ideal generated by all pe -th powers of elements in I. Because
char(R) is positive this is in general an ideal distinct from the normal pe -th power of the ideal.
The Frobenius endomorphism of R is the map F : R → R which takes r 7→ rp . We will write R for
the domain of this map and F∗ R for the codomain, when it is important to distinguish the two (although
they are isomorphic as abelian groups). We define an R-module structure on F∗ R by r.x = rp x for r ∈ R
and x ∈ F∗ R. A reduced ring R is F-finite if F∗ R is finitely generated as an R-module. Similarly, we can
consider the e-th iterated Frobenius map F e : R → R and define an R-module structure for this iterated
Frobenius map; the codomain is denoted here F∗e R. Note that if R is F-finite, then F∗e R is finitely generated
for all e ∈ N.
Important classes of F-finite rings include polynomial rings over a field in positive characteristic, quotients,
and localizations of these rings. For example consider S = k[x1 , . . . , xn ]. Because F∗ S is generated as an
S-module by the products Πni=1 xdi i with each di ≤ p − 1, we conclude that S is F-finite. Similarly, when R is
an F-finite ring and I ⊆ R is an ideal, then F∗ (R/I) is generated over R/I by the images of the generators of
F∗ R over R. Thus any quotient of an F-finite ring is also F-finite. We also have that localizations of F-finite
rings are again F-finite (see R. Fedder, Lemmas 1.4 and 1.5 from [4] for more information.)
e
When R is reduced, we naturally identify F∗e R with the ring R1/p of pe -th roots of elements of R by
1/pe
1/pe
a
sending r 7→ r
. Next, we will decompose this module as R
= R ⊕ M where M has no free summands
e
of R. We define ae to be the maximal rank of any free decomposition of R1/p . A famous result of E. Kunz
e
(which we refer to as Kunz’s Theorem) states that for all p , we have that ae ≤ ped with equality if and only
if R is regular if and only if the Frobenius map is flat. This result prompted C. Huneke and G. Leuschke to
define [6] the F-signature as the following limit.
1.1. Definition. Let (R, m, k) be an F-finite, d-dimensional reduced local ring and k = R/m a field of
positive characteristic p. The F-signature of R is
ae
s(R) := lim ed
e→∞ p
Recently, K. Tucker showed that this limit exists in full generality [11]. It is 1 if and only if R is regular,
and so the F-signature serves as a measure to which R fails to be regular in comparison to other local rings
of the same dimension. If for some e (and thus every e) we have that ae 6= 0 then we say that R is F-pure.
Often it happens that the limit s(R) is zero. I. Aberbach and F. Enescu defined [1] the Frobenius splitting
Key words and phrases. F-signature, F-splitting ratio, F-pure, p-fractals, syzygy gaps.
1
dimension (F-splitting dimension) sdim(R) = m to be the greatest integer such that
ae
lim em
e→∞ p
is greater than zero. This limit exists by [11]. The corresponding limit is defined in [1] to be the Frobenius
splitting ratio (F-splitting ratio) denoted as rF (R). Note that if the splitting dimension sdim(R) = dim(R),
then rF (R) = s(R).
In [2] M. Blickle, K. Schwede, and K. Tucker introduce the concept of F-signature to the pair (R, f t ),
where t ∈ [0, ∞) is a real number and f ∈ R is nonzero.
1.2. Definition. Let f ∈ R be a nonzero element in an F-finite regular local ring (R, m, k). Let d = dim(R)
be the Krull dimension of R, and t ∈ [0, ∞) a positive real number. The F-signature of the pair (R, f t ) is
defined to be the limit
R
1
t
s(R, f ) := lim ed `R
e→∞ p
m[pe ] : f dt(pe −1)e
The supremum over all t such that s(R, f t ) is F-pure is called the F-pure threshold of f and will be
denoted in this paper as FPT(f ).
An important theorem regarding computation of the F-signature when t = a/ps is Proposition 4.1 found
in [3], which states that
1
R
a/ps
s(R, f
) = sd `R
p
m[ps ] : f a
s
which is to say that in this case, we do not need to take a limit: a single length suffices to compute s(R, f a/p ).
In much the same way, we may now define the F-splitting dimension of the pair (R, f t ), where t is a
rational number whose denominator is not divisible by p, to be the greatest integer m such that
1
R
lim sup em `R
m[pe ] : f dt(pe −1)e
e→∞ p
is nonzero. We again define this limit to be the Frobenius splitting ratio (F-splitting ratio) of the
pair (R, f t ), denoted here rF (R, f t ). The first result in this paper relates an approximate higher-order
left derivative of s(R, f t ) at c = F P T (f ) to a constant multiple of rF (R, f c ) when p does not divide the
denominator of c. This is a generalization of Theorem 2.1 in the next section, which can be found in [3] as
Theorem 4.2 relating the first left derivative D− s(R, f 1 ) to the F-signature s(R/hf i).
Computing the F-splitting dimension is aided by formation of a special ideal, called the splitting prime
e
e
of (R, f t ). This ideal is defined to be the maximal ideal J ⊆ R such that f dt(p −1)e J ⊆ J [p ] for all e > 0.
As the name suggests, when it is a proper ideal of R it is a prime ideal. This result can be found in ([8],
Corollary 6.4) but it is presented again here for the reader’s convenience, albeit with a different proof.
1.3. Lemma. Let (R, m, k) be an F-finite regular local ring, and f ∈ R a nonzero element of R. Take
t ∈ [0, ∞) and let P be the splitting prime of the pair (R, f t ). If P is proper, then it is a prime ideal.
Proof. Suppose P 6= R is the splitting prime of (R, f t ) and let c ∈ R \ P . We wish to show that (P : c) = P
which implies that P is a prime ideal.
e
e
I claim that if J is an ideal of R satisfying f dt(p −1)e J ⊆ J [p ] , then for any r ∈ R \ J we also have that
e
e
e
e
e
f dt(p −1)e (J : r) ⊆ (J [p ] : rp ). To see this, let g ∈ (J : r) so that gr ∈ J. Then f dt(p −1)e gr ∈ J [p ] by
e
e
e
e
e
e
the assumption we made on J. Then of course f dt(p −1)e grp ∈ J [p ] , and so f dt(p −1)e g ∈ (J [p ] : rp ),
establishing the claim.
By Kunz’s theorem, we know that the Frobenius endomorphism on R is flat when R is regular, so we may
tensor the exact sequence
0 −−−−→ R/(J : r) −−−−→ R/J −−−−→ R/(J + r) −−−−→ 0
with F∗e R
dt(pe −1)e
e
e
e
to conclude that in this case (J [p ] : rp ) = (J : r)[p ] . Therefore, whenever J is an ideal satisfying
e
e
e
f
J ⊆ J [p ] and r ∈ R \ J is any element, we have that f dt(p −1)e (J : r) ⊆ (J : r)[p ] .
Because P 6= R is the splitting prime for the pair (R, f t ), it is contained in no other ideal satisfying
e
dt(pe −1)e
f
J ⊆ J [p ] . This implies that (P : c) = P and so P is prime.
2
The splitting prime is related to the F-splitting dimension sdim(R, f t ) by the following theorem, which
can be found in [2].
1.4. Theorem ([2] Theorem 4.2). Let (R, m, k) be an F-finite d-dimensional regular local ring, f ∈ R a
nonzero element, t ∈ [0, ∞) and let P be the splitting prime of the pair (R, f t ). Then
sdim(R, f t ) = dim(R/P )
In the third section of this paper, we provide an application of Monsky and Teixeira’s work on p-fractals
to compute the F-signature when f is a homogeneous polynomial in two variables. Specifically, we prove that
s(R, f t ) converges uniformly to a quadratic polynomial in t as p → ∞. We provide an example where we
can explicitly compute the F-pure threshold for f = xy(x + y) and apply the p-fractal techniques mentioned
before to compute the left derivative at the F-pure threshold.
Finally, several computational examples (using Macaulay2) are included in the second-to-last section of
this paper; algorithms that were used to compute these examples comprise the final section.
Special thanks to Karl Schwede for many insightful and encouraging discussions over the course of this
work and preparation of this document. In particular, the proofs of 1.3 and 2.2 were discussed with him. I
would like to thank Kevin Tucker for a stimulating discussion of some of the results presented here and his
suggestions regarding this paper, and I would also like to thank Florian Enescu, who suggested applying the
results of Teixeira’s thesis to F-signature in two variables. Thanks to Wenliang Zhang and Lance Miller for
their useful critiques of this paper.
2. F-Splitting Ratio of Principal Ideals
In this section, we assume that R is an F-finite regular local domain with Krull dimension d. The
F-signature of the pair (R, f t ) was shown recently in [3] to be continuous and convex on [0, ∞), thus
differentiable almost everywhere on the domain [0, ∞). Indeed, the authors of [3] proved the following
theorem:
2.1. Theorem ([3] Theorem 4.3). If (R, m, k) is an F-finite d-dimensional local domain and f ∈ R is a
nonzero element, then
D− s(R, f 1 ) = −s(R/hf i)
D+ s(R, f 0 ) = −eHK (R/hf i)
In this section, we generalize the first part of the above result to an approximate left nth -derivative at the
F-pure threshold of our nonzero element f . Let F P T (f ) = c be the F-pure threshold of f , and denote by
n = d − sdim(R, f c ). We make the following assumption on c in this section only.
Assume that c is a rational number whose denominator is not divisible by p.
In particular, the pair (R, f c ) is sharply F-pure [8] and so the splitting prime is proper. Because p does
not divide the denominator of c, we can write c = a/(ps − 1) for an appropriate a and s. To arrive at an
approximation of c, we write
a
−1
a(p(e−1)s + p(e−2)s + · · · + 1)
= s
(p − 1)(p(e−1)s + p(e−2)s + · · · + 1)
c=
=
ps
a(p(e−1)s + p(e−2)s + · · · + 1)
pes − 1
Now let Ke = (p(e−1)s + p(e−2)s + · · · + 1), so that above we have c = aKe /(pes − 1). Define then
te =
aKe
pes
3
and note that te → c as e → ∞. We compute the following limit, which serves as a sort of approximate left
nth -derivative:
es s
n
p (p − 1)
R
1
s(R, f te )
=
lim
sup
−
`
lim sup
R
n
a
pesd
m[pes ] : hf i(pes −1)c
e→∞
te →c (te − c)
s
n
R
p −1
1
= −
lim sup es(d−n) `R
a
m[pes ] : hf iaKe
e→∞ p
= (−c)−n rF (R, f c )
Since n is the smallest integer such that the above limit is nonzero, if sdim(R, f t ) < dim(R) − 1 then we
have n = dim(R) − sdim(R, f t ) ≥ 2. Because we know that s(R, f t ) is differentiable almost everywhere, if
F P T (f ) = c as above and we have sdim(R, f c ) < dim(R) − 1. Therefore the left derivative D− s(R, f c ) = 0.
The proof in the non-square-free case was suggested by Wenliang Zhang.
2.2. Theorem. Suppose f ∈ R and (R, f c ) is sharply F-pure for c < 1. Write f = uf1n1 · · · frnr , where u is
a unit and each fi is irreducible. If f is not square free, assume that F P T (fi ) = ci < 1 for each i such that
ni ≥ 2. Then the left derivative D− s(R, f c ) = 0.
Proof. Let P be the splitting prime of the pair (R, f c ). I claim that with the hypotheses of the theorem,
we have dim(R/P ) < dim(R) − 1. Because (R, f c ) is sharply F-pure, this implies that P is proper and so
prime. Towards a contradiction assume that ht(P ) = 1. We will need the following lemma:
2.3. Lemma. If P is the splitting prime of (R, f c ) then P RP is the splitting prime of (RP , fPc ) where fP is
the image of f in RP .
dc(pe −1)e
e
(P RP ) ⊆ (P RP )[p ] by definition of P . Also, prime ideals of
Proof of lemma 2.3. P RP satisfies fP
RP are in bijective correspondence with primes of R contained in P ; thus because P RP is maximal in RP ,
it must be the splitting prime of (RP , fPc ).
We now consider two cases: either f is square free, or F P T (fi ) = ci < 1 for all i such that ni ≥ 2.
Because ht(P ) = 1, this implies that P = hfi i for some i. Localize at P and suppose that f is square-free
or ni = 1. This tells us that P RP = f RP , since RP is a DVR with maximal ideal P RP and fP cannot
be a unit, since fP = (uf1n1 · · · fi1 · · · frnr )hfi i = vfi , where v ∈ RP is a unit and the image of fi generates
the maximal ideal. Note that by assumption that c < 1 is a rational number whose denominator is not
divisible by p, there exist infinitely many e such that c(pe − 1) is an integer re , and each such re satisfies
that c(pe − 1) < re + 1 < (pe − 1) + 1. This implies that
fPre P RP 6⊆ (P RP )[p
e
]
contradicting that P RP is the splitting prime of RP .
Suppose then that P = hfi i and ni ≥ 2, and recall that by assumption F P T (fi ) = ci < 1. This gives
that c ≤ ci /ni < 1/ni , so cni (pe − 1) + 1 ≤ ci (pe − 1) + 1 < pe . The same argument as the previous case
leads to a contradiction of lemma 2.3. Thus, ht(P ) ≥ 2 and by Theorem 1.2, we have
dim(R/P ) = sdim(R, f c )
so dim(R/P ) < dim(R) − 1, implying that n ≥ 2 and so D− s(R, f c ) = 0.
It follows immediately that because s(R, f r ) = 0 for all r ≥ c, D+ s(R, f c ) = 0 and so the F-signature is
differentiable at c = F P T (f ) whenever c < 1 is a rational number whose denominator is not divisible by p.
In the next two sections, we will see examples where the result is false when p divides the denominator of c.
3. F-Signature of Homogeneous Polynomials in Two Variables
We turn our attention now to the case when R = k[x, y]hx,yi and let f ∈ R be a product of r ≥ 2 distinct
linear forms with F P T (f ) = c. Here we relax the condition of the previous section that if c is rational, then
p does not divide the denominator. By the exact sequence
R
R
R
0 −−−−→ [pe ] a −−−−→ [pe ] −−−−→ [pe ]
−−−−→ 0
(1)
m
:f
m
m
+ hf a i
4
s
s
s
we have that s(R, f a/p ) = 1 − p12s `R (R/hxp , y p , f a i). This length has been studied extensively by P.
Monsky and P. Teixeira in their work on p-fractals. We can use theorems found in [9] and [7] to obtain the
following result:
3.1. Theorem. The F-signature of the pair s(R, f t ) where f is the product of at least two distinct linear
2
factors converges uniformly on the interval [0, c] to the polynomial r4 t2 − rt + 1 as p → ∞.
e
e
Proof. The Hilbert Syzygy Theorem implies that the module of syzygies between xp , y p and f a can be
generated by two homogeneous elements of degrees m1 ≥ m2 . Their difference δ = m1 − m2 is called the
e
e
e
e
syzygy gap of (xp , y p , f a ). If we need to consider more than one triplet (xp , y p , f a ) we will write this
e
e
syzygy gap as δ(xp , y p , f a ). Theorem 2.10 in [9] tells us that
e
e
`R (R/hxp , y p , f a i) =
1
δ2
(4rape − (ra)2 ) +
4
4
e
e
e
e
Also in his thesis [9], Teixeira showed the functions pae 7→ p12e `R (R/hxp , y p , f a i) and pae 7→ p1e δ(xp , y p , f a )
defined on [0, 1] ∩ Z[p−1 ] can be extended uniquely to continuous functions on [0, ∞). These extended
functions are denoted φf (t) and δf (t) respectively.
e
e
In [7] Monsky proved an upper bound for δ(xp , y p , f a ) in the case when f is homogeneous of degree ≥ 2.
3.2. Theorem ([7] Theorem 11). Let l1 , . . . , lr be linear forms such that li and lj share no common
Pr non-unit
factor for i 6= j and r ≥ 2. Suppose 0 ≤ a1 , . . . , ar ≤ pe and he ai satisfy the inequalities 2ai ≤ 1 aj ≤ 2pe .
e
e
Then δ(xp , y p , Πr1 liai ) ≤ (r − 2)pe−1 .
In our case, this theorem tells us that for f the product of r ≥ 2 distinct linear forms and ra ≤ 2pe then
e
e
δ(xp , y p , f a ) ≤ (r − 2)pe−1 .
Rearranging ra ≤ 2pe , we see that if (ra)/2 ≤ pe , then Monsky’s bound holds. Note that each term in f a
e
has degree in x or degree in y at least (ra)/2, so if f a 6∈ m[p ] then (ra)/2 ≤ pe . Now, we remember the
exact sequence (1) to compute
R
1
a/pe
s(R, f
) = 1 − 2e `R
p
hxpe , y pe , f a i
1
= 1 − 2e (4rape − r2 a2 + δ 2 )
4p
2
a
δ2
r2 a
−
r
+
1
−
=
4 pe
pe
4p2e
e
Extending s(R, f a/p ) to [0, ∞) we get that
r2
s(R, f ) = t2 − rt + 1 −
4
t
e
δf (t)
2
2
e
and Monsky’s upper bound for δ(xp , y p , f a ) shows that as p → ∞, δf (t) → 0 for t < c and so the F-signature
2
converges uniformly to r4 t2 − rt + 1 on the interval [0, c].
3.3. Question. Is there some geometric significance to the quadratic polynomial to which the F-signature
converges with respect to resolution of singularities?
To finish this section, we use the above method to compute the limiting quadratic polynomial of the Fsignature for three distinct lines in the plane, which we may assume is given by f = xy(x + y). Furthermore,
we can compute not only the exact value of c in characteristic 2 mod 3, but also the left derivative of this
function at c in this case. Because we show that the denominator of c is always divisible by p when p ≡ 2
mod 3, the results of the previous section regarding approximate higher-order left derivatives do not apply.
3.4. Example. Let f = xy(x + y) and suppose that k has characteristic p ≥ 5 congruent to 2 mod 3. Define
1
be = p − p+1
pe−1 so that be /pe = 23 − 3p
. A straightforward calculation using Lucas’ theorem shows that
3
5
e
e
for all e ∈ N, we have that f be ∈ m[p ] but f be −1 6∈ m[p ] . Therefore,
e
`R (R/(m[p ] : f be )) = 0
e
`R (R/(m[p ] : f be −1 )) 6= 0
and so we conclude that the F-pure threshold of xy(x + y) is c = be /pe =
proven limiting polynomial of s(R, f t ) is g(t) = 49 t2 − 3t + 1 and g(c) =
g(c) 6= 0, we can directly compute δf (c):
2
δf (c)
= g(c) − s(R, f c )
2
1
= 2
4p
so then δf (c) =
1
p
=
(3−2)pe−1
pe
2
3
1
− 3p
in this case. The above1
c
4p2 . Because s(R, f ) = 0 but
so by Monsky’s bound, we have that δf (c) achieves a local maximum at c.
3.5. Example. We can also provide an affirmative answer to what the left derivative is at F P T (f ) = c in
this case. Note that the denominator is divisible by p, so we cannot apply Theorem 2.2 from the previous
section. We return again to the work of Teixeira, who proves ([10], Theorem II) that if δf achieves a local
maximum at u ∈ [0, 1] then for all t ∈ [0, 1] such that 3|t − u| ≤ δf (u), we have that
δf (t) = δf (u) − 3|t − u|
This implies that δf (t) is piecewise linear near c so we can apply the techniques of calculus to take the left
derivative at c:
2
2
2
δf (t)
δf (t)
3
9 2
t − 3t + 1 −
t−1 −
=
4
2
2
2
δf (t)
3
δf (t)
3
t−1+
t−1−
=
2
2
2
2
now applying the product rule and substituting c = 2p−1
3p , we have:
3 1
2p − 1
3 2p − 1
2p − 1
1
c
− 1 + δf
D− s(R, f ) =
− D− δf
2 2
3p
2
3p
2
3p
3 1
3 2p − 1
1
2p − 1
2p − 1
+
+ D− δf
− 1 − δf
2 2
3p
2
3p
2
3p
3 1
1
1
3 1
1
1
=
− (3)
1−
−1+
+
+ (3)
1−
−1−
2 2
2p
2p
2 2
2p
2p
3
=−
p
To complete this computation, we used that D− δf (c) = 3. We can see this by recalling Theorem II from [10],
which tells us for t sufficiently close to c, we have δf (t) = δf (c) − 3|t − c|. This gives that the left derivative
at c is 3.
Note that this computation of D− s(R, f c ) = − p3 6= 0 in contrast to Theorem 2.2 in the previous section,
where it was shown that when p does not divide the denominator of c the left derivative D− s(R, f c ) = 0.
4. Computational Examples
In this section, we will use the computational algebra package Macaulay2 to explicitly compute the Fsignature of pairs for several polynomials and graph the data obtained using gnuplot. For the first example,
we analyze the cusp C and provide an affirmative answer for the left derivative of the F-signature at F P T (C)
in characteristics 5, 11, and 17. We also explicitly compute the F-signature function for three and four
distinct linear forms in various characteristics using Macaulay2, and graph them with the quadratic limiting
polynomials for the F-signature functions in these cases. All examples rely on routines defined in the next
section.
6
4.1. Example. (The cusp in characteristic 5, 11, and 17) Let C = y 2 − x3 be the cuspidal cubic.
1
It is known that whenever characteristic p ≡ 2 mod 3, the F-pure threshold of C is 65 − 6p
. If 6 divides
p+1 e−1
5
1
e
p+1 we define be = (p− 6 )p
so that be /p = 6 − 6p . In this example, we will use Macaulay2 to compute
e
s(R, C (be −1)/p ) in characteristics p = 5, 11, and 17 for e = 2 and 3 using a function defined in section 5 of
this paper. The code below will compute the value of the function for p = 5 and e = 2 (so that be − 1 = 19);
by changing the base ring, value of e, and be appropriately, we may use this same code for other p and e to
obtain the corresponding values. Here Fsig is a Macaulay2 routine defined explicitly in the next section; it
e
computes s(R, C a/p ) for a single value of a/pe .
:R = ZZ/5[x,y]
:C = y^2-x^3
:Fsig(2, 19, C)
The following data was collected using the above
p e
5 2
5 3
11 2
11 3
17 2
17 3
code, changing parameters as mentioned above:
e
s(R, C (be −1)/p )
1/125
1/625
1/1331
1/14641
1/4913
1/83521
e
So we have that for each of these values, s(R, C (be −1)/p ) = 1/pe+1 . Using this data, we can compute the
derivative of the cusp C in these characteristics. Let p be either 5, 11, or 17 and e be 2 or 3. We compute
the difference quotient
e
e
s(R, C (be −1)/p ) − s(R, C be /p )
(be − 1)/pe − be /pe
e
1
and arrive at −1/p for each value of p and e. For these computations, notice be /pe = 65 − 6p
so s(R, C be /p ) =
1
1
1
0. This shows that the points b2p−1
, b3p−1
, and 56 − 6p
, 0 are colinear points on the convex func2 , p3
3 , p4
h
i
5
1
tion s(R, C t ). By convexity of s(R, f t ) we have that the F-signature is linear on the interval b2p−1
2 , 6 − 6p .
Therefore, we can affirmatively say that the derivative of s(R, C t ) at t = F P T (C) is −1/p.
4.2. Example. (Four distinct lines in characteristic 29) Let k = Z/29Z and consider f = xy(x + y)(x + 2y) ∈
k[x, y].
We will use the Macaulay2 functions defined in the next section to generate a graph of the F-signature of
(R, f t ) for 0 ≤ t ≤ 21 . Here, GenPlot is a function that computes the F-signature of (R, f t ) at values of t of
the form 0 ≤ b/pe ≤ F P T (f ) for some fixed value of e, passed as the first argument.
:R = ZZ/29[x,y]
:f = x*y*(x+y)*(x+2*y)
:GenPlot(2, f, "∼/c29e2")
Once complete, this operation will compute the length
1
R
1 − 4 `R
29
(x292 , y 292 , f a )
for 0 ≤ a ≤ 421 and output these lengths to a file named c29e2 (so titled for “characteristic 29, e=2”)
which is formatted to be graphed by the program gnuplot. We provide these two graphs of the computed
F-signature and the limiting polynomial g(t) here. Even at such a low characteristic, the two are nearly
indistinguishable if plotted simultaneously.
7
1
1
’c29e2’
4*x*x-4*x+1
0.9
0.8
0.8
0.7
0.6
0.6
0.5
0.4
0.4
0.3
0.2
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0
0
0.1
0.2
0.3
0.4
0.5
Figure 1. Let f = xy(x + y)(x + 2y) as in example 4. The left picture here is the plot of
s(R, f t ) for 0 ≤ t ≤ 12 generated using Macaulay2. The right picture is the plot of 4t2 −4t+1
on the same interval.
4.3. Example. (Three distinct lines in characteristic 5) Let f = xy(x + y) ∈ k[x, y] where k = Z/5Z.
Notice that we are in the case of the example at the end of the last section: f = xy(x + y) and characteristic
1
5 ≡ 2 mod 3, so we can compute explicitly that the F-pure threshold is 23 − 15
= 35 . Using code similar to
the above example, we generate a plot for the F-signature of f , the limiting polynomial, and also provide a
plot of 14 δf (t)2 on [0, 35 ].
1
1
’c5e4’
2.25*x*x-3*x+1
0.9
0.8
0.8
0.7
0.6
0.6
0.5
0.4
0.4
0.3
0.2
0.2
0.1
0
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0
0.1
0.2
0.3
0.4
Figure 2. Let f = xy(x + y) as in example 5. The left picture here is the plot of s(R, f t )
for 0 ≤ t ≤ 35 generated using Macaulay. The right picture is the plot of 94 t2 − 3t + 1 on the
same interval.
8
0.5
0.6
0.01
’./sgc5e4’
0.008
0.006
0.004
0.002
0
0
0.1
0.2
0.3
0.4
0.5
0.6
Figure 3. This is the plot of the term 41 δf (t)2 on the interval [0, 53 ] which was obtained by
computing 94 t2 − 3t + 1 − s(R, f t ) with f = xy(x + y) as in example 5.
5. Routines for Computing F-signatures using Macaulay2
The results presented here were significantly influenced by experimental data gathered using the computational algebra package Macaulay2 [5]. This final section provides the source code for functions referenced
in the examples from the previous section. The first function defined here accepts an ideal I in a polynomial
e
ring R and returns the eth Frobenius power I [p ] .
fpow = (I, e) ->
(
L:= first entries gens I;
p := char ring I;
J:= ideal(L#0^(p^e));
for i from 1 to (length L)-1 do
J = J + ideal(L#i^(p^e));
J
)
e
e
1
`R (R/(xp1 , . . . , xpd , f a )), where R is a polynomial ring
ed
p
in variables x1 , . . . , xd and f is some polynomial in this ring.
This second function returns a single length 1 −
9
Fsig = (e, a, f) ->
(
R1:=ring f;
p:= char ring f;
I = fpow(ideal(first entries vars R1), e);
1-(1/p^(dim(R1)*e))*degree(I+ideal(f^a))
)
We can now build on this function to compute the F-signature of specific polynomials and output these
lengths to a file. The first function will compute the values of the F-signature for some homogeneous
polynomial f (specified as the second argument when the function is called) at each value a/pe (e is specified
as the first argument) such that 0 ≤ a/pe ≤ F P T (f ). This is accomplished by repeatedly calling Fsig(e,
a, f). The values computed are then written to a file named f ileN (the third argument passed to the
function) which should be enclosed in quotation marks and give the full path name of the file. The data
is stored in the correct format for use with the program gnuplot to produce images like those found in the
previous section and a new window is opened which contains a plot of the data just computed.
GenPlot = (e, f, fileN) ->
(
cL = for i from 0 to (char (ring f))^e list
q := Fsig(e, i, f) do (stdio<<i<<", "<<q<<endl<<"============="<<endl;
if q==0 then break;)
fp = toString(fileN)<<" ";
for i from 0 to (length cL)-1 do
fp<<toRR(i/(char (ring f))^e)<<" "<<toRR(cL#i)<<endl;
fp<<close;
fp="plotComm"<<"plot ’"<<toString(fileN)<<"’ with lines";
fp<<close;
run "gnuplot -p plotComm";
run "rm plotComm";
)
You can find this code on my website: www.math.unl.edu/∼ecanton2/.
References
[1] Ian M. Aberbach and Florian Enescu. The structure of F-pure rings, 2005.
[2] M. Blickle, K. Schwede, and K. Tucker. F-signature of pairs and the asymptotic behavior of Frobenius splittings, July 2011.
[3] M. Blickle, K. Schwede, and K. Tucker. F-signature of pairs: Continuity, p-fractals and minimal log discrepancies, November
2011.
[4] Richard Fedder. F-purity and rational singularity, 1983.
[5] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in algebraic geometry. Available at
http://www.math.uiuc.edu/Macaulay2/.
[6] Craig Huneke and Graham J. Leuschke. Two theorems about maximal cohen-macaulay modules, 2002. 10.1007/s00208002-0343-3.
[7] Paul Monsky. Mason’s theorem and syzygy gaps, 2006.
[8] Karl Schwede. Centers of F -purity, 2010.
[9] Pedro Teixeira. p-fractals and hilbert-kunz series, 2002.
[10] Pedro Teixeira. Syzygy gap fractals i. some structural results and an upper bound, 2012.
[11] Kevin Tucker. F-signature exists. 10.1007/s00222-012-0389-0.
10
Department of Mathematics, University of Nebraska - Lincoln, Lincoln, NE 68588
E-mail address: [email protected]
11
| 0math.AC
|
Exponentiated Generalized Pareto Distribution: Properties and applications
towards Extreme Value Theory
arXiv:1708.01686v1 [math.ST] 4 Aug 2017
Se Yoon Lee ∗
Joseph H. T. Kim †
Abstract
The Generalized Pareto Distribution (GPD) plays a central role in modelling heavy tail phenomena
in many applications. Applying the GPD to actual datasets however is a non-trivial task. One
common way suggested in the literature to investigate the tail behaviour is to take logarithm to the
original dataset in order to reduce the sample variability. Inspired by this, we propose and study the
Exponentiated Generalized Pareto Distribution (exGPD), which is created via log-transform of the
GPD variable. After introducing the exGPD we derive various distributional quantities, including
the moment generating function, tail risk measures. As an application we also develop a plot as an
alternative to the Hill plot to identify the tail index of heavy tailed datasets, based on the moment
matching for the exGPD. Various numerical analyses with both simulated and actual datasets show
that the proposed plot works well.
Key Words: Extreme Value Theory; Generalized Pareto Distribution (GPD); Exponentiated Generalized Pareto Distribution; Hill plot
1. Introduction
The Generalized Pareto Distribution (GPD) has recently emerged as an influential distribution in modelling heavy tailed datasets in various applications in finance, operational
risk, insurance and environmental studies. In particular, it is widely used to model sample
exceedances beyond some large threshold, a procedure commonly known as the peaks-overthreshold (POT) method in the Extreme Value Theory (EVT) literature; see, e.g., Embrechts
et al. (1997) and Beirlant et al. (2006). The heavy tail phenomenon and the GPD are linked
through the famous Pickands-Balkema-de Hann theorem (Balkema and De Haan (1974)
and Pickands III (1975)) which states that, for an arbitrary distribution of which the sample
maximum tends to a non-degenerate distribution after suitable standardization, the distribution function of its exceedances over a large threshold converges to the GPD. To this extent,
extensive research has been carried out in the literature to characterize the GPD and apply
it to the EVT framework; see, for example, de Zea Bermudez and Kotz (2010) for a comprehensive survey on various estimation procedures for the GPD parameter.
In this paper we introduce a two-parameter exponentiated Generalized Pareto Distribution (exGPD in short) and study its distributional properties, which is the first contribution
of the present paper. The term ‘exponentiated’ is used because this new distribution is obtained via log-transform of the GPD random variable. There are ample examples where a
new distribution is constructed through logarithm or exponential transformation of an existing one. Such examples include normal and log-normal, gamma and log-gamma, and
Pareto and shifted exponential distributions. Thus introducing the exGPD is interesting
on its own right from statistical viewpoint, but we have further motivations of considering
∗
†
Texas A & M University, 400 Bizzell St, College Station, TX 77843
Yonsei University, Seoul, Korea, 120-749
such a distribution in connection to EVT. In various graphical tools offered by EVT one
frequently investigates the tail behaviour with log-transformed data rather than the original
data, as seen in, for example, Hill (1975) plot and the estimator of Pickands III (1975). As
log transformation greatly reduces the variability of extreme quantiles, this practice naturally allows one to investigate the tail behaviour in a more stable manner. Therefore, given
that the GPD is the central distribution in EVT modelling, it makes sense to create an alternative plotting tool using the exGPD directly. This is our second contribution to this paper.
In particular, we develop a new plot as an alternative to the Hill plot to identify the tail index
of heavy tailed datasets. The proposed Log Variance (LV) plot is based on the idea of the
sample variance of log exceedances to be matched to the variance of the exGPD. Through
various numerical illustrations with both simulated and actual datasets, it is shown that the
LV plot works reasonably well compared to the Hill plot, elucidating the usefulness of the
exGPD.
This article is organized as follows. In Section 2 we define the exGPD and investigate
its distributional quantities including the moment generating function, from which the moments can be obtained. We show that the moment of all orders are finite for the exGPD,
unlike the GPD. Section 3 derives popular risk measures, including the Value-at-Risk and
the conditional tail expectation. In Section 4 we develop the LV plot that is derived from
the exGPD variance. We use both simulated and actual datasets to illustrate the proposed
plot and compare it to the Hill plot. Section 5 concludes the article.
2. Exponentiated Generalized Pareto Distribution
2.1
Definition
The distribution function (df) of the two-parameter GPD with parameter (σ, ξ) is defined as
ξx −1/ξ
GX (x) = 1 − 1 +
;
σ
ξ 6= 0
(1)
where the support is x ≥ 0 for ξ > 0 and 0 ≤ x ≤ −σ/ξ for ξ < 0. Here, σ and ξ are
called the scale and shape parameter, respectively. For the case of ξ = 0, the df is defined
as
GX (x) = 1 − e−x/σ ; ξ = 0,
(2)
an exponential distribution defined on x ≥ 0 with scale parameter σ. The density function
is then
1
ξx −1/ξ−1
1
+
,
ξ 6= 0
σ
σ
(3)
gX (x) =
1
e−x/σ ,
ξ = 0.
σ
The GPD is contains three distributions. When ξ > 0 the GPD is an Pareto distribution
of the second kind (or Lomax distribution in the insurance literature) with a heavy tail decaying at a polynomial speed; when ξ = 0 the GPD is an exponential distribution with a
medium tail decaying exponentially; for ξ < 0, the GPD becomes a short-tailed distribution
the upper bound of the distribution support is finite. The kth moment of the GPD is existent
for ξ < 1/k; for instance, the mean and variance are finite only when ξ < 1 and ξ < 0.5,
respectively.
Now we define the exponentiated GPD (exGPD) as the logarithm of the GPD random
variable. That is, when X is GPD distributed, random variable Y = log X is said to be an
exGPD random variable. A simple algebra gives its df as
FY (y) = P (Y ≤ y) = P (log X ≤ y) = P (X ≤ ey )
ξey −1/ξ
= GX (ey ) = 1 − 1 +
; ξ 6= 0,
σ
(4)
of which the support is −∞ < y < ∞ for ξ > 0, and −∞ < y ≤ log(−σ/ξ) for ξ < 0.
When ξ = 0, we have
FY (y) = 1 − e−e
y /σ
= 1 − exp(−ey−log σ );
ξ = 0, −∞ < y < ∞,
(5)
which is the Type III extreme value distribution (or the distribution of −Y is that of Gumbel)
with location parameter log σ. Combining these, we can write the density of the exGPD as
y
ξey −1/ξ−1
e
1
+
, ξ 6= 0
σ
σ
(6)
fY (y) =
1
y
ey−e /σ ,
ξ = 0.
σ
We note that the role of σ changes from a scale parameter under the GPD to a location
parameter under the exGPD, as seen from ey /σ = ey−log σ in (6) for all ξ. In what follows
we denote X and to be the GPD variable, and Y for the exGPD to avoid confusion.
An alternative way to create the exGPD is to define Y through Y = log(W − d)
conditional on W > d, where W is the Pareto random variable with df FW (w) = 1 −
(w/β)−α , w > β. The survival function of Y is then given by
P (W − d > ey |W > d)
P (Y > y) = P (log(W − d) > y|W > d) =
P (W > d)
y −α
d+e
P (W > d + ey )
ey −α
β
=
=
=
1
+
,
d −α
P (W > d)
d
(7)
β
which corresponds to exGPD(d/α, 1/α).
In Figure 1 we compare the densities of the GPD and exGPD side by side for selected
parameter choices for ξ ≥ 0. The most dramatic change is the shape itself, where the GPD is
always decreasing on its support (0, ∞) whereas the exGPD, defined on the entire real line,
has a peak in the middle of the distribution. Obviously, all the realized GPD values between
0 and 1 are mapped to negative numbers under the exGPD. Also, as seen from the densities
(3) and (6), the polynomial right tail of the GPD changes to an exponentially decaying
tail under the exGPD through log transformation. Later we will prove the moments of all
orders are actually finite for the exGPD. From the figure, we have the following additional
comments:
• As σ increases the density of the exGPD shifts to the right because log σ is the location parameter. In fact, the mode can be shown to be log σ, provided it exists; the
proof will be given shortly. As σ is strictly positive, the mode can take any real values
with σ = 1 being the boundary of the sign of the mode.
• As ξ increases the exGPD density becomes less peaked and shows larger dispersion
or variance. This implies that the shape parameter ξ of the GPD roughly plays the
role of a scale parameter under the exGPD. This is also hinted from the density (6);
if the constant 1 is removed from the density function, ξ becomes a scale parameter.
• The visual advantage of the exGPD over the GPD is clear from the figure. The exGPD
expresses the tail thickness, represented by ξ, in a much clearer way in that the tail
decaying angle becomes steeper as ξ gets smaller. Thus one can quickly investigate
how heavy the tail is by looking at, e.g., the histogram of the log data, which is less
straightforward with the GPD densities.
0.4
PDF of exGPD with σ=1
1.0
PDF of GPD with σ=1
ξ=0
ξ=1
ξ=2
0.0
0.2
0.0
0.2
0.1
0.4
f(y)
g(x)
0.6
0.3
0.8
ξ=0
ξ=1
ξ=2
0
2
4
6
8
10
−5
0
x
10
0.4
PDF of exGPD with σ=5
0.20
PDF of GPD with σ=5
ξ=0
ξ=1
ξ=2
0.0
0.05
0.1
0.2
f(y)
0.15
0.3
ξ=0
ξ=1
ξ=2
0.10
g(x)
5
y
0
2
4
6
x
8
10
−5
0
5
10
y
Figure 1: Density comparison: GPD vs. exGPD for ξ ≥ 0
In Figure 2, the densities of the GPD and exGPD are compared for ξ < 0. Again, the
shape of the density changes substantially by log-transform. Note that there is an upper
limit in the support of both distributions in this case. However, when ξ is a small negative
value, the shape looks not that different from ξ ≥ 0 case in that its right tail gradually
decays as if there is no upper limit to our eyes. As |ξ| gets larger, but not greater than 1,
there is no smooth landing around the upper limit of the support; the right tail abruptly drops
to zero without hesitation. When ξ = −1, the exGPD density takes a drastically different
shape, by shooting upside at its upper limit. Thus, when the shape parameter is negative,
its magnitude can completely change the shape of the exGPD density, as is the case for the
GPD distribution. We note that the role of σ is two-fold when ξ < 0; it acts as the location
parameter and also controls the upper limit of the distribution support.
1.0
PDF of exGPD with σ=1
1.0
PDF of GPD with σ=1
0.6
0.8
ξ = − 0.1
ξ = − 0.5
ξ = −1
0.4
0.2
0.0
0.0
0.2
0.4
f(y)
g(x)
0.6
0.8
ξ = − 0.1
ξ = − 0.5
ξ = −1
0
2
4
6
8
10
−6
−4
−2
x
0
2
4
y
1.0
PDF of exGPD with σ=5
0.20
PDF of GPD with σ=5
ξ = − 0.1
ξ = − 0.5
ξ = −1
f(y)
0.0
0.00
0.2
0.05
0.4
g(x)
0.10
0.6
0.15
0.8
ξ = − 0.1
ξ = − 0.5
ξ = −1
0
10
20
30
40
50
−6
−4
x
−2
0
2
4
y
Figure 2: Density comparison: GPD vs. exGPD for ξ < 0
The following result formally shows how the shape parameter of the exGPD relates to
its density shape, which is also linked to the existence of the mode.
Lemma 1. The density of the exGPD(σ, ξ) in (6) is bounded only for ξ ≥ −1, in which
case there is a unique mode at log σ.
Proof: To find the mode, a simple algebra yields that
ey
ξey −1/ξ−1 ey
1
ξey −1/ξ−2 ξey
1+
+
− −1 1+
·
σ
σ
σ
ξ
σ
σ
−1/ξ−1
−1
ey
ξey
1
ξey
ξey
=
1+
1+ − −1 1+
σ
σ
ξ
σ
σ
1 − ey /σ
= fY (y)
.
1 + ξey /σ
fY0 (y) =
(8)
Note that in the last expression, both fY (y) and 1 + ξey /σ are non-negative regardless of
the sign of ξ. We now investigate whether fY0 (y) = 0 gives a sensible solution for different
ranges of ξ.
(1) ξ < −1 case: Using the distribution support −∞ < y ≤ log(−σ/ξ) for negative ξ, we
see that the numerator of the second term in (8) is bounded by
1 + 1/ξ ≤ 1 − ey /σ < 1.
(9)
When ξ < −1, the lower bound of this inequality becomes a strictly positive number,
so fY0 (y) = 0 has no solution as seen in (8). In fact, the density value in this case gets
indefinitely large as y approaches its upper limit since
lim
y→log(−σ/ξ)−
ey
ξey −1/ξ−1
1+
σ
y→log(−σ/ξ)− σ
−1
ξey −1/ξ−1
=
1+
= +∞.
lim
ξ y→log(−σ/ξ)−
σ
fY (y) =
lim
(10)
The last equality holds because −1/ξ −1 < 0 when ξ < −1. Thus the density is unbounded
in this range.
(2) ξ = −1 case: The density reduces to fY (y) = ey /σ, the exponential function shifted
by log σ. This function, defined on −∞ < y ≤ log σ, is increasing in y and attains its
maximum of 1, at its upper limit of the support, y = log σ, which is the mode.
(3) −1 < ξ < 0 case: The lower bound of (9) is negative in this range, so fY0 (y) = 0 has
a unique solution at log σ from (8). Note that the mode in this case is strictly inside the
distribution support because −∞ < log σ < log(−σ/ξ), in contrast to the ξ = −1 case
where the maximum occurs at the upper boundary of the support.
(4) ξ ≥ 0 case: In this case the distribution support is −∞ < y < ∞, which yields
− ∞ < 1 − ey /σ < 1.
(11)
Thus fY0 (y) = 0 has a unique solution y = log σ, which is the mode.
Now we turn to the hazard function of the exGPD which is given by
fY (y)
ey
ξ y −1
ey /σ
1
hY (y) =
=
1+ e
=
= −y
,
y
1 − FY (y)
σ
σ
1 + ξe /σ
σe + ξ
ξ 6= 0.
(12)
Figure 3 compares the hazard function of the GPD and exGPD. The hazard function of
the GPD can increase or decrease depending on the sign of ξ, with heavy tail implied for
ξ > 0 as the hazard function decreases in y. However the exGPD always has an increasing
hazard function, a.k.a. increasing fairlure rate (IFR), regardless of the sign of ξ, as seen
from (12) and the figure. According to the standard theory, therefore, we see that exGPD is
also DMRL (decreasing mean residual lifetime), indicating a light-tailed distribution.
10
Hazard Function of exGPD with σ=1
10
Hazard Function of GPD with σ=1
2
4
h(y)
6
8
ξ = − 0.1
ξ = − 0.5
ξ = −1
0
0
2
4
h(x)
6
8
ξ = − 0.1
ξ = − 0.5
ξ = −1
0
2
4
6
x
8
10
12
−6
−4
−2
0
y
2
4
6
2.0
Hazard Function of GPD with σ=1
Hazard Function of exGPD with σ=1
ξ=0
ξ=1
ξ=2
h(y)
0
0.0
1
0.5
2
1.0
h(x)
3
1.5
4
5
ξ=0
ξ=1
ξ=2
0
1
2
3
4
5
−5
0
x
5
10
y
Figure 3: Hazard function comparison between GPD and exGPD with different parameters
2.2
Moment Generating Function
In what follows, we denote X to be a GPD(σ, ξ) random variable and Y = log X to be
the corresponding exGPD(σ, ξ) variable. In order to derive the moment generating function
(mgf) of Y , we note the following relationship:
s
MY (s) = E[esY ] = E[es log X ] = E[elog X ] = E[X s ],
s ∈ IR.
(13)
from which we have the next result.
Lemma 2. For the exGPD(σ, ξ) distribution defined in (4), the moment generating function
is given by
σ s
1
−
B(s + 1, −1/ξ); s ∈ (−1, ∞), ξ < 0
−
ξ
ξ
s
(14)
MY (s) = 1 σ B(s + 1, 1/ξ − s); s ∈ (−1, 1/ξ), ξ > 0
ξ
ξ
s
σ Γ(1 + s); s ∈ (−1, ∞), ξ = 0,
where B is the beta function defined as
Z 1
Γ(x)Γ(y)
B(x, y) =
tx−1 (1 − t)y−1 dt =
,
Γ(x + y)
0
x > 0, y > 0.
Proof: We prove this for three cases depending on the sign of ξ.
(1) ξ < 0 case: From the support of X ∼ GPD(σ, ξ), we have
−1/ξ−1
Z −σ/ξ
ξ
s
s 1
1+ x
dx.
MY (s) = E[X ] =
x ·
σ
σ
0
If we let
(15)
(16)
ξ
x = −t, we have
σ
σ s 1
σ
MY (s) =
− t · (1 − t)−1/ξ−1 · − dt
ξ
σ
ξ
0
s Z 1
1
σ
=−
−
ts (1 − t)−1/ξ−1 dt
ξ
ξ
0
1
σ s
=−
−
B(s + 1, −1/ξ); −1 < s.
ξ
ξ
Z
1
(17)
(2) ξ > 0 case: As X is defined for all positive values,
MY (s) = E[X s ] =
Z
∞
xs ·
0
By letting
ξ
x = t, we have
σ
Z
MY (s) =
−1/ξ−1
1
ξ
1+ x
dx.
σ
σ
σ
σ s 1
t · (1 + t)−1/ξ−1 · dt
ξ
σ
ξ
0
Z
1 σ s ∞ s
=
t (1 + t)−1/ξ−1 dt
ξ ξ
0
1 σ s
=
B(s + 1, 1/ξ − s); −1 < s < 1/ξ,
ξ ξ
∞
where the last equality comes from a property of the beta function
Z 1
Z ∞
x−1
y−1
B(x, y) =
t (1 − t) dt =
tx−1 (1 + t)−x−y dt, x > 0, y > 0.
0
(18)
(19)
(20)
0
It is pointed out that if we compare the two mgf’s (ξ < 0 vs. ξ > 0) in (14), only the second
argument of the beta function is different.
(3) ξ = 0 case: The derivation is straightforward and omitted; alternatively, this can be
adapted and derived from the mgf of Gumbel distribution as shown in, e.g., Ch. 22 of Johnston et al. (1994).
2.3
Moments
From the mgf of the exGPD given in Lemma 2, we may determine the moments by differentiating it with respect to s. We present the first two moments for ξ 6= 0 first. Let us
rewrite mgf in (14) as a function of gamma functions for easier differentiation.
1
σ s
Γ(s + 1)
−
Γ(−1/ξ)
·
−
·
; s ∈ (−1, ∞), ξ < 0.
ξ
Γ(s − 1/ξ + 1)
ξ
MY (s) =
s
1
σ
Γ(1/ξ + 1) ·
· Γ(s + 1) · Γ(1/ξ − s); s ∈ (−1, 1/ξ), ξ > 0.
ξ
ξ
(21)
For ξ < 0 case, we may obtain the first two derivatives of the mgf as follows.
d
1
σ s
1
MY (s) = − Γ(−1/ξ) · −
· Γ(s + 1) ·
ds
ξ
ξ
Γ(s − 1/ξ + 1)
σ
· log −
+ ψ(s + 1) − ψ(s − 1/ξ + 1) .
ξ
d2
1
σ s
1
MY (s) = − Γ(−1/ξ) · −
· Γ(s + 1) ·
2
ds
ξ
ξ
Γ(s − 1/ξ + 1)
2
σ
0
0
·
log −
+ ψ(s + 1) − ψ(s − 1/ξ + 1) + 0 + ψ (s + 1) − ψ (s − 1/ξ + 1) ,
ξ
where ψ is the digamma function. In this derivation, we used Γ0 (s+1) = Γ(s+1)·ψ(s+1)
and
0
1
ψ(s − 1/ξ + 1)
=−
.
Γ(s − 1/ξ + 1)
Γ(s − 1/ξ + 1)
Similarly, when ξ > 0, we use the fact Γ0 (1/ξ − s) = −Γ(1/ξ − s)ψ(1/ξ − s) to get
s
σ
d
1
σ
· Γ(s + 1) · Γ(1/ξ − s) log
+ ψ(s + 1) − ψ(1/ξ − s) .
MY (s) = Γ(1/ξ + 1) ·
ds
ξ
ξ
ξ
s
2
d
1
σ
· Γ(s + 1) · Γ(1/ξ − s)
MY (s) = Γ(1/ξ + 1) ·
2
ds
ξ
ξ
2
σ
0
0
·
log
+ ψ(s + 1) − ψ(1/ξ − s) + 0 + ψ (s + 1) + ψ (1/ξ − s) .
ξ
Hence, by setting s = 0, we have the first moment of the exGPD:
σ
+ ψ(1) − ψ(1 − 1/ξ); ξ < 0.
log −
ξ
d
E[Y ] = MY (s)
= log σ + ψ(1) − ψ(1/ξ); ξ > 0.
ds
s=0
ξ
log σ + ψ(1); ξ = 0.
(22)
Here ψ(1) = −γ, where γ = 0.5772 · · · is the Euler-Mascheroni constant. The case for for
ξ = 0 has been determined separately, but it is easily done. Likewise, the second moment
is given by
2
σ
+ ψ(1) − ψ(1 − 1/ξ) + ψ 0 (1) − ψ 0 (−1/ξ + 1);
log −
ξ
2
2
d
2
σ
=
E[Y ] = 2 MY (s)
log
+
ψ(1)
−
ψ(1/ξ)
+ ψ 0 (1) + ψ 0 (1/ξ); ξ > 0
ds
s=0
ξ
log σ + ψ(1)2 + ψ 0 (1); ξ = 0.
E[Y ]2 + ψ 0 (1) − ψ 0 (−1/ξ + 1); ξ < 0
= E[Y ]2 + ψ 0 (1) + ψ 0 (1/ξ); ξ > 0
E[Y ]2 + ψ 0 (1); ξ = 0.
(23)
ξ<0
From the second raw moment, the variance of the exGPD becomes
2
∞
X
1
π
0 (1) − ψ 0 (−1/ξ + 1)
ψ
; ξ<0
−
6
(k − 1/ξ)2
k=1
2
∞
X
1
V ar[Y ] = ψ 0 (1) + ψ 0 (1/ξ)
= π +
; ξ>0
6
(k + 1/ξ − 1)2
k=1
0
π2
ψ (1)
; ξ = 0.
6
(24)
Note that the variance of exGPD depends only on ξ. Furthermore, since the summation
terms in (24) are always positive, the value of the sample variance of Y , denoted s2Y , can
serve as a quick indicator about the sign of ξ. That is, if s2Y > π 2 /6 ' 1.645, then ξ
is deemed positive; otherwise, ξ is negative. The first two moments yield the method of
moments estimator (MME) of the exGPD parameter. For ξ < 0,
1
,
1 − (ψ 0 )−1 (ψ 0 (1) − s2 )
= −ξˆM M E · eȲ −ψ(1)+ψ(1−1/ξ̂M M E )
ξˆM M E =
σ̂M M E
and for ξ > 0,
ξˆM M E =
1
(ψ 0 )−1 (s2
− ψ 0 (1))
,
σ̂M M E = ξˆM M E · eȲ −ψ(1)+ψ(1/ξ̂M M E ) .
When ξ = 0, we have σ̂M M E = eȲ −ψ(1) .
One may further obtain higher moments by differentiating the mgf repeatedly, but the
task is not straightforward as higher order derivatives can be complicated, involving infinite series, as seen from the derivation of the first two moments above. However, formally
establishing the existence of higher moments of the exGPD is important because the difficulty of handling the GPD, such as sampling variability, is essentially attributed to its tail
heaviness, directly connected to the non-existence of its moments.
Corollary 3. The exGPD defined in (6) has finite moments of all orders.
Proof: Referring to (21), the mgf of the exGPD is a product of three functions in terms
of s, for both ξ > 0 and ξ < 0 cases. The first term is simply an exponential function of
s, which has derivatives of all orders. The second and third terms are gamma functions or
its reciprocal. The derivatives of a gamma function can be written through the polygamma
function. The polygamma function of order k is defined as the (k + 1)-th derivative of the
logarithm of the gamma function
∞
ψ (k) (z) =
X
dk
dk+1
1
ψ(z) = k+1 log Γ(z) = (−1)k+1 k!
,
k
dz
dz
(z + r)k+1
k ≥ 1,
(25)
r=0
which is finite for z > 0. Hence we conclude that the mgf of the exGPD has derivatives of
all orders, each of which in turn gives a finite value when evaluated at 0. A similar argument
holds for ξ = 0 case.
2.4
Further properties
In the GPD literature, various distributional quantities are available in addition to the ordinary moments and MEF; see,e.g., Ch 3 of Embrechts et al. (1997). Here we list similar
properties of the exGPD; some are parallel to those of the GPD, but others are different. We
present the findings first, and then provide the proofs. All findings include both ξ > 0 and
ξ < 0 cases.
(a) For a real value r with rξ > −1,
ξ Y −r
1
E 1+ e
=
;
σ
1 + rξ
rξ > −1
(26)
(b) For a non-negative integer k,
E
k
ξ Y
log 1 + e
= ξ k · k!
σ
(c) For real value r with 1 + r > |ξ|,
r
Y
E e · F Y (Y )
=
σ
(r + 1 − ξ)(r + 1)
(27)
(28)
(d) Assume that N ∼ P oi(λ), independent of the iid sequence (Yi )ni=1 where Yi ∼
exGP D(σ, ξ), i = 1, 2, · · · n. Write MN = max(Y1 , Y2 , · · · , Yn ). Then
P (MN ≤ y) = exp
ξey −1/ξ
−λ 1+
σ
(29)
(e) For a constant c,
Z
∞
Z
F Y (y)dy =
c
c
∞
ξ
1 + ey
σ
−1/ξ
dy =
c /σ)−1 ; 1/ξ, 0 ,
B
(1
+
ξe
ξ > 0,
B 1 + ξec /σ; 1 − 1/ξ, 0 ,
ξ < 0,
(30)
where B(x; a, b) is the incomplete beta function
Z x
B(x; a, b) =
ta−1 (1 − t)b−1 dt.
(31)
0
Proof: Proof for (a):
Z ∞
ξ y −1/ξ−1
ξ Y −r
ξ y −r ey
E 1+ e
=
1+ e
·
· 1+ e
dy
σ
σ
σ
σ
−∞
Z ∞ y
e
ξ y −1/ξ−r−1
· 1+ e
dy.
=
σ
−∞ σ
(32)
For ξ > 0 case, we take 1 + ξey /σ = t, then by integration by substitution, (32) is equal to
Z ∞
1
1
1
1
t−1/ξ−r−1 · dt = ·
=
; r > −1/ξ.
(33)
ξ
ξ 1/ξ + r
1 + rξ
1
When ξ < 0, (32) becomes, with the same substitution,
Z 0
1
1
1
1
t−1/ξ−r−1 · dt = ·
=
;
ξ
ξ
1/ξ
+
r
1
+
rξ
1
r < −1/ξ,
(34)
completing the proof. Note that the different conditions for ξ has been combined to rξ >
−1. We comment thus that the condition used in Theorem 4.3.13 of Embrechts et al. (1997)
is not correct.
Proof for (b):
k Z ∞
k y
ξ Y
ξ y
e
ξ y −1/ξ−1
E log 1 + e
log 1 + e
=
·
· 1+ e
dy.
σ
σ
σ
σ
−∞
(35)
For ξ > 0 case, if we take log(1 + ξey /σ) = t, then by integration by substitution, (35) is
equal to
Z
∞
ξ −1
tk · e−t/ξ dt = ξ k · k!.
(36)
0
For ξ < 0 case, we obtain the same result with the same substitution, but via a slightly
different integration as ξ < 0 and t < 0. Note that depending on k being odd or even, the
quantity can be negative or positive for ξ < 0.
Proof for (c):
Z
r
Y
E e · F Y (Y )
=
ξ y −1/ξ−1
ξ y −1/ξ r ey
·
· 1+ e
dy
e ·
1+ e
σ
σ
σ
−∞
Z ∞
ξ y −(r+1)/ξ−1
y 1
=
e ·
1+ e
dy.
(37)
σ
σ
−∞
∞
y
For ξ > 0 case, if we take 1 + ξey /σ = t, then (37) is equal to
Z ∞
Z
σ
σ ∞ −(r+1)/ξ
−(r+1)/ξ−1 1
· (t − 1)t
· dt = 2
(t
− t−(r+1)/ξ−1 )dt
ξ
ξ
ξ
1
1
σ
ξ
ξ
−
= 2
ξ r+1−ξ r+1
σ
=
; 1 + r > ξ.
(r + 1 − ξ)(r + 1)
For ξ < 0 case, with the same substitution, (37) becomes
Z 0
σ
1
σ
· (t − 1)t−(r+1)/ξ−1 · dt =
;
ξ
ξ
(r
+
1
−
ξ)(r + 1)
1
r + 1 > 0.
(38)
(39)
Thus the conditions differ depending on the sign of ξ even though the results are identical.
If one wishes to merge these for convenience sake, then 1+r > |ξ| would serve the purpose.
Proof for (d): Note that
P (MN ≤ y) = E[P (MN ≤ y|N = n)] =
∞
X
n=0
P (MN ≤ y|N = n) · P (N = n).
On the other hand, we have
P (MN ≤ y|N = n) = P (Mn ≤ y) = P (Y1 ≤ y, Y2 ≤ y, · · · , Yn ≤ y)
= P (Y1 ≤ y)P (Y2 ≤ y) · · · P (Yn ≤ y) = (FY (y))n .
Therefore, we have
P (MN ≤ y) =
∞
X
∞
X (λ · FY (y))n
λn e−λ
(FY (y)) ·
= e−λ
n!
n!
n
n=0
−λ
n=0
·e
=e
= e−λ·F Y (y)
ξey −1/ξ
.
= exp − λ 1 +
σ
=e
λ·FY (y)
−λ(1−FY (y))
(40)
This property is related to the POT framework. If transformed back, this is the generalized
extreme value (GEV) distribution.
Proof for (e): For ξ > 0 case, use integration by substitution t = (1 + ξey /σ)−1 to get
Z
∞
Z
∞
ξey
1+
σ
F Y (y)dy =
−1/ξ
Z
dy =
c
= B (1 + ξex /σ)−1 ; 1/ξ, 0 ,
c
(1+ξec /σ)−1
t1/ξ−1 · (1 − t)−1 dt
0
(41)
which can be evaluated using the hypergeometric function. Similarly, for ξ < 0 case, we
set t = 1 + ξey /σ to get
Z
log(−σ/ξ)
log(−σ/ξ)
Z
ξey
1+
σ
F Y (y)dy =
c
c
1+ξec /σ
Z
=
−1/ξ
Z
0
dy =
t−1/ξ · (t − 1)−1 dt
1+ξec /σ
t−1/ξ · (1 − t)−1 dt = B 1 + ξec /σ; 1 − 1/ξ, 0 .
0
(42)
For ξ = 0 case, we may use integration by substitution t = ey−log σ to get
Z ∞
Z ∞
Z ∞
−ey−log σ
F Y (y)dy =
e
t−1 e−t dt = Γ(0, ec−log σ ),
dy =
c
(43)
ec−log σ
c
where Γ(s, x) is the incomplete gamma function
Z ∞
Γ(s, x) =
ts−1 e−t dt.
(44)
x
Sometimes the three-parameter GPD is found in the literature, which is created by adding
a location parameter µ to the GPD. The df of this distribution is defined as G(x; µ, σ, ξ) =
G(x − µ; σ, ξ), x > µ, where G(x; σ, ξ) is the GPD in (1). We comment that most results in
this section can be readily applied to the three-parameter GPD without additional difficulty.
2.5
Maximum likelihood estimation
Let y1 , · · · , yn be an iid sample from exGPD(σ, ξ). Then from its density (6) the loglikelihood of the exGPD can be written as
n
n
X
X
n log σ 1
−
+1
log(σ + ξeyi ).
(45)
l(σ, ξ) =
yi +
ξ
ξ
i=1
i=1
ˆ may be found from solving
and the MLE (σ̂, ξ)
n
1
X
∂l
n
1
=
−
+1
=0
∂σ
σξ
ξ
σ + ξeyi
i=1
n
n
X
∂l
n log σ
eyi
1 X
=−
−
+
1
+
log(σ + ξeyi ) = 0.
∂ξ
ξ2
ξ
σ + ξeyi
ξ2
1
i=1
i=1
Clearly, the MLE has to be solved numerically. Due to the condition that σ + ξeyi > 0 for
all i, one must have y(n) < log(−σ/ξ) for ξ < 0 where y(n) is the sample maximum. Then,
for any given ξ < −1, we see that
lim
+
log(−σ/ξ)→y(n)
l(σ, ξ) = ∞,
(46)
implying that there is no finite MLE. So the MLE exists only when ξ > −1, the same
condition for the MLE of the GPD to exist as pointed out in, e.g., Grimshaw (1993). In fact,
it turns out that the MLEs for the GPD and exGPD are identical for a given sample. To see
this, consider the densities of these two distributions with the same parameter (σ, ξ), that is,
gX (x) and fY (y) from (3) and (6), respectively. When a sample x1 , · · · , xn is given from
the GPD, we can obtain a corresponding exGPD sample y1 , · · · , yn with yi = log xi . The
log-likelihood function for the exGPD (ξ 6= 0) is then shown from the density to be
n
X
log fY (yi |σ, ξ) =
i=1
n
X
i=1
=
n
X
i=1
yi +
n
X
log gX (eyi |σ, ξ)
i=1
log xi +
n
X
log gX (xi |σ, ξ)
(47)
i=1
In the last expression, the first term involves no parameter and the second term stands for
the log-likelihood function of the GDP. Hence maximizing the likelihood functions for both
distributions leads to the identical parameter estimate. This implies that, even though the
log transform may stablize the volatility of the GPD sample in terms of, e.g., moments,
the MLE
does not offer additional benefits from
this stabilization. Consequently, if we let
σ̂
σ
for exGPD(σ, ξ), we have
θ̂ = ˆ be the MLE of the parameter θ =
ξ
ξ
√
d
n(θ̂ − θ) ∼ BV N (0, [I1 (θ)]−1 ),
where the covariance matrix is given by
2
2σ (1 + ξ) −σ(1 + ξ)
[I1 (θ)]−1 =
,
−σ(1 + ξ) (1 + ξ)2
ξ > −0.5.
(48)
(49)
The condition ξ > −0.5 is needed for the information matrix to be defined properly; see,
e.g., Smith (1984) or Embrechts et al. (1997).
3. Risk measures and mean excess function
Modern financial risk management in the EVT framework often requires determination
of some well-known tail risk measures to summarize the riskiness of the underlying loss
distribution. Two popular such measures are the Value-at-Risk (VaR) and the conditional
tail expectation (CTE). The VaR of a continuous random variable Y at the 100p% level,
with df FY , is the 100p quantile of the distribution of Y , denoted by
yp = FY−1 (p).
(50)
The VaR is a widely-accepted standard risk measure used in solvency and risk analyses in
financial and insurance industry, with p close to 1. From (4), the VaR of the exGPD is easily
shown to be
σ
−ξ
(51)
yp = log
(1 − p) − 1 , 0 < p < 1,
ξ
or, using that quantiles are preserved under a monotone transform, we can also write yp =
log xp , where xp is the 100p quantile of the GPD. Despite of its popularity, the quantile
VaR measure has also been criticized for failing to meet one of the properties that any
desirable risk measure is expected to satisfy, the criteria known as the coherent risk measure
axioms; see Artzner et al. (1999). In this regard, the CTE1 has received much attention as
an alternative, coherent tail risk measure. The idea of the CTE is to measure the average
severity of the loss when the extreme loss does occurs, where the extreme loss is represented
by the VaR. The CTE of the random variable Y at 100p% level is defined as
CT Ep (Y ) = E(Y |Y > yp ).
(52)
Though the CTE is relatively new, it is closely related to the excess variable Y − u|Y > u
which has long been used in statistics and reliability studies to examine the tail behaviour
of a given distribution. Its expectation e(u) = E(Y − u|Y > u), known as the mean excess
function (MEF) or the mean residual lifetime, is connected to the CTE through
CT Ep (Y ) = E(Y |Y > yp ) = yp + E(Y − yp |Y > yp ) = yp + e(yp ).
(53)
Thus the CTE is immediate if we can obtain the MEF.
For the exGPD we first derive the df of the excess variable, denoted Y 0 = Y − u|Y > u
for a constant u. Referring to the df of the exGPD (4), we have
FY 0 (y 0 ) = P (Y 0 < y 0 ) = P (Y − u ≤ y 0 |Y > u) =
P (u < Y < u + y 0 )
F Y (u) − F Y (u + y 0 )
=
P (Y > u)
F Y (u)
0
ξeu+y −1/ξ
0
σ + ξeu+y −1/ξ
σ
=1−
,
=1−
σ + ξeu
ξeu −1/ξ
1+
σ
1+
y 0 ≥ 0, ξ 6= 0. (54)
Unfortunately this is not an exGPD df. So, unlike the GPD case, the exGPD does not have
the stability property. Nonetheless the MEF can be obtained as the mean of the df (54). We
1
a.k.a. Tail VaR, Conditional VaR and Expected Shortfall in the literature.
note that the expectation of any nonnegative continuous random variable X can be written
as
Z ∞
Z ∞
Z ∞
∞
1 − FX (x)dx =
xfX (x)dx = −x(1 − FX (x)) +
E[X] =
F X (x)dx.
0
0
0
0
(55)
The last equality holds as long as limx→∞ xF X (x) = 0. For the excess variable Y 0 for the
exGPD, which is nonnegative, this condition is met for both ξ < 0 and ξ > 0; the former
case has a finite upper limit, and for the latter case the tail F Y 0 (y 0 ) decays exponentially as
seen from (54). Therefore the MEF of the exGPD becomes
Z ∞
0
σ + ξeu ey −1/ξ 0
0
eY (u) = E[Y ] = E[Y − u|Y > u] =
dy
σ + ξeu
0
Z ∞
0
σ + ξeu ey −1/ξ 0
σ + ξeu 1/ξ
·
dy
=
σ
σ
0
Z ∞
0
σ + ξeu 1/ξ
ξey −1/ξ 0
=
·
1 + −u
dy
(56)
σ
σe
0
To evaluate the integration term in the last expression, let us consider an exGPD variable
Y ∗ with parameter (σ ∗ , ξ) where σ ∗ = σe−u > 0. Then We recognize the the integrand
in (59) as the survival function of another exGPD variable Y ∗ , so that using Property (e) of
the previous section,
Z
σ + ξeu 1/ξ ∞
eY (u) =
F Y ∗ (y ∗ )dy ∗
σ
0
ξeu 1/ξ
∗ )−1 ; 1/ξ, 0 , ξ > 0,
1+
B
(1
+
ξ/σ
σ
=
ξeu 1/ξ
1+
B 1 + ξ/σ ∗ ; 1 − 1/ξ, 0 , ξ < 0,
σ
−1 B (1 + ξeu /σ)−1 ; 1/ξ, 0 , ξ > 0,
(
F̄
(u))
Y
=
(57)
(F̄ (u))−1 B 1 + ξeu /σ; 1 − 1/ξ, 0 , ξ < 0,
Y
For ξ = 0 case, the distribution of Y 0 = Y − u|Y > u is similarly obtained as
FY 0 (y 0 ) = 1 −
F Y (u + y 0 )
u+y 0 −log σ +eu−log σ
u−log σ (ey 0 −1)
= 1 − e−e
= 1 − e−e
,
F Y (u)
y 0 ≥ 0,
(58)
and its MEF is given by
Z ∞
Z ∞
u+y 0 −log σ +eu−log σ
u+y 0 −log σ
e(u) = E[Y 0 ] =
e−e
dy 0 = exp(eu−log σ )
e−e
dy 0
0
0
Z ∞
0
u−log σ
−ey −log σ
0
u−log σ
u
= exp(e
)
e
dy = exp(e
) Γ(0, e /σ),
(59)
u
where the last equality comes from (43).
4. An application: Finding tail index
In this section we illustrate how the proposed exGPD distribution can be used in the EVT
framework. In particular, we focus on finding the tail thickness, characterized as the tail
index α, for a given dataset. Here α corresponds to 1/ξ in the GPD. Determining the
tail index is a problem of great importance as it characterizes the degree of tail risk by
assigning the given distribution to a proper maximum domain of attraction, as well as its
moment existence range. To formally motivate, let us denote the tail (or survival) function
of a distribution by F̄ (x) = 1 − F (x), −∞ < x < ∞. Then we say that F̄ (x) is regularly
varying with index −α < 0, and write F̄ ∈ R−α , if
lim
x→∞
F̄ (xλ)
= λ−α ,
F̄ (x)
λ > 0.
(60)
When α = 0 the tail is called slowly varying, or F̄ ∈ R0 . Using this we can represent
a regularly varying distribution as F̄ (x) ∼ L(x) x−α where L(·) ∈ R0 . We note that
f (x) ∼ g(x) means limx→∞ f (x)/g(x) = 1. Thus the tail of regularly varying functions
can be represented by power functions multiplied by slowly varying functions. In many applications it is of interest to accurately estimate the tail index α from the given dataset; we
refer the reader to, e.g., Embrechts et al. (1997) and Beirlant et al. (2006) for further details.
The method of Hill (1975) is one of the most studied methods to estimate the tail index
α. When F (x) is a regularly varying distribution with α > 0, the Hill method is designed
to find α using a sample X1 , ..., Xn from F (x). If we let the order statistics be Xn,n ≤
... ≤ X1,n , the Hill estimator is defined as
α̂k,n =
X
−1
k
1
log Xj,n − log Xk,n
,
k
2 ≤ k ≤ n.
(61)
j=1
−1
After computing α̂k,n for different k values, one can draw the Hill plot {(k, α̂k,n
)}, k =
−1
2, 3, ..., n on the plane. The goal is then to find an area where α̂k,n is stable in the plot over
different but relatively small k values. Note that log Xk,n serves as the threshold for the
POT method. Other alternative tail index estimators also available in the literature, but the
performance for these alternative estimators crucially depend on the interplay between the
tail index and the so-called second order parameter, and the Hill estimator seems to be most
popular due to its simplicity. It is known that (e.g., Drees et al. (2000)) the Hill plot works
most effectively when the underlying loss distribution F (x) is Pareto with
F (x) = 1 −
x −α
σ
,
x > σ > 0.
(62)
In fact, the Hill estimator implicitly uses the fact that the Pareto variable transforms to a
shifted exponential via taking logarithm, and can be seen as an approximate MLE for the
exponential distribution with rate α, which is the sample mean. Indeed the Hill estimator
has minimum mean squared error when the second order parameter governing the rate of
convergence of (60) is zero which corresponds to the Pareto case. However, as F (x) departs from Pareto, or equivalently, as the log-transformed data departs from exponential,
its performance gets poorer and the plot becomes harder to decipher. Considering that the
log-transformed variable of GPD is that of exGPD, not of exponential, the detrimental performance of the Hill plot in identifying the GPD tail is not surprising.
To this extent, we propose to use the exGPD directly, which is the correct distribution of
the GPD after log-transform. Among other properties of the exGPD, we employ the second
moment result given in (24). The rationale is as follows. When X is a random variable
with a heavy tailed distribution, and its GPD realm starts from some large threshold u > 0,
we assert that X − u|X > u ∼ GP D(σ, ξ) in the spirit of the POT framework. Furthermore, according to the stability property of the GPD, it is also true that X − u0 |X > u0
is GP D(σ + ξ(u0 − u), ξ) distributed, for any higher threshold u0 > u. Hence, from
the definition of the exGPD, we can say that log(X − u|X > u) ∼ exGP D(σ, ξ) and
log(X − u0 |X > u0 ) ∼ exGP D(σ + ξ(u0 − u), ξ) for u0 > u. Since our target parameter ξ is invariant under the threshold change within the GPD realm, and the variance of
log(X − u0 |X > u0 ) involves ξ only for any u0 > u as seen in (24), we are able to estimate
ξ by matching the sample variance of log-exceedances to the theoretical one over different
thresholds, and produce a series of different ξˆ and plot them. This is always possible because the has a finite variance. A stable area is then identified in the plot, just like in the
Hill plot. In particular, we adopt the following algorithm for a given heavy tailed dataset
ordered as Xn,n ≤ ... ≤ X1,n to draw a plot to estimate ξ (or α = 1/ξ):
1. Let ui = Xi,n be the ith threshold.
2. For all Xj,n greater than ui , obtain the exceedances Xj,n − ui and get log(Xj,n − ui ),
j = 1, . . . , i − 1. Equate the sample variance of these i − 1 values to the theoretical
variance in (24) to get an estimate of ξ. Denote this by ξ˜i .
3. Repeat above two steps for i = n, . . . , 3 to yield ξ˜n , ξ˜n−1 , . . . , ξ˜3 . (Note that i stops
at 3 because the sample size needs to be at least two for the sample variance to exist)
P
4. Plot (k, ξˆk ) where ξˆk = (k − 2)−1 ki=3 ξ˜i , k = n, . . . , 3.
We shall call the resulting plot the log variance (LV) plot as it is based on the sample variance of the log-transformed exceedances. Note that Step 4 uses an average of ξ˜ value to
smooth volatile sample variance values in the plot. This is similar to the idea found in
Resnick and Stǎricǎ (1997) where the averaged Hill estimators are plotted to reduce the
variability of the standard plot. In the following two subsections we compare the performances of the proposed algorithm and the Hill estimator for simulated and real datasets
from heavy tailed distributions, respectively.
4.1
Comparison between Hill plot and LV plotSimulated data from GPD and GEV
Our simulated datasets are from the GPD and GEV, two most prominent distributions in the
EVT studies. To compare the performance of the Hill and LV plots, we consider several
GPD and GEV distributions with different choices for parameters ξ, σ and µ. The location
parameter µ has also been added in our simulation study to allow diverse heavy tail shapes.
As our LV plot algorithm above is invariant under the location shifting in X, it can handle
these simulated data with no modification. For each GPD and GEV chosen, we generate
samples of size 2,000 and present two selected sample paths for illustration. We have tried
larger sample sizes and more repetitions, but the conclusions are similar. As distributions
with finite upper bounds hardly concern us in real applications, we focus on ξ > 0 cases
only and negative estimates of ξ are accordingly replaced with 0 in computing ξ˜ in the algorithm; this is also consistent with the Hill plot which works only for ξ > 0. The results
are depicted in Figures 4 to 9 where the first three figures used GPD samples and the last
three used GEV samples. In each figure, the upper and lower panels represent the Hill and
LV plots, respectively. The vertical line shows the value of estimated ξ. The horizontal
axis stands for the ordered data count from the largest observation; for example, 200 in the
horizontal axis means the largest 200 observations in calculating the estimate. Note that we
have limited the range of the figures to the largest 1,000 observations even though the same
size is 2,000. So, basically the left end area in each figure is the region where the GPD
phenomenon takes place and we wish to find the tail index in this area.
From these six figures we see that, depending the combination of the parameter sets,
the Hill plot can be stable but also can have a trend of increasing or decreasing over the
whole dataset, confirming that the Hill plot should be read only over a small fraction of the
dataset on the left end side of the figure, or the tail region. However, since we do not know
exactly where the tail starts, any trend makes the Hill plot harder to decipher. Practical
experience (e.g., Section 7.2 in McNeil et al. (2005)) suggests the 5% upper tail to be used
to determine the tail index with the Hill plot. When focused on this region in each figure,
the Hill plot is sometimes able to reveal the true tail index (e.g., Figures 5 and 7), but in
other cases there is hardly any stable region in the plot (e.g., Figures 6 and 9). Thus the
Hill plot, despite its popularity, remains difficult to use in practical situations. If we turn
to the varaince plot, finding the tail index is relatively easier because it has been smoothed
via averaging, though both plots can be qualitatively poor when the sample itself has large
variation in the tail. For example, in Figures 4 and 8, both plots are relatively stable at the
left end region but they point at the wrong tail index. In Figures 6 and 9, both plots are either
unstable or biased at the left end area; in particular, the Hill plot is highly volatile and the
LV plot has a considerable bias. Compared to the Hill plot, however, a marked advantage
of the LV plot throughout all figures is that it exhibits a quite stable region in not so extreme
quantile levels without any substantial trend, and in this region the estimated value is fairly
close to the true tail index, again, relative to the Hill plot. These observations suggest that
the LV plot can be most useful when one reads the plot between, say, upper 5% and 20%
quantile regions. This is advantageous for practitioners to interpret the plot as the volatility
of extreme quantiles can be avoided in this non-extreme quantile range.
1.0
0.5
Hill plot
0.0
200
400
0
200
400
600
800
1000
600
800
1000
0.5
0.0
LV plot
1.0
0
Order Statistics
1.0
0.5
Hill plot
1.5
Figure 4: Comparison of ξ for GPD with µ = 10, σ = 1, ξ = 0.5
200
400
0
200
400
600
800
1000
600
800
1000
1.0
0.5
LV plot
1.5
0
Order Statistics
2.0
1.5
Hill plot
2.5
Figure 5: Comparison of ξ for GPD with µ = 0, σ = 1, ξ = 1
200
400
0
200
400
600
800
1000
600
800
1000
2.0
1.5
LV plot
2.5
0
Order Statistics
Figure 6: Comparison of ξ for GPD with µ = 10, σ = 1, ξ = 2
1.0
0.5
Hill plot
0.0
200
400
0
200
400
600
800
1000
600
800
1000
0.5
0.0
LV plot
1.0
0
Order statistics
1.0
0.5
Hill plot
1.5
Figure 7: Comparison of ξ for GEV with µ = 0, σ = 1, ξ = 0.5
200
400
0
200
400
600
800
1000
600
800
1000
1.0
0.5
LV plot
1.5
0
Order statistics
2.0
1.5
Hill plot
2.5
Figure 8: Comparison of ξ for GEV with µ = 0, σ = 1, ξ = 1
200
400
0
200
400
600
800
1000
600
800
1000
2.0
1.5
LV plot
2.5
0
Order statistics
Figure 9: Comparison of ξ for GEV with µ = 10, σ = 1, ξ = 2
4.2
Real datasets
We now repeat the same exercise for two actual datasets publicaly available: the Danish
fire data and the BMW stock price log-returns. The famous Danish fire data, collected at
Copenhagen Reinsurance, comprise 2,167 fire losses between 1980 and 1990, both years
inclusive. The numbers have been adjusted for inflation to reflect 1985 values and are expressed in millions of Danish Krone. The BMW data consists of daily log-returns during
January 2, 1973 and July 23, 1996, with a total of 6,146 observations. Both datasets can be
obtained from fExtremes package in R software, and further description of these datasets
can be found in Embrechts et al. (1997).
The Hill and LV plots for these two datasets are presented in Figure 10. For the Danish
data (top panel), previously two choices for the tail index have been suggested. According
to Section 6.4 and 6.5 of Embrechts et al. (1997), the first threshold is at a loss of 10 (109
exceedances) with the corresponding Hill estimate α̂−1 = 0.618, indicating that only the
first moment exists. For the second choice, the GPD threshold is set at 18 (47 exceedances)
and the Hill estimate is α̂−1 = 0.497, suggesting the existence of the first two moments. So
essentially α−1 = ξ is indicated to be in [0.5, 0.62] by the Hill plot. However, the chosen
GPD thresholds, 10 and 18, were suggested based on the mean excess plot because the Hill
plot itself does not show any stable region within the upper 5% quantile range as seen from
the figure. When we use the LV plot, the path is more tamed and easier to investigate. In
fact, following the recommendation from the previous simulation study, we search the plot
between upper 5% and 20% quantile regions and estimate that the true ξ lies in [0.45, 0.65].
This is largely in agreement with the previously suggested interval above, though our interval is slightly wider. Regarding the BMW data, we first observe that both plots are relatively
stable at the left end, implying that the extreme tail variability is not substantial. The Hill
plot suggests that ξ lies roughly in [0.25, 0.3] by inspecting the upper 5% quantile range,
indicating that only the first 3 or 4 moments are existent. In contrast, the LV plot indicates a
slightly different interval of [0.19, 0.25] based on the overall plot, indicating a less heavier
tail with finite moments up to 4 or 5. As the LV plot in this data has a very narrow interval
for ξ value over the entire data range, we have more confidence in estimating the tail index.
We comment that, as observed from previous simulation studies, the interval for ξ in the LV
plot can be identified in a much clearer manner for these real datasets compared to the Hill
plot.
0.9
Hill plot
Danish data
0.7
0.5
0.3
Estimates
LV plot
200
400
600
800
1000
1.0
0
0.8
0.4
0.6
LV plot
0.0
0.2
Estimates
Hill plot
BMW data
0
200
400
600
800
1000
Order statistics
Figure 10: Plot comparison of ξ for Danish (top) and BMW data (bottom)
5. Concluding remarks
In this paper we propose and study the Exponentiated Generalized Pareto Distribution
(exGPD), which is created via log-transform of the Generalized Pareto Distribution (GPD),
an influential distribution in Extreme Value Theory. For this distribution we derive various
distributional quantities, including the moment generating function, tail risk measures and
quantities related to order statistics. As an application we develop a plot as an alternative to
the Hill plot to identify the tail index of heavy tailed datasets. The proposed plot is based
on the idea of the sample variance of log exceedances to be matched to the variance of the
exGPD. Through various numerical illustrations with both simulated and actual datasets, it
is shown that the proposed plot works reasonably well compared to the Hill plot, elucidating
the usefulness of the exGPD.
References
Artzner, P., Delbaen, F., Eber, J. M., and Heath, D. (1999). Coherent measure of risk.
Mathematical Finance, 9, 203–228.
Balkema, A. A. and De Haan, L. (1974). Residual life time at great age. The Annals of
Probability, pages 792–804.
Beirlant, J., Goegebeur, Y., Segers, J., and Teugels, J. (2006). Statistics of extremes: theory
and applications. John Wiley & Sons.
Castillo, E. and Hadi, A. S. (1997). Fitting the generalized Pareto distribution to data.
Journal of the American Statistical Association, 92(440), 1609–1620.
de Zea Bermudez, P. and Kotz, S. (2010). Parameter estimation of the generalized pareto
distribution—part i. Journal of Statistical Planning and Inference, 140(6), 1353–1373.
Drees, H., De Haan, L., and Resnick, S. (2000). How to make a Hill plot. Annals of
Statistics, pages 254–274.
Embrechts, P., Kluppelberg, C., and Mikosch, T. (1997). Modelling Extremal Events for
Insurnace and Finance. Springer-Verlag.
Grimshaw, S. D. (1993). Computing maximum likelihood estimates for the generalized
pareto distribution. Technometrics, 35(2), 185–191.
He, X. and Fung, W. (1993). Method of medians for lifetime data with weibull models.
Stat. Med. v18, 2009.
Hill, B. M. (1975). A simple general approach to inference about the tail of a distribution.
Annals of Statistics, 3(5), 1163–1174.
Hosking, J. R. and Wallis, J. R. (1987). Parameter and quantile estimation for the generalized pareto distribution. Technometrics, 29(3), 339–349.
Johnston, N., Kotz, S., and Balakrishnan, N. (1994). Continuous Univariate Distributions,
volume 2. New York: Wiley.
Lieblein, J. (1953). On the exact evaluation of the variances and covariances of order
statistics in samples from the extreme-value distribution. The Annals of Mathematical
Statistics, pages 282–287.
McNeil, A. J., Frey, R., and Embrechts, P. (2005). Quantitative Risk Management. Princeton University Press, New Jersey.
Park, M. H. and Kim, J. H. (2016). Estimating extreme tail risk measures with generalized
pareto distribution. Computational Statistics & Data Analysis, 98, 91 – 104.
Pickands III, J. (1975). Statistical inference using extreme order statistics. the Annals of
Statistics, pages 119–131.
Resnick, S. and Stǎricǎ, C. (1997). Smoothing the Hill estimator. Advances in Applied
Probability, pages 271–293.
Smith, R. L. (1984). Threshold methods for sample extremes. In Statistical extremes and
applications, pages 621–638. Springer.
Song, J. and Song, S. (2012). A quantile estimation for massive data with generalized pareto
distribution. Computational Statistics & Data Analysis, 56(1), 143–150.
Zhang, J. (2007). Likelihood moment estimation for the generalized pareto distribution.
Australian & New Zealand Journal of Statistics, 49(1), 69–77.
| 10math.ST
|
Convergence in Infinitary Term Graph Rewriting Systems is
Simple (Extended Abstract)∗
Patrick Bahr
Department of Computer Science, University of Copenhagen
Universitetsparken 5, 2100 Copenhagen, Denmark
[email protected]
In this extended abstract, we present a simple approach to convergence on term graphs that allows
us to unify term graph rewriting and infinitary term rewriting. This approach is based on a partial
order and a metric on term graphs. These structures arise as straightforward generalisations of the
corresponding structures used in infinitary term rewriting. We compare our simple approach to a
more complicated approach that we developed earlier and show that this new approach is superior
in many ways. The only unfavourable property that we were able to identify, viz. failure of full
correspondence between weak metric and partial order convergence, is rectified by adopting a strong
convergence discipline.
1
Introduction
In infinitary term rewriting [17] we study infinite terms and infinite rewrite sequences. Typically, this
extension to infinite structures is formalised by an ultrametric on terms, which yields infinite terms by
metric completion and provides a notion of convergence to give meaning to infinite rewrite sequences.
In this paper we extend infinitary term rewriting to term graphs. In addition to the metric approach, we
also consider the partial order approach to infinitary term rewriting [4] and generalise it to the setting of
term graphs.
One of the motivations for studying infinitary term rewriting is its relation to non-strict evaluation,
which is used in programming languages such as Haskell [18]. Non-strict evaluation defers the evaluation of an expression until it is “needed” and thereby allows us to deal with conceptually infinite data
structures and computations. For example, the function from defined below constructs for each number
n the infinite list of consecutive numbers starting from n:
from(n) = n :: from(s(n))
This construction is only conceptual and only results in a terminating computation if it is used in a context
where only finitely many elements of the list are “needed”. Infinitary term rewriting provides us with an
explicit limit construction to witness the outcome of an infinite computation as it is, for example, induced
by from. After translating the above function definition to a term rewrite rule from(x) → x :: from(s(x)),
we may derive an infinite rewrite sequence
from(0) → 0 :: from(s(0)) → 0 :: s(0) :: from(s(s(0))) → . . .
which converges to the infinite term 0 :: s(0) :: s(s(0)) :: . . . , which represents the infinite list of numbers
0, 1, 2, . . . – as intuitively expected.
∗ The
full version of this paper will appear in Mathematical Structures in Computer Science [7].
R. Echahed and D. Plump (Eds.): 7th International Workshop on
Computing with Terms and Graphs
EPTCS 110, 2013, pp. 17–28, doi:10.4204/EPTCS.110.4
c P. Bahr
This work is licensed under the
Creative Commons Attribution License.
18
Convergence in Infinitary Term Graph Rewriting Systems is Simple (Extended Abstract)
Non-strict evaluation is rarely found in isolation, though. Usually, it is implemented as lazy evaluation [14], which complements a non-strict evaluation strategy with sharing. The latter avoids duplication of subexpressions by using pointers instead of copying. For example, the function from above
duplicates its argument n – it occurs twice on the right-hand side of the defining equation. A lazy evaluator simulates this duplication by inserting two pointers pointing to the actual argument.
While infinitary term rewriting is used to model the non-strictness of lazy evaluation, term graph
rewriting models the sharing part of it. By endowing term graph rewriting with a notion of convergence
like in infinitary term rewriting, we aim to unify the two formalisms into one calculus, thus allowing us
to model both aspects within the same calculus.
Contributions & Outline At first we recall the basic notions of infinitary term rewriting (Section 2).
Afterwards, we construct a metric and a partial order on term graphs and show that both are suitable
as a basis for notions of convergence in term graph rewriting (Section 3). Based on these structures we
introduce notions of convergence (weak and strong variants) for term graph rewriting and show correspondences between metric-based and partial order-based convergence (Section 4.1 and 4.2). We then
present soundness and completeness properties of the resulting infinitary term graph rewriting calculi
w.r.t. infinitary term rewriting (Section 4.3). Lastly, we compare our calculi with previous approaches
(Section 5).
2
Infinitary Term Rewriting
Before starting with the development of infinitary term graph rewriting, we recall the basic notions of
infinitary term rewriting. Rewrite sequences in infinitary rewriting, also called reductions, are sequences
of the form (φι )ι<α , where each φι is a rewrite step from a term tι to tι+1 in a term rewriting system
(TRS) R, denoted φι : tι →R tι+1 . The length α of such a sequence can be an arbitrary ordinal. For
f )
f
example, the infinite reduction indicated in Section 1 is the sequence (φif : tif →R f ti+1
i<ω , where ti =
0 :: . . . :: si−1 (0) :: from(si (0)) for all i < ω and R f is the TRS consisting of the single rule from(x) →
x :: from(s(x)).
2.1
Metric Convergence
The above definition of reductions ensures that consecutive rewrite steps are “compatible”, i.e. the result
term of the ι-th step, viz. tι+1 , is the start term of the (ι + 1)-st step. However, this definition does not
relate the start terms of steps at limit ordinal positions to the terms that preceded it. For example, we
can extend the abovementioned reduction (φif )i<ω of length ω, to a reduction (φif )i<ω+1 of length ω + 1
using any reduction step φωf , e.g. φωf : from(0) → 0 :: from(s(0)). In our informal notation this reduction
(φif )i<ω+1 reads as follows:
from(0) → 0 :: from(s(0)) → 0 :: s(0) :: from(s(s(0))) →
...
from(0) → 0 :: from(s(0))
Intuitively, this does not make sense since the sequence of terms that precedes the last step intuitively
converge to the term 0 :: s(0) :: s(s(0)) :: . . . , but not from(0).
In infinitary term rewriting such reductions are ruled out by a notion of convergence and a notion of
continuity that follows from it. Typically, this notion of convergence is derived from a metric d on the
set of (finite and infinite) terms T ∞ (Σ): d(s,t) = 0 if s = t, and d(s,t) = 2−d otherwise, where d is the
P. Bahr
19
minimal depth at which s and t differ. Using this metric, we may also construct the set of (finite and
infinite) terms T ∞ (Σ) by metric completion of the metric space (T (Σ), d) of finite terms.
The mode of convergence in the metric space (T ∞ (Σ), d) is the basis for the notion of weak mconvergence of reductions: a reduction S = (φι : tι →R tι+1 )ι<α is weakly m-continuous if limι→λ tι = tλ
m
for all limit ordinals λ < α; it weakly m-converges to a term t, denoted S : t0 ,→
R t, if it is weakly mb is the length of the underlying sequence of terms (tι )ι<αb . For
continuous and limι→αb tι = t, where α
example, the reduction (φif )i<ω weakly m-converges to the term 0 :: s(0) :: s(s(0)) :: . . . ; but the sequence
(φif )i<ω+1 does not weakly m-converge, it is not even weakly m-continuous as limι→ω tιf is not from(0).
Weak m-convergence is quite a general notion of convergence. For example, given a rewrite rule
a → a, we may derive the reduction a → a → . . . , which weakly m-converges to a even though the rule
a → a is applied again and again at the same position. This generality causes many desired properties to
break, such as unique normal form properties and compression [16]. That is why Kennaway et al. [16]
introduced strong m-convergence, which in addition requires that the depth at which rewrite steps take
place tends to infinity as one approaches a limit ordinal: Let S = (φι : tι →πι tι+1 )ι<α be a reduction,
where each πι indicates the position at which the step φι takes place and |πι | denotes the length of the
position πι . The reduction S is said to be strongly m-continuous (resp. strongly m-converge to t, denoted
S : t0 m
t) if it is weakly m-continuous (resp. weakly m-converges to t) and if (|πι |)ι<λ tends to infinity
for all limit ordinals λ < α (resp. λ ≤ α). For example, the reduction (φif )i<ω also strongly m-converges
to the term 0 :: s(0) :: s(s(0)) :: . . . . On the other hand, the reduction a → a → . . . indicated above weakly
m-converges to a, but it does not strongly m-converge to a.
2.2
Partial Order Convergence
Alternatively to the metric approach illustrated in Section 2.1, convergence can also be formalised using
a partial order ≤⊥ on terms. The idea to use this partial order for infinitary rewriting goes back to
Corradini [12]. The signature Σ is extended to the signature Σ⊥ by adding a fresh constant symbol ⊥.
When dealing with terms in T ∞ (Σ⊥ ), we call terms that do not contain the symbol ⊥, i.e. terms that
are contained in T ∞ (Σ), total. We define s ≤⊥ t iff s can be obtained from t by replacing some subterm
occurrences in t by ⊥. Interpreting the term ⊥ as denoting “undefined”, ≤⊥ can be read as “is less defined
than”. The pair (T ∞ (Σ⊥ ), ≤⊥ ) is known to form a complete semilattice [13], i.e. it has a least element
F
⊥, each directed set D in (T ∞ (Σ⊥ ), ≤⊥ ) has a least
upper
bound
(lub)
D, and every non-empty set B
d
∞
in (T (Σ⊥ ), ≤⊥ ) has greatest lower bound (glb) B. In particular, this means
d that for
any sequence
F
∞
(tι )ι<α in (T (Σ⊥ ), ≤⊥ ), its limit inferior, defined by lim infι→α tι = β <α
β ≤ι<α tι , exists.
In the same way that the limit in the metric space gives rise to weak m-continuity/-convergence, the
limit inferior gives rise to weak p-continuity and weak p-convergence; simply replace lim by lim inf. We
p t if a reduction S starting with term t weakly p-converges to t. The defining difference
write S : t0 ,→
0
between the two approaches is that p-continuous reductions always p-converge. The reason for that lies
in the complete semilattice structure of (T ∞ (Σ⊥ ), ≤⊥ ), which guarantees that the limit inferior always
exists (in contrast to the limit in a metric space).
The definition of the strong variant of p-convergence is a bit different from the one of m-convergence,
but it follows the same idea: a reduction (φi : ti →πi ti+1 )i<ω weakly m-converges iff the minimal depth
di at which two consecutive terms ti ,ti+1 differ tends to infinity. The strong variant of m-convergence is
a conservative approximation of this condition; it requires |πi | to tend to infinity. This approximation is
conservative since |πi | ≤ di ; differences between consecutive terms can only occur below the position at
which a rewrite rule was applied.
20
Convergence in Infinitary Term Graph Rewriting Systems is Simple (Extended Abstract)
In the partial order approach we can make this approximation more precise since we have the whole
term structure at our disposal instead of only the measure provided by the metric d. In the case of
m-convergence, we replaced the actual depth of a minimal difference di with its conservative underapproximation |πi |. For p-convergence, we replace the glb ti u ti+1 , which intuitively represents the
common information shared by ti and ti+1 , with the conservative under-approximation ti [⊥]πi , which
replaces the redex at position πi in ti with ⊥. This term ti [⊥]πi – called the reduction context of the
step φi : ti → ti+1 – is a lower bound of ti and ti+1 w.r.t. ≤⊥ and is, thus, also smaller than ti u ti+1 . The
definition of strong p-convergence is obtained from the definition of weak p-convergence by replacing
lim infι→λ tι with lim infι→λ tι [⊥]πι .
A reduction S = (φι : tι →πι tι+1 )ι<α is called strongly p-continuous if lim infι→λ ti [⊥]πi = tλ for all
p t, if it is strongly p-continuous and
limit ordinals λ < α; it strongly p-converges to t, denoted S : t0
either lim infι→α ti [⊥]πi = t in case α is a limit ordinal, or t = tα+1 otherwise.
Example 2.1. The previously mentioned reduction (φif )i<ω both strongly and weakly p-converges to
the infinite term 0 :: s(0) :: s(s(0)) :: . . . – like in the metric approach. However, while the reduction
a → a → . . . does not strongly m-converge, it strongly p-converges to the term ⊥.
The partial order approach has some advantages over the metric approach. As explained above, every
p-continuous reduction is also p-convergent. Moreover, strong p-convergence has some properties such
as infinitary normalisation and infinitary confluence of orthogonal systems [4] that are not enjoyed by
strong m-convergence.
Interestingly, however, the partial order-based notions of convergence are merely conservative extensions of the metric-based ones:
Theorem 2.1 ([2, 4]). For every reduction S in a TRS, the following equivalences hold:
m t
(i) S : s ,→
iff
p t in T ∞ (Σ).
S : s ,→
(ii) S : s m
t
iff
p t in T ∞ (Σ).
S: s
The phrase “in T ∞ (Σ)” means that all terms in S are total (including t). That is, if restricted to total
terms, m- and p-convergence coincide.
3
Graphs and Term Graphs
In this section, we present our notion of term graphs and generalise the metric d and the partial order ≤⊥
from terms to term graphs.
Our notion of graphs and term graphs is largely taken from Barendregt et al. [8].
Definition 3.1 (graphs). A graph over signature Σ is a triple g = (N, lab, suc) consisting of a set N (of
nodes), a labelling function lab : N → Σ, and a successor function suc : N → N ∗ such that |suc(n)| =
ar(lab(n)) for each node n ∈ N, i.e. a node labelled with a k-ary symbol has precisely k successors. If
suc(n) = hn0 , . . . , nk−1 i, then we write suci (n) for ni .
The successor function suc defines, for each node n, directed edges from n to suci (n). A path from a
node m to a node n is a finite sequence he0 , . . . , el i of numbers such that n = sucel (. . . suce0 (m)), i.e. n is
reached from m by taking the e0 -th edge, then the e1 -th edge etc.
Definition 3.2 (term graphs). A term graph g over Σ is a tuple (N, lab, suc, r) consisting of an underlying
graph (N, lab, suc) over Σ whose nodes are all reachable from the root node r ∈ N. The class of all term
graphs over Σ is denoted G ∞ (Σ). A position of n ∈ N in g is a path in the underlying graph of g from
r to n. The set of all positions of n in g is denoted Pg (n). The depth of n in g, denoted depthg (n), is
the minimum of the lengths of the positions of n in g, i.e. depthg (n) = min {|π| | π ∈ Pg (n) }. The term
P. Bahr
21
graph g is called a term tree if each node in g has exactly one position. We use the notation N g , labg ,
sucg and rg to refer to the respective components N,lab, suc and r of g. Given a graph or a term graph h
and a node n in h, we write h|n to denote the sub-term graph of h rooted in n.
The notion of homomorphisms is crucial for dealing with term graphs. For greater flexibility, we
will parametrise this notion by a set of constant symbols ∆ for which the homomorphism condition is
suspended. This will allow us to deal with variables and partiality appropriately.
Definition 3.3 (∆-homomorphisms). Let Σ be a signature, ∆ ⊆ Σ(0) , and g, h ∈ G ∞ (Σ). A ∆-homomorphism
φ from g to h, denoted φ : g →∆ h, is a function φ : N g → N h with φ (rg ) = rh that satisfies the following
equations for all for all n ∈ N g with labg (n) 6∈ ∆:
labg (n) = labh (φ (n))
φ (sucgi (n))
=
suchi (φ (n))
(labelling)
g
for all 0 ≤ i < ar(lab (n))
(successor)
Note that, for ∆ = 0,
/ we get the usual notion of homomorphisms on term graphs (e.g. Barendsen [9])
and from that the notion of isomorphisms. The nodes labelled with symbols in ∆ can be thought of as
holes in the term graphs that can be filled with other term graphs.
We do not want to distinguish between isomorphic term graphs. Therefore, we use a well-known
trick [19] to obtain canonical representatives of isomorphism classes of term graphs.
Definition 3.4. A term graph g is called canonical if n = Pg (n) holds for each n ∈ N g . That is, each
node is the set of its positions in the term graph. The set of all (finite) canonical term graphs over Σ
is denoted GC∞ (Σ) (resp. GC (Σ)). For each term graph h ∈ GC∞ (Σ), its canonical representative C (h) is
obtained from h by replacing each node n in h by Ph (n).
This construction indeed yields a canonical representation of isomorphism classes. More precisely:
g∼
= C (g) for all g ∈ G ∞ (Σ), and g ∼
= h iff C (g) = C (h) for all g, h ∈ G ∞ (Σ).
We consider the set of terms T ∞ (Σ) as the subset of canonical term trees of GC∞ (Σ). With this
correspondence in mind, we can define the unravelling of a term graph g as the unique term U (g) such
that there is a homomorphism φ : U (g) → g. For example, g0 from Figure 1 is the unravelling of g1 ,
and h0 and gω from Figure 2 both unravel to the infinite term @( f , @( f , . . . )). Term graphs that unravel
to the same term are called bisimilar.
3.1
A Simple Partial Order on Term Graphs
In this section, we want to establish a partial order suitable for formalising convergence of sequences of
canonical term graphs similarly to weak p-convergence on terms.
Weak p-convergence on term rewriting systems is based on the partial order ≤⊥ on T ∞ (Σ⊥ ), which
instantiates occurrences of ⊥ from left to right, i.e. s ≤⊥ t iff t is obtained by replacing occurrences of ⊥
in s by arbitrary terms in T ∞ (Σ⊥ ). Analogously, we consider the class of partial term graphs simply as
term graphs over the signature Σ⊥ = Σ ] {⊥}. In order to generalise the partial order ≤⊥ to term graphs,
we need to formalise the instantiation of occurrences of ⊥ in term graphs. For this purpose, we shall
use ∆-homomorphisms with ∆ = {⊥}, or ⊥-homomorphisms for short. A ⊥-homomorphism φ : g →⊥ h
maps each node in g to a node in h while “preserving its structure”. Except for nodes labelled ⊥ this
also includes preserving the labelling. This exception to the homomorphism condition allows the ⊥homomorphism φ to instantiate each ⊥-node in g with an arbitrary node in h. Using ⊥-homomorphisms,
we arrive at the following definition for our simple partial order ≤S⊥ on term graphs:
Definition 3.5. For each g, h ∈ GC∞ (Σ⊥ ), define g ≤S⊥ h iff there is some φ : g →⊥ h.
22
Convergence in Infinitary Term Graph Rewriting Systems is Simple (Extended Abstract)
One can verify that ≤S⊥ indeed generalises the partial order ≤⊥ on terms. Considering terms as
canonical term trees, we obtain the following characterisation of ≤⊥ on terms s,t ∈ T ∞ (Σ⊥ ):
s ≤⊥ t ⇐⇒ there is a ⊥-homomorphism φ : s →⊥ t.
The first important result for ≤S⊥ is that the semilattice structure that we already had for ≤⊥ is
preserved by this generalisation:
Theorem 3.1. The partially ordered set (GC∞ (Σ⊥ ), ≤S⊥ ) forms a complete semilattice.
For terms, we already know that the set of (potentially infinite) terms can be constructed by forming
the ideal completion of the partially ordered set (T (Σ⊥ ), ≤⊥ ) of finite terms [11]. More precisely, the
ideal completion of (T (Σ⊥ ), ≤⊥ ) is order isomorphic to (T ∞ (Σ⊥ ), ≤⊥ ).
An analogous result can be shown for term graphs:
Theorem 3.2. The ideal completion of (GC (Σ⊥ ), ≤S⊥ ) is order isomorphic to (GC∞ (Σ⊥ ), ≤S⊥ ).
3.2
A Simple Metric on Term Graphs
Next, we shall generalise the metric d from terms to term graphs. To achieve this, we need to formalise
what it means for two term graphs to coincide up to a certain depth, so that we can reformulate the definition of the metric d for term graphs. To this end, we follow the same idea that the original definition of
d on terms from Arnold and Nivat [1] was based on. In particular, we introduce a truncation construction
that cuts off nodes below a certain depth:
Definition 3.6. Let g ∈ G ∞ (Σ⊥ ) and d ≤ ω. The simple truncation g†d of g at d is the term graph defined
as follows:
N g†d = n ∈ N g depthg (n) ≤ d
rg†d = rg
(
(
g
lab
(n)
if
depth
(n)
<
d
sucg (n) if depthg (n) < d
g
labg†d (n) =
sucg†d (n) =
⊥
if depthg (n) = d
hi
if depthg (n) = d
The definition of the simple metric d† follows straightforwardly:
Definition 3.7. The simple distance d† : GC∞ (Σ) × GC∞ (Σ) → R+
0 is defined as follows:
(
0
if g = h
d† (g, h) =
2−d if g 6= h and d = max {e < ω | g†e ∼
= h†e }
Again, we can verify that d† generalises d. In particular, we can show that our notion of truncation
coincides with that of Arnold and Nivat [1] if restricted to terms.
As desired, this generalisation retains the complete ultrametric space structure:
Theorem 3.3. The pair (GC∞ (Σ), d† ) forms a complete ultrametric space.
The metric space analogue to ideal completion is metric completion. On terms, we already know that
we can construct the set of (potentially infinite) terms T ∞ (Σ) by metric completion of the metric space
(T (Σ), d) of finite terms [10]. More precisely, the metric completion of (T (Σ), d) is the metric space
(T ∞ (Σ), d). This property generalises to term graphs as well:
Theorem 3.4. The metric completion of (GC (Σ), d† ) is the metric space (GC∞ (Σ), d† ).
P. Bahr
23
f
f
f
f
f
c
c
(g0 )
c
(g1 )
c
c
(g2 )
c
(g4 )
c
c
(gω )
Figure 1: Limit inferior in the presence of acyclic sharing.
r @
l @
x
Y
r @
l @
@
Y
x
@
@
ρ2
f
Y
f
Y
(ρ1 )
(ρ2 )
(g0 )
(a) Term graph rules that unravel to Y x → x (Y x).
@
@
f
Y
ρ1
f
(b) A single ρ2 -step.
@
@
ρ1
f
@
(g1 )
ρ1
@
f
@
@
Y
(g0 )
(h0 )
Y
(g2 )
@
(gω )
(c) A strongly m-convergent term graph reduction over ρ1 .
Figure 2: Implementation of the fixed point combinator as a term graph rewrite rule.
4
Infinitary Term Graph Rewriting
In this paper, we adopt the term graph rewriting framework of Barendregt et al. [8]. In order to represent
placeholders in rewrite rules, we use variables – in a manner much similar to term rewrite rules. To this
end, we consider a signature ΣV = Σ ] V that extends the signature Σ with a set V of nullary variable
symbols.
Definition 4.1 (term graph rewriting systems). Given a signature Σ, a term graph rule ρ over Σ is a
triple (g, l, r) where g is a graph over ΣV and l, r ∈ N g such that all nodes in g are reachable from l or
r. We write ρl resp. ρr to denote the left- resp. right-hand side of ρ, i.e. the term graph g|l resp. g|r .
Additionally, we require that for each variable v ∈ V there is at most one node n in g labelled v, and we
have that n 6= l and that n is reachable from l in g. A term graph rewriting system (GRS) R is a pair
(Σ, R) with Σ a signature and R a set of term graph rules over Σ.
The notion of unravelling straightforwardly extends to term graph rules: the unravelling of a term
graph rule ρ, denoted U (ρ), is the term rule U (ρl ) → U (ρr ). The unravelling of a GRS R = (Σ, R),
denoted U (R), is the TRS (Σ, {U (ρ) | ρ ∈ R }).
Example 4.1. Figure 2a shows two term graph rules which both unravel to the term rule ρ : @(Y, x) →
@(x, @(Y, x)) that defines the fixed point combinator Y . Note that sharing of nodes is used both to refer
to variables in the left-hand side from the right-hand side and in order to simulate duplication.
24
Convergence in Infinitary Term Graph Rewriting Systems is Simple (Extended Abstract)
Without going into all details of the construction, we describe the application of a rewrite rule ρ
with root nodes l and r to a term graph g in four steps: at first a suitable sub-term graph of g rooted
in some node n of g is matched against the left-hand side of ρ. This matching amounts to finding a
V -homomorphism φ from the left-hand side ρl to g|n , the redex. The V -homomorphism φ allows us to
instantiate variables in the rule with sub-term graphs of the redex. In the second step, nodes and edges
in ρ that are not in ρl are copied into g, such that each edge pointing to a node m in ρl is redirected to
φ (m). In the next step, all edges pointing to the root n of the redex are redirected to the root n0 of the
contractum, which is either r or φ (r), depending on whether r has been copied into g or not (because it
is reachable from l in ρ). Finally, all nodes not reachable from the root of (the now modified version of)
g are removed. With h the result of the above construction, we obtain a pre-reduction step ψ : g 7→n h
from g to h.
The definition of term graph rewriting in the form of pre-reduction steps is very operational. While
this style is beneficial for implementing a rewriting system, it is problematic for reasoning on term
graphs modulo isomorphism, which is necessary for introducing notions of convergence. However, one
can easily see that the construction of the result term graph of a pre-reduction step is invariant under
isomorphism, which justifies the following definition of reduction steps:
Definition 4.2. Let R = (Σ, R) be GRS, ρ ∈ R and g, h ∈ GC∞ (Σ) with n ∈ N g and m ∈ N h . A tuple
φ = (g, n, h) is called a reduction step, written φ : g →n h, if there is a pre-reduction step φ 0 : g0 7→n0 h0
with C (g0 ) = g, C (h0 ) = h, and n = Pg0 (n0 ). We also write φ : g →R h to indicate R.
In other words, a reduction step is a canonicalised pre-reduction step. Figure 2b and Figure 2c
illustrate some (pre-)reduction steps induced by the rules ρ1 respectively ρ2 shown in Figure 2a.
4.1
Weak Convergence
In analogy to infinitary term rewriting, we employ the partial order ≤S⊥ and the metric d† for the purpose
of defining convergence of transfinite term graph reductions.
Definition 4.3. Let R = (Σ, R) be a GRS.
(i) Let S = (gι →R gι+1 )ι<α be a reduction in R. S is weakly m-continuous in R if limι→λ gι = gλ
m
for each limit ordinal λ < α. S weakly m-converges to g ∈ GC∞ (Σ) in R, written S : g0 ,→
R g, if it
is weakly m-continuous and limι→αb gι = g.
(ii) Let R⊥ be the GRS (Σ⊥ , R) over the extended signature Σ⊥ and S = (gι →R⊥ gι+1 )ι<α a reduction
in R⊥ . S is weakly p-continuous in R if lim infι<λ gi = gλ for each limit ordinal λ < α. S
p
weakly p-converges to g ∈ GC∞ (Σ⊥ ) in R, written S : g0 ,→
R g, if it is weakly p-continuous and
lim infι<αb gi = g.
Example 4.2. Figure 2c illustrates an infinite reduction derived from the rule ρ1 in Figure 2a. Since
gi †(i + 1) ∼
= gω †(i + 1) for all i < ω, we have that limi→ω gi = gω , which means that the reduction
weakly m-converges to the term graph gω . Moreover, since each node in gω eventually appears in a term
graph in (gi )i<ω and remains stable afterwards, we have lim infi→ω gι = gω . Consequently, the reduction
also weakly p-converges to gω .
Recall that weak p-convergence for TRSs is a conservative extension of weak m-convergence (cf.
Theorem 2.1). The key property that makes this possible is that for each sequence (tι )ι<α in T ∞ (Σ),
we have that limι→α tι = lim infι→α tι whenever (tι )ι<α converges, or lim infι→α tι is a total term. Sadly,
this is not the case for the metric space and the partial order on term graphs: the sequence of term graphs
depicted in Figure 1 has a total term graph as its limit inferior, viz. gω , although it does not converge in
P. Bahr
25
the metric space. In fact, since the sequence in Figure 1 alternates between two distinct term graphs, it
does not converge in any Hausdorff space, i.e. in particular, it does not converge in any metric space.
This example shows that we cannot hope to generalise the compatibility property that we have for
terms: even if a sequence of total term graphs has a total term graph as its limit inferior, it might not
converge. However, the converse direction of the correspondence does hold true:
Theorem 4.1. If (gι )ι<α converges, then limι→α gι = lim infι→α gι .
From this property, we obtain the following relation between weak m- and p-convergence:
Theorem 4.2. Let S be a reduction in a GRS R. If
m
S : g ,→
R h
then
p
S : g ,→
R h.
As indicated above, weak m-convergence is not the total fragment of weak p-convergence as it is the
case for TRSs, i.e. the converse of the above implication does not hold in general:
Example 4.3. There is a GRS that yields the reduction shown in Figure 1, which weakly p-converges
to gω but is not weakly m-convergent. This reduction can be produced by alternately applying the rules
ρ1 , ρ2 , where the left hand side of both rules and the right-hand side of ρ1 is g0 , and the right-hand side
of ρ2 is g1 .
4.2
Strong Convergence
The idea of strong convergence is to conservatively approximate the convergence behaviour somewhat
independently from the actual rewrite rules that are applied. Strong m-convergence in TRSs requires
that the depth of the redexes tends to infinity thereby assuming that anything at the depth of the redex
or below is potentially affected by a reduction step. Strong p-convergence, on the other hand, uses a
better approximation that only assumes that the redex is affected by a reduction step – not however other
subterms at the same depth. To this end strong p-convergence uses a notion of reduction contexts –
essentially the term minus the redex – for the formation of limits. The following definition provides the
construction for the notion of reduction contexts that we shall use for term graph rewriting:
Definition 4.4. Let g ∈ G ∞ (Σ⊥ ) and n ∈ N g . The local truncation of g at n, denoted g\n, is obtained
from g by labelling n with ⊥ and removing all outgoing edges from n as well as all nodes that thus
become unreachable from the root.
Proposition 4.1. Given a reduction step g →n h, we have g\n ≤S⊥ g, h.
This means that the local truncation at the root of the redex is preserved by reduction steps and is
therefore an adequate notion of reduction context for strong p-convergence [3]. Using this construction
we can define strong p-convergence on term graphs analogously to strong p-convergence on terms. For
strong m-convergence, we simply take the same notion of depth that we already used for the definition
of the simple truncation g†d and thus the simple metric d† .
Definition 4.5. Let R = (Σ, R) be a GRS.
(i) The reduction context c of a graph reduction step φ : g →n h is the term graph C (g\n). We write
φ : g →c h to indicate the reduction context of a graph reduction step.
(ii) Let S = (gι →nι gι+1 )ι<α be a reduction in R. S is strongly m-continuous in R if limι→λ gι = gλ
and (depthgι (nι ))ι<λ tends to infinity for each limit ordinal λ < α. S strongly m-converges to g
in R, denoted S : g0 m
R g, if it is strongly m-continuous and either S is closed with g = gα or S is
open with g = limι→α gι and (depthgι (nι ))ι<α tending to infinity.
26
Convergence in Infinitary Term Graph Rewriting Systems is Simple (Extended Abstract)
(iii) Let S = (gι →cι gι+1 )ι<α be a reduction in R⊥ = (Σ⊥ , R). S is strongly p-continuous in R if
lim infι→λ cι = gλ for each limit ordinal λ < α. S strongly p-converges to g in R, denoted
p
S : g0
R g, if it is strongly p-continuous and either S is closed with g = gα or S is open with
g = lim infι→α cι .
Example 4.4. As explained in Example 4.2, the reduction in Figure 2c both weakly m- and p-converges
to gω . Because contraction takes place at increasingly large depth, the reduction also strongly mconverges to gω . Moreover, since each node in gω eventually appears also in the sequence of reduction contexts (ci )i<ω of the reduction and remains stable afterwards, we have that lim infi→ω ci = gω .
Consequently, the reduction also strongly p-converges to gω .
Remarkably, one of the advantages of the strong variant of convergence is that we regain the correspondence between m- and p-convergence that we know from infinitary term rewriting:
Theorem 4.3 ([5]). Let R be a GRS and S a reduction in R⊥ . We then have that
S: g m
R h
iff
p
∞
S: g
R h in GC (Σ).
In particular, the GRS given in Example 4.3 that induces the reduction depicted in Figure 1 does not
provide a counterexample for the “if” direction of the above equivalence – in contrast to weak convergence. The reduction in Figure 1 does not strongly m-converge but it does strongly p-converge to the
term graph ⊥, which is in accordance with Theorem 4.3 above.
4.3
Soundness and Completeness
In order to assess the value of the modes of convergence on term graphs that we introduced in this paper,
we need to compare them to the well-established counterparts on terms. Ideally, we would like to see
a strong connection between converging reductions in a GRS R and converging reductions in the TRS
U (R) in the form of soundness and completeness properties. For example, for m-convergence we want
m
m
m
to see that g ,→
R h implies U (g) ,→U (R) U (h) – i.e. soundness – and vice versa that U (g) ,→U (R) t
m
implies g ,→R h with U (h) = t – i.e. completeness.
Completeness is already an issue for finitary rewriting [15]: a single term graph redex may correspond to several term redexes due to sharing. Hence, contracting a term graph redex may correspond to
several term rewriting steps, which may be performed independently.
In the context of weak convergence, also soundness becomes an issue. The underlying reason for
this issue is similar to the phenomenon explained above: a single term graph rewrite step may repres1
ent several term rewriting steps, i.e. g →R h implies U (g) →+
U (R) U (h). When we have a converging term graph reduction (φι : gι → gι+1 )ι<α , we know that the underlying sequence of term graphs
(gι )ι<αb converges. However, the corresponding term reduction does not necessarily produce the sequence (U (gι ))ι<αb but may intersperse the sequence (U (gι ))ι<αb with additional intermediate terms,
which might change the convergence behaviour.
While we cannot prove soundness for weak convergence due to the abovementioned problems, we
can show that the underlying modes of convergence are sound in the sense that convergence is preserved
under unravelling.
Theorem 4.4.
(i) limι→α gι = g implies limι→α U (gι ) = U (g) for every sequence (gι )ι<α in (GC∞ (Σ), d† ).
(ii) U (lim infι→α gι ) = lim infι→α U (gι ) for every sequence (gι )ι<α in (GC∞ (Σ⊥ ), ≤S⊥ ).
1 If
the term graph g is cyclic, the corresponding term reduction may even be infinite.
P. Bahr
27
Note that the above theorem in fact implies soundness of the modes of convergence on term graphs
with the modes of convergence on terms since both d† and ≤S⊥ specialise to d respectively ≤⊥ if restricted
to term trees.
However, we can observe that strong convergence is more well-behaved than weak convergence. It
is possible to prove soundness and completeness properties for strong p-convergence:
Theorem 4.5 ([5]). Let R be a left-finite GRS.
p
p
(i) If R is left-linear and g
R h, then U (g) U (R) U (h).
p
p
p
(ii) If R is orthogonal and U (g)
U (R) t, then there are reductions g R h and t U (R) U (h).
Note that the above completeness property is not the one that one would initially expect, namely
p
p
U (g)
U (R) t implies g R h with U (h) = t. But this general completeness property is known to
already fail for finitary term graph rewriting [15].
The soundness and completeness properties above have an important practical implication: GRSs
that only differ in their sharing, i.e. they unravel to the same TRS, will produce the same results, i.e. the
same normal forms up to bisimilarity. GRSs with more sharing may, however, reach a result with fewer
steps. This can be observed in Figure 2, which depicts two rules ρ1 , ρ2 that unravel to the same term rule.
Rule ρ1 reaches gω in ω steps whereas ρ2 reaches a term graph h0 , which is bisimilar to gω , in one step.
The situation for strong m-convergence is not the same as for strong p-convergence. While we do
have soundness under the same preconditions, i.e. g m
R h implies U (g) m
U (R) U (h), the completeness
property we have seen in Theorem 4.5 fails. This behaviour was already recognised by Kennaway et
al. [15]. Nevertheless, we can find a weaker form of completeness that is restricted to normalising
reductions:
Theorem 4.6 ([5]). Given an orthogonal, left-finite GRS R that is normalising w.r.t. strongly m-converging
reductions, we find for each normalising reduction U (g) m
U (R) t a reduction g m
R h such that t =
U (h).
5
Concluding Remarks
We have devised two independently defined but closely related infinitary calculi of term graph rewriting.
This is not the first proposal for infinitary term graph rewriting calculi; in previous work [6] we presented
a so-called rigid approach based on a metric and a partial order different from the structures presented
here.
There are several arguments why the simple approach presented in this paper is superior to the rigid
approach. First of all it is simpler. The rigid metric and partial order have been carefully crafted in order
to obtain a correspondence result in the style of Theorem 2.1 for weak convergence on term graphs.
This correspondence result of the rigid approach is not fully matched by the simple approach that we
presented here, but we do regain the full correspondence by moving to strong convergence.
Secondly, the rigid approach is very restrictive, disallowing many reductions that are intuitively convergent. For example, in the rigid approach the reduction depicted in Figure 2c, would not p-converge
(weakly or strongly) to the term graph gω as intuitively expected but instead to the term graph obtained
from gω by replacing f with ⊥. Moreover, this sequence would not m-converge (weakly or strongly) at
all.
Lastly, as a consequence of the restrictive nature of the rigid approach, the completion constructions
of the underlying metric and partial order do not yield the full set of term graphs – in contrast to our
findings here in Theorem 3.2 and 3.4.
28
Convergence in Infinitary Term Graph Rewriting Systems is Simple (Extended Abstract)
Unfortunately, we do not have solid soundness or completeness results for weak convergence apart
from the preservation of convergence under unravelling and the metric/ideal completion construction of
the set of term graphs. However, as we have shown, this shortcoming is again addressed by moving to
strong convergence.
References
[1] A. Arnold & M. Nivat (1980): The metric space of infinite trees. Algebraic and topological properties.
Fundam. Inf. 3(4), pp. 445–476.
[2] P. Bahr (2009): Infinitary Rewriting - Theory and Applications. Master’s thesis, Vienna University of Technology, Vienna.
[3] P. Bahr (2010): Abstract Models of Transfinite Reductions. In C. Lynch, editor: RTA’10, 6, pp. 49–66,
doi:10.4230/LIPIcs.RTA.2010.49.
[4] P. Bahr (2010): Partial Order Infinitary Term Rewriting and Böhm Trees. In C. Lynch, editor: RTA’10, 6, pp.
67–84, doi:10.4230/LIPIcs.RTA.2010.67.
[5] P. Bahr (2012): Infinitary Term Graph Rewriting is Simple, Sound and Complete. In A. Tiwari, editor:
RTA’12, 15, pp. 69–84, doi:10.4230/LIPIcs.RTA.2012.69.
[6] P. Bahr (2012): Modes of Convergence for Term Graph Rewriting. Logical Methods in Computer Science
8(2):6, doi:10.2168/LMCS-8(2:6)2012.
[7] P. Bahr (2013): Convergence in Infinitary Term Graph Rewriting Systems is Simple. Math. Struct. in Comp.
Science, to appear.
[8] H.P. Barendregt, M.C.J.D. van Eekelen, J.R.W. Glauert, R. Kennaway, M.J. Plasmeijer & M.R. Sleep (1987):
Term graph rewriting. In Philip C. Treleaven Jaco de Bakker, A. J. Nijman, editor: PARLE’87, LNCS 259,
Springer, pp. 141–158, doi:10.1007/3-540-17945-3 8.
[9] E. Barendsen (2003): Term Graph Rewriting. In Terese, editor: Term Rewriting Systems, chapter 13, Cambridge University Press, pp. 712–743.
[10] M. Barr (1993): Terminal coalgebras in well-founded set theory. Theor. Comput. Sci. 114(2), pp. 299 – 315,
doi:10.1016/0304-3975(93)90076-6.
[11] G. Berry & J.-J. Lévy (1977): Minimal and optimal computations of recursive programs. In: POPL’77, pp.
215–226, doi:10.1145/512950.512971.
[12] A. Corradini (1993): Term rewriting in CTΣ . In M.-C. Gaudel & J.-P. Jouannaud, editors: TAPSOFT’93, pp.
468–484, doi:10.1007/3-540-56610-4 83.
[13] J.A. Goguen, J.W. Thatcher, E.G. Wagner & J.B. Wright (1977): Initial Algebra Semantics and Continuous
Algebras. J. ACM 24(1), pp. 68–95, doi:10.1145/321992.321997.
[14] P. Henderson & J.H. Morris, Jr. (1976):
doi:10.1145/800168.811543.
A lazy evaluator.
In:
POPL’76, pp. 95–103,
[15] R. Kennaway, J.W. Klop, M.R. Sleep & F.-J. de Vries (1994): On the adequacy of graph rewriting for simulating term rewriting. ACM Trans. Program. Lang. Syst. 16(3), pp. 493–523, doi:10.1145/177492.177577.
[16] R. Kennaway, J.W. Klop, M.R. Sleep & F.-J. de Vries (1995): Transfinite Reductions in Orthogonal Term
Rewriting Systems. Inf. Comput. 119(1), pp. 18–38, doi:10.1006/inco.1995.1075.
[17] R. Kennaway & F.-J. de Vries (2003): Infinitary Rewriting. In Terese, editor: Term Rewriting Systems, 1st
edition, chapter 12, Cambridge University Press, pp. 668–711.
[18] S. Marlow (2010): Haskell 2010 Language Report.
[19] D. Plump (1999): Term graph rewriting. In H. Ehrig, G. Engels, H.-J. Kreowski & G. Rozenberg, editors:
Handbook of Graph Grammars and Computing by Graph Transformation, 2, World Scientific Publishing Co.,
Inc., pp. 3–61, doi:10.1142/9789812815149 0001.
| 6cs.PL
|
arXiv:1703.00674v2 [cs.AI] 21 Oct 2017
Adaptive Matching for Expert Systems with
Uncertain Task Types
Virag Shah, Lennart Gulikers, Laurent Massoulié
Milan Vojnović
Microsoft Research-Inria Joint Centre
Palaiseau, France, 91120
{virag.shah, lennart.gulikers, laurent.massoulie}@inria.fr
Dept. of Statistics, London School of Economics
London, UK, WC2A 2AE
[email protected]
Abstract—Online two-sided matching markets such as Q&A
forums (e.g. StackOverflow, Quora) and online labour platforms
(e.g. Upwork) critically rely on the ability to propose adequate
matches based on imperfect knowledge of the two parties to be
matched. This prompts the following question: Which matching
recommendation algorithms can, in the presence of such uncertainty, lead to efficient platform operation?
To answer this question, we develop a model of a task / server
matching system. For this model, we give a necessary and sufficient condition for an incoming stream of tasks to be manageable
by the system. We further identify a so-called back-pressure
policy under which the throughput that the system can handle
is optimized. We show that this policy achieves strictly larger
throughput than a natural greedy policy. Finally, we validate
our model and confirm our theoretical findings with experiments
based on logs of Math.StackExchange, a StackOverflow forum
dedicated to mathematics.
I. I NTRODUCTION
Online platforms that enable matches between trading partners in two-sided markets have recently blossomed in many areas: LinkedIn and Upwork facilitate matches between employers and employees; Uber allows matches between passengers
and car drivers; Airbnb and Booking.com connect travelers
and housing facilities; Quora and Stack Exchange facilitate
matches between questions and either answers, or experts able
to provide them.
All these systems crucially rely on the ability to propose
adequate matches based on imperfect knowledge of the characteristics of the two parties to be matched. For example, in
the context of online labour platforms, there is uncertainty
about both the skill sets of candidate employees and the
job requirements. Similarly, in the context of online Q&A
platforms, there is uncertainty about both question types and
users’ ability to provide answers.
This naturally leads to the following question: which matching recommendation algorithms can, in the presence of such
uncertainty, lead to efficient platform operation? A natural
measure of efficiency is the throughput that the platform
achieves, i.e. the rate of successful matches it allows. To
address this question, one thus needs first to characterize
fundamental limits on the achievable throughput.
In this paper, we progress towards answering these questions
as follows.
First, we propose a simple model of such platforms, which
features a static collection of servers, or experts on the one
hand, and a continuous stream of arrivals of tasks, or jobs, on
the other hand. In our model, the platform’s operation consists
of servers iteratively attempting to solve tasks. After being
processed by some server, a task leaves the system if solved;
otherwise it remains till successfully treated by some server.
To model uncertainty about task types, we assume that for
each incoming task we are given the prior distribution of this
task’s “true type”. Servers’ abilities are then represented via
the probability that each server has to solve a task of given
type after one attempt at it.
In a Q&A platform scenario, tasks are questions, and servers
are experts; a server processing a task corresponds to an
expert providing an answer to a question. A task being solved
corresponds to an answer being accepted. In an online labour
platform, tasks could be job offers, and a server may be a pool
of workers with similar abilities. A server processing a task
then corresponds to a worker being interviewed for a job, and
the task is solved if the interview leads to a hire. We could
also consider the dual interpretation when the labour market is
constrained by workers rather than job offers. Then a task is
a worker seeking work, while a server is a pool of employers
looking for hires.
An important feature of our model consists in the fact that
when a task’s processing does not lead to failure, it does
however affect uncertainty about the task’s type. Indeed, the a
posteriori distribution of the task’s type after a failed attempt
on it by some server differs from its prior distribution. For
instance in a Q&A scenario, a question which an expert in
Calculus failed to answer either is not about Calculus, or is
very hard.
For our model, we then determine necessary and sufficient
conditions for an incoming stream of task arrivals to be
manageable by the servers, or in other words, determine
achievable throughputs of the system. In the process we
introduce candidate policies, in particular the greedy policy
according to which a server choses to serve tasks for which
its chance of success is highest. This scheduling strategy is
both easy to implement and is based on a natural motivation.
Surprisingly perhaps, we show that it is not optimal in the
throughput it can handle. In contrast, we introduce a so-called
backpressure policy inspired from the wireless networking
literature [1], which we prove to be throughput-optimal.
We summarize contributions of this paper as follows:
We propose a new model of a generic task-expert system
that allows for uncertainty of task types, heterogeneity of
skills, and recurring attempts of experts in solving tasks.
• We provide a full characterization of the stability region,
or sustainable throughputs, of the task-expert system under
consideration. We establish that a particular backpressure
policy is throughput-optimal, in the sense that it supports
maximum task arrival rate under which the system is stable.
• We show that there exist instances of task-expert systems
under which simple matching policies such as a natural
greedy policy and a random policy can only support a much
smaller maximum task arrival rate, than the backpressure
policy.
• We report the results of empirical analysis of the popular
Math.StackExchange Q&A platform which establish heterogeneity of skills of experts, with experts knowledgeable
across different types of tasks and others specialized in
particular types of tasks. We also show numerical evaluation
results that confirm the benefits of the backpressure policy
on greedy and random matchmaking policies.
The remainder of the paper is structured as follows. Section II presents our system model. In Section III, we present
results for two baseline matchmaking policies, namely Greedy
and Random. Section IV presents the characterization of task
arrival rates that can be supported under which the system
is stable and prove the superiority of backpressure policy
over Random and Greedy. In Section V, we present our
experimental results. Related work is discussed in Section VI.
We conclude in Section VII. Proofs of the results are provided
in Section VIII. In the main text, we summarize the key ideas
for proofs in brief.
•
II. P ROBLEM S ETTING
Let C = {c1 , . . . , ck } be the set of task types. Each task is in
the system is of a particular type in C. Let S = {s1 , . . . , sm }
be the set of servers (or experts) present in the system. When
a server s ∈ S attempts to resolve a task of type c ∈ C, the
outcome is 1 (a success) with probability ps,c and it is 0 (a
failure) with probability 1−ps,c . Upon success we say that the
task is resolved. In the context of online hiring platform, this
is equivalent to successful hiring of an employee for a job. In
the context of Q&A platform, this is equivalent to an answer
by an expert being accepted by the asker of the question.
We consider a Bayesian setting where we have a prior
distribution z = {zc }c∈C ∈ C for a task’s type, where C is the
set of all distributions. Note, different tasks may have different
prior distributions. Clearly, if server s processes a task with
prior distribution z then the probability that it fails is given by
X
ψs (z) =
zc (1 − ps,c ).
(1)
c∈C
Further, upon failure, the posterior distribution of task’s type
is given by
zc (1 − ps,c )
.
(2)
φs (z) =
ψs (z)
c∈C
Note that the posterior distribution of a task’s type upon
failure by a subset of servers does not depend on the sequence
in which these servers resolve the task, i.e., for each s, s0 ∈ S
we have φs ◦ φs0 = φs0 ◦ φs . At any point in time a task
is associated with a ‘mixed-type’ which is defined as the
posterior distribution of its type given the past attempts.
We allow a task to be attempted sequentially by multiple
servers until it is resolved. We would like to resolve the tasks
as quickly as possible.
A. Single Task Scenario
Before considering the setting of online task arrivals, for
ease of exposition we first consider a toy scenario with single
task. Suppose that time t ∈ Z+ is discrete. A task arrives
at time t = 0. Let the prior distribution of its type upon
arrival (equivalently, its mixed-type at time t = 0) be z. At a
time, only one server attempts to resolve a task. Consider the
problem of designing a sequence of servers (s(t) : 0 ≤ t ≤ τ )
such that the probability that the task is resolved within a
fixed time τ is maximized. Let z(0) = z, and for each t ≥ 1
let z(t) = φs(t−1) (z(t − 1)), i.e., z(t) is the mixed-type of
the task at time t given that it was not resolved upon previous
attempts. Then the probability
Qτ that the task is resolved by time
τ is given as f (τ ) = 1 − t=0 ψs(t) (z(t))
Contrast this with the Bayesian active learning setting in
[3], [4] where the goal is to reduce uncertainty in true
hypothesis via outcome from several experiments. Using a
diminishing returns property called adaptive submodularity the
authors in [3] obtain a policy which is competitive with the
optimal. In our setting, f (τ ) is a submodular function. Thus
a greedy policy where s(t) for each t is chosen to be from
arg mins ψs (z(t)) is 1 − 1/e-competitive [5].
The feedback assumed in [3] is more general than the binary
feedback assumed in our case. However we keep the binary
feedback assumption since for hiring platforms the outcome of
a job-employee match is binary and is notified to the platform
by the agents involved. Further feedback on the performance
of the agents may also be reported, but such a feedback is
not reliable as it is often biased towards the larger values [6].
Even for a Q&A setting the model is relevant since only one
answer from the available responses may be accepted by the
asker, such as in StackOverflow.
Further, in this paper we add an extra dimension to the
problem which was not considered in the [3], [4], namely,
we consider the setting of online task arrivals where tasks of
different mixed-types may compete for the servers resources
before they leave upon being resolved. We design throughput
optimal policies under such a setting.
B. Online Task Arrivals
We consider a continuous time setting, i.e., t ∈ R+ . Tasks
arrive at a rate of λ per time unit on average. The subsequent
mixed types Zi of incoming tasks are assumed i.i.d., taking
values in a countable subset Z of C, and we let πz = P (Zi =
z) for all z ∈ Z. Finally, the time for server s ∈ S to complete
an attempt on a job takes on average 1/µs time units, and such
attempt durations are i.i.d.. All involved sources of randomness
are independent.
We assume that Z is closed under φs (·), i.e., for each z ∈
Z, φs (z) ∈ Z. This loses no generality, as the closure of
a countable set with respect to a finite number of maps φs
remains countable.
We assume that a given task may be inspected several times
by a given server and assume that the outcomes success /
failure are independent at each inspection. This can be justified
if a label s in fact represents a collection of experts with similar
abilities, in which case multiple processings by s correspond
to processing by distinct individual experts.
For such a setting we would like to minimize the expected
sojourn time of a typical task, i.e., the expected time between
the arrival and the resolving of a typical task. Recall that the
success probabilities ps,c are assumed to be arbitrary. Under
such a heterogeneous setting minimizing expected sojourn
time is a hard problem. In fact, this is true even when there
is no uncertainty in task types. As a proxy to sojourn time
optimal policies, we will strive for throughput optimal policies.
In particular, we will characterize the arrival rates λ for which
the system can be stabilized, i.e. for which there exists a
scheduling policy which induces a time-stationary regime of
the system’s behavior. Indeed for a stable system the long term
task resolution rate coincides with the task arrival rate λ, and
thus throughput-optimal policies must make the system stable
whenever this is possible. Note that for an unstable system
the number of outstanding tasks accumulate over time and the
expected sojourn time tends to infinity.
Finally, for simplicity we assume more specifically that
the tasks arrive at the instants of a Poisson process with
intensity λ, and that the time for server s to complete an
attempt at a task follows an Exponential distribution with
parameter µs . These are continuous time analog of i.i.d.
arrivals and independent departures per time slot in discreet
time setting. These assumptions will imply that the system
state at any given time t can be represented as a Markov
process, which simplifies throughput analysis. The system
throughput is often insensitive to such statistical assumptions
on arrival and service times [7].
We close the section with some additional assumptions and
notations which will aid our analysis.
At any time t let Nz (t) represent the number of tasks of
mixed-type z present in the system and N (t) = {Nz (t)}z∈Z .
We also let z(s, t) denote the mixed type of the task that
server s works on at time t. For strategies such that the servers
select which task to handle based uniquely on the vector N (t),
the process {N (t)}t≥0 forms a continuous-time Markov chain
(CTMC) [8], [9]. The policies considered in this paper are
studied by analyzing the associated CTMC.
We allow a task to be assigned to multiple experts at a given
time. Further, we allow pre-emptive service, i.e., an expert may
drop service of a task should a new task arrive into a system
or an existing task receive a response.
III. BASELINE P OLICIES
Following the discussion in Section II-A, could it be that a
greedy approach may work well even under the online setting?
From throughput perspective, a natural Greedy policy is one
where each expert is assigned a task which best suits its skills,
i.e., among the outstanding tasks, an expert s is assigned a task
of a mixed-type z which minimizes ψs (z). We now show a
surprising result that for a class of asymmetric systems the
Greedy policy is as suboptimal as a Random policy where
each server chooses a task uniformly at random. Let us first
formally define these policies.
Definition 1 (Greedy Policy): A policy is Greedy if given
the system state each expert is assigned an outstanding task
which maximizes its success probability, i.e., for each N such
that |N | > 0 we have
z(s) ∈ As (N ) = arg min ψs (z),
z:Nz >0
where ties could be broken uniformly at random among this
set.
The transition rates for the CTMC under greedy policy can
be given as follows. Let q(n, n0 ) be the transition rate from
state n to state n0 . Let ez denote the vector with all coordinates
equal to 0 except for the z-coordinate which equals 1. Fix a
state n. For each z ∈ Z we have
q(n, n + ez ) = λπz ,
X
q(n, n − ez ) =
µs (1 − ψs (z))
s:z∈As (N )
q(n, n − ez + eφs (z) ) =
X
s:z∈As (N )
1
,
|As (N )|
µs ψs (z)
1
.
|As (N )|
Transition rate q(n, n0 ) for each (n, n0 ) which is not as given
above is equal to 0.
Definition 2 (Random Policy): A policy is Random if each
expert s is assigned a task chosen uniformly at random from
the pool of outstanding tasks.
We will consider the following class of task-expert system.
Definition 3 (Asymmetric(a) System): Suppose that there are
two task types C = {c1 , c2 } and two experts S = {s1 , s2 }.
Each arrival is equally likely to be of both types, i.e., πz0 = 1
where z 0 satisfies zc0 = 1/2 for each c ∈ C, and πz = 0
if z 6= z 0 . Both experts provide responses at unit rate, i.e.,
µs = 1 for each s. Further, for class c1 we have ps,c1 = 1
for each s ∈ S, and for class c2 we have ps1 ,c2 = a < 1,
and ps2 ,c2 = 0. We refer to such a task-expert system as a
Asymmetric(a) system with parameter a.
For this class of systems, if a task of mixed-type z 0 receives
a failure from either of the experts, then its mixed type
becomes z 00 where zc001 = 0 and zc002 = 1. Thus, it is
sufficient to assume that Z = {z 0 , z 00 } where zc0 = 21 for
each c ∈ C, and zc00 = 1{c = c2 }. Further, it is easy to check
that ψs1 (z 0 ) = (1 − a)/2, ψs1 (z 00 ) = 1 − a, ψs2 (z 0 ) = 1/2,
and ψs2 (z 00 ) = 1. We then have the following result.
Theorem 1: For a given 0 < a < 1, consider the
Asymmetric(a) system as defined in Definition 3. The following statements hold:
(i) The Greedy stabilizes the system if and only if λ <
4a/(2 + a).
(ii) The Random stabilizes the system if and only if λ <
4a/(2 + a).
(iii) There exists a policy which stabilizes the system if λ <
min {3a/(a + 1), 2a}.
(iv) If λ > min {3a/(a + 1), 2a} then no policy can stabilize
the system.
Proof of the theorem is provided in Section VIII. We
provide below the key ides, in brief. For part (i) we use a
suitably chosen linear Lyapunov function. For stability, we
show that the Lyapunov function has negative drift for all but
finitely many states and then use a variant of Lyapunov-Forster
Theorem which follows from Theorem 8.13 in [10]. Part (ii)
follows from the following more general result.
Theorem 2: Under Random policy, a system is stable if and
only if it satisfies the following:
!−1
P
X
z∈Z zc πz
P
,
λ<
s∈S µs ps,c
c∈C
To prove this result we use fluid limit approach developed
in [11]–[13]. Let Xc (t) be the number of tasks in the system
of pure-type c. Let X(t) = {Xc (t)}c . Roughly, given initial
condition X(0) = x, we let limk→∞ k1 X(0) = x, and study
limk→∞ k1 X(kt). We use the following Lyapunov function:
X
Xc
P
L(X) =
Xc log
,
γc c0 Xc0
c
P
z π
z
.
where γc , λ P z∈Zµscps,c
s∈S
Parts (iii) and (iv) of Theorem 1 follow from Theorem 3 in
the Section IV.
Note that for an Asymmetric(a) system the stability threshold for task arrival rates under optimal policy can be up to
25% higher (namely, when a = 1/2) than that under either
Greedy or Random.
While it is intuitive that Random may perform poorly as
compared to an optimal policy, it is counterintuitive that
Greedy may perform as sub-optimally as Random. The reason
for the poor performance of Greedy can be explained as
follows. In the Asymmetric(a) system, we have a flexible
expert, i.e. an expert for tasks of all pure-types, and a specialized expert, i.e. an expert only for pure-type c1 . Under
Greedy policy, all experts prioritize the newly arriving tasks
as it optimizes the probability of achieving success in the
short term. However, a larger long-term throughput could be
achieved if the flexible expert could focus more on the lagging
tasks, i.e., the tasks of pure-type c2 .
IV. O PTIMAL S TABILITY
Main goal of this section is to provide necessary and
sufficient conditions for stability of the system, and to provide
explicit policies which stabilize the system when the sufficient
conditions are satisfied.
We obtain stability conditions via capacity constraints and
flow conservation constraints which capture the flow of tasks
from one type to another upon service by an expert. For
instance, if νs,z represents the flow of tasks of mixed-type
z served by expert s, a fraction 1 − ψs (z) of it leaves the
system due to success while the rest gets converted into a flow
of type φs (z).
flow of mixed-type z,
P The total arrival rate of
0
i.e., λπz + s∈S,z
0 ∈φ−1 (z) νs,z 0 ψs (z ), must match the total
P s
service rate, i.e.,
P s∈S νs,z . Further, the total flow service rate
expert s, i.e., z∈Z νs,z , must be less than its service capacity
µs . The following is the main result of this section.
Theorem 3: Suppose there exists s such that minc ps,c > 0.
If there exist non-negative real numbers νs,z , for s ∈ S, and
z ∈ Z and positive real numbers δs , for s ∈ S such that the
following hold:
X
X
∀z ∈ Z, λπz +
νs,z0 ψs (z 0 ) =
νs,z , (3)
s∈S,z 0 ∈φ−1
s (z)
∀s ∈ S,
X
s∈S
νs,z + δs ≤ µs ,
(4)
z∈Z
then there exists a policy under which the system is stable. If
there does not exist non-negative real numbers νs,z , for s ∈ S,
z ∈ Z and non-negative real numbers δs for s ∈ S such that
the above constraints hold, then the system cannot be stable.
We use the condition of existence of an expert s such that
minc ps,c > 0 only for a technical reason to simplify our proof.
We believe that the result holds even when this condition is
not true.
We now describe the policy which achieves optimal stability.
We need some more notation to describe the policy. Consider
a set Y ⊂ Z. Let X(t) be the number of tasks in the system at
time t which have been of mixed-type z ∈ Z\Y. For z ∈ Y,
let X̃z (t) and Ñz (t) be the tasks of mixed-type z which have
and have not had mixed-type in Z\Y. Also, for convenience,
for each z ∈ Z\Y, let X̃z be the tasks of mixed-typePz, i.e.,
X̃z = Nz for each z ∈ Z\Y. Thus, we have X = z X̃z .
Consider the following policy.
Definition 4 (Backpressure(Y) policy): For a given Y, let
X, and {Ñz }z∈Y be as defined above. For each s ∈ S, z ∈ Y
let
(
Ñz − ψs (z)Ñφs (z) , if φs (z) ∈ Y
ws,z (Ñ , X) =
.
Ñz − ψs (z)X,
if φs (z) ∈ Z\Y
Define
Bs (Ñ , X) = arg
max
z 0 ∈Y:Ñz0 >0
ws,z (Ñ , X).
If
X
s
µs
max
z∈Y:Ñz >0
ws,z (Ñ , X) ≥ X min
c∈C
X
µs ps,c
s
then each expert chooses a task in Ñz where z ∈ Bs (Ñ , X) ⊂
Y with ties broken arbitrarily. Else, each expert serves a task
in X chosen uniformly at random.
The following theorem establishes throughput optimality of
the Backpressure(Y) policy.
Theorem 4: Suppose there exists s such that minc ps,c > 0.
If the sufficient conditions for stability as given in the statement of Theorem 3 are satisfied, then there exists a finite
subset Y of Z such that the policy Backpressure(Y) stabilizes
the system.
In particular, the Backpressure(Z) policy is optimally stable
for Asymmetric(a) system as defined in Definition 3.
To prove Theorem 4, we again use LyapunovForster
theorem. Then,
under Backpressure(Y) policy,
{Ñz }z∈Y , {X̃z }z∈Z is a CTMC. We use the following
Lyapunov function to show stability:
!2
X
X
X
L(Ñ , X̃) =
Ñz2 +
Xz
=
Ñz2 + X 2 .
z∈Y
z∈Z
z∈Y
As such, proving this result requires significantly different
approach as compared to stability proofs via quadratic Lyapunov functions of classical constrained queueing networks
with finite number of queues. In particular, the flow equations
do not directly give a stabilizing static policy. To avoid this
pitfall, we find finite set Y such that the overall arrival rate
into Z\Y is small, and ‘pool’ the slack capacity at the servers
to serve the infinite number of queues in Z\Y.
The stability part of Theorem 3 follows from Theorem 4.
For the converse statement in Theorem 3, we use ergodicity
of the underlying Markov chain.
We now provide an alternative policy which achieves stability under a weaker condition that ps,c are bounded away
from 0 and 1, but with the advantage that it does not rely
on the precise numbers of jobs Nz sharing the same mixed
type z, but rather on ‘local averages’. As such it may remain
optimally stable even when the distribution of mixed types of
incoming jobs is no longer assumed to be discrete.
Definition 5 (Backpressure() policy): Partition set C into
finitely many subsets Ai , i = 1, . . . , l, such that each Ai has
diameter
at most , that is for all z, z 0 ∈ Ai weP
have |z −z 0 | =
P
0
c |zc − zc | ≤ . We then define N (Ai ) :=
z∈Ai Nz , and
the backpressure with respect to server s of a given z as
ws,z (N ) := N (Ai ) − ψs (z)N (Aj ),
where i and j are such that z ∈ Ai and φs (z) ∈ Aj . Then,
each expert chooses a task with mixed-type in
As (N ) = arg
max
z∈Z:Nz >0
ws,z (N ),
with ties broken uniformly at random
We then have the following:
Theorem 5: Suppose that there exists α > 0 such that for
each s, c we have
ps,c ∈ [α, 1 − α].
(5)
Suppose further that the sufficient conditions for stability as
given in the statement of Theorem 3 are satisfied. Then, there
exists an > 0 sufficiently small such that the Backpressure()
policy stabilizes the system.
PFor its 2proof, we use the Lyapunov function L(N ) =
i N (Ai ) . Again, the proof involves a significantly different approach as compared to stability proofs for standard
constrained queuing networks with finite number of queues.
In particular, we develop and use new flow equations which
account for not only the sets Ai associated with the mixedtypes of the tasks but also the lengths of the history of the
tasks.
Unlike backpressure policy proposed in [1] under a different
setting, which was agnostic to system arrival rates, a set Y
(or the ) such that the policy Backpressure(Y) (or policy
Backpressure()) stabilizes the system may depend on the
value of λ. While the policies as stated may be complex to
implement, it allows us to develop implementable heuristics
which significantly outperform greedy policy. We demonstrate
this in the next section.
V. E XPERIMENTAL R ESULTS
In this section, we present our empirical results obtained
by using data from Math.Stack-Exchange Q&A platform. In
this platform, users post tagged questions that are answered
by other users. The asker may select one of the submitted
answers as the best answer for the given question, which is
made public information in the platform. We will use this data
to estimate the success probabilities of experts in answering
questions, and use these parameters in simulations to compare
the throughputs that can be achieved by greedy, random,
and backpressure policies. As we will see, a substantially
larger throughput can be achieved by backpressure policy than
greedy and random.
a) Dataset: The dataset consists of around 702, 286
questions and 994, 138 answers. It was retrieved on February
2nd, 2017. The top 11 most common tags are given in Table I
in decreasing order of popularity. Among these tags, the
most common is calculus which covers 61, 184 questions, and
the least common is complex analysis which covers 22, 813
questions. In our analysis, we used only questions that are
tagged with at least one of the 11 most popular tags, which
amounts to a total of 381, 239 questions and 544, 267 answers.
b) Estimated skill sets: The success probabilities of
answering questions are estimated as follows. For a given usertag pair, the success probability is estimated by the empirical
frequency of the accepted answers by this user for questions of
given tag, conditional on that the user had at least 5 accepted
answers for questions of the given tag, and otherwise we
estimate the success probability is set to be equal to zero.
These success probabilities are estimated for 2, 000 users
with the most accepted answers. Among these users, the user
with the most accepted answers had 4, 665 accepted answers,
and the user with the least number of accepted answers had
13 accepted answers. 712 users had more than 50 answers
accepted. In order to form clusters of users with similar
success probabilities for different tags, we ran the k-means
clustering algorithm.
The estimated success probabilities are shown in Table I.
The columns correspond to different centroids of the clusters
TABLE I: Skills of experts estimated by using data from the Math.Stack-Exchange Q&A platform.
Tag
calculus
real-analysis
linear-algebra
probability
abstract-algebra
integration
sequences-and-series
general-topology
combinatorics
matrices
complex-analysis
Size
1
.32
.17
.46
.07
.02
.09
.05
.02
.03
.27
.02
165
2
.39
.41
.29
.49
.05
.43
.32
.10
.14
.15
.19
188
Clusters
3
4
.30
.35
.25
.32
.05
.36
.02
.33
.03
.32
.05
.19
.16
.31
.03
.16
.06
.43
.02
.31
.08
.16
313 200
and give average success probabilities for different tags. In the
bottom row, we give the sizes of the corresponding clusters.
For instance, the 165 persons in cluster 1 have on average
32% of their calculus, and 46% of their linear algebra answers
accepted.
There is a pronounced heterogeneity in user expertise. We
highlighted the success probabilities of value at least 35%. A
subset of users, namely cluster 6, are got at all topics whereas
other clusters are good at a different subsets of topics.
There is also a prevalence of questions with different
combinations of tags, that is, mixed types. When a question of
mixed type arrives with multiple tags, we associated with it a
mixed-type which is the uniform distribution of the associated
tags. We kept only those combinations of tags that occur for
at least 1% of the total number of questions. This results in
16 tag combinations among which 11 are singletons and 5
are combinations of 2 tags. These are the mixed types z with
postive πz . We observed that roughly 19% of the questions
are tagged with multiple tags, showing the relevance of our
model.
c) Simulation setup: We assumed the experts to have unit
service rates. We make this approximation as we do not have
the information about times at which experts begin to respond
a question. We examined the system for increasing values of
task arrival rates. We simulate our CTMC via a custom discrete
event simulator.
We implement the Backpressure(Y) policy where the set
Y consists of all 11 pure types, the 5 most frequently seen
mixed types upon arrival as described above, and the mixed
types that result from an attempt by an expert exactly once.
Note that pure types can be attempted multiple times without
changing its type. We thus have |Y| = 16 + 5 · 10 = 66. Our
choice of Y is a result of a compromise between performance
and complexity. Choosing a larger set of Y may increase
the stability region by a small fraction, but may significantly
increase the complexity of the Backpressure policy.
Further, while serving the tasks in X, instead of choosing
tasks at random, we choose tasks greedily, i.e., each server
5
.37
.23
.14
.02
.02
.44
.20
.02
.04
.02
.14
179
6
.47
.49
.48
.50
.38
.45
.45
.43
.37
.44
.50
183
7
.28
.40
.26
.06
.23
.03
.09
.50
.02
.06
.09
231
8
.16
.10
.31
.02
.50
.01
.04
.07
.06
.11
.05
187
9
.26
.10
.07
.46
.01
.06
.06
.02
.19
.02
.01
178
10
.41
.44
.43
.04
.27
.37
.33
.31
.05
.34
.44
176
is assigned a task in X which maximizes its probability
of success. Empirically, this improves the performance over
random selection of tasks in X.
d) Performance comparison of different policies: We
examined the evolution of the number of tasks in the system
waiting to be served over time for greedy and backpressure
policy for respective task arrival rates, 3.78 and 3.83 (Figure
1 left) and respective task arrival rates 3.83 and 4.08 (Figure
1 right). Even under stability, backpressure may outperform
greedy. Further, while greedy is unstable at λ = 3.83, the
backpressure is stable even at λ = 4.08.
We evaluated the time-averaged number of tasks in the
system for different task arrival rates for the three policies
under our consideration. By Little’s law, the time-averaged
number of tasks is a good proxy for the expected sojourn
time of a typical task. These results are shown in Figure
2. We observe that the task arrival rates at which random,
greedy, and backpressure become unstable are, 2.2, 3.80 and
4.10, respectively. Thus, the backpressure policy achieves
throughput improvement of about 8% over the greedy policy.
While greedy may perform better than backpressure at lower
arrival rates, the backpressure significantly outperforms greedy
at higher arrival rates.
VI. R ELATED W ORK
Bayesian Active Learning [3], [4], [14], [15] aims at
learning true hypothesis by adaptively selecting sequence of
experiments. In [14] labels are obtained for a batch of items at
a time. In [15] a stream based budgeted setting is considered
where a finite number of items arrive in a random order. In
contrast we allow infinite stream of tasks and are interested
in maximizing the task resolution throughput under capacity
constraints at the servers. The crowdsourcing works such as
[16]–[19] consider task assignment problems for classification
with unknown ground truths, however they consider a static
model. In [20] labeling tasks arrive dynamically and their exit
is tied to the expert allocation decisions, in that a task leaves
Fig. 1: Total number of tasks in the system over time for the greedy and backpressure policy. The task arrival rates are as
indicated in the figures.
4500
Number of tasks in the system
4000
3500
Random
Greedy
Backpressure
3000
2500
2000
1500
1000
500
0
0.0
0.5
1.0
1.5
2.0 2.5
Arrival rate
3.0
3.5
4.0
4.5
Fig. 2: Comparison of Random, Greedy, and Backpressure
policies.
once the probability of error in the label estimate falls below
a threshold.
Our work is also broadly related to that of multi-arm
bandits, e.g., see [21]–[24] and citations therein, in the sense
of optimizing the trade-off between exploration, to learn job
types, and exploitation, to optimize task performance. It also
has some relation with collaborative filtering systems such
as those studied in [25]–[27], which can be interpreted as
expert-task systems where success probabilities admit a lowrank matrix structure. Unlike our work, there good matches
are inferred from observed assignments of tasks to experts,
which are according to a given statistical model, and there are
no resources constraints imposed on the experts.
A related line of work is that on stochastic online matching,
e.g., [28]–[30]. The stochastic online matching can be interpreted as a task-expert system where each expert is associated
with a budget constraint that allows to solve at most one
task. Unlike our work where the task types are uncertain,
uncertainty in these models come from the arbitrariness of the
future task arrivals and the monotonically decreasing available
resource budgets.
Another related literature is that of constrained queueing
systems, where arriving tasks are to be served by heterogeneous servers subject to resource constraints, e.g., [31]–[40].
The goal is to efficiently utilize server resources while providing good performance in servicing tasks, e.g., optimizing
task delays. Our matching policy is of a flavor similar to the
stability-optimal backpressure policy first proposed in [1]. The
setting close to ours is the one studied in [40] for routing
queries in peer-to-peer networks. Here, the types of the queries
are known but the locations of nodes where the queries may
by successfully resolved are uncertain. More technically, we
associate queues with each prior distribution which may be
infinite in number. This makes the stability analysis much
more challenging. Another related work is that on scheduling
flexible servers [38], [39], which allows for tasks of different
types and servers of different skills. It has been established
that a so called max-weight policy is optimal in a heavy traffic
regime. The main difference from our work is that all these
works assume that task types are known.
In [41], the authors considered a task-expert system where
task types are of two difficulty levels (hard or easy) and expert
skills are of two levels (senior or junior). Seniors may serve
any task, but juniors may only serve easy tasks. The hardness
of each task is unknown upon arrival. In comparison, we allow
for much more generality with respect to the heterogeneity of
skills of experts. In their model, a task upon service can only
become progressively harder, which amounts to a feed-forward
system, unlike our model.
The work in [42] considers a model where the job types are
known but the expert types are unknown. They consider the
problem of matching while simultaneously learning the expert
types. A key idea is to use a shadow price which simultaneously accounts for resource utilization and type uncertainties.
They consider an asymptotic regime where each expert is
allowed to work on a large number of tasks, a vanishingly
small amount of which could be used to accurately learn the
expert types, and the rest can be served optimally. In the limit,
the learning aspect is decoupled form the expert utilization,
and it is thus different from our work.
VII. C ONCLUSION
We studied matching of tasks and experts in a system with
uncertain task types. We established a complete characterization of the stability region of the system, i.e. the set of task
arrival rates that can be supported by a matching policy such
that the expected number of tasks waiting to be served is finite.
We showed that any task arrival rate in the stability region
can be supported by a back-pressure matching policy. We also
compared with two baseline matching polices, and identified
instances under which there is a substantial gap between the
maximum task arrival rates that can be supported by these
policies and that of the optimum back-pressure matching
policy.
R EFERENCES
[1] L. Tassiulas and A. Ephremides, “Stability properties of constrained
queueing systems and scheduling policies for maximum throughput in
multihop radio networks,” IEEE Transactions on Automatic Control,
vol. 37, pp. 1936–1948, 1992.
[2] V. Shah, L. Gulikers, L. Massoulié, and M. Vojnović, “Adaptive
matching for expert systems with uncertain task types,” CoRR, vol.
abs/1703.00674, 2017. [Online]. Available: http://arxiv.org/abs/1703.
00674
[3] D. Golovin, A. Krause, and D. Ray, “Near-optimal bayesian active
learning with noisy observations,” in NIPS’10, 2010, pp. 766–774.
[4] S. Javdani, Y. Chen, A. Karbasi, A. Krause, D. Bagnell, and S. S.
Srinivasa, “Near optimal bayesian active learning for decision making.”
in AISTATS’14, 2014, pp. 430–438.
[5] G. L. Nemhauser and L. A. Wolsey, Integer and combinatorial optimization. Wiley, 1988, vol. 18.
[6] M. Daltayanni, L. De Alfaro, and P. Papadimitriou, “Workerrank: Using
employer implicit judgements to infer worker reputation,” in WSDM,
2015.
[7] H.-Q. Ye, J. Ou, and X.-M. Yuan, “Stability of data networks: Stationary
and bursty models,” Operations Research, vol. 53, no. 1, pp. 107–125,
2005.
[8] P. Brémaud, Markov chains: Gibbs fields, Monte Carlo simulation, and
queues. Springer Science & Business Media, 2013, vol. 31.
[9] A. Kumar, “Discrete event stochastic processes: Lecture notes
for an engineering curriculum,” 2014. [Online]. Available: http:
//www.ece.iisc.ernet.in/∼anurag/books/anurag/spqt.pdf
[10] P. Robert, “Stochastic networks and queues, stochastic modelling and
applied probability series, vol. 52,” 2003.
[11] A. N. Rybko and A. L. Stolyar, “Ergodicity of stochastic processes describing the operation of open queueing networks,” Problemy Peredachi
Informatsii, vol. 28, no. 3, pp. 3–26, 1992.
[12] J. G. Dai, “On positive harris recurrence of multiclass queueing networks: A unified approach via fluid limit models,” The Annals of Applied
Probability, vol. 5, no. 1, pp. 49–77, 1995.
[13] L. Massoulié, “Structural properties of proportional fairness: Stability
and insensitivity,” Annals of Applied Probability, vol. 17, no. 3, pp.
809–839, 2007.
[14] Y. Chen and A. Krause, “Near-optimal batch mode active learning and
adaptive submodular optimization.” in ICML’13, 2013, pp. 160–168.
[15] K. Fujii and H. Kashima, “Budgeted stream-based active learning via
adaptive submodular maximization,” in NIPS’16, 2016, pp. 514–522.
[16] D. R. Karger, S. Oh, and D. Shah, “Budget-optimal task allocation for
reliable crowdsourcing systems,” Operations Research, vol. 62, no. 1,
pp. 1–24, 2014.
[17] N. B. Shah, D. Zhou, and Y. Peres, “Approval voting and incentives in
crowdsourcing,” in ICML’15, 2015, pp. 10–19.
[18] Y. Zhang, X. Chen, D. Zhou, and M. I. Jordan, “Spectral methods meet
em: A provably optimal algorithm for crowdsourcing,” J. Mach. Learn.
Res., vol. 17, no. 1, pp. 3537–3580, Jan. 2016.
[19] C. Gao, Y. Lu, and D. Zhou, “Exact exponent in optimal rates for
crowdsourcing,” in ICML’16, 2016, pp. 603–611.
[20] L. Massoulié and K. Xu, “On the capacity of information processing
systems,” in 29th Annual Conference on Learning Theory, New York,
United States, Jun. 2016.
[21] T. Lai and H. Robbins, “Asymptotically efficient adaptive allocation
rules,” Advances in Applied Mathematics, vol. 6, no. 1, pp. 4 – 22,
1985.
[22] P. Auer, N. Cesa-Bianchi, and P. Fischer, “Finite-time analysis of the
multiarmed bandit problem,” Machine Learning, vol. 47, no. 2, pp. 235–
256, 2002.
[23] J. Gittins, K. Glazebrook, and R. Weber, Multi-armed bandit allocation
indices. John Wiley & Sons, 2011.
[24] S. Agrawal and N. Goyal, “Analysis of thompson sampling for the
multi-armed bandit problem.” in Proceedings of the 25th Conference
on Learning Theory, 2012.
[25] J. Kleinberg and M. Sandler, “Convergent algorithms for collaborative
filtering,” in Proceedings of the 4th ACM Conference on Electronic
Commerce, 2003.
[26] ——, “Using mixture models for collaborative filtering,” in Proceedings
of the 36th annual ACM Symposium on Theory of Computing, 2004.
[27] D. H. Stern, R. Herbrich, and T. Graepel, “Matchbox: Large scale online
bayesian recommendations,” in Proceedings of the 18th International
Conference on World Wide Web, 2009.
[28] A. Mehta and D. Panigrahi, “Online matching with stochastic rewards,”
in IEEE 53rd Annual Symposium on Foundations of Computer Science,
Oct 2012, pp. 728–737.
[29] A. Mehta, B. Waggoner, and M. Zadimoghaddam, Online Stochastic
Matching with Unequal Probabilities, 2015, pp. 1388–1404.
[30] C.-J. Ho and J. W. Vaughan, “Online task assignment in crowdsourcing
markets,” in AAAI’12, 2012, pp. 45–51.
[31] M. J. Neely, E. Modiano, and C. E. Rohrs, “Dynamic power allocation
and routing for time varying wireless networks,” in Proc. IEEE INFOCOM, 2003.
[32] L. Georgiadis, M. J. Neely, and L. Tassiulas, “Resource allocation and
cross-layer control in wireless networks,” Foundations and Trends in
Networking, vol. 1, no. 1, pp. 1–144, 2006.
[33] L. Ying, S. Shakkottai, A. Reddy, and S. Liu, “On combining shortestpath and back-pressure routing over multihop wireless networks,”
IEEE/ACM Transactions on Networking, vol. 19, no. 3, pp. 841–854,
Jun. 2011.
[34] L. Bui, R. Srikant, and A. Stolyar, “A novel architecture for reduction of
delay and queueing structure complexity in the back-pressure algorithm,”
IEEE/ACM Transactions on Networking, vol. 19, no. 6, pp. 1597–1609,
Dec 2011.
[35] M. Alresaini, M. Sathiamoorthy, B. Krishnamachari, and M. Neely,
“Backpressure with adaptive redundancy (bwar),” in Proc. IEEE INFOCOM, March 2012, pp. 2300–2308.
[36] B. Ji, C. Joo, and N. Shroff, “Delay-based back-pressure scheduling in
multihop wireless networks,” IEEE/ACM Transactions on Networking,
vol. 21, no. 5, pp. 1539–1552, Oct 2013.
[37] Y. Cui, E. Yeh, and R. Liu, “Enhancing the delay performance of
dynamic backpressure algorithms,” IEEE/ACM Transactions on Networking, 2015.
[38] A. Mandelbaum and A. L. Stolyar, “Scheduling flexible servers with
convex delay costs: Heavy-traffic optimality of the generalized cµ-rule,”
Operations Research, vol. 52, no. 6, pp. 836–855, 2004.
[39] S. T. Maguluri, S. K. Burle, and R. Srikant, “Optimal heavy-traffic
queue length scaling in an incompletely saturated switch,” SIGMETRICS
Perform. Eval. Rev., vol. 44, no. 1, pp. 13–24, Jun. 2016.
[40] V. Shah, G. de Veciana, and G. Kesidis, “A stable approach for routing
queries in unstructured p2p networks,” IEEE/ACM Transactions on
Networking, vol. 24, no. 5, pp. 3136–3147, Oct 2016.
[41] K. Bimpikis and M. G. Markakis, “Learning and hierarchies in service
systems,” Unpublished manuscript, 2015.
[42] R. Johari, V. Kamble, and Y. Kanoria, “Matching while learning,”
CoRR, vol. abs/1603.04549, 2017. [Online]. Available: http://arxiv.org/
abs/1603.04549
[43] S. Asmussen, Applied probability and queues, 2nd ed. Springer Science
& Business Media, 2003.
VIII. P ROOFS
A. Proof of Theorem 1
Proof of Part (i): We first show that if λ < 4a/(2 + a) then the system is stable. For each t let t + τ (t) be the time at which
the first event (arrival or completion of a response) occurs after time t. Let τn = E[τ (t)|N (t) = n], i.e., given that N (t) = n
at time t, τn is the expected time at which the first event occurs after time t. For example, for n = 0 we have τn = 1/λ.
A common approach to show system stability is to use Lyapunov-Foster theorem, see e.g., Proposition I.5.3 on page 21 in
[43]. The idea is to construct a function L(·), called Lyapunov function, such that L(n) tends to infinity as |n| → ∞ and that
has a strictly negative ‘drift’ for all but finite values of n, i.e., there exists a constant > 0 such that
E L(N (t + τ (t))) − L(N (t)) N (t) = n ≤ −τn ,
for all but finite values of n. Intuitively, negative drift condition implies that as L(N (t)) becomes large (i.e., as N (t) becomes
large) the system dynamics is such that L(N (t)) tends to decrease in expectation. This prevents the L(t) from blowing up to ∞
as t increases and thus keeps the system stable. Roughly, the Lyapunov-Foster theorem states that the negative drift condition
is indeed sufficient to ensure that that system is positive recurrent and thus stable. We will use a variant of Lyapunov-Foster
theorem, provided below, which follows from Theorem 8.13 in [10].
Theorem 6: Consider an irreducible CTMC N (t) that takes values in a countable state-space. Let τ (t) and τn be as defined
above. If there exists a function L(·), and constants K > 0 and > 0 such that for L(n) > K we have
E L(N (t + τ (t))) − L(N (t)) N (t) = n ≤ −τn ,
and if {n : L(n) ≤ K} is finite, then N (t) is positive recurrent.
2−a
2−a
λ < a. Thus, there exists δ > 0 such (1+δ) 2(2−λ)
λ<
Now suppose that λ < 4a/(2+a). Then, it can be checked that 2(2−λ)
a. Now, consider the following candidate Lyapunov function: for each n, we have
1
2−a
nz0 + nz00 ,
L(n) = (1 + δ)
τn
2(2 − λ)
where δ is a constant obtained as above.
Let
D(n) = E L(N (t + τ (t))) − L(N (t)) N (t) = n
(6)
Consider a states n such that nz0 > 0. For these states, we obtain
1
2−a
D(n) = (1 + δ)
(λ − µs1 − µs2 ) + (µs1 ψs1 (z 0 ) + µs2 ψs2 (z 0 ))
τn
2(2 − λ)
2−a
1−a 1
2−a
= (1 + δ)
(λ − 2) +
+ = −δ
< 0.
2(2 − λ)
2
2
2
Now, consider states n such that nz0 = 0 and nz00 > 0. For these states we have
1
2−a
λ − µs1 a < 0.
D(n) = (1 + δ)
τn
2(2 − λ)
2−a
Thus, the conditions of Theorem 6 are satisfied with K = D((1, 1))/τ(1,1) and = min(δ(2 − a)/2), a − (1 + δ) 2(2−λ)
λ).
Hence, N (t) is positive recurrent if λ < 4a/(2 + a).
2−a
λ ≥ a. Thus, drift
We now show the only if part. Suppose that λ ≥ 4a/(2 + a). Then, there exists δ ≤ 0 such (1 + δ) 2(2−λ)
is non-negative for all but finite values of n. Further, since L(·) is bounded from below, and since the maximum change in
L(·) upon an arrival or a departure is bounded, using Proposition I.5.4 on page 22 in [43], we establish the only if part.
Proof of Part (ii): Part (ii) is a special case of Theorem 2.
Proof of Part(iii): We use Theorem 3 from Section IV in the paper. For Asymmetric(a) system we have Z = {z 0 , z 00 } where
0
zc = 21 for each c ∈ C, and zc00 = 1{c = c2 }. The flow conservation constraints in Theorem 3 can be given as follows:
X
X
X
X
λ=
νs,z0 , and
νs,z0 ψs (z 0 ) +
νs,z00 ψs (z 00 ) =
νs,z00
s
Suppose a ≥
1
2.
s
There exists an > 0 such that λ =
s
3a(1−)
a+1 .
νs2 ,z0 = 1 − , νs2 ,z00 = 0, νs1 ,z0 =
s
It can be checked that {νsz }s,z where
2−a
2a − 1
(1 − ), νs1 ,z00 =
(1 − )
a+1
a+1
and {δs }s∈S where δs = for each s satisfy sufficient conditions of Theorem 3.
Now suppose a < 12 . There exists an > 0 such that λ = 2a(1 − ). It can be checked that {νs,z }s,z where
νs2 ,z0 = 2a(1 − ), νs2 ,z00 = 0, νs1 ,z0 = 0, νs1 ,z00 = (1 − )
and {δs }s∈S where δs = for each s satisfies sufficient conditions of Theorem 3.
The result then follows from the proof of Theorem 3 by taking Y as Z
Proof of Part(iv): We again use Theorem 3. From the necessary conditions in Theorem 3, we have the following:
λ=
X
νs,z0 ,
(7)
s
1
(1 − a)
νs1 ,z0 + νs2 ,z0 + (1 − a)νs1 ,z00 = νs1 ,z00 ,
2
2
νs1 ,z0 + νs1 ,z00 ≤ 1,
νs2
,z 0
+ νs2
,z 00
≤ 1.
(8)
(9)
(10)
From (8) we get:
νs1 ,z00 =
1
a
(1 − a)
1
νs1 ,z0 + νs2 ,z0
2
2
.
By substituting in (9) this the above expression for νs1 ,z00 , we get
1
1 (1 − a)
0
0
0
νs1 ,z + νs2 ,z ≤ 1
νs1 ,z +
a
2
2
Upon simplifying, we get
νs1 ,z0
2a
≤
1+a
1
1 − νs2 ,z0 .
2a
(11)
Further, we need νs1 ,z0 to be non-negative. Thus, we need νs2 ,z0 ≤ 2a.
Substituting (11) in (7) we get
2a
1
λ ≤ max
1 − νs2 ,z0 , 0 + νs2 ,z0 .
1+a
2a
Suppose a ≥ 21 . Then, subject to (10) and νs2 ,z0 ≤ 2a, the right hand side of the above is maximized when νs2 ,z0 = 1 and
3a
νs2 ,z00 = 0. We thus obtain λ ≤ a+1
. Similarly, if a < 12 , then subject to (10) and νs2 ,z0 ≤ 2a, the right hand side of the above
3a
is maximized when νs2 ,z0 = 2a and νs2 ,z00 = 0, from which we obtain λ ≤ 2a. Thus, overall, we have λ ≤ min( a+1
, 2a).
Hence the result follows.
B. Proof of Theorem 2
Note that the system under random policy is equivalent to the one where pure-type of a task is revealed upon arrival, i.e.,
there is no uncertainty in task types. This is true since the random policy does not use the information of type (pure or mixed).
We thus let the pure-type of each task be revealed upon arrival. Let Xc (t) be the number of tasks in the system of pure-type
c. Let X(t) = {Xc (t)}c . For each c ∈ C, the arrival rate into queue Xc (t) is equal to
X
λc ,
λzc πz .
z∈Z
P
P λc
We first show the if part of the result. Suppose that we have
< 1. We use the fluid limit approach
c∈C
s∈S µs ps,c
developed in [11]–[13]. Roughly, given initial condition X(0) = x, the fluid trajectories of the state process X(t) can be
obtained by scaling initial conditions, speeding time, and then studying the rescaled process; i.e., letting limk→∞ k1 X(0) = x,
and studying limk→∞ k1 X(kt).
Using arguments similar to those used in [13], the fluid limits for the number of tasks in each class can be shown to satisfy
the following at almost all times t: for each c ∈ C and Xc > 0 we have
X
d
Xc
Xc = λc −
µs ps,c P
.
(12)
0
dt
c0 Xc
s∈S
Define a function L on R
C
as
L(X) =
X
c
Xc log
γc
X
Pc
c0
Xc0
,
(13)
where γc ,
P
λc
µs ps,c .
s∈S
d
Further, by following the arguments similar to [13], if we have that L(X) → ∞ and dt
L(X) ≤ − for all X such that
|X| = 1 under fluid limits then the stability of the original system follows. We show below that both these limits hold.
Using (12) and (13), we obtain
X d
d
X
Pc
L(X) =
Xc log
,
(14)
dt
dt
γc c0 Xc0
c
!
X
X
Xc
Xc
=
λc −
µs ps,c P
− log γc ,
(15)
log P
0
0
c0 Xc
c0 Xc
c
s∈S
!
X X
Xc
Xc
λc
P
−P
log P
− log γc ,
(16)
=
µs ps,c
0
0
s∈S µs ps,c
c0 Xc
c0 Xc
c
s
!
X X
=
µs ps,c (γc − Yc ) log(Yc /γc ),
(17)
c
s
P Xc
c0 Xc0
where Yc :=
. Now, (17) is negative and strictly bounded away from zero. This can be seen as follows. Firstly, all
terms in the sum are non-positive. Therefore,
P it suffices to show that there exists aPδ > 0 such that there always exists a c for
which (γc − Yc ) log(Yc /γc ) ≤ −δ. Since, c Yc = 1 and, for some fixed > 0, c γc = 1 − , it follows that there exists c
such that γc −P
Yc ≤ −. For this c, we thus also have Yc /γP
c ≥ 1 + . Consequently, (γc − Yc ) log(Yc /γc ) ≤ − log(1 + ).
Let θ = 1/ c γc and γ̂c = θγc for each c ∈ C. Since c γc < 1, we have θ > 1. Let D(p||q) be the Kullback-Leibler
divergence between two Bernoulli distributions with parameters p and q, i.e., D(p||q) = p log( pq ) + (1 − p) log( 1−p
1−q ). Now,
we can write
θXc
P
L(X) =
Xc log
γ̂c c0 Xc0
c
X
X
Xc
1
=
Xc log θ +
Xc log P
.
0
c0 Xc γ̂c
c
c
!
!
X
X
Xc
P
{γ̂c }c∈C
=
Xc log θ +
Xc D
c0 Xc0 c∈C
c
c
X
which converges to ∞ as |X| grows large.
Hence, the if part of the result follows.
P
We now show that the system is unstable if c∈C
limits. Consider the following function:
P
λc
µs ps,c
X
1
Xc .
µ
s s ps,c
P
c
Clearly, K(X) → ∞ as X → ∞. Define D(.) as in (6), but for K instead of L. Then, we have
X
1
λ
P c
D(0̄) =
,
τ0̄
s µs ps,c
c
and for X 6= 0̄, we have
X
X
1
Xc
1
P
D(X) =
λc −
µs ps,c P
0
τX
s µs ps,c
c0 Xc
c
s∈S
X
X
λ
X
P c
P c
=
−
0
µ
p
s s s,c
c0 Xc
c
c
X
λ
P c
=
− 1,
s µs ps,c
c
≥ 0.
(19)
(20)
≥ 1. We consider the original system instead of the fluid
s∈S
K(X) =
(18)
!
,
Thus, the drift is non-negative for all but finite number of states. Further, since K(X) is bounded from below, the maximum
change in K(X) upon an arrival or a departure is also bounded, using Proposition I.5.4 on page 22 in [43], we get the only
if part.
C. Proof of Theorem 3 and Theorem 4
We first show stability under sufficient conditions provided in the statement of Theorem 3. In the process, we prove Theorem 4.
In networked systems, e.g. see [1], [32], a standard approach towards proving stability of a backpressure type policy is to
design a ‘static’ policy using flow variables {νsz }s,z and the slacks {δs }s which provides a fixed service rate to each queue
Nz such that its drift is sufficiently negative for each. However, in our setup the total number of queues {Nz }z∈Z could be
countable, while the total available slack is finite. Thus, it is not possible to design a static policy such that the drift in each
individual queue is bounded from above by a negative constant. This is unlike any finite-server queueing system considered
in the previous literature.
We thus take a different approach, which can be explained roughly as follows. Since the total exogenous arrival rate λ, and
the total endogenous
arrival rate, i.e. arrival into a queue due to failure at another queue, are both finite (they are bounded
from
P
P
above by s µs ), there exists a finite set Y ⊂ Z such that the total arrival rate into Z\Y is less than minc∈C s∈S δ4s ps,c .
Each task which enters a queue Nz where z ∈ Z\Y is instead sent to a virtual queue X, and stays there until there is a
success. If X is ‘large’ compared to the other queues then all the servers focus on X. The finite number of remaining queues
are operated via a backpressure policy which accounts for the ‘expected backlog’ seen in these queues.
More formally, consider {νs,z }s,z and positive constants {δs }s as postulated in the theorem. Without loss of generality,
assume that there exists a constant 0 < < 1 such that δs = µs for each s ∈ S. Let Y be a finite subset of Z such that
X
X
X δs
X
λπz +
ps,c .
(21)
νs,z0 ψs (z 0 ) ≤ min
c∈C
4
−1
0
s∈S z ∈φs (z)∩Y
z∈Z\Y
P
s∈S
P
Since λ + s∈S,z∈Z νs,z ≤ 2 s µs , such a Y exists.
Let X be the number of tasks in the system which are or have been in past of type z ∈ Z\Y. Once a task enters queue
X it does not leave it until success. There may be tasks in it with mixed-type in Y. Note, our policy will depend on X and
thus {z(s, t)}s will not be N (t) measurable. In turn, N (t) will not be a CTMC. For z ∈ Y, let X̃z and Ñz be the tasks of
mixed-type z which have and have not had mixed-type in Z\Y. Also, for convenience
for each z ∈Z\Y, let X̃z be the tasks
of mixed-type z, i.e., Nz = X̃z for each z ∈ Z\Y. We now formally define σ {X̃z }z∈Z , {Ñz }z∈Y -measurable backpressure
policy. Thus, {Ñz }z∈Y , {X̃z }z∈Z is a CTMC.
We now show stability of the system under this policy for Backpressure(Y) as given in Definition 4. Below we will assume
that the ties in selecting z from Bs (Ñ , X) are broken uniformly at random for simplicity of exposition. The proof can be
easily extend to any other tie breaking approach. Consider the following Lyapunov function.
!2
X
X
X
L(Ñ , X̃) =
Ñz2 +
Xz
=
Ñz2 + X 2 .
z∈Y
z∈Z
z∈Y
For each t, let t + τ (t) be the time at which the first event (arrival or completion of a response) occurs after time t. Clearly,
τ (t) is a stopping time. Further, let τñ,x̃ (t) = E[τ (t)|(Ñ , X̃) = (ñ, x̃)].
Let
h
i
1
E L(Ñ (t + τ ), X̃(t + τ )) − L(Ñ (t), X̃(t)) Ñ (t) = ñ, X̃(t) = x̃ .
D(ñ, x̃) ,
τñ,x̃
D(ñ, x̃) is called drift in state n. We would like to show that there exists a positive integer K and positive constant such
that
D(ñ, x̃) ≤ − ∀(ñ, x̃) s.t. max(|ñ|∞ , x) ≥ K.
Let for each s ∈ S and z ∈ Y let
(
∗
νs,z
= 1 x min
c∈C
)
X
s
µs ps,c >
X
s
µs
max
z∈Y:ñz >0
ws,z (ñ, x) 1{z ∈ Bs (n)}
1
.
|Bs (n)|
Then, one can check that
1
τñ,x̃
E[Ñz (t + τ )2 − Ñz (t)2 Ñ (t) = ñ, X̃(t) = x̃] = (2ñz + 1) λπz +
X
X
s∈S z 0 ∈φ−1
s (z)∩Y
∗
0
νsz
+ (−2ñz + 1)
0 ψs (z )
X
s
∗
νs,z
.
Further, let
(
)
∗
ν = 1 x min
X
c∈C
X
µs ps,c >
s
µs
s
max
z∈Y:ñz >0
ws,z (ñ, x) .
Then, we have that
1
E[X(t + τ )2 − X(t)2 Ñ (t) = ñ, X̃(t) = x̃]
τñ,x̃
X
≤ (2x + 1)
λπz +
X
X
∗
0
νs,z
+ (−2x + 1)ν ∗ min
0 ψs (z )
c
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Z\Y
X
µs ps,c
s
Thus, we get
D(ñ, x̃) ≤
X
(2ñz + 1) λπz +
X
X
∗
0
νs,z
+ (−2ñz + 1)
0 ψs (z )
∗
µs νs,z
s
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Y
X
X
+ (2x + 1)
λπz +
X
X
0
∗
+ (−2x + 1)ν ∗ min
νs,z
0 ψs (z )
c
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Z\Y
X
µs ps,c .
s
Upon arranging terms, we obtain
D(ñ, x̃) ≤
X
2ñz λπz +
X
X
∗
0
νsz
0 ψs (z ) −
X
s
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Y
+ 2x
∗
νs,z
X
λπz +
z∈Z\Y
X
X
s∈S
z 0 ∈φ−1
s (z)∩Y
∗
0
νsz
− ν ∗ min
0 ψs (z )
X
c
µs ps,c
s
+ λ +
XX
X
zZ s∈S
z 0 ∈φ−1
s (z)∩Y
∗
0
νs,z
0 ψs (z ) +
XX
z∈Y
∗
νs,z
+ ν ∗ min
c
s
X
µs ps,c
s
P
The last of the above three terms can be bounded by a constant, say
and z ∈ Y let
P α1 = 10 s µs . For each s ∈ S P
∗
∗
∗
∗
. Further, let ν̂ ∗ = minc s (µs − 3δs /4)ps,c ν ∗ and ν̃ ∗ = minc s (δs /4)psc ν ∗ .
= (δs /4)νs,z
and ν̃sz
= (µs − 3δs /4)νs,z
ν̂s,z
Then,
X
X
X
X
∗
0
∗
ν̂sz
ν̂s,z
D(ñ, x̃) ≤
2ñz λπz +
0 ψs (z ) −
s
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Y
X
+ 2x
λπz +
X
X
∗
0
ν̂s,z
− ν̂ ∗ + α1
0 ψs (z )
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Z\Y
+
X
z∈Y
2ñz
X
X
∗
ν̃s,z
ψs (z 0 ) −
X
∗
ν̃sz
+ 2x
s
s∈S z 0 ∈φ−1
s (z)∩Y
X
X
∗
0
∗
ν̃s,z
0 ψs (z ) − ν̃
(22)
s∈S z 0 ∈φ−1
s (z)∩Y
Consider the following lemma. Its proof is given in Section VIII-D.
Lemma 1: Recall the {νs,z }s,z as postulated by the theorem. For Θ = {θs,z }s∈S,z∈Y ∪ {θ}, where θ and θs,z for each s, z
are reals, let
X
X
X
X
X
X
X
λπz +
f (Θ) =
2ñz λπz +
θsz0 ψs (z 0 ) −
θs,z + 2x
θsz0 ψs (z 0 ) − θ .
z∈Y
s
s∈S z 0 ∈φ−1
s (z)∩Y
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Z\Y
Then,
!
∗
}s∈S,z∈Y ∪ ν̂
f {ν̂s,z
∗
≤f
{νs,z }s∈S,z∈Y ∪ {min
c
X
s∈S
(δs /4)psc } .
From definition of νs,z , we get that the first term in f (Θ) for Θ = {νs,z }s∈S,z∈Y ∪ {minc
and, from (21) we have that the second termin it is less than or equal to 0.
∗
Thus, we have that f {ν̂s,z
}s∈S,z∈Y ∪ ν̂ ∗ ≤ 0. From (22) we in turn obtain
X
X
X
X
X
X
∗
∗
D(ñ, x̃) ≤ α1 +
2ñz
ν̃s,z
ψs (z 0 ) −
ν̃s,z
+ 2x
s
s∈S z 0 ∈φ−1
s (z)∩Y
z∈Y
P
s∈S (δs /4)psc }
is equal to 0,
∗
0
∗
.
ν̃s,z
0 ψs (z ) − ν̃
s∈S z 0 ∈φ−1
s (z)∩Y
Fix > 0. We now show that there exist a positive integer K such that if x > K or if |ñ|∞ > K then D(ñ, x̃) ≤ −. Upon
rearranging terms, we obtain
D(ñ, x̃) ≤ α1 − 2
X
X
∗
ν̃s,z
(ñz − ψs (z)ñφs (z) ) − 2
s∈S z∈Y:φs (z)∈Y
= α1 − 2
XX
X
X
∗
ν̃s,z
(ñz − ψs (z)X) − 2xν̃ ∗
s∈S z∈Y:φs (z)∈Z\Y
∗
ν̃s,z
ws,z (ñ, x)
∗
− 2ν̃ x,
s∈S z∈Y
!
= α1 − max 2
XX
∗
ν̃s,z
ws,z (ñ, x), 2ν̃ ∗ x
s∈S z∈Y
Thus we get,
D(ñ, x̃) ≤ α1 − x min
c∈C
X δs
s∈S
4
ps,c .
P
Hence, for any (ñ, x) such that x > (α1 + ) minc∈C s∈S δ4s ps,c , we have D(ñ, x̃) ≤ −.
We also have that
X δs
D(ñ, x̃) ≤ α1 − 2
max ws,z (ñ, x).
4 z∈Y
s∈S
Thus,
X
δs
δs X
max ws,z (ñ, x) ≤ α1 − 2 min
max
ws,z (ñ, x).
D(ñ, x̃) ≤ α1 − 2 min
z∈Y
s∈S 4
z∈Y
s∈S 4
s∈S
s∈S
P
Now suppose that x ≤ α2 , (α1 + ) minc∈C s∈S δ4s ps,c . Then, if we show that maxz∈Y s∈S ws,z (ñ, x) → ∞ as
0
0
|ñ|∞ → ∞,
P then we have that D(ñ, x̃) ≤ − a positive integer K such that |ñ|∞ > K . We now show that, under x ≤ α2 ,
we have s∈S maxz∈Y ws,z (ñ, x) → ∞ as |ñ|∞ → ∞.
Let z ∗ ∈ arg maxz∈Y ñz . Then we have
X
X
ws,z∗ (ñ, x) ≤
(ñz∗ − ψs (z ∗ ) max(α2 , ñz∗ )
P
s
s∈S
= |S|ñz∗ − max(α2 , ñz∗ )
X
ψs (z)
s
which tends to infinity because
X
s
ψs (z ∗ ) =
XX
s
zc∗ (1 − ps,c ) = |S| −
XX
c
s
≤ |S| − max
c
≤ |S| −
X
zc∗ ps,c
c
zc∗ ps,c ≤ |S| − max zc∗ min
s
c
c
X
ps,c
s
X
1
min
ps,c < |S|.
|C| c s
Thus, there exist positive constants K and such that if x > K or if |ñ|∞ > K then D(ñ, x̃) ≤ −.
Let A , {(ñ, x̃) : max(|ñ|∞ , x) ≤ K}. Then, using a version of Lyapunav-Foster Theorem, namely Theorem 8.6 in [10],
we have that, from any state (ñ, x̃) such that |ñ| + x < ∞, the expected time to return to A, i.e., τA (ñ, x̃) is finite. Further,
T ,
sup τA (ñ, x̃) < ∞.
(ñ,x̃)∈A
Thus, starting with any state in A, we return to A in a finite expected time. We will be done if we show that expected time
to return to state (0, 0) is also finite. We do this as follows. Fix a constant β > 0. Since there exists s such that minc ps,c > 0,
we have that for any interval of time of size β the probability that no arrival happens in this interval and that a task leaves
the system is finite.
Suppose that system is in a state (ñ, x̃) ∈ A at time t = 0. Now consider renewal times Ti , i = 0, 1, 2, . . . ,, where T0 = 0
and for each i > 0, Ti is defined as follows: Ti is equal to Ti−1 + β if indeed no arrival happens and a task leaves the system
in the interval [Ti−1 , Ti−1 + β), else Ti is the first time of return to A after Ti−1 . Clearly E[Ti ] since T as defined above is
finite. Further probability that a task leaves system in time Ti − Ti−1 is finite, say α. Thus, time for system emptying after
first reaching A can be upper-bounded by sum of K geometric random variables with rate α. Thus expected time to return to
state (0, 0) is finite. Hence, the system is stable.
Now suppose that the system is stable. Then, the necessary conditions can be shown to hold by the ergodicity of the system,
and letting νs,z for each s, z to be the long-term fraction of times a server s attempts a task in Nz .
D. Proof of Lemma 1
Upon rearrangement of terms in the expression of f (Θ) we obtain
X
X
X
f (Θ)/2 = −
θs,z (nz − ψs (z 0 )nφs (z) ) −
s
s
z∈Y:φs (z)∈Y
X
θs,z (nz − ψs (z 0 )x) − xθ.
z∈Y:φs (z)∈Z\Y
By using the definition of weights ws,z , we obtain
f (Θ)/2 = −
XX
s
θs,z ws,z (ñ, x) − xθ ≥ −
X
s
z∈Y
X
max ws,z (ñ, x)
θs,z − xθ.
z∈Y
z∈Y
Thus,
f ({νsz }s∈S,z∈Y ∪ {min
c
X
(δs /4)psc })/2
s∈S
X
X
X
≥−
max ws,z (ñ, x)
νs,z − x min
(δs /4)psc
z∈Y
s
z∈Y
c
s∈S
X
X
≥−
(µs − δs /2) max wsz (ñ, x) − x min
(δs /4)ps,c
z∈Y
s
≥ −1
(
X
s
c
s∈S
)
max ws,z (ñ, x)(µs − 3δs /4) ≥ x(min
X
c
z∈Y
s
(
−1
X
s
max ws,z (ñ, x)(µs − 3δs /4)
z∈Y
)
X
s
=f
(µs − 3δs /4)ps,c )
max ws,z (ñ, x)(µs − 3δs /4) < x(min
c
z∈Y
∗
({ν̂sz
}s∈S,z∈Y
X
X
(µs − 3δs /4)ps,c ) x min
(µs − 3δs /4)ps,c
s
c
s
∗
∪ ν̂ ) /2.
Hence, the lemma holds.
E. Proof of Theorem 5
Suppose that the sufficient conditions as given in Theorem 3 are satisfied. Then, in the proof of Theorem 3 we showed
existence of a policy such that the system is ergodic. In fact, since we have a strict slack δs > 0 for capacity constraint at
each server, using proof of Theorem 3 we can design a policy for a system which achieves stability even when the server
capacities are modified as µ0s = µs − R, where 0 < R < mins δs . Under such a policy, for each s ∈ S, z ∈ Z, t ≥ 1, let
µs,z (t) represent the long-term fraction of times a server s attempts a task in Nz which has been attempted t − 1 times in the
past. Then, the following hold.
P
i)
λ
Pz = s µs,z (1),−
P
ii)
µ s− (t)ψs (zs ) = s µs,z (t + 1),
Ps s,z
P
(23)
iii)
µs,z (t) ≤ µs ,
t≥1P z∈ZP
iv) ∃s0 : t≥1 z∈Z µs0 ,z (t) = µs0 − R, R > 0.
The inequalities in (23) can be strengthened to achieve positive slack for each server’s capacity, but (23) as mentioned is
sufficient for our purposes. Using existence of (µs,z (t) : s ∈ S, z ∈ Z) which solves
P (23), we now show that, for Backpressure()
policy, provided > 0 has been chosen small enough, the function L(n)P
:= i n(Ai )2 is a Lyapunov function in the sense
that its drift is negative, bounded away from 0 except for states n with z nz ≤ N for some threshold N . This will imply
the announced result by the same arguments as in the proof of Theorem 4.
Let n = {nz } be given. For each Ai such that n(Ai ) > 0, we pick arbitrarily one point zi in Ai such that nzi > 0. We
then define the projection operator P (z) which maps z to zi if z ∈ Ai . For z ∈ Ai such that n(Ai ) = 0 we say that P (z) is
undefined. We shall also consider for each z ∈ Z the operator
Pst (z) := P (φs (Pst−1 (z))).
(24)
This is defined so long as all the involved projections are defined, i.e. the constructed sequence only visits sets Ai with
n(Ai ) > 0. We also let φts denote the application resulting from t applications of φs .
We now define for each s, z, t, zi the following rates:
νs,zi (1; z)
νs,zi (t; z)
:=
µs,z (1)1IP (z)=zi ,
µs,φts (z) (t)1IPst (z)=zi
:=
0
if Pst (z) is defined,
otherwise.
(25)
Finally, we define the following rates for all t ≤ T , where 0 and β are constants to be specified shortly:
P
rs,zi (1) := Pz∈Z νs,zi (1; z),
rs,zi (t) := z∈Z νs,zi (t; z) + 0 (1 + β t )1Is=s0 1In(Ai )>0 .
(26)
We extend the definition of the rates rs,zi (t) for t > T by induction as follows. First, for s 6= s0 we let rs,zi (t) = 0. For
server s0 , we let
X
X
νs,zi (T + 1; z) + 0 (1 + β T +1 )
(27)
rs0 ,zi (T + 1) :=
s
z:PsT +1 (z)=zi
and for t > T :
X
rs0 ,zi (t + 1) =
(1 − α)rs0 ,zj (t).
(28)
j:Ps1 (zj )=zi
0
The functions ψs are all Lipschitz-continuous. Under the assumption (5), it is easily verified that the functions φs are also
Lipschitz-continuous. Let K be such that all these functions are K-Lipschitz-continuous.
It is readily established by induction on t that for all s, so long as Pst (z) is defined, one has
|Pst (z) − φts (z)| ≤
Kt − 1
.
K −1
(29)
Indeed, one has
|Pst (z) − φts (z)| = |P (φs (Pst−1 (z))) − φts (z)|
≤ |P (φs (Pst−1 (z))) − φs (Pst−1 (z))| + |φs (Pst−1 (z)) − φts (z)|
≤ + K|Pst−1 (z) − φt−1
s (z)|,
and (29) follows by induction.
We now exploit these properties to show that for suitable choices of β, 0 , the previously defined rates rs,zi (t) verify the
following inequalities for all i such that n(Ai ) > 0 and thus zi is defined:
P
P
Pz∈A
Pi λz ≤ s rs,zi (1),
P
(30)
s
zj :Ps1 (zj )=zi rs,zj (t)ψs (zj ) ≤
s rs,zi (t + 1)
The first equation in (30) reads, in view of (25), (26):
X
XX
λz ≤
µs,z (1),
z∈Ai
z∈Ai
s
which holds with equality by (23) i).
The left-hand side of the second equation in (30) reads for t ≤ T :
X
X
X
µs,φts (z) (t)ψs (zj ) + 0 (1 + β t ).
s
zj :Ps1 (zj )=zi z:Pst (z)=zj
Using the Lipschitz property of ψ, the bound (29) established between Pst (z) and φts (z), and letting Λ :=
larger than
X
X
X
Kt − 1
µs,φts (z) (t)ψs (φts (z)) + 0 (1 + β t ) + Λ(1 − α)t K
.
K −1
1
t
s
zj :Ps (zj )=zi z:Ps (z)=zj
P
z
λz , this is no
Indeed, the sum
P
z
µs,z (t) of rates at step t is at most Λ(1 − α)t . This last expression can be rearranged to give
X
X
s
z:Pst+1 (z)=zi
µs,φts (z) (t)ψs (φts (z)) + 0 (1 + β t ) + Λ(1 − α)t K
Kt − 1
.
K −1
In view of (23) ii), the first summation is equal to
X
X
X
µs,φt+1
rs,zi (t + 1) − 0 (1 + β t+1 ).
(z) =
s
s
s
z:Pst+1 (z)=zi
The difference between the right-hand side and the left-hand side of the second equation in (30) is therefore lower-bounded
by
Kt − 1
Kt − 1
0 (1 + β t+1 ) − 0 (1 + β t ) − Λ(1 − α)t K
= 0 β t (β − 1) − Λ(1 − α)t K
.
K −1
K −1
Assuming K ≥ 2, β = K + 1, and 0 = Λ, this difference is at least
Λ(K + 1)t K − ΛK(K t − 1) ≥ KΛ[(K + 1)t − K t + 1] ≥ KΛ.
Letting δ := KΛ, we have in fact shown a strengthening of the second equation in (30), namely:
X
X
X
rs,zj (t)ψs (zj ) ≤ δ +
rs,zi (t + 1)
t ∈ 2, . . . , T and n(Ai ) > 0 ⇒
s
(31)
s
zj :Ps1 (zj )=zi
Consider now t ≥ T + 1. The left-hand side of the second equation in (30) verifies
X
X
rs0 ,zj (t)ψs (zj ) ≤
rs0 ,zj (t)(1 − α),
zj :Ps1 (zj )=zi
zj :Ps1 (zj )=zi
0
0
by the lower-bound of α on the psc . This implies that the announced inequality also holds for t > T .
We now verify that, provided was chosen small enough, the constructed rates rs,zi (t) satisfy the capacity constraints of
the servers. For s 6= s0 , this is easily verified, as by (23) iii),
XX
XX
rs,zi (t) ≤
µs,z (t) ≤ µs .
t
z
i
t
Consider now server s0 . We then have
XX
t
rs0 ,zi (t) ≤
XX
z
i
µs0 ,z (t) +
t
T
+1
X
0 (1 + β t ) +
t=1
Λ(1 − α)T +1 + 0 (1 + β T +1 )
·
α
Thus by (23) iv) this meets the capacity constraint of server s0 provided
T
+1
X
0 (1 + β t ) +
t=1
Λ(1 − α)T +1 + 0 (1 + β T +1 )
≤ R.
α
This can clearly be achieved by first choosing T such that Λ(1 − α)T +1 ≤ Rα/2, and then such that
T
+1
X
0 (1 + β t ) +
t=1
0 (1 + β T +1 )
≤ R/2.
α
It finally remains to prove that thePFoster-Lyapunov stability criterion holds for our proposed backpressure policy. Assume that
each server s dedicates capacity t≥1 rs,zi (t) to jobs of type zi . This does not exceed servers’ capacities as we just showed.
Moreover, in view of (30) and (31), under this allocation the drift of any n(Ai ) such that n(Ai ) > 0 reads
X
XX
XX
X
λz −
rs,zi (t) +
rs,zj (t)ψs (zj ) ≤ −T δ1In(Ai )>0 .
z∈Ai
t≥1
s
t≥1
s
j:Ps1 (zj )=zi
For an arbitrary policy, let µi denote the service rate it devotes to those z in Ai , and λ0i denote the overall arrival rate of jobs
with type
Pz in Ai whether from external arrivals or unsuccessful treatments. The drift for our candidate Lyapunov function
L(n) = i n(Ai )2 then reads
X
X
X
(λ0i + µi ) + 2n(Ai )[λ0i − µi ] ≤ Λ + 2
µs + 2
n(Ai )[λ0i − µi ],
i
s
i
where we used the fact that the overall arrival rate cannot be larger than the exogeneous arrival rate plus the overall service
rate.
P
P
Under the allocations t≥1 rs,zi (t) we just considered, the summation in the right-hand side is at most −2δT i n(Ai ).
Since the backpressure policy we have introduced
this summation among all feasible policies, it guarantees a drift
P minimizes P
for the Lyapunov function L of at most P
Λ + 2 s µs − 2δT Pi n(Ai ). We can therefore rely on Foster’s criterion to deduce
that the return time to the set A = {n : i n(Ai ) ≤ (Λ + 2 s µs )/(δT )} has bounded expectation. We will be done if we
show that the system empties infinitely often. For this, we use the argument similar to that used in Theorem 4.
Fix a constant β > 0. Since α > 0, we have that for any interval of time of size β the probability that no arrival happens
in this interval and that a task leaves the system is finite.
Suppose that system is in a state n ∈ A at time t = 0. Now consider renewal times Ti , i = 0, 1, 2, . . . ,, where T0 = 0 and
for each i > 0, Ti is defined as follows: Ti is equal to Ti−1 + β if indeed no arrival happens and a task leaves the system
in the interval [Ti−1 , Ti−1 + β), else Ti is the first time of return to A after Ti−1 . Clearly E[Ti ] since T as defined above is
finite. Further probability that a task leaves system in time Ti − Ti−1 is finite, say α̃. Thus, time for system emptying after
first reaching A can be upper-bounded by sum of K geometric random variables with rate α̃. Thus expected time to return to
state 0 is finite. Hence, the system is stable.
| 2cs.AI
|
arXiv:1701.07402v2 [math-ph] 29 Jul 2017
Smallest eigenvalue density for regular or
fixed-trace complex Wishart-Laguerre ensemble and
entanglement in coupled kicked tops
Santosh Kumar & Bharath Sambasivam & Shashank Anand
Department of Physics, Shiv Nadar University, Gautam Buddha Nagar, Uttar
Pradesh – 201314, India
E-mail: [email protected]
Abstract. The statistical behaviour of the smallest eigenvalue has important
implications for systems which can be modeled using a Wishart-Laguerre ensemble, the
regular one or the fixed trace one. For example, the density of the smallest eigenvalue of
the Wishart-Laguerre ensemble plays a crucial role in characterizing multiple channel
telecommunication systems. Similarly, in the quantum entanglement problem, the
smallest eigenvalue of the fixed trace ensemble carries information regarding the nature
of entanglement.
For real Wishart-Laguerre matrices, there exists an elegant recurrence scheme
suggested by Edelman to directly obtain the exact expression for the smallest
eigenvalue density. In the case of complex Wishart-Laguerre matrices, for finding exact
and explicit expressions for the smallest eigenvalue density, existing results based on
determinants become impractical when the determinants involve large-size matrices.
In this work, we derive a recurrence scheme for the complex case which is analogous to
that of Edelman’s for the real case. This is used to obtain exact results for the smallest
eigenvalue density for both the regular, and the fixed trace complex Wishart-Laguerre
ensembles. We validate our analytical results using Monte Carlo simulations. We also
study scaled Wishart-Laguerre ensemble and investigate its efficacy in approximating
the fixed-trace ensemble. Eventually, we apply our result for the fixed-trace ensemble
to investigate the behaviour of the smallest eigenvalue in the paradigmatic system of
coupled kicked tops.
1. Introduction
Wishart-Laguerre ensembles constitute an important class of random matrix
ensembles [1, 2] and have found diverse applications in the field of multivariate
statistics [3–5], problems related to time series [6–8], low energy quantum
chromodynamics [9, 10], multiple-channel telecommunication [11–13], quantum
entanglement [14–20], etc. Often the smallest eigenvalue distribution plays a crucial role
in investigating the behaviour of the system under study. For instance, in the context
of multiple input multiple output (MIMO) communication, the smallest eigenvalue of
Wishart-Laguerre ensemble determines the minimum distance between the received
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
2
vectors [21], and also the lower bound on the channel capacity that eventually is used
in antenna selection techniques [22]. Similarly, the smallest eigenvalue density of fixed
trace Wishart-Laguerre ensemble serves as an important metric for characterizing the
entanglement in bipartite systems [23, 24].
Matrices governed by Wishart distribution are parametrized by their size (n) and
the degree of freedom (m) [3–5]; see section 2. In this paper we use the term regular
to mean unrestricted trace Wishart matrices with m ≥ n. The smallest eigenvalue
of Wishart matrices have been studied extensively, both for regular, and fixed trace
scenarios. Moreover, finite-dimension, as well as large-dimension asymptotic cases have
been explored. Our focus here is on the finite-dimensional scenario with the primary
objective to obtain explicit expressions for the smallest eigenvalue density.
In the case of regular Wishart-Laguerre ensemble, for real matrices at finite
n, m, Edelman, among other things, has provided a recursion-based scheme to obtain
the smallest eigenvalue density [25, 26]. For complex matrices, the result for the
cumulative distribution of the smallest eigenvalue goes back to Khatri, who worked
out a determinant-based expression [27]. Forrester and Hughes have given closed
expressions for the density of the smallest and second-smallest eigenvalues in terms
of Wronskian and Toeplitz determinants [28]. Further generalizations applicable to a
wider symmetry class of Wishart matrices have been considered in [29–31]. Damgaard
and Nishigaki have derived the probability distribution of the kth smallest eigenvalue of
Dirac operator in the microscopic scaling limit for real, complex as well as quaternion
cases and demonstrated the universality of the results [32]. These eigenvalues have
direct relationship with those of the Wishart-Laguerre ensemble. In [33] Akemann et
al. have further explored the smallest eigenvalue distribution of real Wishart-Laguerre
matrices and validated the universality in the microscopic limit for the correlated case
also. Moreover, in a very recent work by Edelman, Guionnet, and Péché [34], the
behaviour of the smallest eigenvalue coming from finite random matrices (including
Wishart) based on non-Gaussian entries has been investigated.
For the fixed trace case, Chen, Liu and Zhou [35] have derived the smallest
eigenvalue density in terms of sum of Jack polynomials. Moreover, for the complex case
this expression has been simplified to inverse Laplace transform of a certain determinant.
In [36], for the real Wishart matrices, Edelman’s recursive approach for the regular
Wishart-Laguerre ensemble has been used by Akemann and Vivo to obtain the smallest
eigenvalue density for the fixed trace Wishart-Laguerre ensemble.
For the complex case, the exact result for the smallest eigenvalue density is
available in terms of determinants involving n-dimensional [21, 27, 37] or α-dimensional
matrices [28, 35], where α = m − n is the rectangularity. These results have been used
for asymptotic analysis when n → ∞ for α fixed and these correspond to eigenvalue
distributions comprising Bessel kernel [28,32,38]. On the other hand, if both n, α → ∞,
an analysis involving Fredholm determinant with Airy kernel is possible and that leads
to the celebrated Tracy-Widom (TW) distribution [39–41]. In [42] the transition regime
between the Bessel and Airy densities has also been investigated. While these asymptotic
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
3
results give a wealth of information regarding the universal aspects, if one desires
explicit expressions for the smallest eigenvalue density for large but finite n, α then
these determinant based results turn out to be impractical, even with the availability
of modern computational packages. We should remark that the smallest eigenvalue
density has also been worked out for correlated Wishart-Laguerre ensembles, both for
real and complex cases [37, 43, 44]. However, the exact results are, again, in terms of
determinants or Pfaffians.
The iterative scheme provided by Edelman [25, 26] is quite an effective and
convenient way to calculate the density for the case of real matrices, and one can handle
large values of dimensionality n and rectangularity α. For the complex Wishart-Laguerre
ensemble, Forrester and Hughes have worked out an iterative scheme for the cumulative
distribution of the smallest eigenvalue. However, to the best of our knowledge, an
iterative scheme analogous to that of Edelman’s, for direct evaluation of the probability
density of the smallest eigenvalue has hitherto remained unavailable. In the present
work, we derive the recurrence scheme that applies to the complex Wishart-Laguerre
ensemble. These results involve an ‘exponential times polynomial’ structure. Since the
fixed trace ensemble is related to the regular Wishart-Laguerre ensemble via a Laplace
transform, the structure of the smallest eigenvalue density in the latter leads to a very
convenient calculation of density in the former case as well [36]. Moreover, arbitrary
moments of the smallest eigenvalue are also readily obtained. In addition, for the regular
Wishart-Laguerre ensemble we also indicate a relation between the recurrence relation
and the determinantal results of Forrester and Hughes [28], and explicitly demonstrate
the equivalence between the two results for rectangularity α = 0, 1. Similarly, for
the fixed-trace scenario we prove the equivalence of the recursion-based expression and
the result of Chen, Liu and Zhou [35] based on the inverse Laplace transform of a
determinant, again for α = 0, 1.
Finally, we use the smallest eigenvalue density of the fixed trace ensemble to study
entanglement formation in the paradigmatic system of coupled kicked tops. We should
note that although the Floquet operator involved in this system belongs to the circular
orthogonal ensemble (COE) [1,2], the dynamically generated states are not random real
vectors in the Schmidt basis [45]. Rather, they are complex, and hence, the results for
the complex fixed trace Wishart-Laguerre ensemble are applicable.
2. Wishart-Laguerre ensemble
Consider complex matrices A of dimensions n × m taken from the density PA (A) ∝
exp − tr AA† . We assume without loss of generality that n ≤ m. Then, the
non-negative definite matrices W = AA† constitute the (regular) Wishart-Laguerre
ensemble with the probability density
PW (W) ∝ (det W)m−n exp (− tr W) .
(1)
The joint probability density of unordered eigenvalues (λj ≥ 0, j = 1, ..., n) of W is
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
4
given by [1, 2] ‡
P (λ1 , ..., λn ) =
Cn,α ∆2n ({λ})
n
Y
λαj e−λj ,
(2)
j=1
with α = m − n, and
−1
Cn,α
=
n
Y
Γ(j + 1)Γ(j + α).
(3)
j=1
Q
Also, ∆n ({λ}) = 1≤k<j≤n (λj − λk ) is the Vandermonde determinant. For this classical
ensemble, exact expression for correlation functions of all orders are known [1, 2]. For
example, the first order marginal density (one-level density),
Z ∞
Z ∞
p(λ) =
dλ2 · · ·
dλn P (λ, λ2 , ..., λn ),
(4)
0
0
is given by
n−1
1 −λ α X Γ(j + 1) (α) 2
Lj (λ)
p(λ) = e λ
n
Γ(j + α + 1)
j=0
=
Γ(n) −λ α (α)
(α+1)
e λ [Ln−1 (λ)L(α+1)
(λ) − L(α)
n
n (λ)Ln−1 (λ)].
Γ(m)
(5)
(γ)
Here Li (λ) represents the associated Laguerre polynomials [46].
We now focus on the statistics of the smallest eigenvalue of W. The probability
density for the smallest eigenvalue can be calculated using the joint probability
density (2) as [25, 26, 28]
Z ∞
Z ∞
dλn P (x, λ2 , ..., λn ).
(6)
dλ2 · · ·
f (x) = n
x
x
As shown in Appendix A, this can be brought down to the form
f (x) = cn,m e−nx xα gn,m (x).
(7)
Here the normalization factor cn,m is given by
n−1
cn,m
Y Γ(i + 2)
1
=
,
Γ(n)Γ(m) i=1 Γ(i + α)
(8)
and gn,m (x) is obtained using the following recurrence scheme:
I. Set S0 = gn,m−1 (x), S−1 = 0
II. Iterate the following for i = 1 to n − 1 :
x dSi−1
Si = (x + m − i + 1)Si−1 −
n − i dx
m−i
+ x (i − 1)
Si−2
n−i
III. Obtain gn,m (x) = Sn−1
‡ We should note that m × m–dimensional matrices A† A share the eigenvalues λ1 , ..., λn of AA† . The
other m − n eigenvalues are all zeros. The joint density in this case can be written by introducing deltafunctions for these zero-eigenvalues and implementing proper symmetrization among all eigenvalues.
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
5
The initial case (m = n) is given by gn,n (x) = 1. Thus, starting from the square case,
for which the result is simple (f (x) = ne−nx ), we can go up to any desired rectangularity
by repeated application of the above recurrence scheme. We note that (7) is of the form
f (x) =
αn+1
X
hj xj−1 e−nx ,
(9)
j=α+1
where hj are n, m dependent rational coefficients. The lower and upper limits of the
summation in (9) are α + 1 and αn + 1, respectively. This is because the recurrence
scheme applied for rectangularity α gives gn,m (x) as a polynomial of degree α(n − 1),
and there is the factor xα in (7). The coefficients hj can be extracted once the above
recursion has been applied.
The above simple structure for the probability density gives easy access to the η–th
moment of the smallest eigenvalue of the Wishart-Laguerre ensemble. We obtain
Z ∞
αn+1
X hj
η
η
Γ(j + η).
(10)
hx i =
x f (x) dx =
nj+η
0
j=α+1
We would like to remark that this relationship holds not only for non-negative integer
values of η, but also for any complex η such that Re(η) > −α − 1.
In Appendix B we provide simple Mathematica [47] codes that produce exact results
for the density as well as the η–th moment for any desired value of n, m by implementing
the above results.
In figure 1 we consider the smallest eigenvalue density and compare the analytical
results with Monte Carlo simulations using 105 matrices for n = 8, 15, and several α
values. We find excellent agreement in all cases.
8
1.4
(a)
1.2
α=0
6
� (�)
α=1
4
α=2
α=3
2
(b)
1.0
α=5
0.8
α = 10
0.6
α = 15
0.4
0.2
0
0.0
0.5
1.0
1.5
2.0
0.0
0
�
1
2
3
4
5
6
7
Probability density of the smallest eigenvalue for the WishartLaguerre ensemble with (a) n = 8, (b) n = 15, and α values as indicated. The
solid lines are analytical predictions based on (7), while the symbols (filledcircles, squares, triangles) represent results of Monte Carlo simulations.
Figure 1.
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
Forrester and Hughes’ result for the smallest eigenvalue density reads [28]
i
Γ(n + 1) −nx α h α+j−k−1 (3−α)
e x det Dt
Lm−2 (t)|t=−x
,
f (x) = (−1)α(α−1)/2
Γ(m)
j,k=1,...,α
where Dt ≡ d/dt. Comparing this result with (7), we find that
h
i
α+j−k−1 (3−α)
α(α−1)/2 Γ(n + 1)
det Dt
Lm−2 (t)|t=−x
.
gn,m (x) = (−1)
Γ(m) cn,m
j,k=1,...,α
6
(11)
(12)
Therefore, the recurrence scheme essentially leads to the evaluation of the above
determinant, which otherwise becomes difficult to evaluate directly for large α values.
Demonstrating the equality of the two sides in (12) directly seems challenging for
arbitrary n, m, if at all feasible. However, as shown below, for α = 1, we find that gn,m (x)
does lead to the associated Laguerre polynomial as evaluated by the determinantal
expression. Before analyzing the results of α = 1, we also consider α = 0, which
corresponds to the square case.
2.1. Results for α = 0
In this case gnm = 1 and in the expression (9), there is just one term in the sum, viz.
j = 1. The corresponding value of the coefficient hj is n. Thus, the smallest eigenvalue
density reads
f (x) = ne−nx .
(13)
Also, the moment-expression is given by
Γ(η + 1)
.
(14)
nη
These expressions agree with those derived in [25,28], as they should. We note that (11)
leads to the correct density, as the determinant part has to be taken as 1 for α = 0.
hxη i =
2.2. Results for α = 1
This is a nontrivial case. As shown in Appendix C, in this case, Si in the
(x). Consequently, gn,n+1 (x) =
recurrence relation can be identified with Γ(i + 1)Ln−i+1
i
(2)
Γ(n)Ln−1 (−x). Also cn,n+1 = 1/Γ(n), which leads to the smallest eigenvalue expression
(2)
f (x) = e−nx xLn−1 (−x).
(15)
This agrees with (11) when evaluated for α = 1. The use of the expansion of the
Laguerre polynomial [46] leads to the coefficient hj in (9) as
hj =
Γ(n + 2)
,
Γ(n − j + 2)Γ(j + 1)Γ(j − 1)
j = 2, 3, ..., n + 1.
(16)
The η–th moment follows as
η
hx i = Γ(n + 2)
n+1
X
j=2
nj+η Γ(n
Γ(j + η)
.
− j + 2)Γ(j + 1)Γ(j − 1)
(17)
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
7
3. Fixed trace Wishart-Laguerre ensemble
Fixed trace ensembles constitute a special class of random matrices and can take care
of system dependent constraints. For the Wishart-Laguerre case, the corresponding
fixed trace ensemble arises naturally in the quantum entanglement problem in bipartite
systems [15, 17, 18]. With the trace value fixed at unity, it models the reduced density
matrix; see section 5. Using the Wishart matrices W from the preceding section, the
fixed trace ensemble can be realized by considering the matrices F = W/ tr W [18, 20].
The corresponding probability density is
PF (F) ∝ (det F)α δ(tr F − 1).
(18)
The joint density of unordered eigenvalues (0 ≤ µj ≤ 1; j = 1, ..., n) of F is obtained
as [15, 17, 18]
!
n
n
X
Y
F
2
PF (µ1 , ..., µn ) = Cn,α δ
µi − 1 ∆n ({µ})
µαj ,
(19)
i=1
j=1
F
Cn,α
= Γ(nm) Cn,α [48]. The corresponding marginal density has been derived
where
in [49] as a single sum over hypergeometric 5 F4 , and as a double sum over polynomials
in [50]. In [48] it has been given as a single sum over the Gauss hypergeometric function
(2 F1 ):
n−1
X
pF (µ) =
Ki µi+α (1 − µ)−i+nm−α−2
×
i=0
1−n,i−nm+α+1
n Fα+1
−n,i−nm+α+1
− (n − i − 1)Fα+1
.
(20)
µ
)/Γ(c). Also, the coefficient Ki is given
Here we used the notation Fca,b := 2 F1 (a, b; c; 1−µ
by
(−1)i Γ(m + 1)Γ(nm)
Ki =
.
(21)
nΓ(i + 1)Γ(n − i)Γ(i + α + 2)Γ(nm − α − i − 1)
Figure 2 shows the comparison between analytical and Monte Carlo results for the
marginal density of the fixed trace ensemble. We find excellent agreement.
Using Selberg’s integral and its properties [1, 2], it can be shown that both the
average and the variance of the trace for the regular Wishart-Laguerre ensemble is mn.
Therefore, if we consider the ensemble of matrices W/mn, the corresponding eigenvalues
are 1/mn times the eigenvalues of W. Moreover, while individually these scaled matrices
may not have trace one, on average, it is one. In addition, the variance of trace for this
scaled ensemble is 1/mn, which becomes negligible for large n, m. Therefore, it is
expected that this scaled ensemble will approximate the behaviour of the fixed trace
ensemble. For instance, the marginal density for this scaled ensemble,
pe(µ) = mn p(mnµ),
(22)
should serve as an approximation to pF (µ). We can also use Marčenko-Pastur
density [51] to write down an expression for pe(µ) valid for large n, m:
p
p
(1 ± n/m)
m (µ+ − µ)(µ − µ− )
peMP (µ) =
; µ± =
.
(23)
2π
µ
n
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
8
Marginal density for fixed trace Wishart-Laguerre ensemble
using (20) with n = 8, and α values as indicated. The solid lines are analytical
predictions based on (20), and the symbols correspond to Monte Carlo results.
Figure 2.
This relation of the fixed trace ensemble with the scaled ensemble has been used
in [19, 52–54]. In figure 3 we plot the exact one-eigenvalue density (20) for the fixed
trace ensemble, as well as the densities (22), (23) based on the scaled ensemble. We
find that while the density obtained using the scaled ensemble is not able to capture the
oscillations, it does capture the overall shape of the density quite well.
25
�� ( μ )� � ( μ )� � ��( μ )
(a)
(b)
25
20
20
15
10
5
20
(c)
15
15
10
10
5
5
0
0
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.00
0.05
0.10
0.15
0
0.00 0.02 0.04 0.06 0.08 0.10
�
Figure 3. Marginal density of eigenvalues for the fixed trace Wishart-Laguerre
ensemble: Comparison between exact (solid black), scaled (dashed red), scaled
Marčenko-Pastur (dotted blue) as given by (20), (22) and (23), respectively.
(a) n = m = 15, (b) n = 20, m = 30, (c) n = 25, m = 75.
The exact result for the density of the smallest eigenvalue for the fixed trace
ensemble can be obtained using (7), (9), and the Laplace-inversion result
L−1 {s−a e−nsx }(t = 1) = (1 − nx)a−1 Θ(1 − nx)/Γ(a),
with Θ(z) being the Heaviside-theta function. We have
fF (x) = Γ(nm)L−1 {s1−nm f (sx)}(t = 1)
(24)
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
500
9
400
(a)
400
α=0
300
α=5
α=1
300
� � (�)
(b)
α=2
200
α = 10
200
α = 15
α=3
100
100
0
0
0.000 0.005 0.010 0.015 0.020 0.025 0.030 0.000
�
0.005
0.010
0.015
0.020
Figure 4. Probability density of the smallest eigenvalue for the fixed trace
Wishart-Laguerre ensemble with (a) n = 8, (b) n = 15, and α values as
indicated. The solid lines are analytical predictions based on (25), while the
symbols (filled- circles, squares, triangles) represent results of Monte Carlo
simulations.
= Γ(nm)
αn+1
X
j=α+1
hj
(1 − nx)nm−j−1 xj−1
Θ(1 − nx).
Γ(nm − j)
(25)
F
/Cn,α . In [36] a
The prefactor Γ(nm) comes from the ratio of normalizations, viz. Cn,α
similar strategy has been used for the real case. In figure 4 we show the comparison
between the analytical prediction and the numerical simulation for the smallest
eigenvalue density. They are in excellent agreement.
The idea of using scaled Wishart-Laguerre ensemble, as discussed above, can be
applied here as well. Therefore, an approximate density for the smallest eigenvalue can
be written using (7) as
fe(x) = mnf (mnx).
(26)
In figure 5 we compare this approximation with the exact result given by (25). The
approximation works pretty well. This approximate relation between the two densities
is also the reason behind the very similar shapes of the curves in figures 1 and 4,
respectively.
We also find that, using the first equality in (25), it follows that the η–th moment
of the smallest eigenvalue for the fixed trace ensemble is related to that of the regular
Wishart-Laguerre ensemble as
hxη iF =
Γ(nm)
hxη i.
Γ(nm + η)
(27)
This, similar to (10), holds for Re(η) > −α − 1.
Mathematica [47] codes to obtain explicit results for the above smallest eigenvalue
density of the fixed trace ensemble, as well as the moments are given in Appendix B.
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
120
(a)
� � (�)� � (�)
100
80
(b)
(c)
400
60
80
300
40
60
500
10
200
40
20
20
100
0
0
0
0.00 0.01 0.02 0.03 0.04 0.05 0.00 0.01 0.02 0.03 0.04 0.05 0.000 0.002 0.004 0.006 0.008 0.010
�
Figure 5. Comparison between the exact (fF (x): solid black) and approximate
(fe(x): dashed red) probability densities for the smallest eigenvalue of the fixed
trace Wishart-Laguerre ensemble, as given by (25) and (26), respectively. The
parameter values used are (a) n = m = 5, (b) n = 8, m = 13, and (c)
n = 20, m = 30.
Similar to the unrestricted trace case, we discuss below the cases α = 0, 1 for the
fixed trace scenario.
3.1. Results for α = 0
For α = 0 we just have one term in the series (25), and h1 = n. Therefore, we arrive at
2 −2
fF (x) = n(n2 − 1)(1 − nx)n
Θ(1 − nx).
(28)
Also, the expression for the η-th moment is given by
Γ(η + 1)Γ(n2 )
hx iF = η
.
n Γ(n2 + η)
η
(29)
These expressions are in agreement with those obtained in [23, 24].
3.2. Results for α = 1
In this case use of the result (16) for hj in (25) leads to the smallest eigenvalue density
expression
2
fF (x) = Γ(n +n)Γ(n+2)
n+1
X
j=2
2
(1 − nx)n +n−j−1 xj−1
Θ(1−nx).(30)
Γ(n − j + 2)Γ(j + 1)Γ(j − 1)Γ(n2 + n − j)
Also, the η–th moment follows as
n+1
Γ(n2 + n)Γ(n + 2) X
Γ(j + η)
hx iF =
.
2
j+η
Γ(n + n + η) j=2 n Γ(n − j + 2)Γ(j + 1)Γ(j − 1)
η
(31)
Chen, Liu and Zhou have provided the result for the cumulative distribution § of
the smallest eigenvalue for the complex case in terms of an inverse-Laplace transform
§ More appropriately, the survival function or the reliability function.
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
11
involving a determinant [35]:
n
o 1 − nx
1
(k)
mn−1 −1
−mn
Q(x) = Γ(mn)x
L
s
det[Ln+j−k (−s)]j,k=0,...,α−1
; 0 < x ≤ .(32)
x
n
We set α = 1 in this expression and use the expansion for associated Laguerre
polynomial [46], later on. The inverse Laplace transform can then be explicitly
performed leading us to
2
n
X
xj (1 − nx)n +n−j−1
2
.
(33)
Q(x) = Γ(n + 1)Γ(n + n)
Γ2 (j + 1)Γ(n − j + 1)Γ(n2 + n − j)
j=0
The probability density of the smallest eigenvalue follows upon using fF (x) =
−dQ(x)/dx as
2
n
X
nxj (1 − nx)n +n−j−2
2
fF (x) = Γ(n + 1)Γ(n + n)
Γ2 (j + 1)Γ(n − j + 1)Γ(n2 + n − j − 1)
j=0
2
−Γ(n + 1)Γ(n + n)
n+1
X
j=1
2
xj−1 (1 − nx)n +n−j−1
.
Γ(j)Γ(j + 1)Γ(n − j + 1)Γ(n2 + n − j)
(34)
In the second term we start the sum from j = 1 as j = 0 term is zero because of the
diverging gamma function Γ(j) in the denominator. Moreover, we have added a term
j = n + 1 which, again, is zero because of the diverging Γ(n − j + 1) in the denominator.
Next, we consider j → j − 1 in the summand of the first term, and hence the sum runs
from j = 1 to n + 1. The two terms can then be combined to yield (30) by noticing that
the n = 1 term is zero. We note that (32) also produces the correct result for α = 0 if
the determinant value in this case is interpreted as 1.
4. Large n, α evaluations and comparison with Tracy-Widom density
The universality aspects of the regular Wishart-Laguerre ensemble have been explored in
several notable works [29–34,38–41,55–60]. For the fixed trace case, the local statistical
properties of the eigenvalues have been studied in [35, 61]. In particular, it has been
shown that the fixed trace and the regular Wishart-Laguerre ensembles share identical
universal behaviour for large n at the hard edge, in the bulk and at the soft edge for α
fixed [61].
For the complex Wishart-Laguerre ensemble, in the square case (α = 0), the
smallest eigenvalue scaled by n gives rise to an exponential density [23, 25, 26].
Interestingly, this is true for all n in this case, as evident from (13). For large n it
has been shown that this result holds even if the matrices W are constructed from nonGaussian A [55] (cf. section 2) with certain considerations. For the fixed trace case,
in view of its connection to the scaled Wishart-Laguerre ensemble (25), as discussed
in section 3, the smallest eigenvalue has to be scaled by n3 to obtain the exponential
density [23, 24]. This can be easily verified to be true from (28). Furthermore, very
recently, 1/n corrections to the scaled smallest eigenvalue has been worked our for close
to square cases [34, 59, 60].
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
0.5
0.5
0.5
����������� �������
(a)
(b)
(c)
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0.0
-6
-4
-2
0
2
0.5
0.0
-6
-4
-2
0
2
0.5
0.0
-6
(d)
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
2
0.0
-6
2
(f)
0.4
-2
-2
(e)
0.4
-4
-4
0.5
0.4
0.0
-6
12
-4
-2
�
0
2
0.0
-6
-4
-2
0
2
Figure 6. Comparison of Tracy-Widom density (solid black) with densities
−σf (σx + η) (dashed red) and −(σ/mn)fF ((σx + η)/mn) (dotted blue) for
(a) n = 25, m = 125, (b) n = 25, m = 225, (c) n = 25, m = 425, (d)
n = 50, m = 150, (e) n = 50, m = 250, and (f) n = 50, m = 450. It should be
noted that the rectangularity α varies as 100 for (a), (d); 200 for (b), (e); and
400 for (c), (f). Also, the aspect ratio n/m is 1/5 for (a), (e), and 1/9 for (b),
(f), respectively.
For the rectangular case, Feldheim and Sodin [56] have shown that, in the limit m →
∞, n → ∞ with n/m bounded away from 1, the shifted and scaled smallest eigenvalue,
(λmin − η)/σ, leads to the Tracy-Widom density [39, 40]. Here η = (n1/2 − m1/2 )2 and
σ = (n1/2 − m1/2 )(n−1/2 − m−1/2 )1/3 < 0. The convergence, however, is slower when
α = m − n is close to 0. This can be attributed to the fact that the hard-edge behaviour
is prevalent unless α is large [58]. We should also mention that the Tracy-Widom
density captures the largest typical fluctuations of the smallest eigenvalue, while the
larger atypical fluctuations are described by large deviation results, as derived in [57]
by Katzav and Castillo.
As a consequence of identical universal behaviour of spectra of the regular and
fixed-trace ensembles [61], the Tracy-Widom density is also expected in the case of fixed
trace ensemble. The proper scaling in this case can be inferred from the connection
with the scaled Wishart-Laguerre ensemble, as discussed in Sec. 3. This implies that
the density of (mnµmin − η)/σ will converge to the Tracy-Widom result.
The recursion scheme given in section 2 enables us to work out the exact results
for the smallest eigenvalue density for large values of n and α and hence to explore
the above limit. In view of the scaling and shift indicated above, −σf (σx + η)
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
13
and −(σ/mn)fF ((σx + η)/mn) should coincide with the Tracy-Widom density of
the unitarily-invariant class. We examine this in figure 6. We can see that as the
rectangularity α increases the agreement improves for both n = 25 and n = 50. This
is because the hard-edge effect is diminished with increasing α. We also find that for a
given aspect ratio n/m < 1, as expected, the agreement is better for larger n.
5. Entanglement in bipartite systems
Consider a bipartite partition of an N1 N2 –dimensional Hilbert space H(N1 N2 ) consisting
(N )
(N )
of subsystems A and B, which belong to Hilbert spaces HA 1 and HB 2 , respectively,
(N )
(N )
such that H(N1 N2 ) = HA 1 ⊗ HB 2 . A general state |ψi of H(N1 N2 ) is given in terms of
(N )
(N )
the orthonormal states |iA i of HA 1 , and |αB i of HB 2 as
|ψi =
N1 X
N2
X
xi,α |iA i ⊗ |αB i,
(35)
i=1 α=1
PN1 PN2
2
where xi,α are complex coefficients, such that hψ|ψi =
i=1
α=1 |xi,α | = 1. The
density matrix for the composite system, considering a pure state scenario, is given by
ρ = |ψihψ| =
N1 X
N2
X
xi,α x∗j,β |iA ihj A | ⊗ |αB ihβ B |,
(36)
i,j=1 α,β=1
with tr[ρ] = 1. The reduced density matrix for subsystem, say A, can be obtained by
tracing out the subsystem B as
ρA =
N2
X
0
0
hα |ρ|β i =
α0 ,β 0 =1
N1
X
Fi,j |iA ihj A |,
(37)
i,j=1
P 2
∗
where Fi,j = N
α=1 xi,α xj,α can be viewed as the matrix elements of some N1 × N1 –
dimensional matrix F = XX† . Here X is a rectangular matrix of dimension N1 × N2
that has xi,α as its elements. In the eigenbasis of F, (37) can be written as
ρA =
N1
X
µi |µAi ihµAi |.
(38)
i=1
The eigenvalues µi of F are referred to as the Schmidt eigenvalues or Schmidt numbers.
Due to the trace condition, they satisfy
N1
X
µi = tr XX† = tr F = 1.
(39)
i=1
Suppose N1 ≤ N2 . Now, if we sample all normalized density matrices with equal
probabilities, i.e., if we choose the coefficients xi,α randomly using the Hilbert-Schmidt
density PX (X) ∝ δ(tr XX† − 1), then F defined here is statistically equivalent to the
F defined in (18), and the statistics of the Schmidt eigenvalues are described exactly
by the joint eigenvalue density of the fixed trace Wishart-Laguerre ensemble (19), with
N1 = n, N2 = m [19, 20]. It should be noted that the reduced density matrix for
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
14
the subsystem B will correspond to the matrix X† X, which will share the eigenvalues
µ1 , ..., µn , and will have the rest of its m − n eigenvalues as zero. As such, it carries the
same amount of information as F.
The Schmidt eigenvalues can be used to study various entanglement measures
such as von-Neumann entropy, Renyi entropies, concurrence, purity etc. As a
consequence, fixed trace Wishart-Laguerre ensemble has been extensively used to model
the reduced density matrices arising in the study of entanglement formation in bipartite
systems [14–16, 20, 23, 24, 35, 36, 48–50, 52–54, 61–67]. These works have explored several
aspects such as moments and distributions of Schmidt eigenvalues and entanglement
measures.
The density of the minimum eigenvalue in the present context not only sheds light
on the nature of the entanglement, but also provides important information about the
degree to which the effective dimension of the Hilbert space of the smaller subsystem
can be reduced [23, 24]. The smallest eigenvalue assumes values from 0 to 1/n. In
P
the extreme case of 1/n, it follows from the trace constraint ni=1 µi = 1, that all the
eigenvalues must have the same value 1/n. Consequently, the von-Neumann entropy,
P
− ni=1 µi ln µi , assumes its maximum value ln n, thereby making the corresponding
state maximally entangled. In the other extreme of the smallest eigenvalue being 0 (or
very close to 0), while it does not provide information regarding entanglement, from
the Schmidt decomposition it follows that the effective Hilbert space dimension of the
subsystem gets reduced by one.
In the next section we consider a system of coupled kicked tops and explore to what
extent the behaviour of the smallest Schmidt eigenvalue is described by the fixed trace
Wishart-Laguerre ensemble.
6. Coupled kicked tops
The kicked top system has been a paradigm for studying chaos, both classically and
quantum mechanically [68, 69]. Remarkably, it has also been realized experimentally
using an ensemble of Caesium atoms [70]. In the study of entanglement formation in
bipartite systems, a coupled system of two kicked tops has turned out to be of great
importance and has been investigated by a number of researchers [45, 53, 54, 65, 71, 72].
The full Hamiltonian of the coupled kicked top system is
H = H1 ⊗ 1N2 + 1N1 ⊗ H2 + H12 .
(40)
∞
kr 2 X
π
J
δ(t − ν), r = 1, 2,
Hr = Jyr +
2
2jr zr ν=−∞
(41)
Here,
represent the Hamiltonians for the individual tops, and
∞
X
H12 = √
(Jz ⊗ Jz2 )
δ(t − ν)
j1 j2 1
ν=−∞
(42)
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
15
is the interaction term. Also, 1Nr represents identity operator that acts on Nr dimensional Hilbert space H(Nr ) . The Hamiltonians H1 and H2 correspond respectively
to N1 (= 2j1 + 1)-dimensional, and N2 (= 2j2 + 1)-dimensional Hilbert spaces H(N1 )
and H(N2 ) , respectively. The Hamiltonian for the coupled kicked tops corresponds
to an N1 N2 -dimensional Hilbert space H(N1 N2 ) = H(N1 ) ⊗ H(N2 ) . Also, (Jxr , Jyr , Jyr )
are angular momentum operators for the rth top and satisfy the usual commutation
relations. The parameter kr controls the chaotic behaviour of the individual tops. The
parameter takes care of the coupling between the two tops.
The unitary time evolution operator (Floquet operator) corresponding to the
Hamiltonian (40) is
U = (U1 ⊗ U2 )U12 ,
(43)
with
ιkr 2
ιπ
J
, r = 1, 2;
Ur = exp − Jyr −
2
2jr zr
ι
U12 = exp − √
Jz ⊗ Jz2 .
j1 j2 1
(44)
(45)
√
Here ι = −1 represents the imaginary unit. The initial state for the individual
tops is chosen as a generalized SU(2) coherent state or the directed angular
(r)
(r)
momentum state [68,
is given in |jr , mr i basis as hjr , mr |θ0 , φ0 i =
r69], which
−j
(r)
(r)
2jr
(1 + |γr |2 ) r γrjr −mr
with γr ≡ exp(ιφ0 ) tan(θ0 /2). For later use, we define
jr +mr
Nr -dimensional vectors given by
(r)
(r)
χr = [hjr , mr |θ0 , φ0 i]mr =−jr ,...,+jr .
(46)
For the coupled top, the initial state is taken as the tensor-product of the states
(1)
(1)
(2)
(2)
of the individual tops: |ψ(0)i = |θ0 , φ0 i ⊗ |θ0 , φ0 i. We can implement the time
evolution to obtain the state |ψ(ν)i starting from |ψ(0)i using the iteration scheme
|ψ(ν)i = U |ψ(ν − 1)i = (U1 ⊗ U2 )U12 |ψ(ν − 1)i, which, when written in hj1 , s1 ; j2 , s2 |
basis, is [72]
hj1 , s1 ; j2 , s2 |ψ(ν)i = exp −ι √
s1 s2
j1 j2
+j1
+j2
X
X
×
hj1 , s1 |U1 |j1 , m1 ihj2 , s2 |U2 |j2 , m2 ihj1 , m1 ; j2 , m2 |ψ(ν − 1)i.
m1 =−j1 m2 =−j2
A convenient approach for implementing this iteration scheme and eventually
calculating the reduced density matrix involves writing the states as N1 × N2 matrices:
Ψ(ν) = V ◦ (U1 Ψ(ν − 1)UT2 ).
(47)
Here ‘◦’ represents the Hadamard product and ‘T ’ the transpose. V is an N1 × N2
matrix given by
V = exp −ι √
ab
.
(48)
a=−j1 ,...,+j1
j1 j2
b=−j2 ,...,+j2
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
(a)
50
(1)
(1)
(1)
(2)
(2)
(1)
(1)
(2)
(2)
θ0 =3.08,ϕ0 =2.35,
30
(c)
12
(1)
(1)
10
θ0 =1.62,ϕ0 =1.53
8
θ0 =1.71,ϕ0 =1.42,
θ0 =0.34,ϕ0 =0.47
10
0
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
(i)
(2)
(1)
(2)
(2)
θ0 =1.33,ϕ0 =2.12
6
20
(2)
(1)
(d)
12
(1)
(1)
10
θ0 =2.35,ϕ0 =0.83
8
θ0 =2.53,ϕ0 =1.61,
(2)
(2)
(1)
(1)
(2)
(2)
10
θ0 =2.81,ϕ0 =1.72
8
θ0 =1.73,ϕ0 =1.21,
(2)
(2)
(1)
(1)
(2)
(2)
θ0 =1.64,ϕ0 =2.52
θ0 =1.81,ϕ0 =1.52
6
(1)
θ0 =2.55,ϕ0 =0.63,
θ0 =1.22,ϕ0 =2.13,
θ0 =2.11,ϕ0 =0.96,
θ0 =2.43,ϕ0 =0.89
40
pF ( μ)
(b)
12
θ0 =0.75,ϕ0 =1.71,
16
6
4
4
4
2
2
2
0
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
0
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
0
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35
μ
140
(ii)
120
(iii)
(iv)
120
120
400
100
100
100
� � (�)
300
80
80
80
60
60
60
200
40
40
40
100
20
0
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0
0.000
20
20
0.005
0.010
0.015
0.020
0.025
0.030
�
0
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0
0.000
0.005
0.010
0.015
0.020
0.025
0.030
Figure 7. Comparison of marginal density (top row) and the smallest eigenvalue
density (bottom row) of the coupled kicked top system with the fixed trace
ensemble results for N1 = 11, N2 = 21 and = 1. For plots (a)-(d) as well as
(i)-(iv), the parameters (k1 , k2 ) vary as (0.5, 1), (0.5, 8), (2.5, 3), (7, 8). In
each case the solid (black) curves correspond to the analytical results, while the
dotted (blue) and dashed (red) curves correspond to different initial conditions.
The θ0 and φ0 values used for different initial conditions are mentioned in (a)(d), and hold, respectively, for (i)-(iv) also.
Also, Ur is an Nr × Nr dimensional matrix
kr 2
(jr ) π
Ur = exp −ι a da,b
.
2jr
2 a,b=−jr ,...,+jr
(49)
(j )
Here da,br represents the Wigner (small) d matrix elements. We use the inbuilt function
in Mathematica [47] for Wigner (big) D matrix to evaluate it. The initial N1 × N2 dimensional state matrix is given by
Ψ(0) = χ1 ⊗ χT2 .
(50)
Eventually, the reduced density matrix of dimension N1 × N1 can be constructed as
ρd = Ψ(ν)Ψ(ν)† .
(51)
The eigenvalues of this matrix are the sought after Schmidt eigenvalues, whose statistics
is of interest to us. To obtain an ensemble of states we proceed as follows. We begin
with an initial state and apply (47) iteratively. After ignoring initial 500 states to safely
avoid the transient regime [54], we start considering states separated by 20 time steps
to put off any unwanted correlations. In all, we consider 50000 states for statistical
analysis.
In figure 7 we set N1 = 11, N2 = 21, = 1, and examine the effect of different choices
of k1 , k2 on one level density and smallest eigenvalue density for the coupled kicked tops.
For (a), (i) we have k1 = 0.5, k2 = 1 for which the classical phase spaces of the individual
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
14
pF (μ )
14
(a)
12
14
(b)
12
10
10
8
8
8
6
6
6
4
4
4
2
2
0.1
0.2
0.3
0.4
0.5
0.6
25
2
0
0.0
0.1
0.2
0.3
0.4
0
0.0
0.5
25
(d)
20
20
15
10
10
10
5
5
5
0.10
0.15
0.20
0.25
0.30
0.2
0.3
0.4
0
0.00
0.05
0.10
0.15
0.20
0.25
(f)
20
15
0.05
0.1
25
(e)
15
0
0.00
(c)
12
10
0
0.0
17
0.30
0
0.00
0.05
0.10
0.15
0.20
μ
2000
350
(i)
1500
(ii)
300
(iii)
150
250
100
200
1000
150
50
100
500
f F (x)
50
0
0.000
0.001
0.002
0.003
0.004
0.005
10 000
0
0.000
0.005
0.010
(v)
(iv)
0.01
0.02
0.03
0.04
(vi)
500
400
1000
6000
0
0.00
600
1500
8000
0.015
300
4000
200
500
2000
0
0.0000
100
0.0002
0.0004
0.0006
0.0008
0
0.0000 0.0005 0.0010 0.0015 0.0020 0.0025
0
0.000
0.002
0.004
0.006
0.008
x
Figure 8. Effect of varying on marginal density ((a)-(f)), and the smallest
eigenvalue density ((i)-(vi)). The solid lines (black) correspond to analytical
results, while the dotted (red), dot-dashed (blue) and dashed (green) curves
result from coupled top simulation for = 0.05, 0.1 and 0.5, respectively. The
parameters k1 , k2 are fixed at 7, 8, while dimension parameters (N1 , N2 ) vary
for the figures (a)-(f) as well as (i)-(vi) as (11, 11), (11, 15), (11, 25), (21, 21),
(21, 25), (21, 35).
tops consist mostly of regular orbits [54]. In this case, we can see deviations from the
(r)
(r)
fixed trace ensemble results with strong sensitivity to initial conditions, i.e. θ0 and φ0
values. In (b), (ii) we set k1 = 0.5, k2 = 8. In this case highly chaotic phase space [54]
of the second top leads to an agreement with the results of the fixed trace ensemble,
even though the phase space of the first top is mostly regular. Moreover, there is a
weak sensitivity to initial conditions. In (c), (iii) we consider k1 = 2.5, k2 = 3, both of
which correspond to mixed type phase space [54]. Here we observe deviations, along
with some sensitivity to initial conditions. Finally, in (d), (iv) we take k1 = 7, k2 = 8,
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
18
for which phase spaces of both the tops are highly chaotic. In this case, we have very
good agreement with the random matrix results and very weak sensitivity to the initial
conditions.
In figure 8 we consider a chaotic regime (k1 = 7, k2 = 8) and examine the effect of
varying for various combinations of n and α. We observe that for a given , increase in
n or α leads to a better agreement with the fixed trace ensemble results. Recent studies
in a similar direction have investigated the universal aspects of spectral fluctuations and
entanglement transitions in strongly chaotic subsystems [73, 74].
A quantifier to measure the fraction of close to maximally entangled states can be
R 1/n
the cumulative probability R(δ) = 1/n−δ fF (x) dx [24], that turns out to be vanishingly
small for δ << 1/n and thus, implies that the actual fraction of such states is extremely
small. For example, using (25), we obtain R(δ = 0.1/n) value to be roughly (i) 8 × 10−6
for n = 3, m = 11, (ii) 1×10−35 for n = 7, m = 19, and (iii) 5×10−91 for n = 11, m = 25.
7. Summary and conclusion
We considered complex Wishart-Laguerre ensemble, both with and without the fixed
trace condition, and provided an easily implementable recurrence scheme to obtain the
exact result for the density of the smallest eigenvalue. This method also gives access to
arbitrary moments of the smallest eigenvalue. The recursion-based approach for exact
and explicit expressions for the density is preferable to the results based on determinants
which are difficult to handle with increasing dimensionality n or rectangularity α.
We also demonstrated the equivalence of the recurrence scheme and the determinantbased results for α = 0 and 1. We validated our analytical results using Monte Carlo
simulations and also used large n and α evaluations to compare with the Tracy-Widom
density. As an application to quantum entanglement problem we explored the behaviour
of Schmidt eigenvalues of the coupled kicked top system. Among other things, we found
that in the chaotic regime, the fixed trace ensemble describes the behaviour of the
Schmidt eigenvalues very well if sufficient coupling is provided between the constituent
tops.
Acknowledgments
This work initiated from a project that was carried out at Shiv Nadar University under
the Opportunities for Undergraduate Research (OUR) scheme. The authors are grateful
to the anonymous reviewers for fruitful suggestions.
Appendix A. Recurrence scheme
We begin with (6) and apply the shift λi → λi + x. This results in
Z ∞
Z ∞
n
Y
Y
α −nx
2
f (x) = nCn,α x e
dλ2 · · ·
dλn
(λj − λk )
λ2i (λi + x)α e−λi .
0
0
2≤k<j≤n
i=2
(A.1)
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
19
To derive the recurrence relation, we will proceed parallel to the steps in chapter 4
of [25], or as in [26]. To this end, we shift the indices of the integration variables as
λi → λi−1 , and also introduce the measure dΩi = λ2i e−λi dλi . Consequently, we arrive
at the following expression:
Z ∞
Z ∞
n−1
Y
2
α −nx
dΩn−1 ∆n−1 ({λ})
(λi + x)α . (A.2)
dΩ1 · · ·
f (x) = nCn,α x e
0
0
i=1
Next, we define
α
Ii,j
Z
∞
Z
∞
dΩn−1 ∆2n−1 ({λ})u(x),
dΩ1 · · ·
=
(A.3)
0
0
where the integrand u(x) is
(λ + x)α · · · (λi + x)α (λi+1 + x)α−1 · · · (λi+j + x)α−1
| 1
{z
}|
{z
}
i terms
j terms
α−2
α−2
.
× (λi+j+1 + x)
· · · (λn + x)
{z
}
|
(A.4)
n−i−j−1 terms
We also define the operator
Z ∞
Z
α
Ii,j [v] =
dΩ1 · · ·
0
∞
dΩn−1 ∆2n−1 ({λ})u(x) v.
(A.5)
0
Using the above notation the smallest eigenvalue density can be written as
α
f (x) = nCn,α xα e−nx In−1,0
.
(A.6)
With these, Lemma 4.2 of [25] (or, equivalently, Lemma 4.1 of [26]) holds in the complex
case also:
(
α
α
Ii+1,j−1
− xIi,j
if i < k ≤ i + j,
α
Ii,j [λk ] =
(A.7)
α
α
Ii,j+1 − xIi,j
if i + j < k < n.
The above result follows by writing λk as (λk + x) − x and then using the operator
defined in (A.5). Next, if the terms (λk + x) and (λl + x) share the same exponent in
the integrals (i.e., both k and l fall within one of the closed intervals [1, i], [i + 1, i + j],
or [i + j + 1, n − 1]), then
λk λl
α
Ii,j
= 0,
(A.8)
λk − λl
λk
1 α
α
Ii,j
= Ii,j
,
(A.9)
λk − λl
2
λ2k
α
α
Ii,j
= Ii,j
[λk ].
(A.10)
λk − λl
Equation (A.8) follows because of the asymmetry in λk and λl , while (A.9) is obtained
using the identity λk /(λk −λl )+λl /(λl −λk ) = 1 and using symmetry. The result (A.10)
is obtained using the identity λ2k /(λk − λl ) = λk + λk λl /(λk − λl ) and (A.8).
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
20
The crucial difference occurs in the first equation of Lemma 4.3 [25] (or Lemma
4.2 [26]), which reads for the present case as
α
α
α
α
Ii,j
= (x + α + j + 2k + 2)Ii−1,j+1
− x[k + (α − 1)]Ii−1,j
+ (i − 1)xIi−2,j+2
α
I0,j
=
α−1
Ij,n−j−1
,
(A.11)
(A.12)
with k = n − i − j − 1. The second equation of this set, (A.12), follows readily from the
definition (A.3). This first equation of this set, (A.11), is derived using
α
α
α
Ii,j
= xIi−1,j+1
+ Ii−1,j+1
[λi ],
(A.13)
which is a consequence of (A.7). The difference in the result compared to the real case
α
occurs due to the term Ii−1,j+1
[λi ]. For the complex case, this involves observing the
following:
Z ∞
Z ∞
Y
Y
d
α−1
2 3 −λi
[(λi +x)α−1 (λl −λi )2 λ3i ]e−λi dλi .(A.14)
(λi +x)
(λl −λi ) λi e dλi =
dλi
0
0
i<l
i<l
Also, using the result
α
dIi−1,j+1
α
α
= (i − 1)αIi−2,j+2
+ (j + 1)(α − 1)Ii−1,j
(A.15)
dx
for i + j = n − 1, as given in the proof for Lemma 4.4 of [25] (or Lemma 4.3 of [26]),
yields in the present case
α
x d α
α
α
α
Ii−1,j+1 + x(i − 1) 1 +
Ii−2,j+2
. (A.16)
Ii,j = (x + α + j + 2)Ii−1,j+1 −
j + 1 dx
j+1
α−1
α
in view of (A.12). Equation (A.16)
Now, we can begin with In−1,0
, which is same as I0,n−1
α
can be used with j = n−i−1 repeatedly for i = 1 to n−1 to arrive at In−1,0
, starting from
α
α
I0,n−1 . We note that In−1,0 is the term needed to obtain the smallest eigenvalue density
expression (A.6) explicitly. This is essentially what has been employed in the recurrence
0
0
α
= 1/Cn−1,2 .
involving Si := Ii,n−i−1
/In−1,0
for gn,m (x) in (7). We also note that In−1,0
The constant cn,m of (7) is therefore nCn,α /Cn−1,2 .
Appendix B. Mathematica codes
The following code can be implemented in Mathematica [28] to obtain exact expressions
for the smallest eigenvalue density for the Wishart-Laguerre ensemble:
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
21
� = ��� � = ���
�[��_� ��_] �=
�
�����[��] �����[��]
�������
�����[� + �]
�����[� + �� - ��]
� {�� �� �� - �}�
α = � - �� �[-�] = �� �[�] = �� �[� - �] = ��
���� = �� � < α + �� � ++�
� = � + ��
���� = �� � < �� � ++�
�[�] = ������(� + � - � + �) �[� - �] -
�
�-�
�[�[� - �]� �] + � (� - �)
�-�
�-�
�[� - �]�
�[-�] = �� �[�] = �[� - �]�
�[�_ ] = �[�� �] ⅇ-� � �α ������[�[� - �]]
For generating the smallest eigenvalue density for the fixed trace Wishart-Laguerre
ensemble, the following code can be used along with the above.
� = ���������������ⅇ� � �[�]� ��
��[�_ ] = �����������[� �] ����[[�]]
(� - � �)� �-�-� ��-�
�����[� � - �]
� {�� α + �� α � + �} ��������������[� - � �]
We can also directly implement the inverse Laplace transform function built in
Mathematica:
��[�_ ] = �����������[� �] �������������������������-� � �[� �]� �� �
The ‘Factor’ option in the above codes is for printing compact expressions on-screen.
For computation involving large n or α values, it may be removed, since factoring very
large expressions may result in a large computation time.
The moments of the smallest eigenvalue of the regular or the fixed-trace WishartLaguerre ensemble can be obtained using the following functions:
���[η_ ] �= ���
����[η_ ] �=
�[[�]]
��+η
�����[� + η ]� {�� α + �� α � + �}
�����[� �]
�����[� � + η ]
���[η ]
Appendix C. Relation with associated Laguerre polynomial
The associated Laguerre polynomials satisfy the following relations [46]:
(k)
(k+1)
(k+2)
iLi (−x) = (x + k + 1)Li−1 (−x) + xLi−2 (−x),
d (k)
(k+1)
Li (−x) = Ln−1 (−x).
dx
These two can be combined to obtain the following relation:
x d (k+1)
xk (k+2)
(k)
(k+1)
iLi (−x) = (x + k + 1)Li−1 (−x) −
Li−1 (−x) +
L
(−x).
k − 1 dx
k − 1 i−2
(C.1)
(C.2)
(C.3)
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
22
Considering k = n − i + 1 gives
(n−i+1)
iLi
(n−i+2)
(−x) = (x + n − i + 2)Li−1
(−x) −
x d (n−i+2)
L
(−x)
n − i dx i−1
x(n − i + 1) (n−i+3)
Li−2
(−x).
(C.4)
n−i
Multiplying this equation by Γ(i) and then calling Si = Γ(i + 1)Li (n − i + 1)(−x), we
get
+
x dSi−1
(n − i + 1)
+ x(i − 1)
Si−2 . (C.5)
n − i dx
n−i
This recurrence relation is the same as that given in section 2 when used for m = n + 1.
(2)
Hence, gn,n+1 (x) = Sn−1 = Γ(n)Ln−1 (−x).
Si = (n − i + 2 + x)Si−1 −
Appendix D. Some explicit results
For α = m − n = 0, the smallest eigenvalue density expressions valid for all n are quite
compact and are already provided in (13) and (28), respectively, for the regular WishartLaguerre ensemble and for the fixed trace Wishart-Laguerre ensemble. For a few other
cases we tabulate the exact results in Tables D1 and D2 using the above Mathematica
codes. This includes the α = 1 case for which closed-form results for any n are given in
(15) and (30). In the case of fixed trace ensemble there is a Θ(1 − nx) term in each of
the probability density expressions that we have not shown in the table for the sake of
compactness.
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
n
23
m
f (x)
3
e−2x x (x + 3)
4
.
e−2x x2 x2 + 6x + 12 6
5
e−2x x3 x3 + 9x2 + 36x + 60 /72
6
e−2x x4 x4 + 12x3 + 72x2 + 240x + 360 /1440
4
.
e−3x x x2 + 8x + 12 2
5
.
e−3x x2 x4 + 16x3 + 96x2 + 240x + 240 48
6
.
e−3x x3 x6 + 24x5 + 252x4 + 1440x3 + 4680x2 + 8640x + 7200 2880
7
.
e−3x x4 x8 + 32x7 + 480x6 + 4320x5 + 25200x4 + 97920x3 + 253440x2 + 403200x + 302400 345600
5
.
e−4x x x3 + 15x2 + 60x + 60 6
6
.
e−4x x2 x6 + 30x5 + 360x4 + 2160x3 + 6840x2 + 10800x + 7200 720
e−4x x3 x9 + 45x8 + 900x7 + 10380x6 + 75600x5 + 360720x4 + 1130400x3 + 2268000x2
.
+2721600x + 1512000 259200
e−4x x4 x12 + 60x11 + 1680x10 + 28800x9 + 334800x8 + 2773440x7 + 16790400x6 + 74995200x5
.
+246456000x4 + 586656000x3 + 972518400x2 + 1016064000x + 508032000 217728000
2
3
4
7
8
6
.
e−5x x x4 + 24x3 + 180x2 + 480x + 360 24
7
.
e−5x x2 x8 + 48x7 + 960x6 + 10320x5 + 64800x4 + 241920x3 + 524160x2 + 604800x + 302400 17280
e−5x x3 x12 + 72x11 + 2340x10 + 45120x9 + 572400x8 + 5019840x7 + 31157280x6 + 137894400x5
.
+432734400x4 + 943488000x3 + 1371686400x2 + 1219276800x + 508032000 43545600
e−5x x4 x16 + 96x15 + 4320x14 + 120480x13 + 2323440x12 + 32780160x11 + 349493760x10 + 2870380800x9
+18353563200x8 + 91755417600x7 + 358177075200x6 + 1083937075200x5 + .
2506629888000x4
5
8
9
+4316239872000x3 + 5267275776000x2 + 4096770048000x + 1536288768000
Table D1. Results for the Wishart-Laguerre ensemble
292626432000
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
n
m
fF (x)
3
2
60x 1 − x 1 − 2x
4
2
2
420x2 1 − x
1 − 2x
5
3
2
2520x3 1 − x
1 − 2x
6
4
2
13860x4 1 − x
1 − 2x
4
7
660x 1 − 3x2 1 − 3x
5
7
10920x2 1 − x − x2 − 9x3 + 15x4 1 − 3x
6
7
28560x3 5 − 12x + 12x2 − 48x3 − 48x4 + 432x5 − 411x6 1 − 3x
7
7
1627920x4 1 − 4x + 8x2 − 16x3 + 320x6 − 756x7 + 489x8 1 − 3x
5
3420x 1 + 5x − 20x2 + 4x3 1 − 4x)14
2
3
6
4
7
8
6
7
5
8
9
14
106260x2 1 + 6x + x2 − 204x3 + 486x4 − 424x5 + 356x6 1 − 4x
491400x3 5 + 27x + 51x2 − 683x3 − 5286x4 + 35910x5 − 85295x6 + 116895x7
14
−79980x8 − 9196x9 1 − 4x
6796440x4 7 + 28x + 86x2 − 540x3 − 6775x4 − 18416x5 + 440876x6 − 2012008x7
14
+4901710x8 − 7145600x9 + 5855692x10 − 3288592x11 + 2386196x12 1 − 4x
23
12180x 1 + 16x − 39x2 − 140x3 + 220x4 1 − 5x
23
628320x2 1 + 22x + 142x2 − 1234x3 − 580x4 + 4676x5 + 29788x6 − 92420x7 + 75355x8 1 − 5x
23030280x3 1 + 24x + 243x2 + 280x3 − 19962x4 + 50208x5 − 31022x6 + 649056x7 − 1420095x8
23
−7867032x9 + 35763831x10 − 53675640x11 + 27627140x12 1 − 5x
97740720x4 7 + 168x + 1968x2 + 9642x3 − 75517x4 − 1457898x5 + 10143328x6 − 31939648x7
+134132583x8 − 323536148x9 − 511260568x10 + 786421818x11 + 22191959881x12
23
−105911938466x13 + 211492028376x14 − 203837200540x15 + 80216630930x16 1 − 5x
Table D2. Results for the fixed trace Wishart-Laguerre ensemble
24
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
25
References
[1] Mehta M L 2004 Random Matrices (New York: Academic Press)
[2] Forrester P J 2010 Log-Gases and Random Matrices (LMS-34) (Princeton University Press,
Princeton, NJ)
[3] Anderson T W 2003 An Introduction to Multivariate Statistical Analysis (John Wiley & Sons)
[4] Muirhead R J 2005 Aspects of Multivariate Statistical Theory (Wiley Interscience)
[5] James A T 1964 Ann. Math. Statist. 35 475
[6] Gnanadesikan R 1997 Methods for Statistical Data Analysis of Multivariate Observations (John
Wiley & Sons)
[7] Plerou V, Gopikrishnan P, Rosenow B, Amaral L A N, Guhr T and Stanley H E 2002 Phys. Rev.
E 65 066126
[8] Vinayak and Pandey A 2010 Phys. Rev. E 81 036202
[9] Verbaarschot J 1994 Phys. Rev. Lett. 72 2531
[10] Verbaarschot J J M and Wettig T 2000 Ann. Rev. Nucl. Part. Sci. 50 343
[11] Tulino A M and Verdú S 2004 Random Matrix Theory and Wireless Communications (Now
Publishers Inc)
[12] Foschini G J and Gans M J 1998 Wireless Pers. Commun. 6 311
[13] Telatar I E 1999 Europ. Trans. Telecommun. 10 585
[14] Lubkin E 1978 J. Math. Phys. (N.Y.) 19 1028
[15] Lloyd S and Pagels H 1988 Ann. Phys. (NY) 188 186
[16] Page D N 1993 Phys. Rev. Lett. 71 1291
[17] Hall M J W 1998 Phys. Lett. A 242 123
[18] Życzkowski K and Sommers H-J 2001 J. Phys. A: Math. Gen. 34 7111
[19] Sommers H-J and Życzkowski K 2004 J. Phys. A: Math. Gen 37 8457
[20] Osipov V A, Sommers H-J and Życzkowski K 2010 J. Phys. A: Math. Theor. 43 055302
[21] Burel G 2002 Proc. of the WSEAS Int. Conf. on Signal, Speech and Image Processing (ICOSSIP
2002)
[22] Park C S and Lee K B 2008 IEEE Trans. Wireless Commun. 7 4432
[23] Majumdar S N, Bohigas O and Lakshminarayan A 2008 J. Stat. Phys. 131 33
[24] Majumdar S N 2011 Handbook of Random Matrix Theory, eds. Akemann G, Baik J and Di
Francesco P (Oxford Press, New York)
[25] Edelman A 1989 Eigenvalues and Condition Numbers of Random Matrices Ph.D. thesis, MIT
[26] Edelman A 1991 Lin. Alg. Appl. 159 55
[27] Khatri C G 1964 Ann. Math. Statist. 35 1807
[28] Forrester P J and Hughes T D 1994 J. Math. Phys. 35 6736
[29] Forrester P J 1993 Nucl. Phys. B 402 709
[30] Forrester P J 1994 J. Math. Phys. 35 2539
[31] Nagao T and Forrester P J 1998 Nucl. Phys. B 509 561
[32] Damgaard P H and Nishigaki S M 2001 Phys. Rev. D 63 045012
[33] Akemann G, Guhr T, Kieburg M, Wegner R and Wirtz T 2014 Phys. Rev. Lett. 113 250201
[34] Edelman A, Guionnet A and Péché S 2016 Ann. App. Prob. 26 1659
[35] Chen Y, Liu D-Z and Zhou D-S 2010 J. Phys. A: Math. Theor. 43 315303
[36] Akemann G and Vivo P 2011 J. Stat. Mech. 2011 P05020
[37] Zanella A, Chiani M and Win M Z 2009 IEEE Trans. Commun. 57 1050
[38] Nishigaki S M, Damgaard P H and Wettig T 1998 Phys. Rev. D 58 087704
[39] Tracy C A and Widom H 1993 Phys. Lett. B 305 115
[40] Tracy C A and Widom H 1994 Commun. Math. Phys. 159 151
[41] Tracy C A and Widom H 1994 Commun. Math. Phys. 161 289
[42] Borodin A and Forrester P J 2003 J. Phys. A: Math. Gen. 36 2963
[43] Forrester P J 2007 J. Phys. A: Math. Theor. 40 11093
Smallest eigenvalue density for complex Wishart-Laguerre ensemble
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
26
Wirtz T and Guhr T 2013 Phys. Rev. Lett. 111 094101
Trail C M, Madhok V and Deutsch I H 2008 Phys. Rev. E 78 046211
Szego G 1975 Orthogonal Polynomials (American Mathematical Society, Providence)
Wolfram Research Inc. Mathematica Version 10.0 (Wolfram Research Inc.: Champaign,
Illinois)
Kumar S and Pandey A 2011 J. Phys. A: Math. Theor. 44 445301
Adachi S, Toda M and Kubotani H 2009 Ann. Phys. 324 2278
Vivo P 2010 J. Phys. A: Math. Theor. 43 405206
Marčenko V A and Pastur L A 1967 Math. USSR-Sb. 1 457
Žnidarič M 2007 J. Phys. A: Math. Theor. 40 F105
Bandyopadhyay J N and Lakshminarayan A 2002 Phys. Rev. Lett. 89 060402
Bandyopadhyay J N and Lakshminarayan A 2004 Phys. Rev. E 69 016201
Tao T and Vu V 2010 Geom. Funct. Anal. 20 260
Feldheim O N and Sodin S 2010 Geom. Funct. Anal. 20 88
Katzav E and Castillo I P 2010 Phys. Rev. E 82 040104(R)
Wirtz T, Kieburg M and Guhr T 2015 EPL 109 20005
Perret A and Schehr G 2016 Random Matrices: Theory Appl. 05 1650001
Bornemann F 2016 Ann. Appl. Probab. 26 1942
Liu D-Z and Zhou D-S 2011 Int. Math. Res. Notices 2011 725
Giraud O 2007 J. Phys. A: Math. Theor. 40 F1053
Facchi P, Marzolino U, Parisi G, Pascazio S and Scardicchio A 2008 Phys. Rev. Lett. 101 050502
Vivo P, Pato M P and Oshanin G 2016 Phys. Rev. E 93 052106
Kubotani H, Adachi S and Toda M 2013 Phys. Rev. E 87 062921
Nadal C, Majumdar S N and Vergassola M 2010 Phys. Rev. Lett. 104 110501
Nadal C, Majumdar S N and Vergassola M 2011 J. Stat. Phys. 142 403
Haake F, Kuś M and Scharf R 1987 Z. Phys. B - Condensed Matter 65 381
Haake F 2010 Quantum Signatures of Chaos 3rd ed. (Springer-Verlag, Berlin)
Chaudhury S, Smith A, Anderson B E, Ghose S and Jessen P S 2009 Nature 461 768
Fujisaki H, Miyadera T and Tanaka A 2003 Phys. Rev. E 67 066201
Miller P A and Sarkar S 1999 Phys. Rev. E 60 1542
Srivastava S C L, Tomsovic S, Lakshminarayan A, Ketzmerick R and Bäcker A 2016 Phys. Rev.
Lett. 116 054101
Lakshminarayan A, Srivastava S C L, Ketzmerick R, Bäcker A and Tomsovic S 2016 Phys. Rev.
E 94 010205(R)
| 10math.ST
|
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
Domain Adaptation with L2 constraints for
classifying images from different endoscope systems
arXiv:1611.02443v2 [cs.CV] 2 Feb 2018
Toru Tamaki, Shoji Sonoyama, Takio Kurita, Tsubasa Hirakawa, Bisser Raytchev, Kazufumi Kaneda,
Tetsushi Koide, Shigeto Yoshida, Hiroshi Mieno, Shinji Tanaka, and Kazuaki Chayama
Abstract—This paper proposes a method for domain adaptation that extends the maximum margin domain transfer (MMDT)
proposed by Hoffman et al., by introducing L2 distance constraints between samples of different domains; thus, our method
is denoted as MMDTL2. Motivated by the differences between the
images taken by narrow band imaging (NBI) endoscopic devices,
we utilize different NBI devices as different domains and estimate
the transformations between samples of different domains, i.e.,
image samples taken by different NBI endoscope systems. We
first formulate the problem in the primal form, and then derive
the dual form with much lesser computational costs as compared
to the naive approach. From our experimental results using NBI
image datasets from two different NBI endoscopic devices, we
find that MMDTL2 is better than MMDT and also support
vector machines without adaptation, especially when NBI image
features are high-dimensional and the per-class training samples
are greater than 20.
Index Terms—Domain adaptation; Dual formulation; Kernels;
NBI endoscopy
I. I NTRODUCTION
In many hospitals, endoscopic examinations (i.e., colonoscopies) using narrow band imaging (NBI) systems are widely
performed to diagnose colorectal cancer [1], which is a major
cause of cancer deaths worldwide [2]. During examinations,
endoscopists observe and examine a polyp based on its visual
appearance, including via NBI magnification findings [3], [4],
as shown in Figure 1. To support proper diagnosis during
examinations, a computer-aided diagnostic system based on
the textural appearance of polyps would be helpful; thus,
numerous patch-based classification methods for endoscopic
images have been proposed [5]–[11].
This paper focuses on the inconsistencies between training
and test images. As with other frequently used machine
learning approaches, training classifiers assumes that the distribution of features extracted from both training and test
image datasets are the same; however, different endoscope
systems may be used to collect training and test datasets,
causing such an assumption to be violated. Further, given the
rapid development of medical devices (i.e., endoscopies in this
case), hospitals can introduce new endoscopes after training
images have already been taken. In addition, classifiers may
be trained with a training dataset collected by a certain type of
endoscope in one hospital, while another hospital might use
the same classifiers for images taken by a different endoscope.
In general, such inconsistencies lead to the deterioration of
classification performance; hence, collecting new images for a
new training dataset may be necessary or is at least preferred.
However, this is not the case with medical images. It is
Type A!
Microvessels are not observed or extremely opaque.
Type B!
Fine microvessels are observed around pits, and clear pits
can be observed via the nest of microvessels.
1
Microvessels comprise an irregular network, pits observed
via the microvessels are slightly non-distinct, and vessel
diameter or distribution is homogeneous.
Type C! 2
Microvessels comprise an irregular network, pits observed
via the microvessels are irregular, and vessel diameter or
distribution is heterogeneous.
3
Pits via the microvessels are invisible, irregular vessel
diameter is thick, or the vessel distribution is
heterogeneous, and a vascular areas are observed.
Fig. 1. NBI magnification findings [3].
impractical to collect enough sets of images for all types and
manufacturers of endoscopes. Active learning [12] deals with
small samples; however, this method is not helpful because it
selects small samples from unlabeled large training samples.
Figure 2 shows an example of differences between textures
captured by different endoscope systems. More specifically,
the images shown in Figures 2(a) and 2(b) are the same
scene from a printed sheet of a colorectal polyp image taken
by different endoscope systems at approximately the same
distance to the sheet from the endoscopes. Even for the same
manufacture (e.g., Olympus) and the same modality (e.g.,
NBI), images may differ in terms of factors such as resolution,
image quality, sharpness, brightness, and viewing angle. These
differences may impact classification performance.
To address this problem, Sonoyama et al. [13] proposed
a method based on transfer learning [14]–[17] to estimate a
transformation matrix between feature vectors of training and
test datasets captured by different (i.e., old and new) devices.
In this prior study, we formulated the problem as a constraint
optimization problem and developed an algorithm to estimate
a linear transformation; however, a key limitation is that
corresponding datasets are required, i.e., each test image (i.e.,
taken by a new device) must have a corresponding training
image (i.e., taken by an old device). Further, these images
must capture the same polyp to properly estimate the linear
transformation. These restrictions are rather strong, causing
our system to be somewhat impractical.
Therefore, this paper proposes an improved method for a
task that does not require image-by-image correspondences
between training and test datasets. More specifically, we
extend the transfer learning method proposed by Hoffman
et al. [21], [22], called maximum margin domain transfer
(MMDT). Their approach was a domain adaptation technique
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
(a)
(b)
Fig. 2. Example showing differences in appearance of images captured by
different endoscope systems: (a) an image taken by an older system (i.e.,
video system center: OLYMPUS EVIS LUCERA CV-260 [18]; endoscope:
OLYMPUS EVIS LUCERA COLONOVIDEOSCOPE CF-H260AZI [19]);
(b) an image of the same scene taken by a newer system (i.e., video
system center: OLYMPUS EVIS LUCERA ELITE CV-290 [20]; endoscope:
OLYMPUS CF-H260AZI [19]).
to handle the domain shift problem, in which distributions
of classes (or categories) in one domain, called the source,
change in another domain, called the target. This situation
occurs in various applications, and hence, domain adaptation
and transfer learning have already been widely studied.
Compared to previous studies, MMDT had the following
advantages: (1) applicability to multiclass problems with onevs-rest support vector machines (SVMs); (2) the ability to use
different feature dimensions in the source and target domains;
(3) the ability to handle unlabeled target samples by estimating
global class-independent affine transformation matrix W ; and
(4) scalability to the number of constraints, i.e., MMDT solves
KLs constraints as compared to Lt Ls in the previous studies,
where K is the number of classes and Ls and Lt are the
feature dimensions of the source and target domains.
In this paper1 , we therefore propose a non-trivial extension
to MMDT for handling inconsistencies between NBI endoscopic devices. We summarize the contributions of this paper
as follows.
•
•
First, we add L2 distance constraints to MMDT; our
method is thus called MMDTL2. The original formulation of MMDT uses the Frobenius norm of transformation
W as a regularizer; however, pulling transformation W
into a zero matrix is not intuitive and might not have a
good effect on transferring samples. Other regularizers
were discussed by [22], e.g., an identity matrix when
Ls = Lt , but no examples were given for cases of
Ls 6= Lt . For the latter cases, the target samples in one
category should be transformed into the same category as
that of the source domain. Therefore, we propose using
the L2 distances between the source and transformed
target samples to regularize W .
Second, we explicitly formulate MMDTL2 as a quadratic
programming (QP) problem. In [21], [22], the MMDT
problem was described, but not in QP form. In this paper,
we explicitly formulate MMDTL2 in the standard primal
1 A conference version of this paper was presented [23]. This paper extends
that version with more rigorous derivations, the compact form of the dual
formulation, and the kernelization of the method, as well as experiments from
different aspects.
2
QP form, which includes MMDT as a special case (i.e.,
where no L2 constraints are used).
• Third, we derive the dual QP problem of MMDTL2,
where the QP form mentioned above is the primal form.
The computational costs of MMDT and the primal QP
form of MMDTL2 can be very large for samples with
large dimensions because:
– for MMDT, affine transformation matrix W , which
is of size Ls × Lt (where Ls and Lt are the feature
dimensions of the source and target domains), can
be very large;
– for the primal QP form of MMDTL2, a matrix of size
Ls (Lt + 1) × Ls (Lt + 1) appears in the computation,
which is much larger than W .
In contrast, our derived dual QP form of MMDTL2 is
more scalable because it involves a KM × KM matrix,
where M is the number of target samples and K is
the number of classes. Typically, M is limited or much
smaller than the number of source samples, which is
common and therefore reasonable for domain adaptation
problems, including our problem of NBI device inconsistency.
• Finally, we show that the primal QP solution can be converted from the dual QP solution with much lower computational cost. With our derived formula, this conversion
needs matrices of size KM ×KM at most, while without
elaboration, a matrix of size Ls (Lt + 1) × Ls (Lt + 1)
appears in the conversion.
• In addition, we derive a kernelization of the dual QP
formulation, which enables us to use a nonlinear transformation as W for better performance.
Note that our dual form is different from the dual form
derived by Rodner et al. [24]. Their motivation was to make
MMDT scalable in terms of the number of target samples,
because in their study, they attempted to adapt large datasets
such as ImageNet. Therefore, their dual form still suffers from
the large feature dimensions of the source and target samples.
In contrast, our formulation is scalable in terms of feature
dimensions.
The rest of the paper is organized as follows. We formulate
problems of MMDT and MMDTL2 in Section 2, and then
derive the primal form in Section 3. In Section 4, we show the
dual form, and in Section 5, we obtain the primal solution from
the dual solution. In Section 6, we present our experimental results using datasets of actual NBI endoscopic images. Finally,
in Section 7, we conclude this paper and provide avenues for
future work.
II. P ROBLEM FORMULATION
In this section, we introduce the problems of MMDT [22]
and our proposed MMDTL2.
Problem 1 (MMDT). Suppose we are given a training set
Ls
χs = {xsn , yns }N
× {1, 2, . . . , K} in the source
n=1 ⊂ R
t M
domain and another set χt = {xtm , ym
}m=1 ⊂ RLt ×
{1, 2, . . . , K} in the target domain for a K-class classification
problem.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
3
MMDT solves the following optimization problem:
1
1
min kW k2F + kΘ̂k2F + ct L(W, Θ̂, χt ) + cs L(Θ̂, χs ), (1)
2
W,Θ̂ 2
where ct and cs are weights and
X
W x̂tm
t
L(W, Θ̂, χt ) =
max 0, 1 − ykm
θ̂kT
1
and
1
min kΘ̂k2F + ct L(W, Θ̂, χt ) + cs L(Θ̂, χs )
2
Θ̂
with the same initialization as problem 2, where
(2)
cd L(W, χs , χt ) =
k,m
and
s
L(Θ̂, χ ) =
X
max 0, 1 −
s
ykn
θ̂kT x̂sn
(8)
(3)
k,n
are hinge loss functions. Here, θ̂k ∈ RLs +1 is an SVM
hyperplane parameter (including weights θk ∈ RLs and bias
bk ∈ R, i.e., θ̂k = (θkT , bk )T ) for the kth class stacked into a
t
t
matrix Θ̂ = (θ̂1 , θ̂2 , . . . , θ̂K ), ykm
= 2δ(ym
, k)−1 ∈ {−1, 1}
is a label in terms of the kth hyperplane, and ct and cs are
weights.
Note that x̂ denotes
an augmented vector with 1 as its last
x
element, i.e., x̂ =
.
1
In this problem, we simultaneously obtain K SVM classifiers and transformation W ∈ RLs ×Lt . One-vs-all SVMs
are used for multiclass classification; thus, K hyperplane
parameters θ̂k are obtained in the source domain. Target
samples xtm are transformed by W from the target domain
to the source domain, and then the loss function causes them
to be classified by the SVMs.
Because this problem is non-convex, an alternating optimization approach was used in [22].
Problem 2 (MMDT with iteration). MMDT solves problem 1
by iteratively solving subproblems
1
(4)
min kW k2F + ct L(W, Θ̂, χt ),
W 2
and
1
min kΘ̂k2F + ct L(W, Θ̂, χt ) + cs L(Θ̂, χs ),
(5)
Θ̂ 2
Problem 3 (MMDTL2). MMDT with L2 constraints solves
problem 1 by iteratively solving subproblems
1
min cf kW k2F + ct L(W, Θ̂, χt ) + cd L(W, χs , χt ), (7)
W 2
(9)
t
where snm = cd δ(yns , ym
) is a weight between samples xtm ∈
Lt
s
Ls
R and xn ∈ R and cd is the balance weight.
MMDTL2 reduces to MMDT when cf = 1 and cd = 0.
We add weight cf to the first term of Eq. (7) because ct is
used for both Eqs. (7) and (8). In both equations, ct modifies
the balance between the first and second terms. Therefore,
without cf , the first terms of Eqs. (7) and (8) have the same
weight, which might not be suitable in general. We will
investigate the effect of cf in section 4.4.
The SVM subproblem (8) can be solved by common SVM
solvers, as is done for (5) by MMDT. Therefore, in the
following sections, we focus on deriving the primal and dual
forms of subproblem (7) as standard QP problems.
III. P ROPOSED METHOD
In this section, we highlight the main results only. Full
derivations and proofs are given in the appendix.
A. Primal problem
First, we rephrase subproblem (7) with inequality constraints instead of loss functions.
Problem 4 (Estimation of W ). We want to find W ∈
RLs ×(Lt +1) that minimizes the following objective function:
K X
M
X
1
t
cf kW k2F + cT
ξkm
W,{ξkm } 2
m=1
min
t
k=1
+
initializing Θ̂ with
1
(6)
arg min kΘ̂k2F + cs L(Θ̂, χs ).
Θ̂ 2
As noted in the Introduction, the use of Frobenius norm
kW k2F is not intuitive for the transformation matrix. Further,
it might not be a good choice for small values of ct because
the obtained solution is pulled toward a zero matrix; however,
the use of large values of ct impacts the SVM subproblem (5)
because C-SVM solvers are known to be unstable for large
values of parameter C.
In this paper, we therefore propose the following problem,
which we call MMDTL2. MMDTL2 incorporates additional
constraints of L2 distances to pull target samples to source
samples of the same category.
M
N
1 XX
snm kW x̂tm − xsn k22 ,
2 m=1 n=1
s.t.
M
X
N
X
1
snm kW x̂tm − xsn k22
(10)
2 m=1 n=1
W x̂tm
T
t
t
t
ξkm ≥ 0, ykm θ̂k
− 1 + ξkm
≥ 0.
1
(11)
B. Primal QP problem
The problem can be written in the form of a canonical QP
problem.
Lemma 1. The problem can be written as
1 T T V 0
1
w
w
T
T
min (w , ξ )
+ (−q , cT 1KM )
+ s
0 0
ξ
ξ
w,ξ 2
2
(12)
0
0
IKM
w
s.t.
≥
,
(13)
ξ
Y ΦT IKM
1KM − Y b̃
where variables are defined in the proof.
This primal QP problem involves very large matrices that
are impractical to compute. More precisely, V is a matrix of
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
4
size Ls (Lt + 1) × Ls (Lt + 1), which can be very large when
the dimensions of features (Ls and Lt ) are large. Next, we
therefore derive the dual form of the problem, which we expect
to be less expensive to compute.
C. Dual problem
The dual of the problem is given in the corollary below.
Corollary 1. The dual form of the original primal problem is
given by
max
a
−1 T T
a Y ((ΘT Θ) ⊗ G)Y a
2
T
+ (1T − (Y b̃)T − vec ΘT X s SG Y )a
s.t.
cT 1 ≥ a ≥ 0.
(14)
(15)
D. Retrieving the primal solution
After solving the dual problem with a QP solver, we need
to convert the dual solution a to w and b by
w = V −1 (q + ΦY a),
(16)
then finally to W .
Here, we again face the problem of large matrix V , that
should not be used. In the corollary below, we show primal
solution W directly, i.e., avoiding conversions from a to w,
and then to W .
Corollary 2. The solution to the primal problem is given by
W = X s S + Θ(Υ Λ)T (X t )T A−1 ,
(17)
where
is element-wise multiplication and variables are
given in the proof.
E. Kernelization
We derive the kernel version of the dual formulation. To do
so, we apply W to target sample xt by multiplying it from the
left, i.e., W x̂t ; Therefore, all computations with target samples
are inner products, which means we can use kernels to replace
the inner products.
To transform target sample xt with W , we have
W x̂t = X s S + Θ(Υ Λ)T
! k(xt1 , xt )
1
1
1
..
−1
IM − 2 K t (SM
+ K t )−1
.
.
cf
cf
cf
t
t
k(xM , x )
(18)
IV. R ESULTS AND DISCUSSIONS
In this section, we show experimental results by using
datasets consisting of two NBI endoscopic devices to understand how our proposed method behaves given differing
numbers of target domain training samples.
We used the following parameter values throughout the
experiments unless stated otherwise: cf = ct = cs = cd = 0.1
for MMDTL2 based on our preliminary experiments, and
cs = 0.05 and ct = 1 for MMDT, as used in the code of
the original MMDT [21], [22], [25].
A. NBI endoscopic image dataset
The NBI dataset used in this experiment consisted of two
domains. For the first (i.e., source) domain, we used the NBI
image dataset consisting of 908 NBI patches collected from
endoscopic examinations at Hiroshima University by using
OLYMPUS EVIS LUCERA endoscope system [18]; patches
were labeled based on NBI magnification findings [3], [4],
which categorizes appearances of tumors into types A, B,
and C, with type C further sub-classified into C1, C2, and
C3 based on microvessel structures (see Figure 1). In this
study, we used only types A, B, and C3 in accordance with
our previous work [11], [13], [26], [27]. In general, a patch
is trimmed from a larger frame of the entire endoscopic
image such that the trimmed rectangular region represents
the typical texture pattern of the colorectal polyp appearing
in the frame. To align the size of patches between source
and target domains, we further trimmed the center of each
patch to 180 × 180 and created 734 patches with 289 in
type A, 365 in type B, and 80 in type C3. Note that this
study was conducted with the approval from the Hiroshima
University Hospital ethics committee, and informed consents
were obtained from the patients and/or family members for
the endoscopic examinations.
For the second (i.e., target) domain, we used another NBI
image dataset consisting of 279 patches with 92 in type A
and 187 in type B. These images were taken by OLYMPUS
EVIS LUCERA ELITE endoscope system [20], which is a
newer model than that of the system of the source domain.
Due to the limited number of endoscopic examinations using
this newer endoscope system, we trimmed the center square to
180 × 180 from video frames of 41 NBI examination videos;
hence, there are two factors of domain shift here: (1) the
NBI endoscopic devices and (2) the differences between still
images (i.e., source) and video frames (i.e., target). Note that
type C3 samples obtained using the new endoscopy are not
yet available because the number of type C3 samples, which
corresponds to a developed cancer [11], is smaller in general,
and the new endoscopy of our facility has not yet had the
opportunity to capture an image of a developed cancer of this
type.
From these two domains, we computed convolutional neural
network (CNN) features extracted using CaffeNet [28]; more
specifically, this is the fc6 feature of 4096 dimensions, which
is known to work well for many tasks [29]. These features
were used without dimensionality reduction.
To see the effect of the number of target samples, we
prepared training/test sets as follows. First, we randomly split
the source and target domain samples into half; we therefore
had a source training set of 367 source samples and a target
training set of 139 target samples. Next, we kept a specified
number of target samples per category (up to 40) in each of
the target training sets, discarding the rest. For the test set, we
used 140 target samples. We created 10 training/test sets and
reported average performance.
Figure 3 shows performance results of different methods
over the different numbers of target training samples. As a
baseline, we also show two SVM results, i.e., source SVM
5
100
100
90
90
80
80
accuracy
accuracy
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
70
60
60
sourceSVM
targetSVM
MMDTL2 (linear)
MMDTL2 (RBF)
MMDTL2 (poly)
MMDT
50
40
70
0
10
20
number of target training samples per category
30
40
Fig. 3. Experimental results for the NBI datasets with the fc6 feature.
and target SVM. Source SVM estimates an SVM classifier
with source and target training samples (and only source
samples when no target samples are used). Target SVM does
the same, but only with target training samples. MMDT results
were obtained using the code provided by [25], just as in
our first experiment. There were three results for MMDTL2,
i.e., a linear version and two kernel versions with RBF and
polynomial kernels.
Even with no target samples, source SVM performed relatively better and increased its performance as the number of
target training samples increased. Target SVM started below
70%, but caught up to source SVM when 20 training target
samples were given. The results indicate that our proposed
MMDTL2 behaves between source and target SVMs. When
one or two target samples are given, MMDTL2 behaves
similarly to target SVM and far below the source SVM, but it
quickly gains from the benefits of adaptation. With the RBF
kernel, MMDTL2 is the best when 10 and 15 training samples
are given, but the linear and polynomial kernels become better
when more samples are given. MMDTL2 with the linear kernel
and target SVM approach one another, which is expected
because a sufficient number of target training samples are
considered to be the best for classifying target domain samples.
Overall, MMDTL2 with a polynomial kernel works the best.
To make the discussion above more quantitative, we performed the one-tailed Welch’s unequal variances t-test [30],
[31] to show the statistical significance of the proposed method
with respect to sourceSVM and targetSVM. Table I shows
the same performance results presented in Figure 3, but with
results of the statistical test. Tests with 1% and 5% significance
levels (i.e., p < 0.01 and p < 0.05) are indicated by ∗∗ and
∗, respectively, with respect to sourceSVM. Similarly, ++
and + indicate the significance of test results with respect
to targetSVM. MMDTL2 with a polynomial kernel is better
than sourceSVM and targetSVM when M ≥ 20 (M is is
the number of target samples given per category). MMDTL2
with the RBF kernel works well only when 20 ≥ M ≥ 5.
MMDTL2 is better than sourceSVM when M ≥ 20.
50
40
sourceSVM
targetSVM
MMDTL2 (linear)
MMDTL2 (RBF)
MMDTL2 (poly)
0
10
20
number of target training samples per category
30
40
Fig. 4. Experimental results for the NBI datasets with the conv3 feature.
B. NBI endoscopic image dataset with high-dimension features
Figure 4 shows performance results of the given methods
as in Figure 3 with the same protocols for training and
evaluation. The difference here is the set of features used; more
specifically, we use the conv3 features of 64,896 dimensions
rather than fc6. For NBI patch classification problems [32], we
have seen that conv3 features worked better than fc6 features.
Therefore we choose conv3 features for this experiment.
However, transformation matrix W for the conv3 features
could be very large without our efficient dual formulation. The
ability to handle such large dimensions of features is the key
advantage of our proposed dual MMDTL2. In contrast, it is
not possible to train MMDT because it involves a matrix W of
size 64, 896 × 64, 897 for storage (approximately 31 GB for
double-type elements) and requires more space for working
memory. The test phase is also impractical because MMDT
uses W expertly for converting target samples. Figure 4 shows
that the sourceSVM is the best when a few target samples are
available, just as in Figure 3. Further, our proposed MMDTL2
with a polynomial kernel becomes better at the right half of the
plot, and the differences between sourceSVM and targetSVM
are much more significant than those in Figure 3.
Table II shows the same performance results as shown in
Figure 4 along with the results of the one-tailed Welch’s t-test
[30], [31]. The meaning of the marks is the same as that in
Table I. Again, MMDTL2 with a polynomial kernel is better
than sourceSVM and targetSVM when M ≥ 20. MDTL2
with the linear kernel becomes better than sourceSVM and
targetSVM when M ≥ 35.
C. NBI endoscopic image dataset with features of different
dimensions
Figure 5 shows another set of performance results. Here,
we use different feature dimensions for source and target
samples. In particular, we use the conv3 features of 64,896
dimensions for the source samples and the conv5 features of
43,264 dimensions for the target samples. In this case, MMDT
does not work because of its high memory cost as described in
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
6
TABLE I
E XPERIMENTAL RESULTS FOR THE NBI DATASETS WITH THE FC 6 FEATURE .
1
2
3
4
5
10
15
20
25
30
35
40
sourceSVM
77.31±1.05
78.01±1.34
80.50±1.67
82.69±1.94
84.33±1.69
89.15±0.97
91.21±0.64
91.77±0.61
91.99±0.59
92.34±0.56
92.62±0.62
93.26±0.57
++
++
++
++
++
++
targetSVM
69.01±5.90
61.99±3.52
67.30±4.98
71.21±5.98
77.80±3.39
85.67±1.83
90.21±1.39
92.41±0.93
92.77±0.78
93.48±0.63
93.97±0.53
94.61±0.35
*
**
**
**
MMDT
68.30±3.10
79.65±4.22
75.67±5.11
76.38±4.92
77.87±4.17
77.23±3.52
80.00±3.06
87.73±2.49
86.74±3.23
87.09±2.23
91.21±0.89
90.07±1.05
++
++
MMDTL2
(linear)
70.28±5.31
68.65±3.18
72.13±5.01
74.54±5.80
81.91±2.68
88.09±1.73
91.21±1.33
93.55±0.74
93.12±0.69
93.76±0.30
94.26±0.48
94.96±0.48
++
+
+
**+
**
**
**
**
MMDTL2
(RBF)
65.60±7.46
59.65±4.16
70.07±5.63
73.12±6.38
81.99±2.69
90.07±2.22
93.40±0.89
92.84±0.58
92.55±0.70
92.91±0.76
92.91±0.63
92.62±0.65
+
++
**++
**
MMDTL2
(poly)
70.99±4.81
71.28±2.96
73.76±4.91
75.96±5.38
82.62±2.58
88.09±1.76
91.56±1.02
93.90±0.69
93.90±0.66
94.26±0.40
95.18±0.48
95.96±0.34
++
+
++
+
**++
**++
**++
**++
**++
TABLE II
E XPERIMENTAL RESULTS FOR THE NBI DATASETS WITH THE CONV 3 FEATURE .
sourceSVM
83.97±1.25
83.76±1.02
84.75±0.84
85.82±1.12
86.74±0.96
89.72±0.78
90.78±0.83
91.63±0.78
92.48±0.64
93.12±0.83
93.40±0.69
93.76±0.68
1
2
3
4
5
10
15
20
25
30
35
40
targetSVM
55.18±6.74
50.50±3.98
56.38±4.13
63.33±5.08
68.09±3.39
77.02±1.53
86.03±1.01
89.15±1.14
90.71±1.14
92.06±1.19
93.62±0.90
94.61±0.69
++
++
++
++
++
++
++
++
++
MMDTL2
(linear)
57.23±5.67
56.95±4.39
62.13±3.84
69.01±5.06
75.60±2.75
81.84±1.65
88.44±1.09
91.13±1.10
92.27±0.91
93.19±0.91
94.68±0.68
95.46±0.69
100
90
accuracy
80
70
60
50
40
targetSVM
MMDTL2 (linear)
MMDTL2 (RBF)
MMDTL2 (poly)
0
10
20
number of target training samples per category
30
40
Fig. 5. Experimental results for the NBI datasets with the different features
in source and target domains.
the previous section, and furthermore, sourceSVM cannot be
used because of the difference of feature dimensions. Table III
shows the same performance results along with the results of
the one-tailed Welch’s t-test [30], [31] using the same marks
as those in Table I. MMDTL2 with a polynomial kernel is
better than targetSVM again, when n ≥ 10. MMDTL2 with
the linear and RBF kernels becomes better than targetSVM
when 25 ≥ n ≥ 10.
D. Effects of parameter values
We introduced parameters cf and cd for our MMDTL2
formulation in Section 2. Figure 6 shows performance results
++
++
+
++
++
++
++
++
**+
**+
MMDTL2
(RBF)
52.41±7.46
52.27±4.43
60.28±4.04
68.94±5.42
78.94±2.64
90.28±2.41
92.62±1.06
92.70±0.50
91.63±0.98
94.04±0.92
92.55±1.03
93.40±1.00
++
++
**++
**++
++
MMDTL2
(poly)
51.70±6.60
52.84±4.39
60.07±4.04
68.09±5.26
75.67±2.89
84.54±1.74
91.28±1.07
93.69±0.91
94.18±0.88
94.75±0.88
96.03±0.62
96.52±0.52
++
++
++
**++
**++
**++
**++
**++
for different values of cf , whereas other parameters are fixed
to their default values. The results are almost identical when
cf > 10−7 ; however, the performance becomes unsatisfactory
for smaller values of cf . This is because the inverse of cf
is used in the formulation of MMDTL2 (see the Appendix).
This suggests that a small amount of regularization for kW kF
might help improve performance. Note that the instability of
the results when cf ≤ 10−7 may be due to the fact that
variables of single precision have 8 digits of precision, and
the inverses of smaller values lead to numerical instability.
Therefore, we can say that results are insensitive to the choice
of cf as long as larger values are used.
Figure 7 shows the performance results for different values
of cd , whereas other parameters are fixed to their default
values. To see the differences between the different values,
the results for 20 target samples are shown in Figure 8. The
results are almost identical when cd ≥ 10−6 ; however, the
performance decreases for 10−7 ≥ cd > 10−10 . A detailed
investigation of this effect is one of our future tasks; however,
it is clear that the performance is better and stable for larger
values of cd .
E. Computation time
Figure 9 shows the computation time for the training phases
of each method over different feature dimensions (Ls = Lt )
using MATLAB implementations on an Intel Core i7 3.4 GHz
CPU with 16 GB memory. The primal MMDTL2 involves
quite a large matrix for V , and it is not practical when Ls >
180 in terms of memory and computation time. The cost of
the dual MMDTL2 depends on the number of target samples;
however, it is quite fast compared to the primal MMDTL2 for
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
7
TABLE III
E XPERIMENTAL RESULTS FOR THE NBI DATASETS WITH THE DIFFERENT FEATURES IN SOURCE AND TARGET DOMAINS .
targetSVM
1
2
3
4
5
10
15
20
25
30
35
40
64.33±5.86
55.67±4.81
55.39±4.40
64.18±5.97
69.22±4.51
79.93±2.21
86.24±1.41
88.65±1.16
89.43±1.17
91.28±1.32
92.48±0.69
93.55±0.55
MMDTL2
(linear)
68.30±5.39
61.56±4.70
61.13±4.83
68.30±5.90
72.84±3.91
83.40±2.28
88.01±1.28
90.07±1.21
90.78±1.21
92.13±1.14
92.91±0.73
93.76±0.57
+
+
++
+
+
+
MMDTL2
(RBF)
48.44±5.04
47.45±5.69
55.18±5.46
64.33±7.13
75.53±3.68
89.65±2.34
92.98±1.28
93.33±0.75
93.19±0.82
92.91±0.66
92.84±0.45
92.84±0.84
++
++
++
++
++
++
MMDTL2
(poly)
50.43±5.32
47.66±5.68
54.54±5.32
63.62±7.03
70.85±4.46
83.69±2.19
88.87±1.35
91.63±0.98
92.48±0.89
93.19±0.99
94.04±0.45
94.61±0.48
++
++
++
++
++
++
++
100
100
90
95
accuracy
accuracy
80
70
85
60
Cf = 1e − 10
Cf = 1e − 9
Cf = 1e − 8
Cf = 1e − 7
Cf = 1e − 6
Cf = 0. 1
Cf = 1000
50
40
90
0
10
20
number of target training samples per category
30
80
10 -19 10 -17 10 -15 10 -13 10 -11 10 -9
cd
40
10 -7
10 -5
10 -3
10 -1
10 1
10 3
Fig. 8. Effect of cd for the NBI datasets with the fc6 feature and 20 training
samples.
Fig. 6. Effect of cf for the NBI datasets with the fc6 feature.
100
MMDT (10)
MMDTL2 primal (10)
MMDTL2 dual(10)
MMDTL2 dual (40)
10 2
accuracy
80
70
60
Cd = 1e − 20
Cd = 1e − 10
Cd = 1e − 9
Cd = 1e − 8
Cd = 1e − 7
Cd = 5e − 7
Cd = 1e − 6
Cd = 1e − 4
Cd = 0. 1
Cd = 1000
50
40
0
10
20
number of target training samples per category
30
40
Fig. 7. Effect of cd for the NBI datasets with the fc6 feature.
both 10 and 40 target training samples. The dual MMDTL2
can deal with much higher feature dimensions, while MMDT
exceeds the memory limitation when Ls > 10000 because
matrix W must be explicitly stored.
V. C ONCLUSIONS
In this paper, we proposed MMDT with L2 constraints, i.e.,
MMDTL2, deriving the dual formulation with much lesser
computational costs as compared to the naive QP problem.
computation time [sec]
90
10 1
10 0
10 -1
10 1
10 2
10 3
feature dimension Ls = Lt
10 4
Fig. 9. Computation time. The numbers of target samples used are shown in
braces.
Further, we showed the kernelization of our method. Experimental results with NBI datasets from two different endoscopic
devices showed that our proposed MMDTL2 with linear and
polynomial kernels performed better than the given baselines
(i.e., source and target SVMs). Our future work includes using
other loss functions for problem formulation. We observed
that the one-vs-rest multiclass classification by SVMs was
a performance bottleneck of MMDTL2 in our experiments.
Therefore, instead of relying on maximum margin loss func-
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
8
tions, multiclass logistic loss might be better here. In the
future, we plan to explore this idea and report performance
results for the NBI dataset as well. In addition, we plan to
investigate parameter tuning with cross validation and compare
the proposed method with other adaptation methods such as
that in [33].
A PPENDIX A
P RIMAL PROBLEM
In this section, we rephrase subproblem (7) with inequality
constraints instead of loss functions.
Problem 4 (Estimation of W ). We want to find W ∈
RLs ×(Lt +1) that minimizes the following objective function:
K X
M
X
1
t
cf kW k2F + cT
ξkm
W,{ξkm } 2
m=1
min
t
k=1
+
M
N
1 XX
snm kW x̂tm − xsn k22
2 m=1 n=1
Proof. First, we
T
w1 x̂
w1T x̂
W x̂ = .
..
T
wL
x̂T
s
have
x̂T w1
x̂T w2
= ..
.
T
T
x̂
=
(20)
(21)
First, we rewrite the objective function in a matrix form.
To this end, we introduce the vec operator and some formulas
below.
A. Operator vec
Here, we define a vectorized operator for rearranging
matrix-vector products.
Definition 1. For a given matrix W ∈ RLs ×(Lt +1) , denoted
by a set of row vectors wi ∈ RLt as
T
w1
w2T
W = . ,
(22)
..
T
wL
s
we define operator vec, which vectorizes W in the row-major
order as
w1
w2
vec(W ) = . ∈ RLs (Lt +1) .
(23)
..
w1
w2
..
.
x̂T
..
.
wLs
x̂T
(27)
x̂ wLs
= (ILs ⊗ x̂T )w.
(28)
Using this equation, we have
x̂
x̂
x̂T W T W x̂ = wT
..
.
(19)
s.t.
t
ξkm
≥ 0,
W x̂tm
t
T
t
ykm θ̂k
− 1 + ξkm
≥ 0.
1
Here, w = vec(W ), ILs ∈ RLs ×Ls , is an identity matrix and
⊗ is the tensor product.
T
x̂
x̂T
..
x̂x̂T
..
.
x̂x̂T
T
w
T
= w (ILs ⊗ x̂x̂ )w.
Also, we have
T
x̂
z T W x̂ = z T
.
x̂T
(29)
x̂
T
x̂x̂
= wT
w
(30)
(31)
x̂T
..
T
T
w = (z1 x̂ , z2 x̂ , . . .)w
.
(32)
T
z1 x̂T
z2 x̂T
T T
T
T
= vec
w = vec(z x̂ ) w = w vec(z x̂ ).
..
.
(33)
For later use, we also define
U (x̂) = (ILs ⊗ x̂x̂T ).
(34)
wLs
This definition is different from the one used in the literature, which is defined in the column-major order, for example,
in [34].
Next, we can rewrite matrix-vector multiplications using the
vec operator, as summarized in the following lemma.
Lemma 1. For given matrix W ∈ RLs ×(Lt +1) and vectors
x ∈ RLt +1 and z ∈ RLs , the following equations hold:
W x̂ = (ILs ⊗ x̂T )w
T
T
T
(24)
T
x̂ W W x̂ = w (ILs ⊗ x̂x̂ )w
T
T T
z W x̂ = vec(z x̂ ) w
(25)
(26)
B. Rewriting terms with vec operator
In this subsection, we rewrite the L2 term in the objective
function using the lemma below.
Lemma 2. The L2 constraint term of MMDTL2 can be written
as
M
N
1 XX
1
snm kW x̂tm − xsn k22 = (wT U w − 2q T w + s),
2 m=1 n=1
2
(35)
where U , q, and s are given in the proof below.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
9
Proof. A single L2 term can be rewritten with lemma 1 as
kW x̂tm − xsn k22 = (W x̂tm − xsn )T (W x̂tm − xsn )
and weights
(36)
= (x̂tm )T W T W x̂tm − 2(xsn )T W x̂tm + kxsn k22
(37)
= wT U (x̂tm )w − 2vec(xsn (x̂tm )T )T w + kxsn k22
(38)
T
kxsn k22 ,
(39)
Um = U (x̂tm ) = (ILs ⊗ x̂tm (x̂tm )T )
(40)
= w Um w −
T
2qnm
w
+
s11
..
S= .
···
sN 1
···
sm =
N
X
snm ,
s1M
.. ,
.
(54)
sN M
sn =
n=1
M
X
snm ,
m=1
SM = diag(s1 , . . . , sm , . . . , sM ).
(55)
where
Next, we rewrite the conditions in the problem as shown
below.
and
qnm = vec(xsn (x̂tm )T ).
(41)
Lemma 3. The condition in problem 4 can be written as
By summing the terms with weights, we have
t
t
ykm
(φTkm w + bk ) − 1 + ξkm
≥ 0,
N
M
1 XX
T
snm (wT Um w − 2qnm
w + kxsn k22 )
(42)
2 m=1 n=1
!
N
M X
X
1
= w
snm Um w
2
m=1 n=1
!
M X
N
M
N
X
1 XX
T
−
snm qnm w +
snm kxsn k22 (43)
2
m=1 n=1
m=1 n=1
1
= (wT U w − 2q T w + s).
2
(44)
(56)
where φkm = vec(θk (x̂tm )T ).
Proof.
t
ykm
θ̂kT
W x̂tm
1
t
− 1 + ξkm
≥0
(57)
t
t
ykm
(θkT W x̂tm + bk ) − 1 + ξkm
≥0
(58)
t
ykm
(vec(θk (x̂tm )T )T w
t
ykm
(φTkm w
+ bk ) − 1 +
+ bk ) − 1 +
t
ξkm
t
ξkm
≥0
(59)
≥0
(60)
Here, U , q, and s are the corresponding factors; we further
rewrite them into the simpler forms shown below.
U=
M X
N
X
snm Um =
m=1 n=1
M
X
M
X
sm Um
(45)
C. Primal QP problem
m=1
In this subsection, we write the problem in the form of a
canonical
QP problem.
=
sm (ILs ⊗ x̂tm (x̂tm )T ) = (ILs ⊗
sm x̂tm (x̂tm )T )
m=1
m=1
Lemma 4. Problem 4 can be written as
(46)
1 T T V 0
1
w
w
= (ILs ⊗ X t SM (X t )T )
(47)
T
T
min (w , ξ )
+ (−q , cT 1KM )
+ s
0
0
ξ
ξ
w,ξ 2
2
M X
N
M X
N
X
X
(61)
q=
snm qnm =
snm vec(xsn (x̂tm )T ) (48)
m=1 n=1
m=1 n=1
s.t.
M X
N
X
0
0
IKM
w
= vec(
snm xsn (x̂tm )T )
(49)
≥
,
(62)
ξ
Y ΦT IKM
1KM − Y b̃
m=1 n=1
M
X
= vec(X s S(X t )T )
s=
M X
N
X
snm kxsn k22 =
m=1 n=1
(50)
N
X
sn kxsn k22
(51)
n=1
Here, we use data matrices
s
X =
X =
(x̂t1 , x̂t2 , . . . , x̂tM )
Proof. First, we define two matrices V ∈ RLs (Lt +1)×Ls (Lt +1)
and A ∈ R(Lt +1)×(Lt +1) as follows:
V = cf ILs (Lt +1) + U
(xs1 , xs2 , . . . , xsN )
∈R
Ls ×N
(52)
and
t
where variables are defined in the proof below.
∈R
(Lt +1)×M
(53)
(63)
t
t T
= cf ILs (Lt +1) + (ILs ⊗ X SM (X ) )
(64)
= ILs ⊗ (cf ILt +1 + X t SM (X t )T )
(65)
= ILs ⊗ A
(66)
t
t T
A = cf ILt +1 + X SM (X )
(67)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
10
Then, we rewrite the objective function (19) as
1
cf kwk22 + cT
2
Proof. The Lagrangian of problem (61) is given by
K X
M
X
k=1
1
t
ξkm
+ (wT U w − 2q T w + s)
2
m=1
(68)
1 T
1
= w V w − q T w + s + cT 1TKM ξ
(69)
2
2
1
1
V 0
w
w
= (wT , ξ T )
+ (−q T , cT 1TKM )
+ s,
0
0
ξ
ξ
2
2
(70)
where
L=
K X
M
X
1
1 T
t
w V w − q T w + s + cT
ξkm
2
2
m=1
k=1
−
K X
M
X
t
µkm ξkm
−
k=1 m=1
K X
M
X
t
t
akm (ykm
(φTkm w + bk ) − 1 + ξkm
),
k=1 m=1
(81)
where akm ≥ 0 and µkm ≥ 0 are Lagrange multipliers.
ξ=
t
t
t
t
t
(ξ11
, ξ12
, . . . , ξ1M
, ξ21
, . . . , ξKM
)T .
(71)
To rewrite conditions (20) and (21), we turn these constraints into vector form with a generalized inequality. The
first constraint, i.e., (20), can be written as follows:
0
ξ≥0
w
IKM
≥ 0.
ξ
(72)
(73)
The second constraint, i.e., (21), is
Y (ΦT w + b̃) − 1KM + ξ ≥ 0
w
(Y ΦT , IKM )
≥ 1KM − Y b̃,
ξ
(74)
To simplify the derivation, we convert it into vector form
1
1
L = wT V w − q T w + s + cT 1TKM ξ − µT ξ
2
2
− aT (Y (ΦT w + b̃) − 1KM + ξ)
(82)
1 T
= w V w − (q + ΦY a)T w + (cT 1KM − µ − a)T ξ
2
1
+ aT (1KM − Y b̃) + s,
(83)
2
where
(75)
where 1KM is a vector of KM ones and
b = (b1 , b2 , . . . , bK )T
(84)
µ = (µ11 , µ12 , µ1M , µ21 , . . . , µKM )T ,
(85)
(76)
T
b̃ = (b ⊗ 1M ) = (b1 , b1 , . . . , b1 , b2 , . . . , bK )
(77)
Φ = (φ11 , φ12 , . . . , φ1M , φ21 , . . . , φKM )
(78)
Y =
a = (a11 , a12 , a1M , a21 , . . . , aKM )T
t
t
t
t
t
diag(y11
, y12
, . . . , y1M
, y21
, . . . , yKM
).
(79)
By combining these inequalities (73) and (75), we have the
conditions in a single form, as claimed.
This primal QP problem involves very large matrices that
are impractical to compute. More precisely, V is a matrix of
size Ls (Lt + 1) × Ls (Lt + 1), which can be very large when
the dimensions of features (Ls and Lt ) are large. In the next
section, we therefore derive the dual form of the problem,
which we expect to be less expensive to compute.
and
with a ≥ 0 and µ ≥ 0.
Next, we take the derivatives of the Lagrangian as follows.
For w, we have
∂L
= V w − (q + ΦY a) = 0
∂w
V w = q + ΦY a
w=V
−1
(q + ΦY a).
(86)
(87)
(88)
A PPENDIX B
D UAL PROBLEM
In this section, we derive the dual of the problem.
A. Lagrangian
Lemma 5 (Lagrangian). The Lagrangian of problem (61) is
given by
For ξ, we have
∂L
= cT 1 − µ − a = 0
∂ξ
cT 1 − a = µ ≥ 0
cT 1 − a ≥ 0
1
L = − aT Y ΦT V −1 ΦY a
2
T
T
T
+ (1 − b̃ Y − q V
−1
1
1
ΦY )a − q T V −1 q + s.
2
2
(80)
cT 1 ≥ a ≥ 0
for µ ≥ 0 and a ≥ 0.
(89)
(90)
(91)
(92)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
11
By incorporating w and cT into the Lagrangian and using
V T = V , we have
1
L = wT V w − (q + ΦY a)T w + (cT 1 − µ − a)T ξ
2
1
+ aT (1 − Y b̃) + s
2
1
= (V −1 (q + ΦY a))T V (V −1 (q + ΦY a))
2
(93)
1
− (q + ΦY a)T (V −1 (q + ΦY a)) + aT (1 − Y b̃) + s
2
(94)
1
1
= − (q + ΦY a)T V −1 (q + ΦY a) + aT (1 − Y b̃) + s
2
2
(95)
1 T −1
= − (q V q + 2q T V −1 ΦY a + aT Y ΦT V −1 ΦY a)
2
1
+ aT (1 − Y b̃) + s
(96)
2
1
= − aT Y ΦT V −1 ΦY a + (1T − b̃T Y − q T V −1 ΦY )a
2
1
1
− q T V −1 q + s.
(97)
2
2
Proof.
vec(abT )T V −1 vec(cdT )
−1
A
T
T
= (a1 b , a2 b , . . .)
(102)
A−1
c1 d
c2 d
..
..
.
.
c1 A−1 d
−1
= (a1 bT , a2 bT , . . .) c2 A d
..
.
X
T −1
T
=
ai ci b A d = (a c)bT A−1 d
Lemma 7. Given matrix Φ ∈ RLs (Lt +1)×KM , we have
ΦT V −1 Φ = (ΘT Θ) ⊗ G,
where Θ = (θ1 , θ2 , . . . , θK ) ∈ R
latter given in the proof below.
Ls×K
and G ∈ R
A
t
(98)
= (cf ILt +1 + X SM (X ) )
(99)
1
1 t −1
1
= ILt +1 − 2 X (SM + (X t )T X t )−1 (X t )T .
cf
cf
cf
(100)
Note that the second form of A−1 is obtained using Woodbury’s formula only if SM is non-singular and cf 6= 0; this is
usually the case, because diagonal elements of SM are sums
of (non-negative) weights.
Lemma 6. Given V of the structure above and vectors a, c ∈
RLs and b, d ∈ RLt , we have
vec(abT )T V −1 vec(cdT ) = (aT c)bT A−1 d.
(101)
(108)
By stacking the above equation for m = 1, . . . , M , we have
T
φk1
.. −1
φk 0 1 · · · φk 0 M
(109)
. V
φTkM
(x̂tM )T A−1 x̂t1
t T −1
, the
φTkm V −1 φk0 m0 = vec(θk (x̂tm )T )T V −1 vec(θk0 (x̂tm0 )T )
(107)
B. Lagrangian with a compact form
−1
(106)
M ×M
Proof. Using the above lemma, we have
(x̂t1 )T A−1 x̂t1
..
= (θkT θk0 )
.
V −1 = (ILs ⊗ A)−1 = ILs ⊗ A−1
(105)
Next, we explore φT V −1 φ via the lemma below.
To remove the large matrix V , we use the structure of V
and rewrite terms involving V (φT V −1 φ and q T V −1 Φ).
First, the inverses of V and A are as follows:
(104)
i
= (θkT θk0 )(x̂tm )T A−1 x̂tm0 .
This is indeed the dual form, but it still involves a large
matrix V . In the next subsection, by utilizing the structure of
V , we will write L in such a way that it involves only smaller
matrices.
(103)
···
···
(x̂t1 )T A−1 x̂tM
..
.
(x̂tM )T A−1 x̂tM
(110)
= (θkT θk0 )(X t )T A−1 X t
=
(θkT θk0 )G,
(111)
(112)
where G = (X t )T A−1 X t .
Finally, by stacking the above equation for k = 1, . . . , K,
we obtain compact form
T
φ11
..
.
T
φ −1
T −1
1M
φ11 , · · · φ1M , φ21 , · · · φKM
Φ V Φ=
φT21 V
.
..
φTKM
(113)
T
T
T
(θ1 θ1 )G (θ1 θ2 )G · · · (θ1 θK )G
(θ2T θ1 )G (θ2T θ2 )G · · · (θ2T θK )G
=
..
..
..
..
.
.
.
.
T
T
(θK
θ1 )G (θK
θ2 )G · · ·
= (ΘT Θ) ⊗ G,
T
(θK
θK )G
(114)
(115)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
12
where Θ = (θ1 , θ2 , . . . , θK ).
By adding and stacking the equation, we have
!
M X
N
X
T
V −1 φk0 1 , · · · φk0 M ,
snm qnm
m=1 n=1
M X
N
X
Note that we can rewrite G further as
=
G =(X t )T A−1 X t
(116)
1
1
1
−1
=(X t )T ( ILt +1 − 2 X t (SM
+ (X t )T X t )−1 (X t )T )X t
cf
cf
cf
(117)
1
= (X t )T X t
cf
1
1
−1
− 2 (X t )T X t (SM
+ (X t )T X t )−1 (X t )T X t
cf
cf
(118)
1
1
1
−1
+ K t )−1 K t ,
= K t − 2 K t (SM
(119)
cf
cf
cf
where K t = (X t )T X t is a kernel matrix. If (K t )−1 exists,
we obtain
(126)
snm (θkT0 xsn )((x̂tm )T A−1 x̂t1 , . . . , (x̂tm )T A−1 x̂tM )
m=1 n=1
(127)
=
=
=
=
=
=
M
X
N
X
m=1 n=1
M X
N
X
snm (θkT0 xsn )((x̂tm )T A−1 (x̂t1 , . . . , x̂tM ))
(128)
snm (θkT0 xsn )(xtm )T A−1 X t
(129)
m=1 n=1
M X
N
X
θkT0
snm xsn (x̂tm )T A−1 X t
m=1 n=1
!
M X
N
X
T
s
t T
θk 0
snm xn (x̂m )
A−1 X t
m=1 n=1
θkT0 X s S(X t )T A−1 X t
θkT0 X s SG.
(130)
(131)
(132)
(133)
Finally, by stacking the equation, we have
G = (cf (K t )−1 + SM )−1
(120)
T
q T V −1 Φ = (θ1T X st A−1 X t , . . . , θK
X st A−1 X t )
(134)
T
= (θ1T X s SG, . . . , θK
X s SG)
T
T
θ1
= vec ... X s SG
(135)
by applying the Woodbury formula.
In summary, G ∈ RM ×M is
t −1
−1
(cf (K ) + SM ) ,
−1
G = c1f K t − c12 K t (SM +
t T −1f t
(X ) A X ,
1
cf
(136)
T
θK
t −1
if (K ) exists,
−1
K t )−1 K t , if (SM
+ c1f K t )−1 exists,
T
= vec ΘT X s SG .
(137)
otherwise,
(121)
depending on the existence of the inverse of kernel matrix
K t ∈ RM ×M . It exists when the dimension of the column
space of X t is M , i.e., when the target samples are linearly
independent. If not, the second option can be used, i.e., the
−1
inverse of SM
+ c1f K t , which can be interpreted as the
−1
regularization of kernel K t with diagonal weight matrix SM
.
T −1
In the next lemma, we rewrite q V Φ.
Lemma 8. Given matrix Φ ∈ RLs (Lt +1)×KM and vector q ∈
RLs (Lt +1) , we have
T
q T V −1 Φ = vec ΘT X s SG .
(122)
Proof. Using the above lemma in a similar way as with the
previous lemma, we have
T
qnm
V −1 φk0 m0 = vec(xsn (x̂tm )T )T V −1 vec(θk0 (x̂tm0 )T )
(123)
= ((xsn )T θk0 )(x̂tm )T A−1 x̂tm0
(124)
(θkT0 xsn )(x̂tm )T A−1 x̂tm0 .
(125)
=
We now have the final form of the dual and present it in
the corollary below.
Corollary 1. The dual form of the original primal problem is
given by
−1 T T
a Y ((ΘT Θ) ⊗ G)Y a
2
T
+ (1T − (Y b̃)T − vec ΘT X s SG Y )a
max
a
s.t.
cT 1 ≥ a ≥ 0.
(138)
(139)
Proof. According to the lemmas derived above, we can write
the Lagrangian as
−1 T T T −1
a Y Φ V ΦY a + (1T − q T V −1 ΦY )a
2
−1 T −1
1
+
q V q+ s
(140)
2
2
−1 T T
=
a Y ((ΘT Θ) ⊗ G)Y a
2
T
+ (1T − (Y b̃)T − vec ΘT X s SG Y )a
−1 T −1
1
+
q V q + s.
(141)
2
2
By omitting the last two terms, which do not involve a, we
have the dual problem as claimed.
L=
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
13
Note that this dual form involves matrices of size at most
KM × KM , which is reasonable when the number of categories K and the number of samples in the target domain M
are both small.
If all snm = 0, then the problem reduces to MMDT, i.e.,
L=
−1 T T
a Y ((ΘT Θ) ⊗ G)Y a + (1T − (Y b̃)T )a, (142)
2
t T
Next, we derive (137) by first stacking
V −1 φk0 m0 = V −1 vec(θk0 (x̂tm0 )T )
V −1 φk0 1 · · · φk0 M
where G = (X ) X , since A = I.
=
=
A PPENDIX C
R ETRIEVING THE PRIMAL SOLUTION
=
After solving the dual problem with a QP solver, we need
to convert the dual solution a to w and b by
=
(143)
then finally to W .
Here, we again face the problem of large matrix V . We,
therefore, derive the core parts V −1 q and V −1 Φ as shown in
the lemmas below.
c1 d
c2 d
..
..
.
.
A−1
c1 A−1 d
−1
= c2 A d = c ⊗ (A−1 d)
..
.
(145)
M X
N
X
(θk0 ⊗ (A x̂t1 ), . . . , θk0 ⊗ (A−1 x̂tM ))
θk0 ⊗ (A−1 x̂t1 , . . . , A−1 x̂tM )
θk0 ⊗ (A−1 (x̂t1 , . . . , x̂tM ))
θk0 ⊗ (A−1 X t ).
(156)
(157)
(158)
(159)
V −1 Φ = V −1 φ11 , · · · φ1M , φ21 , · · · φKM
= (θ1 ⊗ (A
−1
t
X ), . . . , θK ⊗ (A
−1
t
X ))
= (θ1 , . . . , θK ) ⊗ (A−1 X t )
= Θ ⊗ (A
−1
(160)
(161)
(162)
t
X ).
(163)
Finally, we show primal solution W directly, i.e., avoiding
conversions from a to w, and then to W . In the corollary
below, we construct matrix W from a with much less computational costs.
Corollary 2. The solution to the primal problem is given by
W = X s S + Θ(Υ
(146)
Λ)T (X t )T A−1 ,
(164)
where
is element-wise multiplication and variables are
given in the proof below.
Proof. We first use the lemmas above to obtain
Lemma 10. Given V , q, and Φ, we have
V −1 q =
(155)
−1
(144)
Proof.
−1
A
−1
T
V vec(cd ) =
Then, we further stack the above equation for k = 1, . . . , K
to obtain
Lemma 9. Given V of the structure above and vectors c ∈
RLs and d ∈ RLt , we have
V −1 vec(cdT ) = c ⊗ (A−1 d).
(154)
for m = 1, . . . , M to obtain
t
w = V −1 (q + ΦY a),
(153)
= θk0 ⊗ (A−1 x̂tm0 )
snm xsn ⊗ (A−1 x̂tm )
(147)
m=1 n=1
w = V −1 (q + ΦY a)
=
and
M X
N
X
(165)
!
snm xsn ⊗ (A−1 x̂tm )
+ Θ ⊗ (A−1 X t ) Y a.
m=1 n=1
V
−1
Φ = Θ ⊗ (A
−1
(166)
t
X ).
(148)
For the ith part of w, we have
Proof. For (136), we have
V
−1
q=V
−1
= V −1
M X
N
X
m=1 n=1
M X
N
X
!
snm qnm
snm vec(xsn (x̂tm )T )
=
=
m=1 n=1
M X
N
X
m=1 n=1
(167)
(150)
The first term here can be written as
snm V −1 vec(xsn (x̂tm )T )
⊗ (A
−1
x̂tm ).
(152)
xs1,i
xs
−1 t T 2,i
= A X S . (168)
..
(151)
A−1
snm xsn
!
snm xsn,i (A−1 x̂tm )
m=1 n=1
+ (θ1,i , . . . , θK,i ) ⊗ (A−1 X t ) Y a.
!
m=1 n=1
M X
N
X
wi =
(149)
M X
N
X
M X
N
X
m=1 n=1
!
snm xsn,i x̂tm
xsN,i
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
14
and the second term as
as
(θ1,i A−1 X t , . . . , θK,i A−1 X t )Y a
(169)
= θ1,i A−1 X t Y1 a1 + · · · + θK,i A−1 X t YK aK
(170)
−1
t
X (θ1,i Y1 a1 + · · · + θK,i YK aK )
θ1,i
= A−1 X t (Y1 a1 , . . . , YK aK ) ...
=A
(171)
(172)
θK,i
=A
−1
t
X (Υ
θ1,i
..
Λ) . ,
θK,i
(173)
where
ak1
ak2
ak = .
..
(174)
akM
a11
..
Λ = (a1 , . . . , aK ) = .
···
a1M
···
t
t
Yk = diag(yk1
, . . . , ykM
)
t
t
y11 · · · yK1
.. .
Υ = ...
.
aK1
..
.
(175)
aKM
(176)
W = X s S + Θ(Υ
= X s S + Θ(Υ
Λ)T (X t )T A−1
Λ)T (X t )T
1
1
1
−1
+ (X t )T X t )−1 (X t )T
IL +1 − 2 X t (SM
cf t
cf
cf
(182)
!
(183)
= X s S + Θ(Υ
Λ)T
1
1
1
−1
(X t )T − 2 (X t )T X t (SM
+ (X t )T X t )−1 (X t )T
cf
cf
cf
(184)
s
T
= X S + Θ(Υ Λ)
!
1
1 t −1
1 t −1
IM − 2 K (SM + K )
(X t )T .
(185)
cf
cf
cf
We apply W to target sample xt by multiplying it from the
left, i.e., W x̂t ; Therefore, all computations with target samples
are inner products, which means we can use kernels to replace
the inner products.
In the dual form, we write matrix G with the kernel version
using kernel matrix K t as
k(xt1 , xt1 ) · · · k(xt1 , xtM )
..
..
..
(186)
Kt =
,
.
.
.
k(xtM , xt1 ) · · ·
k(xtM , xtM )
(177)
where k() is a kernel function. To transform target sample xt
with W , we have
W x̂t = X s S + Θ(Υ Λ)T
Combining these two terms, we obtain
! k(xt1 , xt )
s
1
1
1
..
−1
x1,i
+ K t )−1
IM − 2 K t (SM
.
.
θ1,i
cf
cf
cf
xs2,i
t
t
k(xM , x )
wi = A−1 X t S T . + A−1 X t (Y1 a1 , . . . , YK aK ) ... .
..
(187)
θK,i
xsN,i
Note that the nonlinearity introduced by this kernelization
(178) appears only in the transformation part; target samples are
transformed nonlinearly to the source domain. Only linear
By stacking the ith part for i = 1, . . . , Ls , we yield the SVMs in the source domain can be used here because the
matrix directly as
primal solutions (i.e., hyperplane parameters) of the sourcedomain SVMs are explicitly used in the estimation of W .
Target samples are therefore linearly classified in the source
W T =(w1 , . . . , wLs ) = A−1 X t S T (X s )T
−1 t
T
+ A X (Y1 a1 , . . . , YK aK )Θ
(179) domain after being nonlinearly transformed from the target
domain.
=A−1 X t S T (X s )T + (Y1 a1 , . . . , YK aK )ΘT (180)
=A−1 X t S T (X s )T + (Υ Λ)ΘT .
(181)
ACKNOWLEDGMENT
t
y1M
···
t
yKM
Part of this work was supported by JSPS KAKENHI Grant
Numbers JP26280015 and JP14J00223, and by the Research
Center for Biomedical Engineering, and was with the help of
a grant by Chugoku Industrial Innovation Center.
A PPENDIX D
K ERNELIZATION
R EFERENCES
In this section, we derive the kernel version of the dual
formulation. The obtained transformation is further rewritten
[1] S. Tanaka, T. Kaltenbach, K. Chayama, and R. Soetikno, “Highmagnification colonoscopy (with videos),” Gastrointest Endosc, vol. 64,
no. 4, pp. 604–13, Oct 2006.
!
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
[2] Cancer
Research
UK.
(2016)
Bowel
cancer
statistics.
[Online].
Available:
http://www.cancerresearchuk.org/cancer-info/
cancerstats/types/bowel/
[3] H. Kanao, S. Tanaka, S. Oka, M. Hirata, S. Yoshida, and K. Chayama,
“Narrow-band imaging magnification predicts the histology and invasion
depth of colorectal tumors.” Gastrointest Endosc, vol. 69, no. 3 Pt 2,
pp. 631–636, Mar 2009.
[4] S. Oba, S. Tanaka, S. Oka, H. Kanao, S. Yoshida, F. Shimamoto, and
K. Chayama, “Characterization of colorectal tumors using narrow-band
imaging magnification: combined diagnosis with both pit pattern and
microvessel features,” Scand J Gastroenterol, vol. 45, no. 9, pp. 1084–
92, Sep 2010.
[5] M. Häfner, A. Gangl, M. Liedlgruber, A. Uhl, A. Vécsei, and F. Wrba,
“Classification of endoscopic images using delaunay triangulation-based
edge features,” in Image Analysis and Recognition, ser. Lecture Notes in
Computer Science, A. Campilho and M. Kamel, Eds. Springer Berlin
Heidelberg, 2010, vol. 6112, pp. 131–140.
[6] M. Häfner, A. Gangl, M. Liedlgruber, A. Uhl, A. Vecsei, and F. Wrba,
“Endoscopic image classification using edge-based features,” in Pattern
Recognition (ICPR), 2010 20th International Conference on, Aug 2010,
pp. 2724–2727.
[7] R. Kwitt, A. Uhl, M. Häfner, A. Gangl, F. Wrba, and A. Vécsei, “Predicting the histology of colorectal lesions in a probabilistic framework,”
in Computer Vision and Pattern Recognition Workshops (CVPRW), 2010
IEEE Computer Society Conference on, June 2010, pp. 103–110.
[8] S. Gross, T. Stehle, A. Behrens, R. Auer, T. Aach, R. Winograd,
C. Trautwein, and J. Tischendorf, “A comparison of blood vessel features
and local binary patterns for colorectal polyp classification,” in Proc.
SPIE, vol. 7260, 2009, pp. 72 602Q–72 602Q–8.
[9] T. Stehle, R. Auer, S. Gross, A. Behrens, J. Wulff, T. Aach, R. Winograd,
C. Trautwein, and J. Tischendorf, “Classification of colon polyps in nbi
endoscopy using vascularization features,” in Proc. SPIE, vol. 7260,
2009, pp. 72 602S–72 602S–12.
[10] J. J. W. Tischendorf, S. Gross, R. Winograd, H. Hecker, R. Auer,
A. Behrens, C. Trautwein, T. Aach, and T. Stehle, “Computer-aided
classification of colorectal polyps based on vascular patterns: a pilot
study,” Endoscopy, vol. 42, no. 3, pp. 203–7, Mar 2010.
[11] T. Tamaki, J. Yoshimuta, M. Kawakami, B. Raytchev, K. Kaneda,
S. Yoshida, Y. Takemura, K. Onji, R. Miyaki, and S. Tanaka, “Computeraided colorectal tumor classification in NBI endoscopy using local
features,” Medical Image Analysis, vol. 17, no. 1, pp. 78–100, 2013.
[12] B. Settles, Active Learning, ser. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers, June
2012.
[13] S. Sonoyama, T. Hirakawa, T. Tamaki, T. Kurita, B. Raytchev,
K. Kaneda, T. Koide, S. Yoshida, Y. Kominami, and S. Tanaka,
“Transfer learning for Bag-of-Visual words approach to NBI endoscopic
image classification,” in 2015 37th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC).
IEEE, aug 2015, pp. 785–788. [Online]. Available: http://ieeexplore.
ieee.org/lpdocs/epic03/wrapper.htm?arnumber=7318479
[14] S. J. Pan and Q. Yang, “A survey on transfer learning,” Knowledge and
Data Engineering, IEEE Transactions on, vol. 22, no. 10, pp. 1345–
1359, 2010.
[15] R. Raina, A. Battle, H. Lee, B. Packer, and A. Y. Ng, “Self-taught
learning: transfer learning from unlabeled data,” in Proceedings of the
24th international conference on Machine learning. ACM, 2007, pp.
759–766.
[16] W. Dai, Q. Yang, G.-R. Xue, and Y. Yu, “Boosting for transfer learning,” in Proceedings of the 24th international conference on Machine
learning. ACM, 2007, pp. 193–200.
[17] D. Silver, G. Bakir, K. Bennett, R. Caruana, M. Pontil, S. Russell, and
P. Tadepalli, “Nips workshop on “inductive transfer: 10 years later”,”
Whistler, Canada, 2005.
[18] OLYMPUS. (2006, 5) EVIS LUCERA SPECTRUM Endoscopic Video
Imaging System. [Online]. Available: http://www.olympus-global.com/
en/news/2006a/nr060515evise.jsp
[19] ——. (2006, 12) EVIS LUCERA colonovideoscope instructions.
[Online]. Available: https://www.olympus.co.uk/medical/en/medical
systems/hidden/download jsp.jsp?link=/medical/rmt/media/content/
content 1/documents 1/brochures 1/Evis Lucera GIF-260 Series
operation manual 001 V1-en GB 20000101.pdf
[20] ——. (2012, 5) CV-290 EVIS LUCERA ELITE Processor.
[Online]. Available: https://www.olympus.co.uk/medical/en/medical
systems/products services/product details/product details 63297.jsp
[21] J. Hoffman, E. Rodner, T. Darrell, J. Donahue, and K. Saenko, “Efficient
learning of domain-invariant image representations,” In Proceedings of
15
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
the 1st International Conference on Learning Representations, pp. 1–9,
2013.
J. Hoffman, E. Rodner, J. Donahue, B. Kulis, and K. Saenko, “Asymmetric and category invariant feature transformations for domain adaptation,” International Journal of Computer Vision, vol. 109, no. 1-2, pp.
28–41, 2014.
S. Sonoyama, T. Tamaki, T. Hirakawa, B. Raytchev, K. Kaneda,
T. Koide, S. Yoshida, H. Mieno, and S. Tanaka, “Transfer Learning for
Endoscopic Image Classification,” in The Korea-Japan joint workshop
on Frontiers of Computer Vision (FCV2016), 2016, pp. 258–262.
E. Rodner, J. Hoffman, J. Donahue, T. Darrell, and K. Saenko. (2013)
Towards Adapting ImageNet to Reality: Scalable Domain Adaptation
with Implicit Low-rank Transformations. [Online]. Available: http:
//arxiv.org/abs/1308.4200
J. Hoffman, E. Rodner, J. Donahue, K. Saenko, T. Darrell, and
B. Kulis, “Domain Adaptation Project,” 2013. [Online]. Available:
https://www.eecs.berkeley.edu/∼jhoffman/domainadapt/
T. Hirakawa, T. Tamaki, B. Raytchev, K. Kaneda, T. Koide,
Y. Kominami, S. Yoshida, and S. Tanaka, “SVM-MRF segmentation of
colorectal NBI endoscopic images,” in 2014 36th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society,
vol. 2014. IEEE, aug 2014, pp. 4739–4742. [Online]. Available: http:
//ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=6944683
S. Sonoyama, T. Tamaki, T. Hirakawa, B. Raytchev, K. Kaneda,
T. Koide, Y. Kominami, S. Yoshida, and S. Tanaka, “Tradeoff between speed and performance for colorectal endoscopic
NBI image classification,” in Proc. SPIE, S. Ourselin and
M. A. Styner, Eds., vol. 9413, mar 2015, p. 94132D. [Online]. Available: http://dx.doi.org/10.1117/12.2081928http://proceedings.
spiedigitallibrary.org/proceeding.aspx?doi=10.1117/12.2081928
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for
fast feature embedding,” arXiv preprint arXiv:1408.5093, 2014.
A. Sharif Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “CNN
Features Off-the-Shelf: An Astounding Baseline for Recognition,” in The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
Workshops, 2014.
B. L. Welch, “On the comparison of several mean values: An alternative
approach,” Biometrika, vol. 38, no. 3/4, pp. 330–336, 1951. [Online].
Available: http://www.jstor.org/stable/2332579
G. K. Kanji, 100 Statistical Tests, 3rd ed. SAGE Publications Ltd, Aug
2006.
T. Tamaki, S. Sonoyama, T. Hirakawa, B. Raytchev, K. Kaneda,
T. Koide, S. Yoshida, H. Mieno, and S. Tanaka, “Computer-Aided Colorectal Tumor Classification in NBI Endoscopy Using CNN Features,”
in The Korea-Japan joint workshop on Frontiers of Computer Vision
(FCV2016), 2016, pp. 61–65.
E. Tzeng, J. Hoffman, T. Darrell, U. C. Berkeley, K. Saenko, and
U. Lowell, “Simultaneous Deep Transfer Across Domains and Tasks,”
ICCV2015, pp. 0–0, 2015.
D. A. Harville, Matrix Algebra From a Statistician’s Perspective. New
York: Springer-Verlag, 1997. [Online]. Available: http://www.springer.
com/us/book/9780387949789
| 1cs.CV
|
Waveform Relaxation for the Computational Homogenization of
Multiscale Magnetoquasistatic Problems
I. Niyonzimaa,b,∗, C. Geuzainec , S. Schöpsa,b
a Institut
arXiv:1607.05429v1 [math.NA] 19 Jul 2016
für Theorie Elektromagnetischer Felder, Technische Universitaet Darmstadt,
Schlossgartenstrasse 8, D-64289 Darmstadt, Germany.
b Graduate School of Computational Engineering, Technische Universitaet Darmstadt,
Dolivostrasse 15, D-64293 Darmstadt, Germany.
c Applied and Computational Electromagnetics (ACE), Université de Liège,
Montefiore Institute, Quartier Polytech 1, Allée de la découverte 10, B-4000 Liège, Belgium
Abstract
This paper proposes the application of the waveform relaxation method to the homogenization of multiscale magnetoquasistatic problems. In the monolithic heterogeneous multiscale method, the nonlinear
macroscale problem is solved using the Newton–Raphson scheme. The resolution of many mesoscale problems per Gauß point allows to compute the homogenized constitutive law and its derivative by finite
differences. In the proposed approach, the macroscale problem and the mesoscale problems are weakly
coupled and solved separately using the finite element method on time intervals for several waveform
relaxation iterations. The exchange of information between both problems is still carried out using the
heterogeneous multiscale method. However, the partial derivatives can now be evaluated exactly by solving
only one mesoscale problem per Gauß point.
Keywords: Cosimulation method, Eddy currents, Finite element method, FE2 , HMM, Homogenization,
Multiscale modeling, Nonlinear problems, Magnetoquasistatic problems, Waveform relaxation method.
1. Introduction
The recent use of the heterogeneous multiscale method (HMM [1]) in electrical engineering has allowed
to accurately solve magnetoquasistatic (MQS) problems with multiscale materials, e.g. microstructured
composites with ferromagnetic inclusions exhibiting hysteretic magnetic behavior [2, 3]. The method
requires the solution of one macroscale and mesoscale problems at each Gauß point of the macroscale
problem (see Figure 1) in a coupled formulation based on the Finite Element (FE) method. In [2, 3]
the coupled problem was monolithically time discretized by using equal step sizes at all scales and the
resulting nonlinear problem was solved by an inexact parallel multilevel Newton-Raphson scheme.The
finite-difference approach involves the resolution of 4 mesoscale problems in the three-dimensional case
(respectively 3 mesoscale problems in two-dimensions) for computing the approximated Jacobian at each
Gauß point.
The use of different time steps becomes important for problems involving different dynamics at both
scales. In the case of the soft ferrite material studied in [4], for example, it was shown that capacitive
effects occurring at the mesoscale could be accounted for by upscaling proper homogenized quantities in
the macroscopic MQS formulation. Another relevant case involves perfectly isolated laminations and soft
magnetic composites (SMC) with eddy currents at the mesoscopic level (scales of the sheet/metalic grain)
but without the resulting macroscopic eddy currents. The application of the HMM to problems involving
such materials leads to a formulation featuring magnetodynamic problems at the mesoscopic scale and a
magnetostatic problem at the macroscopic level. Thus, small time steps should be used at the mesoscale
to resolve the eddy currents (especially with saturated hysteretic materials) while large time steps could be
∗ Corresponding
author.
addresses: [email protected] (I. Niyonzima), [email protected] (C. Geuzaine), [email protected] (S. Schöps)
∗∗ Email
Preprint submitted to Journal of Computational Physics
March 12, 2018
downscaling
upscaling
Figure 1: Scale transitions between macroscale (left) and mesoscale (right) problems. Downscaling (Macro to meso): obtaining
proper boundary conditions and the source terms for the mesoscale problem from the macroscale solution. Upscaling (meso
to Macro): effective quantities for the macroscale problem calculated from the mesoscale solution [2, 3].
used to discretize the rather slowly-varying exciting source current at the macroscale level. Obviously, in
such cases of different dynamics, the use of different time steps can help to reduce the overall computational
cost.
In this paper we propose a novel approach that provides a natural setting for the use of different
time steps. The approach applies the waveform relaxation method [5, 6] to the homogenization of MQS
problems: the macroscale problem and the mesoscale problems are solved separately on time intervals and
their time-dependent solutions are exchanged in a fixed point iteration. The decoupling of the macroscale
and the mesoscale problems and the independent resolution of these problems on time intervals has the
potential to significantly reduce both the computation and communication cost of the multiscale scheme;
in particular it allows to compute the Jacobian exactly at each Gauß point of the macroscale domain by
solving only one mesoscale problem. As a drawback, waveform relaxation iterations are needed for the
overall problem to converge in addition to the Newton–Raphson iterations on the meso- and macroscale.
The latter exhibits quadratic convergence, while the fixed point iteration only leads to a linear convergence
but is applied to waveforms instead of classical vector spaces. We present both approaches and compare
the computational and the communication costs for both the monolithic and the waveform relaxation
HMM.
The article is organized as follows: in Section 2 we introduce Maxwell’s equations and the MQS problem.
The weak form of the MQS problem is then derived using the modified vector potential formulation. Section
3 deals with the multiscale formulations of the HMM for the MQS problem along the lines of the works [2, 3]
with an emphasis on the coupling between the macroscale and the mesoscale problems. These formulations
are valid for the monolithic and the waveform relaxation (WR) HMM. In Section 4 we develop a novel
theoretical framework for the monolithic HMM. Using this framework we derive a reduced Jacobian from
the Jacobian of the full problem using the Schur complement, similar as it has been proposed for the
Variational Multiscale Method in [4]. Section 5 gives a short overview of the waveform relaxation method.
The notion of weak and strong coupling are explained in the general context of coupled systems. The
method is then used in Section 6 in combination with the HMM and gives rise to the newly developed
WR–HMM. Section 7 is dedicated to the estimation of the computational cost for both the monolithic
HMM and the WR-HMM. Formulae for the computation of costs for the monolithic HMM and the WRHMM are derived and analyzed to give a hint on a possible reduction of the computational cost of both
methods. Section 8 deals with an application case. We consider an application involving idealized soft
magnetic materials (SMC) without global eddy currents. Convergence of the method as a function of the
waveform relaxation iterations and the macroscale/mesoscale time stepping is numerically investigated.
2
Γ = Γh ∪ Γe
Ωs
Ωc
j
ΩC
c
Figure 2: Bounded domain Ω and its subregions [2, 3]. The domain Ω can be split into the conductiong region Ωc (with
σ > 0) and the non-conducting region ΩC
c = Ω\Ωc (with σ = 0) which contains inductors Ωs where the current density j s
is imposed. The boundary of the domain Γ is such that Γ = Γe ∪ Γh with Γe ∩ Γh = ∅. The region Γe is the part of the
boundary where the tangential trace of e (resp. the normal trace of b) is imposed and Γh is the part of the boundary where
the tangential trace of h (resp. the normal trace of d or j) is imposed.
2. The magnetoquasistatic problem
3
In an open, bounded domain Ω = Ωc ∪ ΩC
c ⊂ R (see Figure 2) and t ∈ I = (t0 , tend ] ⊂ R, the evolution
of electromagnetic fields is governed by the following Maxwell’s equations on Ω × I, i.e.,
curl h = j + ∂t d,
and the constitutive laws, e.g. [7]
j(x, t) = J e(x, t), x ,
curl e = −∂t b,
div d = ρ,
d(x, t) = D e(x, t), x ,
div b = 0,
h(x, t) = H b(x, t), x .
(2.1 a-c)
In these equations, h is the magnetic field [A/m], b the magnetic flux density [T], e the electric
field [V/m], d the electric flux density [C/m2 ], j the electric current density [A/m2 ], and ρ the electric
charge density [C/m3 ]. The domain Ωc contains conductors whereas the domain ΩC
c contains insulators.
Additionally, suitable initial conditions and boundary conditions must be imposed for the problem to be
well posed.
In this paper we consider only the ‘magnetoquasistatic’ (MQS) case; it is derived from Maxwell’s
equations by neglecting the displacement currents with respect to the eddy currents ∂t d j. This can
be justified if L λ and δ ' L with L, the characteristic length of the system, λ the wavelength of the
exciting source and δ the skin depth. A more rigorous analysis can be found in [8]. The resulting eddy
current problem can be defined by the following MQS approximation of Maxwell’s equations [9]
curl h = j,
curl e = −∂t b,
div b = 0,
(2.2 a-c)
and the relevant constitutive laws for j (see equation (2.1 a)) and h (see equation (2.1 b)). For the
applications treated in this paper the first (electric) constitutive law will be considered of the form j(x, t) =
σ(x) e(x, t) + j s (x, t), with σ the (anisotropic) electric conductivity and j s an imposed (source) electric
2
current density in Ωs ⊂ ΩC
c [A/m ]. The second (magnetic) constitutive law can be linear, nonlinear
reversible or nonlinear irreversible (i.e. with hysteresis). Typical nonlinear reversible models include
Brauer’s model [10], Rougé’s formula [11] or splines.
Boundary conditions on the tangential component of the magnetic field (or on the normal component
of j) and on the normal component of the magnetic flux density (or on the tangential component of e)
3
are imposed on complementary parts Γh and Γe of the boundary Γ = ∂Ω = Γe ∪ Γh :
n × h|Γh = ht ,
n · b|Γe = bn .
(2.3 a-b)
In this paper, we use the modified vector potential formulation and write b and e as:
e = −∂t a
b = curl a ,
(2.4 a-b)
with a the magnetic vector potential [V s/m]. Therefore the essential boundary condition n · b|Γe =
n · (curl a)|Γe = bn leads to the cancellation of the normal component of the curl and can be fulfilled by
imposing n × a|Γe = 0.
The MQS problem (2.2 a-c) together with the constitutive laws (2.1 a) and (2.1 c) can be solved
using the finite element method. To do this, a Galerkin formulation of the problem must be developed.
Existence of the (weak) solutions presupposes some regularity assumptions on the data of the problem.
The conductivity σ is defined such that the mapping J defined in (2.1 a) is monotone, nondecreasing
and continuous in e. For the linear electric law used in this paper, these conditions are fulfilled if σ is
bounded, i.e., if σ ∈ L∞ (Ω). The mapping H is assumed to be maximal monotone which presupposes
that ∂H/∂b is positive definite. This is the case for linear and nonlinear magnetic mappings but does
not hold for hysteretic magnetic materials. The excitation term j s needs also to be regular enough, e.g.,
j s ∈ L2 ((0, T ); V (Ω)∗ ). The space V (Ω) := H e (curl; Ω) is the appropriate function space for the vector
potential with boundary conditions on Γe , and the superscript ∗ is used to denote its dual (see [9, 12]).
Using these assumptions, the weak form of (2.2 a) reads [12, 13]: find a ∈ L2 ((0, T ); V (Ω)) with
∂t a ∈ L2 ((0, T ); V (Ω)∗ ) such that
(σ∂t a, a0 )Ωc + (H(curl a), curl a0 )Ω + hht , a0 iΓh − (j s , a0 )Ωs = 0
(2.5)
holds for all test functions a0 ∈ V0 (Ω), where the subscript 0 is used to denote homogeneous boundary
Dirichlet conditions. More regularity in time and space for the solution can be obtained by imposing more
regularity on the data of the problem [14] Round brackets (·, ·) are used for volume integrals whereas angle
brackets h·, ·i are used for surface integrals. The field a derived from (2.5) must be gauged on ΩC
c to ensure
its uniqueness. This can mathematically be achieved by factoring the space H 0e (curl; Ω) by gradients of
scalar potentials, e.g., [12].
2.1. Multiscale
Following [2, 3, 15], we use the subscript ε = l/L to denote quantities with rapid fluctuations. The
length l denotes the length of the periodic cell and the length L denotes the characteristic length of the
material or the minimum wavelength of the exciting source current j s (t). This wavelenght is defined
as λmin = c/fmax where c is the speed of light and fmax is the highest frequency obtained when j s (t)
is decomposed using the Fourier transform. The homogenized computational domain is assumed to be
located far from the boundary Γ such that the boundary term in (2.5) is independent of ε.Using this
convention we can define the equivalent multiscale weak form for equation (2.5).
Problem 2.1 (Multiscale finescale problem). Find aε ∈ L2 ((0, T ); V (Ω)) with ∂t aε ∈ L2 ((0, T ); V (Ω)∗ )
such that
(σ ε ∂t aε , a0 ε )Ωc + (Hε (curl aε ), curl a0 ε )Ω + hht , a0 ε iΓh − (j s , a0 ε )Ωs = 0
(2.6)
holds for all test functions a0 ε ∈ V0 (Ω).
This ‘finescale’ weak form is used as the reference solution for problems involving multiscale materials.
The conductivity σ ε and the material mapping Hε are defined by
x
and Hε (bε (x, t), x) = H(bε (x, t), x, x/ε)
(2.7)
σ ε (x) = σ
ε
for all x = (x1 , x2 , x3 ) in ΩH , where ΩH is the multiscale computational domain. The mapping H is
used to represent two-scale composite materials for which the characteristic length at the mesoscale is ε
[15]. By abuse of notation, we use H instead of H in the rest of the text. In (2.7), slow variations of the
4
material law are accounted for by the term x while the fast fluctuations are accounted for by the term x/ε
for ε 1 (see [16, 17, 15]).
As an illustration, consider a two-dimensional linear magnetic material law H(b(x, t), x) = µ(x) b(x, t)
with the magnetic permeability defined as
(
µ1
if(x1 mod T ) ≤ λ,
µ(x) =
µ2
if(x1 mod T ) > λ,
for all x ∈ ΩH , positive µ1 , µ2 and λ < T . The magnetic permeability µ is periodic with period T and is
representative of a stack of laminations made of two materials. The division by the parameter ε allows to
make the period smaller (from T to ε T ). In the previous case, the permeability becomes:
(
µ1
if(x mod εT ) ≤ λ,
µ(x/ε) =
µ2
if(x mod εT ) > λ,
and is ε-periodic. Therefore the material law Hε (bε (x, t), x, x/ε) = µ(x/ε) bε (x, t) is rapidly fluctuating
for ε 1.
In the following sections, the indices M, m and c are used for denoting the macroscale, the mesoscale
and correction terms, respectively. The variables x ∈ Ω and y = (y1 , y2 , y3 ) ∈ Ωm are the macroscale
and the mesoscale coordinates and the mesoscale coordinates are only defined on the cell domain with the
origin at the barycenter.
3. The heterogeneous multiscale method
Developments of this section are derived along the lines of ideas in [2, 3]. The resolution of the finescale
reference Problem 2.1 is computationally expensive for small values of ε. Hence, multiscale methods as
HMM [1] are necessary to reduce the computational costs and eventually make realistic simulations feasible.
For the MQS problems, HMM is based on the scale separation assumption (ε 1) and was already
illustrated in Figure 1. When HMM is applied, the finescale problem is replaced by one macroscale problem
defined on a coarse mesh covering the entire domain and accounting for the slow variations of the finescale
solution, and by many mesoscale problems defined on small, finely meshed areas around some points of
interest of the macroscale mesh (e.g. numerical quadrature points), and are used for computing missing
information. The transfer of information between these problems is done during the upscaling and the
downscaling stages.
Equations governing the macroscale and the mesoscale are derived from the finescale problem using
the asymptotic homogenization theories. The macroscale problem is derived from the finescale problem
using the classical weak convergence theory whereas the mesoscale problem is derived using the two-scale
convergence [15]. For these convergence theories to be applied, the solution of the finescale Problem 2.1
must exist and belong to appropriate function spaces (e.g.: reflexive or separable Banach spaces). This
imposes some regularity conditions on σ ε , Hε and on the excitation j s which have been stated in Section
2. However, the HMM has been numerically used for hysteretic magnetic laws that do not fulfill the
monotonicity assumptions on Hε [3].
3.1. The macroscale problem
The macroscale weak form of the problem in a-formulation can be derived from equation (2.5) as follows [3]:
Problem 3.1 (Macroscale weak problem). Find aM ∈ L2 ((0, T ); V (Ω)) with ∂t aM ∈ L2 ((0, T ); V (Ω)∗ )
such that
D
E
σM ∂t aM , a0M
+ HM (bM , x), curlx a0M + n × hM , a0M
− j s , a0M
=0
(3.1)
Ωc
holds for all
a0M
Ω
∈ V0 (Ω), bM := curlx aM and t ∈ I = (t0 , tend ].
5
Γh
Ωs
Thanks to the linearity of the electric law, the macroscopic conductivity σM is obtained using the asymptotic expansion method [16]
!
Z
X
∂χj (y)
1
σi j −
σi k
dy,
(3.2)
(σM )i j =
|Ωm | Ωm
∂yk
k
where χj is obtained by solving the cell problem: find χj ∈ VG such that
(grady ψ)T , σ(grady χj − ej )
= 0, ∀ψ ∈ VG .
(3.3)
Ωm
The space VG is the space H 1 (Ωm ) with periodic boundary conditions while ej is the unit vector in the j th
spatial direction. The field bc = curly ac is the magnetic correction field obtained by solving the mesoscale
problem corresponding to points x ∈ Ω. The mesoscale fields depend on the associated macroscale field
and vice versa due to coupling of scales, i.e.,
ac (x, y, t) = Ac y, aM (x, t)
(3.4)
where Ac denotes the solution operator of the mesoscale problem to be described in the following section
3.2. It is defined in analogy to the corrector operator defined in [18] for nonlinear scalar elliptic problems.
The macroscopic magnetic law HM in (3.1) is computed at each point (x, t) ∈ Ω × I using both scales
according to the averaging formula from the two-scale convergence theory [15]
Z
1
HM (bM (x, t), x) :=
H bc (x, y, t) + bM (x, t), x, y dy ∀(x, t) ∈ Ω × I
(3.5)
|Ωm | Ωm
where bM := curlx aM and bc = curly Ac (y, aM ). The underlying constitutive law H is known for the
heterogeneous phases at the mesoscale level.
3.2. Mesoscale problems
The following weak form of the mesoscale is defined from equation (2.5) [15, 3]
Problem 3.2 (Mesoscale weak problem). Find the mesoscale correction ac ∈ L2 ((0, T ); Vper (Ωm )) with
∂t ac ∈ L2 ((0, T ); Vper (Ωm )∗ ) using periodic boundary conditions such that
= 0,
(3.6)
− σeM , a0c
+ H(bc + bM , x, y), curly a0c
σ∂t ac , a0c
Ωm
Ωmc
Ωmc
for all a0c ∈ Vper (Ωm ), magnetic correction field bc = curly ac , and periodic boundary conditions and bM
given. The subscript per is used to denote the use of periodic boundary conditions.
The macroscale magnetic and electric fields are defined as
bM (x, t) := curlx aM (x, t)
eM (x, t) := −∂t aM (x, t) − κ(∂t bM × y)T
with κ = 1 for two-dimensional problems and κ = 1/2 for three-dimensional problems. Existence and
uniqueness of the mesoscale correction ac motivates the introduction of the solution operator in (3.4). It can
be formally deduced based on standard theory for nonlinear elliptic-parabolic problems, e.g. [12, 19, 20, 21].
3.3. Space and time discretization
Macroscale and mesoscale equations are solved using the finite element method. The first step consists
h
in discretizing the computational domain into elements. The fields aH
M and ac are approximation of
M
the continuous fields aM and ac on the discretized computational domain and aH
M ∈ (0, T ] × W H,0 and
6
M
m
ahc ∈ (0, T ] × W m
h,0 where W H,0 and W h,0 are H(curl)-conforming edge finite elements. In this paper we
consider only lowest order H(curl)-conforming edge finite elements as test and basis functions, i.e.,
o
n
I
M
v
∈
N
(K
)
∀K
∈
T
WM
:=
v
∈
H(curl;
Ω)
M
M
0
H ,
H,0
o
n
(3.7)
I
m
Wm
h,0 := v ∈ H(curl; Ωm ) v ∈ N0 (Km ), ∀Km ∈ Th
where N0I (K) := a + b × x | a, b ∈ R3 , see e.g. [9]. The triangulations THM and Thm are defined on the
macroscale and mesoscale domains, respectively.
The weak forms (3.1) and (3.6) can then be computed using numerical quadrature rules. This implies that the quantities involved in the integrations (e.g. the homogenized material law) be known at
Gauß points (i = 1, . . . , NGP ). Omitting the superscript i used for the numbering of the Gauß in the
approximation of the mesoscale field ahc (y, t), the discrete spaces give rise to the approximations
aM (x, t) ≈ aH
M (x, t) =
NM
X
αM,p (t)aM,p (x)
Nc
X
h
a(i)
c (y, t) ≈ ac (y, t) =
and
p=1
αc,p (t)ac,p (y).
(3.8)
p=1
Testing (3.1) and (3.6) yields the macroscale mass matrix
M M := σM aM , a0M
Ωc
which is singular due to σM = 0 on ΩC
c and the stiffness term
D
E
F M (αM , αc ) := HM curly ac + curlx aM , x , curlx a0M + ht , a0M
Ω
(1)
(N
Γh
− j s , a0M
(3.9)
Ωs
)
with ac = [ac , . . . , ac GP ]. Similar definitions hold for M m and F m on the mesoscale. The extension to
higher order edge elements or nodal elements for 2D problems is straightforward. Following the classical
approach, numerical quadrature rules are used to compute the weak forms. For the macroscale problem,
we use numerical quadrature with one Gauß point which is enough to capture the slow variations of the
missing material law at Gauß points. The missing material law can also be computed at the barycenter of
the element [22].
(1)
(NGP )
Problem 3.3 (Semidiscrete multiscale problem). Find waveforms [αM (t), αc (t), . . . , αc
M M ∂t αM + F M (αM , αc ) = 0,
(t)] such that
(3.10)
and for the mesoscale problems i = 1, . . . , NGP
(i)
(i)
(i)
M m ∂t α(i)
c + F m (αc , αM , ∂t αM ) = 0
(1)
(NGP )
for a given set of initial values [αM (t0 ), αc (t0 ), . . . , αc
(3.11)
(t0 )].
Finally, the time-dependent Problem (3.10-3.11) can be solved using any classical (implicit) time integration scheme followed by a nonlinear solution method. In the simplest case, i.e., using the backward
Euler scheme, the following nonlinear problem has to be solved for for each time step:
Problem 3.4 (Nonlinear, discrete multiscale problem). Find the solutions
(k)
GP ,k)
[αM , α(1,k)
, . . . , α(N
] ∈ IRNM +NGP ·Nc
c
c
such that
(k)
(k−1)
α − αM
(k)
(k)
+ F M αM , α(k)
= 0,
RM αM , α(i,k)
:= M M M
c
c
∆tk
7
(3.12)
and for the mesoscale problems i = 1, . . . , NGP
Rm
(k)
αM , α(i,k)
c
(i,k)
:= M m
αc
(i,k)
(i,k−1)
− αc
∆tk
+ Fm
(i,k)
α(i,k)
, αM ,
c
αM
(i,k−1)
− αM
∆tk
!
= 0,
(3.13)
(k)
where the superscript k is used to denote the approximations at time instants tk ∈ [t0 , tend ] , e.g. αM ≈
αM (tk ) and ∆tk := tk+1 − tk is the corresponding time step size.
The following loops are defined for the monolithic and the waveform relaxation methods: the loop for
the number of time windows (TW) with 1 ≤ n ≤ NTW , the loop for the number of waveform relaxation
iterations (WR) with 1 ≤ l ≤ NWR , the loop for the number of time stepping (TS) with 1 ≤ k ≤ NTS ,
the loop for the number of Newton–Raphson nonlinear iterations (NR) with 1 ≤ j ≤ NNR and the loop
for the number of Gauß points (GP) with 1 ≤ i ≤ NGP . Table 1 summarizes the loops, the letter used for
indexing them and the total number of iterations for each loop.
Table 1: Loops involved in the monolithic and waveform relaxation HMM.
Type of loop
Time window
Waveform relaxation
Time stepping
Newton–Raphson
Gauß points
Loop
TW
WR
TS
NR
GP
Index
n
l
k
j
i
Maximum number of iterations
NTW
NWR
NTS
NNR
NGP
4. Monolithic HMM
In the following a rigorous interpretation of the time-stepping procedures proposed in the context
of HMM is given in terms of Problem 3.4. These derivations are an important building block for the
comparison with the waveform relaxation approach in Section 5.
In [3] the Algorithm 1 was proposed. For each time step, a nonlinear system on the macroscale is solved
using the Newton–Raphson method until convergence is reached. In each Newton iteration the material
law (3.5) is evaluated
Z
1
(j,k)
(i,j,k)
H b(i,j,k)
, curlx aM , x, y dy
HM
:=
c
|Ωm | Ωm
(j,k)
= curly Ac (y, aM ) is obtained from the discretized version of the nonlinear solution operator
where b(i,j,k)
c
given in (3.4). This is implemented by solving the nonlinear equation (3.13) again by the Newton–Raphson
m
method using NNR
iterations, cf. Algorithm 2. This relaxation within the Newton scheme corresponds
to a monolithic time-stepping scheme although it features parallel evaluations at the Gauß points at the
level of the nonlinear solver. The two nested new Newton loops (inner and outer) are a special case of
a parallel multilevel Newton scheme as they are used for example in circuit simulation [23]. Let us state
the equivalence for the case in which only one inner iteration of a simplified Newton–Raphson scheme is
carried out. This is closely related to the Newton–Raphson scheme developed in [18] which involves the
evaluation of the Fréchet derivative of the nonlinear corrector operator.
(j,k)
Let αM
(i,j,k)
and αc
(j,k)
FM
:=
denote the j th Newton–Raphson iterates. Then we define
(j,k)
F M αM , α(j,k)
c
(i,j,k)
,
F (i,j,k)
m
:= F m
8
(i,j,k)
α(i,j,k)
, αM ,
c
αM
(i,j,k−1)
− αM
∆tk
!
.
(4.1)
Proposition 4.1. Solving the monolithic system (3.12)-(3.13)
the Jacobian
(j,k)
∂F M
0
···
0
M M
∂α(j,k)
M
∂F (1,j,k)
m
0
Mm 0
0
1
(j,k)
+ ∂α(j,k)
JR :=
M
∆tk .
..
..
..
.
0
0
.
(NGP ,j,k)
∂F
m
0
0
0 M m
(j,k)
∂αM
with the Newton–Raphson scheme using
(j,k)
∂F M
(1,j,k)
∂αc
∂F (1,j,k)
m
(1,j,k)
(j,k)
···
∂F M
(NGP ,j,k)
0
0
(NGP ,j,k)
∂F m
∂αc
0
∂αc
0
..
0
0
.
(4.2)
(NGP ,j,k)
∂αc
is equivalent to the scheme proposed in Algorithms 1 and 2 if
m
(a) no inner Newton iterations on the mesoscale are carried out, i.e., NNR
= 1 and
(i,j,k)
(b) the sensitivity of the mesoscale Jacobian w.r.t. to the macroscale is disregarded, i.e., ∂Jm
0
(j,k)
/∂αM
=
Proof. The equivalence is easily established by comparing the solution operator Ac as used in Algorithms 1(j,k)
2 to the Schur complement of the Jacobian JR as given in (4.2) and already proposed in [4] for the
Variational Multiscale Method. The latter reads
!−1
N
(j,k)
(j,k)
GP
(i,j,k)
(i,j,k)
X
∂F M
MM
Mm
∂F m
∂F m
(j,k)
∂F M
.
J¯R :=
−
+
+
(4.3)
(j,k)
(i,j,k)
(i,j,k)
(j,k)
∆tk
∆t
k
∂αM
∂αc
∂αc
∂αM
i=1
m
Since Assumption (a), i.e., NNR
= 1 in Algorithm 2 holds, the discretized version of the solution operator
Ac applied to the linearized problem (3.13) can be explicitly given as
−1
M m ∂F (i,j,k)
(j,k)
(j,k)
(i)
m
(i,j,k)
(i,j,k)
(i,j,k)
(i,j,k)
(α
)
=
α
−
J
α(i,j+1,k)
=
A
R
,
α
,
α
+
with
J
:=
m
c
c
m
c
c
m
M
M
∆tk ∂α(i,j,k)
c
which yields immediately the derivative with respect to the macro scale
−1 ∂F (i,j,k)
∂A(i)
(j,k)
m
c
(i,j,k)
(αM ) ≈ − Jm
.
(j,k)
(j,k)
∂αM
∂αM
(i,j,k)
where the contribution from the Jacobian Jm
contributions
(j,k)
Ac (αM )
(4.4)
is disregarded due to Assumption (b). Summing up all
=
N
GP
X
(j,k)
A(i)
c (αM ),
i=1
plugging them into the macroscale stiffness matrix (3.9) and exploiting (4.4) concludes the proof
MM
dF M (j,k)
!
(j,k)
(j,k)
(j,k)
JM =
+
αM , Ac (αM ) = J¯F .
∆tk
dαM
In practice the assumptions (a) and (b) can be violated and one will end up with a different variant of
the Newton–Raphson scheme. For example: the computation (4.4) involves the derivative of the correction
terms Ac with respect to the macroscale magnetic density αM , In [3], it was proposed to solve several
m
mesoscale problems per Gauß point in parallel: for a two-dimensional problem, Ndim
= 3 problems were
(i)
solved to approximate the Jacobian, the first one with the nominal macroscale source (e.g. αM ) and the
(i)
other two with a perturbated source magnetic density αM + δ (i) αi where δ (i) is a small perturbation and
m
αi is a vector oriented along the x, y or z-axes. Similarly, in 3D Ndim
= 4 problems need to be solved.
On the other hand, one can use a fixed point iteration scheme as suggested in [24], which is actually the
limit case of a a waveform relaxation approach, where time window size equals time step size.
9
1
2
3
4
5
6
7
8
9
Input: macroscale source j s and mesh.
Output: fields (macro/meso), global quantities.
begin
t ← t0 , initialize the macroscale field aM |t0 = aM0 ,
# begin the macroscale time loop (index k)
for (k ← 1 to NTS ) do
# begin the macroscale NR loop (index j)
M
for (j ← 1 to NNR
) do
# parallel resolution of mesoscale problems (index i)
for (i ← 1 to NGP ) do
(i,j,k−1)
downscale the sources αM
,
(i,j,k)
(i,j,k)
= Ac (y, αM
10
compute αc
11
compute the homogenized law HM
(i,j,k)
13
upscale the homogenized law
end
14
assemble the Jacobian JM
12
15
16
17
), see Algorithm 2
(j,k)
=
(i,j,k)
HM
(i,j,k)
and ∂HM
and
(i,j,k)
/∂bM
,
(i,j,k)
(i,j,k)
∂HM /∂bM ,
dF M (j,k)
1
(j,k)
MM +
αM , Ac (αM ) to solve (3.12),
∆tk
dαM
end
end
end
Algorithm 1: Pseudocode for the monolithic FE-HMM
(i,j,k)
1
2
3
4
5
6
7
8
9
10
11
Input: macroscale sources αM
and the mesoscale mesh.
(i,j,k)
m
Output: homogenized law HM , per Gauß point for Ndim
problems.
begin
prescribe periodic boundary conditions, impose sources,
t ← tM , initialize the correction ac |tM ,
m
solve Ndim
mesoscale problems for the k th time step,
m
for (p ← 1 to Ndim
) do
# begin the mesoscale NR loop (index j)
m
for (j ← 1 to NNR
) do
(i,j,k)
assemble the Jacobian Jm
to solve (3.13).
end
end
end
Algorithm 2: Pseudocode for one mesoscale problem
10
5. The waveform relaxation method
Waveform relaxation methods solve time dependent problems iteratively, i.e., they generalize the classical ideas of Gauß–Seidel and Jacobi iteration to the time domain. The method starts with an initial
guess of the solution over a time interval and computes iteratively approximations of increasing accuracy
[25]. Typically the problem is decomposed into subproblems and each subproblem is solved separately.
Let us consider the two ordinary differential equations
∂t y1 = f1 (y1 , y2 )
∂t y2 = f2 (y1 , y2 ).
A monolithic or strongly coupled approach discretizes the problem in time as one system of equations. On
the other hand, an iterative Gauß–Seidel type scheme
(l)
(l)
(l−1)
)
(l)
(l)
(l)
)
∂t y1 = f1 (y1 , y2
∂t y2 = f2 (y1 , y2
(l)
will resolve both equations subsequently, e.g., the first one for the unknown y1 (t) on t ∈ I while considering
(l−1)
(0)
y2
(t) on t ∈ I given and vice-versa. The very first iteration requires an initial guess y2 , which is
typically obtained by constant extrapolation, [26]. In the simplest case an implicit Euler method can be
chosen for time stepping, e.g.
(k,l)
y1
(k,l)
y2
(k−1,l)
− y1
∆tk
(k,l)
, y2
(k,l)
, y2
= f1 (y1
(k,l−1)
)
(k,l)
).
(k−1,l)
− y2
∆tk
= f2 (y1
(k,l)
where y1 describes the unknown y1 at time tk and iteration l; ∆tk denotes the k-th time step size for both
problems. Obviously, the iteration scheme allows to combine different time integrators with independent
time step sizes. It is therefore often referred to as co-simulation or weak coupling. The convergence is
well understood and unconditionally guaranteed for systems of ordinary differential equations [27, 28].
However, already in the case of simple differential algebraic equations, e.g. the system
(l)
(l)
(l)
(l−1)
∂t y1 = f1 (y1 , z1 , y2
0=
(l)
)
(l) (l) (l−1) (l−1)
g1 (y1 , z1 , y2
, z2
)
(l)
(l)
(l)
∂t y2 = f2 (y1 , z1 , y2
0=
(l−1)
, z2
(l) (l) (l)
g2 (y1 , z1 , y2
(l)
, z2
(l)
, z2
with det
∂g1
!
6= 0
(l)
∂z1
)
)
with det
∂g2
(l)
!
6= 0
∂z2
the convergence of the fixed point iteration is conditional. In particular the dependence of algebraic
(l)
equations on old algebraic iterates is critical, i.e., the Jacobian ∂g1 /∂z2 must be sufficiently small, [29].
The convergence for more complex problems, possibly with higher DAE-index, is even more involved, [30].
Waveform relaxation has been originally applied in the simulation of electrical networks but has been
applied in various disciplines. Recently, the method was rediscovered to cosimulate coupled problems [6].
The method converges particularly fast on small intervals and hence it is common to subdivide the time
interval of interest into so called time windows and to apply the method on each time window separately.
This subdivision does not hinder the overall convergence since the error propagation from windows to
window can be controlled [31]. Waveform relaxation is a particular parallel-in-time methods and hence
closely linked to Parareal [32, 33] which has also recently been applied to multiscale problems [34].
11
6. Waveform Relaxation HMM
We employ a waveform relaxation-based approach with windowing [5]. Weak forms similar to (3.1) for
the macroscale and (3.6) for the mesoscale problem are solved on a series of time windows In = (tn−1 , tn ] ⊂
I (n = 1, 2, . . . , NTW ). On each time window, macroscale and mesoscale problems are solved separately
in time-domain, such that waveforms, e.g., αM (t), are obtained. Afterwards the coupling between the
problems is introduced by exchanging the waveforms, and solving the system iteratively. In each waveform
(l)
relaxation iteration l, the resolution of mesoscale problems (for instance with solutions αc (t)), is followed
(l)
by the resolution of the macroscale problem (for instance with the solution αM (t)) until convergence is
(l−1)
(l)
reached (for instance kαM − αM kL∞ (0, T ;L2 (Ω)) < tolM ). In the rest of the section, we will often omit
(j,k,l,n)
(i,j,k,l,n)
(j,k,l)
the time window index n to simplify notation (for instance αM
and αc
become αM
and
(i,j,k,l)
αc
, respectively).
For any given waveform relaxation iteration l, NGP mesoscale problems are solved (in parallel) using
the macroscale source terms from the previous waveform iteration l − 1. In the following section we discuss
these two problems starting with the macroscale.
6.1. The macroscale problem
(i,k,l,n)
are
The waveform relaxation starts with the resolution of mesoscale problems. The solutions αc
then used for computing the homogenized constitutive law needed by the nonlinear macroscale problem
derived from the semi-discrete equations (3.12):
(k,l)
MM
αM
(k−1,l)
− αM
∆tk
(k,l)
+ F M αM , α(k,l)
=0
c
(6.1)
(k,l)
at time points tk . The macroscale and the mesoscale problems
with known (mesoscale) corrections αc
(k,l)
are decoupled and the homogenized constitutive law HM used in (6.1) is upscaled using the formula
Z
1
(k,l)
HM (x, t, bM (x, t)) =
H(x, y, b(k,l)
(x, y, t) + bM (x, t))dy.
c
|Ωm | Ωm
with bM = curlx aM and where the mesoscale field b(k,l)
= curly ac (k,l) is obtained by solving mesoscale
c
problems for a waveform relaxation iteration l as explained in Section 6.2. The decoupling between the
macroscale and the mesoscale problems allows to compute the homogenized Jacobian directly by
Z
(k,l)
∂HM
1
∂H (k,l)
∂bM
=
(b
+ bM )
dy,
(6.2)
∂αM
|Ωm | Ωm ∂bM c
∂αM
since H is known as a closed-form expression. This results from the independence of the mesoscale solutions
(k,l)
(k,l)
αc
on the macroscale source αM . Indeed, each mesoscale problem corresponding to the Gauß point
denoted by i is computed using the macroscale fields from the previous waveform relaxation iteration
(k,l−1)
α(k,l)
= Ac (αM
c
).
Unlike the mesoscale problem in (3.4) which was strongly coupled with the macroscale problem, the
(i,k,l−1)
function Ac is evaluated at each αM
and there is no need to evaluate the derivative ∂Ac /∂αM using
the finite difference method as done in (4.4). Note however that one mesoscale field computation per Gauß
point is needed for each waveform relaxation iteration.
Let us consider the case of a quasi-linear law as e.g. Brauer’s model [10]:
H(b) = µ(b) b
with b = bM + b(k,l)
, the integrand in (6.1) is calculated using
c
∂µ ∂|b|2
∂µ
∂H(b)
= µ(b) +
⊗ b = µ(b) + 2
b⊗b
∂bM
∂|b|2 ∂bM
∂|b|2
where ⊗ denotes the square dyadic product.
12
6.2. Mesoscale problems
Starting from the mesoscale semi-discrete equations (3.13) of Problem 3.4, the following nonlinear
mesoscale problems are derived for the lth waveform iteration:
(i,k,l)
Mm
αc
(i,k−1,l)
− αc
∆tk
(k,l−1)
+ F m α(i,k,l)
, αM
= 0,
c
i = 1, . . . , NGP .
(6.3)
(k,l−1)
with known (macroscale) waveforms αM
(given at time points tk ). In this equation, the macroscale
(i,l−1)
source per Gauß point αM
is considered to be known and taken from the previous waveform relaxation
iteration l − 1. The mesoscale problems defined in (6.3) can then be solved in parallel on the time window
In and the mesoscale solutions stored for all Gauß points of the macroscale grid. These solutions are
later used for computing the homogenized material law and the Jacobian as described in Section 6.1. The
mesocale corrections bc appearing in (6.1) is independent from the macroscale field bM .
If an implicit Euler scheme is used for the macroscale and the mesoscale problems, the overall discretized
system consists in solving the following problem:
Problem 6.1 (Nonlinear, discrete, WR multiscale problem). Find a series of solutions
(k,l)
GP ,k,l)
[αM , α(1,k,l)
, . . . , α(N
] ∈ IRNM +NGP ·Nc
c
c
such that
(k,l)
(k−1,l)
α
− αM
(k,l)
(k,l)
+ FM αM , α(k,l)
=0
RM αM , α(i,k,l)
= MM M
c
c
∆tk
and for the mesoscale problems i = 1, . . . , NGP
(i,k−1,l)
(i,k,l)
− αc
αc
(k,l−1)
(k,l−1)
(i,k,l)
Rm α M
, α(i,k,l)
=
M
+
F
α
,
α
= 0.
m
m
c
c
M
∆tk
(j,k,l)
Let αM
(j,k,l)
FM
(i,j,k,l)
and αc
:=
denote the j th Newton–Raphson iterates. Then we define
(j,k,l)
F M αM , α(j,k,l)
c
(i,j,k,l)
,
F (i,j,k,l)
m
:= F m
(i,j,k,l)
α(i,j,k,l)
, αM
,
c
The following total Jacobian must be computed to resolve Problem 6.1:
(j,k,l)
(j,k,l)
∂FM
0
···
0 ∂FM
M M
∂α(j,k,l) ∂α(1,j,k,l)
c
M
(1,j,k,l)
∂F
m
0
M
0
0
m
0
1
(j,k,l)
(1,j,k,l)
+
=
JR
∂αc
∆tk
..
..
..
.
.
.
0
0
0
0
0
0
0
0 M m
αM
(i,j,k−1,l)
− αM
∆tk
(j,k,l)
···
∂FM
!
. (6.4)
(NGP ,j,k,l)
∂αc
0
..
0
.
0
.
0
(N ,j,k,l)
∂Fm GP
(6.5)
(NGP ,j,k,l)
∂αc
The decoupling of mesoscale from the macroscale solution makes all the elements of the first column equal
(j,k,l)
(j,k,l)
to zero except for ∂FM
/∂αM . Let us state the following result for the sake of completeness:
Proposition 6.1. The use of the Jacobian (6.5) for solving the waveform relaxation problem is equivalent
to the resolution of the following decoupled system for each waveform relaxation iteration l:
(j,k,l)
∂FM
MM
(j,k,l)
J¯F
=
+
(j,k,l)
∆tk
∂αM
13
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
Input: macroscale source j s and mesh.
Output: fields (macro/meso), global quantities.
begin
# begin loops over time windows (index n)
for (n ← 1 to NTW ) do
# begin waveform relaxation (WR) loop (index l)
for (l ← 1 to NWR ) do
# 1. Parallel resolution of meso-problems (index i)
for (i ← 1 to NGP ) do
(l−1)
downscale the sources αM ,
solve mesoscale problems (1 per Gauß point) on [tk , tk+1 ],
save the solution,
end
# 2. Resolution of the macro-problem on [tk , tk+1 ]
t ← tk , initialize the macro-field aM |tk = aMk ,
# begin the macroscale time loop(index k)
for (k ← 1 to NTS ) do
# begin the macroscale NR loop (index j)
M
) do
for (j ← 1 to NNR
# parallel updating of the homogenized law (index i)
for (i ← 1 to NGP ) do
Read the mesoscale fields b(l)
m,
(l)
(l)
update ∂HM /∂bM using (6.1),
(l)
(l)
(l)
upscale the law HM , ∂HM /∂bM ,
end
assemble the matrix and solve,
end
end
end
end
end
Algorithm 3: Pseudocode for the waveform relaxation FE-HMM.
14
Proof. The application of the Schur complement to (6.5) leads to the conclusion.
The implementation of the waveform relaxation method is illustrated in Algorithm 3. We propose a
Gauß–Seidel type scheme between the macro- and the mesoscale, where the mesoscale is solved in parallel.
It starts with a loop over time windows and for each time window, a waveform relaxation loop involving
weakly coupled macroscale and mesoscale problems is carried out. The mesoscale are solved in parallel on
(l−1)
the time interval In using the macroscale sources from the previous WR iteration αM
in the mesoscale
problems. Mesoscale solutions are then stored for later use in the evaluation of the homogenized magnetic
field and of the Jacobian. Then the macroscale problem is solved on In until convergence. The resolution
involves a time discretization that leads to a nonlinear problem that is solved using the Newton–Raphson
method with the Jacobian evaluated using the previously stored mesoscale fields.
Ultimately, the solution of the entire multiscale problem is obtained as follows (only superscripts of
(l,n)
(l,n)
the waveform relaxation and the time window are involved, i.e., the fields involved are ac
and aM ):
(0,n)
the macroscale field aM is initialized with constant extrapolation. The mesoscale problem can then be
(1,n)
solved for the first iteration, i.e., one obtains ac . This is then used for successive waveform relaxation
iterations
(0,1)
→
a(1,1)
c
→ aM
(0,2)
→
a(1,2)
c
→ aM
(0,NTW )
→
TW )
a(1,N
c
→ aM
aM
aM
aM
(1,1)
→
a(2,1)
c
→
...
→ aM WR
(1,2)
→
..
.
a(2,2)
c
→
...
→ aM WR
TW )
a(2,N
c
→
...
→ aM WR
(1,NTW )
→
(N
,1)
(N
,2)
(N
,NTW )
,
.
In addition to the flexible use of different FE bases and meshes at both scales, this approach also
provides a natural setting for the use of different integrators and time step sizes. Communication costs can
also be reduced in the case of parallel computations. As a drawback, the number of iterations for solving
both the macroscale and the mesoscale problems may increase. A rigorous convergence analysis requires a
structural analysis of the coupled problem (3.10)–(3.11). In particular, the DAE index [35] must be known
to guarantee convergence, c.f. [30]. This analysis is beyond the scope of this paper. However, the total
costs for the monolithic and the waveform relaxation methods are estimated in the next Section.
7. Estimation of the computational cost
In this section, we evaluate and compare total computational costs of the monolithic and the waveform
relaxation algorithms. The total cost comprises the computational and the communication costs, which
M
m
M
and Csol
for the one time step costs for
differ from one algorithm to the other. We use Csol
, Ccom , Cass
the mesoscale computations (including Newton iterations), the mesoscale–macroscale communications, the
m
macroscale assembling and the macrocale resolution, respectively. Csol
is the most expensive and dominant
operation. The total number of time steps is denoted by NTS and the number of time windows NTW , such
m
TS
Csol
.
that the computational costs for a time window with NTS /NTW time steps are roughly NNTW
7.1. Monolithic HMM
The total cost of the monolithic algorithm is given by
m
M
CMono = CMono
+ CMono
,
(7.1)
and has two contributions: the mesoscale contribution
m
M
m
m
CMono
= NTS NNR
NGP (Ndim
Csol
+ Ccom ),
(7.2)
and the macroscale contribution
M
M
M
M
CMono
= NTS NNR
(Cass
+ Csol
),
15
(7.3)
M
m
where NTS is the total number of time steps, NNR
and NNR
are the average numbers of Newton–Raphson
m
iterations for macroscale and mesoscale problems to converge, Ndim
is the number of problems solved to
approximate the Jacobian and the other N? are defined in Table 1.
The total cost in (7.1) can be understood from the Algorithms 1 and 2. The overall problem is
discretized in NTS macroscale time steps and a macroscale nonlinear system is solved for each time step
M
using the Newton–Raphson scheme. Therefore, NNR
nonlinear (Newton–Raphson) iterations are performed
for each time step. Each of these iterations involves the evaluation of the Jacobian at NGP Gauß points
m
by solving Ndim
= 3 or 4 mesoscale problems. The macroscale linear system is then assembled and solved.
m
m
The mesoscale cost Csol
involves the resolution of nonlinear mesoscale problems over one time step. NNR
M
nonlinear iterations are needed for each time step at the mesoscale and 2NTS NNR NGP communications
involving the parallel transfer of small chunks of information are needed for the overall time interval.
The macroscale assembling and resolution costs can be disregarded with respect to the mesosclae costs,
therefore
m
CMono ≈ CMono
.
7.2. Waveform relaxation HMM
The total cost for the WR algorithm is given by:
m
M
CWR = CWR
+ CWR
.
(7.4)
In the general case where NTW time windows In = (tn , tn+1 ] are used. The two contributions in (7.4) are
the mesoscale cost
NTW
m
m
Ccom
(7.5)
CWR
= NTS NWR NGP Csol
+
NTS
m
= NTS NWR NGP Csol
+ NTW NWR NGP Ccom ,
(7.6)
and the macroscale cost
M
M
m
M
M
CWR
= NTW NWR NNR
NGP Cjac
+ Cass
+ Csol
.
(7.7)
m
Additionally, Cjac
is the cost for reading a pre-stored mesoscale field map and evaluating the Jacobian
M
is the average number
for all time steps of the time window using (6.1) for each time step, and NNR
of Newton–Raphson iterations for macroscale problems to converge. These costs can be made small
m
compared to the mesoscale computational cost Cdim
and the communication cost Ccom by the use of a
smart implementation.
The total cost in (7.4) can be understood from the Algorithm 3. The overall problem is discretized and
solved on NTW time windows and for each time window In , a waveform relaxation loop involving NWR
WR iterations during which mesoscale problems are solved and stored. The communication cost involves
TS
the transfer of NNTW
NGP NNR communications of bigger chunks of informations for each time window.
M
The nonlinear macroscale problem is solved using the Newton–Raphson scheme. Therefore, NNR
nonlinear iterations are carried for each time step and for each Newton–Raphson iteration, the Jacobian
is computed for NGP Gauß points. This is done by reading mesoscale field maps for each Gauß point
and then evaluating the homogenized law using (6.1). The macroscale linear system is then assembled
and solved. The reading of mesoscale fields maps and the update of the homogenized law are one of the
leverage for accelerating computations in the context of the waveform relaxation method.
Neglecting the macroscale assembling and resolution costs, equation (7.4) can be approximated by
NTW
M
m
m
NNR
Cjac
+ Ccom ,
CWR ≈ NTS NWR NGP Csol
+
NTS
where NTS is the total number of time steps. The following theorem allows to compare computational
costs for the monolithic and the waveform relaxation approaches.
16
Theorem 7.1. The computational costs for the monolithic and the waveform relaxation methods are
respectively given by the following approximations:
M
m
m
CMono ≈ NTS NNR
NGP ( Ndim
Csol
+ Ccom ) ,
(7.8)
and
CWR
NTW
m
M
m
≈ NTS NWR NGP Csol +
NNR Cjac + Ccom .
NTS
(7.9)
Proof. The theorem results from the developments of Sections 7.1 and 7.2. The high level of parallelization
as explained in Algorithms 1-2 and 3 results from the independence of mesoscale problems.
Moreover, the two approaches can easily be parallelized.
Remark 7.2. The computational cost of the waveform relaxation method can be decreased by minimizing
the cost related to the reading of the mesoscale fields and the update of the homogenized law.
Assume that there exists κ ∈ (0, 1) such that
NTW
NTW M m
m
N C = κ(Csol
+
Ccom )
NTS NR jac
NTS
which is a reasonable assumption because the cost due to the computation of the material law using the
pre-stored mesoscale maps on the whole time window is small compared to the cost due to the computation
of mesoscale problems on the same time window and the communication. Then the relationship
NTW
NTW
M
m
m
m
NWR Csol +
NNR Cjac + Ccom
= NWR (1 + κ)Csol +
(1 + κ) Ccom
NTS
NTS
M
m
m
< NNR
(Ndim
Csol
+ Ccom )
between (7.8) and (7.9) shows that the waveform relaxation method is more efficient if
NWR <
m
Ndim
NM
(1 + κ) NR
(7.10)
and each time window consists of at least NTS > 2 time steps which is a rather technical assumption. As
can be seen from relation (7.10), reducing the number of time windows (NTW ) reduces the communication
cost between the mesoscale and macroscale problems. Additionally, the reduction of cost due to the
evaluation of the Jacobian minimizes the overall cost of the waveform relaxation method.
8. Application
We use a soft magnetic composite (SMC) material to test the ideas developed in the previous sections.
An idealized 2D periodic SMC (with 20 × 20 grains) surrounded by an inductor is considered.
For the first series of numerical tests we use the SMC structure depicted in Figure 3 (only 10 × 10
grains are shown). The geometry has been chosen such that the vector potential formulation b = curl a
as described in Section 2 with a = (0, 0, az ) can be used. The magnetic flux density b = (bx , by , 0) lives
in the xy-plane. Only a quarter of the structure is considered for numerical computations thanks to the
symmetry (see Figure 4 - left for the reference geometry and Figure 4 - right for the geometry used for the
homogenized problem). In both cases, the following boundary conditions are imposed on Γinf , Γh and Γv :
(n · b)|Γinf = 0 ⇐ (n × a)|Γinf = 0,
(n · b)|Γh = 0 ⇐ (n × a)|Γh = 0,
(n × h)|Γv = 0.
Using Ampères equation (2.2 a), the source current j s must be imposed perpendicular to the xy-plane
j s = (0, 0, js ) with js = js0 s(t) = js0 sin(2πf t). We consider the operating frequency f = 50 kHz which
corresponds to λ = 6000m). The wavelength of the source is much larger compared to the length of the
structure (' 500µm) so that the assumption of a magnetoquasistatic problem can be made.
17
. j
Inductor
SMC
egap
Air
j
L
ea
ei
Figure 3: Soft magnetic composite two-dimensional used geometry. Two
√ opposite source current are imposed in the top and
bottom inductors. The lengths are given by L = 1000 µm, ea = 150 2/2 µm, ei = 100 µm and egap = 100 µm. Only 100
grains out of 400 are drawn on the image.
Γinf
Γinf
Γv
Γv
Inductor
Inductor
SMC
SMC
Air
Air
Γh
Γh
Figure 4: Geometry used for computations. Only a quarter of the geometry is used thanks to the symmetries. Left: reference
geometry. Only 25 grains out of 100 are drawn on the image. Right: homogenized geometry.
18
We consider isotropic materials and therefore, the magnetic field h has only xy components. We use
the same material properties as those used in [2]: the insulation material is linear isotropic (with µr = 1
and σ = 0). The conductor has an isotropic electric conductivity σ = 5 MS and is also governed by the
following nonlinear magnetic law [10]:
H(bε ) = α + β exp(γ||bε ||2 ) bε
with α = 388, β = 0.3774 and γ = 2.97. The reference solution is obtained by solving a FE problem on an
extremely fine mesh (347,324 elements) of the whole SMC structure. The mesoscale problems are solved
on a square elementary cell meshed with (4,215 elements). Results obtained using the newly developed
waveform relaxation (WR) approach subscripted WR are compared to the reference results (subscripted
“Ref”) obtained solving the reference problem (2.6) on a very fine mesh and results of the monolithic
approach (subscripted“Mono”) obtained solving problem (3.1) and (3.6).
Quantities of interest (global quantities and errors) are defined and used for numerical validation. The
global quantities are the reference, the monolithic and the WR eddy currents losses:
τ PRef (t) = q(aε ),
τ Pmono (t) = q(am ),
where the functionals q is defined as:
Z
(σ|∂t uε (x, t)|) dx
Ω
q(u) = Z c 1 Z
2
(σ|∂
u
(x,
y,
t)|
)
dy
dx
t m
|Ωm | Ωmc
Ω
τ PlWR (t) = q(alm ),
if u is the reference solution
if u is the solution of the multiscale method
Equivalent quantities can be defined in terms of the magnetic energy and the magnetic power. Two types
of errors are defined: the relative error on global quantities
Errlτ P = δrel (τ PlWR , τ PMono ),
and the relative error on the fields uM
ErrlWmag = δrel (WmaglWR , WmagMono ),
(l)
Erru = δ̄rel (uM )
(8.1)
(8.2)
where u stand for the fields bM , ∂t aM and ∂t bM . The functions δrel and δ̄rel are defined by:
δrel (v, w) =
||v − w||L∞ (0,T )
,
||w||L∞ (0,T )
δ̄rel (u) =
||ul − ul−1 ||L∞ (0,T ;L∞ (Ω))
.
||u0 ||L∞ (0,T ;L∞ (Ω))
8.1. Numerical convergence analysis
The monolithic HMM and the WR HMM algorithms have been implemented in the open source software
GetDP [36] using a finite element formulation. Two cases are considered for the numerical validation of the
method: (a) the case with 1 time window and the same time stepping at the macroscale and the mesoscale;
(b) the case with 1 time window and different time stepping at the macroscale and the mesoscale.
For case (a), an excellent agreement is obtained between the WR solutions and the monolithic solutions
to which the monolithic solutions are expected to converge for the eddy current losses (Figure 5 - left) and
the magnetic energy (Figure 5 - right). Table 2 depicts the evolution of the relative L∞ error defined in
(8.1) on eddy currents losses and the magnetic energy between the WR and the monolithic approach. As
can be seen from this table, the error between the WR and the monolithic cases is smaller than the error
between the monolithic and the reference cases after the third WR iteration. These findings were verified
for all the numerical tests that were run. Figure 6 - left. Left depicts the evolution of the relative L∞
error on the electromagnetic fields defined in (8.2) as a function of the WR iterations. This criterion is to
be used to control the error in case the monolithic solution is not available.
Also for case (b) a good agreement is obtained between the waveform relaxation solutions and the
monolithic solutions. Figure 6 - right illustrates the case with different macroscale time discretization (10,
25 and 50 time steps per period) but with the same time discretization at the mesoscale (50 time steps
per period). In particular, a good agreement is observed for values of the error up to 10−4 . Beyond this
19
0.1
Ref
Mono
WR l = 1
WR l = 2
4
Magnetic energy (J)
Eddy current losses (W)
6
2
0
0
1 · 10−5
2 · 10−5
Time (s)
3 · 10−5
4 · 10−5
8 · 10
Ref
Mono
WR l = 1
WR l = 2
−2
6 · 10−2
4 · 10−2
2 · 10−2
0
0
1 · 10−5
2 · 10−5
Time (s)
3 · 10−5
4 · 10−5
Figure 5: Instantaneous eddy current losses (left) and magnetic power (right) for the reference, the monolithic and the WR
approaches. An overall mesh of 8722 elements with 25 elements for the homogenized domain and a time step ∆t = 2 · 10−7 s
were used. Only results for the first two WR iterations are shown.
Table 2: Convergence of the eddy current losses and the magnetic energy as a function of the WR iterations. The relative
errors Errτ P and ErrWmag between the reference and the monolithic approaches are 2.12 · 10−2 and 1.24 · 10−2 , respectively.
WR iteration l
1
2
3
4
5
6
7
8
9
10
Errlτ P
6.68 · 10−1
8.08 · 10−2
1.17 · 10−2
1.87 · 10−3
3.78 · 10−4
1.41 · 10−4
1.05 · 10−4
9.0 · 10−5
9.8 · 10−5
9.8 · 10−5
Errlτ P − Errl−1
τP
−
5.87 · 10−1
6.90 · 10−2
9.84 · 10−3
1.49 · 10−3
2.35 · 10−4
3.80 · 10−5
6.00 · 10−6
< 10−6
< 10−6
20
ErrlWmag
1.17 · 10−1
1.45 · 10−2
2.08 · 10−3
3.18 · 10−4
5.01 · 10−5
8.00 · 10−6
1.30 · 10−6
2.00 · 10−6
< 10−6
< 10−6
ErrlWmag − Errl−1
Wmag
−
1.03 · 10−1
1.24 · 10−2
1.77 · 10−3
2.68 · 10−4
4.21 · 10−5
6.70 · 10−6
1.10 · 10−6
< 10−6
< 10−7
T S10
T S25
T S50
10−2
ErrlWmag
10−1
Errlf
10−1
bM
∂t aM
∂t bM
10−3
−5
10
10−3
10−4
10−5
10−7
10−6
2
4
6
8
10
2
4
Waveform relaxation l
6
8
10
Waveform relaxation l
Figure 6: Left: Convergence of the macroscale waveforms for successive WR iterations (case with 1 time window and the
same time discretization at the macroscale and the mesoscale). Right: Relative error on magnetic energy as a function of the
WR iterations. The macroscale time grid comprise 10, 25 and 50 time steps per period, respectively whereas the time grid
for the mesoscale problems comprise 50 time steps per period. In both cases, a macroscale mesh with 8722 elements with 25
elements for the homogenized domain was used. The relative error between the reference and the monolithic plots is 0.01243.
Joule losses
Magnetic energy
10−1
Relative errors
Relative errors
10−1
10−2
10−2
Joule losses
Magnetic energy
10−3
5
10
15
20
25
2
4
Time steps ∆t (s)
Number of mesoproblems
6
8
·10−7
Figure 7: Relative errors on eddy current losses and magnetic energy between the waveform relaxation and the reference
cases. Left: Relative errors as a function of the number of mesoproblems (1 to 25 mesoproblems are considered for the
homogenized domain). Right: Relative errors as a function of the time discretization. The time step ranges from 2.5 · 10−8 s
to 8 · 10−7 s.
value, smaller errors are even obtained for finer macroscale time grids. This underlines the possibility for
efficient and consistent usage of multirate time stepping
Figure 7 depicts the convergence of the eddy current losses and the magnetic field to the reference
solutions when the spatial grid and time grid are refined. As can be seen in Figure 7 - left, the relative
errors decrease as the number of mesoproblems is increased from 1 to 25. One Gauss point was considered
for each element and therefore the macroscale mesh for the homogenized domain contains the same number
of elements. Figure 7 - right depicts the same evolution for the case for different time discretizations. In
this case, a linear convergence is observed for the eddy current losses (with the time derivative) whereas
the curve for magnetic energy exhibits a faster convergence. A good agreement was also observed in the
case of many time windows.
An empirical comparison of the computational cost between the two methods can be made on the
basis of the formula (7.10). In our numerical computations, we have always found that the errors on eddy
currents losses and on the magnetic energy (Errlτ P and ErrlWmag ) become smaller than the errors between
monolithic and the reference quantities already at the third waveform relaxation iteration (see Table 2).
The first iteration is not computationally costly as it involves the initialization of the mesoscale solution
to zero. Therefore, it is reasonable to assume that the errors of both methods become comparable for
21
NWR = 2. Neglecting the mesoscale costs related to the reading and the updating of the homogenized
m
m
M
constitutive law (i.e.: κ → 0) and considering NNR
= 3 for a residue of 10−6 so that NNR
= NNR
, then
M
m
M
m
m
NWR < NNR Ndim with NWR = 2, NNR = 3, Ndim = 4 for a three-dimensional problem (resp. Ndim
=3
for a two-dimensional problem) and a theoretical speed-up of 4.5 for two-dimensional problems (resp. 6 for
three-dimensional problems) can be gained. However, the current proof-of-concept implementation only
allows a speed up of 2.
9. Conclusion
In this paper the heterogeneous multiscale method was combined with the waveform relaxation method.
An efficient algorithm exploiting exact Jacobian information based on Schur-complements was proposed.
Estimates have shown that an optimal implementation of the algorithm can be expected to be up to 6
times faster than a comparably monolithic approach. In the case of multirate behavior, even higher speedup are expected. Convergence and efficiency have been numerically investigated using a challenging test
example. Finally, optimization of our implementation and applying the available convergence analysis of
waveform relaxation for higher index differential algebraic systems is subject of future research.
Acknowledgment
This work was supported by the German Funding Agency (DFG) by the grant ‘Parallel and Explicit
Methods for the Eddy Current Problem’ (SCHO-1562/1-1), the ‘Excellence Initiative’ of the German
Federal and State Governments and the Graduate School CE at Technische Universität Darmstadt and
the Belgian Science Policy under grant IAP P7/02 (Multiscale modelling of electrical energy system).
22
References
[1] W. E, B. Engquist, Heterogeneous multiscale methods, Communications in Mathematical Sciences
1 (1) (2003) 87–132.
[2] I. Niyonzima, R. Sabariego, P. Dular, F. Henrotte, C. Geuzaine, Computational homogenization for
laminated ferromagnetic cores in magnetodynamics, IEEE Transactions on Magnetics 49 (5) (2012)
2049–2052.
[3] I. Niyonzima, Multiscale finite element modeling of nonlinear quasistatic electromagnetic problems,
Ph.D. thesis, Université de Liège – Belgium (September 2014).
[4] O. Bottauscio, A. Manzin, Comparison of multiscale models for eddy current computation in granular
magnetic materials, Journal of Computational Physics 253 (2013) 1–17.
[5] J. K. Jacob K. White, F. Odeh, A. L. Sangiovanni-Vincentelli, A. E. Ruehli, Waveform relaxation:
Theory and practice, Technical Report UCB/ERL M85/65, University of California, Berkley (1985).
[6] S. Schöps, H. De Gersem, A. Bartel, A cosimulation framework for multirate time-integration of
field/circuit coupled problems, IEEE Transactions on Magnetics 46 (2010) 3233–3236.
[7] J. D. Jackson, Classical electrodynamics, 3rd Edition, John Wiley & Sons, 1998.
[8] K. Schmidt, O. Sterz, R. Hiptmair, Estimating the eddy-current modeling error, IEEE Transactions
on Magnetics 44 (6) (2008) 686–689.
[9] A. Bossavit, Computational Electromagnetism. Variational Formulations, Complementarity, Edge
Elements, Academic Press, 1998.
[10] J. Brauer, Simple equations for the magnetization and reluctivity curves of steel, IEEE Transactions
on Magnetics 11 (1) (1975) 81.
[11] R. R. R., Représentation algébrique des caractéristiques magnétiques, Bulletin de la Société Française
des Electriciens - Tome VI 5 (69) (1936) 881–892.
[12] F. Bachinger, U. Langer, J. Schöberl, Numerical analysis of nonlinear multiharmonic eddy current
problems, Numerische Mathematik 100 (4) (2005) 593–616.
[13] A. Abdulle, Numerical homogenization methods for parabolic monotone problems, MATHICSE Technical Report 33.2015, EPFL, Lausanne (2015).
[14] H. Brezis, Functional analysis, Sobolev spaces and partial differential equations, Springer Science &
Business Media, 2010.
[15] A. Visintin, Electromagnetic processes in doubly-nonlinear composites, Communications in Partial
Differential Equations 33 (2008) 804–841.
[16] A. Bensoussan, J.-L. Lions, G. Papanicolaou, Asymptotic Analysis for Periodic Structures, American
Mathematical Society, 2011.
[17] A. Visintin, Homogenization of doubly-nonlinear equations, Rendiconti Lincei - Matematica E Applicazioni 17 (2006) 211–222.
[18] P. Henning, M. Ohlberger, A Newton-scheme framework for multiscale methods for nonlinear elliptic
homogenization problems, in: Proceedings of the Conference Algoritmy, 2015, pp. 65–74.
[19] V. Barbu, Nonlinear semigroups and differential equations in Banach spaces, Leyden-Noordhoff, 1976.
[20] H. Brezis, Opérateurs maximaux monotones et semi-groupes de contractions dans les espaces de
Hilbert, North-Holland, Amsterdam, 1973.
23
[21] F. E. Browder, Existence theorems for nonlinear partial differential equations, in: Proc. Sympos. Pure
Math, Vol. 16, 1970, pp. 1–60.
[22] A. Abdulle, G. Vilmart, Coupling heterogeneous multiscale FEM with Runge–Kutta methods for
parabolic homogenization problems: A fully discrete spacetime analysis, Mathematical Models and
Methods in Applied Sciences 22 (06) (2012) 1250002/1–1250002/40.
[23] R. Grb, M. Gnther, U. Wever, Q. Zheng, Optimization of parallel multilevel-Newton algorithms on
workstation clusters, in: Euro-Par’96 Parallel Processing, 1996, pp. 91–96.
[24] A. Abdulle, M. E. Huber, V. Gilles, Linearized numerical homogenization method for nonlinear monotone parabolic multiscale problems, Multiscale Modeling & Simulation 13 (3) (2015) 916–952.
[25] M. J. Gander, Waveform relaxation, in: Encyclopedia of Applied and Computational Mathematics,
Springer, 2012.
[26] M. Arnold, M. Günther, Preconditioned dynamic iteration for coupled differential-algebraic systems,
BIT Numerical Mathematics 41 (1) (2001) 1–25.
[27] E. Lelarasmee, A. E. Ruehli, A. L. Sangiovanni-Vincentelli, The waveform relaxation method for timedomain analysis of large scale integrated circuits, IEEE Transactions on Computer-Aided Design of
Integrated Circuits and Systems 1 (3) (1982) 131–145.
[28] K. Burrage, Parallel and sequential methods for ordinary differential equations, Oxford University
Press, Oxford, 1995.
[29] U. Miekkala, O. Nevanlinna, Convergence of dynamic iteration methods for initial value problems,
SIAM Journal on Scientific Computing 8 (1987) 459–482.
[30] M. L. Crow, M. D. Ilic, The waveform relaxation method for systems of differential/algebraic equations, Mathematical and Computer Modelling 19 (12) (1994) 67–84.
[31] A. Bartel, M. Brunk, M. Günther, S. Schöps, Dynamic iteration for coupled problems of electric
circuits and distributed devices, SIAM Journal on Scientific Computing 35 (2) (2013) B315–B335.
[32] J.-L. Lions, Y. Maday, G. Turinici, A parareal in time discretization of PDE’s, Comptes Rendus de
lAcadmie des Sciences - Series I - Mathematics 332 (2001) 661–668.
[33] M. J. Gander, S. Vanderwalle, Analysis of the parareal time-parallel time-integration method, SIAM
Journal on Scientific Computing 29 (2) (2007) 556–578.
[34] M. Astorino, F. Chouly, A. Quarteroni, Multiscale coupling of finite element and lattice Boltzmann
methods for time dependent problems, MOX–Report 47/2012, Politecnico di Milano (2012).
[35] E. Hairer, G. Wanner, Solving Ordinary Differential Equations: Stiff and differential-algebraic problems, Springer Verlag, 2010.
[36] P. Dular, C. Geuzaine, F. Henrotte, W. Legros, A general environment for the treatment of discrete
problems and its application to the finite element method, IEEE Transactions on Magnetics 34 (5)
(1998) 3395–3398.
24
| 5cs.CE
|
On the Direction of Discrimination:
An Information-Theoretic Analysis of Disparate
Impact in Machine Learning
Hao Wang, Berk Ustun, and Flavio P. Calmon
arXiv:1801.05398v1 [cs.IT] 16 Jan 2018
John A. Paulson School of Engineering and Applied Sciences
Harvard University
Emails: hao [email protected], {berk,flavio}@seas.harvard.edu
Abstract—In the context of machine learning, disparate impact
refers to a form of systematic discrimination whereby the output
distribution of a model depends on the value of a sensitive
attribute (e.g., race or gender). In this paper, we present an
information-theoretic framework to analyze the disparate impact
of a binary classification model. We view the model as a fixed
channel, and quantify disparate impact as the divergence in
output distributions over two groups. We then aim to find
a correction function that can be used to perturb the input
distributions of each group in order to align their output
distributions. We present an optimization problem that can be
solved to obtain a correction function that will make the output
distributions statistically indistinguishable. We derive closed-form
expression for the correction function that can be used to compute
it efficiently. We illustrate the use of the correction function for
a recidivism prediction application derived from the ProPublica
COMPAS dataset.
I. I NTRODUCTION
Machine learning (ML) models aim to exploit biases in
the training data to predict an outcome variable of interest.
In many real-world applications, however, effective prediction
should not be achieved by discriminating on a sensitive
attribute, such as race or gender [1, 2].
Discrimination can occur directly when a sensitive attribute
is used as an input to the model, known as disparate treatment.
More pervasive today is a phenomenon known as disparate
impact [3], where a sensitive attribute is omitted from the
model, but still affects its predictions through correlations with
“proxy” variables (e.g., income, education level). The potential
to discriminate by proxy manifests itself in contexts outside of
ML. In the United States, for example, racial minorities were
indirectly denied financial services by exploiting correlations
between race, address, and income – a practice known as
redlining [4].
Disparate impact is likely to arise as an artifact of empirical
loss minimization when a sensitive attribute is valuable for
the purposes of prediction, and has proxy variables in the
training data. A large body of recent work has documented this
phenomenon in real-world applications ranging from online
advertising [5] to recidivism prediction [6]. When used in
human or algorithmic decision-making, models with disparate
impact may not only violate anti-discrimination laws [3], but
inadvertently amplify societal biases [7].
Such issues have motivated a growing stream of technical
work on disparate impact in ML, focusing on topics such as:
(i) how to identify and quantify disparate impact [8, 9, 10, 11];
(ii) how to train models that mitigate disparate impact [12, 13,
14]; and (iii) how to identify causal factors of discrimination
[15]. The present work is inscribed within the first research
direction.
In this paper, we consider the disparate impact problem
from an information-theoretic perspective. Our goal is to
derive a correction function that can be used to identify
features that act as “proxies” of a sensitive attribute for a
fixed prediction model. We first present a correction function
that has an information-theoretic interpretation in terms of
error exponents of binary hypothesis testing (Section II). We
then derive closed-form expressions for the correction function
that can easily be computed using the prediction model,
which uses the features X to predict the outcome Y , and a
group membership distribution, which uses the features X to
“predict” the sensitive attribute S (Section III). Our approach
is inspired by recent work in information-theoretic privacy
[16], which takes a similar route to analyzing the behavior of
error exponents under small perturbations. We illustrate our
framework on a recidivism prediction problem derived from
the ProPublica COMPAS dataset [6] (Section IV).
II. F RAMEWORK
We consider a channel WY |X , which takes as input a
vector of d random variables X = (X1 , . . . , Xd ) ∈ X and
produces as output a random variable Y ∈ Y. We assume
that the support sets X and Y are finite. In practice, WY |X
represents a predictive model (e.g., a linear classifier to predict
recidivism), X represents a vector of features (e.g., Age,
Salary), Y represents the predicted output of WY |X given X
(e.g., Y = 1 iff the model predicts that a prisoner with features
X will commit a crime after being released from prison).
We seek to characterize differences in the output distribution
of the channel WY |X with respect to a sensitive attribute S.
We focus on the case where the sensitive attribute is binary
S ∈ {0, 1}, and use PX|S=0 , PX|S=1 and PY |S=0 , PY |S=1
to denote the conditional distributions of inputs and outputs,
respectively. We say that a channel WY |X has disparate impact
with respect to S when PY |S=0 6= PY |S=1 .We assume that
output space
input space
WY |X
PX|S=0
PY |S=0
QX
QY
PY |S=1
PX|S=1
Fig. 1: Illustration of correction paths on the probability simplex. The
green path corresponds to aligning only output distributions. The red path
corresponds to aligning the input distributions and output distributions.
WY |X does not use the sensitive attribute S, as doing so would
violate legal constraints in applications such as hiring and
credit scoring (see, e.g., [3]). In this setting, the Markov condition S → X → Y ensures that PY |S=0 = WY |X ◦ PX|S=0
and PY |S=1 = WY |X ◦PX|S=1 . Thus, disparate impact occurs
when PX|S=0 6= PX|S=1 .
Given a fixed channel WY |X , disparate impact can be
reduced by perturbing PX|S=0 into a new distribution QX such
that the corresponding output distribution QY = WY |X ◦ QX
is “closer” to PY |S=1 (cf. Fig. 1). Intuitively, larger disparities
between output distributions require larger perturbations, and
the direction between QX and PX|S=0 identifies which components of X contribute to this disparity. We formally define
this setup next.
Definition 1. Let J(·k·) be a convex divergence measure (e.g.
total variation, KL-divergence), and PS,X,Y be fixed, where
S → X → Y . For λ , (λ1 , λ2 , λ3 , λ4 ) with λi ≥ 0, we
define the objective function as follows.
Lλ (QX ) , λ1 J(QX kPX|S=0 ) + λ2 J(QX kPX|S=1 )
+ λ3 J(QY kPY |S=0 ) + λ4 J(QY kPY |S=1 )
(1)
where QY , WY |X ◦ QX .
The correction path between PX|S=0 and PX|S=1 is given
by the solution of the optimization problem
Lλopt
, min Lλ (QX ).
QX
(2)
From standard convexity results, e.g. [17], as the values
of λi are changed, the distribution QX which achieves the
minimal value given in (2) describes the lower boundary of
the set
{(J(QX kPX|S=0 ), J(QX kPX|S=1 ),
J(QY kPY |S=0 ), J(QY kPY |S=1 )) | supp(QX ) ⊆ X }.
The solutions of (2) form different paths on the probability
simplex, as illustrated in Fig. 1. For λ2 = λ4 = 0, the optimal
solution is QX = PX|S=0 . When λ2 = λ3 = 0, QX will
traverse the shortest path (as measured by J(·k·)) between
PX|S=0 and the set {PX | J(WY |X ◦ PX kPY |S=1 ) = 0},
corresponding to the green line in Fig. 1. Note that this path
transforms PX|S=0 into a distribution devoid of disparate
impact, but potentially different from PX|S=1 . Perhaps of
greater interest is the path delineated by setting λ3 = 0
and varying λ1 , λ2 and λ4 . This corresponds to the red path
in Fig. 1, traversing between PX|S=0 and PX|S=1 , while
controlling the similarity of the induced distribution on Y .
Our goal is to produce a correction function that indicates
which components of X are contributing to the observed disparate impact. More precisely, the correction function is a local
(multiplicative) perturbation of PX|S=0 (see Definition 2) that
decreases the objective function (1) the most. This definition
leads to correction functions that can be cast in terms of
predictive models for S and Y given X, as illustrated in
Section III.
Definition 2. For a given function f ∈ L(PX|S=0 ), we define
the perturbed distribution PeX|S=0 as
PeX|S=0 (x) , PX|S=0 (x)(1 + f (x)),
(3)
where > 0 is sufficiently small so that PeX|S=0 is a valid
probability distribution, and
L(PX|S=0 ) ,
f : X → R | E [f (X)|S = 0] = 0, E f (X)2 |S = 0 = 1 .
Moreover, PeY |S=0 , WY |X ◦ PeX|S=0 .
Next, we denote ∆λ (f ) as the decrease in the objective
function (1) by locally perturbing the distribution PX|S=0 .
Definition 3. For a given λ , we define ∆λ (f ) as
∆λ (f ) , lim+
→0
Lλ (PeX|S=0 ) − Lλ (PX|S=0 )
.
(4)
The function f which achieves the minimal value of ∆λ (f )
is defined as the correction function next.
Definition 4. For a given λ , the correction function fe is
defined as
fe = argmin ∆λ (f ).
(5)
f ∈L(PX|S=0 )
Connections to Binary Hypothesis Testing
In the remainder of this paper, we consider Y binary
and KL-divergence as the divergence measure, i.e., J(·k·) =
DKL (·k·). In this case, the objective function (1) can be
expressed as:
Lλ (QX ) = λ1 DKL (QX kPX|S=0 ) + λ2 DKL (QX kPX|S=1 )
+ λ3 DKL (QY kPY |S=0 ) + λ4 DKL (QY kPY |S=1 ).
(6)
Our choice of DKL (·k·) is motivated by its relationship
with the error exponent in hypothesis testing ([17, Chap. 11],
[18, 19]). Specifically, when λ3 = λ4 = 0 in (2), the correction
path describes the best trade-off (in terms of the first order term
in the exponent) between the Type I and Type II error for a
hypothesis test that seeks to distinguish S given an observation
of X. The optimal value of (2) can then be expressed in
terms of the Rényi’s α-divergence (cf. [19, Section II-A]),
i.e., Dα (PX|S=0 kPX|S=1 ). In other words, under the choice
of KL-divergence the correction path can be understood as the
trade-off between error exponents of two (independent) binary
hypothesis tests: one to distinguish S from X; and another to
distinguish S from Y .
We refer to the functions (fm , gm ) that attain the maximum
as the principal functions. The maximal correlation can be
equivalently given by
r h
i
III. M AIN R ESULTS
Theorems 1 and 2 characterize the correction function fe.
Theorem 1 shows that the correction function is the linear
combination of the log-likelihood ratio function fl and the
principal function fm . Theorem 2 shows that fl and fm can
be expressed in terms of the group membership distribution
PS|X and the channel WY |X , respectively.
In this section, we derive closed-form expressions for the
correction function under the objective given in (6). We
leverage the connection between small perturbations of KLdivergences and maximal correlation (noted, for example, in
[20, 21]).
Our main results consist of Theorems 1 and 2, where we
prove that the correction function that minimizes (5) is a
linear combination of two components: fl , which “aligns”
the perturbed input distribution with PX|S=1 ; and fm , which
“aligns” the corresponding output distribution with PY |S=1 .
In Theorem 2, we show that fl and fm (and thus fe) can be
expressed in terms of the group membership distribution PS|X
and the channel WY |X . This result has an important practical
benefit: it allows us to compute the correction function fe
directly using only PS|X and WY |X , without computing
the complete joint distribution PS,X,Y . In what follows, we
present a formal statement of these results.
We start our derivation of the correction function by providing a simplified expression for ∆λ (f ) in the following lemma.
Lemma 1. For a given λ , ∆λ (f ) can be simplified as
#
"
PX|S=0 (X)
S=0
∆λ (f ) = λ2 E f (X) log
PX|S=1 (X)
"
#
PY |S=0 (Y )
+ λ4 E g(Y ) log
S=0 ,
PY |S=1 (Y )
2
E E [gm (Y )|X, S = 0] S = 0 .
Theorem 1. Given λ , the correction function fe has the form
fe = nl fl + nm fm ,
(10)
where nl and nm are constants computed as:
nl = q
nm = q
−λ2
λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1
2
,
+ (λ2 a2 )2
−λ4 ρm (PX|S=0 ; WY |X )b1
,
2
λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1 + (λ2 a2 )2
a1 = E [fl (X)fm (X)|S = 0] ,
r h
i
2
a2 = E (fl (X) − a1 fm (X)) S = 0 ,
b1 = E [gl (Y )gm (Y )|S = 0] .
Proof. See Appendix A-B.
(7)
where g(y) = E [f (X)|Y = y, S = 0].
Theorem 2. The log-likelihood ratio function fl and the
principal function fm can be expressed as
"
#
PS|X (0|x)
PS|X (0|X)
fl (x) = log
− E log
S = 0 , (11)
PS|X (1|x)
PS|X (1|X)
(gm (1) − gm (0))WY |X (1|x) + gm (0)
.
(12)
ρm (PX|S=0 ; WY |X )
p
p
Here, gm (0) =
p/(1 − p) and gm (1) = − (1 − p)/p
where p , PY |S=0 (1).
fm (x) =
Proof. See Appendix A-A.
We now present definitions of log-likelihood ratio functions
and principal functions that will be used to derive the correction function (see, e.g., [22] for further details on maximal
correlation).
Definition 5. The log-likelihood ratio functions fl and gl are
given by
"
#
PX|S=0 (x)
PX|S=0 (X)
− E log
S = 0 , (8)
fl (x) , log
PX|S=1 (x)
PX|S=1 (X)
"
#
PY |S=0 (y)
PY |S=0 (Y )
gl (y) , log
− E log
S = 0 . (9)
PY |S=1 (y)
PY |S=1 (Y )
Definition 6. The maximal correlation between X and Y
given S = 0 is defined as
ρm (PX|S=0 ; WY |X ) ,
ρm (PX|S=0 ; WY |X ) =
max
f ∈L(PX|S=0 )
g∈L(PY |S=0 )
E [f (X)g(Y )|S = 0] .
Proof. See Appendix A-C.
Combining Theorems 1 and 2, we obtain a closed-form
expression for the correction function fe. Note that, due to our
definition of fe in terms of local perturbations, the correction
function does not depend on λ1 and λ3 in (6). Corollary 1
states the corresponding value of ∆λ (fe).
Corollary 1. For a given λ , ∆λ (fe) is given by
q
2
∆λ (fe) = − λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1 + (λ2 a2 )2
where a1 , a2 , b1 are defined in Theorem 1.
We instantiate our results for λ2 = 0 next. This corresponds
to the green line in Fig. 1, where our main objective is to align
only the output distributions. Unsurprisingly, the following
corollary shows that the correction function fe under this
scenario is (up to a sign difference) the principal function fm .
Corollary 2. When λ2 = 0, the correction function fe has the
form
fe = −sign (E [gl (Y )gm (Y )|S = 0]) fm
(13)
and
∆λ (fe)
v
#
"
u
u
P
(Y
)
Y
|S=0
S=0 .
= −λ4 ρm (PX|S=0 ; WY |X )tVar log
PY |S=1 (Y )
We use the full dataset to train two logistic regression
models: (i) WY |X , which uses the features to predict the
outcome; (ii) PS|X , which uses the features to predict group
membership. Although WY |X does not use S as an input,
it has significant disparate impact, assigning higher scores
on average to African Americans compared to Caucasians
(E[Y |S = 0] = 0.543 vs. E[Y |S = 1] = 0.438). Using
WY |X and PS|X , we apply Theorems 1 and 2 to compute
the correction function fe for λ2 = λ4 = 1.
Results
In Fig. 2, we show how the correction function can identify
features that contribute to the disparate impact of a predictive
We conclude this section with Example 1, where we commodel. Here, we plot the conditional distribution of fe for
pute fe when PS|X can be expressed by a logistic regression.
distinct values of PriorCounts and LengthOfStay. As shown, the
Example 1. When PS|X (1|x) = (1 + exp(θ0 + hθ, xi)−1 , the distribution of fe is similar across all values of LengthOfStay,
log-likelihood ratio function fl is the linear function
which suggests that LengthOfStay does not affect the disparate
impact of the model. In contrast, the distribution of fe differs
d
h
i
X
fl (x) = hθ, xi −
E θi Xi S = 0 .
(14) based on the value of PriorCounts (see e.g., the differences
between PriorCounts > 3). This suggests that the model may be
i=1
using PriorCounts to discriminate between African Americans
Combining (10), (12) and (14), fe can be expressed in terms and Caucasians.
of a simple linear combination of WY |X , a linear function
In Table 1, we show prototypical examples for the correction
of x, and a constant term. In addition, when PS|Y (1|y) =
function,
which correspond to feature vectors for which the
(1 + exp(γ0 + γ1 y))−1 , ∆λ (f ) can be simplified for any
correction
function fe attains its maximum value/minimum
perturbation f (including fe) as
value/value closest to 0. Recalling that the fe represents a local
d
perturbation that minimizes the objective in (4), we see that
X
∆λ (f ) = λ2
θi E [f (X)Xi |S = 0] + λ4 γ1 E [g(Y )Y |S = 0] . the disparity in the output distributions of WY |X is maximal
i=1
for African American males that are under 25 years old, with
> 3 priors, and charged with a felony (middle column). This
See Appendix A-D for details
IV. N UMERICAL E XPERIMENTS
We discuss next a numerical experiment where we compute
correction functions for a recidivism prediction model. We
consider the ProPublica COMPAS dataset [6], which contains
information on the criminal history and demographic makeup
of prisoners in Brower County, Florida from 2013–2014. Our
goal is to illustrate the technical feasibility of our approach
on a real-world dataset, and to show that the correction
function can be computed using standard predictive models
for S and Y given X (i.e., without the need to compute
the distribution PS,X,Y ). We make no comment on associated
societal interpretations. Code to reproduce our analysis can be
found in [23].
Setup
We restrict our analysis to individuals who are African
American (S = 0) or Caucasian (S = 1). We mildly
process the raw dataset by dropping records with missing information, and converting categorical variables to numerical values. Our final dataset contains 5278 records
(3175 African American + 2103 Caucasian), where the
record for individual i consists of a feature vector xi =
(Age, ChargeDegree, Sex, PriorCounts, LengthOfStay), and an outcome variable, set as yi = 1 iff he/she is arrested for a crime
within 2 years of release from prison.
Fig. 2: Empirical distributions of the correction function fe conditioned on
distinct values of PriorCounts (top) and LengthOfStay (bottom). We use a
kernel density estimator to smooth the histograms.
F EATURES
Age
ChargeDegree
Sex
PriorCounts
LengthOfStay
argmin |fe(x)|
argmax fe(x)
argmin fe(x)
> 45
< 25
Misdemeanor
Felony
Felony
Female
Male
Male
25 to 45
0
>3
1 to 3
< 3 Months
< Week
< Week
Table 1: Prototypical examples for the correction function fe. We show the
features of African American entries that attain the maximum, minimum,
and smallest values of fe.
[10]
[11]
[12]
[13]
is to be expected, as the training dataset for WY |X shows a
large correlation between PriorCounts and the outcome, and
higher PriorCounts for African Americans on average.
[14]
V. D ISCUSSION
Disparate impact in machine learning is a critical issue with
important societal implications. In this paper, we proposed an
information-theoretic framework to study disparate impact. We
derived a correction function to link components of X to the
disparity observed in the output distributions, and illustrated
its value on a recidivism prediction application derived from
a real-world dataset. Interesting directions for future work
include extending our analysis to a broader class of predictive
models, and using correction functions to design machine
learning algorithms that mitigate disparate impact. We are
confident that information-theoretic tools can inspire exciting
new solutions to the problem.
[15]
[16]
[17]
[18]
[19]
[20]
ACKNOWLEDGMENTS
F.P. Calmon would like to thank the Harvard Dean’s Competitive Fund for Promising Scholarship for supporting this
research.
R EFERENCES
[1] A. Romei and S. Ruggieri, “A multidisciplinary survey on
discrimination analysis,” The Knowledge Engineering Review,
vol. 29, no. 5, pp. 582–638, 2014.
[2] S. Hajian, F. Bonchi, and C. Castillo, “Algorithmic bias: from
discrimination discovery to fairness-aware data mining,” in Proceedings of the 22nd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining. ACM, 2016, pp.
2125–2126.
[3] S. Barocas and A. Selbst, “Big data’s disparate impact,” 2016.
[4] D. B. Hunt, “Redlining,” Encyclopedia of Chicago, 2005.
[5] L. Sweeney, “Discrimination in online ad delivery,” Queue,
vol. 11, no. 3, p. 10, 2013.
[6] J. Angwin, J. Larson, S. Mattu, and L. Kirchner, “Machine bias,”
ProPublica, 2016.
[7] D. Ensign, S. A. Friedler, S. Neville, C. Scheidegger, and
S. Venkatasubramanian, “Decision making with limited feedback: Error bounds for recidivism prediction and predictive
policing,” 2017.
[8] M. Feldman, S. A. Friedler, J. Moeller, C. Scheidegger, and
S. Venkatasubramanian, “Certifying and removing disparate
impact,” in Proceedings of the 21th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM,
2015, pp. 259–268.
[9] P. Adler, C. Falk, S. A. Friedler, G. Rybeck, C. Scheidegger,
B. Smith, and S. Venkatasubramanian, “Auditing black-box
[21]
[22]
[23]
models for indirect influence,” in Proceedings of ICDM 2016.
IEEE, 2016, pp. 1–10.
J. Adebayo and L. Kagal, “Iterative orthogonal feature projection for diagnosing bias in black-box models,” arXiv preprint
arXiv:1611.04967, 2016.
C. Simoiu, S. Corbett-Davies, and S. Goel, “The problem of
infra-marginality in outcome tests for discrimination,” Working
paper available at https://arxiv. org/pdf/1607.05376. pdf, Tech.
Rep., 2017.
F. Kamiran and T. Calders, “Classifying without discriminating,” in 2009 2nd International Conference on Computer,
Control and Communication. IEEE, 2009, pp. 1–6.
R. Zemel, Y. Wu, K. Swersky, T. Pitassi, and C. Dwork,
“Learning fair representations,” in Proceedings of the 30th International Conference on Machine Learning (ICML-13), 2013,
pp. 325–333.
F. P. Calmon, D. Wei, B. Vinzamuri, K. N. Ramamurthy, and
K. R. Varshney, “Optimized pre-processing for discrimination
prevention,” in Advances in Neural Information Processing
Systems, 2017, pp. 3995–4004.
N. Kilbertus, M. Rojas-Carulla, G. Parascandolo, M. Hardt,
D. Janzing, and B. Schölkopf, “Avoiding discrimination through
causal reasoning,” arXiv preprint arXiv:1706.02744, 2017.
J. Liao, L. Sankar, V. Y. Tan, and F. P. Calmon, “Hypothesis
testing in the high privacy limit,” in Proc. of 54th IEEE
Annual Allerton Conference on Communication, Control, and
Computing. IEEE, 2016, pp. 649–656.
T. M. Cover and J. A. Thomas, “Elements of information theory
2nd edition,” 2006.
R. Blahut, “Hypothesis testing and information theory,” IEEE
Trans. on Info. Theory, vol. 20, no. 4, pp. 405–417, 1974.
E. Tuncel, “On error exponents in hypothesis testing,” IEEE
Trans. on Info. Theory, vol. 51, no. 8, pp. 2945–2950, 2005.
A. A. Gohari and V. Anantharam, “Evaluation of marton’s inner
bound for the general broadcast channel,” IEEE Trans. on Info.
Theory, vol. 58, no. 2, pp. 608–619, 2012.
V. Anantharam, A. Gohari, S. Kamath, and C. Nair, “On
maximal correlation, hypercontractivity, and the data processing inequality studied by Erkip and Cover,” arXiv preprint
arXiv:1304.6133, 2013.
A. Rényi, “On measures of dependence,” Acta mathematica
hungarica, vol. 10, no. 3-4, pp. 441–451, 1959.
“Disparate Impact Repository,” GitHub, 2018. [Online].
Available: http://www.github.com/ustunb/disparate-impact
Also,
A PPENDIX A
P ROOFS
DKL (PeY |S=0 kPY |S=0 ) = O(2 ).
A. Proof of Lemma 1
Proof. First, note that we can compute the distribution PeY |S=0
by passing PeX|S=0 through the given channel WY |X and get
the following expression.
X
PeY |S=0 (y) =
WY |X (y|x)PX|S=0 (x)(1 + f (x))
x
=PY |S=0 (y) +
X
=PY |S=0 (y) +
X
x
x
WY |X (y|x)PX|S=0 (x)f (x)
(18)
Combining (15), (16), (17), (18) together and letting → 0,
we obtain
#
"
PX|S=0 (X)
S=0
∆λ (f ) =λ2 E f (X) log
PX|S=1 (X)
"
#
PY |S=0 (Y )
+ λ4 E g(Y ) log
S=0 .
PY |S=1 (Y )
PX|Y,S=0 (x|y)PY |S=0 (y)f (x)
=PY |S=0 (y) (1 + E [f (X)|Y = y, S = 0])
B. Proof of Theorem 1
=PY |S=0 (y)(1 + g(y)).
By the definition of PeX|S=0 , we can compute the KLdivergence between PeX|S=0 and PX|S=1 in the following way.
DKL (PeX|S=0 kPX|S=1 )
X
PX|S=0 (x)(1 + f (x))
=
PX|S=0 (x)(1 + f (x)) log
PX|S=1 (x)
x
Proof. Note that when Y is a binary random variable,
for any function g(y) with E [g(Y )] = 0, we have that
g(Y ) = E [g(Y )gm (Y )] gm (Y ). Furthermore, for any function
f (x) with E [f (X)] = 0, if E [f (X)fm (X)] = 0, then
E [f (X)|Y ] = 0.
We define
fL (x) ,
= DKL (PX|S=0 kPX|S=1 )
X
PX|S=0 (x)
+
PX|S=0 (x)f (x) log
PX|S=1 (x)
Xx
PX|S=0 (x)(1 + f (x)) log(1 + f (x))
+
fl (x) − a1 fm (x)
,
a2
(19)
when a2 6= 0. When a2 = 0, we can choose an
arbitrary function fL such that fL ∈ L(PX|S=0 ) and
E [fL (X)fm (X)|S = 0] = 0. Note that fL ∈ L(PX|S=0 )
from the definition. Furthermore, we have gl (y) = b1 gm (y)
with b1 = E [gl (Y )gm (Y )|S = 0]. Similarly,
x
= DKL (PX|S=0 kPX|S=1 )
"
#
PX|S=0 (X)
+ E f (X) log
S=0
PX|S=1 (X)
X
+
PX|S=0 (x)(1 + f (x))f (x) + O(2 )
f (x) = m1 fm (x) + m2 fL (x) + m3 fr (x)
(20)
where,
x
= DKL (PX|S=0 kPX|S=1 )
"
#
PX|S=0 (X)
+ E f (X) log
S=0
PX|S=1 (X)
+ E [f (X)|S = 0] + 2 E f (X)2 |S = 0 + O(2 )
m1 = E [f (X)fm (X)|S = 0] ,
m2 = E [f (X)fL (X)|S = 0] ,
r h
i
2
m3 = E (f (X) − m1 fm (X) − m2 fL (X)) S = 0 .
= DKL (PX|S=0 kPX|S=1 )
"
#
PX|S=0 (X)
+ E f (X) log
S = 0 + O(2 ).
PX|S=1 (X)
fr (x) ,
(15)
Following similar computation, we have
DKL (PeX|S=0 kPX|S=0 ) = O(2 ).
(16)
Since E [g(Y )|S = 0] = 0, we have
DKL (PeY |S=0 kPY |S=1 ) − DKL (PY |S=0 kPY |S=1 )
"
#
PY |S=0 (Y )
= E g(Y ) log
S = 0 + O(2 ).
PY |S=1 (Y )
(17)
f (x) − m1 fm (x) − m2 fL (x)
,
m3
when m3 6= 0. When m3 = 0, we can choose an
arbitrary function fr such that fr ∈ L(PX|S=0 ) and
E [fr (X)fm (X)|S = 0] = E [fr (X)fL (X)|S = 0] = 0. Note
that fr ∈ L(PX|S=0 ) following the definition. Since, by the
definition of fL and fr , E [fm (X)fL (X)|S = 0] = 0 and
E [fm (X)fr (X)|S = 0] = 0, then
g(y) = E [f (X)|Y = y, S = 0]
= ρm (PX|S=0 ; WY |X )m1 gm (y).
Therefore, following previous discussions and using Lemma
1, we have
∆λ (f )
#
PX|S=0 (X)
S=0
= λ2 E f (X) log
PX|S=1 (X)
"
#
PY |S=0 (Y )
+ λ4 E g(Y ) log
S=0
PY |S=1 (Y )
"
#
"
"
#
= λ2 E f (X)fl (X) S = 0 + λ4 E g(Y )gl (Y ) S = 0
= λ2 a1 m1 + λ2 a2 m2 + λ4 ρm (PX|S=0 ; WY |X )b1 m1 .
Since f ∈ L(PX|S=0 ), we have that m21 + m22 + m23 = 1. Accordingly, we can minimize ∆λ (f ) by solving the optimization
problem:
min
λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1 m1 + λ2 a2 m2
m1 ,m2
s.t.
m21
+
m22
+
m23
= 1.
Byqthe Cauchy-Schwarz inequality, the minimal value is
2
− λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1 + (λ2 a2 )2 which is
achieved by setting
− λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1
m1 = q
,
2
λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1 + (λ2 a2 )2
m2 = q
#
"
PX|S=0 (x)
PX|S=0 (X)
S=0
fl (x) = log
− E log
PX|S=1 (x)
PX|S=1 (X)
#
"
PS|X (0|x)
PS|X (0|X)
= log
S=0
− E log
PS|X (1|x)
PS|X (1|X)
#
"
1 − PS|X (1|x)
1 − PS|X (1|X)
= log
S=0 .
− E log
PS|X (1|x)
PS|X (1|X)
D. Details for Example 1
When PS|X (1|x) = (1 + exp(θ0 + hθ, xi)−1 , then
"
#
1 − PS|X (1|x)
1 − PS|X (1|X)
fl (x) = log
− E log
S=0
PS|X (1|x)
PS|X (1|X)
h
i
= θ0 + hθ, xi − E θ0 + hθ, Xi S = 0
= hθ, xi −
λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1
2
+ (λ2 a2 )2
Therefore, the function f , which achieves this minimal value,
is nm fm + nl fl where
−λ4 ρm (PX|S=0 ; WY |X )b1
nm = q
2
λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1 + (λ2 a2 )2
and
−λ2
λ2 a1 + λ4 ρm (PX|S=0 ; WY |X )b1
2
.
+ (λ2 a2 )2
d
h
i
X
E θi Xi S = 0 .
i=1
−λ2 a2
m3 = 0.
nl = q
Note that
PX|S=0 (x)
PS|X (0|x)PS (1)
log
= log
PX|S=1 (x)
PS|X (1|x)PS (0)
PS|X (0|x)
PS (1)
= log
+ log
.
PS|X (1|x)
PS (0)
Next,
"
#
PX|S=0 (X)
E f (X) log
S=0
PX|S=1 (X)
"
#
PS|X (0|X)
= E f (X) log
S=0
PS|X (1|X)
"
#
PS (1)
+ E f (X) log
S=0
PS (0)
= E [f (X) (θ0 + hθ, Xi) |S = 0]
= θ0 E [f (X)|S = 0] +
d
X
θi E [f (X)Xi |S = 0]
i=1
C. Proof of Theorem 2
Proof. Suppose that gm (0) = a ≥ 0 and gm (1) = b. Then
E [gm (Y )|S = 0] = a(1
implies that
− p) 2+ bp = 0 which
2
b = −(1−p)
a.
Since
E
g
(Y
)
|S
=
0
=
a
(1
−
p) + b2 p =
m
p
q
2
p
a2 (1 − p) + a2 (1−p)
= 1, then a = 1−p
.
p
Next,
fm (x)
E [gm (Y )|X = x, S = 0]
=
ρm (PX|S=0 ; WY |X )
gm (1)WY |X (1|x) + gm (0)WY |X (0|x)
=
ρm (PX|S=0 ; WY |X )
gm (1) − gm (0)
gm (0)
=
WY |X (1|x) +
.
ρm (PX|S=0 ; WY |X )
ρm (PX|S=0 ; WY |X )
=
d
X
θi E [f (X)Xi |S = 0] .
i=1
Similarly, if we also assume that PS|Y (1|y) = (1 + exp(γ0 +
γ1 y))−1 , then
"
#
PY |S=0 (Y )
E g(Y ) log
S = 0 = γ1 E [g(Y )Y |S = 0] .
PY |S=1 (Y )
Thus, we can express,
∆λ (f )
= λ2
d
X
i=1
θi E [f (X)Xi |S = 0] + λ4 γ1 E [g(Y )Y |S = 0] .
| 7cs.IT
|
Energy Clustering
Guilherme França∗ and Joshua T. Vogelstein†
Johns Hopkins University
Abstract
arXiv:1710.09859v1 [stat.ML] 26 Oct 2017
Energy statistics was proposed by Székely in the 80’s inspired by the Newtonian gravitational
potential from classical mechanics, and it provides a hypothesis test for equality of distributions. It
was further generalized from Euclidean spaces to metric spaces of strong negative type, and more
recently, a connection with reproducing kernel Hilbert spaces (RKHS) was established. Here we
consider the clustering problem from an energy statistics theory perspective, providing a precise
mathematical formulation yielding a quadratically constrained quadratic program (QCQP) in the
associated RKHS, thus establishing the connection with kernel methods. We show that this QCQP
is equivalent to kernel k-means optimization problem once the kernel is fixed. These results imply
a first principles derivation of kernel k-means from energy statistics. However, energy statistics
fixes a family of standard kernels. Furthermore, we also consider a weighted version of energy
statistics, making connection to graph partitioning problems. To find local optimizers of such
QCQP we propose an iterative algorithm based on Hartigan’s method, which in this case has
the same computational cost as kernel k-means algorithm, based on Lloyd’s heuristic, but usually
with better clustering quality. We provide carefully designed numerical experiments showing the
superiority of the proposed method compared to kernel k-means, spectral clustering, standard
k-means, and Gaussian mixture models in a variety of settings.
∗
[email protected]
†
[email protected]
1
I.
INTRODUCTION
Energy statistics [1] is based on a notion of statistical potential energy between probability
distributions, in close analogy to Newton’s gravitational potential in classical mechanics. It
provides a model-free hypothesis test for equality of distributions which is achieved under
minimum energy. When probability distributions are different the statistical potential energy
diverges as sample size increases, while tends to a nondegenerate limit distribution when
probability distributions are equal. Energy statistics has been applied to several goodnessof-fit hypothesis tests, multi-sample tests of equality of distributions, analysis of variance
[2], nonlinear dependence tests through distance covariance and distance correlation, which
generalizes the Pearson correlation coefficient, and hierarchical clustering [3] by extending
Ward’s method of minimum variance. Moreover, in Euclidean spaces, an application of
energy statistics to clustering was already proposed [4]. We refer the reader to [1], and
references therein, for an overview of energy statistics theory and its applications.
In its original formulation, energy statistics has a compact representation in terms of
expectations of pairwise Euclidean distances, providing straightforward empirical estimates.
The notion of distance covariance was further generalized from Euclidean spaces to metric
spaces of strong negative type [5]. Furthermore, the missing link between energy distance
based tests and kernel based tests has been recently resolved [6], establishing an equivalence
between generalized energy distances to maximum mean discrepancies (MMD), which are
distances between embeddings of distributions in reproducing kernel Hilbert spaces (RKHS).
This equivalence immediately relates energy statistics to kernel methods often used in machine learning, and form the basis of our approach.
Clustering has such a long history in machine learning, making it impossible to mention
all important contributions in a short space. Perhaps, the most used method is k-means [7–
9], which is based on Lloyd’s heuristic [7] of assigning a data point to the cluster with closest
center. The only statistical information about each cluster comes from its mean, making it
sensitive to outliers. Nevertheless, k-means works very well when data is linearly separable
in Euclidean space. Gaussian mixture models (GMM) is another very common approach,
providing more flexibility than k-means, however, it still makes strong assumptions about
the distribution of the data.
To account for nonlinearities, kernel methods were introduced [10, 11]. A mercer kernel
2
[12] is used to implicitly map data points to a RKHS, then clustering can be performed in
the associated Hilbert space by using its inner product. However, the kernel choice remains
the biggest challenge since there is no principled theory to construct a kernel for a given
dataset, and usually a kernel introduces hyperparameters that need to be carefully chosen.
A well-known kernel based clustering method is kernel k-means, which is precisely k-means
formulated in the feature space [11]. Furthermore, kernel k-means algorithm [13, 14] is still
based on Loyd’s heuristic [7] of grouping points that are closer to a cluster center. We refer
the reader to [15] for a survey of clustering methods.
Although clustering from energy statistics, in Euclidean spaces, was considered in [4], the
precise optimization problem behind this approach remains elusive, as well as the connection with kernel methods. The main theoretical contribution of this paper is to fill this gap.
Since the statistical potential energy is minimum when distributions are equal, the principle
behind clustering is to maximize the statistical energy, enforcing probability distributions
associated to each cluster to be different from one another. We provide a precise mathematical formulation to this statement, leading to a quadratically constrained quadratic
program (QCQP) in the associated RKHS. This immediately establishes the connection between energy statistics based clustering, or energy clustering for short, with kernel methods.
Moreover, our formulation holds for general semimetric spaces of negative type. We also
show that such QCQP is equivalent to kernel k-means optimization problem, however, the
kernel is fixed by energy statistics. The equivalence between kernel k-means, spectral clustering, and graph partitioning problems is well-known [13, 14]. We further demonstrate how
these relations arise from a weighted version of energy statistics.
Our main algorithmic contribution is to use Hartigan’s method [16] to find local solutions
of the above mentioned QCQP, which is NP-hard in general. Hartigan’s method was also
used in [4], but without any connection to kernels. More importantly, the advantages of
Hartigan’s over Lloyd’s method was already demonstrated in some simple settings [17, 18],
but apparently this method did not receive the deserved attention. To the best of our
knowledge, Hartigan’s method was not previously employed together with kernel methods.
We provide a fully kernel based Hartigan’s algorithm for clustering, where the kernel is fixed
by energy statistics. We make clear the advantages of this proposal versus Lloyd’s method,
which kernel k-means is based upon and will also be used to solve our QCQP. We show that
both algorithms have the same time complexity, but Hartigan’s method in kernel spaces
3
offer several advantages. Furthermore, in the examples considered in this paper, it also
provides superior performance compared spectral clustering, which is more expensive and
in fact solves a relaxed version of our QCQP.
Our numerical results provide compelling evidence that Hartigan’s method applied to
energy clustering is more accurate and robust than kernel k-means algorithm. Furthermore,
our experiments illustrate the flexibility of energy clustering, showing that it is able to
perform accurately on data coming from very different distributions, contrary to k-means
and GMM for instance. More specifically, the proposed method performs closely to k-means
and GMM on normally distributed data, however, it is significantly better on data that is
not normally distributed. Its superiority in high dimensions is striking, being more accurate
than k-means and GMM even on Gaussian settings.
II.
BACKGROUND ON ENERGY STATISTICS AND RKHS
In this section we introduce the main concepts from energy statistics and its relation to
RKHS which form the basis of our work. For more details we refer the reader to [1] and [6].
iid
iid
Consider random variables in RD such that X, X 0 ∼ P and Y, Y 0 ∼ Q, where P and Q
are cumulative distribution functions with finite first moments. The quantity
E(P, Q) ≡ 2EkX − Y k − EkX − X 0 k − EkY − Y 0 k,
(1)
called energy distance [1], is rotationally invariant and nonnegative, E(P, Q) ≥ 0, where
equality to zero holds if and only if P = Q. Above, k · k denotes the Euclidean norm in RD .
Energy distance provides a characterization of equality of distributions, and E 1/2 is a metric
on the space of distributions.
The energy distance can be generalized as, for instance,
Eα (P, Q) ≡ 2EkX − Y kα − EkX − X 0 kα − EkY − Y 0 kα
(2)
where 0 < α ≤ 2. This quantity is also nonnegative, Eα (P, Q) ≥ 0. Furthermore, for
0 < α < 2 we have that Eα (P, Q) = 0 if and only if P = Q, while for α = 2 we have
E2 (P, Q) = 2kE(X) − E(Y )k2 which shows that equality to zero only requires equality of the
means, and thus E2 (P, Q) = 0 does not imply equality of distributions.
The energy distance can be even further generalized. Let X, Y ∈ X where X is an
arbitrary space endowed with a semimetric of negative type ρ : X × X → R, which is
4
required to satisfy
n
X
i,j=1
where Xi ∈ X and ci ∈ R such that
ci cj ρ(Xi , Xj ) ≤ 0,
Pn
i=1 ci
= 0. Then, X is called a space of negative type.
We can thus replace RD → X and kX − Y k → ρ(X, Y ) in the definition (1), obtaining the
generalized energy distance
E(P, Q) ≡ 2Eρ(X, Y ) − Eρ(X, X 0 ) − Eρ(Y, Y 0 ).
(3)
For spaces of negative type there exists a Hilbert space H and a map ϕ : X → H such that
ρ(X, Y ) = kϕ(X) − ϕ(Y )k2H . This allows us to compute quantities related to probability
distributions over X in the associated Hilbert space H. Even though the semimetric ρ may
not satisfy the triangle inequality, ρ1/2 does since it can be shown to be a proper metric. Our
energy clustering formulation, proposed in the next section, will be based on the generalized
energy distance (3).
There is an equivalence between energy distance, commonly used in statistics, and distances between embeddings of distributions in RKHS, commonly used in machine learning.
This equivalence was established in [6]. Let us first recall the definition of RKHS. Let H be
a Hilbert space of real-valued functions over X . A function K : X × X → R is a reproducing
kernel of H if it satisfies the following two conditions:
1. hx ≡ K(·, x) ∈ H for all x ∈ X .
2. hhx , f iH = f (x) for all x ∈ X and f ∈ H.
In other words, for any x ∈ X and any function f ∈ H, there is a unique hx ∈ H that
reproduces f (x) through the inner product of H. If such a kernel function K exists, then
H is called a RKHS. The above two properties immediately imply that K is symmetric
and positive definite. Indeed, notice that hhx , hy i = hy (x) = K(x, y), and by definition
hhx , hy i∗ = hhy , hx i, but since the inner product is real we have hhy , hx i = hhx , hy i, or
P
equivalently K(y, x) = K(x, y). Moreover, for any w ∈ H we can write w = ni=1 ci hxi
P
where {hxi }ni=1 is a basis of H. It follows that hw, wiH = ni,j=1 ci cj K(xi , xj ) ≥ 0, showing
that the kernel is positive definite. If G is a matrix with elements Gij = K(xi , xj ) this is
equivalent to G being positive semidefinite, i.e. v > G v ≥ 0 for any vector v ∈ Rn .
5
The Moore-Aronszajn theorem [19] establishes the converse of the above paragraph. For
every symmetric and positive definite function K : X ×X → R, there is an associated RKHS
HK with reproducing kernel K. The map ϕ : x 7→ hx ∈ HK is called the canonical feature
map. Given a kernel K, this theorem enables us to define an embedding of a probability
R
measure P into the RKHS as follows: P 7→ hP ∈ HK such that f (x)dP (x) = hf, hP i for
R
all f ∈ HK , or alternatively, hP ≡ K( · , x)dP (x). We can now introduce the notion of
distance between two probability measures using the inner product of HK , which is called
the maximum mean discrepancy (MMD) and is given by
γK (P, Q) ≡ khP − hQ kHK .
(4)
2
γK
(P, Q) = EK(X, X 0 ) + EK(Y, Y 0 ) − 2EK(X, Y )
(5)
This can also be written as [20]
iid
iid
where X, X 0 ∼ P and Y, Y 0 ∼ Q. From the equality between (4) and (5) we also have
hhP , hQ iHK = E K(X, Y ).
Thus, in practice, we can estimate the inner product between embedded distributions by
averaging the kernel function over sampled data.
The following important result shows that semimetrics of negative type and symmetric
positive definite kernels are closely related [21]. Let ρ : X × X → R and x0 ∈ X an arbitrary
but fixed point. Define
K(x, y) ≡ 21 [ρ(x, x0 ) + ρ(y, x0 ) − ρ(x, y)] .
(6)
Then, it can be shown that K is positive definite if and only if ρ is a semimetric of negative
type. We have a family of kernels, one for each choice of x0 . Conversely, if ρ is a semimetric
of negative type and K is a kernel in this family, then
ρ(x, y) = K(x, x) + K(y, y) − 2K(x, y)
= khx − hy k2HK
(7)
and the canonical feature map ϕ : x 7→ hx is injective [6]. When these conditions are satisfied
we say that the kernel K generates the semimetric ρ. If two different kernels generate the
same ρ they are said to be equivalent kernels.
6
Now we can state the equivalence between the generalized energy distance (3) and inner
products on RKHS, which is one of the main results of [6]. If ρ is a semimetric of negative
type and K a kernel that generates ρ, then replacing (7) into (3), and using (5), yields
2
(P, Q).
E(P, Q) = 2 [E K(X, X 0 ) + E K(Y, Y 0 ) − 2E K(X, Y )] = 2γK
Due to (4) we can compute the energy distance E(P, Q) between two probability distributions
using the inner product of HK .
Finally, let us recall the main formulas from generalized energy statistics for the test
statistic of equality of distributions [1]. Assume we have data X = {x1 , . . . , xn } where
S
xi ∈ X , and X is a space of negative type. Consider a disjoint partition X = kj=1 Cj ,
with Ci ∩ Cj = ∅. Each expectation in the generalized energy distance (3) can be computed
through the function
g(Ci , Cj ) ≡
1 XX
ρ(x, y),
ni nj x∈C y∈C
i
(8)
j
where ni = |Ci | is the number of elements in partition Ci . The within energy dispersion is
defined by
W ≡
k
X
nj
j=1
2
g(Cj , Cj ),
(9)
and the between-sample energy statistic is defined by
S≡
where n =
Pk
j=1
X
ni nj
[2g(Ci , Cj ) − g(Ci , Ci ) − g(Cj , Cj )] ,
2n
1≤i<j≤k
(10)
nj . Given a set of distributions {Pj }kj=1 , where x ∈ Cj if and only if x ∼ Pj ,
the quantity S provides a test statistic for equality of distributions [1]. When the sample size
is large enough, n → ∞, under the null hypothesis H0 : P1 = P2 = · · · = Pk we have that
S → 0, and under the alternative hypothesis H1 : Pi 6= Pj for at least two i 6= j, we have
that S → ∞. Note that this test does not make any assumptions about the distributions
Pj , thus it is said to be non-parametric or distribution-free.
One can make a physical analogy by thinking that points x ∈ Cj form a massive body
whose total mass is characterized by the distribution function Pj . The quantity S is thus a
potential energy of the from S(P1 , . . . , Pk ) which measures how different the distribution of
these masses are, and achieves the ground state S = 0 when all bodies have the same mass
distribution. The potential energy S increases as bodies have different mass distributions.
7
III.
CLUSTERING BASED ON ENERGY STATISTICS
This section contains the main theoretical results of this paper, where we formulate an
optimization problem for clustering based on energy statistics and RKHS introduced in the
previous section.
Due to the previous test statistic for equality of distributions, the obvious criterion for
clustering data is to maximize S which makes each cluster as different as possible from
the other ones. In other words, given a set of points coming from different probability
distributions, the test statistic S should attain a maximum when each point is correctly
classified as belonging to the cluster associated to its probability distribution. The following
straightforward result shows that maximizing S is, however, equivalent to minimizing W
which has a more convenient form.
Lemma 1. Let X = {x1 , . . . , xn } where each data point xi lives in a space X endowed with a
S
semimetric ρ : X × X → R of negative type. For a fixed integer k, the partition X = kj=1 Cj ,
where Ci ∩ Cj = ∅ for all i 6= j, maximizes the between-sample statistic S, defined in equation
(10), if and only if
min W (C1 , . . . , Ck ),
C1 ,...,Ck
(11)
where the within energy dispersion W is defined by (9).
Proof. From (9) and (10) we have
k
k
k
X
1 X
1 X
nj ni g(Ci , Ci )
ni nj g(Ci , Cj ) +
n−
S+W =
2n i,j=1
2n i=1
j6=i=1
i6=j
=
k
1 X
1 XX
n
ni nj g(Ci , Cj ) =
ρ(x, y) = g(X, X).
2n i,j=1
2n x∈X y∈X
2
Note that the right hand side of this equation only depends on the pooled data, so it is a
constant independent of the choice of partition. Therefore, maximizing S over the choice of
partition is equivalent to minimizing W .
For a given k, the clustering problem amounts to finding the best partition of the data by
minimizing W . Notice that this is a hard clustering problem as partitions are disjoint. The
optimization problem (11) based on energy statistics was already proposed in [4]. However,
it is important to note that this is equivalent to maximizing S, which is the test statistic for
8
equality of distributions. In this current form, the relation with kernels and other clustering
methods is obscure. In the following, we show what is the explicit optimization problem
behind (11) in the corresponding RKHS, establishing the connection with kernel methods.
Based on the relation between kernels and semimetrics of negative type, assume that the
kernel K : X × X → R generates ρ. Define the Gram matrix
K(x1 , x1 ) K(x1 , x2 ) · · · K(x1 , xn )
K(x2 , x1 ) K(x2 , x2 ) · · · K(x2 , xn )
.
G≡
..
..
..
..
.
.
.
.
K(xn , x1 ) K(xn , x2 ) · · · K(xn , xn )
(12)
Let Z ∈ {0, 1}n×k be the label matrix, with only one nonvanishing entry per row, indicating
to which cluster (column) each point (row) belongs to. This matrix satisfies Z > Z = D,
where the diagonal matrix D = diag(n1 , . . . , nk ) contains the number of points in each
cluster. We also introduce the rescaled matrix Y ≡ ZD−1/2 . In component form they are
given by
Zij ≡
1 if xi ∈ Cj
Yij ≡
0 otherwise
√1
nj
0
if xi ∈ Cj
.
(13)
otherwise
Throughout the paper, we use the notation Mi• to denote the ith row of a matrix M , and
M•j denotes its jth column. Our next result shows that the optimization problem (11) is
NP-hard since it is a quadratically constrained quadratic program (QCQP) in the RKHS.
Proposition 2. The optimization problem (11) is equivalent to
max Tr Y > G Y
Y
s.t. Y ≥ 0, Y > Y = I, Y Y > e = e,
(14)
where e = (1, 1, . . . , 1)> ∈ Rn is the all-ones vector, and G is the Gram matrix (12).
Proof. From (7), (8), and (9) we have
k
k X
X
1X 1 X
1 X
W (C1 , . . . , Ck ) =
K(x, x) −
ρ(x, y) =
K(x, y) .
2 j=1 nj x,y∈C
nj y∈C
j=1 x∈C
j
j
(15)
j
Note that the first term is global so it does not contribute to the optimization problem.
Therefore, minimizing (15) is equivalent to
k
X
1 X
max
K(x, y).
C1 ,...,Ck
n
j=1 j x,y∈C
j
9
(16)
But
X
K(x, y) =
n X
n
X
Zpj Zqj Gpq = (Z > G Z)jj ,
p=1 q=1
x,y∈Cj
−1
where we used the definitions (12) and (13). Notice that n−1
= Djj
, where the diagonal
j
matrix D = diag(n1 , . . . , nk ) contains the number of points in each cluster, thus the objective
P
−1
function in (16) is equal to kj=1 Djj
Z > GZ jj = Tr D−1 Z > GZ . Now we can use the
cyclic property of the trace, and by the definition of the matrix Z in (13), we obtain the
following integer programing problem:
max Tr
Z
ZD−1/2
>
G ZD−1/2
s.t. Zij ∈ {0, 1},
Pk
j=1
Zij = 1,
Pn
i=1
Zij = nj .
(17)
Now we write this in terms of the matrix Y = ZD−1/2 . The objective function immedi
ately becomes Tr Y > G Y . Notice that the above constraints imply that Z T Z = D, which
in turn gives D−1/2 Y T Y D−1/2 = D, or Y > Y = I. Also, every entry of Y is positive by
definition, Y ≥ 0. Now it only remains to show the last constraint in (14), which comes
from the last constraint in (17). In matrix form this reads Z T e = De. Replacing Z = Y D1/2
we have Y > e = D1/2 e. Multiplying this last equation on the left by Y , and noticing that
Y D1/2 e = Ze = e, we finally obtain Y Y > e = e. Therefore, the optimization problem (17)
is equivalent to (14) .
Based on Proposition 2, to group data X = {x1 , . . . , xn } into k clusters we first compute
the Gram matrix G and then solve the optimization problem (14) for Y ∈ Rn×k . The
ith row of Y will contain a single nonzero element in some jth column, indicating that
xi ∈ Cj . This optimization problem is nonconvex, and also NP-hard, thus a direct approach
is computational prohibitive even for small datasets. However, one can find approximate
solutions by relaxing some of the constraints, or obtaining a relaxed SDP version of it. For
instance, the relaxed problem
max Tr Y > G Y
Y
s.t. Y > Y = I
has a well-known closed form solution Y ? = U R, where the columns of U ∈ Rn×k contain
the top k eigenvectors of G corresponding to the k largest eigenvalues λ1 ≥ λ2 ≥ . . . ≥ λk ,
and R ∈ Rk×k is an arbitrary orthogonal matrix. The resulting optimal objective function
P
assumes the value max Tr Y ? > G Y ? = ki=1 λi . Spectral clustering is based on the above
10
approach, where one further normalize the rows of Y ? , then cluster the resulting rows as
data points. A procedure on these lines was proposed in the seminal papers [22, 23].
Note that the optimization problem (14) based on energy statistics is valid for data living
in an arbitrary space of negative type, where a semimetric ρ, and thus the kernel K, are
assumed to be known. The standard energy distance (2) fixes a family of choices in Euclidean
spaces given by
ρα (x, y) = kx − ykα ,
Kα (x, y) = 21 (kxkα + kykα − kx − ykα ) ,
for 0 < α ≤ 2 and we fix x0 = 0 in (6). The same would be valid for data living in a more
general semimetric space (X , ρ) where ρ fixes the kernel. In practice, the clustering quality
strongly depend on the choice of a suitable ρ. Nevertheless, if prior information is available
to make this choice, it can be immediately incorporated in the optimization problem (14).
Relation to Kernel k-Means
One may wonder how energy clustering relates to the well-known kernel k-means problem1
which is extensively used in machine learning. For a positive semidefinite Gram matrix G,
as defined in (12), there exists a map ϕ : X → HK such that K(x, y) = hϕ(x), ϕ(y)i. Kernel
k-means optimization problem, in feature space, is defined by
k X
X
2
min J(C1 , . . . , Ck ) ≡
kϕ(x) − ϕ(µj )k
C1 ,...,Ck
where µj =
1
nj
P
x∈Cj
(18)
j=1 x∈Cj
x is the mean of cluster Cj in the ambient space. Notice that the above
objective function is strongly tied to the idea of minimizing distances between points and
cluster centers, which arises from k-means objective function based on Lloyd’s method [7].
It is known [13, 14] that problem (18) can be cast into a trace maximization in the same
form as (14). The next result makes this explicit, showing that (11) and (18) are actually
equivalent.
Proposition 3. For a fixed kernel, the clustering optimization problem (11) based on energy statistics is equivalent to the kernel k-means optimization problem (18), and both are
equivalent to (14).
1
When we refer to kernel k-means problem we mean specifically the optimization problem (18), which
should not be confused with kernel k-means algorithm that is just one possible recipe to solve (18). The
distinction should also be clear from the context.
11
Proof. Notice that kϕ(x)−ϕ(µj )k2 = hϕ(x), ϕ(x)i−2hϕ(x), ϕ(µj )i+hϕ(µj ), ϕ(µj )i, therefore
k X
X
1 X
2 X
K(x, y) + 2
K(y, z) .
J=
K(x, x) −
nj y∈C
nj y,z∈C
j=1 x∈C
j
j
(19)
j
The first term is global so it does not contribute to the optimization problem. Notice that
P
P
P
the third term gives x∈Cj n12 y,z∈Cj K(y, z) = n1j y,z∈Cj K(y, z), which is the same as the
j
second term. Thus, problem (18) is equivalent to
k
X
1 X
K(x, y)
C1 ,...,Ck
n
j
j=1
x,y∈C
max
j
which is exactly the same as (16) from the energy statistics formulation. Therefore, once the
kernel K is fixed, the function W given by (9) is the same as J in (18). The remaining of the
proof proceeds as already shown in the proof of Proposition 2, leading to the optimization
problem (14).
The above result shows that kernel k-means optimization problem is equivalent to the
clustering problem formulated in the energy statistics framework, when operating on the
same kernel. Notice, however, that energy statistics is valid for arbitrary semimetric spaces
of negative type, fixing the kernel function in the associated RKHS, which is guaranteed to
be positive definite. On the other hand, kernel k-means (18) by itself is just an heuristic
approach that does not make any explicit mention to the kernel. Based on Proposition 3
one may view kernel k-means as being derived from the energy statistics framework.
Kernel k-means, spectral clustering, and graph partitioning problems such as ratio association, ratio cut, and normalized cut are all equivalent to a QCQP of the form (14)
[13, 14]. One can thus use kernel k-means algorithm to solve these problems as well. This
correspondence involves a weighted version of problem (14), that will be demonstrated in
the following from the perspective of energy statistics.
IV.
CLUSTERING BASED ON WEIGHTED ENERGY STATISTICS
We now generalize energy statistics to incorporate weights associated to each data point.
Let w(x) be a weight function associated to point x ∈ X . Define
g(Ci , Cj ) ≡
1 XX
w(x)w(y)ρ(x, y),
si sj x∈C y∈C
i
j
12
si ≡
X
x∈Ci
w(x).
(20)
Replace this function in the formulas (9) and (10), with ni → si and n → s, where s =
Pk
j=1 sj . With these changes Proposition 1 remains the unaltered, so the clustering problem
becomes
k
X
sj
g(Cj , Cj )
min W (C1 , . . . , Ck ) ≡
C1 ,...,Ck
2
j=1
where now g is given by (20). Define the following matrices and vector:
√1
if xi ∈ Cj
sj
Yij ≡
,
W ≡ diag(w1 , . . . , wn ),
H ≡ W 1/2 Y,
0
otherwise
(21)
ω ≡ We, (22)
where wi = w(xi ) and e ∈ Rn is the all-ones vector.The analogous of Proposition 2 is as
follows.
Proposition 4. The weighted energy clustering given by problem (21) is equivalent to
max Tr H > (W 1/2 GW 1/2 )H
s.t. H ≥ 0, H > H = I, HH > ω = ω,
H
(23)
where G is the Gram matrix (12), ω = (w1 , . . . , wn )T contains the weights of each point, and
W = diag(ω).
Proof. Replacing (7) and eliminating the global terms which do not contribute, the optimization problem (21) becomes
k
X
1 XX
w(x)w(y)K(x, y).
C1 ,...,Ck
s
j
j=1
x∈C y∈C
max
j
j
This objective function can be written as
√
k
n
n
k X
n X
n
>√
X
X
Zjp
wp 1/2
wq Zqj
1 XX
1/2
wp wq Zpj Zqj Gpq =
wp Gpq wq
√
√
s
sj
sj
j=1 j p=1 q=1
j=1 p=1 q=1
=
k
X
j=1
H > W 1/2 GW 1/2 H
jj
= Tr H > W 1/2 GW 1/2 H .
To obtain the constraints, note that Hij ≥ 0 by definition, and
(H > H)ij =
n
n
X
1
δij X
w` Z`i Z`j =
w` Z`i = δij ,
Y`i W`` Y`j = √ √
s
s
i sj
i
`=1
`=1
`=1
n
X
13
where δij = 1 if i = j and δij = 0 if i 6= j is the Kronecker delta. Therefore, H > H = I. This
is a constraint on the rows of H. To obtain a condition on its columns observe that
√
wp wq
k
if both xp , xq ∈ Ci
X
Z
Z
√
pj qj
si
=
H > H pq = wp wq
sj
0
j=1
otherwise.
Therefore, (H > HW 1/2 )pq =
√
wp wq s−1
if both points xp and xq belong to the same cluster,
i
which we denote by Ci for some i ∈ {1, . . . , k}, and (H > HW 1/2 )pq = 0 otherwise. Thus, the
pth line of this matrix is nonzero only on entries corresponding to points that are in the same
Pn
√
√
wp ,
cluster as xp . If we sum over the columns of this line we obtain wp s−1
i
q=1 wq Zqi =
or equivalently HH > W 1/2 e = W 1/2 e, which gives the constraint HH > ω = ω.
Connection with Graph Partitioning
The relation between kernel k-means and graph partitioning problems is known [13, 14].
For conciseness, we repeat a similar analysis due to the relation of these problems to energy
statistics and RKHS, which provides a different perspective.
Consider a graph G = (V, E, A) where V is the set of vertices, E the set of edges, and A
is an affinity matrix of the graph, which measures the similarities between pairs of nodes.
Thus, Aij 6= 0 if (i, j) ∈ E, and Aij = 0 otherwise. We also associate weights to every vertex,
P
wi = w(i) for i ∈ V, and let sj = i∈Cj wi , where Cj ⊆ V is one partition of V. Let
links(C` , Cm ) ≡
X
Aij .
i∈C` ,j∈Cm
We want to partition the set of vertices V into k disjoint subsets, V =
generalized ratio association problem is given by
max
Ci ,...,Ck
Sk
j=1
k
X
links(Cj , Cj )
j=1
sj
Cj . The
(24)
and maximizes the within cluster association. The generalized ratio cut problem
min
Ci ,...,Ck
k
X
links(Cj , V\Cj )
sj
j=1
(25)
minimizes the cut between clusters. These two problems are equivalent, in analogous way
as minimizing (9) is equivalent to maximizing (10) as shown in Proposition 1. Here this is
14
due to the equality links(Cj , V\Cj ) = links(Cj , V) − links(Cj , Cj ). Several graph partitioning
methods [22, 24–26] can be seen as a particular case of (24) or (25).
Consider the ratio association problem (24), whose objective function can be written as
k X
k
n X
n
>
X
X
Zjp
Zqj
1 XX
>
Apq =
√ Apq √ = Tr Y AY ,
s
sj
sj
j=1 p=1 q=1
j=1 j p∈C q∈C
j
j
with Z defined in (13) and Y in (22). Therefore, the ratio association problem can be written
in the form (23), i.e.
max Tr H > W −1/2 AW −1/2 H
H
s.t. H ≥ 0, H > H = I, HH > ω = ω.
This is exactly the same problem as weighted energy clustering with G = W −1 AW −1 .
Assuming this matrix is positive semidefinite, this generates a semimetric (7) for graphs
given by
ρ(i, j) =
Aii Ajj
2Aij
+ 2 −
2
wi
wj
wi wj
or
ρ(i, j) = −
2Aij
wi wj
(26)
for vertices i, j ∈ V, and where in the second equation we assume the graph has no self-loops,
i.e. Aii = 0. Using (26) in the energy statistics formulation allows one to make inference on
graphs. Above, the weight wi = w(i) of node i ∈ V can be, for instance, its degree wi = d(i).
V.
TWO-CLASS PROBLEM IN ONE DIMENSION
Before stating a general algorithm to solve the optimization problem (14) we first consider
the simplest possible case which is one-dimensional data and a two-class problem. This will
be useful to test energy clustering on a simple setting.
Fixing ρ(x, y) = |x − y| according to the standard energy distance, we can actually
compute the function (8) in O(n log n) and minimize W directly. This is done by noting
that
|x − y| = (x − y)1x≥y − (x − y)1x<y
= x (1x≥y − 1x<y ) + y (1y>x − 1y≤x )
where we have the indicator function defined by 1A = 1 if A is true, and 1A = 0 otherwise.
Let C be a partition with n elements. Using the above distance we have
g (C, C) =
1 XX
x (1x≥y + 1y>x − 1x≥y − 1x<y ) .
n2 x∈C y∈C
15
Algorithm 1 E 1D -clustering algorithm to find local solutions to the optimization problem
(11) for a two-class problem in one dimension.
input data X
output label matrix Z
e = [x1 , . . . , xn ]
1: sort X obtaining X
2: for j ∈ [1, . . . , n] do
3:
4:
Ce1,j ← [xi : i = 1, . . . , j], and Ce2,j ← [xi : i = j + 1, . . . , n]
W (j) ← W Ce1,j , Ce2,j , from (27)
5: end for
6: j ? ← arg minj W (j)
7: Zj• ← (1, 0) if j ≤ j ? , and Zj• ← (0, 1) otherwise, for j = 1, . . . , n
The sum over y can be eliminated since each term in the parenthesis is simply counting the
number of elements in C that satisfy the condition of the indicator function. Assuming that
we first order the data in C, obtaining Ce = [xj ∈ C : x1 ≤ x2 ≤ · · · ≤ xn ], we get
n
2 X
e
e
(2` − 1 − n)x` .
g C, C = 2
n `=1
e Ce is O(n) and the cost of sorting the data is at the
Note that the cost of computing g C,
S
most O(n log n). Assuming that each partition is ordered, X = kj=1 Cej , the within energy
dispersion can be written explicitly as
W Ce1 , . . . , Cek =
nj
k X
X
2` − 1 − nj
j=1 `=1
nj
x` .
(27)
For a two-class problem we can use the formula (27) to cluster the data through a simple
e Then we compute (27)
algorithm as follows. We first order the entire dataset, X → X.
e and pick the point which gives the minimum value of W . This
for each possible split of X
procedure is described in Algorithm 1 and called E 1D -clustering. Note that this algorithm
is deterministic, however, it only works for one-dimensional data with Euclidean distance.
The total complexity of E 1D -clustering is O(n log n + n2 ) = O(n2 ).
Assuming the true label matrix Z is available, a direct measure of how different the
16
estimated matrix Ẑ is from Z, up to label permutations, is given by
n
k
1 XX
accuracy(Ẑ) ≡ max
Ẑiσ(j) Zij
σ n
i=1 j=1
(28)
where σ is a permutation of the k cluster groups. The accuracy is always between [0, 1],
where 1 corresponds to all points correctly clustered, and 0 to all points wrongly clustered.
For a balanced two-class problem the value 1/2 correspond to chance.
We now consider two simple experiments where we sample n points from a two-class
mixture. We plot the average accuracy (28) versus n, with error bars indicating standard
error. The data is clustered using E 1D -clustering algorithm, GMM and k-means. For GMM
and k-means we use the implementations from the well-known scikit-learn library in Python
[27], where k-means is initialized through k-means++ procedure [28], and GMM is initialized
with the output of k-means. We run both algorithms 5 times with different initializations
and pick the answer with best objective function value. Notice that E 1D -clustering does not
require random initialization so we only run it once. For each n we use use 100 Monte Carlo
runs. In Fig. 1a we have the results for data sampled from the Gaussian mixture
iid
µ1 = 1.5, σ1 = 0.3, µ2 = 0, σ2 = 1.5.
x ∼ 21 N µ1 , σ12 + 12 N µ2 , σ22 ,
(29)
In this case the optimal accuracy obtained from Bayes classification error is ≈ 0.88, indicated
by the dashed line in the plot. The three methods perform closely, with a slight advantage
of GMM, as expected since it is a consistent model to the data, and E 1D -clustering performs
slightly better than k-means. In Fig. 1c we show a density estimation from clustering
1000 points from this mixture using the three algorithms. Notice that all of them are able
to distinguish the two classes. On the other hand, in Fig. 1b we consider a mixture of
lognormal distributions,
iid
x ∼ 21 exp N µ1 , σ12 + 12 exp N µ2 , σ22 ,
µ1 = 1.5, σ1 = 0.3, µ2 = 0, σ2 = 1.5. (30)
The optimal Bayes accuracy is again ≈ 0.88. We can now see that E 1D -clustering is still
very accurate, while GMM and k-means basically cluster at chance. Density estimation
after clustering 1000 points this mixture using the three algorithms are is shown in Fig. 1d.
Note that only E 1D -clustering was able to distinguish the two classes. k-means and GMM
put most of the points in a single cluster, and points on the tail of the second component of
(30) in the other cluster. The experiments of Fig. 1 illustrate how energy clustering is more
flexible compared to k-means and GMM.
17
0.90
0.90
0.80
0.85
accuracy
accuracy
0.85
1D
E -clustering
k-means
GMM
0.80
0.75
E 1D -clustering
k-means
GMM
0.70
0.65
0.60
0.55
0.75
100
200
300
400
500
600
700
800
0.50
100
200
300
0.8
truth
E 1D
k-means
GMM
700
800
0.5
0.4
0.3
0.4
0.2
0.2
0.1
−4
truth
E 1D
k-means
GMM
0.6
0.6
0.0
−6
600
0.7
1.4
1.0
500
(b)
(a)
1.2
400
# points
# points
−2
0
2
4
6
x
0.0
−2
0
2
4
6
8
10
12
14
x
(c)
(d)
FIG. 1. E 1D -clustering versus k-means and GMM. (a,b) We plot the mean accuracy (28) over
100 Monte Carlo trials, versus the number of sampled points. Error bars are standard error. The
dashed line indicates Bayes accuracy (≈ 0.88 in both cases). (a) Clustering results for data normally
distributed as in (29). (b) Data lognormally distributed as in (30). (c) Density estimation of each
component in the mixture (29) after clustering 1000 sampled points using the three algorithms,
compared to the ground truth. (d) The same but for lognormal data (30).
VI.
ITERATIVE ALGORITHMS FOR ENERGY CLUSTERING
In this section we introduce an iterative algorithm to find a local maximizer of the optimization problem (14). Due to Proposition 3 we can also find an approximate solution
by the well-known kernel k-means algorithm based on Lloyd’s heuristic [13, 14], which for
convenience will also be restated in the present context.
Consider the optimization problem (16) written as
max
{C1 ,...,Ck }
k
X
Qj
Q=
,
nj
j=1
18
Qj ≡
X
x,y∈Cj
K(x, y),
(31)
where Qj represents an internal energy cost of cluster Cj , and Q is the total energy cost
where each Qj is weighted by the inverse of the number of points in Cj . For a data point xi
we denote its own energy cost with the entire cluster C` by
Q` (xi ) ≡
X
y∈C`
K(xi , y) = Gi• · Z•` ,
where we recall that Gi• (G•i ) denotes the ith row (column) of matrix G.
Lloyd’s Method for Energy Clustering
To optimize kernel k-means objective function (19) we remove the global term and define
the function
J (`) (xi ) ≡
2
1
Q − Q` (xi ).
2 `
n`
n`
(32)
We are thus solving
min
Z
n X
k
X
Zi` J (`) (xi ).
i=1 `=1
One possible strategy is to assign xi to cluster Cj ? according to
j ? = arg min J (`) (xi ).
`=1,...,k
This is done for every data point xi and repeated until convergence, i.e. until no new
assignments are made. The entire procedure is described in Algorithm 2, which we name
E L -clustering to emphasize that we are optimizing the within energy function W based on
Lloyd’s method [7]. It can be shown that this algorithm converges provided G is positive
semidefinite.
E L -clustering is precisely kernel k-means algorithm [13, 14] but written more concisely and
with the kernel induced by energy statistics. Indeed, recalling that K(x, y) = hϕ(x), ϕ(y)i
where ϕ : X → HK is the feature map, we have from (32) that
J (`) (xi ) = hϕ(µ` ), ϕ(µ` )i − 2hϕ(xi ), ϕ(µ` )i = kϕ(xi ) − ϕ(µ` )k2 − kϕ(xi )k2 ,
where µ` =
1
n`
P
x∈C`
x is the mean of cluster C` . Therefore, min` J (`) (xi ) = min` kϕ(xi ) −
ϕ(µ` )k2 , i.e. we are assigning xi to the cluster with closest center (in feature space), which
is the familiar Lloyd’s heuristic approach that kernel k-means is based upon.
19
Algorithm 2 E L -clustering is Lloyd’s method for energy clustering, which is precisely
kernel k-means algorithm, with the kernel induced by energy statistics. This procedure
finds local solutions to the optimization problem (14).
input number of clusters k, Gram matrix G, initial label matrix Z ← Z0
output label matrix Z
1: q ← (Q1 , . . . , Qk )> have the costs of each cluster, defined in (31)
2: n ← (n1 , . . . , nk )> have the number of points in each cluster
3: repeat
4:
for i = 1, . . . , n do
5:
let j be such that xi ∈ Cj
6:
j ? ← arg min`=1,...,k J (`) (xi ), where J (`) (xi ) is defined in (32)
7:
if j ? 6= j then
8:
move xi to Cj ? : Zij ← 0 and Zij ? ← 1
9:
update n: nj ← nj − 1 and nj ? ← nj ? + 1
10:
update q: qj ← qj − 2Qj (xi ) and qj ? ← qj ? + 2Qj ? (xi )
11:
end if
12:
end for
13: until convergence
To check the complexity of E L -clustering, notice that to compute the second term of
J (`) (xi ) in (32) requires O(n` ) operations, and although the first term requires O(n2` ) it
only needs to be computed once outside loop through data points (step 1 of Algorithm 2).
Therefore, the time complexity of E L -clustering is O(nk max` n` ) = O(kn2 ). For a sparse
Gram matrix G having n0 nonzero elements this complexity can be further reduced to O(kn0 ).
Hartigan’s Method for Energy Clustering
We now consider Hartigan’s method [16] applied to the optimization problem in the
form (31), which gives a local solution to the QCQP defined in (14). The method is based
in computing the maximum change in the total cost function Q when moving each data
20
point to another cluster. More specifically, suppose point xi is currently assigned to cluster
Cj yielding a total cost function denoted by Q(j) . Moving xi to cluster C` yields another
total cost function denoted by Q(`) . We are interested in computing the maximum change
∆Qj→` (xi ) ≡ Q(`) − Q(j) , for ` 6= j. From (31), by explicitly writing the costs related to
these two cluster we obtain
−
∆Q
j→`
Qj
Qj
Q`
Q+
`
+
−
−
(xi ) =
n` + 1 nj − 1
nj
n`
−
where Q+
` denote the cost of the new `th cluster with the point xi added to it, and Qj is
the cost of new jth cluster with xi removed from it. Noting that Q+
` = Q` + 2Q` (xi ) + Gii
and Q−
j = Qj − 2Qj (xi ) + Gii , we get the formula
Qj
1
Q`
1
j→`
− 2Qj (xi ) + Gii −
− 2Q` (xi ) − Gii .
∆Q (xi ) =
nj − 1 nj
n` + 1 n`
(33)
Therefore, if ∆Qj→` (xi ) > 0 we get closer to a maximum of (31) by moving xi to C` , otherwise
we keep xi in Cj .
We thus propose the following algorithm. We start with an initial configuration for the
label matrix Z, then for each point xi we compute the cost of moving it to another cluster
C` , i.e. ∆Qj→` (xi ) for ` = 1, . . . , k with ` 6= j, where j denotes the index of its current
partition, x ∈ Cj . Hence, we choose
j ? = arg max ∆j→` (xi ).
`=1,...,k | `6=j
?
If ∆Qj→j (xi ) > 0 we move xi to cluster Cj ? , otherwise we keep xi in its original cluster Cj .
This process is repeated until no points are assigned to new clusters. The entire procedure
is explicitly described in Algorithm 3, which we denote E H -clustering to emphasize that it is
based on Hartigan’s method. This method automatically ensures that the objective function
is monotonically increasing at each iteration, and consequently the algorithm converges in
a finite number of steps.
The complexity analysis of E H -clustering is the following. Computing the Gram matrix
G requires O(Dn2 ) operations, where D is the dimension of each data point and n is the
data size. However, both algorithms E L - and E H -clustering assume that G is given. There
are more efficient methods to compute G, specially if it is sparse, but we will not consider
this further and just assume that G is given. The computation of each cluster cost Qj has
complexity O(n2j ), and overall to compute q we have O(n21 + · · · + n2k ) = O(k maxj n2j ).
21
Algorithm 3 E H -clustering is Hartigan’s method for energy clustering. This algorithm
finds local solutions to the optimization problem (14). The steps 6 and 10 are different
than E L -clustering described in Algorithm 2.
input number of clusters k, Gram matrix G, initial label matrix Z ← Z0
output label matrix Z
1: q ← (Q1 , . . . , Qk )> have the energy costs of each cluster, defined in (31)
2: n ← (n1 , . . . , nk )> have the number of points in each cluster
3: repeat
4:
for i = 1, . . . , n do
5:
let j be such that xi ∈ Cj
6:
j ? ← arg max`=1,...,k | `6=j ∆Qj→` (xi ) using (33)
7:
if ∆Qj→j (xi ) > 0 then
?
8:
move xi to Cj ? : Zij ← 0 and Zij ? ← 1
9:
update n: nj ← nj − 1 and nj ? ← nj ? + 1
10:
update q: qj ← qj − 2Qj (xi ) + Gii and qj ? ← qj ? + 2Qj ? (xi ) + Gii
11:
end if
12:
end for
13: until convergence
These operations only need to be performed a single time. For each point xi we need to
compute Qj (xi ) once, which is O(nj ), and we need to compute Q` (xi ) for each ` 6= j. The
cost of computing Q` (xi ) is O(n` ), thus the cost of step 6 in Algorithm 3 is O(k max` n` ) for
` = 1, . . . , k. For the entire dataset this gives a time complexity of O(nk max` n` ) = O(kn2 ).
Note that this is the same cost as in E L -clustering, or kernel k-means algorithm. Again, if
G is sparse this can be reduced to O(kn0 ) where n0 is the number of nonzero entries of G.
In the following we mention some important known results about Hartigan’s method.
Theorem 5 (Telgarsky-Vattani [17]). Hartigan’s method has the cost function strictly decreasing in each iteration. Moreover, if n > k then
1. the resulting partition has no empty clusters, and
22
2. the resulting partition has distinct means.
Neither of these two conditions are guaranteed to be satisfied by Lloyd’s method, and
consequently by E L -clustering algorithm. The next result indicates that Hartigan’s method
can potentially escape local optima of Lloyd’s method.
Theorem 6 (Telgarsky-Vattani [17]). The set of local optima of Hartigan’s method is a
(possibly strict) subset of local optima of Lloyd’s method.
The above theorem implies that E L -clustering cannot improve on a local optima of E H -
clustering. On the other hand, E H might improve on a local optima of E L . Lloyd’s method
forms Voronoi partitions, while Hartigan’s method groups data in regions formed by the
intersection of spheres called circlonoi cells. It can be shown that the circlonoi cells are
contained within a smaller volume of a Voronoi cell, and this excess volume grows exponentially with the dimension of X [17, Theorems 2.4 and 3.1]. Points in this excess volume
force Hartigan’s method to iterate, contrary to Lloyd’s method. Therefore, Hartigan’s can
escape local optima of Lloyd’s. Moreover, this improvement should be more prominent as
dimension increases. Also, the improvement grows as the number of clusters k increases.
The empirical results of [17] show that an implementation of Hartigan’s method has comparable execution time to an implementation of Lloyd’s method, but no explicit complexity
was provided. We show that both E L - and E H -clustering have the same time complexity. To
the best of our knowledge, Hartigan’s method was not previously considered together with
kernels, as we are proposing in E H -clustering algorithm.
In [18], Hartigan’s method was applied to k-means problem with any Bregman divergence.
It was shown that the number of Hartigan’s local optima is upper bounded by O(1/k) [18,
Proposition 5.1]. In addition, it was provided examples where any initial partition correspond to a local optima of Lloyd’s method, while the number of local optima in Hartigan’s
method is small and correspond to true partitions of the data. Empirically, the number of
Hartigan’s local optima was considerably smaller than the number of Lloyd’s local optima.
The above results indicate that Hartigan’s method provides several advantages over
Lloyd’s method, a fact that will also be supported by our numerical experiments in the
next section where E H outperforms of E L (kernel k-means) in several settings, specially in
high dimensions.
23
VII.
NUMERICAL EXPERIMENTS
The main goal of this section is threefold. First, we want to compare E H -clustering in
Euclidean space to k-means and GMM. Second, we want to compare E H -clustering, based
on Hartigan’s method, to E L -clustering or kernel k-means, based on Lloyd’s method, and
also to spectral clustering, when they all operate on the same kernel. Third, we want to
illustrate the flexibility provided by energy clustering, which is able to cluster accurately in
different settings while keeping the same kernel.
The following experimental setup holds unless specified otherwise. We consider E H -
clustering, E L -clustering and spectral clustering with the following semimetrics and corresponding generating kernels:
ρα (x, y) = kx − ykα ,
Kα (x, y) = 21 (kxkα + kykα − kx − ykα ) ,
(34)
ρeσ (x, y) = 2 − 2e−
e σ (x, y) = e−
K
kx−yk
2σ ,
(35)
kx−yk2
2σ 2 .
(36)
ρbσ (x, y) = 2 − 2e−
kx−yk
2σ ,
kx−yk2
2σ 2 ,
b σ (x, y) = e−
K
The relation between kernel and semimetric is given by formula (6) where we fix x0 = 0. The
standard ρ1 , from the original energy distance (1), will always be present in the experiments
as a reference, being the implied choice unless explicitly mentioned. For k-means, GMM
and spectral clustering we use the robust implementations of scikit-learn library [27], where
k-means is initialized with k-means++ [28], and GMM with the output of k-means, making
it more robust and preventing it from breaking in high dimensions. Spectral clustering
implementation is based on [22]. We implemented E L -clustering as described in Algorithm 2,
and E H -clustering as described in Algorithm 3. Both will also be initialized with k-means++.
We run the algorithms 5 times with different initializations, picking the result with best
objective function value. We evaluate clustering quality by the accuracy (28) based on the
true labels. For each setting we show the average accuracy over 100 Monte Carlo trials, with
error bars indicating standard error.
We briefly mention that we compared E H -clustering, as described in Algorithm 3, to
E 1D -clustering, described in Algorithm 1, for several univariate distributions. Both perform
very closely. However, we omit these results since we will analyse more interesting scenarios
in high dimensions.
From the results of [17], summarized in the end of the previous section, we expect the
24
x1
−3
5
4
3
2
1
0
−1
−2
−3
x2
4
3
2
1
0
−1
−2
−3
−4
x2
8
6
4
2
0
−2
−4
−6
−8
x3
4
4
3
2
1
0
−1
−2
−3
−4
3
x1
2
1
0
−1
6
4
x3
2
−4
−6
4
10
3
8
2
6
1
4
0
−1
2
0
−2
−2
4
4
3
3
2
2
−3
−4
1
x5
x5
0
−2
x4
x4
−2
0
1
0
−1
−1
−3
−3 −2 −1 0
−3
−2 −1
−2
−2
1
x1
2
3
4
−4 −3 −2 −1 0 1 2 3 4
x2
−3 −2 −1
0
x3
1
2
3
−3 −2 −1 0
1
x4
2
3
4
−3 −2 −1 0
1
x5
2
3
4
0
1
x1
2
3
4
−6 −4 −2
0
x2
2
4
(a)
6
−4−3−2−1 0 1 2 3 4 5
x3
−4 −2
0
2
x4
4
6
8
−3 −2 −1 0
1
x5
2
3
4
(a)
FIG. 2. Pair plots for the first 5 dimensions. (a) Data normally distributed as in (37). (b) Data
normally distributed as in (38). We sample 200 points for both cases. We can see that there is a
considerable overlap between the clusters.
improvement of Hartigan’s over Lloyd’s method to be more accentuated in high dimensions.
Thus, we analyze how the algorithms degrade as the number of dimensions increase while
keeping the number of points in each cluster fixed. Consider data from the Gaussian mixture
iid
x ∼ 21 N (µ1 , Σ1 ) + 12 N (µ2 , Σ2 ),
Σ1 = Σ2 = ID ,
µ1 = (0, . . . , 0)> ,
| {z }
×D
µ2 = 0.7(1, . . . , 1, 0, . . . , 0)> .
| {z } | {z }
×10
(37)
×(D−10)
To get some intuition about how separated data points from each class are, we show scatter
plots between the first 5 dimensions in Fig. 2a. Note that the Bayes error is fixed as D
increases, yielding an optimal accuracy of ≈ 0.86. We sample 200 points on each trial. The
results are shown in Fig. 3a. We can see that E H and spectral clustering have practically
the same performance, which is higher than E L -clustering (kernel k-means). Moreover, E H
outperforms k-means and GMM, where the improvement is noticeable specially in high
dimensions. Note that in this setting k-means and GMM are consistent models to the data,
however, energy clustering degrades much less as dimension increases.
Still for a two-class Gaussian mixture, we now allow the diagonal entries of one of the
25
0.95
0.85
0.90
0.75
0.70
0.65
0.60
0.55
0.85
accuracy
accuracy
0.80
EH
EL
spectral
k-means
GMM
0.80
0.75
0.70
0.65
0.60
50
100
150
200
0.55
EH
EL
spectral
k-means
GMM
100
# dimensions
200
300
400
500
600
700
# dimensions
(a)
(b)
FIG. 3. Comparison of E H -clustering, E L -clustering (kernel k-means), spectral clustering, k-means
and GMM in high dimensional Gaussian settings. We plot the mean accuracy versus the number
of dimensions, with error bars indicating standard error from 100 Monte Carlo runs. (a) Data
normally distributed as in (37), with Bayes accuracy ≈ 0.86, over the range D ∈ [10, 200]. (b)
Data normally distributed as in (38), with Bayes accuracy ≈ 0.95, over the range D ∈ [10, 700].
covariances to have different values by choosing
iid
x ∼ 12 N (µ1 , Σ1 ) + 12 N (µ2 , Σ2 ),
µ1 = (0, . . . , 0)> ,
| {z }
×D
µ2 = (1, . . . , 1, 0, . . . , 0)> ,
| {z } | {z }
×10
×(D−10)
Σ1 = ID ,
Σ2 =
e 10
Σ
0
0
ID−10
e 10 = diag(1.367, 3.175, 3.247, 4.403, 1.249, 1.969, 4.035, 4.237, 2.813, 3.637).
Σ
,
(38)
We simply chose a fixed set of 10 numbers uniformly at random on the interval [1, 5] for the
e 10 , and any other choice would give analogous results. We show pair plots of
diagonal of Σ
this data in Fig. 2b. We sample a total of 200 points from (38) on each trial. The Bayes error
is kept fixed when increasing D yielding an optimal accuracy ≈ 0.95. In Fig. 3b we see that
GMM performs better in low dimensions, but it quickly degenerates as D increases. The
same is true for k-means and E L -clustering. However, E H and spectral clustering remains
much more stable in high dimensions. Notice that a naive implementation of GMM should
not be able to estimate the covariances when D & 100, however, scikit-learn library uses
k-means output as initialization, therefore the output of GMM in this implementation is at
least as good as k-means and the algorithm is more robust in high dimensions.
26
Consider sampling data from the following Gaussian mixture in R20 :
iid
x ∼ 21 N (µ1 , Σ1 ) + 12 N (µ2 , Σ2 ),
µ1 = (0, . . . , 0)> ,
| {z }
×20
µ2 = 12 (1, . . . , 1, 0, . . . , 0)> ,
| {z } | {z }
5
Σ1 = 12 I20 ,
Σ2 = I20 .
(39)
15
The optimal accuracy based on Bayes classification error is ≈ 0.90. We increase the sample
size n ∈ [10, 400] and show the accuracy versus n for the different kernels (34) and (35)
within E H -clustering algorithm, which are compared to k-means and GMM. The results are
in Fig. 4a. Note that for small n all methods are superior than GMM, which slowly catches
up and tend to optimal Bayes, as expected since it is a consistent model to the data. Note
e 1 is as accurate as GMM for large number of points,
also that E H -clustering with kernel K
however, it is superior for small number of points. Still for the same setting, in Fig. 4b we
show the difference in accuracy provided by E H minus E L and E H minus spectral clustering,
e 1 . Note that E H was always superior than kernel k-means and
when using the kernel K
spectral clustering, otherwise there would be points negative values on the y-axis.
Consider the same experiment but now with a lognormal mixture,
iid 1
2
x∼
exp {N (µ1 , Σ1 )} + 12 exp {N (µ2 , Σ2 )} ,
µ1 = (0, . . . , 0)> ,
| {z }
×20
µ2 = 12 (1, . . . , 1, 0, . . . , 0)> ,
| {z } | {z }
5
Σ1 = 12 I20 ,
Σ2 = I20 .
(40)
15
The results are in Fig. 4c. Energy clustering still performs accurately, with any of the
utilized kernels, providing better results than k-means and GMM on this non-normal data.
e 1 still provides the best results for small number of points, but its performance is
The kernel K
eventually achieved by K1/2 , indicating that α ≈ 1/2 in the standard energy distance should
be more appropriate for skewed distributions. In Fig. 4d we show the difference between E H e 1 . Again, the accuracy
clustering to kernel k-means and spectral clustering, with the kernel K
provided by E H is higher than the other methods, although not much higher than spectral
clustering in this example. The two experiments of Fig. 4 illustrate how energy clustering is
more flexible, performing well in different settings with the same kernel, contrary to k-means
and GMM.
In Fig. 5a–c we have complex two dimensional datasets. The two parallel cigars in (a)
have 200 points each. The concentric circles in (b) and (c) have 400 points for each class.
We apply E H -clustering with the kernels (34), (35) and (36). We also consider the best
27
0.90
0.85
0.85
0.80
E H , ρ1
0.75
E H , ρ1/2
E H , ρe1
0.70
50
100
150
200
250
300
350
k-means
GMM
0.70
50
100
150
200
250
# points
# points
(a)
(c)
300
350
400
0.25
difference in accuracy
difference in accuracy
E H , ρe1
0.75
0.55
400
0.25
0.20
0.15
0.10
0.00
E H , ρ1/2
0.80
0.60
0.30
0.05
E H , ρ1
0.65
k-means
GMM
0.65
0.60
accuracy
accuracy
0.90
EH − EL
E H − spectral
50
100
150
200
250
300
350
400
EH − EL
0.20
E H − spectral
0.15
0.10
0.05
0.00
50
100
150
200
250
# points
# points
(b)
(d)
300
350
400
FIG. 4. E H -clustering with kernels (34) and (35) versus k-means and GMM. In both settings Bayes
accuracy is ≈ 0.9. We show average accuracy (error bars are standard error) versus number of
points for 100 Monte Carlo trials. (a,b) Gaussian mixture (39). (c,d) Lognormal mixture (40).
The plots in (c) and (d) consider the difference in accuracy between E H versus E L (kernel k-means)
e1.
and spectral clustering, with the kernel K
(a)
(b)
(c)
(d)
FIG. 5. (a) Parallel cigars. (b) Two concentric circles with noise. (c) Three concentric circles with
noise. (d) MNIST handwritten digits. Clustering results are in Table I and Table II.
28
TABLE I. Clustering data from Fig. 5a–c.
Fig. 5a
ρ1
E H -clustering
spectral-clustering
k-means
GMM
0.705 ± 0.065
ρ1/2 0.952 ± 0.048
ρ1
Fig. 5b
Fig. 5c
0.521 ± 0.005 ρ1
0.393 ± 0.020
ρ1/2 0.522 ± 0.004 ρ1/2 0.486 ± 0.040
ρe2
0.9987 ± 0.0008 ρe1
0.778 ± 0.075 ρe2
0.666 ± 0.007
ρe2
0.557 ± 0.014
ρb1
0.732 ± 0.002 ρb2
0.364 ± 0.004
0.522 ± 0.004 7
0.368 ± 0.005
7
0.595 ± 0.011 7
0.465 ± 0.030
ρb2
0.956 ± 0.020
7
0.550 ± 0.011
7
0.903 ± 0.064
ρb1
7
1.0 ± 0.0
ρb2
0.676 ± 0.002
TABLE II. Clustering MNIST data from Fig. 5d.
{0, 1, . . . , 4}
{0, 1, . . . , 6}
{0, 1, . . . , 8}
{0, 1, . . . , 9}
σ
10.41
10.41
10.37
10.19
ρ1
0.873 ± 0.025
0.731 ± 0.016
0.687 ± 0.016
0.581 ± 0.011
ρ1/2 0.874 ± 0.027
0.722 ± 0.017
0.647 ± 0.017
0.600 ± 0.009
0.847 ± 0.031
0.695 ± 0.023
0.657 ± 0.014
0.584 ± 0.013
Class Subset
parameter
E H -clustering
ρeσ
ρbσ
0.891 ± 0.009 0.759 ± 0.011 0.704 ± 0.011 0.591 ± 0.012
0.769 ± 0.012
0.678 ± 0.014
0.649 ± 0.018
0.565 ± 0.009
k-means
7
0.878 ± 0.010
0.744 ± 0.008
0.695 ± 0.012
0.557 ± 0.012
GMM
7
0.839 ± 0.015
0.694 ± 0.010
0.621 ± 0.009
0.540 ± 0.009
spectral-clustering
ρbσ
kernel choice for each example for spectral clustering. Moreover, we consider k-means and
GMM. We perform 10 Monte Carlo runs for each example. The results are in Table I. For
(a) we initialize all algorithms with k-means++, and for (b) and (c) we initialize at random.
E H has superior performance in every example, and in particular better than the spectral
clustering. In (a) the standard kernel from energy statistics in Euclidean space, K1 and
K1/2 , are able to provide accurate results, however, for the examples in (b) and (c) the
b 1 and K
b 2 provide a significant improvement.
kernel choice is more sensitive, where K
Next, we consider the infamous MNIST handwritten digits as illustrated in Fig. 5d. Each
29
data point is an 8-bit gray scale image forming a 784-dimensional vector corresponding to
the digits {0, 1, . . . , 9}. We compute the parameter
n
1 X
kxi − xj k2 ,
σ = 2
n i,j=1
2
from a separate training set, to be used in the kernels (35) and (36). We consider subsets
of {0, 1, . . . , 9}, sampling 100 points for each class. The results are shown in Table II,
where kernels and parameters are indicated. E H -clustering performs slightly better than
k-means and GMM, however the difference is not considerable. Unsupervised clustering on
MNIST without any feature extraction is not trivial. For instance, the same experiment
was performed in [29] where a low-rank transformation is learned then subsequently used
in subspace clustering, providing very accurate results. It would be interesting to explore
analogous methods for learning a better representation of the data and subsequently apply
E H -clustering.
VIII.
DISCUSSION
We proposed clustering from the perspective of generalized energy statistics, valid for
arbitrary spaces of negative type. Our mathematical formulation of energy clustering reduces to a QCQP in the associated RKHS, as demonstrated in Proposition 2. We showed
that the optimization problem is equivalent to kernel k-means, once the kernel is fixed; see
Proposition 3. Energy statistics, however, fixes a family of standard kernels in Euclidean
space, and more general kernels on spaces of negative type can also be obtained. We also
considered a weighted version of energy statistics, whose clustering formulation establishes
connections with graph partitioning. We proposed the iterative E H -clustering algorithm
based on Hartigan’s method, which was compared to kernel k-means algorithm based on
Lloyd’s heuristic. Both have the same time complexity, however, numerical and theoretical results provide compelling evidence that E H -clustering is more robust with a superior
performance, specially in high dimensions. Furthermore, energy clustering, with standard
kernels from energy statistics, outperformed k-means and GMM on several settings, illustrating the flexibility of the proposed method which is model-free. In many settings, the
iterative E H -clustering also surpassed spectral clustering, which is solves a relaxation of the
original QCQP, and in other settings performed closely but never worse. Note that spec30
accuracy
1.00
0.95
0.90
0.85
0.80
0.75
0.70
0.65
0.60
0.55
EH
EL
spectral
k-means
GMM
0
50
100
150
200
# unbalanced points
FIG. 6. Comparison of energy clustering algorithms to k-means and GMM on unbalanced clusters.
The data is normally distributed as (41), where we vary m ∈ [0, 240], and in each case we do 100
Monte Carlo runs showing the average accuracy with standard error.
tral clustering is more expensive than our iterative method, going up to O(n3 ), and finding
eigenvectors of very large matrices is problematic.
A limitation of the proposed methods for energy clustering is that it cannot handle
accurately highly unbalanced clusters. As an illustration, consider the following Gaussian
mixture:
n1
n1
N (µ1 , Σ1 ) +
N (µ1 , Σ1 ), µ1 = (0, 0, 0, 0)> , µ2 = 1.5 × (1, 1, 0, 0)> ,
2N
2N
1
I2 0
, n1 = N − m, n2 = N + m, N = 300.
Σ1 = I4 , Σ2 = 2
0 I2
iid
x∼
(41)
We then increase m ∈ [0, 240] making the clusters progressively more unbalanced. We plot
the average accuracy over 100 Monte Carlo runs for each m, with error bars indicating
standard error. The results are shown in Fig. 6. For highly unbalanced clusters we see that
GMM performs better than the other methods, which have basically similar performance.
Based on this experiment, an interesting problem would be to extend E H -clustering algorithm
to account for highly unbalanced clusters.
Moreover, it would be interesting to formally demonstrate cases where energy clustering
is a consistent in the large n limit. A soft version of energy clustering is also an interesting
extension. Finally, kernel methods can benefit from sparsity and fixed-rank approximations
of the Gram matrix, and there is plenty of room to make E H -clustering algorithm more
scalable.
31
ACKNOWLEDGMENTS
We would like to thank Carey Priebe for discussions. We would like to acknowledge
the support of the Transformative Research Award (NIH #R01NS092474) and the Defense
Advanced Research Projects Agencys (DARPA) SIMPLEX program through SPAWAR contract N66001-15-C-4041.
[1] G. J. Székely and M. L. Rizzo. Energy Statistics: A Class of Statistics Based on Distances.
Journal of Statistical Planning and Inference, 143:1249–1272, 2013.
[2] M. L. Rizzo and G. J. Székely. DISCO Analysis: A Nonparametric Extension of Analysis of
Variance. The Annals of Applied Statistics, 4(2):1034–1055, 2010.
[3] G. J. Székely and M. L. Rizzo. Hierarchical Clustering via Joint Between-Within Distances:
Extending Ward’s Minimum Variance Method. Journal of Classification, 22(2):151–183, 2005.
[4] S. Li. k-Groups: A Generalization of k-Means by Energy Distance. PhD Thesis, Bowling
Green State University, 2015.
[5] R. Lyons. Distance Covariance in Metric Spaces. The Annals of Probability, 41(5):3284–3305,
2013.
[6] D. Sejdinovic, B. Sriperumbudur, A. Gretton, and K. Fukumizu. Equivalence of DistanceBased and RKHS-Based Statistic in Hypothesis Testing. The Annals of Statistics, 41(5):2263–
2291, 2013.
[7] S. P. Lloyd. Least Squares Quantization in PCM. IEEE Transactions on Information Theory,
28(2):129–137, 1982.
[8] J. B. MacQueen. Some Methods for Classification and Analysis of Multivariate Observations.
In Proceedings of the 5th Berkeley Symposium on Mathematical Statistics and Probability,
volume 1, pages 281–297. University of California Press, 1967.
[9] E. Forgy. Cluster Analysis of Multivariate Data: Efficiency versus Interpretabiliby of Classification. Biometrics, 21(3):768–769, 1965.
[10] B. Schölkopf, A. J. Smola, and K. R. Müller. Nonlinear Component Analysis as a Kernel
Eigenvalue Problem. Neural Computation, 10:1299–1319, 1998.
[11] M. Girolami. Kernel Based Clustering in Feature Space. Neural Networks, 13(3):780–784,
32
2002.
[12] J. Mercer. Functions of Positive and Negative Type and their Connection with the Theory of
Integral Equations. Proceedings of the Royal Society of London, 209:415–446, 1909.
[13] I. S. Dhillon, Y. Guan, and B. Kulis. Kernel K-means: Spectral Clustering and Normalized
Cuts. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD ’04, pages 551–556, New York, NY, USA, 2004. ACM.
[14] I. S. Dhillon, Y. Guan, and B. Kulis. Weighted Graph Cuts without Eigenvectors: A Multilevel
Approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(11):1944–
1957, 2007.
[15] M. Filippone, F. Camastra, F. Masulli, and S. Rovetta. A Survey of Kernel and Spectral
Methods for Clustering. Pattern Recognition, 41:176–190, 2008.
[16] J. A. Hartigan and M. A. Wong. Algorithm AS 136: A k-Means Clustering Algorithm. Journal
of the Royal Statistical Society. Series C (Applied Statistics), 28(1):100–108, 1979.
[17] M. Telgarsky and A. Vattani. Hartigan’s Method: k-Means Clustering without Voronoi.
In Proceedings of the 13th International Conference on Artificial Intelligence and Statistics
(AISTATS), volume 9, pages 313–319. JMLR, 2010.
[18] N. Slonim, E. Aharoni, and K. Crammer. Hartigan’s k-Means versus Lloyd’s k-Means —
Is it Time for a Change? In Proceedings of the 20th International Conference on Artificial
Intelligence, pages 1677–1684. AAI Press, 2013.
[19] N. Aronszajn. Theory of Reproducing Kernels. Transactions of the American Mathematical
Society, 68(3):337–404, 1950.
[20] A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola. A Kernel Two-Sample
Test. Journal of Machine Learning Research, 13:723–773, 2012.
[21] C. Berg, J. P. R. Christensen, and P. Ressel. Harmonic Analysis on Semigroups: Theory of
Positive Definite and Related Functions. Graduate Text in Mathematics 100. Springer, New
York, 1984.
[22] J. Shi and J. Malik. Normalized Cust and Image Segmentation. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 22(8):888–905, 2000.
[23] A. Y. Ng, M. I. Jordan, and Y. Weiss. On Spectral Clustering: Analysis and an Algorithm. In
Advances in Neural Information Processing Systems, volume 14, pages 849–856, Cambridge,
MA, 2001. MIT Press.
33
[24] B. Kernighan and S. Lin. An Efficient Heuristic Procedure for Partitioning Graphs. The Bell
System Technical Journal, 49(2):291–307, 1970.
[25] P. Chan, M. Schlag, and J. Zien. Spectral k-Way Ratio Cut Partitioning. IEEE Transactions
on Computer-Aided Design of Integrated Circuits and Systems, 13:1088–1096, 1994.
[26] S. X. Yu and J. Shi. Multiclass Spectral Clustering. In Proceedings Ninth IEEE International
Conference on Computer Vision, volume 1, pages 313–319, 2003.
[27] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel,
P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher,
M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine
Learning Research, 12:2825–2830, 2011.
[28] D. Arthur and S. Vassilvitskii. k-means++: The Advantage of Careful Seeding. In Proceedings
of the Eighteenth annual ACM-SIAM Symposium on Discrete Algorithms, pages 1027–1035,
Philadelphia, PA, USA, 2007. Society for Industrial and Applied Mathematics.
[29] Q. Qui and G. Sapiro. Learning Transformations for Clustering and Classification. Journal
of Machine Learning Research, 16:187–225, 2015.
34
| 1cs.CV
|
1
Modelling of Facial Aging and Kinship: A Survey
arXiv:1802.04636v1 [cs.CV] 13 Feb 2018
Markos Georgopoulos, Yannis Panagakis, and Maja Pantic,
1
Abstract—Computational facial models that capture properties of facial cues related to aging and kinship increasingly attract the
attention of the research community, enabling the development of reliable methods for age progression, age estimation, age-invariant
facial characterization, and kinship verification from visual data. In this paper, we review recent advances in modelling of facial aging
and kinship. In particular, we provide an up-to date, complete list of available annotated datasets and an in-depth analysis of geometric,
hand-crafted, and learned facial representations that are used for facial aging and kinship characterization. Moreover, evaluation
protocols and metrics are reviewed and notable experimental results for each surveyed task are analyzed. This survey allows us to
identify challenges and discuss future research directions for the development of robust facial models in real-world conditions.
✦
I NTRODUCTION
Humans possess explicit, cue-based, and often culturally
determined systems for perceiving the facial appearance of
their peers [248]. Facial appearance is a primary source of information regarding the person’s identity, gender, ethnicity,
affective state, head pose, age and kinship relations. Hence,
the perception of facial attributes governs person perception, interpersonal attraction, and consequently prosocial
and social behaviour [5], [169].
Human face has been thoroughly studied from different,
but complementary, perspectives across several disciplines
such as neuroscience e.g., [70], psychology e.g., [23], [156],
sociology e.g., [61], anthropometry e.g, [66], medicine e.g.,
[51] and computer science. From a computational point of
view, in particular, advances in computational face modelling enabled the development of reliable methods for
automatic detection of faces [247], recognition of identity
[257], [107], [2], gender [158], and ethnicity [72]; detection of
salient facial features [221], [58] and estimation of head pose
[50], analysis of facial expressions [189], [168], [249] from
visual data. Notably, recently proposed methods match or
even achieve better accuracy than humans in several tasks
e.g., [97]. This progress herald a surge of novel applications in communication, entertainment, cosmetology, and
biometrics, to name a few, while facilitating basic research
in social sciences and medicine e.g., [157]. A thorough list
of machine learning and computer vision methods solving
the aforementioned face modelling and analysis tasks can
be found in the comprehensive survey papers [247], [257],
[107], [158], [72], [221], [50], [189], [249].
Research towards the development of more detailed
computational facial models that capture properties of facial cues related to aging and kinship increasingly attracts
the attention of the community. Indeed, by capitalizing on
recent advances in machine learning, computer vision, and
•
M. Georgopoulos, Y. Panagakis and M. Pantic are with the Dept. of
Computing, Imperial College London, UK.
Y. Panagakis is also with the Dept. of Computer Science, Middlesex
University London.
Corresponding
author:
M.
Georgopoulos,
email:
[email protected].
the available massive collections of facial data, significant
progress has been made towards addressing the following
problems:
i
ii
iii
iv
Age Progression: that is, the process of transforming a
facial visual input, in order to model it across different ages. The change of the age can be bidirectional,
so that the facial output can appear either younger
or older than the input.
Age Estimation: refers to the process of labelling a
facial signal with an age or age group. The input signal can be 2D, 3D or image sequences. The problems
that fall into this category can be divided further
into two subcategories, depending on the labels of
the training data: (a) real age or (b) apparent age
estimation, which refers to the age that is inferred
by humans based on the individual’s appearance.
Age-Invariant Facial Characterization: involves the
process of building a signal representation that is
invariant to the facial transformations and appearance changes caused by aging.
Kinship Verification: is defined as the process of determining whether the individuals in a pair of facial
visual inputs are blood related.
Early models for facial age progression and estimation
date back to 1994-95 [118], [43], while the problem of face
recognition across ages was first investigated in 2000 [121].
More recently, since 2010, methods for kinship verification
have emerged [65]. Since then, the development of 1) robust and computationally efficient models (e.g., AAMs [40],
CLMs [42] etc) and descriptors (e.g., SIFT [144], HoGs [44],
LPPs [99], SURF [12], DAISY[213] etc) of facial appearance
2) effective machine learning methods such as Boosting
[71] and Support Vector Machines [218], and 3) manually
annotated facial datasets e.g., MORPH2 (2006), FG-NET
(2004), FERET (1998), YGA (2008), Gallagher’s Images of
Groups (2009), Ni’s Web-Collected Images (2009) , facilitated
the deployment of reliable computational models for facial
aging and kinship. Models, methods, and data for facial
aging modelling which have been published before 2010
are thoroughly surveyed in [178], [73], while an overview
of research efforts for facial kinship modelling is currently
missing.
2
This paper aims to provide an up-to-date literature survey of the work done 1) towards the development of facial
aging models, complementing previous studies in [73], [178]
in several ways and 2) in the emerging topic of facial kinship
modelling. Concretely, the aims of this survey are organized
as follows:
•
•
•
A complete catalogue of publicly available datasets
with manual annotations for facial age and kinship
modelling tasks is listed in Section 3. We put particular emphasis on data collected in naturalistic, realworld (in the wild) conditions by providing reviews
for 9 recently collected datasets for age modelling
and all the available (i.e., 16) collections of facial
images for kinship verification.
A comprehensive review of recent as well as seminal
methodologies for age progression, age estimation,
age-invariant facial characterization, and kinship
verification is provided in Sections 4-7. In particular,
we provide an in-depth analysis of both geometric,
hand-crafted and learned facial representations for
the aforementioned tasks and discuss the type of
information they encode as well as their advantages
and limitations. We further elaborate methods that
rely on deep discriminative (e.g., CNNs [125]) and
generative (e.g., GANs [85]) models and appear to
be highly effective. Moreover, we review evaluation
protocols and metrics, and analyze the most notable
experimental results for each surveyed task.
The review of data, and computational methods for
facial aging and kinship reveal useful practices as
well as challenges that are yet to be solved. These
along with drawn conclusions are discussed in Section 9.
To begin with, a number of modern applications of computational models for facial aging and kinship are discussed
in the following section.
updating archive images of wanted criminals as well as
images of lost children. Additionally, cases such as matching
orphaned or lost children and finding kins of a victim, to
name a few, demand the verification of kin relationship of
two people. To that end, automatic genealogical research can
significantly aid the work of law enforcement agencies.
Medicine and Cosmetology: Being able to model aging and
kinship and simulating the transformations on the face is
vital for modern medicine and cosmetology. Medical home
systems that are used to monitor elderly people can aid
medical diagnosis by detecting premature aging. On the
other hand, automatic rejuvenation of the face can serve
as a guide for cosmetic surgery. Particularly in the case of
children, the parents’ craniofacial aging patterns can be used
to predict the child’s head growth, so that injury-related
cosmetic surgery can have optimal long-term results.
Commercial use: The ever-growing usage of social media
and availability of personal photos have led to the rapid
integration of facial analysis by businesses. Automatically
estimating the customers’ age can help with efficient customer profiling and age-oriented decision making, e.g., ageoriented advertisements. Likewise, targeted ads can be more
effective when taking kin relation into considerations, as
people’s preferences can be affected by their relatives.
Entertainment: Visual effects that can age or rejuvenate
the actors are already being used in the film making industry. These effects are not limited to movies but are also
widely applied to photo editing. The imminent integration
of such tools into popular design software will make for
more realistic retouche of photos. Make-up artists that specialize in transforming the face can leverage the construction
of person, age and kin specific morphable models. Guided
by those models, the artists will transform the face of the
actor for roles that demand sibling-like similarity between
actors.
3
2
A PPLICATIONS
In this section we present the most significant applications
of modelling facial aging as well as kinship in biometrics,
forensics, medicine, cosmetology, business and entertainment.
Biometrics: The physical, physiological or behavioural
cues based on which a person is recognized, e.g., iris,
fingerprint, face are referred to as biometrics [108]. Age
and kinship comprise soft biometrics [109], [45], [163] as
they can be used to boost the effectiveness of recognition.
Besides improving face recognition accuracy there is a need
for robustness towards aging and kinship. Passport checks
demand age-invariance in case of large age gap between
the passport image and the person in question. Similarly,
kinship invariance can potentially boost automatic face
recognition, in particular towards distinguishing between
kins that look alike.
Forensics: Forensics include a set of scientific techniques
that are used for crime detection. Among these techniques,
forensics art demonstrates the challenging task of producing
a lifelike image of a person. In some cases, forensics experts
face the need to change the age of a face. Such cases include
DATASETS
The availability of labelled datasets is a cornerstone towards
the development of facial models for aging and kinship.
In this section, we review recently released datasets which
contain annotations for age progression, age estimation, ageinvariant facial characterization, and kinship verification.
Particular emphasis is put on data capturing naturalistic,
real-world conditions, often referred as in the wild [105]. A
complete catalogue of the available datasets for facial age
modeling is listed in Table 1 while Table 2 contains the
available facial data for kinship verification.
3.1 Datasets with age labels
The vast majority of the available datasets in Table 1 for
facial age modeling contain still images and apart from
the FACES, IRIP, LHI, and YGA face datasets, they are not
balanced with regards to the gender and age of the subjects.
Also, while containing an abundance of different annotated
faces, many of these datasets do not contain a considerable
number of images of the same person at different ages,
which is essential for training methods for age progression
and age-invariant facial characterization.
3
TABLE 1: Datasets with age labels
Databases
#data
#subjects
age-range
precision
In the Wild
Year
modality
AgeDB [155]
IMDB-WIKI [184]
AFAD [161]
IRIP [243]
OUI-Adience [57]
CACD [32]
UvA-Nemo [49]
VADANA [202]
LHI Face Dataset [1]
HOIP [69]
FACES [55]
Web Image Db [159]
Images of Groups [77]
YGA [88]
Iranian Face [10]
Brown Sisters [162]
Scherbaum’s Db [191]
MORPH 2 [180]
WIT-BD [215]
AI&R [76]
FRGC [172]
Lifespan Database [154]
FG-NET [119],[122]
PIE [199]
FERET [173]
Caucasian Face Db [24]
16,458
523,051
164,432
2,100
26,580
163,446
1,247
2,298
8,000
306,600
1026
219,892
28,231
8,000
3,600
16
438
55.134
26,222
34
50,000
576
1.002
41,638
14,126
147
568
20,284
N/A
N/A
2,284
2,000
400
43
8,000
300
171
N/A
28,231
1,600
616
4
438
13.618
5,500
17
568
576
82
68
1,199
147
1-101
0-100
14-40+
1-70
0-60+
N/A
8-76
0-78
9-89
15-64
19-80
1-80
0-66+
0-93
2-85
15-62
8-18 or adult
16-77
3-85
22-66
18-70
18-93
0-69
N/A
N/A
20-62
exact age
exact age
exact age
exact age
age group
exact age
exact age
age group
exact age
age group
exact age
exact age
age group
exact age
exact age
exact age
both
exact age
age group
exact age
exact age
age group
exact age
exact age
exact age
exact age
yes
yes
yes
yes
yes
yes
no
yes
no
no
no
yes
yes
yes
yes
yes
no
no
yes
no
partially
yes
yes
no
partially
no
2017
2016
2016
2016
2014
2014
2014
2011
2010
2010
2010
2009
2009
2008
2007
2007
2007
2006
2006
2006
2005
2004
2004
2002
1998
1995
images
images
images
images
images
images
videos
images
images
images
images
images
images
images
images
images
3D
images
images
images
images, 3D
images
images
images
images
images
TABLE 2: Datasets with kinship labels
Databases
#data
#kin pairs
#relationships
In the Wild
same photo
Year
modality
KFVW [240]
WVU Kinship Db [115]
Families In the Wild [183]
TSKinface [175]
Sibling-Face [96]
Group-Face [96]
HQfaces [219]
LQfaces [219]
KinFaceW-I [149]
KinFaceW-II [149]
Family 101 [64]
IIITD-Kinship [114]
VADANA [202]
UvA-Nemo [49]
UBKinface [234], [193]
CornellKin [65]
836
904
11,193
787
N/A
106
184
196
1,066
2,000
14,816
544
2,298
515
600
300
418
113
418,060
1,015x2
720
395
92
98
533
1,000
206
272
69
95
400
150
4
7
11
4
3
7
3
3
4
4
4
7
7
7
4
4
yes
yes
yes
yes
yes
yes
no
yes
yes
yes
yes
yes
yes
no
yes
yes
N/A
N/A
partially
yes
no
yes
no
no
partially
yes
no
N/A
yes
no
no
partially
2017
2017
2016
2015
2014
2014
2014
2014
2014
2014
2013
2012
2012
2012
2011
2010
videos
images
images
images
images
images
images
images
images
images
images
images
images
videos
images
images
For all aging modelling tasks, the most widely used
benchmarks, and their respective aging databases, are FGNET and MORPH (alboum 2). The Face and Gesture recognition Network database (FG-NET) contains 1,002 images of
82 subjects, which are mostly Caucasians. The age of the
subjects ranges from newborns to 69 years old, while the
gap between the age of the same person in different images
ranges from 0 to 54 years. This dataset was collected from
personal photo albums and contains coordinates of facial
landmarks for each facial image. The MORPH dataset was
released in 2006 and its second album contains 55,134 images of 13,618 people. The subjects are mostly African and
there are 4 images of each person on average.
3.1.1 Data suitable for age estimation
The databases used for the task of age estimation vary
greatly in terms of sample size, number of subjects, and
age-range, as indicated in Table 1. Nevertheless, the recent
success of deep learning-based models in computer vision
has created a need for larger datasets with age annotations.
Towards this end, the IMDB-WIKI, AFAD and OUI-Adjience
datasets have been collected.
OUI-Adience [57]: The OUI-Adience dataset contains
26,580 facial images from the albums in the website
Flickr.com. These photos were made publicly available
through the Creative Commons license. The photos were
collected from about 200 albums, while the Viola-Jones [220]
detector was used to detect the faces. The age-range of the
dataset is 0-60+, although the ages above 48 years old are
less represented.
UvA-NEMO [49]: The UvA-NEMO dataset consists of
1,240 videos of 400 people smiling. The smiles are both
posed and spontaneous and were captured under controlled
conditions. The age of the faces varies from 8 to 76 years.
AFAD [161]: The Asian Face Age Dataset (AFAD) was
introduced by Niu et al. and contains 164,432 face images.
These images were collected from the RenRen Social Net-
4
work which is widely used by Asian students. Therefore,
most subjects are Asian and under 30 years old. The date of
birth of every subject was provided by the respective user
account.
IMDB-WIKI [184]: The largest age-annotated dataset is
the IMDB-WIKI dataset. The dataset consist of 523,051 facial
images that were crawled from the Wikipedia and IMDB
websites. All metadata were also collected from the above
mentioned web-sites. The average age of the dataset is about
32 years old.
The age annotations to all the datasets in Table 1 refer
to real age, and therefore cannot be used for the task
of apparent age estimation. In order to address the lack
of such datasets Escalera et al. [59] built the ChalearnAgeGuess dataset, which, to our knowledge, is the first
dataset with apparent age annotations. This dataset contains
4,691 images and was used for the Chalearn competition.
The annotations were acquired using the AgeGuess online
voting platform. Using the same platform, as well as Mechanical Turk workers, the APPA-REAL [3] dataset was
introduced by Agustsson et al. This dataset contains 7,591
faces with real and apparent age annotations in more than
7,000 images. The age range is between 0 and 95 years.
3.1.2 Datasets for Age Progression and Age-Invariant Facial Characterization
The tasks of age progression and age-invariant facial characterization require data containing the same person at
different ages. Therefore, the collection of such databases is
a challenging procedure. Besides the FG-NET and MORPH
datasets, the recently created AgeDB, VADANA and CACD
datasets are also widely used for age progression and ageinvariant facial characterization tasks. A brief overview of
these datasets is presented below.
AgeDB [155]: AgeDB contains 16,488 images of 568 subjects, manually collected from Google Images. Avoiding to
collect the images in a semi-automatic manner ensures that
the labels are not noisy. Each subject in the dataset has
29 photos on average, while the average age-range of the
subjects is 50.3 years.
VADANA [202]: The VADANA dataset contains 2,293
images from only 43 subjects (26 males and 17 females) that
are mostly South-Asian. It contains 3-300 images per subject
with wide variation in pose, illumination and occlusions,
while the maximum age gap between two images of the
same subject is 37 years. The annotations of this dataset
include among other kin relations and occlusions (facial hair
and glasses).
CACD [32]: The Cross-Age Celebrity Dataset (CACD)
was published in 2014 and contains 163,496 images from
2,000 celebrities. It was collected from Google Images using
information from IMDB. The dataset contains more than 80
images per subject on average, but the maximum age gap
between these images is 10 years. This is due to the fact that
the collected images were from 2004-2013. The subjects are
mostly Caucasian.
Table 1 reveals that age modelling has been primarily
studied by employing datasets captured in the wild. The
first major benchmark in the wild is FG-NET and traces
back to 2004. Ever since, the majority of publicly available
datasets are captured under unconstrained conditions, albeit
having a small number of images. Images from the FGNET or Morph2 datasets present limited pose and image
quality variation. These issues deteriorate the real-world
performance of age modelling systems, especially when
dealing with extreme poses or small and blurry images. The
recent advances in face detection technology [153], [252] has
resulted in much larger automatically collected datasets e.g.,
IMDB-WIKI, CACD, OUI-Adience. These datasets contain a
much larger number of images and present greater variation with regards to pose, expression and image quality.
The reported baseline performance for the OUI-Adience is
much lower (i.e., classification accuracy of 45.6%) than older
datasets, which is characteristic of the challenging nature of
the benchmark.
3.2 Datasets with kinship labels
Kinship verification from faces is a relatively new task,
aiming to verify whether the individuals in a pair of facial
visual inputs are blood related. There are eleven different
kinds of blood relations: (i) Father- Daughter, (ii) MotherDaughter, (iii) Father- Son, (iv) Mother- Son, (v) Sister- Sister,
(vi) Sister- Brother, (vii) Brother- Brother. (viii) GrandfatherGrandson, (ix) Grandmother- Grandson, (x) GrandfatherGranddaughter and (xi) Grandmother- Granddaughter. A
catalogue of dataset with kinship annotations is listed in
Table 2. The vast majority of the available data are annotated
in terms of only 4 kin relations.
CornellKin [65]: The first widely used dataset with kinship annotation is the CornellKin Database. The dataset
consists of 300 images and 150 kin pairs, gathered through
online searches.
UB KinFace [234], [193]: In 2011, Shao et al. published
the UB KinFace Ver 2.0 dataset, which was an extension to
UB KinFace Ver 1.0. The UB KinFace consists of 600 webcollected images and 400 kin pairs, half of which are Asian.
Every kin pair is represented by an image of the child, an
image of the parent at a young age and an image of the
parent at an older age.
IIITD Kinship Database [114]: The IIITD Kinship Database
was released in 2012 and contains 544 images of 272 kin
pairs in unconstrained environment. The dataset also includes 272 non-kin pairs. The faces are mainly Indian and
American, while other ethinicities include Asian and AfroAmerican. The dataset contains 7 kin relations.
Family 101 [64]: The Family 101 dataset was published
in 2013 and consists of 14,816 images of 607 people, collected from Amazon Turk workers. It is also the first one
to indicate the structure of the families. In particular, the
dataset contains whole family trees, each containing 1 to 7
nuclear families. There are 206 families, each having 3 to
9 family members. The subjects depicted in the data are
mostly Caucasian.
KinFaceW I & II [149]: The most widely used kinshipannotated datasets are the KinFaceW I & II albums that
were intoduced by Lu et al. in 2012. The KinFaceW I album
contains 1,066 images and 533 kin pairs, while the KinFaceW
II contains 2,000 images of 1,000 pairs. All the images are
collected from the web and the difference between the
two albums is that KinFaceW II contains kin-related faces
cropped from the same image.
5
HQ & LQ Faces [219]: In order to investigate the ability
to generalize from high-quality images captured under controlled condition to low-quality images gathered from the
Internet, Vieira et al. introduced the HQfaces and LQfaces
datasets. The HQfaces dataset contains frontal images of
92 sibling pairs captured by a professional photographer
in neutral expression. The LQfaces dataset consists of 98
sibling pairs collected from the Internet with varying resolution. The subjects in the two datasets are mainly Caucasian.
Sibling-Face & Group-Face [96]: In 2014, motivated by the
lack of large number of sibling pairs, Guo et al. introduced
the Sibling-Face dataset. The images were collected from
sites like Flickr, while the annotations were obtained from
the tags and descriptions of the images. Along with this
dataset, the Group-Face dataset was also introduced, containing images of groups of kin-related people.
TSKinface [175]: Qin et al. introduced the TSKinface
dataset, which consists of 787 images and 1015 family
groups. These groups contain the father, the mother and
child or children. The are 274 photos of Father-MotherDaughter families, 285 Father-Mother-Son families and 228
Father-Mother-Daughter-Son families. The majority of the
subjects included in this dataset are Asian.
Families in the Wild [182]: One of the most recent efforts
in kinship annotated datasets is the Families in the Wild
(FiW) dataset, containing 11,193 images. The FiW dataset
is the largest kinship annotated dataset to date, containing 30,725 face-images of 10,676 individuals. The dataset
includes 11,193 family photos, while it is the first to have
2,060 annotated pairs of grandparents and grandchildren.
WVU Kinship Database [115]: The WVU Kinship Database
was introduced in 2017 by Kohli et al. and contains 113 kinrelated pairs. The dataset allows for intra-class variation, as
it contains 4 images per person. The dataset is not balanced
and consists of: 22 Brother-Brother pairs, 9 Brother- Sister
pairs, 12 Sister -Sister pairs, 12 Father- Daughter pairs, 34
Father- Son pairs, 12 Mother- Daughter pairs and 8 MotherSon pairs.
Kinship Face Videos in the Wild Dataset [240]: The Kinship
Face Videos in the Wild (KFVW) dataset contains 418 face
videos from TV shows. Each video is 100-500 frames long
and presents great variation in pose, occlusions and expressions. The baseline experiments on the dataset achieved
poor performance and thus, the authors note the challenges
of advancing video-based kisnhip verification with this
dataset.
Similar to age related tasks, the problem of kinship modelling has been attacked using mainly in the wild datasets.
The nature of the labels makes the task of gathering and
annotating such datasets very laborious. As a result most of
the datasets in Table 2 contain a small number of images,
which is not ideal for modelling large variations. What is
more, these images, more often than not, originate from the
same photo. Cropping the data from the same image can
significantly bias the task of classification, by adding factors
like the environment, lighting, chrominance and the image
quality [143]. Therefore, datasets that contain such images
are considered biased and can not be used for the task of
kinship verification. This is evident when comparing the
state-of-the-art performance between classic benchmarks
like KinFaceW-II (i.e., classification accuracy 96.2%) and
more recent ones like FIW (i.e., classification accuracy 71%).
It is worth mentioning that, VADANA and UvA-NEMO
Smile Databases contain both kinship and age annotation.
In particular, for the Smile Dataset, the kinship annotations
were obtained only from the names of the subjects.
In the following sections, both seminal and recent methods for age progression, age estimation, age-invariant facial characterization, and kinship verification are reviewed.
Particular emphasis is put on the facial representations
employed. That is, for each of the aforementioned tasks,
the presented methods are grouped according to the type
of information captured by the employed facial representations. Broadly, geometric, hand-crafted, and learned facial
representations are considered. Discussion on corresponding classification and regression methods is also provided.
Furthermore, for each task evaluation protocols and metrics
are reviewed, and the most notable experimental results are
analyzed.
4
AGE ESTIMATION
A significant volume of research has been done in age
estimation from facial visual data. Since the labels of the
various age-annotated datasets are either discrete ages or
correspond to age-groups, the problem can be naturally
cast as a multiclass classification problem, where the classes
represent discrete ages or age-groups. However, neighbouring labels (ages, or age groups) might share important
information which is neglected by classification methods.
This is addressed by regression methods which appear
to perform better. Nevertheless, appearance changes more
rapidly during youth and slower in adults. To alleviate
this, non-stationary kernels can be employed; yet learning
with kernels is prone to overfitting. A different approach
to deal with this challenge is to adopt ranking methods
e.g., [28],[29], that learn an individual classifier for each
age class. In the following subsections several methods for
age estimation are reviewed. As already mentioned they
are organized according to the type of facial representation
employed.
4.1 Geometric facial representation-based methods
The tasks of automatic age estimation from facial images
introduced in 1994 by Kwon and Lobo [118]. Inspired by
anthropometric studies [5] that describe the growth of the
human head from infancy to adulthood, six facial distance
ratios are used to discriminate between infants and adults.
The adult faces are further classified into young adults and
older adults by using snakelets (i.e., deformable curves)
[110], which capture wrinkles on certain regions.
The use of such anthropometric models, that represent
the shape of the human head, introduces a number of
challenges for age estimation. For instance, geometric descriptors are sensitive to pose variations and can therefore
be applied only to frontal images. Several methods deal
with this challenge by normalizing the faces via Procrustes
analysis [112]. Invariant to small changes in camera location
and head pose geometric facial representation has been
proposed in [228] and [210] by employing the properties of
Grassmann manifold [56]. Age estimation is obtained via
6
Support Vector Machine (SVM)-based regression in [228]
and Relevance Vector Machine (RVM) [212] in [210].
Experimental results on the FG-NET dataset in [228] indicate that even thought geometric information is sufficient
for age estimation in young ages, texture information is
needed for accurate age estimation in adults. To alleviate
this, fusion of geometric and texture information (represented by Gabor phase patterns (HGPP)) is applied [250].
4.2
Appearance-based methods
Appearance models allow to capture texture information of
the face along with its shape. A wide variety of appearance
models have been proposed in the literature. Below, we
present those that have been extensively used in the context
of age estimation.
Active Appearance Models: Active Appearance Model
(AAM) is a generative facial model introduced in [40] by
Cootes et al. AAMs employ Principal Component Analysis
(PCA) to learn a linear model for shape and appearance
from images and a set of landmarks. This representation
was first used for age estimation in [123]. In order to capture
95 per cent of the data variation, 50 parameters of AAM
are used and age estimation is formulated as a regression
problem in the parameter space. Besides age estimation, the
effectiveness of the model in simulating the aging process is
evaluated on age progression as well as age-invariant face
recognition.
Most of the datasets in Table 1 are imbalanced; that is,
the labels of the data are sparse and not evenly distributed.
To overcome this problem, the age labels are encoded in
a cumulative manner in [37] using Cumulative attribute
(CA) vectors. If the face is older than the ith age, the ith
element of its CA vector is 1 and 0 otherwise. In this way,
data points with similar labels have a similar cumulative
attribute vector. The CA is obtained from the parameters
of AAM and Support Vector Regression (SVR) is performed
in the CA space. The experiments on FG-NET indicate that
shape features play a very import role in age estimation
of young faces, while texture features become increasingly
more discriminative after the age of 20.
In order to make use of the ordinal information of the
age labels, Chang et al. [28], [29] introduced a ranking
approach for the task of age estimation in 2010 using AAMs.
A ranking model gradually splits the feature space as a
binary classification problem is solved for each age label
k. The splits are produced by conducting a query ’is this
face older than age k?’. The final estimation is then produced
by a ranking rule based on the outputs of the classifiers.
Nevertheless, learning a large number of classifiers can be
computationally intensive. More recent ranking approaches
introduce AAM-based methodologies to reduce the computational cost, e.g., [138], [137]. In particular, a Partial Least
Squared-based Ranker, which learns all classifiers jointly, is
introduced in [138]. Also, a more time-efficient algorithm is
proposed in [137]. A Linear Canonical Correlation Analysisbased Ranker is employed and competitive results are obtained with a lower number of parameters. The smaller
complexity and model size make such approaches ideal for
real-time applications.
In the above papers the effectiveness of the texture and
shape parameters of AAM as a facial representation for the
task of age estimation have been demonstrated. However,
AAM uses PCA and can therefore only account for linear
modes of variation, which introduces a set of challenges
when generalizing to unseen data. In order to capture nonlinear variations such as expressions and poses, Duong et
al. [54] introduced the Deep Appearance Model (DAM).
This generative model is based on Deep Boltzman Machines
[188], a hierarchical method that can capture variations in
shape and texture that could be higher than second order.
Experiments in the FG-NET dataset indicate the ability of
DAM to outperform AAM, while being more robust to
additive noise. Nonetheless, having more parameters, this
method is more resource intensive in comparison to traditional AAM. Deep hierarchical representations are analyzed
further in subsection 4.4.
Local Binary Pattern: The Local Binary Patterns Histogram (LBP) [164] is an effective texture descriptor which
has been widely employed for texture classification and
face recognition [4]. LBPs are binary codes computed in
neighbours of pixels and their histogram is obtained for
different image-patches. Some variations of the descriptor
[227] compare the values of three (Three Patch-LBP) or
four (Four Patch-LBP) patches to produce a code for each
pixel. Yang et al. [245] employed this facial representation
to classify faces into three age-classes. In particular, LBP
Histograms are extracted from a large set of possible face
patches and AdaBoost [190] is applied next to choose the
most discriminative ones. Several different regression methods are applied to the extracted representations, exhibiting
promising results. In [57], fusion of LBP and FP-LBP [227]
is exploited in age group and gender classification via SVM
classifier.
Wavelets-based Features: Wavelets-based features have
been widely used in facial analysis, due to their ability
to robustly capture texture information. Among the most
widely used types of wavelets in age estimation are the
Haar wavelets [151], yielding robust to appearance variation
features. Haar-like features are employed for the task of age
estimation in [258]. Age estimation is posed as a regression
problem with a regularized L2 loss function using boosting.
Gabor wavelets have also been widely applied in texture
analysis [132]. Importantly, these wavelets have a biological
significance, since their kernels are similar to the receptive
field of the mammalian cortical simple cells. The magnitude
of the coefficients, which is obtained by convolving facial
images with Gabor wavelets across different scales and
orientations, is used as facial representation in [78]. Agegroup classification is performed using Fuzzy LDA. Experiments on a private dataset indicate improved performance
compared to that obtained by employing LBPs.
To capture curvature information, like wrinkles, extended curvature Gabor (ECG) filters are used in [113].
After feature selection, Random Forest (RF) Regression is
used for age estimation. Another interesting wavelet based
representation is the scattering transform (ST) [150]. ST is
calculated by cascading wavelet modulus operators along
different paths in a deep convolution network. A number of
recent works use Gabor wavelets with ST for age estimation.
A ranking approach using ST representations is introduced
in [27]. The experiments reveal that AAMs performs better
than ST in FG-NET dataset, while the opposite stands for
7
the MORPH2 dataset. This is attributed to the fact that, FGNET has significantly more images per person and fewer
people than MORPH2, capturing long-term person-specific
age information. Hence, the AAM can efficiently model
the intra-person variance, while the ST appears to model
person-invariant age information better.
Biologically Inspired Features: Besides the biologically relevant Gabor wavelets, the Biologically Inspired Features
(BIF) [91] have also been proposed as suitable facial representation for age estimation. Inspired by the HMAX model
[181], the BIF pipeline consists of simple (S1) and complex
(C1) layers imitating the classical Hubel and Wiesel model
[106] for the primary virtual cortex. Gabor filters are used
for S1 units, while the C1 units pool over the S1 units
with a non-linear maximum operator ’MAX’ and normalizes
with a standard deviation operator ”STD”. It is argued that
the ”STD” operator reveals local variations capturing vital
aging information, like wrinkles and eyelid bags. Experiment are conducted on the FG-NET and YGA datasets
using SVM classification and SVR regression. Experimental
results indicate that classification accuracy drops on small
unbalanced datasets like the FG-NET.
A variety of different regression methods have been
proposed using the BIF descriptor. In particular, a non-linear
extension of Partial Least Squares (PLS) [79] is introduced
in [90]. Kernel PLS (KPLS) regression is used on BIF feature
space for simultaneous dimensionality reduction and function learning. Another BIF-based method, particularly an
age estimator that is robust to pose variation, is introduced
in [203]. The multi-view property of this method is obtained
by using video information in a semi-supervised manner.
That is, a regression problem is solved with a regularization factor that enforces output consistency throughout an
unlabelled video sequence.
A fusion of classifiers and regressors is used in [97],
where an automatic demographic estimator is introduced.
A hierarchical fusion of SVM classifiers and SVR regression
is used for age estimation. That is, a face is passed through
a series of binary classifiers, indicating different age groups
in a coarse-to-fine manner, before going through an SVR
regressor. A comparative study between the automated system and human workers is also presented in [97], revealing
the inability of the humans to perform accurate real age
estimation, as well as the challenges of crowd-sourcing such
efforts.
A set of regression methods that take into account the
correlation between age classes is proposed in [81], [80].
Based on the fact that aging is a slow and smooth process,
the information extracted from a face does not only describe
the exact age of the subject, but also the neighbouring ages.
Therefore, based on that assumption the model assigns a
label distribution to the data p(y|x), instead of the exact label.
For that purpose two label distribution learning algorithms,
namely IIS-LLD [81] and CPNN [81], are proposed and
tested in FG-NET and MORPH2 datasets, using AAMs and
BIFs respectively.
4.3
Subspace learning-based methods
The aforementioned facial representations are extracted
from each and every image individually. Nevertheless, the
images of the same person at different ages are not uncorrelated and form a pattern to be recognized. To investigate
this concept, the AGing pattErn Subspace (AGES) method
is introduced in [82]. In order to represent the transformations caused by aging, aging patterns are employed. These
patterns are obtained as sequences of face images sorted
in time order. Each individual in the training set has a
different aging pattern, which is obtained from the training
images and by solving a missing value problem for the missing ages. By applying PCA on the resulting aging pattern
vectors, the so-called Aging Pattern Subspace is obtained.
Each point on this subspace represents an aging pattern.
During testing, an unseen face is projected on the aging
pattern subspace and an aging pattern is selected based on
the reconstruction error. That is, the input face is projected
on all the possible positions in each aging pattern and
the generated face is compared to the input. The selected
position on the selected aging pattern is then returned as
the estimated age. Instead of pixel intensities, the method
uses low-level face representation extracted from a different
model, in particular the AAM.
One of the shortcoming of the AGES method is that it
requires a large number of images of the same person at
different ages to obtain the aging patterns. In practice, most
datasets are very sparse, containing only a few images per
person that do not span a large range of ages. To overcome
this challenge, Fu et al. [74] uses face images from all the
people in the training set to obtain the aging pattern. In
that way, the common underlying aging pattern of the data
is captured. By representing each age label with multiple
images, a discriminative manifold [192] is obtained in a supervised manner. During testing, the input face is embedded
on this low dimensional subspace and a regression problem
is solved. Several methods are applied to visualize the
geometry of the manifold, e.g., Locality Preserving Projections (LPP) [99], Orthogonal Locality Preserving Projections
(OLPP) [25] and Conformal Embedding Analysis (CEA)
[75], with the results indicating a distinct pattern. This is
contrary to results produced by unsupervised methods, e.g.,
PCA, that do not take into account the age information. Different regression methods have been applied on the aging
manifold. Particularly, simple regression methods, i.e. linear,
quadratic and cubic, are used in [74]. A more sophisticated
method is proposed in [88], where Locally Adjusted Robust
Regression (LARR) is introduced.
The aging manifold on the aforementioned methods
spans the image space. That is, pixel intensities are used
as local descriptors. This can deteriorate the discriminative
ability of the model, since the image space represents all
variations and is not particularly age-sensitive. In order to
incorporate the texture information in the aging manifold,
LBP are used as low-level representations in [63] where the
age is obtained using regression methods, e.g., a Neural Networks (NN) and Quadratic Funtion (QF). Similarly, aging
manifold can be obtained from different descriptors, e.g.,
Gabor [127] and BIF [126]. Particularly, an aging manifold
that preserves the ordinal information as well as the geometrical structure of the data is obtained in [127], [126]. The
resulting subspace is well-suited for ranking methods and
therefore the OHRank is employed for age estimation.
Most Manifold Learning algorithms assume that the fea-
8
ture space is locally Euclidean and therefore use a Euclidean
metric to determine neighbourhoods. In the aging manifod,
this is not always the case, due to the non-linear nature
of aging. Therefore, a distance metric adjustment to LPP is
introduced in [30]. The new metric is learned using Labelsensitive Relevant Component Analysis (RCA) [9]. Ordinal
information is also incorporated in the aging manifold and
k-Nearest Neighbours (kNN) and SVR are used for age
estimation on the embedded subspace. Similarly, a subspace
with ordinal information is captured in [39]. Pairwise age
ranking is utilized to obtain a subspace that minimizes the
distance between images of the same label, under a set of
ranking constraints.
4.4
Deep learning-based methods
Among the different sources of facial variation, the nonlinear transformations, such as deformations due to expressions, pose and age, are the most challenging to model. To
model such non-linear variations, some of the aforementioned methods extract simpler features at different levels in
a hierarchical manner. Since deep architectures build nonlinearities on top of each other, they can learn non-linear
transformations more efficiently than simple models and
prove to perform better in most tasks, albeit being more
vulnerable to overfitting. A classical deep architecture is
the feed-forward Artificial Neural Network (ANN). This
model uses pixel intensities (PI) or low-level descriptors
as inputs and builds a subspace in a supervised manner
using backpropagation. A variation of ANN, namely the
Compositional Pattern Producing Network (CPPN) [204],
is used in [62] for the task of age estimation. Contrary to
typical ANNs, the four-layer ANN proposed in [62] has
different transfer functions for each neuron.
Similar to the ANNs, Convolutional Neural Networks
(CNN) [125] are used to extract high-level features, or perform regression or classification as an end-to-end model.
Instead of using handcrafted features, a CNN applies filters
on the raw images in a hierarchical supervised manner.
Convolution and subsampling are applied iteratively to
create each layer’s feature map. The filters are then optimized using backpropagation, while the architectures used
can vary greatly. Deep CNNs have outperformed classical
methods in most tasks, including age estimation. Yang et al.
[244] is the first to use a 5-layer CNN for age estimation. The
model has multiple outputs for gender, age and race, while
the age estimation performance is worse than that obtained
by BIF. In order to improve the performance, Yi et al.
[246] adopts methodologies from traditional facial analysis
and used a fully connected ensemble of 23 CNNs. Each
CNN is applied on different face patches and the method
succeeds in improving the state-of-the-art performance on
the MORPH 2 dataset.
Recently, the large availability of data and computational
power have boosted the popularity of deeper and more
complex architectures (e.g., ResNet [98], DenseNet [104]).
A hybrid deep multi-task CNN is employed in [238] to
predict age, race and gender. This architecture incorporates
the gender and race predictions by training a different age
estimator for each group. Another deep CNN architecture
that leverages weakly labelled data is proposed in [103]. The
proposed model performs age estimation with assistance
of age difference information (AEAD). The age difference
information is obtained from image pairs of the same subject
with known age gap.
Deep ranking methods have found great success in age
estimation. A deep ordinal regressor is trained on the very
large AFAD dataset in [161]. This ranking approach uses
multiple output CNNs (MO-CNN) and outperforms the
classical ranking algorithms on the MORPH2 dataset. Similarly, a Ranking-CNN approach, that utilizes multiple deep
binary classifiers, is proposed in [38]. This method employs
an ensemble of CNNs that are fused with aggregation. The
inconsistency of the binary outputs is proved not to affect
the overall performance as the output error is bounded by
the maximum error of the binary classifiers [38]. Therefore,
as long as the maximum error is decreased, the inconsistent
labels of the classifiers do not matter.
A different age ranker for separate age-groups is
adopted in [134] to learn a series of aging patterns. A
deep CNN architecture is employed to minimize the distance of faces within the same age-group, while maximizing
the distance between faces from different groups. The age
prediction is then acquired using the OHRank algorithm
for each age-group. To effectively maximize the distance
between age groups, label-sensitive deep metric learning
(LSDML) is introduced [136]. This method optimizes the
procedure by jointly learning a discriminative metric and
mining hard negative pairs. In order to mitigate the effect
of sparse and unbalanced datasets, the method is extended
to multisource LSDML that maximizes the cross population
correlation between different datasets.
A large number of deep methods are applied to apparent
age estimation, mainly hosted by the Chalearn Looking
at People (LAP) challenge [59], [60]. In particular, a deep
network consisting of 10 convolutional layers, 5 pooling
layer and 1 fully connected layer [36] is used in [179]. For the
aging function, a test image is classified into 3 age groups
and a 3-layer ANN regresses the apparent age. Each of the
regressors is trained on a different dataset.
The highest positions in the LAP challenge are populated by variation of popular deep learning architectures.
Such variations include ensembles of multiple networks and
fusion of different models. The most widely adopted ones
are the VGG-16 and GoogLeNet [208] Deep CNN, due to
their success in the Imagenet object recognition challenge
[185]. Modified versions of the original VGG-16 model are
used for apparent age estimation in [184], [117], [7] and
[216]. These models are trained and finetuned on several
different datasets to achieve significant generalization, while
a variety of classification and regression techniques are
employed on the deep representations. Particularly, an SVM
based classifier is used in [216] and improves the results
of the regression method used in [184]. Lastly, a fusion of
GoogLeNet-based classifier and regressor is used in [141]
and hierarchical SVM age grouping followed by SVR and
RF regression on the deep features is employed in [260].
4.5 Other representations
Besides the above facial representation, several other approaches have been proposed. Such methodologies either
9
utilize multiple descriptors, e.g., [11], [33], [67]; that is,
multiple features are fused to capture more discriminative
information, or introduce novel extensions to these representations, e.g., [148], [222]. In particular, the Neighborhood
Centroid Difference Vector (NCDV) from the LBP descriptor
is used in [148] and the Cost-Sensitive Local Binary Features
(CS-LBFL) are introduced. A multi-feature learning extension is also proposed and the final prediction is obtained
using the OHRank algorithm.
Most descriptors in this section take into account agesensitive information, e.g., wrinkles, smoothness, bags under eyes, facial hair, etc. Since many age groups share the
same attributes, such information is not always accurate
and can lead to misclassifications. Nevertheless the presence
of these attributes is different between the age groups. In
[222], a Learning Using Privileged Information (LUPI) [217]
approach is proposed, so that such inaccuracies are used
to boost the model’s generalization ability. That is, for each
attribute, the age groups are ordered based on the presence
of the specific attribute. The problem is solved using relative
attribute SVM+ (raSVM+).
In order to obtain a discriminative mapping of the
feature space, a number of the above methods have introduced hierarchical features. The notion of hierarchical
representations is not limited to deep learning methods, but
is generally used in representations that accumulate information in multiple layers. A simple and efficient method to
extract hierarchical features for age estimation is presented
in [116]. The method, namely MidFea, consists of k-means,
convolution, max-pooling, vector quantization and random
projection operations. The MidLevel features are then used
as input to a Neuron Selectivity layer (NS) [15] to obtain the
final representation.
All of the above representations originate from 2Dimensional images; that is, they are either extracted from
pixel intensities or facial landmarks. Nevertheless, other
modalities, e.g., 3D [232], video [48], context [77], contain
age information and a number of methodologies are proposed to capture it. In particular, Gallagher et al. [77] assign
contextual features that capture local pairwise information
and the global position of the person. A Gaussian Maximum
Likelihood (GML) classifier is used to classify the faces into 7
age groups. Interestingly, the contextual features achieve an
accuracy of more than double random chance, while gender
is correctly estimated two-thirds of the time.
An interesting combination of different modalitites is
that of appearance with facial dynamics. This is investigated
in [48], where temporal face dynamics are extracted from
smile and disgust videos. The videos that are used include posed and spontaneous expression and were recorded
under controlled conditions. The experiments in [48] reveal that dynamic features are not as discriminative as
appearance features for age estimation. Towards that end,
several appearance estimators are used, i.e. dynamic, intensity based encoded features (IEF), gradient-based encoded features (GEF), BIF and LBP, while feature selection
is performed using the Min-Redundancy Max-Relevance
(mRMR) algorithm. The combination of the appearance and
dynamic features achieves the best results, indicating that
dynamics can help significantly towards age estimation.
4.6 Evaluation protocols, metrics, and results for age
estimation
Protocols: A number of different evaluation protocols have
been used to evaluate Age Estimation models. Classical
methods like k-fold cross-validation (c-v) and 80-20 split
have been used in this context as well. Contrary to c-v, when
applying an 80-20 split, a predefined 80% of the data is used
for training and 20% is used for testing, without reshuffling
the data. Besides these widely adopted protocols, specific
protocols have been proposed for specific datasets.
One of the most widely used protocols is the LeaveOne-Person-Out (LOPO) protocol, proposed for the FGNET
dataset. According to this protocol, experimental evaluation
is performed using images of previously unseen individuals. Therefore, training is done using all subjects in the
database apart from the subject whose age we are estimating.
Other dataset-specific protocols include the Images of
Groups (Groups) protocol and the Chalearn apparent age
dataset protocol (chalearn). The data split for the Groups
is a random selection of 3,500 training images and 1,050
testing images. The age group classification accuracy (acc.)
is calculated both for exact match (AEM) as well as for
allowing error of one category (AEO). On the other hand,
the Chalearn dataset is split into 2,476 images for training,
1,136 images for validation and 1087 images for testing.
Metrics: The large number of methods in age estimation
has exposed the need for a common evaluation protocol.
The most widely adopted measures are the Mean Absolute
Error (MAE) and the Cumulative Score (CS). The MAE is
defined as the average of the distance between the predicted
age labels and the ground truth,
M AE =
N
X
|ȳ − y|/M.
(1)
=1
On the other hand, CS(j ) is the percentage of the samples,
where the predicted age of model did not deviate from the
ground truth more than j years,
CS(j) = Ne≤j /N × 100%.
(2)
Lastly, specifically for the apparent age estimation models
[59], the adopted measure is the error calculated as follows:
error = 1 − e−
(y−yi )2
2σ2
.
(3)
Results: Table 3 includes the methods that have reported
the highest scores for each dataset in this subsection, using
comparable and widely used protocols. A chronological
overview of the most important methods and result on Age
Estimation is tabulated in Table 4.
4.7 Discussion on age estimation
There is a number of factors affecting the accuracy of age
estimation algorithms. The choice of facial representation
plays a vital role in this, with the advantages of each approach having been described above. Besides that, the types
of variation within each dataset determines the success of
the facial representation and the algorithm. Such variations
include image quality, race, gender and facial expressions.
10
TABLE 3: Best reported results on Age Estimation for each dataset
Dataset
method
metric
protocol
score
FG-NET
MORPH2
Groups
Adience
AFAD
Chalearn
UvA-NEMO
[135]
[135]
[127]
[135]
[161]
[7]
[48]
MAE
MAE
AEM/AEO acc.
acc.
MAE,CS(≤ 10)
error
MAE
LOPO
80-20
groups
5-fold c-v
80-20
chalearn
10-fold c-v
3.74
2.89
48.5% , 88%
60.2% ± 5.3
3.34, ∼ 95%
0.2411
4.33
An experimental study of age estimation under changes in
image quality was conducted in [6]. Image quality affects
texture information that captures aging cues, like wrinkles.
Therefore, the results confirmed the deterioration of the
accuracy with declining image quality.
People of different gender do not follow the same aging
patterns. A similar assumptions can be made for people
of different race, eg. Caucasian and Asian. Several popular
datasets, including MORPH2, contain faces from people of
different gender and race. Such non linear variations affect
the accuracy of the facial representations. The studies in [89]
, [95] and [14] indicate this problem and propose transfer
learning and domain adaptation solutions to deal with it.
A similar problem arises when performing age estimation
across different expressions. The quantitative study in [94]
indicates the significant influence of facial expressions on
age estimation, while a cross-expression age estimation
framework is proposed.
In total, different datasets have different modes of variation, since they were collected under different circumstances. That is, image quality, illumination, race, gender
and age among the data can vary significantly. Similar to
the above, using the tools of transfer learning, Su et al.
represented different datasets as different domains and a
cross-database age estimation framework is proposed in
[205].
4.8
Challenges in the wild
Most of the methods presented in this subsection are evaluated on in the wild datasets. Nevertheless, as mentioned
in section 3, these datasets pose different levels of challenges. Table 3 indicates that datasets like Adience and
Groups are indeed more challenging, having great variation
regarding face size, image quality, make-up and occlusions.
Particularly, the median face from the Groups dataset has
18.5 pixels between the eye centers, while the Adience
dataset demonstrates extreme variation in image quality.
On the other hand, FG-NET and MORPH2 contain family
photos and mugshots respectively. While the image quality in the first one varies greatly, neither of them contain
extreme poses. It is therefore expected, that the state-of-theart results for these benchmarks are better compared to the
Adience and Groups benchmarks. Lastly, the best score is
achieved with the MORPH2 dataset, which can be justified
by the fact that the faces in the mugshots are frontal.
5
AGE PROGRESSION
Face synthesis of an individual at different age groups is a
task that has been studied for several years across computer
graphics, anthropometry and computer vision. Although
earlier methods used face models introduced in computer
graphics, recent work has gradually transitioned to computer vision. One of the seminal works is by Burt et al.
[43], which simulates the aging process by employing both
colour and shape information. In this approach, the synthesized image is produced by adding the difference between
the average faces (i.e. the prototypes) of each age group to
the test image. Other early works include anthropometric
growth [176], [177], modelling of wrinkles [230], [231], [229],
[211], aging functions [122] and caricaturing 3D face models
[167], [166]. Notably, not all of the aforementioned earlier
methods produce photorealistic results. More recently proposed methodologies in age synthesis are presented in the
following paragraphs.
5.1 Subspace Learning based Methods
The tools provided by subspace and factor analysis have
been used to model the transformations caused by aging.
A prototype based approach that compensates for illumination variation in the prototypes is introduced in [111]. In
this Illumination Invariant Age Progression (IAAP) method,
the prototype for each age cluster is obtained as a rank-4
approximation using Singular Value Decomposition (SVD)
and are relighted according to the input face. Then the
input face is rendered to a different age by applying the
difference in flow and texture between the source and target
prototypes. The results are evaluated in a large scale user
study on Mechanical Turk, the result of which indicates the
inability of the workers to recognize a face across large age
differences.
In prototype-based approaches (e.g., [43], [111]), the texture transformation is modelled as the difference between
the prototypes of the source and target age groups. That
is, the texture changes are the same for different inputs,
as long as the source and target age are the same. Instead
of a specific prototype, a dictionary of aging patterns for
each age group is learned in [197]. To capture the different
factors of variation, an input face is decomposed into an
aging and a personalized layer. The changes in the aging
layer is modelled using the dictionaries of the target and
the source age groups. A more general approach that is
robust to different kinds of variation, namely Robust Age
Progression (RAP), is introduced in [186]. Each image of
a specified gender is expressed as a superposition of the
age component and the common component. The common
component captures facial variations such as identity, shape,
pose, occlusions, illumination and expressions, while the
age information is captured by the age component. Thus,
by computing an orthonormal bases of the age and common
11
Year
Paper Representation
Method
Dataset(Score)
Metric
1994
[118]
ratios,wrinkles
hand crafted rule
private(100%)
accuracy
2002
[123]
AAM
regression
private(1.88)
MAE
2005
[258]
Haar
regression, boosting
FG-NET(5.81)
MAE
2007
[245]
LBP
Rule-based, Adaboost
2007
[82]
AAM
LDA, AGES
FERET(7.88%)
PIE(12.5%)
FG-NET(6.22)
MORPH(8.07)
error rate
error rate
MAE
MAE
2008
[74]
PI
CEA, multilinear regression
2008
[88]
AAM
OLPP, LAR
YGA-fem(8)
YGA-male(7.8)
FG-NET(5.07)
YGA-mal(5.30)
YGA-fem(5.25)
MAE
MAE
MAE
MAE
MAE
2009
2009
2009
[77]
[78]
[91]
contextual
Gabor
BIF
GML
Fuzzy LDA
PCA,SVM/SVR
Groups(32.9%, 64.4%)
private(91%)
FG-NET(4.77)
MORPH-fem(3.47)
MORPH-mal(3.91)
AEM, AEO
accuracy
MAE
MAE
MAE
2010
[28]
AAM
Threshold ranker
2010
[254]
AAM
Multi-task warped GP
FG-NET(4.67, ∼ 68%)
MORPH(6.49, ∼ 50%)
FG-NET(4.14)
MORPH(4.07)
MAE,CS(5)
MAE,CS(5)
MAE
MAE
2011
[29]
AAM
OHRank
2011
2011
2011
[90]
[203]
[160]
BIF
BIF
BIF
Kernel PLS
multiview regression
RMIR
2011
2011
[62]
[244]
AAM
PI
CPPN, kNN
5-layer CPNN
FG-NET(4.48, 74.7%)
MORPH (5.88, 56.5%)
MORPH(4.18)
Custom(6.94 ±0.07)
FG-NET(8.37)
MORPH(6.06)
FG-NET(4.67)
FG-NET(4.88)
MAE,CS(5)
MAE,CS(5)
MAE
MAE
MAE
MAE
MAE
MAE
2012
2012
2012
[228]
[210]
[127]
geometric, HGPP
geometric HGPP
Gabor , contextual
SVM, PLS
RVM
PLO, OHRank
FG-NET(6.76)
FG-NET(6.2)
FG-NET(4.82)
Groups(48.5%, 88%)
MAE
MAE
MAE
AEM, AEO
2013
[81]
AAM , BIF
MFA, CPNN/IIS-LLD
2013
[37]
AAM
CA, SVR
2013
2013
[30]
[39]
AAM
AAM
LPP, lsRCA, kNN, SVR
ranking SVR
FG-NET(4.76)
MORPH(4.87 ±0.31)
FG-NET(4.67, 74.5%)
MORPH(5.88, 57.9%)
FG-NET(4.38)
FG-NET(4.56)
MORPH1(5.41)
MORPH2-Cau(4.42)
MAE
MAE
MAE,CS(5)
MAE,CS(5)
MAE
MAE
MAE
MAE
2014
[57]
LBP, FPLBP
Dropout-SVM
2014
[246]
PI
CNN
Adience(45.1% ±2.6)
Groups(66.6% ±0.7)
MORPH(3.64)
accuracy
accuracy
MAE
2015
2015
[116]
[48]
MidFea-NS, SVM
mRMR, SVM, SVR
FG-NET(4.73)
UvA-NEMO(4.33 ±4.06)
MAE
MAE
2015
[27]
PI
dynamic, IEF, GEF,
BIF, LBP
ST
CSOHRank
2015
2015
[54]
[126]
PI, landmarks
BIF
DBM
PLO, OHRank
2015
[97]
BIF
Adaboost, SVM,SVR
FG-NET(4.70, ∼ 75%)
MORPH(3.82, ∼ 78%)
FG-NET(5.28)
FG-NET(1.306)
Groups(0.864)
FACES-best(5.16)
FG-NET(3.8 ±4.2)
MORPH(3.5 ±3.0)
PCSO(4.1 ±3.3)
MAE,CS(5)
MAE,CS(5)
MAE
MAE
MAE
MAE
MAE
MAE
MAE
2016
[222]
DSIFT, LUPI
raSVM+, OHRank
2016
[161]
PI
MOCNN
2016
2016
[184]
[7]
PI
PI
DEX
ensemble DEX
FG-NET(4.07)
MORPH(5.05±0.11)
MORPH(3.27)
AFAD(3.34)
Chalearn(3.221, 0.2649)
Chalearn(0.2411)
MAE
MAE
MAE
MAE
MAE, error
error
2017
2017
[38]
[136]
PI
PI
Ranking-CNN
M-LSDML
2017
[103]
PI
AEAD
MORPH(2.96, ∼ 85%)
MORPH(2.89)
Adience(60.2±5.3)
FG-NET(3.74)
FACES(3.11 - 5.01)
Chalearn(0.315)
MORPH(2.78)
FG-NET(2.8)
MAE,CS(5)
MAE
accuracy
MAE
MAE
error
MAE
MAE
TABLE 4: Overview of Age Estimation methods
12
components via SVD, an image can be progressed to another
age group as a linear combinations of these bases at the kth
age group.
In a similar manner, Hidden Factor Analysis (HFA) is
used in [243] to decompose the facial input. Thus, the face is
decomposed into a linear combination of the mean face and
the identity, age and noise factors. The age component at a
different age is then sparsely represented by a dictionary of
age components of the same age group. The shape is progressed by applying the difference between the mean shapes
of the target and source age clusters. In order to disentangle
the common and individual components, a robust extension
to Joint and Individual Variation Explained [142] is used in
[187]. Thus, by modifying these components, an input face
can be reconstructed at a different age group. The method
is evaluated on the FG-NET and AgeDB datasets, while
experiments are conducted on facial expression synthesis
as well.
To deal with the different modes of variation, instead of
a matrix, a tensor is used to represent the data in [225]. The
different modalities of the data, namely the pixels intensities, identities and ages of the face images, are included
along the the dimensions of the tensor. This means that
along each axis of the 3-dimensional structure, the variation
of only one modality changes. Based on the fact that low
frequency image components preserve the face identity,
super-resolution in tensor space is used to map the texture
of a downsampled input image to a different age.
5.2
Sequence Modelling
Since the effects of aging are temporally correlated, i.e. the
transformation is smooth and continuous, it is intuitive
to model it as a sequence across across the different age
groups. In order to describe the evolution of the facial
representation through the age groups, different sequence
modelling methods have been proposed, including Markov
Process [207], [206] and Recurrent Neural Networks (RNN)
[223]. In particular, the transformation of different facial
parts is modelled separately in [206]. Instead of using AAM,
a region based AAM model (RB-AAM) is adopted. The relationships between the different sub-regions are modelled
based on the physical structure of the face. To approximate
the short-term aging of each sub-region, an aging function
approach is used. The aging model over a long period is
formulated as a Markov Process by concatenating the short
period aging functions under smoothness and consistency
constraints. On the other hand, Wang et al. [223] use RNN
[87] to model age progression between neighbouring age
groups. The faces of the neighbouring groups are normalized jointly. An RNN is adopted to perform aging of a face
in the shared eigenface space [214] of two neighbouring age
groups. The output of the RNN is reconstructed, projected
on the shared eigenface space of the next group and used as
input for the next RNN.
5.3
3-Dimensional Representations
The aforementioned methods come with certain limitations
since they are based on 2-dimensional representations. Contrary, 3D facial representations capture both shape and texture information and can potentially obtain pose invariance.
A 3-Dimensional model is used to build an age progression
system for children faces [196]. Different facial components,
such as mouth, eyes, nose and face shape, are extracted and
progressed individually. The basic assumption is that if two
children look similar at a young age, they will continue to
look similar when they grow older. Therefore, each component is compared to a database of facial components and is
progressed according to the corresponding aging pattern.
The selected components are then merged to synthesize
the progressed image. The experimental results on selected
images of the Jackson family indicate the possibility of
synthesizing age progressed faces of children given the faces
of their relatives.
A different approach that incorporates the whole face in
a 3D aging model is proposed in [170]. The aging model
is approximated as a weighted average of all the texture
and shape aging patterns in the training set. The model
is evaluated using a commercial face recognition system
as described in Section 6. Similarly, texture and shape are
modelled separately in [152]. The face shape of each age
group is modelled using age specific 3D Point Distribution
Model (3DPDM) [41], while the texture is modelled using
recursive PCA.
5.4 Deep learning-based representations
Synthesizing a photorealistic face image with arbitrary
modes of variation is a challenging task. Deep learning
methods are able to incorporate knowledge from multiple datasets and are therefore suited to deal with this
complicated problem. In particular, Generative Adversarial Networks (GANs)[85] have proven capable to produce
realistic images and have been successfully applied to face
synthesis, e.g., [139], [256]. The original model consists of
two networks, a Generator and a Discriminator, that are
trained simultaneously. The Generator tries to model the
data distribution and synthesizes images which the Discriminator classifies as ’real’ or ’fake’, that is whether they
come from actual data distribution. The image generation
is conditioned on random input noise z, which follows a
predefined distribution. The optimization procedure can be
considered as a minmax game between the two models, as
the parameters of each network are optimized alternatively.
In some variations of the model, particularly in the Conditional GAN (CGAN) , the generated images are conditioned
on a specific input, rather than random noise.
A Conditional Adversarial Autoencoder (CAAE) for age
progression is introduced in [255]. Instead of randomly
sampling z, it is obtained from an autoencoder, so that incorporates the personality of the face. The model includes two
discriminators, one to enforce p(z) to be uniform; that is, to
force z to evenly populate the latent space, and one to force
the generator to produce realistic images. The generation
of the image is conditioned on both the personality of the
input, as well as the desired age group. The method is evaluated based on survey result, that indicate the ground truth.
A conditional GAN (Age-cGAN) is also employed in [8] to
perform face aging. The model obtains the embeddings of
the input and ouput of the GAN from a face recognition
neural network. In order to preserve the identity of the
input face, the Euclidean distance between the embeddings
is minimized.
13
5.5 Evaluation protocols, metrics, and results for age
progression
age difference of 10 years. A chronological overview of the
methods in this subsection is tabulated in Table 7.
Protocols: There is a wide variety of different protocols
in the literature regarding the evaluation of age progression models, since not all methods focus on the same age
groups, e.g., some focus on children faces, while others
focus on adult faces. That being said, all the methods in
this subsection use the FG-NET dataset to test their methods
qualitatively or quantitatively.
Metrics: The evaluation of age progression methods is
nontrivial, since the task itself is ill-posed and the actual
aging process is highly uncertain. A comparative study of
different age progression evaluation techniques is presented
in [120], while a framework for evaluation is described in
[124]. A short description of the most common evaluation
methods is presented as follows:
a. Human-based evaluation: Directly comparing the synthesized images with the ground truth is not an effective
measure of accuracy, as the images were taken under different sources of variation. A way to evaluate the results of
age progression algorithms is by conducting a study with
human users. The results can then be evaluated based on
the users’ answers to questions comparing the synthesized
image and the ground truth at the target age.
b. Age estimation accuracy: Age progression models are
also evaluated based on their ability to generate faces with
characteristics of the target age group. The accuracy of the
synthesis can be measured in an automated manner, using
an age estimator and the corresponding metrics as described
in Section 4.
c. Preservation of face identity: An age progression method
is evaluated based on whether the synthesized images
can still be recognized as images of the same person. To
determine that, face recognition and verification systems
have been widely used. In the context of age progression
evaluation, common metrics include classification accuracy
(rank n), the ROC curve and the area under it (AUC) as well
as Equal Error Rate (EER). A more detailed description of
Age-Invariant Face Recognition (AIFR) and Cross-Age Face
Verification (CAFV) methods and the corresponding metrics
is included in Section 6.
Results: While the methods in this subsection have
been evaluated on different datasets, we will only use the
common results on FG-NET for comparison. Based on their
quantitative results on FG-NET, the most significant results
are presented in Table 5. The results for preservation of
identity are reported before (score-b) Age Progression as
well as after (score-a) wherever available.
TABLE 6: Best reported results on Age Progression based on
preservation of identity on other datasets, score-B and scoreA correspond to face verification score before and after age
progression
TABLE 5: Best reported results on Age Progression based on
preservation of identity on FG-NET
method
test
metric
score-b
score-a
[197]
[186]
[225]
[223]
CAFV
CAFV
AIFR
CAFV
EER
acc., AUC
accuracy
EER
14.89%
N/A
0.55
∼ 15%
8.53%
0.709, 0.806
0.7
∼ 10%
dataset
meth.
test
metric
score-B
score-A
AgeDB
MORPH
CACD
BROWNS
[187]
[170]
[186]
[170]
CAFV
CAFV
CAFV
CAFV
acc.,AUC
acc.(r=5)
acc.,AUC
acc.(r=5)
0.591,0.624
68%
N/A
45%
0.621,0.654
73%
0.735, 0.798
60%
5.6 Discussion on age progression methods
Since age progression is a challenging task, age clusters are
labelled with whole age-groups and not exact ages. This
formulation also helps with the highly incomplete datasets
that are available. This issue is more dominant in faces of
children, where images are very hard to collect and old
family photos are often of bad quality.
5.7 Challenges in the wild
The number of widely adopted benchmarks for age progression is limited to 5 datasets, all of which are captured
in the wild. The use of the CAFV protocol does not allow
for quantitative comparison between the different datasets,
since different face verification algorithms are used. Therefore, a comparison can only take place using qualitative
standards that are hard to evaluate. The challenging nature
of the task has led research to focus mainly on datasets with
small variation in pose and image quality. The qualitative
differences between the results is largely accounted to the
methods and so, conclusions can not be drawn regarding
the ’wildness’ of the datasets for the task of age progression.
6
AGE - INVARIANT FACIAL CHARACTERIZATION
Age-invariant facial characterization involves two basic
tasks: age-invariant face recognition and cross-age face verification. The goal of methods for age-invariant face recognition is to build a model that is able to recognize the identity
of a face across different ages from a database of faces. On
the other hand, cross-age face verification aims to determine
whether two age separated images are from the same person. These tasks can be approached either with generative
or discriminative methods. In generative approaches, an
input face is transformed to the target age before performing
face recognition, according to the methods described in the
section 5. On the other hand, in discriminative approaches
[130], [16], age-invariant representations are extracted and
a classification problem is solved. The generative methods
are described in the previous section, while the most recent
discriminative methods are presented in the following.
6.1 Age-invariant descriptors
Results on other databases are scarce. Preservation of
identity results on AgeDB, Morph, CACD and Browns are
tabulated in Table 6. The results on AgeDB are reported for
The methods in this subsection focus on obtaining features
that are invariant to the aging process of the face. The local
descriptors that are widely adopted in face recognition, e.g.,
14
Year
Paper
Repr.
Method
Dataset(score-B - score-A)
Test
Metric
1995
[43]
2D
prototype based
private
N/A
N/A
1999
[166]
3D
PCA
private
N/A
N/A
2001
[211]
2D
wavelet, prototype based
private
N/A
N/A
2006
[176]
2D
craniofacial growth
Custom(28% - 37%)
Recogntion
acc.(rank=1)
2008
[177]
2D
craniofacial, wrinkle analysis
Custom(38% - 51%)
Recogntion
acc.(rank=5)
2010
[207]
2D
And-Or Graph, Markov Chain
2010
[170]
3D
weighted average
LHI
MORPH
FG-NET(42% - 55%)
MORPH(68% - 73%)
BROWNS(45% - 60%)
N/A
N/A
Verification
Verification
Verification
N/A
N/A
acc.(rank=5)
acc.(rank=5)
acc.(rank=5)
2011
[196], [195]
3D
metric learning
FG-NET
Jackson Family
N/A
N/A
N/A
N/A
2012
2012
[225]
[206]
2D
2D
tensor space analysis
RB-AAM, Markov Chain
FG-NET(55% - 70%)
Custom
Recogntion
N/A
Accuracy
N/A
2014
[111]
2D
IAAP
FG-NET
N/A
N/A
2015
[197]
2D
dictonary learning
FG-NET(14.89% - 8.53%)
Verification
EER
2016
2016
[243]
[186]
2D
2D
HFA
RAP
FG-NET(∼ 52% − ∼ 54%)
FG-NET(N/A - 0.709)
FG-NET(N/A-0,806)
CACD(N/A - 0.735)
CACD(N/A-0,798)
FG-NET(∼ 15% - ∼ 9%)
Recognition
Verification
Verification
Verification
Verification
Verification
acc.(rank=1)
Accuracy
AUC
Accuracy
AUC
EER
Custom
AgeDB(0.591- 0.621)
AgeDB(0.624-0,654)
IMDB-WIKI-cleaned( - 82.9%)
N/A
Verification
Verification
Recognition
N/A
Accuracy
AUC
accuracy
2016
[223]
2D
RFA
2017
2017
[255]
[187]
2D
2D
CAAE
RJIVE
2017
[8]
2D
Age-cGAN
TABLE 7: Overview of Age Progression Methods, score-B and score-A correspond to face verification score before and after
age progression
LBPs, have been used in an age-invariant context as well. An
experimental evaluation of local descriptors in age-invariant
face recognition is presented in [13]. Nevertheless, these
classical descriptors cannot be used as stand-alone facial
representations, as they do not always capture the ageinvariant information. Therefore, the following methods focus on introducing new facial representations that are robust
to age changes. The face representation used in [131] is a
hierarchical combination of Gradient Orientations (GO) [34]
of each color channel of the image at different scales, called
Gradient Orientation Pyramid (GOP). The hierarchical information captured by this descriptor is classified using
SVM and improves the verification results on adults. On
the other hand, higher order information does not improve
the accuracy significantly for age changes in teenagers.
Using high dimensional LBPs, an age-invariant representation for cross-age face verification is described in [31], [32].
After extracting the descriptors, a reference representation
of every person is obtained at each age. The features are then
encoded into this reference space and maxpooling is used to
normalize the representations of the same person at different
ages. The resulting features are age-invariant, since they
have a high response at the reference person at any year.
The face pairs are then classified according to their cosine
similarity. Lastly, a novel encoder that makes use of binary
patterns, similarly to LBPs, is introduced in [84]. Unlike
other encoders, this descriptor converts binary patterns to
evenly distributed codes. The final representation is obtained by maximizing the entropy of the descriptor. Identity
matching is performed by decomposing the representation
using Identity Factor Analysis (IFA) and classifying the
cosine similarity of the inputs. A similar feature extractor,
called Local Pattern Selection (LPS), is used at multiple
scales in [128]. The multiple scaling and dense sampling
of the method result in a high-dimensional representation,
which is refined using bagging [21] multiple classifiers.
6.2 Age-invariant subspace learning-based methods
Similar to the previous tasks, instead of focusing on ageinvariant local descriptors, an age-invariant subspace can be
learned from simple representations. To build this subspace,
several subspace analysis methods have been employed. In
particular, Multi-Feature Discriminant Analysis (MFDA) is
employed in [129]. The subspace is obtained from Scale
Invariant Feature Transform (SIFT) [144] and Multi-Scale
LBPs features. The classification problem is solved in the
subspace using bagging. The LBP feature space is also employed in [20] to obtain a subspace that captures geometric
features of the data, such as shape. Nonlinear Topological
Component Analysis (NTCA) is introduced to obtain the
low-dimensional age-invariant subspace.
Another subspace method that has been used in other
age modelling tasks, e.g., age progression [243], is the HFA.
Hidden Factor Analysis decomposes facial features into
a linear combination of the age component, the identity
component and the noise term. Contrary to age progression,
for the task of face recognition, the identity component
is taken into account. The feature space is obtained from
Histograms of Oriented Gradients (HOGs) [44]. The final
15
output is obtained based on the cosine similarity of the
identity components.
Lastly, a different method that coordinates cross-age face
verification and cross-face age verification is introduced
in [52]. Motivated by the fact that, the first task needs
age-invariant features while the latter needs age-sensitive
features, the method coordinates the two in a multi-task
learning manner. To achieve this, both tasks share the same
feature pool and feature interaction is encouraged via an
orthogonal regularization. That is, the final features are
selected so that the age-sensitive ones are avoided.
6.3
Deep learning-based methods
Deep Learning models have not found wide use for ageinvariant facial characterization, possibly due to the lack
of large aging datasets. Nevertheless, a latent factor guided
convolutional network is presented in [226]. The fully connected layer of the CNN is designed using Latent Identity
Analysis, separating the identity components from the rest
of the convolutional features. The experiments performed
indicate that simply finetuning a deep CNN on an aging
dataset improves the accuracy significantly. This reveals
that a baseline CNN needs further processing in order to
efficiently learn age-invariant features.
6.4 Evaluation protocols, metrics, and results for ageinvariant facial characterization
Protocols: The most widely used datasets in this subsection are the FG-NET, MORPH and CACD datasets, while
the protocols involve some of the aforementioned ones
like k-fold cross validation and LOPO. Particularly for the
MORPH Album 2 dataset, a common protocol includes
splitting the dataset into a training and a test subset, each
containing images from 10,000 subjects (10k-10k). Lastly, it
should be noted that since the CACD dataset consists of
some noisy labels as well as duplicates, a verification subset
called CACD-VS is introduced in [32] and used for testing.
Metrics : Different methods have been used to evaluate
age-invariant facial recognition and cross-age face verification. Some methods [52], [131] use Equal Error Rate (EER)
to evaluate the performance of the verification. This metric
describes the rate at which both accept and reject errors
agree. Nevertheless, the most popular evaluation metric is
verification accuracy or recognition accuracy. In most cases,
rank-1 or rank-n accuracy is used, where the rank indicates
the number of gallery images that have to be inspected in
order to achieve this performance.
Results: The most accurate method among the ones
included in this subsection is the deep learning model
in [226], which outperforms the rest in all datasets. The
results are tabulated as follows, along with the second-best
performance in each dataset. A chronological overview of
the methods in this subsection is tabulated in Table 9.
6.5 Discussion on age-invariant characterization methods
Similar to classical Face Verification and Recognition methods, pose, illumination, occlusions and expressions are significant variation inducing factors. Therefore, the aforementioned methods should be able to obtain invariant
TABLE 8: Best reported results on age-invariant facial characterization
dataset
method
metric
protocol
score
FG-NET
FG-NET
MORPH
MORPH
CACD-VS
CACD-VS
[226]
[84]
[226]
[128]
[226]
[32]
rank-1 acc.
rank-1 acc.
rank-1 acc.
rank-1 acc.
rank-1 acc.
rank-1 acc.
LOPO
LOPO
10k-10k
10k-10k
10-fold c-v
10-fold c-v
88.1%
76.2%
97.51%
94.87%
98.5%
87.6%
representations, not only to aging, but also these kind of
variation. The issue of inadequate aging datasets, that was
mentioned in the previous subsection, stands for this task
as well. This is important, especially for images of children,
where experiments [131] have shown that verification is
much harder compared to adults. Lastly, face recognition
and verification depend prominently on the age difference
between the probe image and the gallery image [92], where
the accuracy of the algorithm usually decreases with larger
age gaps.
6.6 Challenges in the wild
Similar to age progression, age-invariant facial characterization methods have been evaluated only on three in the
wild datasets. Since all three benchmarks follow similar
protocols, they can be compared based on the the stateof-the-art results. The results on Table 8 validate the nonchallenging nature of the MORPH2 benchmark, particularly
in comparison to the pose and image quality variation in
FG-NET.
7
K INSHIP V ERIFICATION
Kinship Verification has recently gained interest as a task
in machine learning and computer vision community. The
task refers to bi-subject verification; that is, classifying a pair
of input images as kin or non-kin. A smaller number of
methods have explored tri-subject kinship verification [175],
which compares a couple (Mother and Father) against a
child, or whole family classification [64], as well as kinship
recognition [96]. The seminal work on kinship verification
is by Fang et al. [65]. The first kinship annotated dataset is
introduced and facial parts (FP), such as mouth hair and
nose, are extracted using a pictorial structure model [68].
That is, the facial structures are represented as parts in a
deformable configuration. The representation also included
color information, facial distances (FD) and gradient histograms (GH). The difference between the feature vectors
of the query couple are classified using methods like k-NN
and SVM. The experiments performed in the paper indicate
that the most discriminant feature was the color of the eyes.
Moreover, the highest accuracy was achieved for Father-Son
image-pairs, while the algorithm outperformed the human
workers. Such results suggest that kinship verification of
other people is a difficult task for humans. Similar to the
above, this section is organized based on the facial representation used for the task.
7.1 Invariant descriptors to genetic variations
Intuitively, people perceive kinship based on local facial
attributes. These attributes often correspond to facial parts,
16
Year
Paper
Representation
Method
Dataset(score)
Metric
2007
[130]
GOP
SVM
Private I(5.1%)
Private II(10.8%)
EER
EER
2008
[16]
SIFT
Feature Drift
FG-NET(N/A)
ROC
2010
[131]
GOP
SVM
FG-NET(≥ 18)(24.1%)
FG-NET(8-18)(30.5%)
FG-NET(≤ 8)(38.6%)
Private I(5.1%)
Private II(10.8%)
EER
EER
EER
EER
EER
2011
[129]
SIFT, MLBP
MFDA
FG-NET(47.5%)
MORPH(83.9%)
acc.(rank=1)
acc.(rank=1)
2013
[83]
HOG
PCA, LDA, IFA
FG-NET(69%)
MORPH(91.14%)
acc.(rank=1)
acc.(rank=1)
2014
[31], [32]
CARC
PCA,LDA, SVM
MORPH(92.8%)
CACD(87.6%)
acc.(rank=1)
acc.(rank=1)
2015
[84]
MEFD, MLBP, SIFT
PCA, LDA, IFA
2015
[20]
LBP
NTCA
2015
[52]
SIFT, LBP, GOP, BIF
AGCD
FG-NET(76.2%)
MORPH(94.59%)
FG-NET(48.96%)
MORPH(83.8%)
FG-NET(19.4%)
MORPH(5.5%)
acc.(rank=1)
acc.(rank=1)
acc.(rank=1)
acc.(rank=1)
EER
EER
2016
2016
[128]
[226]
LPS
PI
LFDA, HFA
LF-CNN
MORPH(94.87%)
FG-NET(88.1%)
MORPH(97.51%)
CACD(98.5%)
acc.(rank=1)
acc.(rank=1)
acc.(rank=1)
acc.(rank=1)
TABLE 9: Overview of age-invariant facial characterization methods
like nose, eyes and mouth, the heredity of which indicates
kin-related people. Different descriptors have been applied
to capture facial part information for kinship verification.
The eyes, nose and mouth are explored in [93], where the
DAISY [213] descriptor is used to represent each part. A
similarity score is then computed for each part of the two
faces to determine the familial traits. Similarly, 12 facial
parts are explored in [64] using the dense SIFT (dSIFT)
descriptor. In this approach, each part is reconstructed from
a dictionary of facial parts. The dictionary contains facial
parts from the family of the query face, as well as nonrelated faces. In order to determine the family of a test face,
sparsity is enforced on the reconstruction coefficients, based
on the assumption that the query face only inherits features
from its relatives. The final family prediction is obtained
based on the three most dominant facial attributes.
A set of overlapping facial patches is used for feature
extraction in [235]. The face is segmented in 5 layers from
coarse to fine; that is, the first layer contains the whole face
and the other layers contain increasingly smaller parts. Binary attributes that indicate the presence or absence of a trait
(e.g., mustache), as well as relative attributes (e.g., bigger
nose) are taken into consideration. The LIFT algorithm [253]
is employed to extract the binary attributes, while a ranking
function is learned for the relative ones. Finally, the output
is obtained from an SVM classifier.
Instead of modelling different facial parts, several methods employ texture descriptors that are widely adopted
in face verification. In particular, Local Phase Quantization
(LPQ) [165], Three and Four-Patch LBPs and Weber Local
Descriptors (WLD) [35] are employed to represent the face
in [19]. The fusion of multiple features results in improved
classification accuracy. Feature selection is employed to obtain the final representation and the classification problem
is solved using SVM. Similarly, the Weber-normalized [35]
faces are obtained in [114] to alleviate the illumination variation in the data. The features are detected around salient
points, that are extracted using Difference of Gaussian. A
similarity vector is then computed around each pair of
salient points by employing the Self-Similarity Descriptor
(SSD) [194]. The distance measure is then classified by an
SVM classifier.
On the other hand, in order to perform tri-subject kinship verification, each image is partitioned into overlapping
patches and SIFT features are extracted in [175]. Feature
selection is then performed and Symmetric Bilinear Model
(SBM) is used to learn the similarity between the parents
and the child, by learning the pairwise similarities first.
The verification function is modelled as the linear combination of the two similarities. Calculating the tri-subject
similarity in two steps can induce noise to the model, since
the tri-person inheritance is not considered. Based on that,
Zhang et al. [251] introduced a model to compute tri-subject
dissimilarity. An over-complete set of features from High
Dimensional LBP Histograms (HDLBPH) is employed as
facial representation. The experimental results indicate the
effectiveness of modelling tri-subject kinship jointly instead
of using two bi-subject models.
7.2 Subspace learning-based methods
The methods in this subsection focus on obtaining a kinshipinvariant subspace by making use of tools like factor analysis and transfer learning. The aforementioned descriptors
can be used as local descriptors to build the subspace. In
particular, LBP, LPQ and SIFT descriptors are used in [53].
Based on the hypothesis that, symmetry information can
capture non kinship variations like pose and illumination,
the symmetry feature is employed as the kinship unrelated
17
part of an image. The kinship unrelated part is then subtracted from the original image and a classifier performs
verification. The experimental results indicate that LPQ
perform better than the other descriptors on the KinFaceWII dataset, while LBP perform better on the KinFaceW-I
dataset.
Midlevel features are employed in [242], where LBP
and SIFT descriptors are used as low-level image representations. The final subspace is obtained using Prototype
Discriminative Feature Learning (PDFL). For this method,
two datasets are used, one with labelled kin relations and
one without. To obtain the subspace, the objective function
minimizes the difference between kin-related samples and
maximizes the difference between neighbouring non-kin
samples. An SVM classifier is used for verification, while
the experimental results indicate improved performance
compared to the state-of-the-art. Nevertheless, the model
fails to outperform the human observer for some kinds of
kin relations.
Based on the assumption that parents resemble their
children more closely when they are younger compared to
when they are older, transfer subspace tools are used in [236]
and [193] to determine kin relationships in photos. In particular, the divergence between the distributions of children
and old parents is minimized, using the young parents as an
intermediate set. Gabor features and distance ratios from the
anthropometic model [176] are used as descriptors and the
resulting subspace maximizes the child-young parent and
child-old parent similarities. Experiments on context-aware
kinship verification are performed and are presented in the
subsection 7.7.
7.3
Metric learning-based methods
Metric Learning methods [237] aim at learning a distance
metric that gets similar samples closer than dissimilar ones.
A number of methods have successfully used the tools of
Metric Learning to perform kinship verification using different facial features. In particular, Ensemble Metric Learning
with sample and feature selection is employed in [201] to
create a mapping that reduces the distance between images
of related people. The local representations include Spatial
histogram of SIFT (SH-SIFT), as well as a vector containing
intensity , Pyramid Histogram of Gradients (PHOG) [17]
and Gabor wavelets at four scales and six orientations.
The final decision is obtained using SVR. The experiments
reveal the implications of age difference and gender on the
performance of the kinship verification model, which will
be discussed in the next subsection.
In order to learn a metric that not only projects kinship
relations close but also pulls non-relations out of their
neighbourhood, Neighbourhood Repulsed Metric Learning
(NRML) is introduced in [149]. Further to that, Multiview
NRML is proposed to obtain metrics for multiple feature
representations and to deal with multiview data. Four
different descriptors are used, namely LBP, SIFT, TPLBP
and LE [26]. The last one proves to outperform the other
descriptors in the experiments, due to the fact that it is
directly learned from the data. The method is extended to
handle perioculal images in [171]. The block-based NRML
(BNRML) uses the histograms from the block pairs to learn
multiple distance metrics. The method uses LTP [209] features and outperforms the original NRML method on the
KinFaceW benchmarks.
Multiple distance metrics for different sets of features
are learned in [241] and Discriminative Multi-Metric Learning (DMML) is introduced. Three different descriptors are
employed, namely LBP, SIFT and Spatial Pyramid LEarning
(SPLE) [259] and the method learns a metric for each of them
simultaneously. This method [241] proves to perform better
for multiple feature metric learning, as well as single metric
learning, while the SPLE appears to be the best descriptor
for the task. In both methods, kinship relation is decided
using an SVM classifier.
Multiple feature representations are also employed in
[102] and [101]. Instead of learning a distance metric with
concatenated feature vectors, Large Margin Multi-Metric
Learning (LM3 L) [102], [101] is introduced and multiple
distance metrics are learned jointly. In that way, more discriminative and complementary information is exploited, as
the correlation of the different representations is maximized.
Furthermore, to better exploit the local manifold structure of
images, Local Large Margin Multi-Metric Learning (L2 M3 L)
[101] is introduced to incorporate Local Metric Learning
with LM3 L.
Learning a Mahalanobis distance metric is equivalent to
finding a linear transformation that projects the samples to a
subspace, where the Euclidean distance of the similar samples is smaller than the dissimilar ones. Similarly, a linear
transformation that minimizes the correlation, instead of the
euclidean distance, is learned in [239]. In order to learn a set
of non-linear transformations, Discriminative Deep Metric
Learning (DDML) is introduced in [146]. This method uses
neural networks to project the faces to a discriminative
subspace and is also extended to Descriminative Deep Multi
Metric Learning (DDMML) for multiple features. The experimental results show that deep architectures that employ
multiple features perform better for the tasks of face and
kinship verification. Subsection 7.4 features more on deep
hierarchical representations for kinship verification.
7.4 Deep learning-based representations
Similar to the aforementioned tasks in this survey, Deep
Learning methodologies have shown some of the most
promising results for kinship verification. The widely used
architectures, e.g., VGG16, have been used for this task [182]
to exploit large datasets like the Families In the Wild dataset.
More sophisticated deep learning methods introduced for
kinship verification are described as follows.
A deep autoencoder is used in [46] to encode the relationship between the faces in two images. Instead of
encoding each image separately, a relational model that uses
two images as input is introduced. The gated autoencoder
model outperforms most metric learning techniques, as it
learns the features and the metrics jointly. The output is a
statistic about relatedness and resemblance, which is used
to perform kinship verification. The experimental results indicate the genetic inheritance of facial features in the Family
101 dataset. In particular, sons appear to resemble more their
fathers, while the opposite stands for the daughters.
One of the latest works in kinship verification is the
filtered contractive Deep Belief Network (fcDBN) proposed
18
in [115]. A DBN is a deep learning method that consists
of multiple stacked Restricted Boltzman Machines (RBM)
[100]. The deep representations are obtained by applying
filters to capture the inherent structure of the images. In
order to train the fcDBN model a large number of images
from multiple datasets are used, a strategy that had not
been previously employed for this task before. Kinship
verification is performed by a 3-layer Neural Network and
the method succeeds in outperforming the state-of-the-art
methods in multiple datasets.
7.5
Other representations
Additionally to the above, other approaches have been
introduced for the task of Kinship verification. Inheritable
genetic transformation is modelled in [140] and [174]. In
particular, Fisher Vector [200] in Opponent Colour Space are
used in [174] as local representations. The inheritable information is captured by learning a common transformation on
the representations of child and parent. On the other hand,
the SIFT flow algorithm [133] is employed in [174] and the
inheritable transformation is similarly learned so that the
kin-related representations are as close as possible.
Instead of only applying texture descriptors, Geometric
information is also incorporated in [224]. Facial landmarks
are used and the representation is obtained using the methods of the Grassmann manifold. In addition, appearance
features are extracted by using LBP on a Gaussian image pyramid and are fused with the geometric features.
A Gaussian Mixture Model (GMM) is used for similarity
feature extraction from the appearance features. Verification
is achieved using an SVM and the experiments indicate
that even though geometric information is not effective on
its own for kinship verification, it significantly boosts the
performance when fused with appearance information.
Among other modalities, contextual and dynamic features have been studied for kinship verification. In particular, kinship verification in a photo using prior context
knowledge is studied in [236]. The contextual information
included captures gender relation, age difference, relative
distance and kinship score. The experiments show that
contextual information is useful for kinship verification.
Lastly, dynamic features from smiles are employed studied
in [47]. The dynamic features and are used along with
Completed LBP from Three Orthogonal Patterns (CLBPTOP). The experimental results indicate that smile dynamics
are not sufficiently discriminative when used individually,
but like geometric features, they can increase the accuracy
of kinship verification in combination with spatio-temporal
features. Interestingly, the performance of the model is
higher for spontaneous smiles rather than posed ones.
7.6 Evaluation protocols, metrics, and results for kinship verification
Protocols: The most widely adopted Evaluation Protocol
for Kinship Verification is k-fold cross validation, particularly 5-fold cross validation (5-fold c-v). This protocol is
used for most of the datasets in this subsection. Specifically
for the Vadana dataset, a different protocol is proposed.
The dataset is split into 6 subsets, 3 including parent-child
(VADANA-PC) and 3 including sibling (VADANA-S). For
each relationship, two subsets included only adult faces and
one both adult and child faces. Lastly, the aforementioned
leave-one-out protocol protocol is used for UvA-NEMO
dataset.
Metrics: The common evaluation metric for the task of
kinship verification is the classification accuracy (acc.). The
results are usually obtained for every subset, indicating a
different kinship relation, as well as on the whole dataset.
In cases where the accuracy is obtained on more than one
set, instead of the average, we include the best and the
worst scores reported. Aforementioned metrics, like EER,
are also applied to kinship verification. The metric used in
the VADANA protocol is the accuracy at EER (AEER).
Results: The best reported results for each dataset, along
with the evaluation protocols are tabulated in Table 10. The
methods in this subsection are presented in chronological
order in Table 11.
TABLE 10: Best reported results on Kinship Verification (the
results in [96] are for Kinship Recognition)
dataset
method
metric
protocol
score
TSKinFace
CornellKin
FIW
WVU KInship
VADANA
KinFaceW-I
KinFaceW-II
Family 101
UvA-NEMO
Sibling-Face
Group-Face
UBKinFace
IIITD-Kinship
[251]
[115]
[182]
[115]
[84]
[115]
[115]
[224]
[47]
[96]
[96]
[115]
[114]
acc.
acc.
acc.
acc.
AEER
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
5-fold c-v
5-fold c-v
5-fold c-v
5-fold c-v
vadana
5-fold c-v
5-fold c-v
5-fold c-v
leave-one-out
N/A
N/A
5-fold c-v
5-fold c-v
89.8%
89.5%
71% ±2.3
90.8%
60.43-80.18%
96.1%
96.2%
92.03%
67.11%
52.48%
69.25%
91.8%
75.2%
7.7 Discussion on kinship verification
The performance of kinship verification models can be
affected by a number of factors. Some of the factors that
can deteriorate the performance include pose, image quality
[18], lighting conditions [175], as well as gender and age
difference between kins [201]. Experiments are conducted
in the cited studies to define the effect of each factor.
In order to determine the effectiveness of the machine
in kinship verification, a study on kinship verification by
humans is performed in [149]. The results indicate that
humans perform slightly worse than machines, while contextual information such as hair, colour and background can
improve their accuracy. Moreover, a study of the several
factors affecting human accuracy at verifying kins [115]
indicates that women outperform men, while older people
can distinguish kins better. The kin relationships containing
females appear also to be detected more accurately, which
may be attributed occlusions like beard or mustache. The
study also explains how much kin-related information is
captured in different face patches.
7.8 Challenges in the wild
With the oldest benchmark dating only back to 2010, kinship
verification is a field in its infancy. As mentioned in Section
3, the majority of methods in this subsection is evaluated on
19
in the wild data. Nevertheless, Table 10 indicates significant
differences between the state-of-the-art performance on each
dataset (e.g., 96.2% on KinFaceW II and 71% on FIW). These
differences can be interpreted by the bias in experimental
set ups inflicted by gathering images from the same photo
[143]. Additionally, the laborious task of gathering these
images has resulted in small datasets that do not contain
extreme variation with regards to pose, occlusions or image
quality. On the other hand, the much larger FIW dataset
demonstrates larger variation in pose, image quality as well
as age of the subjects. The problem of kinship verification
in the wild has been tackled in two competitions [147],[145].
The second is based on the KinFaceW datasets while a new
dataset was collected for the first one. The difference in the
results indicates that the KinFaceW datasets are indeed less
challenging.
8
AGING AND K INSHIP
The fact that both aging and kinship are genetically encoded
[22] indicates an inherent synergy between the two. This
synergy is used in [198] to perform kinship guided age
progression. The parent face serves as a prior to predict
the aging of the child. The kinship information is incorporated by morphing the age-progressed face to the one
of the parent. On the other hand, old and young parent
face images are used in [233]. The correlation between the
aging of the parent face and their children is leveraged by a
transfer learning model from the young parent-child to the
old parent-child domain.
Besides the possible synergy between them, aging and
kinship modelling face similar issues when it comes to
dealing with faces in the wild. Variations like illumination, pose, expression and image quality deteriorate the
accuracy of such systems. The methods in the previous
sections apply classical face analysis descriptors, e.g., LBP,
SIFT, HOG, as well as state-of-the-art deep CNN to attack
these problems. Furthermore, aging and kinship modelling
share a number of protocols. In particular, cross-age face
verification and kinship verification systems aim to learn a
similarity metric and similar methods can be applied. This
approach is adopted in [149], where the task of cross-age
face verification is tackled as ’self-kinship verification’ using
metric learning.
Nevertheless, deep learning (DL) methods demonstrate
a number of limitations that hinder the effective use of
such methods for modelling complex transformations like
aging and kinship. To approximate such functions DL methods use hierarchical models that need sufficiently dense
sampling from the data distribution, that is, they are data
hungry. The resulting models are highly non-linear and
complex, creating systems that are not mathematically transparent, while being difficult to train with no guarantee of
convergence. This deems DL methods unsuitable for critical
tasks due to their lack of interpretability and sensitivity to
adversarial examples [86].
To further advance the fields of aging and kinship
modelling, the community should focus on diversifying the
modalities of the data, since the vast majority of research
focuses on 2-Dimensional data. Although some papers indicate the possibility of using dynamic representation for
these tasks (e.g., [47], [48]), the lack of labelled videos
captured in the wild delays progress in this problem. Lastly,
the lack of large in the wild datasets with multiple annotations, i.e. both age and kinship, does not allow for research
on the correlation between age and kinship. Since both
kinship and aging transformations are genetically encoded,
the inheritance of aging patterns can be investigated. Such
datasets will also reveal the possibility of incorporating age
or kinship as soft-biometrics to boost the accuracy of face
recognition systems (e.g., in [115]) as well as investigate the
synergies between the two.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
9
C ONCLUSIONS
Motivated by the increasing interest and plethora of real
world applications, the state-of-the-art methods in age and
kinship modelling has been surveyed. The main challenges
and results of the methods are described for each task
individually. The results tabulated in Tables 4, 7, 9 and
11 indicate the superiority of deep learning methods in
modelling the non-linear transformations of aging and kinship. Specifically, the discriminative power of deep learning
representations, in addition to the availability of in the wild
image datasets, have produced results that surpass human
abilities in the tasks of age estimation, age-invariant facial
characterization and kinship verification. On the other hand,
the application of deep generative models (e.g., GANs) for
the task of age progression has produced photorealistic ageprogressed images with impressive high-frequency details.
[7]
[8]
[9]
[10]
[11]
LHI Image Database. 2010.
A. F. Abate, M. Nappi, D. Riccio, and G. Sabatino. 2d and 3d face
recognition: A survey. Pattern recognition letters, 28(14):1885–1906,
2007.
E. Agustsson, R. Timofte, S. Escalera, X. Baro, I. Guyon, and
R. Rothe. Apparent and real age estimation in still images
with deep residual regressors on appa-real database. In IEEE
International Conference on Automatic Face & Gesture Recognition
(FG), 2017.
T. Ahonen, A. Hadid, and M. Pietikainen. Face description with
local binary patterns: Application to face recognition. IEEE transactions on pattern analysis and machine intelligence, 28(12):2037–
2041, 2006.
R. A. Alley. Social and Applied Aspects of Perceiving Faces. Erlbaum,
Hillsdale, NJ, 1988.
F. Alnajar, T. Gevers, and S. Karaoglu. Age estimation under
changes in image quality: An experimental study. In IEEE
International Conference on Image Processing (ICIP), pages 3987–
3991, Sept 2015.
G. Antipov, M. Baccouche, S.-A. Berrani, and J.-L. Dugelay. Apparent age estimation from face images combining general and
children-specialized deep learning models. In IEEE Conference
on Computer Vision and Pattern Recognition Workshops (CVPR-W),
2016.
G. Antipov, M. Baccouche, and J.-L. Dugelay. Face aging with
conditional generative adversarial networks. IEEE International
Conference on Image Proccessing (ICIP), 2017.
A. Bar-Hillel, T. Hertz, N. Shental, and D. Weinshall. Learning
distance functions using equivalence relations. In Proceedings of
the 20th International Conference on Machine Learning (ICML), pages
11–18, 2003.
A. Bastanfard, M. A. Nik, and M. M. Dehshibi. Iranian face
database with age, pose and expression. In Machine Vision, 2007.
ICMV 2007. International Conference on, pages 50–55. IEEE, 2007.
C. Bauckhage, A. Jahanbekam, and C. Thurau. Age recognition in
the wild. In International Conference on Pattern Recognition (ICPR),
pages 392–395, 2010.
20
Year
Paper
Representation
Method
Dataset(score)
Metric
2010
[65]
FP, FD, Colour, GH
kNN, SVM
CornellKin(70.67%)
acc.
2011
[193]
Gabor
CMC,TSL, kNN
UBKinFace(56.5%)
acc.
2012
2012
2012
2012
2012
[93]
[235]
[114]
[236]
[201]
DAISY
binary, relative attributes
DoG salient points
Gabor
PHOG, Gabor, SH-SIFT
Bayes decision
SVM
SSD, SVM
CMC, TSL, kNN
Meric Learning, SVR
private(75%)
UBKinFace(82.5%)
IITD(75.2%)
UBKinFace(56.5%)
VADANA-PC(80.18%)
VADANA-S(75.64%)
acc.
acc.
acc.
acc.
AEER
AEER
2013
2013
[64]
[47]
dSIFT
dynamic, CLBP-TOP
Sparse Group Lasso
mRMR, SVM
Family101(32%)
UvA-NEMO(67.11%)
acc.
acc.
2014
[149]
LBP, SIFT, TPLBP and LE
MNRLM, SVM
2014
[241]
LBP, SIFT, SPLE
DMML, SVM
2014
2014
[102]
[46]
LE, LBP, TPLBP, SIFT
PI
LMMML
gated autoencoder
2014
2014
[224]
[96]
LBP, Grassman Manifold
LBP,BIF, relative age
GMM, geodesic distance, SVM
logistic regression, graph classifier
KinFaceW-I(69.9%)
KinFaceW-II(76.5%)
KinFaceW-I(69.5-75.5%)
KinFaceW-II(76.5-79.5%)
CornellKin(70.5-77.5%)
UBKinFace(70-74.5%)
KinFaceW-II(78.7%)
KinFaceW-I(74.5%)
KinFaceW-II(82.5%)
Family101(92.03%)
Custom(52.48%)
GroupFace(69.25%)
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
2015
[19]
LPQ, TPLBP, FPLBP, WLD
mRMR, SFFS, SVM
2015
2015
[175]
[53]
SIFT
LPQ
RSBM
PCA, feature substraction
2015
[242]
LBP, SIFT
PDFL, SVM
KinFaceW-I(86.3%)
KinFaceW-II(83.1%)
TSKinFace(85.4%)
KinFaceW-I(70.9%)
KinFaceW-II(77.1%)
KinFaceW-I(70.1%)
KinFaceW-II(77%)
CornellKin(71.9%)
UBKinFace(67.3%)
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
2016
[251]
HDLBP
dissimilarity vector, mRMR
TSKinFace(89.7%)
acc.
2017
2017
[101]
[146]
LE, LBP, TPLBP, SIFT
LBP,DSIFT, HOG, LPQ
LLMMML
DDMML
2017
[239]
LE
NRCML
2017
[171]
LTP
BNRML
2017
2017
[183]
[115]
PI
PI
VGG
fcDBN,3-l NN
KinFaceW-II(80%)
KinFaceW-I(83.5%)
KinFaceW-II(84.3%)
TSKinFace(82.5-88.5%)
KinFaceW-I(66.3%)
KinFaceW-II(78.8%)
KinFaceW-I(78.7%)
KinFaceW-II(80.55%)
FIW(71%±2.3)
KinFaceW-I(96.1%)
KinFaceW-II(96.2%)
CornellKin(89.5%)
UBKinFace(91.8%)
WVU(90.8%)
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
acc.
TABLE 11: Overview of Kinship Verification Methods
[12]
[13]
[14]
[15]
[16]
[17]
H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust
features. European Conference onComputer vision (ECCV), pages
404–417, 2006.
M. Bereta, P. Karczmarek, W. Pedrycz, and M. Reformat. Local
descriptors in application to the aging problem in face recognition. Pattern Recognition, 46(10):2634–2646, 2013.
B. Bhattarai, G. Sharma, A. Lechervy, and F. Jurie. A joint learning
approach for cross domain age estimation. In IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP),
pages 1901–1905, March 2016.
E. L. Bienenstock, L. N. Cooper, and P. W. Munro. Theory for
the development of neuron selectivity: orientation specificity and
binocular interaction in visual cortex. Technical report, DTIC
Document, 1981.
S. Biswas, G. Aggarwal, N. Ramanathan, and R. Chellappa. A
non-generative approach for face recognition across aging. In
IEEE International Conference on Biometrics: Theory, Applications and
Systems (BTAS), pages 1–6. IEEE, 2008.
A. Bosch, A. Zisserman, and X. Munoz. Representing shape
with a spatial pyramid kernel. In Proceedings of the 6th ACM
international conference on Image and video retrieval, pages 401–408.
ACM, 2007.
[18]
[19]
[20]
[21]
[22]
[23]
[24]
A. G. Bottino, M. De Simone, A. Laurentini, and T. Vieira. A new
problem in face image analysis: finding kinship clues for siblings
pairs. 2012.
A. Bottinok, I. U. Islam, and T. F. Vieira. A multi-perspective
holistic approach to kinship verification in the wild. In IEEE International Conference and Workshops on Automatic Face and Gesture
Recognition (FG), volume 2, pages 1–6. IEEE, 2015.
D. Bouchaffra. Nonlinear topological component analysis: Application to age-invariant face recognition. IEEE Transactions on
Neural Networks and Learning Systems, 26(7):1375–1387, July 2015.
L. Breiman. Bagging predictors. Machine learning, 24(2):123–140,
1996.
W. S. Browner, A. J. Kahn, E. Ziv, A. P. Reiner, J. Oshima, R. M.
Cawthon, W.-C. Hsueh, and S. R. Cummings. The genetics of
human longevity. The American journal of medicine, 117(11):851–
860, 2004.
V. Bruce and A. Young. In the eye of the beholder: the science of face
perception. Oxford University Press, 1998.
D. M. Burt and D. I. Perrett. Perception of age in adult caucasian
male faces: Computer graphic manipulation of shape and colour
information. Proceedings of the Royal Society of London B: Biological
Sciences, 259(1355):137–143, 1995.
21
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
D. Cai, X. He, J. Han, and H.-J. Zhang. Orthogonal laplacianfaces for face recognition. IEEE transactions on image processing,
15(11):3608–3614, 2006.
Z. Cao, Q. Yin, X. Tang, and J. Sun. Face recognition with
learning-based descriptor. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 2707–2714. IEEE, 2010.
K. Y. Chang and C. S. Chen. A learning framework for age rank
estimation based on face images with scattering transform. IEEE
Transactions on Image Processing, 24(3):785–798, March 2015.
K. Y. Chang, C. S. Chen, and Y. P. Hung. A ranking approach
for human ages estimation based on face images. In International
Conference on Pattern Recognition (ICPR), pages 3396–3399, 2010.
K. Y. Chang, C. S. Chen, and Y. P. Hung. Ordinal hyperplanes
ranker with cost sensitivities for age estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages
585–592, 2011.
W.-L. Chao, J.-Z. Liu, and J.-J. Ding. Facial age estimation based
on label-sensitive learning and age-oriented regression. Pattern
Recognition, 46(3):628–641, 2013.
B.-C. Chen, C.-S. Chen, and W. H. Hsu. Cross-age reference coding for age-invariant face recognition and retrieval. In European
Conference on Computer Vision (ECCV), 2014.
B. C. Chen, C. S. Chen, and W. H. Hsu. Face recognition and
retrieval using cross-age reference coding with cross-age celebrity
dataset. IEEE Transactions on Multimedia, 17(6):804–815, June 2015.
C. Chen, W. Yang, Y. Wang, K. Ricanek, and K. Luu. Facial feature
fusion and model selection for age estimation. In IEEE International Conference and Workshops on Automatic Face and Gesture
Recognition (FG), pages 200–205, March 2011.
H. F. Chen, P. N. Belhumeur, and D. W. Jacobs. In search of illumination invariants. In Computer Vision and Pattern Recognition,
2000. Proceedings. IEEE Conference on, volume 1, pages 254–261.
IEEE, 2000.
J. Chen, S. Shan, C. He, G. Zhao, M. Pietikainen, X. Chen, and
W. Gao. Wld: A robust local image descriptor. IEEE transactions
on pattern analysis and machine intelligence, 32(9):1705–1720, 2010.
J.-C. Chen, V. M. Patel, and R. Chellappa. Unconstrained face
verification using deep cnn features. In IEEE Winter Conference on
Applications of Computer Vision (WACV), pages 1–9. IEEE, 2016.
K. Chen, S. Gong, T. Xiang, and C. C. Loy. Cumulative attribute
space for age and crowd density estimation. In IEEE Conference
on Computer Vision and Pattern Recognition, pages 2467–2474, June
2013.
S. Chen, C. Zhang, M. Dong, J. Le, and M. Rao. Using rankingcnn for age estimation.
Y. L. Chen and C. T. Hsu. Subspace learning for facial age estimation via pairwise age ranking. IEEE Transactions on Information
Forensics and Security, 8(12):2164–2176, Dec 2013.
T. F. Cootes, G. J. Edwards, and C. J. Taylor. Active appearance
models. IEEE Transactions on pattern analysis and machine intelligence, 23(6):681–685, 2001.
T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham. Active
shape models-their training and application. Computer vision and
image understanding, 61(1):38–59, 1995.
D. Cristinacce and T. F. Cootes. Feature detection and tracking
with constrained local models. In BMVC, volume 1, page 3, 2006.
D. I. P. D. Michael Burt. Perception of age in adult caucasian
male faces: Computer graphic manipulation of shape and colour
information. Proceedings: Biological Sciences, 259(1355):137–143,
1995.
N. Dalal and B. Triggs. Histograms of oriented gradients for
human detection. In Computer Vision and Pattern Recognition, 2005.
CVPR 2005. IEEE Computer Society Conference on, volume 1, pages
886–893. IEEE, 2005.
A. Dantcheva, P. Elia, and A. Ross. What else does your biometric
data reveal? a survey on soft biometrics. IEEE Transactions on
Information Forensics and Security, 11(3):441–467, 2016.
A. Dehghan, E. G. Ortiz, R. Villegas, and M. Shah. Who do i
look like? determining parent-offspring resemblance via gated
autoencoders. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pages 1757–1764, 2014.
H. Dibeklioglu, A. Ali Salah, and T. Gevers. Like father, like
son: Facial expression dynamics for kinship verification. In IEEE
International Conference on Computer Vision (ICCV), pages 1497–
1504, 2013.
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
H. Dibeklioğlu, F. Alnajar, A. A. Salah, and T. Gevers. Combining facial dynamics with appearance for age estimation. IEEE
Transactions on Image Processing, 24(6):1928–1943, 2015.
H. Dibeklioğlu, A. Salah, and T. Gevers. Are you really smiling
at me? spontaneous versus posed enjoyment smiles. European
Conference on Computer Vision (ECCV), pages 525–538, 2012.
C. Ding and D. Tao. A comprehensive survey on pose-invariant
face recognition. ACM Transactions on intelligent systems and
technology (TIST), 7(3):37, 2016.
R. O. Dingman and P. Natvig. Surgery of facial fractures. Saunders,
1964.
L. Du and H. Ling. Cross-age face verification by coordinating
with cross-face age verification. In IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pages 2329–2338, June
2015.
X. Duan and Z.-H. Tan. A feature subtraction method for image
based kinship verification under uncontrolled environments. In
IEEE International Conference on Image Processing (ICIP), pages
1573–1577. IEEE, 2015.
C. N. Duong, K. Luu, K. G. Quach, and T. D. Bui. Beyond principal components: Deep boltzmann machines for face modeling.
In 2015 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), pages 4786–4794, June 2015.
N. C. Ebner, M. Riediger, and U. Lindenberger. FACES–A
database of facial expressions in young, middle-aged, and older
women and men: Development and validation. Behavior research
methods, 42(1):351–362, 2010.
A. Edelman, T. A. Arias, and S. T. Smith. The geometry of algorithms with orthogonality constraints. SIAM journal on Matrix
Analysis and Applications, 20(2):303–353, 1998.
E. Eidinger, R. Enbar, and T. Hassner. Age and gender estimation
of unfiltered faces. IEEE Transactions on Information Forensics and
Security, 9(12):2170–2179, Dec 2014.
P. Ekman and W. V. Friesen. Facial action coding system. 1977.
S. Escalera, J. Fabian, P. Pardo, X. Bar, J. Gonzlez, H. J. Escalante,
D. Misevic, U. Steiner, and I. Guyon. Chalearn looking at people 2015: Apparent age and cultural event recognition datasets
and results. In IEEE International Conference on Computer Vision
Workshop (ICCV- W), pages 243–251, Dec 2015.
S. Escalera, M. Torres Torres, B. Martinez, X. Baro, H. Jair Escalante, I. Guyon, G. Tzimiropoulos, C. Corneou, M. Oliu,
M. Ali Bagheri, and M. Valstar. Chalearn looking at people and
faces of the world: Face analysis workshop and challenge 2016.
In The IEEE Conference on Computer Vision and Pattern Recognition
(CVPR) Workshops, June 2016.
R. V. Exline. Explorations in the process of person perception:
Visual interaction in relation to competition, sex, and need for
affiliation. Journal of personality, 31(1):1–20, 1963.
N. Fan. Learning nonlinear distance functions using neural
network for regression with application to robust human age
estimation. In IEEE International Conference on Computer Vision
(ICCV), pages 249–254, Nov 2011.
H. Fang, P. Grant, and M. Chen. Discriminant feature manifold
for facial aging estimation. In International Conference on Pattern
Recognition (ICPR), pages 593–596, Aug 2010.
R. Fang, A. C. Gallagher, T. Chen, and A. Loui. Kinship classification by modeling facial feature heredity. In IEEE International
Conference on Image Processing (ICIP), pages 2983–2987. IEEE,
2013.
R. Fang, K. D. Tang, N. Snavely, and T. Chen. Towards computational models of kinship verification. In IEEE International
Conference on Image Processing (ICIP), pages 1577–1580. IEEE,
2010.
L. G. Farkas. Anthropometry of the Head and Face. Raven Pr, 1994.
E. Fazl-Ersi, M. E. Mousa-Pasandi, R. Laganire, and M. Awad.
Age and gender recognition using informative features of various
types. In IEEE International Conference on Image Processing (ICIP),
pages 5891–5895, Oct 2014.
P. F. Felzenszwalb and D. P. Huttenlocher. Pictorial structures
for object recognition. International journal of computer vision,
61(1):55–79, 2005.
S. J. Foundation. Human and object interaction processing (hoip)
face database. 2014.
W. A. Freiwald, D. Y. Tsao, and M. S. Livingstone. A face
feature space in the macaque temporal lobe. Nature neuroscience,
12(9):1187–1196, 2009.
22
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
[90]
[91]
[92]
[93]
Y. Freund and R. E. Schapire. A desicion-theoretic generalization
of on-line learning and an application to boosting. In European
conference on computational learning theory, pages 23–37. Springer,
1995.
S. Fu, H. He, and Z.-G. Hou. Learning race from face: A survey. IEEE transactions on pattern analysis and machine intelligence,
36(12):2483–2509, 2014.
Y. Fu, G. Guo, and T. S. Huang. Age synthesis and estimation via
faces: A survey. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 32(11):1955–1976, Nov 2010.
Y. Fu and T. S. Huang. Human age estimation with regression on
discriminative aging manifold. IEEE Transactions on Multimedia,
10(4):578–584, June 2008.
Y. Fu, M. Liu, and T. S. Huang. Conformal embedding analysis
with local graph modeling on the unit hypersphere. In IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
pages 1–6. IEEE, 2007.
Y. Fu and N. Zheng. M-face: An appearance-based photorealistic
model for multiple facial attributes rendering. IEEE Transactions
on Circuits and Systems for Video technology, 16(7):830–842, 2006.
A. C. Gallagher and T. Chen. Understanding images of groups
of people. In IEEE Conference on Computer Vision and Pattern
Recognition, pages 256–263. IEEE, 2009.
F. Gao and H. Ai. Face age classification on consumer images
with gabor feature and fuzzy lda method. In International Conference on Biometrics (ICB), pages 132–141. Springer, 2009.
P. Geladi and B. R. Kowalski. Partial least-squares regression: a
tutorial. Analytica chimica acta, 185:1–17, 1986.
X. Geng, Q. Wang, and Y. Xia. Facial age estimation by adaptive
label distribution learning. In International Conference on Pattern
Recognition (ICPR), pages 4465–4470, Aug 2014.
X. Geng, C. Yin, and Z. H. Zhou. Facial age estimation by learning
from label distributions. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 35(10):2401–2412, Oct 2013.
X. Geng, Z. H. Zhou, and K. Smith-Miles. Automatic age
estimation based on facial aging patterns. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 29(12):2234–2240, Dec
2007.
D. Gong, Z. Li, D. Lin, J. Liu, and X. Tang. Hidden factor analysis
for age invariant face recognition. In IEEE International Conference
on Computer Vision (ICCV), pages 2872–2879, Dec 2013.
D. Gong, Z. Li, D. Tao, J. Liu, and X. Li. A maximum entropy
feature descriptor for age invariant face recognition. In 2015 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
pages 5289–5297, June 2015.
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. WardeFarley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in neural information processing systems
(NIPS), pages 2672–2680, 2014.
I. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. 2015.
A. Graves. Supervised sequence labelling. In Supervised Sequence
Labelling with Recurrent Neural Networks, pages 5–13. Springer,
2012.
G. Guo, Y. Fu, C. R. Dyer, and T. S. Huang. Image-based human
age estimation by manifold learning and locally adjusted robust
regression. IEEE Transactions on Image Processing, 17(7):1178–1188,
2008.
G. Guo and G. Mu. Human age estimation: What is the influence
across race and gender? In IEEE Conference on Computer Vision
and Pattern Recognition Workshop (CVPR- W), pages 71–78, June
2010.
G. Guo and G. Mu. Simultaneous dimensionality reduction and
human age estimation via kernel partial least squares regression.
In IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), pages 657–664, June 2011.
G. Guo, G. Mu, Y. Fu, and T. S. Huang. Human age estimation
using bio-inspired features. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 112–119, June 2009.
G. Guo, G. Mu, and K. Ricanek. Cross-age face recognition on
a very large database: The performance versus age intervals
and improvement using soft biometric traits. In International
Conference on Pattern Recognition (ICPR), pages 3392–3395, Aug
2010.
G. Guo and X. Wang. Kinship measurement on salient facial
features. IEEE Transactions on Instrumentation and Measurement,
61(8):2322–2325, 2012.
[94]
[95]
[96]
[97]
[98]
[99]
[100]
[101]
[102]
[103]
[104]
[105]
[106]
[107]
[108]
[109]
[110]
[111]
[112]
[113]
[114]
[115]
[116]
[117]
[118]
[119]
G. Guo and X. Wang. A study on human age estimation under
facial expression changes. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 2547–2553, June 2012.
G. Guo and C. Zhang. A study on cross-population age estimation. In IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), pages 4257–4263, June 2014.
Y. Guo, H. Dibeklioglu, and L. van der Maaten. Graph-based kinship recognition. In International Conference on Pattern Recognition
(ICPR), pages 4287–4292. Citeseer, 2014.
H. Han, C. Otto, X. Liu, and A. K. Jain. Demographic estimation
from face images: Human vs. machine performance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(6):1148–
1161, June 2015.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for
image recognition. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pages 770–778, 2016.
X. He, S. Yan, Y. Hu, P. Niyogi, and H.-J. Zhang. Face recognition
using laplacianfaces. IEEE transactions on pattern analysis and
machine intelligence, 27(3):328–340, 2005.
G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning
algorithm for deep belief nets. Neural computation, 18(7):1527–
1554, 2006.
J. Hu, J. Lu, Y.-P. Tan, J. Yuan, and J. Zhou. Local large-margin
multi-metric learning for face and kinship verification. IEEE
Transactions on Circuits and Systems for Video Technology, 2017.
J. Hu, J. Lu, J. Yuan, and Y.-P. Tan. Large margin multi-metric
learning for face and kinship verification in the wild. In Asian
Conference on Computer Vision, pages 252–267. Springer, 2014.
Z. Hu, Y. Wen, J. Wang, M. Wang, R. Hong, and S. Yan. Facial
age estimation with age difference. IEEE Transactions on Image
Processing, 26(7):3087–3097, 2017.
G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten.
Densely connected convolutional networks. In IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2017.
G. B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller. Labeled
faces in the wild: A database for studying face recognition in
unconstrained environments. Technical report.
D. H. Hubel and T. N. Wiesel. Receptive fields, binocular
interaction and functional architecture in the cat’s visual cortex.
The Journal of physiology, 160(1):106–154, 1962.
R. Jafri and H. R. Arabnia. A survey of face recognition techniques. 2009.
A. Jain, P. Flynn, and A. A. Ross. Handbook of Biometrics. Springer
Science & Business Media, 2007.
A. K. Jain, S. C. Dass, and K. Nandakumar. Soft biometric traits
for personal recognition systems. In Biometric Authentication,
pages 731–738. Springer, 2004.
M. Kass, A. Witkin, and D. Terzopoulos. Snakes: Active contour
models. International journal of computer vision, 1(4):321–331, 1988.
I. Kemelmacher-Shlizerman, S. Suwajanakorn, and S. M. Seitz.
Illumination-aware age progression. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3334–3341, June
2014.
D. G. Kendall. Shape manifolds, procrustean metrics, and complex projective spaces. Bulletin of the London Mathematical Society,
16(2):81–121, 1984.
J. Kim, D. Han, S. Sohn, and J. Kim. Facial age estimation via
extended curvature gabor filter. In IEEE International Conference
on Image Processing (ICIP), pages 1165–1169, Sept 2015.
N. Kohli, R. Singh, and M. Vatsa. Self-similarity representation
of weber faces for kinship classification. In IEEE International
Conference on Biometrics: Theory, Applications and Systems (BTAS),
pages 245–250. IEEE, 2012.
N. Kohli, M. Vatsa, R. Singh, A. Noore, and A. Majumdar. Hierarchical representation learning for kinship verification. IEEE
Transactions on Image Processing, 26(1):289–302, Jan 2017.
S. Kong, Z. Jiang, and Q. Yang. Modeling neuron selectivity
over simple midlevel features for image classification. IEEE
Transactions on Image Processing, 24(8):2404–2414, Aug 2015.
Z. Kuang, C. Huang, and W. Zhang. Deeply learned rich coding
for cross-dataset facial age estimation. In IEEE International
Conference on Computer Vision Workshop (ICCV-W), pages 338–343,
Dec 2015.
Y. H. Kwon et al. Age classification from facial images. In IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
pages 762–767. IEEE, 1994.
A. Lanitis. FG-NET Aging Database. 2002.
23
[120] A. Lanitis. Comparative evaluation of automatic age-progression
methodologies. EURASIP Journal on Advances in Signal Processing,
2008:101, 2008.
[121] A. Lanitis and C. J. Taylor. Towards automatic face identification
robust to ageing variation. In IEEE International Conference and
Workshops on Automatic Face and Gesture Recognition (FG), pages
391–396. IEEE, 2000.
[122] A. Lanitis, C. J. Taylor, and T. F. Cootes. Toward automatic
simulation of aging effects on face images. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 24(4):442–455, 2002.
[123] A. Lanitis, C. J. Taylor, and T. F. Cootes. Toward automatic
simulation of aging effects on face images. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 24(4):442–455, 2002.
[124] A. Lanitis, N. Tsapatsoulis, K. Soteriou, D. Kuwahara, and S. Morishima. Fg2015 age progression evaluation. In IEEE International
Conference and Workshops on Automatic Face and Gesture Recognition
(FG), volume 1, pages 1–6, May 2015.
[125] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based
learning applied to document recognition. Proceedings of the IEEE,
86(11):2278–2324, 1998.
[126] C. Li, Q. Liu, W. Dong, X. Zhu, J. Liu, and H. Lu. Human
age estimation based on locality and ordinal information. IEEE
Transactions on Cybernetics, 45(11):2522–2534, Nov 2015.
[127] C. Li, Q. Liu, J. Liu, and H. Lu. Learning ordinal discriminative
features for age estimation. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 2570–2577, June 2012.
[128] Z. Li, D. Gong, X. Li, and D. Tao. Aging face recognition: A
hierarchical learning model based on local patterns selection.
IEEE Transactions on Image Processing, 25(5):2146–2154, May 2016.
[129] Z. Li, U. Park, and A. K. Jain. A discriminative model for
age invariant face recognition. IEEE transactions on information
forensics and security, 6(3):1028–1037, 2011.
[130] H. Ling, S. Soatto, N. Ramanathan, and D. W. Jacobs. A study of
face recognition as people age. In IEEE International Conference on
Computer Vision (ICCV), pages 1–8. IEEE, 2007.
[131] H. Ling, S. Soatto, N. Ramanathan, and D. W. Jacobs. Face
verification across age progression using discriminative methods.
IEEE Transactions on Information Forensics and Security, 5(1):82–91,
March 2010.
[132] C. Liu and H. Wechsler. Gabor feature based classification
using the enhanced fisher linear discriminant model for face
recognition. IEEE Transactions on Image processing, 11(4):467–476,
2002.
[133] C. Liu, J. Yuen, and A. Torralba. Sift flow: Dense correspondence
across scenes and its applications. In Dense Image Correspondences
for Computer Vision, pages 15–49. Springer, 2016.
[134] H. Liu, J. Lu, J. Feng, and J. Zhou. Group-aware deep feature
learning for facial age estimation. Pattern Recognition, 66:82–94,
2017.
[135] H. Liu, J. Lu, J. Feng, and J. Zhou. Label-sensitive deep metric
learning for facial age estimation. IEEE Transactions on Information
Forensics and Security, 2017.
[136] H. Liu, J. Lu, J. Feng, and J. Zhou. Label-sensitive deep metric
learning for facial age estimation. IEEE Transactions on Information
Forensics and Security, 13(2):292–305, 2018.
[137] H. Liu and X. Sun. Linear canonical correlation analysis based
ranking approach for facial age estimation. In IEEE International
Conference on Image Processing (ICIP), pages 3249–3253, Sept 2016.
[138] H. Liu and X. Sun. A partial least squares based ranker for fast
and accurate age estimation. In IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), pages 2792–2796.
IEEE, 2016.
[139] M.-Y. Liu and O. Tuzel. Coupled generative adversarial networks. In Advances in neural information processing systems, pages
469–477, 2016.
[140] Q. Liu, A. Puthenputhussery, and C. Liu. Inheritable fisher vector
feature for kinship verification. In IEEE International Conference
on Biometrics: Theory, Applications and Systems (BTAS), pages 1–6.
IEEE, 2015.
[141] X. Liu, S. Li, M. Kan, J. Zhang, S. Wu, W. Liu, H. Han, S. Shan,
and X. Chen. Agenet: Deeply learned regressor and classifier for
robust apparent age estimation. In IEEE International Conference
on Computer Vision Workshop (ICCV- W), pages 258–266, Dec 2015.
[142] E. F. Lock, K. A. Hoadley, J. S. Marron, and A. B. Nobel. Joint
and individual variation explained (jive) for integrated analysis
of multiple data types. The annals of applied statistics, 7(1):523,
2013.
[143] M. B. Lopez, E. Boutellaa, and A. Hadid. Comments on the
kinship face in the wild data sets. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 38(11):2342–2344, Nov 2016.
[144] D. G. Lowe. Distinctive image features from scale-invariant
keypoints. International journal of computer vision, 60(2):91–110,
2004.
[145] J. Lu, J. Hu, V. E. Liong, X. Zhou, A. Bottino, I. U. Islam, T. F.
Vieira, X. Qin, X. Tan, S. Chen, et al. The fg 2015 kinship verification in the wild evaluation. In IEEE International Conference
and Workshops on Automatic Face and Gesture Recognition (FG),
volume 1, pages 1–7. IEEE, 2015.
[146] J. Lu, J. Hu, and Y.-P. Tan. Discriminative deep metric learning
for face and kinship verification. IEEE Transactions on Image
Processing, 2017.
[147] J. Lu, J. Hu, X. Zhou, J. Zhou, M. Castrillón-Santana, J. LorenzoNavarro, L. Kou, Y. Shang, A. Bottino, and T. F. Vieira. Kinship
verification in the wild: The first kinship verification competition.
In Biometrics (IJCB), 2014 IEEE International Joint Conference on,
pages 1–6. IEEE, 2014.
[148] J. Lu, V. E. Liong, and J. Zhou. Cost-sensitive local binary feature
learning for facial age estimation. IEEE Transactions on Image
Processing, 24(12):5356–5368, Dec 2015.
[149] J. Lu, X. Zhou, Y.-P. Tan, Y. Shang, and J. Zhou. Neighborhood repulsed metric learning for kinship verification. IEEE transactions
on pattern analysis and machine intelligence, 36(2):331–345, 2014.
[150] S. Mallat. Group invariant scattering. Communications on Pure and
Applied Mathematics, 65(10):1331–1398, 2012.
[151] S. G. Mallat. A theory for multiresolution signal decomposition:
the wavelet representation. IEEE transactions on pattern analysis
and machine intelligence, 11(7):674–693, 1989.
[152] A. Maronidis and A. Lanitis. Facial age simulation using agespecific 3d models and recursive pca. In VISAPP (1), pages 663–
668, 2013.
[153] M. Mathias, R. Benenson, M. Pedersoli, and L. Van Gool. Face
detection without bells and whistles. In European Conference on
Computer Vision (ECCV), pages 720–735. Springer, 2014.
[154] M. Minear and D. C. Park. A lifespan database of adult facial
stimuli. Behavior Research Methods, 36(4):630–633, 2004.
[155] S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kotsia,
and S. Zafeiriou. Agedb: the first manually collected, in-the-wild
age database. In IEEE Conference on Computer Vision and Pattern
Recognition Workshop (CVPR- W), 2017.
[156] E. Moyse. Age estimation from faces and voices: a review.
Psychologica Belgica, 54(3), 2014.
[157] L. Nanni, A. Lumini, and S. Brahnam. Local binary patterns variants as texture descriptors for medical image analysis. Artificial
intelligence in medicine, 49(2):117–125, 2010.
[158] C. B. Ng, Y. H. Tay, and B. M. Goi. Vision-based human gender
recognition: A survey. 2012.
[159] B. Ni, Z. Song, and S. Yan. Web image mining towards universal
age estimator. In Proceedings of the 17th ACM international conference on Multimedia, pages 85–94. ACM, 2009.
[160] B. Ni, Z. Song, and S. Yan. Web image and video mining
towards universal and robust age estimator. IEEE Transactions
on Multimedia, 13(6):1217–1229, Dec 2011.
[161] Z. Niu, M. Zhou, L. Wang, X. Gao, and G. Hua. Ordinal regression with multiple output cnn for age estimation. In The IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
June 2016.
[162] G. Nixon and P. Galassi. The Brown Sisters: Thirty-Three Years.
Museum of Modern Art, 2007.
[163] M. S. Nixon, P. L. Correia, K. Nasrollahi, T. B. Moeslund, A. Hadid, and M. Tistarelli. On soft biometrics. Pattern Recognition
Letters, 68:218–230, 2015.
[164] T. Ojala, M. Pietikainen, and T. Maenpaa. Multiresolution grayscale and rotation invariant texture classification with local binary patterns. IEEE Transactions on pattern analysis and machine
intelligence, 24(7):971–987, 2002.
[165] V. Ojansivu and J. Heikkilä. Blur insensitive texture classification
using local phase quantization. In International conference on image
and signal processing, pages 236–243. Springer, 2008.
[166] A. J. O’Toole, T. Price, T. Vetter, J. Bartlett, and V. Blanz. 3d shape
and 2d surface textures of human faces: the role of averages in
attractiveness and age. Image and Vision Computing, 18(1):9–19,
1999.
24
[167] A. J. O’toole, T. Vetter, H. Volz, and E. M. Salter. Threedimensional caricatures of human heads: distinctiveness and the
perception of facial age. Perception, 26(6):719–732, 1997.
[168] M. Pantic and L. J. M. Rothkrantz. Automatic analysis of facial
expressions: The state of the art. IEEE Transactions on pattern
analysis and machine intelligence, 22(12):1424–1445, 2000.
[169] J. H. Park, M. Schaller, and M. Van Vugt. Psychology of human
kin recognition: Heuristic cues, erroneous inferences, and their
implications. Review of General Psychology, 12(3):215, 2008.
[170] U. Park, Y. Tong, and A. K. Jain. Age-invariant face recognition.
IEEE Trans. Pattern Anal. Mach. Intell., 32(5):947–954, May 2010.
[171] B. Patel, R. Maheshwari, and B. Raman. Evaluation of periocular
features for kinship verification in the wild. Computer Vision and
Image Understanding, 2017.
[172] P. J. Phillips, P. J. Flynn, T. Scruggs, K. W. Bowyer, J. Chang,
K. Hoffman, J. Marques, J. Min, and W. Worek. Overview of the
face recognition grand challenge. In IEEE Conference on Computer
Vision and Pattern Recognition, volume 1, pages 947–954. IEEE,
2005.
[173] P. J. Phillips, H. Wechsler, J. Huang, and P. J. Rauss. The feret
database and evaluation procedure for face-recognition algorithms. Image and vision computing, 16(5):295–306, 1998.
[174] A. Puthenputhussery, Q. Liu, and C. Liu. Sift flow based genetic
fisher vector feature for kinship verification. In IEEE International
Conference on Image Processing (ICIP), pages 2921–2925, Sept 2016.
[175] X. Qin, X. Tan, and S. Chen. Tri-subject kinship verification: Understanding the core of a family. IEEE Transactions on Multimedia,
17(10):1855–1867, 2015.
[176] N. Ramanathan and R. Chellappa. Modeling age progression in
young faces. In IEEE Computer Society Conference on Computer
Vision and Pattern Recognition (CVPR), volume 1, pages 387–394,
June 2006.
[177] N. Ramanathan and R. Chellappa. Modeling shape and textural
variations in aging faces. In IEEE International Conference and
Workshops on Automatic Face and Gesture Recognition (FG), pages
1–8. IEEE, 2008.
[178] N. Ramanathan, R. Chellappa, and S. Biswas. Computational
methods for modeling facial aging: A survey. Journal of Visual
Languages & Computing, 20(3):131 – 144, 2009.
[179] R. Ranjan, S. Zhou, J. C. Chen, A. Kumar, A. Alavi, V. M. Patel,
and R. Chellappa. Unconstrained age estimation with deep
convolutional neural networks. In IEEE International Conference
on Computer Vision Workshop (ICCV- W), pages 351–359, Dec 2015.
[180] K. Ricanek and T. Tesafaye. Morph: A longitudinal image
database of normal adult age-progression. In IEEE International
Conference on Automatic Face and Gesture Recognition (FG), pages
341–345. IEEE, 2006.
[181] M. Riesenhuber and T. Poggio. Hierarchical models of object
recognition in cortex. Nature neuroscience, 2(11):1019–1025, 1999.
[182] J. P. Robinson, M. Shao, Y. Wu, and Y. Fu. Family in the wild
(fiw): A large-scale kinship recognition database.
[183] J. P. Robinson, M. Shao, Y. Wu, and Y. Fu. Families in the wild
(fiw): Large-scale kinship image database and benchmarks. In
Proceedings of the 2016 ACM on Multimedia Conference, MM ’16,
pages 242–246, New York, NY, USA, 2016. ACM.
[184] R. Rothe, R. Timofte, and L. V. Gool. Deep expectation of real
and apparent age from a single image without facial landmarks.
International Journal of Computer Vision, July 2016.
[185] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet
large scale visual recognition challenge. International Journal of
Computer Vision, 115(3):211–252, 2015.
[186] C. Sagonas, Y. Panagakis, S. Arunkumar, N. Ratha, and
S. Zafeiriou. Back to the future: A fully automatic method for
robust age progression. In International Conference on Pattern
Recognition (ICPR), pages 4226–4231, Dec 2016.
[187] C. Sagonas, Y. Panagakis, A. Leidinger, S. Zafeiriou, et al. Robust
joint and individual variance explained. In IEEE Computer Society
Conference on Computer Vision and Pattern Recognition (CVPR),
2017.
[188] R. Salakhutdinov and G. Hinton. Deep boltzmann machines. In
Artificial Intelligence and Statistics, pages 448–455, 2009.
[189] E. Sariyanidi, H. Gunes, and A. Cavallaro. Automatic analysis
of facial affect: A survey of registration, representation, and
recognition. IEEE transactions on pattern analysis and machine
intelligence, 37(6):1113–1133, 2015.
[190] R. E. Schapire and Y. Singer. Improved boosting algorithms
using confidence-rated predictions. Machine learning, 37(3):297–
336, 1999.
[191] K. Scherbaum, M. Sunkel, H.-P. Seidel, and V. Blanz. Prediction
of individual non-linear aging trajectories of faces. In Computer
Graphics Forum, volume 26, pages 285–294. Wiley Online Library,
2007.
[192] H. S. Seung and D. D. Lee. The manifold ways of perception.
science, 290(5500):2268–2269, 2000.
[193] M. Shao, S. Xia, and Y. Fu. Genealogical face recognition based
on ub kinface database. In IEEE Conference on Computer Vision and
Pattern Recognition Workshop (CVPR- W), pages 60–65. IEEE, 2011.
[194] E. Shechtman and M. Irani. Matching local self-similarities across
images and videos. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR). IEEE, 2007.
[195] C.-T. Shen, F. Huang, W.-H. Lu, S.-W. Shih, and H.-Y. M. Liao.
3d age progression prediction in children’s faces with a small
exemplar-image set. J. Inf. Sci. Eng., 30(4):1131–1148, 2014.
[196] C.-T. Shen, W.-H. Lu, S.-W. Shih, and H.-Y. M. Liao. Exemplarbased age progression prediction in children faces. In IEEE International Symposium on Multimedia (ISM), pages 123–128. IEEE,
2011.
[197] X. Shu, J. Tang, H. Lai, L. Liu, and S. Yan. Personalized age progression with aging dictionary. In IEEE International Conference on
Computer Vision (ICCV), pages 3970–3978, Dec 2015.
[198] X. Shu, J. Tang, H. Lai, Z. Niu, and S. Yan. Kinship-guided age
progression. Pattern Recognition, 59:156–167, 2016.
[199] T. Sim, S. Baker, and M. Bsat. The cmu pose, illumination, and
expression (pie) database. In IEEE International Conference on
Automatic Face and Gesture Recognition (FG), pages 53–58. IEEE,
2002.
[200] K. Simonyan, O. M. Parkhi, A. Vedaldi, and A. Zisserman. Fisher
vector faces in the wild.
[201] G. Somanath and C. Kambhamettu. Can faces verify bloodrelations? In IEEE International Conference on Biometrics: Theory,
Applications and Systems (BTAS), pages 105–112. IEEE, 2012.
[202] G. Somanath, M. Rohith, and C. Kambhamettu. Vadana: A dense
dataset for facial image analysis. In IEEE International Conference
on Computer Vision Workshops (ICCV- W), pages 2175–2182. IEEE,
2011.
[203] Z. Song, B. Ni, D. Guo, T. Sim, and S. Yan. Learning universal
multi-view age estimator using video context. In IEEE International Conference on Computer Vision (ICCV), pages 241–248, Nov
2011.
[204] K. O. Stanley. Compositional pattern producing networks: A
novel abstraction of development. Genetic programming and evolvable machines, 8(2):131–162, 2007.
[205] Y. Su, Y. Fu, Q. Tian, and X. Gao. Cross-database age estimation
based on transfer learning. In IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), pages 1270–1273,
March 2010.
[206] J. Suo, X. Chen, S. Shan, W. Gao, and Q. Dai. A concatenational
graph evolution aging model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11):2083–2096, Nov 2012.
[207] J. Suo, S. C. Zhu, S. Shan, and X. Chen. A compositional and
dynamic model for face aging. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 32(3):385–401, March 2010.
[208] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov,
D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with
convolutions. In IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pages 1–9, 2015.
[209] X. Tan and B. Triggs. Enhanced local texture feature sets for face
recognition under difficult lighting conditions. IEEE transactions
on image processing, 19(6):1635–1650, 2010.
[210] P. Thukral, K. Mitra, and R. Chellappa. A hierarchical approach
for human age estimation. In 2012 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), pages 1529–1532.
IEEE, 2012.
[211] B. Tiddeman, M. Burt, and D. Perrett. Prototyping and transforming facial textures for perception research. IEEE computer graphics
and applications, 21(5):42–50, 2001.
[212] M. E. Tipping. Sparse bayesian learning and the relevance vector
machine. Journal of machine learning research, 1(Jun):211–244, 2001.
[213] E. Tola, V. Lepetit, and P. Fua. Daisy: An efficient dense descriptor
applied to wide-baseline stereo. IEEE transactions on pattern
analysis and machine intelligence, 32(5):815–830, 2010.
25
[214] M. A. Turk and A. P. Pentland. Face recognition using eigenfaces.
In IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), pages 586–591. IEEE, 1991.
[215] K. Ueki, T. Hayashida, and T. Kobayashi. Subspace-based agegroup classification using facial images under various lighting
conditions. In Automatic Face and Gesture Recognition, 2006. FGR
2006. 7th International Conference on, pages 6–pp. IEEE, 2006.
[216] M. Uricar, R. Timofte, R. Rothe, J. Matas, and L. Van Gool.
Structured output svm prediction of apparent age, gender and
smile from deep features. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR) Workshops, June 2016.
[217] V. Vapnik and A. Vashist. A new learning paradigm: Learning
using privileged information. Neural networks, 22(5):544–557,
2009.
[218] V. N. Vapnik and V. Vapnik. Statistical learning theory, volume 1.
Wiley New York, 1998.
[219] T. F. Vieira, A. Bottino, A. Laurentini, and M. De Simone. Detecting siblings in image pairs. The Visual Computer, 30(12):1333–1345,
2014.
[220] P. Viola and M. J. Jones. Robust real-time face detection. International journal of computer vision, 57(2):137–154, 2004.
[221] N. Wang, X. Gao, D. Tao, H. Yang, and X. Li. Facial feature point
detection: A comprehensive survey. Neurocomputing, 2017.
[222] S. Wang, D. Tao, and J. Yang. Relative attribute svm+ for learning
for age estimation. IEEE Transactions on Cybernetics, 46(3):827–839,
March 2016.
[223] W. Wang, Z. Cui, Y. Yan, J. Feng, S. Yan, X. Shu, and N. Sebe.
Recurrent face aging. In The IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), June 2016.
[224] X. Wang and C. Kambhamettu. Leveraging appearance and
geometry for kinship verification. In IEEE International Conference
on Image Processing (ICIP), pages 5017–5021. IEEE, 2014.
[225] Y. Wang, Z. Zhang, W. Li, and F. Jiang. Combining tensor
space analysis and active appearance models for aging effect
simulation on face images. IEEE Transactions on Systems, Man,
and Cybernetics, Part B (Cybernetics), 42(4):1107–1118, Aug 2012.
[226] Y. Wen, Z. Li, and Y. Qiao. Latent factor guided convolutional
neural networks for age-invariant face recognition. In The IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
June 2016.
[227] L. Wolf, T. Hassner, and Y. Taigman. Descriptor based methods
in the wild. In European Conference on Computer Vision Workshop
(ECCV- W), 2008.
[228] T. Wu, P. Turaga, and R. Chellappa. Age estimation and face
verification across aging using landmarks. IEEE Transactions on
Information Forensics and Security, 7(6):1780–1788, Dec 2012.
[229] Y. Wu, P. Kalra, L. Moccozet, and N. Magnenat-Thalmann. Simulating wrinkles and skin aging. The visual computer, 15(4):183–198,
1999.
[230] Y. Wu, N. M. Thalmann, and D. Thalmann. A plastic-visco-elastic
model for wrinkles in facial animation and skin aging.
[231] Y. Wu, N. M. Thalmann, and D. Thalmann. A dynamic wrinkle
model in facial animation and skin ageing. Computer Animation
and Virtual Worlds, 6(4):195–205, 1995.
[232] B. Xia, B. B. Amor, M. Daoudi, and H. Drira. Can 3d shape of the
face reveal your age? In 2014 International Conference on Computer
Vision Theory and Applications (VISAPP), volume 2, pages 5–13,
Jan 2014.
[233] S. Xia, M. Shao, and Y. Fu. Kinship verification through transfer
learning.
[234] S. Xia, M. Shao, and Y. Fu. Kinship verification through transfer
learning. In IJCAI Proceedings-international joint conference on
artificial intelligence, volume 22, page 2539, 2011.
[235] S. Xia, M. Shao, and Y. Fu. Toward kinship verification using
visual attributes. In International Conference on Pattern Recognition
(ICPR), pages 549–552. IEEE, 2012.
[236] S. Xia, M. Shao, J. Luo, and Y. Fu. Understanding kin relationships in a photo. IEEE Transactions on Multimedia, 14(4):1046–
1056, 2012.
[237] E. P. Xing, M. I. Jordan, S. J. Russell, and A. Y. Ng. Distance metric
learning with application to clustering with side-information. In
Advances in neural information processing systems, pages 521–528,
2003.
[238] J. Xing, K. Li, W. Hu, C. Yuan, and H. Ling. Diagnosing deep
learning models for high accuracy age estimation from a single
image. Pattern Recognition, 66:106–116, 2017.
[239] H. Yan. Kinship verification using neighborhood repulsed correlation metric learning. Image and Vision Computing, 60:91–97,
2017.
[240] H. Yan and J. Hu. Video-based kinship verification using distance
metric learning. Pattern Recognition, 2017.
[241] H. Yan, J. Lu, W. Deng, and X. Zhou. Discriminative multimetric
learning for kinship verification. IEEE Transactions on Information
forensics and security, 9(7):1169–1178, 2014.
[242] H. Yan, J. Lu, and X. Zhou. Prototype-based discriminative
feature learning for kinship verification. IEEE transactions on
cybernetics, 45(11):2535–2545, 2015.
[243] H. Yang, D. Huang, Y. Wang, H. Wang, and Y. Tang. Face
aging effect simulation using hidden factor analysis joint sparse
representation. IEEE Transactions on Image Processing, 25(6):2493–
2507, 2016.
[244] M. Yang, S. Zhu, F. Lv, and K. Yu. Correspondence driven
adaptation for human profile recognition. In IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), pages 505–512,
June 2011.
[245] Z. Yang and H. Ai. Demographic classification with local binary
patterns. Advances in Biometrics, pages 464–473, 2007.
[246] D. Yi, Z. Lei, and S. Z. Li. Age estimation by multi-scale
convolutional network. In Asian Conference on Computer Vision
(ACCV), pages 144–158. Springer, 2014.
[247] S. Zafeiriou, C. Zhang, and Z. Zhang. A survey on face detection
in the wild: past, present and future. Computer Vision and Image
Understanding, 138:1–24, 2015.
[248] L. A. Zebrowitz and J. M. Montepare. Social psychological
face perception: Why appearance matters. Social and Personality
Psychology Compass, 2(3):1497–1517, 2008.
[249] Z. Zeng, M. Pantic, G. I. Roisman, and T. S. Huang. A survey
of affect recognition methods: Audio, visual, and spontaneous
expressions. IEEE transactions on pattern analysis and machine
intelligence, 31(1):39–58, 2009.
[250] B. Zhang, S. Shan, X. Chen, and W. Gao. Histogram of gabor
phase patterns (hgpp): A novel object representation approach
for face recognition. IEEE Transactions on Image Processing,
16(1):57–68, 2007.
[251] J. Zhang, S. Xia, H. Pan, and A. Qin. A genetics-motivated
unsupervised model for tri-subject kinship verification. In IEEE
International Conference on Image Processing (ICIP), pages 2916–
2920. IEEE, 2016.
[252] K. Zhang, Z. Zhang, Z. Li, and Y. Qiao. Joint face detection
and alignment using multitask cascaded convolutional networks.
IEEE Signal Processing Letters, 23(10):1499–1503, 2016.
[253] M.-L. Zhang and L. Wu. Lift: Multi-label learning with labelspecific features. IEEE transactions on pattern analysis and machine
intelligence, 37(1):107–120, 2015.
[254] Y. Zhang and D. Y. Yeung. Multi-task warped gaussian process
for personalized age estimation. In 2010 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, pages 2622–
2629, June 2010.
[255] Z. Zhang, Y. Song, and H. Qi. Age progression/regression by
conditional adversarial autoencoder. In CVPR, 2017.
[256] J. Zhao, M. Mathieu, and Y. LeCun. Energy-based generative
adversarial networks. In International Conference on Learning
Representations (ICLR), 2017.
[257] W. Zhao, R. Chellappa, P. J. Phillips, and A. Rosenfeld. Face
recognition: A literature survey. ACM computing surveys (CSUR),
35(4):399–458, 2003.
[258] S. K. Zhou, B. Georgescu, X. S. Zhou, and D. Comaniciu. Image
based regression using boosting method. In IEEE International
Conference on Computer Vision (ICCV), volume 1, pages 541–548.
IEEE, 2005.
[259] X. Zhou, J. Hu, J. Lu, Y. Shang, and Y. Guan. Kinship verification
from facial images under uncontrolled conditions. In Proceedings
of the 19th ACM international conference on Multimedia, pages 953–
956. ACM, 2011.
[260] Y. Zhu, Y. Li, G. Mu, and G. Guo. A study on apparent age
estimation. In IEEE International Conference on Computer Vision
Workshop (ICCV-W), pages 267–273, Dec 2015.
| 1cs.CV
|
arXiv:1712.04016v1 [math.AC] 11 Dec 2017
FREE RESOLUTIONS OF DYNKIN FORMAT
AND THE LICCI PROPERTY OF GRADE 3 PERFECT IDEALS
LARS WINTHER CHRISTENSEN, OANA VELICHE, AND JERZY WEYMAN
Abstract. Recent work on generic free resolutions of length 3 attaches to
every resolution a graph and suggests that resolutions whose associated graph
is a Dynkin diagram are distinguished. We conjecture that in a regular local
ring, every grade 3 perfect ideal whose minimal free resolution is distinguished
in this way is in the linkage class of a complete intersection.
1. Introduction
Let Q be a commutative Noetherian ring. Quotient rings of Q that have projective
dimension at most 3 as Q-modules have been and still are investigated vigorously.
A recent development is Weyman’s [13] construction of generic rings for resolutions
of length 3. For a free resolution,
∂
∂
∂
F0 ←−1− F1 ←−2− F2 ←−3− F3 ←− 0 ,
the format is the quadruple (f0 , f1 , f2 , f3 ) with fi = rank Fi . To each resolution
format Weyman associates a graph and a generic ring, and he proves that the
generic ring is Noetherian exactly if the graph is a Dynkin diagram. This suggests
that free resolutions of formats corresponding to Dynkin diagrams play a special
role; here we explore what that role might be in the context of linkage.
Let Q be local and I ⊆ Q be a perfect ideal of grade 3. Let m be the minimal
number of generators of I, the minimal free resolution of Q/I over Q then has
format f = (1, m, m + n − 1, n) for m > 3 and some integer n > 1, which is often
referred to as the type of the ring Q/I. For convenience we refer to the tuple f as
the resolution format, or simply the format, of I. The graph associated to f is
•m−3
···
•1
•
•1
···
•n
•
A complete intersection ideal has resolution format (1, 3, 3, 1), and the corresponding graph is the Dynkin diagram A3 . A Gorenstein ideal that is not complete
intersection has format (1, m, m, 1) for some odd m > 5, and the associated graph
is the Dynkin diagram Dm . There is another way to obtain Dynkin diagrams
of type D, namely from almost complete intersection ideals, which have formats
Date: 11 December 2017.
2010 Mathematics Subject Classification. 13C40; 13D02.
Key words and phrases. Free resolutions, licci ideals, linkage, perfect ideals.
This work is part of a body of research that started during the authors’ visit to MSRI in
Spring 2013 and continued during a months-long visit by L.W.C. to Northeastern University; the
hospitality of both institutions is acknowledged with gratitude. L.W.C. was partly supported by
NSA grant H98230-14-0140 and Simons Foundation collaboration grant 428308, and J.W. was
partly supported by NSF DMS grant 1400740.
2
L. W. CHRISTENSEN, O. VELICHE, AND J. WEYMAN
(1, 4, n + 3, n), corresponding to Dn+3 , for n > 2. It is elementary to verify that
E6 , E7 , and E8 are the only other Dynkin diagrams that can possibly arise in this
manner. In the balance of this paper, we say that a (resolution) format is Dynkin
if the corresponding graph is a Dynkin diagram.
If Q is regular with maximal ideal n, then every grade 2 perfect ideal I ⊆ n2
is licci, that is, in the linkage class of a complete intersection ideal. Not every
perfect ideal of grade 3 is licci, and in a certain sense licci ideals of grade 4 are
rare; see Miller and Ulrich [11]. Works of Kunz [10] and Watanabe [12] show that a
grade 3 perfect ideal is licci if it is Gorenstein or almost complete intersection; these
results predate the term linkage, but Buchsbaum and Eisenbud interpret them in
the introduction to [3]. The purpose of this paper is to motivate the following:
Conjecture. Let (Q, n) be a regular local ring and f = (1, m, m + n − 1, n) be a
resolution format realized by some grade 3 perfect ideal contained in n2 .
I If f is not Dynkin, then there exists a grade 3 perfect ideal I ⊆ n2 of format
f that is not licci.
II If f is Dynkin, then every grade 3 perfect ideal I ⊆ n2 of format f is licci.
The paper is organized as follows. In Section 2 we recall basics on linkage and
prove a few technical results. In Section 3 we collect evidence for part I of the
conjecture; in fact, we prove it for local rings that are obtained by localizing a
polynomial algebra at the irrelevant maximal ideal. Some of our evidence for part
II was already discussed above, and further evidence is provided in Section 4.
2. Linkage of grade 3 perfect ideals
To goal of this section, other than to recall the language of linkage, is to establish a
proposition that allows some control over the resolution format of a directly linked
ideal. To this end we need a folklore application of Prime Avoidance for which we
did not find a citable reference.
(2.1) Lemma. Let Q be a commutative ring and P1 , . . . , Ps be prime ideals in Q.
For every ideal J and element x in Q with J + (x) 6⊆ P1 ∪ P2 ∪ . . . ∪ Ps there exists
an element v ∈ J with v + x ∈
/ P1 ∪ P2 ∪ . . . ∪ Ps .
Proof. Without loss of generality we may assume that there are no containments
among the prime ideals P1 , . . . , Ps and that they are ordered such that there is an
r between 0 and s with x ∈ P1 , . . . , Pr and x ∈
/ Pr+1 , . . . , Ps . Notice that if r = 0,
then one can take v = 0.
Let r > 0. From the assumption on J + (x) one gets J 6⊆ P1 ∪ . . . ∪ Pr ; choose an
element u ∈ J \ (P1 ∪ . . . ∪ Pr ). If r = s, then one can take v = u. In the case r < s
one has Pr+1 ∩ . . . ∩ Ps 6⊆ P1 ∪ . . . ∪ Pr as, indeed, by Prime Avoidance containment
would imply Pr+1 ∩ . . . ∩ Ps ⊆ Pi for some i and then one would have Pj ⊆ Pi
for some j 6= i, which contradicts the assumption on P1 , . . . , Ps . Now choose an
element y ∈ Pr+1 ∩ . . . ∩ Ps \ P1 ∪ . . . ∪ Pr . It is elementary to verify that uy + x is
not in P1 ∪ P2 ∪ . . . ∪ Ps .
In the rest of this section, we keep the setup and notation close to [1] by Avramov,
Kustin, and Miller; the reader may see [1, sec. 1] for more background.
Let I be a perfect ideal of grade 3 in a local ring Q. An ideal J ⊆ Q is said to be
directly linked to I if there exists a regular sequence x1 , x2 , x3 of elements in I with
FREE RESOLUTIONS OF DYNKIN FORMAT
3
J = (x1 , x2 , x3 ) : I. The ideal J is then also a perfect ideal of grade 3. An ideal J
is said to be linked to I if there exists a sequence of ideals I = J0 , J1 , . . . , Jn = J
such that Ji+1 is directly linked to Ji for each i = 0, . . . , n − 1.
(2.2) Proposition. Let Q be a local ring and I ⊆ Q a grade 3 perfect ideal of
resolution format (1, m, m + n − 1, n).
(a) If m > 4 holds, then there exists a grade 3 perfect ideal of resolution format
(1, n + 3, m + n, m − 2) that is directly linked to I.
(b) If m > 5 holds, then there exists a grade 3 perfect ideal of resolution format
(1, n + 3, m + n − 1, m − 3) that is directly linked to I.
Proof. Denote by n the maximal ideal of Q and by k the residue field Q/n. Let
F• be the minimal free resolution of Q/I. For a regular sequence x1 , x2 , x3 in
I the corresponding Koszul complex is denoted by K• . The canonical surjection
Q/(x1 , x2 , x3 ) → Q/I lifts to a morphism ϕ• : K• → F• . As recalled in [1, sec. 1],
the dual of the mapping cone of ϕ• yields a, not necessarily minimal, free resolution
of Q/J. Further, the ranks of the modules in the minimal free resolution G• of Q/J
are given by
rankk G1 = n + 3 − rankk (ϕ3 ⊗ k) − rankk (ϕ2 ⊗ k)
rankk G2 = m + n + 2 − rankk (ϕ2 ⊗ k) − rankk (ϕ1 ⊗ k)
rankk G3 = m − rankk (ϕ1 ⊗ k) .
Set A• = TorQ
• (Q/I, k). The morphism ϕ• ⊗ k : H(K• ) → A• is a morphism of
graded-commutative k-algebras.
(a): By assumption, I is not generated by a regular sequence, so by [1, thm. 2.1]
one has (A1 )3 = 0 and, therefore, rankk (ϕ3 ⊗ k) = 0. It now suffices to show
that one can choose a regular sequence x1 , x2 , x3 such that rankk (ϕ1 ⊗ k) = 2 and
rankk (ϕ2 ⊗ k) = 0 hold. By [1, thm. 2.1] one can choose minimal generators x′1 and
x′2 of I such that the corresponding cycles z1′ and z2′ in F1 satisfy [z1′ ] · [z2′ ] = 0 in
A• . Let P1 , . . . , Ps be the associated prime
Ss ideals of Q. As I has grade 3, the ideal
nI + (x′1 ) is not contained in the union i=1 Pi . It now follows from
Ss Lemma (2.1)
that there exits an element v1 in nI such that x1 = v1 + x′1 is not in i=1 Pi , i.e. not
a zero-divisor. Similarly, there exists a v2 ∈ nI such that x2 = v2 + x′2 is not in the
union of the associated primes of the Q-module Q/(x1 ). Thus, x1 , x2 form a regular
sequence of minimal generators of I, and the corresponding cycles z1 and z2 in F1
satisfy [z1 ] = [z1′ ] and [z2 ] = [z2′ ] in A1 . Finally there exists, again by Lemma (2.1),
an element x3 in nI and not in the union of the associated primes of the Q-module
Q/(x1 , x2 ). With this regular sequence x1 , x2 , x3 one has rankk (ϕ1 ⊗ k) = 2, as the
cycle z3 in F1 corresponding to x3 satisfies [z3 ] = 0 in A1 . In particular, one has
[z1 ] · [z3 ] = 0 = [z2 ] · [z3 ] and hence rankk (ϕ2 ⊗ k) = 0.
(b): By assumption, I is not generated by a regular sequence, so by [1, thm. 2.1]
one has (A1 )3 = 0 and, therefore, rankk (ϕ3 ⊗ k) = 0. It now suffices to show
that one can choose a regular sequence x1 , x2 , x3 such that rankk (ϕ1 ⊗ k) = 3 and
rankk (ϕ2 ⊗ k) = 0 hold. By [1, thm. 2.1] one can choose minimal generators x′1 , x′2 ,
and x′3 of I such that the corresponding cycles z1′ , z2′ , z3′ ∈ F1 satisfy [zi′ ] · [zj′ ] = 0 in
A• . (Notice that the assumption m > 5 is needed in order to choose these elements
in case the multiplicative structure is TE; for the other possible structures, B, G,
and H, such elements can be chosen as long as m is at least 4.) As in the proof of
4
L. W. CHRISTENSEN, O. VELICHE, AND J. WEYMAN
part (a) there exist elements v1 , v2 , v3 ∈ nI such that v1 + x′1 , v2 + x′2 , v3 + x′3 is a
regular sequence in I with the desired properties.
(2.3) Corollary. Let Q be a local ring and I ⊆ Q a grade 3 perfect ideal of
resolution format (1, m, m + n − 1, n).
(a) If m > 4 holds, then there exists a grade 3 perfect ideal of resolution format
(1, m + 1, m + n + 1, n + 1) that is linked to I.
(b) If m > 5 holds, then there exists a grade 3 perfect ideal of resolution format
(1, m, m + n, n + 1) that is linked to I.
(c) If m > 4 and n > 2 holds, then there exists a grade 3 perfect ideal of resolution
format (1, m + 1, m + n, n) that is linked to I.
Proof. Part (a) follows from two consecutive applications of Proposition (2.2)(a).
Part (b) follows from an application of (2.2)(b) followed an application of (2.2)(a).
Finally, part (c) follows by application of (2.2)(a) and (2.2)(b) in that order.
Let I ⊆ Q be a perfect ideal of grade 3. For every regular sequence x1 , x2 , x3
in I one has (x1 , x2 , x3 ) : ((x1 , x2 , x3 ) : I) = I; see Golod [7]. Thus if an ideal J
is directly linked to I, then I is also directly linked to J. It follows that “being
linked” is an equivalence relation. The ideal I is called licci if it is in the linkage
class of a complete intersection ideal.
3. Evidence for Part I
By a Dynkin diagram we always mean a simply laced Dynkin diagram, i.e. a diagram
from the ADE classification. Recall from the Introduction that
•m−3
···
•1
•
•1
···
•n
•
is the graph associated to the resolution format (1, m, m + n − 1, n). The Dynkin
formats that can be realized by grade 3 perfect ideals are
A3
(3.1)
corresponding to
(1, 3, 3, 1)
Dm
—
(1, m, m, 1) for odd m > 5
Dn+3
—
(1, 4, n + 3, n) for n > 2
E6
—
(1, 5, 6, 2)
E7
—
(1, 6, 7, 2) and (1, 5, 7, 3)
E8
—
(1, 7, 8, 2) and (1, 5, 8, 4)
This is straightforward to verify when one recalls that
• (1, 3, 3, 1) is the only possible format with m = 3; this excludes the formats
(1, 3, n + 2, n) corresponding to An+2 for n > 2.
• The minimal number of generators of grade 3 Gorenstein ideal is odd, see [3].
This excludes the formats (1, m, m, 1) corresponding to Dm for even m > 4.
(3.2) Theorem. Let k a field and e > 3; set Q = k[X1 , . . . , Xe ](X1 ,...,Xe ) . For every
resolution format f = (1, m, m + n − 1, n) with m > 3 and n > 1 that is not Dynkin
there exists a grade 3 perfect ideal that is contained in (X1 , . . . , Xe )2 has resolution
format f and is not licci.
FREE RESOLUTIONS OF DYNKIN FORMAT
5
It is implicit in this statement that every format (1, m, m + n − 1, n) that is not
Dynkin is realized by a degree 3 perfect ideal in the ring Q. A proof was sketched
in the last section of [13], we provide the details in the Appendix.
The proof of Theorem (3.2) takes up the balance of this section; indeed the
Theorem follows from Propositions (3.3), (3.5), and (3.7). It proceeds in two steps:
First we prove that it is enough to exhibit non-licci ideals for two specific formats,
(1, 6, 8, 3) and (1, 8, 9, 2), which are not Dynkin. Second we produce such examples,
in fact entire families of them.
(3.3) Proposition. Let Q be a local ring. If there is a resolution format f such that
(1) there exists a grade 3 perfect ideal in Q of format f,
(2) f is not Dynkin, and
(3) every perfect ideal in Q of format f is licci
then (1, 6, 8, 3) or (1, 8, 9, 2) is such a format.
Proof. Assume that there exists at least one resolution format that satisfies conditions (1)–(3). Let f = (1, m, m + n − 1, n) be minimal among all such formats
with regard to the total rank β(f) = 2(m + n) of the modules in the minimal free
resolution. Necessarily, one has m > 4 as the Dynkin format (1, 3, 3, 1) is the only
format with m = 3. It follows that the format f′ = (1, n + 2, m + n − 2, m − 3) is
Dynkin. Indeed, one has β(f′ ) = β(f) − 2 and every perfect ideal of format f′ is by
Proposition (2.2)(a) is directly linked to a perfect ideal of format f and hence licci.
Given the Dynkin formats (3.1), the potential minimal formats f are as follows:
Dynkin format f′
(1, 2l + 1, 2l + 1, 1) for l > 1
(1, 4, l + 3, l) for l > 2
(1, 5, 6, 2)
(1, 5, 7, 3)
(1, 5, 8, 4)
(1, 6, 7, 2)
(1, 7, 8, 2)
Potential minimal format f
(1, 4, 2l + 2, 2l − 1)
(1, l + 3, l + 4, 2)
(1, 5, 7, 3)
(1, 6, 8, 3)
(1, 7, 9, 3)
(1, 5, 8, 4)
(1, 5, 9, 5)
Among the potential minimal formats, (1, 4, 2l + 2, 2l − 1) does not exist for l = 1,
and for l > 2 the format is Dynkin corresponding to D2l+2 . The format fl =
(1, l + 3, l + 4, 2) is for l ∈ {2, 3, 4} Dynkin corresponding to E6 , E7 , and E8 .
Further, the formats (1, 5, 7, 3) and (1, 5, 8, 4) are Dynkin corresponding to E7 and
E8 . Thus, possible minimal formats that are not Dynkin and such that every ideal
of the format is licci are (1, 5, 9, 5), (1, 6, 8, 3), (1, 7, 9, 3) and fl for l > 5.
An ideal of format (1, 5, 9, 5) is by Proposition (2.2)(b) linked to an ideal of
format f5 = (1, 8, 9, 2). An ideal of format (1, 6, 8, 3) is by Corollary (2.3)(c) linked
one of format (1, 7, 9, 3). Similarly, an ideal of format fl is linked to one of format
fl+1 . Thus, if there exists a format that is not Dynkin with the property that every
ideal of that format is licci, then (1, 6, 8, 3) or (1, 8, 9, 2) has that property.
To provide examples of ideals of formats (1, 6, 8, 3) and (1, 8, 9, 2) that are not
licci, we rely on numerical obstructions found by Huneke and Ulrich [8].
6
L. W. CHRISTENSEN, O. VELICHE, AND J. WEYMAN
(3.4) Let k be a field and e > 3; set Q = k[X1 , . . . , Xe ]. Let I be a homogeneous
perfect ideal in Q of grade 3 with minimal free resolution
Q ←−
m
M
Q(−d1,i ) ←−
m+n−1
M
Q(−d2,i ) ←−
Q(−d3,i ) ←− 0 .
i=1
i=1
i=1
n
M
Set Q = Q(X1 ,...,Xe ) and I = I(X1 ,...,Xe ) . Assuming that one has d1,1 6 d1,2 6 · · · 6
d1,m and d3,1 6 · · · 6 d3,n it follows from [8, cor. 5.13] that if the inequality
d3,n 6 2d1,1
holds, then the ideal I is not licci.
Notice that if Q/I is Artinian, i.e. e = 3, then one has Q/I ∼
= Q/I and the
inequality says that 3 plus the socle degree of Q/I does not exceed 2 times the
initial degree of I.
Resolution format (1, 6, 8, 3).
(3.5) Proposition. Adopt the notation from (3.4). Let Φ be a 2 × 4 matrix of
linear forms in Q or a 3 × 3 symmetric matrix of linear forms, and let I be the ideal
of 2 × 2 minors of Φ. The ideal I has resolution format (1, 6, 8, 3) and it is not licci.
Proof. The graded minimal free resolution of Q/I has the form
Q ←− Q(−2)6 ←− Q(−3)8 ←− Q(−4)3 ←− 0 ;
see Eagon and Nothcott [5, thm. 2] and Józefiak [9, thm. 3.1] or Weyman [14,
prop. (6.1.7) and thm. (6.3.1)]. It follows from (3.4) that I is not licci.
The next example illustrates Proposition (3.5) and the linkage argument in the
proof of Proposition (3.3). Recall that an Artinian local ring is called level if the
socle is exactly the highest nonvanishing power of the maximal ideal.
(3.6) Example. Let k be a field; set Q = k[X, Y, Z] and N = (X, Y, Z). The
ideal N 2 is generated by the six quadratic monomials, and the quotient Q/N 2 is
an Artinian local level algebra of socle degree 1 and type 3, so the minimal free
resolution over Q is
Q ←− Q(−2)6 ←− Q(−3)8 ←− Q(−4)3 ←− 0 ;
Notice that N 2 is the ideal of 2 × 2 minors of either matrix
X Y Z
X Y Z 0
or Y 0 X .
0 X Y Z
Z X Y
The linked ideal
I = (X 2 , Y 2 , Z 3 ) : N 2 = (X 2 , Y 2 , XY Z, XZ 2 , Y Z 2 , Z 3 )
defines a local Artinian k-algebra with graded basis
1
x, y, z
xy, xz, yz, z 2 .
Evidently, there are no linear forms in the socle of Q/I, so it is a level algebra of
type 4, whence I has resolution format (1, 6, 9, 4). Next, the ideal
J = (X 3 − Y Z 2 , Y 3 − XZ 2 , Z 3 ) : I
= (X 3 − Y Z 2 , Y 3 − XZ 2 , Z 3 , X 2 Y 2 , X 2 Y Z, XY 2 Z, XY Z 2 )
FREE RESOLUTIONS OF DYNKIN FORMAT
7
defines a local Artinian k-algebra with graded basis
1
x, y, z
x2 , xy, xz, y 2 , yz, z 2
x2 y, x2 z, xy 2 , xyz, xz 2 , y 2 z, yz 2
x2 z 2 , y 2 z 2 .
It is straightforward to verify that the socle of Q/J is generated by xyz, x2 z 2 , and
y 2 z 2 . It follows that J has resolution format (1, 7, 9, 3).
Resolution format (1, 8, 9, 2). Recall that an Artinian k-algebra is called compressed if it has maximal length among all algebras with the same socle polynomial;
see (4.1). This notion has a natural generalization to Cohen–Macaulay algebras,
see Fröberg and Laksov [6], and that is the one we use in the proposition below.
(3.7) Proposition. Adopt the notation from (3.4). Assume that the graded kalgebra Q/I is compressed with with socle polynomial 2t3 ; that is,
(
2 for i = 6
Q
dimk (Tor3 (Q/I, k))i =
0 for i 6= 6 .
The ideal I has resolution format (1, 8, 9, 2) and it is not licci.
Proof. By [6, prop. 6] the Hilbert series of Q/I is
1 + 3t + 6t2 + 2t3
.
(1 − t)e−3
Further, it follows from the equality 32 = 2 31 that Q/I is extremely compressed,
see [6, p. 133]. Now it follows from [6, prop. 16] that the graded free resolution of
Q/I has the form
Q ←− Q(−3)8 ←− Q(−4)9 ←− Q(−6)2 ←− 0 ;
The conclusion now follows from (3.4).
By work of Boij and Laksov [2, thm. (3.4)], Proposition (3.7) applies to generic
Artinian level algebras of embedding dimension 3, socle degree 3, and type 2.
The next example illustrates Proposition (3.7) and the linkage argument in the
proof of Proposition (3.3).
(3.8) Example. Let k be a field. The ideal
I = (X 3 , X 2 Y + Y Z 2 , X 2 Z + XY Z, XY 2 + XY Z, XZ 2 , Y 3 , Y 2 Z, Z 3 )
in k[X, Y, Z] defines an Artinian local k-algebra with graded basis
1
x, y, z
x2 , xy, xz, y 2 , yz, z 2
xyz, yz 2 .
It is elementary to check that there are no quadratic forms in the socle of Q/I, so it
is a compressed level algebra of type 2, and it follows that I has format (1, 8, 9, 2).
The linked ideal
J = (X 3 , Y 3 , Z 3 ) : I = (X 3 , X 2 Y − Y Z 2 , XY Z − XZ 2 − Y 2 Z, Y 3 , Z 3 )
defines an Artinian local k-algebra with graded basis
1
x, y, z
x2 , xy, xz, y 2 , yz, z 2
x2 z, xy 2 , xz 2 , y 2 z, yz 2 .
Again it is elementary to verify that there are no quadratic forms in the socle, so it
is a level algebra of type 5; in particular, the resolution format of J is (1, 5, 9, 5).
Finally, notice that under the flat extension k[X, Y, Z] → k[X, Y, Z, X4 , . . . , Xe ]
the ideal I extends to an ideal that satisfies the hypothesis of Proposition (3.7).
8
L. W. CHRISTENSEN, O. VELICHE, AND J. WEYMAN
4. Evidence for Part II
As discussed in the Introduction, it is known that grade 3 perfect ideals of resolution
format (1, 4, n + 3, n) for n > 2—almost complete intersection ideals—and ideals
of format (1, m, m, 1) for odd m > 3—Gorenstein ideals—are licci. The remaining Dynkin formats are (1, 5, 6, 2) corresponding to E6 , (1, 6, 7, 2) and (1, 5, 7, 3)
corresponding to E7 , and (1, 7, 8, 2) and (1, 5, 8, 4) corresponding to E8 ; see (3.1).
In the previous section, we proved the existence of ideals that are not licci by
invoking the numerical obstructions in (3.4). In this section, we prove that homogeneous grade 3 perfect ideals in k[X1 , X2 , X3 ] of resolution formats corresponding
to E6 , E7 , or E8 avoid this obstruction.
(4.1) Let Q be a polynomial algebra of embedding dimension e. For a finitely
generated graded Q-module M , let hM (·) denote the Hilbert function. Let I be
a homogeneous
P ideal in Q such that Q/I is Artinian with socle degree s; its socle
polynomial sj=1 cj tj is the Hilbert series of the socle of Q/I. For every i > 0
there is an inequality
hQ/I (i) 6 min{hQ (i),
s
X
cj hQ (j − i)}
j=i
s
e − 1 + i X
e−1+j−i
cj
.
= min
,
e−1
e−1
j=i
If equality holds for P
every i, then Q/I is called compressed, see [6, sec. 3].
The number n = sj=1 cj is the type of Q/I and, evidently, one has
(4.1.1)
hQ/I (i) 6 min{hQ (i), nhQ (s − i)}
e−1+i
e−1+s−i
= min
, n
.
e−1
e−1
(4.2) Proposition. Adopt the notation from (3.4) and let e = 3. If the ideal I has
initial degree d1,1 > 2 and resolution format (1, m, m + 1, 2) for 4 6 m 6 7, then
one has d3,2 > d1,1 + d1,2 .
As one has d1,1 + d1,2 > 2d1,1 it follows that ideals I of format (1, 5, 6, 2),
(1, 6, 7, 2), and (1, 7, 8, 2) avoid the numerical obstruction to being licci from (3.4).
Proof. To simplify the notation from (3.4), set
R = Q/I ,
d = d1,1 ,
d′ = d1,2 ,
and s = d3,2 − 3 .
Notice that R is Artinian of socle degree s. The inequalities 2 6 d1,1 and d1,1 + 2 6
d3,2 yield
16s
and d 6 s + 1 .
We assume that s + 3 6 d + d′ holds and aim for a contradiction. To this end
it is by (4.1.1) sufficient to prove that there is a u 6 s such that the inequality
hR (u) > 2hQ (s − u) holds.
FREE RESOLUTIONS OF DYNKIN FORMAT
Case 1. Assume that d <
s+3
2
9
holds. One then has d′ > s + 3 − d > d, and hence
hR (i) > hQ (i) − hQ (i − d) − (m − 1)hQ (i − d′ )
> hQ (i) − hQ (i − d) − (m − 1)hQ (i − s − 3 + d)
> hQ (i) − hQ (i − d) − 6hQ (i − s − 3 + d) .
For d = 2 one has
s+2
s
hR (s) > hQ (s) − hQ (s − 2) =
−
= 2s + 1 > 2 = 2hQ (0) .
2
2
Now let d > 3 and set u = s + 2 − ⌈ d2 ⌉; notice that u 6 s. Straightforward
computations yield
(
1
d(d − 2)
for d even
2hQ (s − u) = 41
4 (d + 1)(d − 1) for d odd
and
hQ (u) − hQ (u − d) − 6hQ (u − s − 3 + d) =
(
1
4 d(4s − 7d + 8)
1
4 (d(4s − 7d + 12)
for d even
+ 3) for d odd.
For d even one now has
d(4s − 7d + 8) d(d − 2)
−
4
4
d(2(s − 2d + 3) − 1)
=
2
d
> ,
2
hR (u) − 2hQ (s − u) >
and for d odd
d(4s − 7d + 12) + 3 (d + 1)(d − 1)
−
4
4
= d(s − 2d + 3) + 1
hR (u) − 2hQ (s − u) >
>d+1.
That is, independent of the parity of d one has hR (u) > 2hQ (s − u).
Case 2. Assume that d > s+3
2 holds; one has
hR (i) > hQ (i) − mhQ (i − d) > hQ (i) − 7hQ (i − d) .
Set u = ⌊ s+d
2 ⌋ and σ = s − d + 3. Notice that one has
1 6 σ 6 d 6 s + 1 and u 6 s .
Straightforward computations yield
(
2hQ (s − u) =
1
2
4 (σ − 1)
1
4 σ(σ + 2)
for s + d even
for s + d odd
and
hQ (u) − 7hQ (u − d) =
(
1
4 (2d(d
1
4 (2d(d
+ σ) − 3(σ 2 − 1))
for s + d even
+ σ − 1) − 3σ(σ − 2)) for s + d odd.
10
L. W. CHRISTENSEN, O. VELICHE, AND J. WEYMAN
The inequality d > σ explains the second inequality in each of the computation
below. For s + d even one has
2d(d + σ) − 3(σ 2 − 1) σ 2 − 1
−
4
4
4σ 2 − 4(σ 2 − 1)
>
4
=1,
hR (u) − 2hQ (s − u) >
and for s + d odd
2d(d + σ − 1) − 3σ(σ − 2) σ(σ + 2)
−
4
4
2σ(2σ − 1) − 4σ(σ − 1)
>
4
σ
= .
2
hR (u) − 2hQ (s − u) >
That is, independent of the parity of d one has hR (u) > 2hQ (s − u).
(4.3) Proposition. Adopt the notation from (3.4) and let e = 3. If the ideal I has
initial degree d1,1 > 2 and resolution format (1, 5, n + 4, n) for 1 6 n 6 4, then one
has d3,2 > d1,1 + d1,2 .
As one has d1,1 + d1,2 > 2d1,1 it follows, in particular, that ideals I of format
(1, 5, 7, 3) and (1, 5, 8, 4) avoid the numerical obstruction to being licci from (3.4).
Proof. As in the proof of Proposition (4.2), set R = Q/I, d = d1,1 , d′ = d1,2 , and
s = d3,2 − 3. Note that R is Artinian of socle degree s, and that one has 1 6 s and
d 6 s + 1. We assume that s + 3 6 d + d′ holds and aim for a contradiction. To this
end it is by (4.1.1) sufficient to prove that there is a u 6 s such that the inequality
hR (u) > nhQ (s − u) holds.
′
Case 1. Assume that d 6 s+3
2 holds. One then has d > s + 3 − d > d, and hence
hR (i) > hQ (i) − hQ (i − d) − 4hQ (i − d′ )
> hQ (i) − hQ (i − d) − 4hQ (i − s − 3 + d) .
For d = 2 one has s > 2 and hR (s) > hQ (s) − hQ (s − 2) = 2s + 1 > 4 > nhQ (0) as
computed in the proof of Proposition (4.2).
Now let d > 3 and set u = s + 2 − ⌈ d2 ⌉; notice that u 6 s. Straightforward
computations yield
(
1
d(d − 2)
for d even
nhQ (s − u) 6 4hQ (s − u) = 12
2 (d + 1)(d − 1) for d odd
and
hQ (u) − hQ (u − d) − 4hQ (u − s − 3 + d) =
(
1
2 d(2s − 3d + 5)
1
2 (d(2s − 3d + 6)
for d even
+ 1) for d odd.
FREE RESOLUTIONS OF DYNKIN FORMAT
11
For d even one now has
d(2s − 3d + 5) d(d − 2)
−
2
2
d(2(s − 2d + 3) + 1)
=
2
3d
>
,
2
hR (u) − nhQ (s − u) >
and for d odd
d(2s − 3d + 6) + 1 (d + 1)(d − 1)
−
2
2
2d(s − 2d + 3) + 2
=
2
>d+1.
hR (u) − nhQ (s − u) >
That is, independent of the parity of d one has hR (u) > nhQ (s − u).
Case 2. Assume that d > s+3
2 holds; one has
hR (i) > hQ (i) − 5hQ (i − d) .
Set u =
⌊ s+d
2 ⌋
and σ = s − d + 3. Notice that one has
26σ+16d6s+1
and u 6 s .
Straightforward computations yield
nhQ (s − u) 6 4hQ (s − u) =
(
1
2
2 (σ − 1)
1
2 σ(σ + 2)
for s + d even
for s + d odd
and
hQ (u) − 5hQ (u − d) =
(
1
2 (d(d
1
2 (d(d
+ σ) − (σ 2 − 1))
for s + d even
+ σ − 1) − σ(σ − 2)) for s + d odd.
The inequality d > σ+1 explains the second inequality in each of the computation
below. For s + d even one has
d(d + σ) − (σ 2 − 1) σ 2 − 1
−
2
2
(σ + 1)(2σ + 1) − 2(σ 2 − 1)
>
2
3(σ + 1)
=
,
2
hR (u) − nhQ (s − u) >
and for s + d odd
d(d + σ − 1) − σ(σ − 2) σ(σ + 2)
−
2
2
(σ + 1)2σ − 2(σ 2 − σ)
>
2
= 2σ .
hR (u) − nhQ (s − u) >
That is, independent of the parity of d one has hR (u) > nhQ (s − u).
12
L. W. CHRISTENSEN, O. VELICHE, AND J. WEYMAN
Appendix A. Realizability of resolution formats
In the power series ring k[[X, Y, Z]] there exists by [3, prop. 6.2] an ideal of resolution
format (1, m, m, 1), i.e. a Gorenstein ideal of grade 3, for every odd m > 3. We
start by recording that this is the case in any local ring of sufficient depth.
(A.1) Lemma. Let Q be a local ring of depth at least 3. For every odd m > 3
there exists a grade 3 perfect ideal of resolution format (1, m, m, 1).
Proof. Denote by n the maximal ideal of Q and let x, y, z be a regular sequence
in n. Let l be an integer with m = 2l + 1 and define a skew symmetric matrix Vl
as in [4, sec. 3]. Let I be the ideal generated by the sub-maximal Pfaffians of Vl .
It follows from Proposition 3.3 in loc. cit. that the radical of I contains x, y, and
z, so I has grade (at least) 3, whence by [3, thm. 2.1] it is a Gorenstein ideal of
format (1, m, m, 1).
(A.2) Theorem. Let (Q, n) be a local ring of depth at least 3. For every resolution
format f = (1, m, m + n − 1, n) with m > 3 and n > 1 that is not (1, m, m, 1) with
m even and not (1, 3, n + 2, n) with n > 2 there exists a grade 3 perfect ideal I ⊆ n2
of resolution format f.
Proof. By Lemma (A.1) there exists for every odd g > 3 a grade 3 perfect ideal
with resolution format (1, g, g, 1). Fix a format f = (1, m, m + n − 1, n) with m > 3
and n > 2. If m = 4 and n is even, then it follows from Proposition (2.2)(b) that
there exists an ideal of format f that is linked to one of format (1, n + 3, n + 3, 1). If
m = 4 and n is odd, then it follows from Proposition (2.2)(a) that there exists an
ideal of format f that is linked to one of format (1, n + 2, n + 2, 1). If m > 5 is odd,
then it follows from n − 1 applications of Corollary (2.3)(b) that there is an ideal of
format f that is linked to one of format (1, m, m, 1). If m > 6 is even, then it follows
from Corollary (2.3)(a) and n−2 applications of Corollary (2.3)(b) that there exists
an ideal of format f that is linked to one of format (1, m − 1, m − 1, 1).
References
[1] Luchezar L. Avramov, Andrew R. Kustin, and Matthew Miller, Poincaré series of modules
over local rings of small embedding codepth or small linking number, J. Algebra 118 (1988),
no. 1, 162–204. MR0961334
[2] Mats Boij and Dan Laksov, Nonunimodality of graded Gorenstein Artin algebras, Proc. Amer.
Math. Soc. 120 (1994), no. 4, 1083–1092. MR1227512
[3] David A. Buchsbaum and David Eisenbud, Algebra structures for finite free resolutions,
and some structure theorems for ideals of codimension 3, Amer. J. Math. 99 (1977), no. 3,
447–485. MR0453723
[4] Lars Winther Christensen, Oana Veliche, and Jerzy Weyman, Trimming a Gorenstein ideal,
J. Commut. Algebra, to appear. Preprint arXiv:1512.02720v2 [math.AC]; 11 pp.
[5] Jack A. Eagon and Douglas G. Northcott, Ideals defined by matrices and a certain complex
associated with them, Proc. Roy. Soc. Ser. A 269 (1962), 188–204. MR0142592
[6] Ralf Fröberg and Dan Laksov, Compressed algebras, Complete intersections (Acireale, 1983),
Lecture Notes in Math., vol. 1092, Springer, Berlin, 1984, pp. 121–151. MR0775880
[7] Evgeniy S. Golod, A note on perfect ideals, Algebra (A.I. Kostrikin, Ed.), Moscow State
Universit Press, 1980, pp. 37–39.
[8] Craig Huneke and Bernd Ulrich, The structure of linkage, Ann. of Math. (2) 126 (1987),
no. 2, 277–334. MR0908149
[9] Tadeusz Józefiak, Ideals generated by minors of a symmetric matrix, Comment. Math. Helv.
53 (1978), no. 4, 595–607. MR0511849
FREE RESOLUTIONS OF DYNKIN FORMAT
13
[10] Ernst Kunz, Almost complete intersections are not Gorenstein rings, J. Algebra 28 (1974),
111–115. MR0330158
[11] Matthew Miller and Bernd Ulrich, Linkage and compressed algebras, Proceedings of the
conference on algebraic geometry (Berlin, 1985), Teubner-Texte Math., vol. 92, Teubner,
Leipzig, 1986, pp. 267–275. MR0922918
[12] Junzo Watanabe, A note on Gorenstein rings of embedding codimension three, Nagoya Math.
J. 50 (1973), 227–232. MR0319985
[13] Jerzy Weyman, Generic free resolutions and root systems, Ann. Inst. Fourier, to appear.
Preprint arXiv:1609.02083 [math.AC].
[14] Jerzy Weyman, Cohomology of vector bundles and syzygies, Cambridge Tracts in Mathematics, vol. 149, Cambridge University Press, Cambridge, 2003. MR1988690
Texas Tech University, Lubbock, TX 79409, U.S.A.
E-mail address: [email protected]
URL: http://www.math.ttu.edu/~ lchriste
Northeastern University, Boston, MA 02115, U.S.A.
E-mail address: [email protected]
University of Connecticut, Storrs, CT 06269, U.S.A.
E-mail address: [email protected]
URL: http://www.math.uconn.edu/~ weyman
| 0math.AC
|
Maximizing the information learned from finite
data selects a simple model
Henry H. Mattinglya , Mark K. Transtrumb , Michael C. Abbottc,1 , and Benjamin B. Machtad,1
a
Lewis-Sigler Institute and Department of Chemical and Biological Engineering, Princeton University, Princeton, NJ 08544, USA; b Department of Physics and Astronomy,
Brigham Young University, Provo, Utah 84602, USA; c Marian Smoluchowski Institute of Physics, Jagiellonian University, Ulica Łojasiewicza 11, 30-348 Kraków, Poland;
d
Lewis-Sigler Institute and Department of Physics, Princeton University, Princeton, NJ 08544, USA
arXiv:1705.01166v3 [physics.data-an] 14 Feb 2018
PNAS (7 February 2018) + SI ≈ arXiv:1705.01166v3
Earlier versions had the title “Rational Ignorance: Simpler Models Learn More Information from Finite Data”
We use the language of uninformative Bayesian prior choice to study
the selection of appropriately simple effective models. We advocate
for the prior which maximizes the mutual information between parameters and predictions, learning as much as possible from limited data.
When many parameters are poorly constrained by the available data,
we find that this prior puts weight only on boundaries of the parameter manifold. Thus it selects a lower-dimensional effective theory in a
principled way, ignoring irrelevant parameter directions. In the limit
where there is sufficient data to tightly constrain any number of parameters, this reduces to Jeffreys prior. But we argue that this limit is
pathological when applied to the hyper-ribbon parameter manifolds
generic in science, because it leads to dramatic dependence on effects invisible to experiment.
The theory and the experiment are together described by a
probability distribution p(x|θ), for each value of the theory’s
parameters θ ∈ Θ. This function encodes both the quality and
quantity of data to be collected.
The mutual information between the parameters and their
expected data is defined as MI = I(X; Θ) = S(Θ) − S(Θ|X),
where S is the Shannon entropy (23). The MI thus quantifies
the information which can be learned about the parameters
by measuring the data, or equivalently, the information about
the data which can be encoded in the parameters (24, 25).
Defining p? (θ) by maximizing this, we see:
i. The prior p? (θ) is almost always discrete (26–30), with
weight only on a finite number
PK K of points, or atoms
(Figures 1 and 2): p? (θ) = a=1 λa δ(θ − θa ).
Effective Theory | Model Selection | Renormalization Group | Bayesian
Prior Choice | Information Theory
ii. When data is abundant, p? (θ) approaches Jeffreys prior
pJ (θ) (31–33). As this continuum limit is approached,
the proper spacing of the atoms shrinks as a power law
(Figure 3).
P
hysicists prefer simple models not because nature is simple, but because most of its complication is usually irrelevant. Our most rigorous understanding of this idea comes from
the Wilsonian renormalization group (1–3), which describes
mathematically the process of zooming out and losing sight
of microscopic details. These details only influence the effective theory which describes macroscopic observables through
a few relevant parameter combinations, such as the critical
temperature, or the proton mass. The remaining irrelevant
parameters can be ignored, as they are neither constrained
by past data nor useful for predictions. Such models can now
be understood as part of a large class called sloppy models
(4–14), whose usefulness relies on a similar compression of a
large microscopic parameter space down to just a few relevant
directions.
This justification for model simplicity is different from the
one more often discussed in statistics, motivated by the desire
to avoid overfitting (15–21). Since irrelevant parameters have
an almost invisible effect on predicted data, they cannot be
excluded on these grounds. Here we motivate their exclusion
differently: we show that simplifying a model can often allow it
to extract more information from a limited data set, and that
this offers a guide for choosing appropriate effective theories.
We phrase the question of model selection as part of the
choice of a Bayesian prior on some high-dimensional parameter space. In a set of nested models, we can always move
to a simpler model by using a prior which is nonzero only on
some subspace. Recent work has suggested that interpretable
effective models are typically obtained by taking some parameters to their limiting values, often 0 or ∞, thus restricting to
lower-dimensional boundaries of the parameter manifold (22).
Our setup is that we wish to learn about a theory by
performing some experiment which produces data x ∈ X.
iii. When data is scarce, most atoms lie on boundaries of
parameter space, corresponding to effective models with
fewer parameters (Figure 4). The resulting distribution of
weight along relevant directions is much more even than
that given by Jeffreys prior (Figure 5).
Significance Statement
Most physical theories are effective theories, descriptions at
the scale visible to our experiments which ignore microscopic
details. Seeking general ways to motivate such theories, we
find an information theory perspective: if we select the model
which can learn as much information as possible from the
data, then we are naturally led to a simpler model, by a path
independent of concerns about overfitting. This is encoded as
a Bayesian prior which is nonzero only on a subspace of the
original parameter space. We differ from earlier prior selection
work by not considering infinitely much data. Having finite data
is always a limit on the resolution of an experiment, and in our
framework this selects how complicated a theory is appropriate.
H.H.M. and M.C.A. performed the numerical experiments. H.H.M., M.K.T., M.C.A. and B.B.M. designed the research, interpreted results, and contributed to writing. M.C.A. and B.B.M. led the
writing of the paper.
Present affiliation for H.H.M.: Department of Molecular Cellular and Developmental Biology, Yale
University, New Haven, CT, 06520
Present affiliations for B.B.M: Department of Physics, Yale University, New Haven, CT, 06520
Systems Biology Institute, Yale University, West Haven, CT, 06516
1
To whom correspondence should be addressed. E-mail: [email protected] and
[email protected]
A
B
C
Fig. 1. Optimal priors for the Bernoulli model (1). Red lines indicate the positions
of delta functions in p? (θ), which are at the maxima of fKL (θ), Eq. (3). As
m → ∞ these coalesce into Jeffreys prior pJ (θ).
After some preliminaries, we demonstrate these properties in
three simple examples, each a stylized version of a realistic
experiment. To see the origin of discreteness, we study the bias
of an unfair coin and the value of a single variable corrupted
with Gaussian noise. To see how models of lower dimension
arise, we then study the problem of inferring decay rates in a
sum of exponentials.
In the Appendix we discuss the alorithms used for finding
p? (θ) (Figures 2 and S3), and we apply some more traditional
model selection tools to the sum of exponentials example
(Figure S1).
Priors and Geometry
Bayes’ theorem tells us how to update our knowledge of θ
upon observing data x, from prior
p(θ) to posterior p(θ|x) =
R
p(x|θ)p(θ)/p(x), where p(x) = dθ p(θ) p(x|θ). In the absence
of better knowledge we must pick an uninformative prior
which codifies our ignorance. The naive choice of a flat prior
p(θ) = const. has undesirable features, in particular making
p(x) depend on the choice of parameterization, through the
measure dθ.
Jeffreys prior pJ (θ) is invariant under changes of parameterizaion, because it is constructedpfrom some properties of
the experiment (34). This pJ (θ) ∝ det gµν is, up to normalization, the volume form arising from the Fisher information
metric (FIM, often -matrix):
Bernardo defined a prior p? (θ) by maximizing the mutual
information between parameters Θ and the expected data X m
from m repetitions, and then a reference prior by taking the
limit m → ∞ (29, 31). Under certain benign assumptions,
this reference prior is exactly Jeffreys prior (31–33), providing
an alternative justification for pJ (θ).
We differ in taking seriously that the amount of data collected is always finite.∗ Besides being physically unrealistic,
the limit m → ∞ is pathological both for model selection
and prior choice. In this limit any number of parameters
can be perfectly inferred, justifying an arbitrarily complicated
model. In addition, in this limit the posterior p(θ|x) becomes
independent of any smooth prior.†
Geometrically, the defining feature of sloppy models is that
they have a parameter manifold with hyper-ribbon structure
(6–9): there are some long directions (corresponding to d relevant, or stiff, parameters) and many shorter directions (D − d
irrelevant, or sloppy, parameter combinations). These lengths
are often estimated using the eigenvalues of gµν , and have
logarithms that are roughly evenly-spaced over many orders
of magnitude (4, 5). The effect of coarse-graining is to shrink
irrelevant directions (here using the technical meaning of irrelevant: a parameter which shrinks under renormalization group
flow) while leaving relevant directions extended, producing a
sloppy manifold (8, 14). By contrast the limit m → ∞ has the
effect of expanding all directions, thus erasing the distinction
between directions longer and shorter than the critical length
scale of (approximately) one standard deviation.
On such a hyper-ribbon, Jeffreys prior has an undesirable
feature: since it is constructed from the D-dimensional notion
of volume, its weight along the relevant directions always
depends on the volume of the D − d irrelevant directions. This
gives it extreme dependence on which irrelevant parameters
are included in the model.‡ The optimal prior p? (θ) avoids
this dependence because it is almost always discrete, at finite
m.§ It puts weight on a set of nearly distinguishable points,
closely spaced along the relevant directions, but ignoring the
irrelevant ones. Yet being the solution to a reparameterizationinvariant optimization problem, the prior p? (θ) retains this
good feature of pJ (θ).
Maximizing the mutual information was originally done to
calculate the capacity of a communication channel, and we
can borrow techniques from rate-distortion theory here: the
algorithms we use were developed there (37, 38), and the discreteness we exploit was discovered several times in engineering
(26–28, 39). In statistics, this problem is more often discussed
as an equivalent minimax problem (40). Discreteness was
∗
~ =
gµν (θ)
Z
~
~
~ ∂ log p(x|θ) ∂ log p(x|θ) .
dx p(x|θ)
µ
ν
∂θ
∂θ
This Riemannian metric defines a P
reparameterization-invariant
D
distance between points, ds2 =
g dθµ dθν . It meaµ,ν=1 µν
sures the distinguishability of the data which θ and θ + dθ are
expected to produce, in units of standard deviations. Repeating an (identical and independently distributed)
experiment
Qm
m times means considering pm (~
x|θ) = j=1 p(xj |θ), which
m
leads to metric gµν
(θ) = m gµν (θ). However the factor mD/2
in the volume is lost by normalizing pJ (θ). Thus Jeffreys prior
depends on the type of experiment, but not the quantity of
data.
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
2
Interned for five years, John Kerrich only flipped his coin 104 times (35). With
computers we can do better, but even the LHC only generated about 1018 bits of
data (36).
†
For simplicity we consider only regular models, i.e. we assume all parameters are
structurally identifiable.
‡
See Figure 5 for a demonstration of this point. For another example, consider a
parameter manifold Θ which is a cone, with Fisher metric ds2 = (50 dϑ)2 +
ϑ2 dΩ2n /4: there is one relevant direction ϑ ∈ [0, 1] of length L = 50, and n
irrelevant directions forming a sphere of diameter ϑ. Then the prior on ϑ alone
~ is p(ϑ) = (n + 1)ϑn , putting most of the weight near to ϑ = 1,
implied by pJ (θ)
dramatically so if n = D − d is large. But since only the relevant direction is visible
to our experiment, the region ϑ ≈ 0 ought to be treated similarly to ϑ ≈ 1. The prior
~ has this property.
p? (θ)
§
We offer both numerical and analytic arguments for discreteness below. The exception
to discreteness is that if there is an exact continuous symmetry, p? (θ) will be constant
along it. For example if our Gaussian model Eq. (2) is placed on a circle (identifying
both θ ∼ θ + 1 and x ∼ x + 1) then the optimum prior is a constant.
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
also observed in other minimax problems (41–43), and later
in directly maximising mutual information (29, 30, 33, 44).
However it does not seem to have been seen as useful, and
none of these papers explicitly find discrete priors in dimension
D > 1, which is where we see attractive properties. Discreteness has been useful, although for different reasons, in the idea
of rational inattention in economics (45, 46). There, market
actors have a finite bandwidth for news, and this drives them
to make discrete choices despite all the dynamics being continuous. Rate-distortion theory has also been useful in several
areas of biology (47–49), and discreteness emerges in a recent
theoretical model of the immune system (50).
We view this procedure of constructing the optimal prior
as a form of model selection, picking out the subspace of Θ
on which p? (θ) has support. This depends on the likelihood
function p(x|θ) and the data space X, but not on the observed
data x. In this regard it is closer to Jeffreys’ perspective
on prior selection than to tools like the information criteria
and Bayes factors, which are employed at the stage of fitting
to data. We discuss this difference at more length in the
Appendix.
log p𝜏(𝜃)
𝜎 = 1/10
2
2
1
e−(x−θ) /2σ .
p(x|θ) = √
2πσ
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
𝜃, with steps Δ𝜃 = 1/300
�
�
log p⭑(𝜃)
dθ p(θ) fKL (θ),
[3]
Z
fKL (θ) = DKL p(x|θ) p(x) =
dx p(x|θ) log
p(x|θ)
p(x)
¶
where DKL is the Kullback–Leibler divergence.
MaximizR
ing MI over all functions p(θ) with dθ p(θ) = 1 gives
fKL (θ) = const. But the maximizing function will not, in
general, obey p(θ) ≥ 0. Subject to this inequality p? (θ) must
satisfy
[1]
{p? (θ) > 0, fKL (θ) = MI} or {p? (θ) = 0, fKL (θ) < MI}
at every θ. With finite data fKL (θ) − MI must be an analytic
function of θ, and therefore must be smooth with a finite
numbers of zeros, corresponding to the atoms of p? (θ) (Figure
1A). See (28, 29, 46) for related arguments for discreteness,
and (41–43) for other approaches.
The number of atoms occurring in p? (θ) increases as the
data improves. For K atoms there is an absolute bound
MI ≤ log K, saturated if they are perfectly distinguishable. In
Figure 2C we observe that the optimal priors instead approach
a line MI → ζ log K, with slope ζ ≈ 0.75. At large L the
length of parameter space is proportional to the number of
distinguishable points, hence MI → log L. Together these
imply K ∼ L1/ζ , and so the average number density of atoms
grows as
ρ0 = K/L ∼ L1/ζ−1 ≈ L1/3 ,
[2]
Repeated
measurements are equivalent to smaller σ (by σ →
√
σ/ m), so we fix m = 1 here. The Fisher metric is gθθ = 1/σ 2 ,
thus L = 1/σ. An optimal prior is shown in Figure 2, and
in Figure 5A along with its implied distribution of expected
data. This p(x) is similar to that implied by Jeffreys prior,
here pJ (θ) = 1.
We calculated p? (θ) numerically in two ways. After discretizing both θ and x, we can use the Blahut–Arimoto (BA)
�
Z
MI = I(X; Θ) =
p
m
,
θ(1−θ)
𝜃, with steps Δ𝜃 = 1/30
◼
◼
algorithm (37, 38). This converges to the global maximum,
which is a discrete distribution: see Figure 2. Alternatively,
using our knowledge that p? (θ) is discrete, we can instead
adjust the positions θa and weights λa of a finite number of
atoms. See the Supplement for more details.
To see analytically why discreteness arises, we write the
mutual information as
thus pJ (θ) = [π θ(1 − θ)]−1 , and
R√
√
proper parameter space length L =
ds2 = π m.
In the extreme case m = 1, the optimal prior is two delta
functions, p? (θ) = 12 δ(θ) + 12 δ(θ − 1), and MI = log 2, exactly
one bit (29, 30, 33). Before an experiment that will only
ever be run once, this places equal weight on both outcomes;
afterwards it records the outcome. As m increases, weight is
moved from the boundary onto interior points, which increase
in number, and ultimately approach the smooth pJ (θ): see
Figures 1 and 2A.
Similar behavior is seen in a second example, in which we
measure one real number x, normally distributed with known
σ about the parameter θ ∈ [0, 1]:
This gives gθθ =
◼
Fig. 2. Convergence of the Blahut–Arimoto algorithm. This is for the oneparameter Gaussian model Eq. (2) with L = 10 (comparable to m = 10 in
Figure 1). On the right θ is discretized into ten times as many points, but pτ (θ)
clearly converges to the same five delta functions.
We begin with some problems with a single bounded parameter, of length L in the Fisher metric. These tractable cases
illustrate the generic behaviour along either short (irrelevant)
or long (relevant, L 1) parameter directions in higherdimensional examples.
Our first example is the Bernoulli problem, in which we
wish to determine the probability θ ∈ [0, 1] that an unfair
coin gives heads, using the data from m trials. It is sufficient
to record the total number of heads x, which occurs with
probability
m!
θx (1 − θ)m−x .
x!(m − x)!
◼
◼
�
One-Parameter Examples
p(x|θ) =
𝜏 = 10 iterations
𝜏 = 100
𝜏 = 104
Exact p⭑(𝜃)
¶
L 1.
[4]
The function fKL (θ) is sometimes called the Bayes risk, as it quantifies how poorly
the prior will perform if θ turns out to be correct. One of the problems equivalent
to maximising the mutual information (40) is the minimax problem for this (see also
Figure 1):
max I(X; Θ) = min max fKL (θ) = min max
p(θ)
p(θ)
θ
R
dθp(θ)DKL [p(x|θ)kq(x)] .
q(x) p(θ)
The distributions we call expected data p(x) are also known as Bayes strategies, i.e.
distributions on X which are the convolution of the likelihood p(x|θ) with some prior
p(θ). The optimal q(x) from this third formulation (with minq(x) . . .) can be shown
to be such a distribution (40).
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
3
B
𝜃
r=
A
0
B
A
p
Bernoulli: L = ⇡ m
r=
=
k2
Gaussian: L = 1/𝜎
k1
C
1
y2
s = 1ê3
s = 1ê7
y1
‡
0.5
ʇ‡
ʇ‡
ʇÊÊÊ
‡‡
‡‡‡‡‡‡‡‡‡ ‡ ‡ ‡‡‡‡‡‡‡‡‡‡‡‡‡‡‡
‡‡ÊÊÊ
‡‡
‡‡
‡‡
‡‡
‡‡‡‡
‡
‡‡
‡
‡‡
‡‡
‡‡
‡‡
‡‡
‡‡‡
Ê
Ê
‡‡
‡‡
ʇÊ
‡Ê‡
ÊÊ
Ê
Ê
d eff
𝜁=1
𝜁 = 3/4
1
0
‡
‡
‡‡‡‡ ‡ ‡
‡‡‡‡
‡‡‡‡
2
Number of atoms K
Ê
ÊÊÊ
‡Ê‡
ʇÊ
ʇ‡
ʇ‡
‡‡‡‡
‡‡‡‡
‡‡‡
‡‡‡
5
Ê
Ê
rá2
Ê
Ê
rá1
rá0
0
2
Ê
Ê
Ê
Bernoulli
Gaussian
Ê
10 20
1ês
Ê
Ê
Ê
Ê
ÊÊ
Ê
s = 1ê19
2
Wr
I(X; ⇥)/ log 2
1
Ê
Ê
Ê Ê
Ê ÊÊ
Ê
ÊÊ
Ê Ê
r=
C
Ê Ê ÊÊ
Ê ÊÊÊÊ
Ê
ÊÊ
Ê Ê
Ê
Ê ÊÊ
s = 1ê50
50 100
1/𝜎
Fig. 3. Behavior of p? (θ) with increasing Fisher length. Panels A and B show
the atoms of p? (θ) for the two one-dimensional models as L is increased (i.e. we
perform more repetitions m or have smaller noise σ ). Panel C shows the scaling of
the mutual A
information (in bits) with the number of atoms K .The dashed line is the
bound MI ≤ log K , and the solid line is the scaling law MI ∼ 3/4 log K .
Thus the proper spacing between atoms shrinks to zero in
the limit of infinite data, i.e. neighboring atoms cease to be
distinguishable.
To derive this scaling law analytically, in a related paper
(51) we consider a field theory for the number density of atoms,
in which the entropy density (omitting numerical factors) is
B
2
S = const. − e−ρ [ρ4 (ρ0 )2 + 1].
From this we find ζ = 3/4, which is consistent with both
examples presented above.
Multi-parameter Example
In the examples above, p? (θ) concentrates weight on the edges
of its allowed domain when data is scarce (i.e. when m is
small or σ is large, hence L is small). We next turn to a multiparameter model in which some parameter combinations are
ill-constrained, and where edges correspond to reduced models.
The physical picture is that we wish to determine the
composition of an unknown radioactive source, from data of xt
Geiger counter clicks at some times t. As parameters we have
the quantities Aµ and decay constants kµ of isotopes µ. The
probability of observing xt should be a Poisson distribution (of
mean yt ) at each time, but we approximate these by Gaussians
to write:‖
p(~x|~
y) ∝
Y
2
e−(xt −yt )
/2σ 2
,
X
yt =
t
‖
Aµ e−kµ t .
[5]
µ
Using a normal distribution of fixed σ here is what allows the metric in Eq. (6) to
be so simple. However the qualitative behavior from the Poisson distribution is very
similar.
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
4
Fig. 4. Parameters and priors for the exponential model (5). Panel A shows the
area of the ~
y plane covered by all decay constants k1 , k2 ≥ 0. Panel B shows the
positions of the delta functions of the optimal prior p? (~
y ) for several values of σ , with
colors indicating the dimensionality r at each point. Panel C shows the proportion of
weight on these dimensionalities.
Fixing σt = σ = const. then brings us to a nonlinear leastsquares model of the type studied by (6, 7). This same model
also arises in other contexts, such as the asymptotic approach
to equilibrium of many dynamical systems (52).
We can see the essential behavior with just two isotopes
in fixed quantities: Aµ = 12 , thus yt = 12 (e−k1 t + e−k2 t ).
Measuring at only two times t1 and t2 , we almost have a
two-dimensional version of Eq. (2), in which the center of the
distribution ~
y = (y1 , y2 ) plays the role of θ above. The mapping between ~k and ~
y is shown in Figure 4A, fixing t2 /t1 = e.
The Fisher information metric is proportional to the ordinary
Euclidean metric for ~
y , but not for ~k:
gµν (~k) =
1 X ∂yt ∂yt
σ2
∂kµ ∂kν
⇐⇒
gst (~
y) =
t
1
δst .
σ2
[6]
Thus Jeffreys prior is a constant on the allowed region of the
~
y plane.
Then we proceed to find the optimum p? (~
y ) for this model,
shown in Figure 4B for various values of σ. When σ is large,
this has delta functions only in two of the corners, allowing
only k1 , k2 = 0 and k1 , k2 = ∞. As σ is decreased, new atoms
appear first along the lower boundary (corresponding to the
one-dimensional model where k1 = k2 ) and then along the
other boundaries. At sufficiently small σ, atoms start filling
in the (two-dimensional) interior.
To show this progression in Figure 4C, we define Ωr as
the total weight on all edges
PD of dimension r, and an effective
dimensionality deff =
r Ωr . This increases smoothly
r=1
from 0 towards D = 2 as the data improves.
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
A
Gaussian model: D = 1
σ = �/��
p(x) from p⭑(𝜃)
�★ (θ)
p(x) from Jeffreys
■
■
■
■
■
Discussion
𝜃
θ
�
B
Exponential
model: D = 2
�
Prior p? (~y)
�★ (�+ )
y+ =
y1 + y2+
p
2
s = 1ê7
σ = �/�
r=0
r=1
■
■
■
■
■
�
to equal decay constants will still have significant weight.
The fact that such edges give human-readable simpler models
(unlike arbitrary submanifolds) was the original motivation
for preferring them in (22), and it is very interesting that our
optimization procedure has the same preference.††
■
0
■
y1 + y2
y+ = p
�+ = (�� +�2� )/ �
�
Fig. 5. Distributions of expected data p(x) from different priors. Panel A is the
one-parameter Gaussian model, with L = 10. Panel B projects the two-parameter
exponential model onto the y1 + y2 direction, for σ = 1/7 where the perpendicular
direction should be irrelevant. The length
is about the same
√ of the relevant direction
p? (~y) of expected
as the one-parameter case: L+ = 7 2. Notice that the distribution
data p(x+ ) from Jeffreys
√prior here is quite different, with almost no weight at the
ends of the range (0 and 2), because this prior still weights the area not the length.
At medium values of σ, the prior p? (~
y ) almost ignores the
width of the parameter
manifold,
and
cares
mostly about its
√
length (L+ = 2/σ along the diagonal). This behavior is very
different to that of Jeffreys prior: in Figure 5B we demonstrate
this by plotting the distributions of data implied by these two
priors. Jeffreys puts almost no weight near the ends of the long
(i.e. stiff, or relevant) parameter’s range, because the (sloppy,
or irrelevant) width happens to be even narrower there than
in the middle. By contrast our effective model puts significant
weight on each end, much like the one-parameter model in
Figure 5A.
The difference between one and two parameters being relevant (in Figure 4B) is very roughly σ = 1/7 to σ = 1/50, a
factor 7 in Fisher length, thus a factor 50 in the number of repetitions m — perhaps the difference between a week’s data and
a year’s. These numbers are artificially small to demonstrate
the appearance of models away from the boundary: more
realistic models often have manifold lengths spread over many
orders of magnitude (5, 8), and thus have some parameters
inaccessible even with centuries of data. To measure these we
need a qualitatively different experiment, justifying a different
effective theory.
The one-dimensional model along the lower edge of Figure
4A is the effective theory with equal decay constants. This
remains true if we allow more parameters k3 , k4 , . . . in Eq. (5),
and p? (~
y ) will still place a similar weight there.∗∗ Measuring xt
also at later times t3 , t4 , . . . will add more thin directions to the
manifold (7), but the one-dimensional boundary corresponding
While the three examples we have studied here are very simple,
they demonstrate a principled way of selecting optimal effective
theories, especially in high-dimensional settings. Following
(45), we may call this rational ignorance.
The prior p? (θ) which encodes this selection is the maximally uninformative prior, in the sense of leaving maximum
headroom for learning from data. But its construction depends on the likelihood function p(x|θ), and thus it contains
knowledge about the experiment through which we are probing
nature. Jeffreys prior pJ (θ) also depends on the experiment,
but more weakly: it is independent of the number of repetitions m, precisely because it is the limit m → ∞ of the
optimal prior (32, 33).
Under either of these prescriptions, performing a second
experiment may necessitate a change in the prior, leading to
a change in the posterior not described by Bayes’ theorem. If
the second experiment is different from the first, then changing to Jeffreys prior for the combined experiment (and then
applying Bayes’ rule just once) will have this effect (56, 57).‡‡
Our prescription differs from Jeffreys in also regarding more
repetitions of an identical experiment as being different. Many
experiments would have much higher resolution if they could
be repeated for all eternity. The fact that they cannot is an
important limit on the accuracy of our knowledge, and our
proposal treats this limitation on the same footing as the rest
of the specification of the experiment.
Keeping m finite is where we differ from earlier work on
prior selection. Bernardo’s reference prior (31) maximizes the
same mutual information, but always in the m → ∞ limit
where it gives a smooth analytically tractable function. Using
I(X; Θ) to quantify what can be learned from an experiment
goes back to Lindley (24). That finite information implies a
discrete distribution was known at least since (26, 27). What
has been overlooked is that this discreteness is useful for
avoiding a problem with Jeffreys prior on the hyper-ribbon
parameter spaces generic in science (5): because it weights the
irrelevant parameter volume, Jeffreys prior has strong dependence on microscopic effects invisible to experiment. The limit
m → ∞ has erased the divide between relevant and irrelevant
parameters, by throwing away the natural length scale on the
parameter manifold. By contrast p? (θ) retains discreteness at
roughly this scale, allowing it to ignore irrelevant directions.
Along a relevant parameter direction this discreteness is no
worse than rounding θ to as many digits as we can hope to
measure, and we showed that in fact the spacing of atoms
decreases faster than our accuracy improves.
††
Edges of the parameter manifold give simpler models not only in the sense of
having fewer parameters, but also in an algorithmic sense. For example, the
Michaelis–Menten model is analytically solvable (53) in a limit which corresponds
to a manifold boundary (54). Stable linear dynamical systems of order n are model
∗∗
boundaries order n + 1 systems (55). Taking some parameter combinations to the
If we have more parameters than measurements then the model must be singular.
extreme can lock spins into Kadanoff blocks (54).
In fact the exponential model of Figure 4 is already slightly singular, since k1 ↔ k2
does not change the data; we could cure this by restricting to k2 ≥ k1 , or by working ‡‡ This view is natural in the objective Bayesian tradition, but see (58) and (59–61) for
with ~
y , to obtain a regular model.
alternative viewpoints.
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
5
Model selection is more often studied not as part of prior
selection, but at the stage of fitting the parameters to data.
From noisy data, one is tempted to fit a model which is more
complicated than reality; avoiding such overfitting improves
predictions. The AIC, BIC (15, 62) and related criteria (19,
20, 63–65) are subleading terms of various measures in the
m → ∞ limit, in which all (nonsingular) parameters of the
true model can be accurately measured. Techniques like MDL,
NML, and cross-validation (63, 66, 67) need not take this limit,
but all are applied after seeing the data. They favor minimally
flexible models close to the data seen, while our procedure
favors one answer which can distinguish as many different
outcomes as possible. It is curious that both approaches can
point towards simplicity. We explore this contrast in more
detail in the Appendix.§§
Being discrete, the prior p? (θ) is very likely to exclude
the true value of the parameter, if such a θtrue ∈ Θ exists.
This is not a flaw: the spirit of effective theory is to focus on
what is relevant for describing the data, deliberately ignoring
microscopic effects which we know to exist (68). Thus the
same effective theory can emerge from different microscopic
physics [as in the universality of critical points describing phase
transitions (69)]. The relevant degrees of freedom are often
quasiparticles [such as the Cooper pairs of superconductivity
(70)] which do not exist in the microscopic theory, but give
a natural and simple description at the scale being observed.
We argued here for such simplicity not on the grounds of the
difficulty of simulating 1023 electrons, nor of human limitations,
but based on the natural measure of information learned.
There is similar simplicity to be found outside of physics.
For example the Michaelis–Menten law for enzyme kinetics
(71) is derived as a limit in which only the ratios of some reaction rates matter, and is useful regardless of the underlying
system. In more complicated systems which we cannot solve
by hand, and for which the symmetries and scaling arguments
used in physics cannot be applied, we hope that our information approach may be useful for identifying the appropriately
detailed theory.
6. M. K. Transtrum, B. B. Machta and J. P. Sethna, Why are Nonlinear Fits
to Data so Challenging?, Phys. Rev. Lett. 104 (2010) 060201
[arXiv:0909.3884].
7. M. K. Transtrum, B. B. Machta and J. P. Sethna, Geometry of nonlinear
least squares with applications to sloppy models and optimization,
Phys. Rev. E83 (2011) 036701 [arXiv:1010.1449].
8. B. B. Machta, R. Chachra, M. K. Transtrum and J. P. Sethna, Parameter
Space Compression Underlies Emergent Theories and Predictive
Models, Science 342 (2013) 604–607 [arXiv:1303.6738].
9. M. K. Transtrum, B. B. Machta, K. S. Brown, B. C. Daniels, C. R. Myers and J. P.
Sethna, Perspective: Sloppiness and emergent theories in physics,
biology, and beyond, J. Chem. Phys. 143 (2015) 010901
[arXiv:1501.07668].
10. T. O’Leary, A. C. Sutton and E. Marder, Computational models in the age
of large datasets, Cur. Op. Neurobio. 32 (2015) 87–94.
11. T. Niksic and D. Vretenar, Sloppy nuclear energy density functionals:
effective model reduction, Phys. Rev. C94 (2016) 024333
[arXiv:1606.08617].
12. D. V. Raman, J. Anderson and A. Papachristodoulou, Delineating
Parameter Unidentifiabilities in Complex Models, arXiv:1607.07705.
13. G. Bohner and G. Venkataraman, Identifiability, reducibility, and
adaptability in allosteric macromolecules, J. Gen. Physiol. 149 (2017)
547–560.
14. A. Raju, B. B. Machta and J. P. Sethna, Information geometry and the
renormalization group, arXiv:1710.05787.
15. H. Akaike, A new look at the statistical model identification, IEEE
Trans. Automat. Contr. 19 (1974) 716–723.
16. N. Sugiura, Further analysts of the data by Akaike’s information
criterion and the finite corrections, Comm. Stat. Th. Meth. 7 (1978)
13–26.
17. V. Balasubramanian, Statistical inference, Occam’s razor, and
statistical mechanics on the space of probability distributions,
Neural Comp. 9 (1997) 349–368.
18. I. J. Myung, V. Balasubramanian and M. A. Pitt, Counting probability
distributions: differential geometry and model selection., PNAS 97
(2000) 11170–11175.
19. D. J. Spiegelhalter, N. G. Best, B. P. Carlin and A. Van Der Linde, Bayesian
measures of model complexity and fit, J. Roy. Stat. Soc. B 64 (2002)
583–639.
20. S. Watanabe, Asymptotic equivalence of Bayes cross validation and
21.
22.
23.
ACKNOWLEDGMENTS.
We thank Vijay Balasubramanian,
William Bialek, Robert de Mello Koch, Peter Grünwald, Jon Machta,
James Sethna, Paul Wiggins, and Ned Wingreen for discussion and
comments. We thank ICTS Bangalore for hospitality.
H.H.M. was supported by NIH grant R01GM107103. M.K.T.
was support by NSF-EPCN 1710727. B.B.M. was supported by a
Lewis-Sigler Fellowship and by NSF PHY 0957573. M.C.A. was
supported by NCN grant 2012/06/A/ST2/00396.
1. L. P. Kadanoff, Scaling laws for Ising models near Tc , Physics 2 (1966)
263–272.
2. K. G. Wilson, Renormalization group and critical phenomena. 1.
Renormalization group and the Kadanoff scaling picture, Phys. Rev.
B4 (1971) 3174–3183.
3. J. L. Cardy, Scaling and renormalization in statistical physics. Cambridge Univ.
Press, 1996.
4. J. J. Waterfall, F. P. Casey, R. N. Gutenkunst, K. S. Brown, C. R. Myers, P. W.
Brouwer and J. P. Sethna, Sloppy-model universality class and the
Vandermonde matrix., Phys. Rev. Lett. 97 (2006) 150601–4.
5. R. N. Gutenkunst, J. J. Waterfall, F. P. Casey, K. S. Brown, C. R. Myers and J. P.
Sethna, Universally Sloppy Parameter Sensitivities in Systems
Biology Models, PLoS Comp. Biol. 3 (2007) e189–8.
§§
Model selection usually starts from a list of models to be compared, in our language
a list of submanifolds of Θ. We can also consider maximising mutual information in
this setting, rather than with an unconstrained function p(θ), and unsurprisingly we
observe a similar preference for highly flexible simpler models. This is also discussed
in the Appendix, at Eq. (S3).
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
6
24.
25.
widely applicable information criterion in singular learning
theory, JMLR 11 (2010) 3571–3594 [arXiv:1004.2316].
P. A. Wiggins and C. H. LaMont, Information-based inference for
singular models and finite sample sizes, arXiv:1506.05855.
M. K. Transtrum and P. Qiu, Model reduction by manifold boundaries,
Phys. Rev. Lett. 113 (2014) 098701–6.
C. E. Shannon, A mathematical theory of communication, Bell Sys.
Tech. J. 27 (1948) 623–656.
D. V. Lindley, On a Measure of the Information Provided by an
Experiment, Ann. Math. Statist. 27 (1956) 986–1005.
A. Rényi, On some basic problems of statistics from the point of
view of information theory, Proc. 5th Berkeley Symp. on Math. Statist.
and Prob. (1967) 531–543. [projecteuclid.org].
26. G. Färber, Die Kanalkapazität allgemeiner Übertragunskanäle bei
begrenztem Signalwertbereich beliebigen Signalübertragungszeiten
sowie beliebiger Störung, Arch. Elektr. Übertr. 21 (1967) 565–574.
27. J. G. Smith, The information capacity of amplitude-and
variance-constrained scalar gaussian channels, Information and
Control 18 (1971) 203–219.
28. S. L. Fix, Rate distortion functions for squared error distortion
measures, Proc. 16th Annu. Allerton Conf. Commun. Control Comput. (1978)
704–711.
29. J. O. Berger, J. M. Bernardo and M. Mendoza, On Priors that Maximize
Expected Information, Recent Developments in Statistics and Their
Applications (1988) 1–20. [www.uv.es/~bernardo/].
30. Z. Zhang, Discrete Noninformative Priors. PhD thesis, Yale University, 1994.
[UMI 9523257].
31. J. M. Bernardo, Reference Posterior Distributions for Bayesian
Inference, J. Roy. Stat. Soc. B 41 (1979) 113–147. [www.uv.es/~bernardo/].
32. B. S. Clarke and A. R. Barron, Jeffreys’ prior is asymptotically least
favorable under entropy risk, J. Stat. Plan. Infer. 41 (1994) 37–60.
33. H. R. Scholl, Shannon optimal priors on independent identically
distributed statistical experiments converge weakly to Jeffreys’
prior, Test 7 (1998) 75–94.
34. H. Jeffreys, An Invariant Form for the Prior Probability in
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
Estimation Problems, Proc. Roy. Soc. A 186 (1946) 453–461.
35. J. E. Kerrich, An Experimental Introduction to the Theory of Probability.
Copenhagen: E Munksgaard, 1946.
36. C. O’Luanaigh, CERN data centre passes 100 petabytes, home.cern
(2013).
37. S. Arimoto, An algorithm for computing the capacity of arbitrary
discrete memoryless channels, IEEE Trans. Inform. Theory 18 (1972)
14–20.
38. R. Blahut, Computation of channel capacity and rate-distortion
functions, IEEE Trans. Inform. Theory 18 (1972) 460–473.
39. K. Rose, A mapping approach to rate-distortion computation and
analysis, IEEE Trans. Inform. Theory 40 (1994) 1939–1952.
40. D. Haussler, A general minimax result for relative entropy, IEEE
Trans. Inform. Theory 43 (1997) 1276–1280.
41. M. N. Ghosh, Uniform Approximation of Minimax Point Estimates,
Ann. Math. Statist. 35 (1964) 1031–1047.
42. G. Casella and W. E. Strawderman, Estimating a Bounded Normal
Mean, Ann. Statist. 9 (1981) 870–878.
43. I. Feldman, Constrained Minimax Estimation of the Mean of the
Normal Distribution with Known Variance, Ann. Statist. 19 (1991)
2259–2265.
44. M. Chen, D. Dey, P. Müller, D. Sun and K. Ye, Frontiers of statistical decision
making and Bayesian analysis. Springer, New York, NY, 2010.
45. C. A. Sims, Rational Inattention: Beyond the Linear-Quadratic
Case, American Economic Review 96 (2006) 158–163.
46. J. Jung, J. Kim, F. Matĕjka and C. A. Sims, Discrete Actions in
Information-Constrained Decision Problems, .
[princeton.edu/~sims/#RIDiscrete].
47. S. Laughlin, A simple coding procedure enhances a neuron’s
information capacity, Z. Naturforsch. 36 c (1981) 910–912.
48. G. Tkačik, C. G. Callan and W. Bialek, Information flow and
optimization in transcriptional regulation, PNAS 105 (2008)
12265–12270 [arXiv:0705.0313].
49. M. D. Petkova, G. Tkačik, W. Bialek, E. F. Wieschaus and T. Gregor, Optimal
decoding of information from a genetic network, arXiv:1612.08084.
50. A. Mayer, V. Balasubramanian, T. Mora and A. M. Walczak, How a
well-adapted immune system is organized, PNAS 112 (2015)
5950–5955 [arXiv:1407.6888].
51. M. C. Abbott and B. B. Machta, An information scaling law: ζ = 3/4,
arXiv:1710.09351.
52. S. H. Strogatz, Nonlinear Dynamics And Chaos. Sarat Book House, 2007.
53. S. Schnell and C. Mendoza, Closed Form Solution for
Time-dependent Enzyme Kinetics, J. Theor. Biol. 187 (1997) 207–212.
54. M. Transtrum, G. Hart and P. Qiu, Information topology identifies
emergent model classes, arXiv:1409.6203.
55. P. E. Paré, A. T. Wilson, M. K. Transtrum and S. C. Warnick, A unified view
model selection, Statistics Surveys 4 (2010) 40–79.
68. P. W. Anderson, More Is Different, Science 177 (1972) 393–396.
69. R. W. Batterman, Philosophical Implications of Kadanoff’s Work on
the Renormalization Group, J. Stat. Phys. 167 (2017) 559–574.
[pitt.edu/~rbatterm/].
70. J. Bardeen, L. N. Cooper and J. R. Schrieffer, Theory of
superconductivity, Phys. Rev. 108 (1957) 1175–1204.
71. L. Michaelis and M. L. Menten, The kinetics of invertin action, FEBS
Letters 587 (2013) 2712–2720. [Translation by T. R. C. Boyde].
72. R. E. Kass and A. E. Raftery, Bayes Factors, J. Amer. Stat. Assoc. 90 (1995)
773.
73. A. Bhadra, J. Datta, N. G. Polson and B. Willard, Default Bayesian
analysis with global-local shrinkage priors, Biometrika 103 (Dec.,
2016) 955–969.
74. D. Simpson, H. Rue, A. Riebler and T. G. Martins, Penalising Model
Component Complexity: A Principled, Practical Approach to
Constructing Priors, Statist. Sci. 32 (2017) 1–28 [arXiv:1403.4630].
75. J. I. Myung, D. J. Navarro and M. A. Pitt, Model selection by normalized
maximum likelihood, Journal of Mathematical Psychology 50 (2006)
167–179.
76. P. D. Grünwald, The Minimum Description Length Principle. MIT Press, 2007.
77. C.-I. Chang and L. D. Davisson, On calculating the capacity of an
infinite-input finite (infinite)-output channel, IEEE Trans. Inform.
Theory 34 (1988) 1004–1010.
78. J. Lafferty and L. Wasserman, Iterative Markov chain Monte Carlo
computation of reference priors and minimax risk, Proc. 17th conf.
Uncert. AI (2001) 293–300 [arXiv:1301.2286].
79. J. Dauwels, Numerical computation of the capacity of continuous
memoryless channels, Proc. 26th Symp. Inf. Th. Benelux (2005).
[www.dauwels.com].
Appendix: Model Selection from Data
The usual discussion of model selection takes place after observing
data x. If we wish to compare some models∗ labeled by d, each
with some prior pd (θ), then one prescription is to choose the model
with the largest p(x). Labelling this explicitly, we write
p(x|d) =
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
pd (θ) > 0 on Θd ⊂ Θ.
[S1]
If the Bayes factor p(x|d)/p(x|d0 ) is larger than one then (absent
any other information) d is preferred over d0 (72).† In the usual
asymptotic limit m → ∞, this idea leads to minimising the Bayesian
information criterion (BIC) (62):
− log p(x|d) ≈ − log p(x|θ̂d ) +
pp. 1989–1994, IEEE, 2015.
56. N. Lewis, Combining independent Bayesian posteriors into a
(2000) 1244–1255.
59. T. Seidenfeld, Why I am not an objective Bayesian; some
reflections prompted by Rosenkrantz, Theor Decis 11 (1979) 413–440.
60. R. E. Kass and L. Wasserman, The selection of prior distributions by
formal rules, J. Amer. Stat. Assoc. 91 (1996) 1343–1370.
61. J. Williamson, Objective Bayesianism, Bayesian conditionalisation
and voluntarism, Synthese 178 (2009) 67–85.
62. G. Schwarz, Estimating the dimension of a model, Ann. Statist. 6
(1978) 461–464.
63. J. Rissanen, Modeling by Shortest Data Description, Automatica 14
(1978) 465–471.
64. C. S. Wallace and D. M. Boulton, An Information Measure for
Classification, The Computer Journal 11 (1968) 185–194.
65. S. Watanabe, A widely applicable Bayesian information criterion,
JMLR 14 (2013) 867–897 [arXiv:1208.6338].
66. P. D. Grünwald, I. J. Myung and M. A. Pitt, Advances in Minimum Description
Length: Theory and Applications. 2009.
67. S. Arlot and A. Celisse, A survey of cross-validation procedures for
dθ p(x|θ) pd (θ),
Θd
of Balanced Truncation and Singular Perturbation
Approximations, in Proceedings of the American Control Conference,
confidence distribution, with application to estimating climate
sensitivity, J. Stat. Plan. Infer. (2017) to appear.
57. N. Lewis, Modification of Bayesian Updating where Continuous
Parameters have Differing Relationships with New and Existing
Data, arXiv:1308.2791.
58. D. Poole and A. E. Raftery, Inference for Deterministic Simulation
Models: The Bayesian Melding Approach, J. Amer. Stat. Assoc. 95
Z
where − log p(x|θ̂d ) =
1 2
χ
2
=
1
2
Pm
i=1
d
log m + O(m0 )
2
[xi − yi (θ̂d )]2 /σ 2 ∼ O(m),
and θ̂d is a maximum likelihood estimator for x, constrained to the
appropriate subspace:
θ̂d (x) = argmax p(x|θ).
θ∈Θd
The term d log m penalises models with more parameters, even
though they can usually fit the data more closely. Despite the name
this procedure is not very Bayesian: one chooses the effective model
(and hence the prior) after seeing the data, rather than simply
updating according to Bayes theorem.‡
∗
†
The word model unfortunately means several things in the literature. We mean
parameter space Θd always equipped with a likelihood function p(x|θ), and usually
with a prior pd (θ). When this is a subspace of some larger model ΘD (whose
likelihood function agrees, but whose prior may be unrelated) then we term the
smaller one an effective model, or a reduced model, although we do not always write
the adjective. The optimal prior p? (θ) defines an effective model in this sense. Its
support will typically be on several boundaries of ΘD . If the boundaries of ΘD (of
all dimensions) are regarded as a canonical list of reduced models, then p? (θ) is
seldom a sub-model of any one of them.
R
If one of the priors is improper, say
dθ pd (θ) = ∞, then p(x|d) will also be
infinite. In this sense the Bayes factor behaves worse than the posterior p(θ|x),
which can still be finite.
‡
Terms penalising more complex models can be translated into shrinkage priors, which
concentrate weight near to simpler models (73). Perhaps the shrinkage priors closest
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
7
Related prescriptions can be derived from minimum description length (MDL) ideas. To allow reconstruction of the data we
transmit both the fitted parameters and the residual errors, and
minimising the (compressed) length of this transmission drives a
tradeoff between error and complexity (63, 64, 66). A convenient
version of this goes by the name of normalized maximum likeihood
(NML) (75, 76), and chooses the model d which maximizes
pNML
(x) =
d
Bayes Factor Comparison
A
1
0
0
0
1
x1
Full model,
pJ(𝜃)
0
x1
Optimal
p⭑(𝜃)
1
NML
1
x2
1
0
0
0
1
x1
Full model,
pJ(𝜃)
0
x1
Three edges,
separately
Zd =
,
Zd
Z
dx0 p x0 |θ̂d (x0 ) .
[S2]
Θd
• Figure S1A compares three models: the complete model (with
Jeffreys prior), the optimal model described by discrete prior
p? (~
y ), and an even simpler model with weight only on the
three vertices ~
y = (0, 0), ( 21 , 12 ), (1, 1).
• Figure S1B instead compares the complete model to three
different one-parameter models (along the three boundaries
of the allowed region of the ~
y plane) and a zero-parameter
model (one point ~
y , in no particularly special place). In terms
of decay rates the three lines are limits k1 = k2 , k1 = 0 and
k2 = ∞.
Corners
only
Bayes
B
This is not Bayesian in origin, and does not depend on the prior on
each effective model d, only its support Θd . The function pNML
(x)
d
is not expected data in the sense of p(x) — it is not the convolution
of the likelihood with any prior.§ In the asymptotic limit pNML
(x)
d
approaches p(x) from Jeffreys prior, and this criterion agrees with
BIC (63), but away from this limit they differ.
In Figure S1 we apply these two prescriptions to the exponential
example treated in the text. At each ~
x ∈ X we indicate which
of a list of models is preferred.¶ Figure S2 instead draws the
distributions being used.
NML Comparison
x2
1
p x|θ̂d (x)
1
One point
Fig. S1. Model selection from data point x. Each large panel here shows which
of a list of effective models is preferred after observing data x ∈ X . On the left the
criterion is maximizing Eq. (S1), on the right the criterion is Eq. (S2). We study the
same exponential model considered above, with σ = 1/10. Panel A compares our
optimal effective model with prior p? (θ) to the full model (with Jeffeys prior) and to an
even simpler model whose prior is just three delta functions. (These are drawn in the
legend below). Panel B compares the full model to three different one-dimensional
models, each allowing only one edge of Θ (with a uniform prior along this, i.e. the
one-dimensional Jeffreys prior) and also to a trivial model (just one point), again with
colors as indicated just below.
Different effective models are preferred for different values of data x.
At a given point x, if several models contain the same θ̂(x) then the
simplest among them is preferred, which in the NML case means
precisely the one with the smallest denominator Zd . In fact a trivial
model consisting of just one point Θ0 = θ̂(x) would always be
preferred if it were among those considered — there is no automatic
preference for models which can produce a wide range of possible
data.
By contrast our prior selection approach aims to be able to
distinguish as many possible outcomes in X as possible. Applied
to the same list of models as in Figure S1, this gives the following
fixed scores (base e):
Ifull = 1.296,
and
I? = 1.630,
Icorners = 1.098
Iupper = 0.852,
Ilower-left = 0.845,
Ilower-right = 1.418,
Ione-point = 0.
[S3]
to this paper’s are the penalised complexity priors of (74). Those are also reparameterization invariant, and also concentrate weight on a subspace of Θ, often a boundary.
However both the subspace (or base model) and the degree of concentration (scaling
parameter) are chosen by hand, rather than being deduced from p(x|θ).
§
This relevant optimization problem here can be described as minimizing worst-case
expected regret, written (75)
NML
pd
p(x|θ̂d (x))
,
q(x)
= argmin max log
x
q
θ̂d (x) ∈ Θd .
Perhaps the closest formulation of our maximum mutual information problem is that
our p? (x), the distribution on X not the prior, can be found as follows (40):
p? = argmin max
q∈B
θ
R
X
dx p(x|θ) log
p(x|θ)
q(x)
where q(x) is constrained to be a Bayes strategy i.e. to arise from some prior p? (θ).
Note the absence of θ̂d (x) and the presence of an integral over X , corresponding
to the fact thats that this maximization takes place without being given a subspace
Θd nor seeing data x. The resulting distributions on X are also different, as drawn in
Figure S2. If plotted on Figure 5B, pNML
(x) from the full model would be somewhere
2
between the two expected data p(x) lines there. But it is not a comparable object; its
purpose is model comparison as in Figure S1.
¶
Recall that ~
x is ~
y corrupted by Gaussian noise, and ~
y is constrained to the area show
in Figure 4A because it arises from decay rates kµ via Eq. (5). We may regard
either yt or kµ as being the parameters, generically θ .
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
8
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
B
A
pNML
(x)
d
x2
p(x|d)
■
■
■
■
�
■
x1
�
x
x2
pNML
(x)
d
■
�
d=
x
Full model, pJ(𝜃)
Optimal p⭑(𝜃)
Corners only
x1
Fig. S2. Distributions over X . Panel A shows p(x|d) Eq. (S1) and pNML
(x)
d
Eq. (S2) for the one-parameter Gaussian example Eq. (2) with σ = 1/10. The
three models being compared are the full model (with a flat prior), the simpler model
defined by p? (θ), and a model with just the endpoints of the line. Under the Bayes
factor comparison, the p? (θ) model would never be preferred here. Panel B draws
pNML
(x) for the two-parameter exponential model, Eq. (5), for the complete model,
d
the effective model defined by the support of p? (θ), and an even simpler model
allowing only three points — the same three models as compared in Figure S1A.
Notice that pNML
(x) is always a constant on the allowed region x ∈ y(Θd ).
d
By definition p? (θ) has the highest score. In second place comes the
line along the lower edge (corresponding to k1 = k2 ). The shorter
lines are strongly disfavored, because they cover a much smaller
range of possible data.
Algorithms
The standard algorithm for maximizing channel capacity (of discrete
memoryless channels) was written down independently by Blahut
(38) and Arimoto (37). This aspect of rate-distortion theory is
mathematically the same as the problem we consider, of maximizing
mutual information by choosing the prior. The algorithm starts
with p0 (θ) = const., and then at each time step updates this by
pτ +1 (θ) =
1 fKL (θ)
e
pτ (θ)
Zτ
0
[S4]
where Zτ =
dθ0 efKL (θ ) pτ (θ0 ) maintains normalization, and
fKL (θ) = DKL [p(x|θ)kp(x)] is computed with pτ (θ). Since this
R
𝜏 = 106 iterations
�
x2
■
■
𝜏 = 1000
Fig. S3. Convergence of the BA algorithm (S4) for the exponential model. This
shows p? (~
y ) for the case σ = 1/50, with ~
y discretized on a grid of spacing 1/100
in the bulk and 1/200 along the boundaries of the allowed region.
x1
■
■
𝜏 = 100
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
is a convex optimization problem, the algorithm is guaranteed to
converge to the global maximum. This makes it a good tool to see
discreteness emerging.
Figures 2 and S3 show the progress of this algorithm for the
one- and two-dimensional models in the text. We stress that the
number and positions of the peaks which form are unchanged when
the discretization of θ is made much finer. Notice also that the
convergence to delta functions happens much sooner near to the
boundaries than in the interior. The convergence to the correct
mutual information I(X; Θ), and towards the optimum distribution
on data space p(x), happens much faster than the convergence to
the correct number of delta functions.
Because θ must be discretized for this procedure, it is poorly
suited to high-dimensional parameter spaces. However once we
know that p? (θ) is discrete it is natural to consider algorithms
exploiting this. With K atoms, we can adjust their positions θ~a
and weights λa using gradients
∂MI
= λa
∂θaµ
Z
dx
~
~
p(x|θ)
∂p(x|θ)
log
µ
∂θ
p(x)
~ θ
~a
θ=
∂MI
= fKL (θ~a ) − 1.
∂λa
[S5]
Figures 1, 2A, 5 and the square plot points in Figure 4 were generated this way. This optimization is not a convex problem (there is
some tendency to place two atoms on top of each other, and thus
use too few points of support) but it can often find the optimum
solution. We can confirm this by calculating fKL (θ) everywhere
— any points for which this is larger than its value at the atoms
indicates that we do not have the optimal solution, and should add
an atom.
Monte Carlo algorithms for this problem have been investigated
in the literature, see (77, 78) and especially (79). (Incidentally, we
observe that (77)’s table 1 contains a version of scaling law Eq. (4),
with ζ ≈ 1/2. No attempt was made there to use the optimal
number of atoms, only to calculate the channel capacity to sufficient
accuracy.)
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
9
| 10math.ST
|
A Comparison of Public Causal Search Packages on
Linear, Gaussian Data with No Latent Variables1
Joseph D. Ramsey and Bryan Andrews
arXiv:1709.04240v2 [cs.AI] 16 Sep 2017
Pittsburgh, PA, USA
Abstract
We compare Tetrad (Java) algorithms to the other public software packages
BNT (Bayes Net Toolbox, Matlab), pcalg (R), bnlearn (R) on the “vanilla”
task of recovering DAG structure to the extent possible from data generated
recursively from linear, Gaussian structure equation models (SEMs) with no
latent variables, for random graphs, with no additional knowledge of variable order or adjacency structure, and without additional specification of
intervention information. Each one of the above packages offers at least one
implementation suitable to this purpose. We compare them on adjacency
and orientation accuracy as well as time performance, for fixed datasets. We
vary the number of variables, the number of samples, and the density of
graph, for a total of 27 combinations, averaging all statistics over 10 runs,
for a total of 270 datasets. All runs are carried out on the same machine and
on their native platforms. An interactive visualization tool is provided for
the reader who wishes to know more than can be documented explicitly in
this report.
Keywords: Causal Search, Linear, Gaussian, benchmarking
1. Introduction
Understanding how variables relate to one another causally has been of
interest in many fields, including biology, neuroscience, climate science, economics, and the social sciences. It has been argued that causal research is
best carried through intervention of variables or through controlled randomized experiments, though this is often prohibitively expensive, unethical, or
1
Revision 1
Preprint submitted to arXiv
September 19, 2017
impossible. An alternative is the discovery of causal relations from observational data alone. Numerous algorithms have been constructed to this end,
which fall generally into one of two groups, constraint-based (like the PC
algorithm) or score-based (like the GES algorithm). In addition to understanding how these algorithms relate to one another theoretically, it is at a
basic level of interest to see how they related to one another in performance.
To this end, we compare Tetrad implementations (in Java)2 to a number of
widely-used implementations from publicly accessible packages: BNT (Matlab), pcalg (R, C),3 bnlearn (R, C++),4 on their native platforms and on
the same machine. Notably, through the Center for Causal Discovery5 , the
Tetrad algorithms have been given wrappers for R and Python, so there is
no need usually for researchers to give up their accustomed platform in order
to use them. We compare them, however, in their native Java.6
We focus on what we call the “vanilla” task–data simulated i.i.d. recursively from linear, Gaussian structural equation models (SEMs), with no
latent variables, from graphs that are not particularly “surprising”, and with
no intervention data. To make random graphs, we simply add edge randomly
in the forward direction to variables in a linear order. Coefficients for edges
are drawn uniformly from (0.2, 0.9); exogenous error variances from (1, 3).7
We study this task, because each of the platforms studied has at least one
algorithm to address it, so the comparisons from one platform to the next on
the same datasets are meaningful. We vary sample size as 100, 500 or 1000;
we vary number of variables in the generating graph as 50, 100, or 500; we
vary the average degree of the graph as 2, 4, or 6. For each combination, we
simulate and store 10 random graphs and datasets, with different coefficients
and exogenous error variances.
We recognize that there are many types of data on which one may wish
to compare causal search algorithms. One may wish to look at linear or
nonlinear connection functions, Gaussian or non-Gaussian exogenous errors,
2
http://www.phil.cmu.edu/tetrad/
https://cran.r-project.org/package=pcalg
4
http://www.bnlearn.com/
5
http://www.ccd.pitt.edu/
6
http://www.phil.cmu.edu/tetrad/
7
A “surprising” graph might be one that is, for example, scale-free, so that some nodes
have many adjacents, others very few. Typically nodes with many adjacents are difficult
to analyze unless all arrows point out from them.
3
2
continuous or discrete variables, or a mixture continuous and discrete.8 One
may wish to assume knowledge of time order or not, or knowledge of adjacent structure. We take the attitude, though, that these problems need to
be taken up one at a time as algorithms become publicly available. Since
the “vanilla” task has taken up so much of the effort of so many researchers
over so many years, it would be good to get a snapshot of progress on it.
We will draw some very basic conclusions from this study after presentation
of results. Our attitude is simply that, while there is great value in comparing algorithms based on their theoretical virtues, there is also at a base
level value in comparing them head to head on a common task, for common
parameter values, on a common machine, just on their performance. This
way, the strategies that are most effective quickly become apparent and may
be pursued further, and strategies that aren’t as effective may be pursued,
after some thought, perhaps less vigorously.
We run each algorithm on its native platform (Matlab, R, or Java), on
exactly the same datasets and on exactly the same machine, and compile
the results into a single table. We believe this renders the results comparable. To our knowledge, there is no such previous comparison of all of these
implementations together, though comparisons of subsets have been pursued.
The table of results that comes out of this is quite large, too large really
to draw good conclusions by eye without a tool to do so. To this end, we offer
an interactive visualization comparison of the results. We give some results
individually in the Results section below, but the reader is encouraged to try
the visualization tool and draw their own conclusions.
In Section 2 we give the simulation task on which we are comparing the
various algorithms. In Section 3, we specify the comparison procedure. In
Section 4, we preview the packages and algorithms we will compare. In
Section 5, we describe the visualization tool we will use. In Section 6, we
give some sample results. In Section 7, we give a discussion of these results.
In Section 8, we draw some conclusions. It should be noted (as we will in
fact note) that there are many more results than are described in the text
of the paper; we give these results in an Appendix. This table of results is
available as an attached file with this report for use with the visualization
tool. See Section 5.
8
See, for example, Andrews et al., 2017
3
2. The Simulation Task
We generate data, at various sample sizes, from models where they struture should be recoverable, given unbounded sample size. Most of the algorithms listed can recover the pattern (CPDAG) of the true DAG, some the
adjacency structure of the true DAG. We are interested to know how the
algorithms scale in three dimensions. First, we’d like to know how they scale
up with sample size. Second, we’d like to know how they scale up with the
number of variables being studied. Third, we’d like to know how they scale
up with the density of the generating graph. We, therefore, create datasets
for models that are varied along these three dimensions. For sample size, we
chose 100, 500, or 1000 samples. 100 we will take to be a “small” sample,
although for some problems it’s large. 1000 is usually a large sample, where
one expects near convergence in the typical case. For numbers of variables,
we choose 50, 100, or 500 variables. 500 is not an upper limit for all of the
algorithms we study, not by any means. But it is already out of reach for
some of them, so in the interest of providing a common set of datasets to
study across as many algorithms as possible, we limit it thus. Likewise, 50
is by no means necessarily a small number of variables; for some problems,
that’s in fact quite large, but we are interested in how these algorithms scale,
so we pick 50 as a relatively small number of variables to study. For the density of graph, we vary the average degree of the graph, which is the average
total number of parents and children each node has in the true DAG. We
choose values of 2, 4, and 6. Again, a graph of average degree 6 is not necessarily the densest graph one may wish to analyze. In fact, for some scientific
problems, for instance in neuroscience, the density of the true model may
be much greater than that. Nevertheless, an average degree of 6 is already
out of reach for many algorithm, including many of the algorithms we are
comparing, so we limit ourselves thus.
This yields a total of 27 combinations. For each combination, we simulate
ten graphs and data sets randomly choosing simulation parameters for each
dataset, for a total of 270 datasets. These are stored in a public GitHub
repository.9 ; readers are encouraged to try their hand at these datasets to
see if they can improve on the performances described here.
9
https://github.com/cmu-phil/comparison-data
4
3. The Comparison Procedure
We adapted the Algorithm Comparison Tool in Tetrad [10] to do the
comparisons. That tool automates the simulation of datasets and the running
of lists of algorithm variants on that data to produce a final tabular report
of performance statistics. For various reasons, this was ill-suited to the task
of comparing algorithms cross-platform, so adjustments were made. First,
elapsed times were in the original tool calculated automatically on the fly for
algorithms without a record being made of them; we explicitly made a record
of these. Second, result graphs for the existing tool were stored internally
and results tabulated over them; as an adjustment we stored these out in
a directory structure to files. Third, since graphs on other platforms are
saved in formats different from Tetrad’s graph format, parsers needed to be
written to read graphs from each platform, sometimes for each algorithm, so
that in the end all graphs could be compared to the correct “true” DAG for
each simulation; we wrote these parsers. Graph outputs and elapsed time
files were stored in parallel directories and classes were added to reload the
original simulation data and true DAGs, the result graphs and elapsed times,
and to calculate performance statistics and present them in a table in the
manner of the original algorithm comparison tool.
Importantly, all algorithms were run on their native platforms, with timing results calculated on those platforms. To make the timing results comparable, we ran all algorithms on the same machine, with 3.3 GHz Intel I7
processor, with 16G RAM. This processor has two cores with two hardware
threads each, so there is some speedup available for algorithms that can take
advantage of parallelization. The version of Matlab used was 9.2.0.556344
(R2017a); the version of R used was 3.4.1; the version of Java used was
1.8.0 131. The version of BNT used was last updated 17 Oct 2007; a newer
version could not be found. The pcalg code used was version 2.5.0-0, packaged 2017-07-11 and built from sources. The version of bnlearn used was 4.2,
published 2017-07-03, and built from sources. The version of Tetrad used was
commit 39b8abd in the Tetrad GitHub repository from 8/3/2017. The end
result of this analysis was a table of simulation results for each algorithm
or algorithm variant, each sample size, each number of variables, and each
average degree. All statistics were averaged over 10 runs.
While the algorithms comparison tool allows arbitrary statistics to be
calculated and recorded in the final table, we settle on the following list of
statistics, which seem to us to be fairly expressive. Let AT P be the number
5
of true positive adjacency judgments, AF P the number of false positive adjacency judgments, AF N the number of false negative adjacency judgments;
let AHT P , AHF P , AHF N be the corresponding judgments for arrowheads
in the graph, where X → Y counts as as true positive arrowhead if it occurs
in both the true DAG and the estimated graph. Then we have the following
measures:
Vars: Number of variables in the graph
Deg: Average degree of the graph
N: Number of simulated samples
P
AP : Adjacency Precision = AT PAT+AF
P
P
AR: Adjacency Recall = AT PAT+AF
N
P
AHP : Arrowhead precision = AHTAHT
P +AHF P
P
AHR: Arrowhead recall = AHTAHT
P +AHF N
M cAdj: Matthew’s correlation coefficient for adjacencies
=√
(AT P ∗AT N )−(AF P ∗AF N )
(AF P +AF P )∗(AT P +AF N )∗(AT N +AF P )∗(AT N +AF N )
M cArrow: Matthew’s correlation coefficient for arrowheads
=√
(AHT P ∗AHT N )−(AHF P ∗AHF N )
(AHF P +AHF P )∗(AHT P +AHF N )∗(AHT N +AHF P )∗(AHT N +AHF N )
E = Elapsed Time in Seconds
Here, the Matthews correlation ranges from -1 to 1, with higher positive values indicating greater agreement between precision and recall, respectively, for adjacency or arrowhead judgments. The Matthews correlation is
designed to give an aggregate accuracy for precision and recall, balancing
well when the sizes of the constituent sets (TP, FP, FN) are very different in
size. We report AP , AR, AHP , AHR, and E in the results section below
but include M cAdj and M cArrow in the tables in the Appendix.
In some cases results are not provided. These are cases where running
time for a single run of the algorithm was greater than 10 minutes.
4. The Packages and their Algorithms Relevant to this Task
We make some general remarks that apply to multiple packages. Algorithms in these packages, as is generally the case, are commonly said to be
of two sorts, constraint-based or score-based. The distinction is blurred, as
has been noted by [7].
6
The PC-style algorithms are all constraint-based because they operate
entirely from considerations of conditional independence, which constitute
constraints on the search. They take a conditional independence test as an
oracle; the data themselves are screened off by the conditional independence
test. In fact, the algorithms could operate without data if the conditional
independence facts alone were available. So a sensible way to check the
correctness of these algorithms is to feed them d-separation facts from a
known true graph and see if they can recover the CPDAG or adjacencies of
the DAG, depending on the algorithm.
The GES algorithm (Chickering 2002), by contrast, is said to be scorebased, on the strength of the assertion that they rely on Bayesian scores
like BIC for their search. However, it can also be carried using conditional
independence tests as oracles. To make this point, we include some runs of
FGES for Tetrad using the Fisher Z test. For the Fisher Z runs, we use α − p
as a “score”, where alpha is the alpha cutoff for PC and p is the p-value for
the conditional independence test. This is not a proper score, but it does
convey the conditional independence information properly, and as can be
seen, it works well. Also, we have undertaken to check the correctness of
FGES from d-separation facts alone. One way to do this is to return a score
of +1 for conditional d-connection and -1 for conditional d-separation. If one
does this, GES should, in fact, give the same graphs back from d-separation
facts alone as PC, and it does, for all of the FGES implementations. We
assume similar checks have been made for pcalg. In general, however, using
a score for GES that prioritizes true colliders changes the priority of addition
and removal of edges in the forward and backward stages, in ways that help
to find true colliders more quickly. This shortens and stabilizes the search.
For the PC-style algorithms, for small samples, there is always the possibility of a collider conflict–that is, a situation where one is given a model
X → Y ← Z − W , where the X − Y − Z triple has been oriented as a
collider, and for similar reasons one wants to orient the triple Y − Z − W as
a collider as well. For the PC algorithm this can happen if the edge X − Z
has been removed from the graph on the strength of the observation that
I(X, Z|S1) where Y is not in S1, but also I(Y, W |S2) where X is not in S2.
(Here, I(X, Y |Z1) is true if and only if X is independent of Y conditional on
set Z of variables.) For models in which there are no latent variables, this is a
small sample error, but it may occur quite often, depending on the algorithm.
There are at least three possible graphs that could result. One could orient
both colliders to get X → Y ↔ Z ← W ; we call this the bidirected rule;
7
this has been pursued in the Tetrad package in previous versions.10 Or one
could orient just the first collider and refuse to orient the second, which we
call the priority rule–that is, X → Y ← Z − W . Or one could simply orient
the second without unorienting the X → Y edge, to get X → Y → Z ← W ,
which we call the overwrite rule. For these runs, Tetrad’s PC algorithms are
all using the priority rule. The pcalg package below uses the overwrite rule.
For constraint-based algorithms we use the default conditional independence test for the package, except where noted. When there is an option
for setting an alpha level, we set it at at 0.01 or 0.001, except where noted.
For the GES algorithms, we use BIC, with a multiplier on the penalty, so
BIC = 2L − cklnN , where L is the likelihood, c the penalty discount, k the
degrees of freedom, and N the sample size. We use penalty discounts 2 and
4; for pcalg, this corresponds to λ’s of 2lnN and 4lnN (see [7].
Scripts used to generate all results are available. For Tetrad, they are
included in the GitHub results repository11 in a directory called “condition2”.
Since we were adapting the Tetrad algorithms comparison tool, we began
with the Tetrad algorithms and proceeded from there.
4.1. The Tetrad Suite
The Tetrad package contains more or less the same variety of algorithms
for the “vanilla” case as pcalg, with some notable variations in particular
cases. It contains almost all of the same variation in PC-style algorithms
(PC, PC-Stable, CPC, CPC-Stable) and also an optimized GES algorithm
(FGES). The package is written in Java. The indices given are the indices of
the variants in the Appendix.
The PC algorithm was defined in [11]; [5] later gave implied orientation
rules which have been adapted in the Tetrad implementation. The PC-Stable
algorithm [3] aims to render the adjacencies in the output independent of
the order of the input variables; we have implemented this based on the
10
The reason this has been pursued is that when latent common causes of pairs of
measured variables affect the data, a bidirected edge of this form indicates the existence of
one or more latent variables, a desired behavior. Here, since we know there are no latent
variables, these bidirected edges are a nuisance, due to small sample errors of judgment,
so we don’t try to orient them.
11
https://github.com/cmu-phil/causal-comparisons; this repository is currently private
because other other incomplete comparisons are included, but will be made available to
readers if a GitHub ID is sent to the authors. But all of the results graphs and scripts are
included there.
8
description. The PC-Stable-Max algorithm alters the CPC algorithm to
orient colliders using maximum p-value conditioning sets and renders both
adjacencies and orientations order-independent. The CPC algorithm ([9])
orients colliders by insisting on uniformity of independence judgments. CPCStable combines CPC orientation with the PC-Stable adjacency procedure.
The “stable” adjacency search ([3]) was adapted from the pcalg package.
Notably, for the CPC algorithm, we interpret all ambiguous markings in
the graph as noncolliders (see [9]). CPC aims to mark all unshielded triple
X−Y −Z as unambiguous colliders, unambiguous noncolliders, or ambiguous.
Thus the output represents a collections of CPDAGs, not a specific CPDAG.
To get a specific CPDAG from the collections, one needs to choose for each
ambiguous triple whether it is a collider or a noncollider, and then apply the
Meek orientation rules. Here, we choose to interpret each ambiguous triple
as a noncollider, as does pcalg. We do this for compatibility.
The FGES algorithm ([8]) modifies the GES algorithm ([2]), with the
intention of making it more scalable. It forgoes the full loop in the forward
phase, using caching of information instead to avoid some repetition of tests.
There are two settings, faithfulness = true and faithfulness = false. If the
faithfulness flag is set to true, edges X − Y with X uncorrelated with Y are
not added to the graph in a first sweep; for paths X − Y − Z that should
be colliders, X − Z is then added to the graph and removed in a subsequent
backward step, orienting the collider. This strategy is similar to the one
described in [7] for the ARGES algorithm and results in significant speedup,
as they attest. With the faithfulness flag set to false, the same first sweep
(forward followed by backward) is performed, but then edges W − Z are
recovered for X that are unconditionally d-connected to Z but not correlated
with Z unconditionally in a second forward sweep, and the backward sweep
is repeated. This allows for some edges with canceling paths to be recovered
and is equivalent to Chickering’s formulation of GES. Additionally, sections
of the algorithm that can be parallelized are treated as parallelized. This
includes in particular the initial step of adding a single edge to an empty
graph, a step which is particularly time-consuming.
As noted, all g of the PC variants for the Tetrad runs are configured to use
the priority rule to resolve colliders. An alpha of 0.01 and 0.001 are used. For
FGES, some smaller, for Fisher Z, some smaller alphas are included, 0.0001
and 0.00000001, since these are advantageous. For FGES with BIC, penalties
2 and 4 are included. The following are the specific variants include. The
indices correspond to the indices in the full in the Appendix.
9
1. PC (“Peter and Clark”), Priority Rule, using Fisher Z test, alpha =
0.01
2. PC (“Peter and Clark”), Priority Rule, using Fisher Z test, alpha =
0.001
3. PC-Stable (“Peter and Clark” Stable), Priority Rule, using Fisher Z
test, alpha = 0.01
4. PC-Stable (“Peter and Clark” Stable), Priority Rule, using Fisher Z
test, alpha = 0.001
5. PC-Stable-Max (“Peter and Clark”), Priority Rule, using Fisher Z
test, alpha = 0.01
6. PC-Stable-Max (“Peter and Clark”), Priority Rule, using Fisher Z
test, alpha = 0.001
7. CPC (Conservative “Peter and Clark”), Priority Rule, using Fisher Z
test, alpha = 0.01
8. CPC (Conservative “Peter and Clark”), Priority Rule, using Fisher Z
test, alpha = 0.001
9. CPC-Stable (Conservative “Peter and Clark” Stable), Priority Rule,
using Fisher Z test, alpha = 0.01
10. CPC-Stable (Conservative “Peter and Clark” Stable), Priority Rule,
using Fisher Z test, alpha = 0.001
11. FGES (Fast Greedy Equivalence Search) using Fisher Z Score, alpha
= 0.001, faithfulnessAssumed = false
12. FGES (Fast Greedy Equivalence Search) using Fisher Z Score, alpha
= 1.0E-4, faithfulnessAssumed = false
13. FGES (Fast Greedy Equivalence Search) using Fisher Z Score, alpha
= 1.0E-8, faithfulnessAssumed = false
14. FGES (Fast Greedy Equivalence Search) using Sem BIC Score, penaltyDiscount = 2, faithfulnessAssumed = false
15. FGES (Fast Greedy Equivalence Search) using Sem BIC Score, penaltyDiscount = 4, faithfulnessAssumed = false
16. FGES (Fast Greedy Equivalence Search) using Fisher Z Score, alpha
= 0.001, faithfulnessAssumed = true
17. FGES (Fast Greedy Equivalence Search) using Fisher Z Score, alpha
= 1.0E-4, faithfulnessAssumed = true
18. FGES (Fast Greedy Equivalence Search) using Fisher Z Score, alpha
= 1.0E-8, faithfulnessAssumed = true
19. FGES (Fast Greedy Equivalence Search) using Sem BIC Score, penaltyDiscount = 2, faithfulnessAssumed = true
10
20. FGES (Fast Greedy Equivalence Search) using Sem BIC Score, penaltyDiscount = 4, faithfulnessAssumed = true
4.2. The bnlearn Package
The blearn package ([13], [14], [15], [1]) consists of a variety of algorithms,
largely centered around the idea of finding Markov blankets and using Markov
blankets to construct estimations of CPDAGs or undirected graph representing the adjacency structure of the true DAG. Judging from their description,
these algorithms are not primarily aimed at the purely continuous case; we
will come back to them in a subsequent study of multinomial data, for which
more complete test results from the authors are available. Nevertheless, we
give the continuous-only results here, using the default conditional independence test.
Two of the algorithms, MMPC and HITON, estimate adjacencies but not
orientations. For this reason, orientation statistics will not be given for them.
The bnlearn algorithms are all written natively in C and included in R
through the usual sort of wrappers.
We test the following algorithms and their variants. (the indices are the
indices of these variants in our output table.) We use an alpha of 0.01 or
0.001. MMHC in the version of the software we tested does not take an alpha
values.
21. MMPC alpha = 0.01
22. MMPC alpha = 0.001
23. GrowShrink alpha = 0.01
24. GrowShrink alpha = 0.001
25. IAMB alpha = 0.01
26. IAMB alpha = 0.001
27. Fast.IAMB alpha = 0.01
28. Fast.IAMB alpha = 0.001
29. Inter.IAMB alpha = 0.01
30. Inter.IAMB alpha = 0.001
31. si.hiton.pc alpha = 0.01
32. si.hiton.pc alpha = 0.001
33. MMHC
Finally, we took one of the bnlearn algorithms, iamb, and varied the
independence test from the available options. The “monte carlo” (mc) tests
did not scale well even to our smallest problems, so we focused on the options
without the Monte Carlo adjustment.
11
34.
35.
36.
37.
38.
39.
40.
41.
iamb
iamb
iamb
iamb
iamb
iamb
iamb
iamb
alpha
alpha
alpha
alpha
alpha
alpha
alpha
alpha
=
=
=
=
=
=
=
=
0.01.test=cor
0.001.test=cor
0.01.test=mi-g
0.001.test=mi-g
0.01.test=mi-g-sh
0.001.test=mi-g-sh
0.01.test=zf
0.001.test=zf
4.3. The pcalg Package
Descriptions for the pcalg package may be found in [4]; for more recent
documentation, please see the documentation provided in the R package
pcalg. We focus just one algorithms that apply directly to the “vanilla”
task: the PC-style algorithms and the GES implementation. As noted above,
pcalg and Tetrad for this case contain very similar implementation. They
differ in implementation style, language, and platform. The pcalg algorithms
are all written natively in C++ and included in R through the usual sort
of wrappers. Exploring differences in implementation style and choices is an
interesting topic, but this lies beyond the scope of this paper; needless to say,
all of the code is publicly available for anyone who is interested.
In pcalg, the pc method is designed to adjust its behavior in response to
parameters to morph into the various PC-style algorithms. Most of these
have counterparts in the Tetrad suite; some do not (like the majority rule).
To make the comparison easy, we will use the Tetrad names for the algorithms. The specific pcalg commands used to run those algorithms are available in the comparison2 directory of the results repository. In all cases, we
use alpha values of 0.01 or 0.001. For PC-stable, which is readily parallelizable, we set the number of cores to 4.
The GES method in the pcalg package uses a λ rather than a penalty
discount parameter, but these are inter-convertible, we use λ = clnN ; here,
c is the penalty discount.
42. PC pcalg defaults alpha = 0.01
43. PC pcalg defaults alpha = 0.001
44. PC-Stable pcalg ncores=4 alpha = 0.01
45. PC-Stable pcalg ncores=4 alpha = 0.001
46. CPC pcalg defaults alpha = 0.01
47. CPC pcalg defaults alpha = 0.001
48. CPC pcalg majority.rule defaults alpha = 0.01
12
49. CPC pcalg majority.rule defaults alpha = 0.001
For GES, we test the following variants:
50. GES pcalg defaults 2*log(nrow(data))
51. GES pcalg defaults 4*log(nrow(data))
4.4. The BNT Package
BNT (“Bayes Net Toolbox”, [6]) contains a number of algorithms, Once
one eliminates algorithms that are for discrete variables only, algorithms that
assume time order of the variables or knowledge of adjacency structure, or
algorithms that are designed for the latent variable case, the only structure
learning algorithm left to compare with is the implementation of the PC
algorithm, learn struct dag pc. This takes an alpha level; we use 0.01 or
0.001. The BNT algorithms are written natively in Matlab. There are two
variants for PC in BNT in our table:
52. learn.struct.pdag.pc bnt alpha = 0.01
53. learn.struct.pdag.pc bnt alpha = 0.001
5. Visualization Tool
The results are shown in the Appendix in tabular form. Since the this
table is quite large, we found it useful to provide an interactive visualization
tool12 to help readers understand the results. We render figures from this
tool in the Results section below. However, we invite interested readers to
explore it on their own, as there is much more data to explore (see Appendix).
One simply needs to download these three files, stats.txt and config.txt, and
std.txt,13 then go to the URL above for the tool and load them the respective
buttons, then set the number of runs to 10. By providing a file of standard
deviations (std.txt) and the number of runs, the tool additionally calculates
95% confidence intervals around the mean statistics; [12] is used to calculate
the inverse t CDF. Then click to the Plot tab and select a combination of
parameters for which a result exists, and bar plots will be rendered as in
the Results section, below. As many or as few results as are desired may be
simultaneously plotted.
12
http://www.pitt.edu/~bja43/causal
These files are currently provided as example files for the visualization tool, but they
are also included in this arXiv download.
13
13
In order for the bar plots to be displayed, one needs to select a combination of parameters for which data is available in the table. One needs to pick
a one or more numbers of variables, average degrees, and sample sizes. One
needs to pick parameters for which they are compared. (Different algorithms
may require different parameters.) Then one needs to pick one or more algorithms from the four packages.14 Provided one has selected a combination
of settings for which data is available, charts will be rendered.
6. Results
There is one algorithm that is implemented on three different of the
platforms–viz., PC. We show results for this in Figure 1. Here we use an
alpha value of 0.001, with 100 variables, average degree 4, sample size 500.
In Figure 2 we show a comparison of all of the PC-style algorithms in
the Tetrad and pcalg packages. Here, we look at the 100 variables case, with
average degree 4, sample size 500, alpha = 0.001.
In Figure 3 we show a comparison of all of the GES algorithms in the
Tetrad and pcalg packages. Here, we look at the 100 variable case, average
degree 4, sample size 500. For algorithms that use the Fisher Z test, an alpha
of 0.001 was used; for algorithms that use a BIC score, a penalty of 2 was
used. (For pcalg, λ in this case is 2lnN .)
In Figure 4 we show a comparison of all of the algorithms tested in the
bnlearn package. Here, we look at the 100 variable case, average degree 4,
sample size 500, alpha = 0.001.
The plots in Figure 1, Figure 2, 3, and 4 are all for the same datasets and
so are comparable. Using the visualization tool one may obtain similar plots
for other cases.
In Figure 5 we show a comparison of elapsed times for all of the above
choices, for common datasets.
7. Discussion
For Figure 1, it is important to note that all datasets have had their
columns randomly rearranged, so that there is no opportunity to gain an
14
It is worth noting that this visualization tool is intended to be useful outside of this
particular project, and already has been. One simply needs to make a stats, config, and
std files in the proper format and load them in. Plots may be edited using Plotly.
14
Figure 1: Adjacency and orientation accuracy of PC across platforms for 100 variables,
average degree 4, 500 samples, alpha = 0.001.
Figure 2: Adjacency and orientation accuracy of PC-style algorithm for Tetrad and pcalg
platforms. This is for 100 variables, average degree 4, sample size 500, alpha = 0.001.
15
Figure 3: Adjacency and orientation accuracy of GES for Tetrad and pcalg packages. This
is for 100 variables, average degree 4, sample size 500. For Fisher Z algorithms, alpha =
0.001; for BIC algorithms, penalty = 2 (for pcalg, λ = 2lnN ).
Figure 4: Adjacency and orientation accuracy of bnlearn algorithms. Algorithms with no
orientation statistics do not estimate orientations. This is for 100 variables, average degree
4, 500 samples, alpha = 0.001, where applicable.
16
Figure 5: Elapsed times for all algorithms, average degree 4, number of variables 100,
sample size 500, alpha 0.001, penalty 2. The elapsed times are on a log scale.
advantage from having the columns in causal order. Under these conditions,
PC (as with most algorithms for coefficients in the range +/-(0.2, 0.9)) gets
nearly 100% precision for adjacencies, with recall about 0.8, and orientation
precision of about 0.6, with orientation recall (judged against the true DAG)
about 0.45. This is a fairly common profile for PC for this sized problem, and
all three of the platforms more or less agree to it. The Tetrad algorithms gets
slightly higher precision for orientations than the others, perhaps enough to
choose it over pcalg, but really the implementations are all very similar.
Figure 2 compares all of the PC-style algorithms from the pcalg and
Tetrad packages. All of these algorithms use the same adjacency search,
so it is not surprising that their adjacency statistics are identical. They
differ only in orientation. Here, the PC and PC-Stable algorithms yield
fairly unusable orientations; the algorithms that attempt to boost collider
orientation accuracy all do better. The algorithms that do the best are the
CPC variants, for both platforms. There is a trade-off of orientation precision
for orientation recall for these algorithms. PC-Stable Max and CPC Majority
get the higher orientations recalls, but both suffer somewhat on precision; the
performance of each is basically the same.
Figure 3 compares all of the GES variants from pcalg and Tetrad. Over17
all, the performances are very similar, though for orientation precision in
particular, the pcalg implementation is to be preferred on these grounds.
Figure 4 gives timing results for all of the algorithms compared in the
previous figures. The elapsed times are given on a log scale, as some of the
times are fairly slow. Generally, the Tetrad algorithms were faster. If we
consider CPC to be the most accurate of the PC-style algorithms, for pcalg,
it comes back in 3.9 seconds, whereas for Tetrad it comes back in .27 seconds.
For GES, the time posted for pcalg was 0.67 seconds, whereas for Tetrad the
best time was 0.08 seconds. For this size problem, of course, all of these times
are entirely reasonable; one should not be hesitant to wait a few seconds to
get a good result.
For scaling up, however, timing may become an issue. In figure 6 we show
the faster algorithms, for the largest models we consider, 500 variables, average degree 6, 1000 samples. These are all variants of the FGES algorithms,
with one contrast, the pcalg GES algorithm. The fastest of these is for FGES
with the faithfulness flag turned on, using the BIC score with penalty 4. In
Figure 7 the accuracies of the same algorithms are given. As can be seen,
this particular algorithm is not only the fastest at this task but is also one
of the most accurate, though the pcalg GES still has a slight advantage in
accuracy.
8. Conclusion
We have taken an orthogonal approach to comparing algorithms, not
based on their theory but based purely on their performance. While we have
focused only on the “vanilla” task (data generated recursively and i.i.d. from
a linear, Gaussian model with no latent variables), this is a task that has
been taken up in several software packages, both public like these, but also
in private, and which has a long history in the literature. It is usually the
first sort of algorithm one tries for a dataset, for better or for worse. We
know of no comparison to date that compares all of the public packages to
one another directly on this task using exactly the same data on exactly the
same machine.
Overall, we find that the Tetrad package fares quite well in this comparison. The PC-style algorithms are comparable to those in other packages.
The best of these in terms of orientation precision, CPC, has a similar profile
to CPC in pcalg. The main difference is that the Tetrad PC-style algorithms
are much faster than their counterparts. There are no doubt algorithmic
18
Figure 6: Elapsed times for faster algorithms for the 500 variables, average degree 6,
sample size 1000 case, alpha = 0.001, penalty 4. The elapsed times are on a log scale
Figure 7: Accuracies for faster algorithms for the 500 variables, average degree 6, sample
size 1000 case. The elapsed times are on a log scale.
19
reasons for this. The pcalg algorithms are implemented in C++, the Tetrad
algorithms in Java; generally from benchmark testing C++ is 2 - 2.5 times
faster than Java with the same implementation on the same machine. We
surmise that if the Java implementation were simply translated intelligently
into C++, one would be at least a 2-fold speedup over Java, in this way.
Instead the C++ implementations are many times slower. For GES, we find
that overall the pcalg implementation has a small but persistent edge over
FGES for accuracy, though once again the Tetrad implementation is many
times faster. Partly this is due to the parallelization, though the particular
machine these algorithms were run on has only two cores, so this doesn’t
account for the 10 to 20 fold speedup we observe. If one takes into account
that C++ is at least two times faster than Java, one might expect that if
the Tetrad implementation were directly translated into C++ there would
be at least a 20 to 40 fold speedup over the pcalg implementation. This of
course suggests a programming project for someone. In any case, this kind
of speedup is important, since it outpaces by several fold the kind of speedup
one might expect from running the algorithm on a supercomputer.
While we compare these packages separately, we recognize that there has
been a considerable amount of “borrowing back and forth” between these
and other projects. The notion of inferring causal graphs from conditional
independence constraints goes back some time now and is used in all of
these projects. The PC algorithm, as noted, originally appearing in an early
version of Tetrad, is used in three of the projects, and algorithms derived
from PC in two of them. The CPC algorithm, originally in Tetrad, was
reimplemented in pcalg, along with some variants there. In bnlearn, the
building blocks of other constraint-based algorithms are shuffled to generate
CPDAGs from considerations of Markov blanket estimations and scoring
algorithms. For GES, Chickering’s (2002) algorithm was adjusted in each of
pcalg and Tetrad; the specific formulation of the BIC score difference was
adapted from ARGES [7] and is responsible for some of the speedup there.
Overall we find that for the PC-style algorithms, the algorithms of choice
for accuracy are versions the ones based on CPC. However, these are dominated on recall by the GES-style algorithms, which among these options
also include the fastest. This is perhaps surprising, since GES has had an
reputation among a number of researchers as being a slow algorithm, but
both FGES and the pcalg implementation of GES prove this reputation to
be undeserved; the algorithm offers many opportunities for optimization.
For future work, if algorithms are identified that should have been in20
cluded in this comparison, we will include them and update the comparison.
Beyond this condition, there are several more conditions to explore, as suggested above. We will take up as time permits.
Appendix A. Results Tables
Tables of results follow. Here, ’Alg’ is the algorithm as indexed in the
text, ’vars’ is the number of variables in each simulation, ’Deg’ is the average
degree of the graph, ’N’ is the sample size, ’AP’ is adjacency precision (as
defined in the text), ’AR’ adjacency recall, ’AHP’ arrowhead precision, ’AHR’
arrowhead recall, ’McAdj’ Mathew’s correlation for adjacencies, ’McArrow’
Matthews correlation for orientations, and ’E’ elapsed time in seconds. For
hardware and software specifications, please see text. There is one table for
each simulation; the conditions for the simulation are in the table and are
common for all algorithms in each table. ’*’ in table indicates the value is
undefined due to division by zero and ’-’ means that the value is zero or that
the simulations in that position in the table took longer than 10 minutes to
complete. Each statistic is an average over 10 runs.
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Vars
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
Deg
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
N
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
AP
0.95
0.99
0.95
0.99
0.95
0.99
0.95
0.99
0.95
0.99
0.95
0.98
0.99
0.91
0.98
0.96
0.98
AR
0.67
0.56
0.66
0.55
0.66
0.55
0.67
0.56
0.66
0.55
0.73
0.63
0.32
0.78
0.57
0.73
0.63
AHP
0.59
0.55
0.56
0.54
0.86
0.86
0.99
0.97
0.99
0.97
0.85
0.9
0.92
0.8
0.88
0.86
0.9
21
AHR
0.28
0.2
0.28
0.19
0.31
0.21
0.15
0.05
0.15
0.05
0.42
0.32
0.04
0.47
0.23
0.41
0.31
McAdj
0.79
0.73
0.79
0.73
0.79
0.73
0.79
0.73
0.79
0.73
0.83
0.78
0.56
0.83
0.74
0.83
0.78
McArrow
0.11
0.06
0.08
0.04
0.35
0.27
0.27
0.12
0.26
0.12
0.4
0.37
0.09
0.4
0.29
0.4
0.37
E
0.03
0.04
0.01
0.01
0.04
0.01
0.01
0.02
0.01
0.01
0.02
8.00E-03
4.30E-03
0.02
0.01
0.03
0.06
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.99
0.91
0.98
0.95
0.99
0.96
1
0.96
1
0.96
1
0.96
1
0.95
0.99
0.83
0.96
1
0.96
1
0.98
1
0.96
1
0.95
0.99
0.95
0.99
0.95
0.99
0.95
0.99
0.98
0.99
0.95
0.99
0.32
0.77
0.57
0.66
0.54
0.67
0.56
0.67
0.55
0.67
0.56
0.67
0.55
0.67
0.56
0.74
0.67
0.55
0.68
0.57
0.65
0.53
0.67
0.56
0.66
0.55
0.66
0.55
0.66
0.55
0.66
0.55
0.67
0.32
0.67
0.56
0.92
0.82
0.88
*
*
0.85
0.87
0.82
0.87
0.83
0.93
0.82
0.87
*
*
0.61
0.82
0.87
0.82
0.88
0.89
0.86
0.82
0.87
0.54
0.53
0.54
0.53
0.98
0.97
0.96
0.97
0.91
0.97
0.58
0.58
22
0.04
0.47
0.23
0.27
0.1
0.28
0.1
0.27
0.1
0.28
0.1
0.54
0.28
0.1
0.28
0.11
0.16
0.06
0.28
0.1
0.3
0.23
0.3
0.23
0.17
0.06
0.25
0.12
0.38
0.1
0.32
0.25
0.56
0.83
0.74
0.78
0.72
0.79
0.74
0.8
0.73
0.79
0.74
0.8
0.73
0.79
0.73
0.78
0.8
0.73
0.8
0.74
0.79
0.72
0.8
0.74
0.79
0.73
0.79
0.73
0.79
0.73
0.79
0.73
0.81
0.55
0.79
0.74
0.09
0.41
0.29
0.3
0.17
0.28
0.16
0.28
0.17
0.28
0.16
0.24
0.28
0.16
0.28
0.18
0.23
0.11
0.28
0.16
0.07
0.03
0.07
0.03
0.28
0.13
0.34
0.22
0.42
0.21
0.11
0.09
4.20E-03
0.02
8.80E-03
0.09
0.07
0.14
0.1
0.06
0.08
0.11
0.09
0.11
0.06
0.03
0.05
0.13
0.08
0.06
0.08
0.06
0.06
0.05
0.08
0.06
0.05
0.04
0.06
0.05
0.14
0.07
0.12
0.07
0.05
0.02
0.3
0.3
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
Vars
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
Deg
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
N
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
AP
0.98
1
0.98
1
0.98
1
0.98
1
0.98
1
0.98
0.99
0.99
0.99
0.99
0.98
0.99
0.99
0.99
0.99
0.98
1
0.97
1
0.97
1
0.97
1
0.97
1
0.98
1
0.9
0.97
1
0.97
AR
0.94
0.91
0.94
0.9
0.94
0.9
0.94
0.91
0.94
0.9
0.97
0.95
0.84
0.95
0.89
0.96
0.95
0.83
0.95
0.89
0.94
0.9
0.95
0.9
0.95
0.9
0.95
0.9
0.95
0.9
0.94
0.91
0.96
0.95
0.9
0.95
AHP
0.75
0.73
0.68
0.67
0.88
0.91
0.98
0.98
0.98
0.98
0.95
0.97
0.95
0.95
0.95
0.95
0.97
0.95
0.95
0.95
*
*
0.91
0.9
0.91
0.9
0.92
0.9
0.91
0.9
*
*
0.71
0.91
0.9
0.91
23
AHR
0.56
0.5
0.53
0.48
0.63
0.59
0.54
0.45
0.54
0.45
0.7
0.69
0.57
0.69
0.63
0.7
0.68
0.56
0.69
0.63
0.65
0.57
0.64
0.56
0.64
0.56
0.64
0.56
0.76
0.64
0.56
0.64
McAdj
0.96
0.95
0.96
0.95
0.96
0.95
0.96
0.95
0.96
0.95
0.97
0.97
0.91
0.97
0.94
0.97
0.97
0.91
0.97
0.94
0.96
0.95
0.96
0.94
0.96
0.95
0.96
0.95
0.96
0.95
0.96
0.95
0.92
0.96
0.95
0.96
McArrow
0.39
0.34
0.3
0.26
0.57
0.57
0.59
0.53
0.6
0.53
0.69
0.69
0.58
0.68
0.63
0.69
0.69
0.57
0.68
0.63
0.61
0.54
0.61
0.54
0.61
0.54
0.61
0.54
0.48
0.61
0.54
0.6
E
0.06
0.02
0.04
0.04
0.03
0.02
0.04
0.03
0.04
0.02
0.02
0.02
0.01
0.02
0.02
0.02
0.02
0.01
0.02
0.02
0.1
0.1
0.33
0.22
0.16
0.13
0.15
0.14
0.15
0.13
0.03
0.06
0.17
0.16
0.13
0.16
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Vars
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
Deg
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
N
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1
0.99
1
0.97
1
0.98
1
0.98
1
0.98
1
0.98
1
0.99
1
0.98
1
AP
0.96
0.99
0.96
1
0.96
1
0.96
0.99
0.96
1
0.95
0.97
0.98
0.97
0.98
0.95
0.97
0.98
0.9 0.9
0.94 0.9
0.89 0.88
0.95 0.91
0.9 0.9
0.94 0.75
0.9 0.72
0.94 0.75
0.9 0.72
0.94 0.98
0.9 0.98
0.94 0.96
0.9 0.98
0.94 0.96
0.85 0.97
0.94 0.78
0.91 0.76
AR
0.97
0.96
0.97
0.95
0.97
0.95
0.97
0.96
0.97
0.95
0.99
0.98
0.96
0.99
0.97
0.98
0.98
0.95
24
0.56
0.62
0.53
0.64
0.56
0.63
0.57
0.63
0.57
0.56
0.47
0.65
0.58
0.7
0.59
0.64
0.59
AHP AHR
0.79 0.61
0.8
0.57
0.78 0.62
0.76 0.58
0.87 0.65
0.91 0.62
0.99 0.61
1
0.53
0.99 0.6
1
0.52
0.89 0.7
0.91 0.69
0.92 0.64
0.92 0.69
0.93 0.66
0.89 0.7
0.92 0.69
0.93 0.63
0.95
0.97
0.94
0.96
0.95
0.96
0.95
0.96
0.95
0.96
0.95
0.96
0.95
0.97
0.92
0.96
0.95
McAdj
0.96
0.97
0.96
0.97
0.96
0.97
0.96
0.97
0.96
0.97
0.97
0.98
0.97
0.98
0.98
0.97
0.98
0.96
0.54
0.58
0.5
0.61
0.54
0.42
0.36
0.42
0.36
0.61
0.54
0.66
0.63
0.7
0.62
0.47
0.42
McArrow
0.47
0.45
0.46
0.41
0.58
0.59
0.65
0.59
0.65
0.59
0.64
0.65
0.62
0.65
0.63
0.64
0.66
0.61
0.12
0.19
0.16
0.16
0.13
0.08
0.07
0.07
0.07
0.37
0.25
0.29
0.27
0.06
0.06
0.38
0.34
E
0.04
0.02
0.05
0.02
0.05
0.03
0.08
0.04
0.08
0.04
0.03
0.03
0.02
0.02
0.01
0.03
0.03
0.02
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
Alg
Vars
Deg
N
0.98
0.98
0.96
1
0.96
1
0.95
1
0.95
1
0.95
1
0.95
0.99
0.9
0.95
1
0.95
1
0.97
1
0.95
1
0.96
1
0.96
1
0.96
1
0.96
1
0.99
0.99
0.96
1
AP
0.98
0.97
0.97
0.95
0.98
0.97
0.98
0.97
0.98
0.96
0.98
0.97
0.97
0.96
0.98
0.98
0.97
0.98
0.97
0.98
0.97
0.98
0.97
0.97
0.95
0.97
0.95
0.97
0.95
0.97
0.95
0.99
0.94
0.97
0.96
AR
0.93
0.93
*
*
0.89
0.93
0.89
0.93
0.89
0.93
0.89
0.93
*
*
0.71
0.89
0.93
0.89
0.93
0.92
0.92
0.89
0.93
0.69
0.68
0.69
0.68
0.99
0.99
0.96
0.96
0.96
0.96
0.71
0.71
0.69
0.65
0.69
0.66
0.7
0.66
0.7
0.65
0.7
0.66
0.77
0.7
0.66
0.7
0.66
0.68
0.65
0.7
0.66
0.61
0.56
0.61
0.56
0.61
0.52
0.68
0.64
0.71
0.64
0.61
0.57
0.98
0.97
0.96
0.97
0.96
0.98
0.96
0.98
0.96
0.98
0.96
0.98
0.96
0.97
0.94
0.96
0.98
0.96
0.98
0.97
0.98
0.96
0.98
0.96
0.97
0.96
0.97
0.96
0.97
0.96
0.97
0.99
0.96
0.96
0.97
AHP
AHR
McAdj
25
0.66
0.63
0.62
0.63
0.63
0.63
0.63
0.63
0.63
0.63
0.48
0.63
0.63
0.63
0.63
0.64
0.62
0.63
0.63
0.35
0.3
0.35
0.3
0.65
0.59
0.68
0.66
0.7
0.65
0.37
0.35
McArrow
0.02
0.01
0.16
0.13
0.53
0.37
0.2
0.18
0.2
0.17
0.2
0.19
0.05
0.08
0.27
0.2
0.17
0.2
0.17
0.32
0.26
0.2
0.16
0.11
0.09
0.11
0.09
0.61
0.37
0.45
0.39
0.07
0.07
0.42
0.36
E
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.98
0.99
0.99
0.99
0.99
0.99
0.98
0.99
0.99
0.99
0.92
0.95
0.98
0.87
0.97
0.93
0.95
0.99
0.9
0.97
0.99
1
0.98
1
0.99
1
0.98
1
0.99
1
0.99
0.99
0.96
0.99
1
0.99
1
0.48
0.39
0.46
0.37
0.46
0.37
0.48
0.39
0.46
0.37
0.66
0.55
0.28
0.72
0.49
0.59
0.51
0.27
0.65
0.45
0.45
0.36
0.48
0.36
0.48
0.37
0.49
0.37
0.48
0.37
0.47
0.38
0.55
0.48
0.37
0.49
0.38
0.57
0.52
0.51
0.48
0.78
0.8
0.94
0.98
0.94
0.98
0.76
0.8
0.84
0.72
0.82
0.77
0.79
0.88
0.73
0.81
*
*
0.7
0.85
0.76
0.83
0.66
0.8
0.76
0.83
*
*
0.68
0.76
0.83
0.7
0.83
26
0.21
0.15
0.19
0.13
0.24
0.14
0.09
0.05
0.09
0.05
0.49
0.37
0.07
0.56
0.32
0.43
0.33
0.06
0.49
0.27
0.23
0.12
0.22
0.11
0.25
0.1
0.22
0.1
0.39
0.22
0.11
0.25
0.11
0.67
0.6
0.66
0.59
0.66
0.59
0.67
0.6
0.66
0.59
0.76
0.71
0.51
0.78
0.67
0.73
0.68
0.5
0.75
0.65
0.65
0.58
0.67
0.58
0.67
0.59
0.68
0.59
0.67
0.59
0.67
0.6
0.71
0.67
0.59
0.68
0.6
0.06
0.01
0.01
-0.02
0.24
0.18
0.19
0.15
0.19
0.15
0.37
0.33
0.14
0.37
0.31
0.34
0.3
0.13
0.34
0.28
0.18
0.18
0.2
0.16
0.15
0.15
0.2
0.16
0.24
0.2
0.16
0.18
0.17
9.70E-03
4.70E-03
9.30E-03
4.20E-03
0.01
4.50E-03
0.01
0.02
0.01
4.50E-03
0.02
0.02
5.20E-03
0.04
0.01
0.01
8.00E-03
3.70E-03
0.02
7.50E-03
0.1
0.08
0.26
0.15
0.12
0.1
0.15
0.11
0.13
0.1
0.03
0.03
0.15
0.14
0.1
0.15
0.09
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
1
0.46 0.73
1
0.35 0.93
0.99 0.48 0.75
1
0.37 0.81
0.99 0.46 0.56
0.99 0.37 0.5
0.99 0.46 0.56
0.99 0.37 0.5
0.99 0.46 0.94
0.99 0.37 0.98
0.99 0.46 0.84
0.99 0.37 0.91
0.97 0.59 0.84
0.99 0.28 0.93
0.98 0.48 0.6
0.99 0.38 0.57
0.2
0.05
0.22
0.11
0.24
0.16
0.24
0.16
0.09
0.05
0.22
0.1
0.44
0.14
0.27
0.19
0.66
0.57
0.67
0.59
0.66
0.59
0.66
0.59
0.66
0.59
0.66
0.59
0.74
0.5
0.67
0.6
0.18
0.14
0.2
0.16
0.07
5.42E-03
0.07
5.42E-03
0.19
0.15
0.27
0.19
0.41
0.24
0.11
0.07
0.1
0.07
0.13
0.1
0.09
0.06
0.09
0.06
0.31
0.11
0.23
0.13
0.08
0.04
0.38
0.31
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Vars
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
Deg
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
N
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
AP
0.98
0.99
0.98
0.99
0.98
0.99
0.98
0.99
0.98
0.99
0.85
0.88
0.93
0.86
0.91
0.86
0.89
0.94
0.86
AHR
0.49
0.42
0.46
0.41
0.58
0.53
0.47
0.38
0.47
0.38
0.78
0.77
0.65
0.77
0.7
0.75
0.72
0.57
0.74
McAdj
0.87
0.83
0.87
0.83
0.87
0.83
0.87
0.83
0.87
0.83
0.89
0.89
0.86
0.89
0.87
0.87
0.87
0.82
0.87
McArrow
0.27
0.22
0.22
0.19
0.48
0.46
0.5
0.44
0.51
0.44
0.57
0.58
0.53
0.57
0.55
0.55
0.56
0.48
0.55
E
0.07
0.04
0.08
0.05
0.11
0.07
0.21
0.1
0.16
0.06
0.13
0.1
0.05
0.05
0.03
0.08
0.08
0.03
0.04
AR
0.79
0.72
0.78
0.72
0.78
0.72
0.79
0.72
0.78
0.72
0.94
0.92
0.81
0.93
0.86
0.91
0.87
0.73
0.9
AHP
0.68
0.66
0.65
0.64
0.82
0.84
0.93
0.93
0.94
0.93
0.76
0.79
0.82
0.77
0.81
0.76
0.8
0.82
0.77
27
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
0.92
0.98
0.99
0.99
1
0.99
1
0.99
1
0.99
1
0.98
0.99
0.96
0.99
1
0.99
1
0.99
1
0.99
1
0.98
0.99
0.98
0.99
0.98
0.99
0.98
0.99
0.92
0.96
0.98
0.99
0.79
0.77
0.71
0.8
0.73
0.8
0.73
0.8
0.74
0.8
0.73
0.78
0.72
0.82
0.8
0.73
0.8
0.74
0.79
0.72
0.8
0.73
0.78
0.72
0.78
0.72
0.78
0.72
0.78
0.72
0.9
0.75
0.79
0.72
0.81
*
*
0.8
0.78
0.82
0.8
0.81
0.79
0.82
0.8
*
*
0.79
0.82
0.8
0.82
0.81
0.82
0.78
0.82
0.8
0.62
0.6
0.62
0.6
0.94
0.94
0.82
0.85
0.85
0.87
0.63
0.62
Alg
1
Vars
50
Deg
4
N
AP AR
1000 0.98 0.86
28
0.63
0.59
0.52
0.59
0.53
0.59
0.53
0.6
0.52
0.68
0.59
0.53
0.6
0.53
0.59
0.5
0.59
0.53
0.48
0.42
0.48
0.42
0.48
0.38
0.59
0.54
0.77
0.61
0.48
0.44
AHP AHR
0.7
0.57
0.84
0.86
0.83
0.88
0.84
0.88
0.85
0.88
0.85
0.88
0.84
0.87
0.83
0.88
0.88
0.85
0.88
0.85
0.88
0.84
0.88
0.85
0.87
0.83
0.87
0.83
0.87
0.83
0.87
0.83
0.9
0.84
0.87
0.83
McAdj
0.91
0.51
0.46
0.4
0.48
0.42
0.48
0.41
0.49
0.42
0.51
0.48
0.42
0.49
0.43
0.48
0.38
0.48
0.43
0.19
0.14
0.19
0.14
0.51
0.44
0.48
0.48
0.64
0.55
0.2
0.18
McArrow
0.34
0.03
0.25
0.23
1.75
0.94
0.67
0.39
0.81
0.41
0.73
0.4
0.09
0.09
0.51
0.7
0.37
0.74
0.37
0.8
0.46
0.7
0.35
0.31
0.22
0.31
0.19
2.55
1.15
1.9
1.28
0.14
0.1
0.71
0.54
E
0.12
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.99
0.99
0.99
0.99
0.99
0.98
0.99
0.99
0.99
0.86
0.89
0.91
0.88
0.91
0.87
0.89
0.91
0.88
0.9
0.99
1
1
1
1
1
1
1
1
1
0.99
0.99
0.98
1
1
1
1
1
1
0.83
0.85
0.82
0.85
0.82
0.86
0.83
0.85
0.82
0.98
0.98
0.92
0.98
0.94
0.97
0.97
0.89
0.97
0.91
0.84
0.81
0.86
0.84
0.86
0.84
0.86
0.84
0.86
0.84
0.86
0.82
0.87
0.86
0.84
0.86
0.84
0.87
0.84
29
0.68
0.66
0.64
0.82
0.83
0.95
0.94
0.95
0.95
0.77
0.8
0.81
0.8
0.82
0.78
0.8
0.81
0.8
0.81
*
*
0.88
0.84
0.88
0.84
0.88
0.83
0.88
0.84
*
*
0.83
0.88
0.84
0.88
0.84
0.88
0.84
0.52
0.53
0.48
0.67
0.64
0.63
0.55
0.63
0.55
0.82
0.82
0.76
0.82
0.78
0.82
0.81
0.73
0.81
0.76
0.7
0.64
0.69
0.64
0.69
0.63
0.69
0.64
0.73
0.69
0.64
0.69
0.65
0.68
0.64
0.9
0.91
0.9
0.91
0.9
0.91
0.9
0.91
0.9
0.91
0.93
0.91
0.92
0.91
0.91
0.92
0.89
0.92
0.9
0.91
0.89
0.92
0.91
0.92
0.91
0.92
0.91
0.92
0.91
0.91
0.9
0.91
0.92
0.91
0.92
0.91
0.92
0.91
0.28
0.26
0.22
0.53
0.52
0.64
0.57
0.64
0.57
0.61
0.64
0.6
0.63
0.62
0.61
0.63
0.58
0.62
0.59
0.62
0.53
0.61
0.53
0.61
0.52
0.61
0.53
0.59
0.61
0.53
0.61
0.54
0.61
0.54
0.06
0.15
0.14
0.23
0.18
0.49
0.25
0.5
0.29
0.15
0.13
0.09
0.07
0.05
0.13
0.11
0.06
0.07
0.04
0.48
0.44
5.3
2.93
2.43
1.22
3.85
1.33
2.75
1.29
0.2
0.18
0.86
2.77
1.17
2.87
1.17
4.06
1.82
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
4
4
4
4
4
4
4
4
4
4
4
4
4
4
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Vars
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
Deg
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
N
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
1
0.86 0.88
1
0.84 0.84
0.99 0.85 0.6
0.99 0.82 0.58
0.99 0.85 0.6
0.99 0.82 0.58
0.99 0.85 0.95
0.99 0.82 0.94
0.99 0.85 0.84
0.99 0.82 0.85
0.92 0.96 0.85
0.94 0.9 0.86
0.98 0.86 0.61
0.99 0.83 0.6
AP
0.94
0.97
0.96
0.98
0.96
0.98
0.94
0.97
0.96
0.98
0.84
0.88
0.93
0.8
0.89
0.87
0.9
0.93
0.84
0.92
0.97
AR
0.35
0.27
0.33
0.26
0.33
0.26
0.35
0.27
0.33
0.26
0.6
0.51
0.24
0.64
0.44
0.5
0.42
0.22
0.54
0.37
0.31
AHP
0.53
0.55
0.52
0.51
0.66
0.74
0.83
0.91
0.83
0.94
0.63
0.64
0.65
0.61
0.65
0.63
0.62
0.65
0.62
0.64
*
30
0.69
0.64
0.51
0.47
0.51
0.47
0.63
0.55
0.68
0.67
0.83
0.76
0.53
0.49
0.92
0.91
0.91
0.9
0.91
0.9
0.91
0.9
0.91
0.9
0.94
0.91
0.91
0.9
AHR
0.16
0.11
0.15
0.1
0.18
0.1
0.09
0.04
0.08
0.04
0.44
0.36
0.11
0.48
0.3
0.35
0.28
0.09
0.39
0.23
-
McAdj
0.55
0.49
0.54
0.48
0.54
0.48
0.55
0.49
0.54
0.48
0.68
0.64
0.44
0.68
0.6
0.63
0.59
0.43
0.64
0.56
0.52
0.61
0.53
0.18
0.14
0.18
0.14
0.64
0.57
0.57
0.56
0.69
0.64
0.2
0.17
McArrow
0.03
0.03
0.02
7.93E-03
0.13
0.12
0.15
0.11
0.14
0.11
0.23
0.2
0.09
0.22
0.17
0.17
0.14
0.08
0.19
0.14
-
2.71
1.13
0.47
0.36
0.51
0.34
3.41
2.35
2.75
2.44
0.16
0.13
0.96
0.74
E
6.70E-03
3.70E-03
9.50E-03
6.30E-03
0.01
5.90E-03
0.01
0.01
0.01
8.60E-03
0.05
0.04
9.60E-03
0.06
0.02
0.03
0.02
3.70E-03
0.03
9.30E-03
0.12
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.99
0.98
0.99
0.99
0.98
0.98
0.99
0.99
0.99
0.95
0.98
0.95
0.99
0.98
0.99
0.98
0.99
0.99
0.99
0.99
0.96
0.98
0.96
0.98
0.96
0.98
0.96
0.98
0.91
0.94
0.94
0.97
0.24
0.34
0.26
0.35
0.25
0.36
0.26
0.35
0.25
0.34
0.27
0.39
0.35
0.25
0.36
0.27
0.34
0.24
0.35
0.26
0.33
0.26
0.33
0.26
0.33
0.26
0.33
0.26
0.51
0.24
0.35
0.28
Alg
1
2
3
Vars
50
50
50
Deg
6
6
6
N
AP AR
500 0.94 0.59
500 0.95 0.54
500 0.96 0.57
*
0.68
0.74
0.67
0.75
0.64
0.73
0.68
0.79
*
*
0.66
0.67
0.75
0.65
0.76
0.65
0.73
0.67
0.77
0.48
0.5
0.48
0.5
0.81
0.93
0.71
0.79
0.73
0.75
0.52
0.52
0.16
0.09
0.18
0.08
0.21
0.08
0.18
0.09
0.27
0.18
0.08
0.19
0.09
0.17
0.06
0.18
0.08
0.16
0.11
0.16
0.11
0.08
0.04
0.16
0.09
0.4
0.15
0.18
0.12
AHP AHR
0.61 0.35
0.61 0.32
0.59 0.32
31
0.46
0.55
0.48
0.56
0.48
0.57
0.49
0.56
0.48
0.55
0.49
0.58
0.56
0.48
0.57
0.49
0.55
0.47
0.57
0.48
0.54
0.48
0.54
0.48
0.54
0.48
0.54
0.48
0.66
0.45
0.55
0.49
0.13
0.12
0.14
0.12
0.13
0.11
0.14
0.13
0.17
0.14
0.12
0.13
0.13
0.12
0.09
0.14
0.12
-9.03E-03
-4.50E-03
-9.03E-03
-4.50E-03
0.14
0.11
0.15
0.14
0.29
0.17
0.02
0.01
0.08
0.42
0.19
0.18
0.12
0.24
0.15
0.2
0.13
0.04
0.03
0.22
0.21
0.11
0.23
0.12
0.14
0.09
0.22
0.11
0.15
0.07
0.13
0.08
0.47
0.15
0.29
0.16
0.12
0.05
0.51
0.34
McAdj
0.72
0.69
0.72
McArrow
0.15
0.13
0.11
E
0.08
0.05
0.17
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
0.97
0.96
0.97
0.94
0.95
0.96
0.97
0.71
0.77
0.87
0.73
0.83
0.71
0.77
0.88
0.73
0.85
0.97
0.98
1
1
1
1
1
1
1
1
0.96
0.96
0.96
1
1
1
1
1
1
1
1
0.53
0.57
0.53
0.59
0.54
0.57
0.53
0.91
0.89
0.78
0.9
0.82
0.89
0.85
0.68
0.87
0.75
0.56
0.51
0.6
0.56
0.61
0.56
0.61
0.56
0.61
0.56
0.57
0.52
0.59
0.61
0.56
0.61
0.56
0.62
0.56
0.61
0.56
0.57
0.74
0.72
0.87
0.86
0.88
0.87
0.59
0.64
0.75
0.6
0.7
0.58
0.64
0.74
0.61
0.71
*
*
0.71
0.69
0.73
0.71
0.72
0.7
0.72
0.72
*
*
0.74
0.73
0.71
0.73
0.71
0.72
0.72
0.73
0.71
32
0.28
0.42
0.37
0.35
0.27
0.35
0.26
0.74
0.72
0.65
0.73
0.68
0.71
0.69
0.55
0.7
0.61
0.38
0.36
0.41
0.37
0.4
0.36
0.4
0.38
0.45
0.41
0.37
0.41
0.36
0.42
0.37
0.41
0.36
0.69
0.72
0.69
0.72
0.69
0.72
0.69
0.78
0.8
0.81
0.78
0.81
0.77
0.78
0.75
0.77
0.77
0.72
0.68
0.75
0.72
0.76
0.73
0.76
0.73
0.76
0.73
0.72
0.69
0.73
0.76
0.73
0.76
0.73
0.77
0.73
0.76
0.73
0.09
0.31
0.27
0.37
0.31
0.38
0.31
0.35
0.4
0.46
0.37
0.43
0.33
0.38
0.39
0.35
0.39
0.25
0.23
0.29
0.25
0.27
0.24
0.27
0.26
0.32
0.29
0.25
0.29
0.24
0.28
0.26
0.29
0.25
0.07
0.16
0.14
0.28
0.1
0.3
0.1
0.46
0.29
0.13
0.28
0.12
0.43
0.26
0.07
0.24
0.09
0.53
0.39
9.47
2.84
2.76
0.95
4.54
1.3
3.07
1
0.23
0.18
1.01
3.49
0.9
4.17
0.91
3.09
1.06
3.06
0.89
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
6
6
6
6
6
6
6
6
6
6
6
6
500
500
500
500
500
500
500
500
500
500
500
500
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
Vars
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
Deg
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
N
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.96
0.97
0.96
0.97
0.96
0.97
0.96
0.97
0.85
0.91
0.95
0.96
AP
0.93
0.93
0.95
0.95
0.95
0.95
0.93
0.93
0.95
0.95
0.59
0.64
0.74
0.62
0.72
0.6
0.64
0.75
0.63
0.73
0.97
0.97
0.1
0.57
0.53
0.57
0.53
0.57
0.53
0.57
0.53
0.84
0.72
0.59
0.54
0.52
0.52
0.52
0.52
0.88
0.85
0.73
0.76
0.72
0.78
0.52
0.53
AR
0.63
0.59
0.61
0.57
0.61
0.57
0.63
0.59
0.61
0.57
0.94
0.92
0.85
0.92
0.86
0.93
0.9
0.79
0.91
0.82
0.6
0.56
0.06
33
0.3
0.28
0.3
0.28
0.35
0.26
0.41
0.39
0.7
0.61
0.32
0.3
AHP AHR
0.6
0.39
0.57 0.34
0.57 0.35
0.54 0.31
0.72 0.45
0.72 0.41
0.85 0.41
0.84 0.35
0.86 0.4
0.84 0.34
0.46 0.7
0.5
0.7
0.58 0.65
0.49 0.7
0.57 0.67
0.47 0.7
0.51 0.69
0.59 0.61
0.5
0.7
0.58 0.64
*
*
0.06 0.03
0.72
0.69
0.72
0.69
0.72
0.69
0.72
0.69
0.83
0.79
0.73
0.7
McAdj
0.74
0.72
0.74
0.71
0.74
0.71
0.74
0.72
0.74
0.71
0.71
0.73
0.76
0.72
0.76
0.7
0.73
0.74
0.72
0.74
0.74
0.71
0.07
0.03
0.03
0.03
0.03
0.38
0.3
0.29
0.31
0.47
0.47
0.04
0.05
McArrow
0.15
0.1
0.1
0.06
0.3
0.28
0.4
0.35
0.4
0.35
0.19
0.24
0.29
0.23
0.28
0.19
0.24
0.28
0.24
0.28
0.02
1.11
0.46
0.76
0.44
3.42
1.77
3.08
2.01
0.25
0.17
1.37
0.93
E
0.22
0.12
0.37
0.16
0.36
0.26
0.77
0.37
0.79
0.36
1.47
1.02
0.37
0.71
0.26
1.38
0.95
0.28
0.7
0.22
1.05
0.76
3.49
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.1 0.06 0.06
0.1 0.06 0.06
0.99 0.62 0.72
0.99 0.64 0.74
0.99 0.61 0.73
0.99 0.64 0.75
0.99 0.62 0.72
0.95 0.61 *
0.95 0.57 *
0.96 0.63 0.7
0.99 0.64 0.75
0.99 0.62 0.72
0.99 0.64 0.74
0.99 0.61 0.71
0.99 0.65 0.74
1
0.62 0.72
0.99 0.64 0.75
0.99 0.62 0.72
0.95 0.61 0.52
0.95 0.57 0.53
0.95 0.61 0.52
0.95 0.57 0.53
0.95 0.61 0.85
0.95 0.57 0.84
0.95 0.61 0.71
0.95 0.57 0.73
0.73 0.88 0.6
0.83 0.79 0.69
0.93 0.63 0.51
0.93 0.59 0.53
Alg
1
2
3
4
5
Vars
100
100
100
100
100
Deg
2
2
2
2
2
N
AP
100 0.89
100 0.97
100 0.91
100 0.98
100 0.91
AR
0.64
0.54
0.64
0.53
0.64
AHP
0.52
0.56
0.48
0.56
0.76
34
0.03
0.03
0.4
0.43
0.41
0.44
0.4
0.46
0.44
0.4
0.44
0.4
0.44
0.41
0.44
0.4
0.34
0.32
0.34
0.32
0.41
0.36
0.45
0.43
0.69
0.64
0.34
0.33
0.07
0.08
0.76
0.78
0.76
0.78
0.76
0.74
0.71
0.75
0.78
0.76
0.78
0.76
0.78
0.77
0.78
0.76
0.74
0.71
0.74
0.71
0.74
0.71
0.74
0.71
0.77
0.79
0.74
0.72
0.01
9.04E-03
0.27
0.31
0.28
0.32
0.28
0.28
0.32
0.27
0.32
0.27
0.31
0.28
0.32
0.27
0.04
0.05
0.04
0.05
0.4
0.35
0.3
0.3
0.33
0.4
0.03
0.05
2.33
1.03
14.59
70.99
15.85
119.75
12.02
0.5
0.35
1.83
101.94
13.09
101
14.95
85.95
11.59
101.31
13.65
1.27
0.76
1.34
0.73
4.76
2.95
4.42
3.38
0.34
0.22
1.87
1.25
AHR McAdj McArrow E
0.27 0.75
0.03
0.01
0.2
0.72
0.06
0.02
0.24 0.76
-8.49E-03 5.30E-03
0.2
0.72
0.05
4.40E-03
0.3
0.76
0.25
7.60E-03
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.98
0.89
0.97
0.91
0.98
0.91
0.98
1
0.83
0.99
0.91
0.98
1
0.84
0.99
0.89
0.98
0.89
0.97
0.98
0.88
0.97
0.89
0.98
0.89
0.98
0.71
0.89
0.98
0.88
0.97
0.94
0.99
0.89
0.97
0.91
0.98
0.53
0.64
0.54
0.64
0.53
0.7
0.59
0.32
0.74
0.54
0.69
0.59
0.32
0.73
0.54
0.63
0.53
0.64
0.54
0.55
0.65
0.54
0.64
0.55
0.64
0.54
0.7
0.64
0.55
0.65
0.55
0.62
0.51
0.64
0.55
0.64
0.53
0.84
0.97
1
0.99
1
0.79
0.87
0.98
0.7
0.89
0.79
0.87
0.98
0.71
0.89
*
*
0.8
0.9
0.9
0.8
0.91
0.79
0.9
*
*
0.48
0.8
0.9
0.78
0.9
0.86
0.95
0.8
0.9
0.5
0.56
35
0.21
0.13
0.05
0.13
0.05
0.4
0.29
0.05
0.48
0.24
0.4
0.29
0.05
0.46
0.24
0.21
0.13
0.13
0.22
0.12
0.22
0.13
0.47
0.23
0.13
0.23
0.13
0.16
0.07
0.23
0.13
0.29
0.23
0.72
0.75
0.72
0.76
0.72
0.79
0.76
0.56
0.78
0.73
0.79
0.76
0.56
0.78
0.72
0.74
0.72
0.75
0.72
0.73
0.75
0.72
0.75
0.73
0.75
0.73
0.7
0.75
0.73
0.75
0.73
0.76
0.71
0.75
0.73
0.76
0.72
0.26
0.25
0.15
0.25
0.15
0.35
0.33
0.15
0.31
0.3
0.34
0.33
0.15
0.31
0.3
0.23
0.22
0.22
0.24
0.21
0.23
0.22
0.08
0.24
0.22
0.23
0.23
0.23
0.17
0.24
0.22
0.02
0.06
5.70E-03
0.02
0.02
7.10E-03
6.20E-03
8.00E-03
6.60E-03
5.40E-03
0.02
6.70E-03
7.70E-03
7.00E-03
4.10E-03
0.01
6.80E-03
0.21
0.2
0.65
0.37
0.17
0.28
0.18
0.31
0.18
0.09
0.07
0.42
0.33
0.15
0.36
0.16
0.21
0.14
0.33
0.16
0.23
0.12
44
45
46
47
48
49
50
51
52
53
100
100
100
100
100
100
100
100
100
100
2
2
2
2
2
2
2
2
2
2
100
100
100
100
100
100
100
100
100
100
0.91
0.98
0.91
0.98
0.91
0.98
0.99
1
0.89
0.97
0.64
0.53
0.64
0.53
0.64
0.53
0.63
0.31
0.64
0.54
0.5
0.56
0.97
1
0.93
0.99
0.9
0.98
0.5
0.56
0.29
0.23
0.14
0.07
0.23
0.12
0.37
0.11
0.3
0.24
0.76
0.72
0.76
0.72
0.76
0.72
0.79
0.55
0.75
0.72
0.02
0.06
0.26
0.18
0.32
0.25
0.41
0.24
0.02
0.07
0.14
0.12
0.27
0.17
0.25
0.2
0.14
0.07
1.16
1.01
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Vars
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
Deg
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
N
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
AP
0.92
1
0.93
1
0.93
1
0.92
1
0.93
1
0.95
0.99
1
0.97
1
0.95
0.99
1
0.97
1
0.93
1
0.93
1
-
AR
0.95
0.91
0.94
0.91
0.94
0.91
0.95
0.91
0.94
0.91
0.98
0.95
0.84
0.97
0.89
0.97
0.95
0.83
0.97
0.89
0.94
0.9
0.96
0.9
-
AHP
0.72
0.77
0.66
0.7
0.83
0.91
0.99
0.99
0.99
0.99
0.93
0.97
0.98
0.96
0.98
0.94
0.97
0.98
0.96
0.98
*
*
0.85
0.89
-
AHR
0.58
0.53
0.54
0.5
0.62
0.58
0.53
0.45
0.53
0.45
0.73
0.67
0.53
0.71
0.6
0.73
0.67
0.53
0.7
0.6
0.65
0.55
-
McAdj
0.93
0.95
0.93
0.95
0.93
0.95
0.93
0.95
0.93
0.95
0.96
0.97
0.91
0.97
0.94
0.96
0.97
0.91
0.97
0.94
0.93
0.94
0.94
0.95
-
McArrow
0.38
0.39
0.29
0.3
0.52
0.57
0.59
0.53
0.59
0.53
0.7
0.68
0.59
0.7
0.63
0.7
0.68
0.59
0.7
0.63
0.56
0.52
-
E
0.05
0.03
0.04
0.02
0.04
0.05
0.07
0.07
0.07
0.05
0.03
0.02
0.02
0.02
0.02
0.02
0.02
0.01
0.02
0.02
0.29
0.29
1.93
1.05
-
36
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Alg
1
2
3
4
5
6
7
Vars
100
100
100
100
100
100
100
Deg
2
2
2
2
2
2
2
N
1000
1000
1000
1000
1000
1000
1000
1
0.93
1
0.93
1
0.92
1
0.8
0.93
1
0.93
1
0.96
1
0.93
1
0.93
1
0.93
1
0.93
1
0.93
1
0.99
1
0.92
1
0.91
0.95
0.91
0.95
0.91
0.95
0.9
0.96
0.95
0.91
0.95
0.91
0.94
0.9
0.95
0.91
0.94
0.91
0.94
0.91
0.94
0.91
0.94
0.91
0.95
0.85
0.95
0.91
0.89
0.86
0.89
0.86
0.89
*
*
0.63
0.86
0.89
0.86
0.89
0.87
0.87
0.86
0.89
0.62
0.65
0.62
0.65
0.99
0.99
0.95
0.97
0.98
0.98
0.65
0.68
0.55
0.64
0.55
0.64
0.55
0.76
0.64
0.55
0.65
0.55
0.62
0.55
0.64
0.55
0.57
0.51
0.57
0.51
0.54
0.46
0.63
0.56
0.67
0.57
0.58
0.53
AP AR AHP AHR
0.94 0.99 0.77 0.64
0.99 0.97 0.81 0.63
0.94 0.99 0.73 0.62
0.99 0.97 0.77 0.6
0.94 0.99 0.89 0.7
0.99 0.97 0.95 0.68
0.94 0.99 1
0.63
37
0.95
0.94
0.95
0.94
0.95
0.93
0.95
0.87
0.94
0.95
0.94
0.95
0.95
0.95
0.94
0.95
0.93
0.95
0.93
0.95
0.93
0.95
0.93
0.95
0.97
0.92
0.93
0.95
0.52
0.56
0.52
0.57
0.52
0.41
0.57
0.52
0.57
0.52
0.55
0.5
0.57
0.52
0.25
0.24
0.25
0.24
0.6
0.54
0.64
0.6
0.69
0.61
0.28
0.29
0.49
0.47
0.51
0.89
0.43
0.13
0.12
0.73
0.68
0.39
0.64
0.4
0.77
0.54
0.62
0.41
0.35
0.2
0.24
0.19
0.91
0.61
0.83
0.67
0.2
0.17
1.25
1.11
McAdj McArrow E
0.96
0.46
0.06
0.98
0.5
0.04
0.96
0.41
0.1
0.98
0.43
0.04
0.96
0.63
0.09
0.98
0.67
0.05
0.96
0.67
0.15
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.99
0.94
0.99
0.95
0.99
1
0.99
0.99
0.96
0.99
1
0.99
0.99
0.94
0.99
0.94
0.99
0.99
0.94
0.99
0.94
0.99
0.94
0.99
0.84
0.94
0.99
0.94
0.99
0.97
1
0.94
0.99
0.94
0.99
0.94
0.99
0.97
0.99
0.97
0.99
0.99
0.95
0.99
0.96
0.99
0.98
0.95
0.98
0.96
0.99
0.96
0.99
0.97
0.97
0.99
0.97
0.99
0.97
0.99
0.97
0.99
0.99
0.97
0.99
0.97
0.99
0.97
0.99
0.97
0.99
0.97
0.99
0.97
38
1
1
1
0.93
0.96
0.97
0.97
0.97
0.93
0.96
0.97
0.97
0.97
*
*
0.91
0.94
0.94
0.91
0.94
0.91
0.94
*
*
0.68
0.91
0.94
0.91
0.94
0.94
0.94
0.91
0.94
0.71
0.74
0.71
0.74
0.57
0.63
0.57
0.74
0.71
0.67
0.72
0.69
0.73
0.7
0.66
0.71
0.68
0.69
0.65
0.66
0.68
0.65
0.68
0.66
0.8
0.68
0.66
0.69
0.66
0.68
0.65
0.68
0.66
0.63
0.61
0.63
0.61
0.98
0.96
0.98
0.97
0.99
0.97
0.99
0.98
0.98
0.99
0.97
0.99
0.98
0.96
0.98
0.96
0.98
0.98
0.96
0.98
0.96
0.98
0.96
0.98
0.91
0.96
0.98
0.96
0.98
0.98
0.98
0.96
0.98
0.96
0.98
0.96
0.98
0.63
0.67
0.63
0.7
0.71
0.68
0.72
0.69
0.7
0.7
0.68
0.71
0.69
0.64
0.64
0.65
0.64
0.64
0.64
0.65
0.48
0.64
0.65
0.64
0.65
0.66
0.64
0.64
0.65
0.39
0.4
0.39
0.4
0.08
0.18
0.1
0.05
0.04
0.03
0.03
0.03
0.04
0.03
0.03
0.03
0.03
0.39
0.39
4.2
2.39
0.8
0.71
0.71
1.11
0.74
0.15
0.17
1.03
0.94
0.64
0.96
0.65
1.62
1.11
0.92
0.62
0.33
0.29
0.3
0.19
46
47
48
49
50
51
52
53
100
100
100
100
100
100
100
100
2
2
2
2
2
2
2
2
1000
1000
1000
1000
1000
1000
1000
1000
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Vars
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
Deg
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
N
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.94
0.99
0.94
0.99
0.99
1
0.94
0.99
AP
0.96
0.99
0.97
0.99
0.97
0.99
0.96
0.99
0.97
0.99
0.91
0.94
0.98
0.84
0.96
0.93
0.95
0.98
0.89
0.97
0.97
1
0.97
0.99
1
0.96
0.99 1
0.97 1
0.99 0.97
0.97 0.99
0.98 0.97
0.94 0.96
0.99 0.73
0.97 0.77
AR
0.5
0.38
0.49
0.37
0.49
0.37
0.5
0.38
0.49
0.37
0.7
0.57
0.26
0.74
0.47
0.62
0.52
0.26
0.66
0.45
0.46
0.35
0.48
0.37
0.37
0.49
AHP
0.59
0.56
0.52
0.49
0.78
0.75
0.97
0.94
0.98
0.95
0.77
0.8
0.83
0.71
0.81
0.78
0.79
0.85
0.74
0.82
*
*
0.66
0.8
0.79
0.66
39
0.64
0.58
0.7
0.66
0.71
0.68
0.65
0.62
0.96
0.98
0.96
0.98
0.99
0.97
0.96
0.98
AHR
0.25
0.16
0.22
0.14
0.26
0.15
0.11
0.04
0.11
0.04
0.55
0.41
0.07
0.6
0.3
0.48
0.36
0.06
0.52
0.27
0.26
0.09
0.11
0.24
McAdj
0.69
0.61
0.68
0.6
0.68
0.6
0.69
0.61
0.68
0.6
0.79
0.72
0.49
0.78
0.66
0.75
0.69
0.49
0.76
0.65
0.66
0.59
0.67
0.6
0.6
0.68
0.69
0.64
0.7
0.69
0.71
0.69
0.42
0.44
McArrow
0.09
0.05
0.02
-0.01
0.25
0.16
0.22
0.13
0.23
0.12
0.41
0.35
0.13
0.4
0.3
0.38
0.32
0.12
0.38
0.29
0.17
0.15
0.15
0.15
1.22
0.94
1.15
0.95
0.21
0.17
1.35
1.21
E
0.03
0.02
0.02
5.90E-03
0.02
8.70E-03
0.03
0.02
0.04
6.90E-03
0.06
0.03
6.20E-03
0.11
0.02
0.02
0.02
5.50E-03
0.03
0.01
0.29
0.25
1.45
0.62
0.32
0.48
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
1
0.97
1
0.96
0.99
0.91
0.97
1
0.97
1
0.99
1
0.97
1
0.97
0.99
0.97
0.99
0.97
0.99
0.97
0.99
0.96
0.99
0.96
0.99
0.37
0.48
0.37
0.49
0.37
0.56
0.48
0.37
0.5
0.39
0.46
0.34
0.48
0.38
0.49
0.37
0.49
0.37
0.49
0.37
0.49
0.37
0.62
0.27
0.5
0.38
0.76
0.7
0.79
*
*
0.69
0.69
0.79
0.67
0.77
0.69
0.9
0.69
0.78
0.51
0.52
0.51
0.52
0.98
0.95
0.86
0.92
0.86
0.88
0.54
0.55
0.11
0.24
0.11
0.42
0.24
0.11
0.28
0.13
0.22
0.06
0.24
0.12
0.24
0.17
0.24
0.17
0.11
0.04
0.23
0.1
0.48
0.14
0.26
0.18
0.6
0.67
0.6
0.68
0.6
0.71
0.67
0.6
0.68
0.61
0.66
0.57
0.68
0.6
0.68
0.6
0.68
0.6
0.68
0.6
0.68
0.6
0.76
0.5
0.69
0.61
0.14
0.18
0.15
0.26
0.17
0.15
0.18
0.15
0.16
0.12
0.17
0.16
0.01
0.02
0.01
0.02
0.23
0.13
0.28
0.19
0.45
0.22
0.05
0.05
0.32
0.7
0.33
0.1
0.09
0.57
0.57
0.27
0.69
0.3
0.38
0.24
0.6
0.28
0.24
0.2
0.26
0.13
0.69
0.29
0.65
0.32
0.34
0.11
1.3
1.16
Alg
1
2
3
4
5
6
7
8
9
Vars
100
100
100
100
100
100
100
100
100
Deg
4
4
4
4
4
4
4
4
4
N
AP
500 0.98
500 0.99
500 0.98
500 0.99
500 0.98
500 0.99
500 0.98
500 0.99
500 0.98
AR
0.83
0.78
0.82
0.77
0.82
0.77
0.83
0.78
0.82
AHP
0.7
0.68
0.63
0.62
0.85
0.86
0.96
0.97
0.97
AHR
0.55
0.48
0.49
0.44
0.66
0.61
0.55
0.44
0.54
McAdj
0.9
0.87
0.9
0.87
0.9
0.87
0.9
0.87
0.9
McArrow
0.33
0.27
0.22
0.18
0.56
0.53
0.58
0.51
0.58
E
0.12
0.06
0.17
0.08
0.28
0.22
0.49
0.27
0.52
40
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
0.99
0.89
0.93
0.96
0.91
0.95
0.9
0.93
0.96
0.92
0.95
0.99
0.99
0.99
1
1
0.99
1
0.99
1
0.98
0.99
0.96
0.99
1
0.98
1
0.99
1
0.99
1
0.98
0.99
0.98
0.99
0.98
0.99
0.77
0.97
0.95
0.84
0.96
0.88
0.94
0.91
0.77
0.93
0.82
0.81
0.76
0.84
0.78
0.78
0.84
0.78
0.83
0.78
0.82
0.77
0.85
0.83
0.78
0.83
0.78
0.83
0.77
0.83
0.78
0.82
0.77
0.82
0.77
0.82
0.77
0.97
0.81
0.85
0.89
0.83
0.88
0.82
0.86
0.88
0.84
0.87
*
*
0.84
0.83
0.84
0.84
0.81
0.84
0.84
*
*
0.82
0.84
0.84
0.84
0.84
0.85
0.84
0.84
0.84
0.6
0.59
0.6
0.59
0.97
0.97
41
0.43
0.84
0.81
0.72
0.83
0.75
0.81
0.77
0.65
0.8
0.69
0.65
0.58
0.59
0.66
0.58
0.65
0.59
0.73
0.65
0.59
0.65
0.59
0.66
0.57
0.65
0.59
0.49
0.45
0.49
0.45
0.55
0.43
0.87
0.92
0.93
0.89
0.93
0.91
0.91
0.92
0.86
0.92
0.88
0.89
0.86
0.91
0.88
0.88
0.91
0.88
0.9
0.88
0.89
0.87
0.9
0.9
0.88
0.9
0.88
0.91
0.87
0.9
0.88
0.9
0.87
0.9
0.87
0.9
0.87
0.5
0.66
0.68
0.65
0.67
0.66
0.65
0.66
0.59
0.66
0.61
0.54
0.48
0.5
0.55
0.47
0.55
0.5
0.58
0.54
0.5
0.54
0.5
0.55
0.48
0.54
0.5
0.17
0.15
0.17
0.15
0.59
0.51
0.26
0.24
0.18
0.12
0.13
0.09
0.14
0.11
0.06
0.09
0.05
0.92
0.63
12.17
5.57
1.67
3.49
1.47
3.94
1.67
0.31
0.29
1.77
3.04
1.42
3.15
1.43
3.63
1.89
2.98
1.42
0.91
0.79
0.96
0.57
5.98
3.9
48
49
50
51
52
53
100
100
100
100
100
100
4
4
4
4
4
4
500
500
500
500
500
500
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
Vars
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
Deg
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
N
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.98
0.99
0.96
0.98
0.98
0.99
AP
0.99
1
0.99
1
0.99
1
0.99
1
0.99
1
0.87
0.9
0.94
0.9
0.93
0.88
0.9
0.94
0.89
0.93
0.99
1
0.3
0.2
1
0.99
1
0.99
0.82
0.77
0.93
0.84
0.83
0.78
0.88
0.9
0.91
0.93
0.63
0.63
AR
0.88
0.85
0.87
0.84
0.87
0.84
0.88
0.85
0.87
0.84
0.99
0.98
0.93
0.99
0.95
0.98
0.97
0.9
0.97
0.92
0.87
0.84
0.27
0.17
0.86
0.89
0.86
0.89
42
0.68
0.62
0.82
0.72
0.52
0.48
AHP AHR
0.73 0.6
0.74 0.58
0.68 0.55
0.66 0.52
0.9
0.72
0.9
0.68
0.97 0.68
0.96 0.6
0.97 0.68
0.96 0.59
0.81 0.87
0.85 0.86
0.88 0.81
0.84 0.86
0.87 0.82
0.82 0.85
0.84 0.84
0.88 0.77
0.83 0.85
0.87 0.79
*
*
0.26 0.22
0.17 0.13
0.88 0.68
0.9
0.75
0.87 0.68
0.89 0.74
0.9
0.87
0.95
0.9
0.9
0.87
McAdj
0.93
0.92
0.93
0.91
0.93
0.91
0.93
0.92
0.93
0.91
0.93
0.94
0.93
0.94
0.94
0.92
0.93
0.92
0.93
0.92
0.93
0.91
0.28
0.18
0.92
0.94
0.92
0.94
0.6
0.58
0.75
0.69
0.22
0.21
McArrow
0.39
0.39
0.3
0.26
0.65
0.62
0.69
0.62
0.69
0.62
0.69
0.72
0.71
0.72
0.71
0.68
0.69
0.68
0.69
0.68
0.19
0.11
0.6
0.67
0.6
0.67
6.12
4.31
0.67
0.5
2.27
1.84
E
0.25
0.18
0.47
0.21
0.6
0.49
1.28
0.71
1.41
0.93
0.48
0.42
0.26
0.19
0.12
0.31
0.25
0.14
0.13
0.09
1.82
1.32
22.24
4.75
6.06
21.54
5.86
19.52
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
Alg Vars
1
100
2
100
3
100
4
100
5
100
6
100
7
100
8
100
9
100
10
100
11
100
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1
0.99
1
0.97
0.99
1
0.99
1
0.99
1
0.99
1
0.99
1
0.99
1
0.99
1
0.99
1
0.95
0.97
0.99
1
0.86
0.88
0.85
0.9
0.89
0.86
0.89
0.86
0.89
0.85
0.89
0.86
0.87
0.84
0.87
0.84
0.87
0.84
0.87
0.84
0.98
0.93
0.88
0.85
Deg
6
6
6
6
6
6
6
6
6
6
6
N
AP AR
100 0.95 0.37
100 0.98 0.3
100 0.96 0.36
100 0.98 0.28
100 0.96 0.36
100 0.98 0.28
100 0.95 0.37
100 0.98 0.3
100 0.96 0.36
100 0.98 0.28
100 0.85 0.65
0.88
*
*
0.85
0.89
0.88
0.89
0.88
0.89
0.86
0.89
0.87
0.64
0.62
0.64
0.62
0.97
0.96
0.91
0.91
0.91
0.94
0.66
0.65
0.68
0.79
0.74
0.68
0.74
0.69
0.74
0.68
0.74
0.68
0.55
0.51
0.55
0.51
0.68
0.6
0.74
0.7
0.86
0.82
0.57
0.54
0.92
0.93
0.92
0.93
0.94
0.92
0.94
0.92
0.94
0.92
0.94
0.92
0.93
0.91
0.93
0.91
0.93
0.91
0.93
0.91
0.96
0.95
0.93
0.91
AHP
0.55
0.52
0.5
0.46
0.71
0.72
0.89
0.94
0.9
0.97
0.71
AHR
0.18
0.12
0.16
0.11
0.21
0.12
0.07
0.03
0.07
0.03
0.52
McAdj
0.58
0.52
0.58
0.51
0.58
0.51
0.58
0.52
0.58
0.51
0.73
43
0.6
0.65
0.67
0.6
0.67
0.6
0.66
0.58
0.67
0.6
0.24
0.2
0.24
0.2
0.69
0.62
0.67
0.65
0.78
0.77
0.28
0.25
McArrow
0.04
0.02
-1.16E-05
-0.03
0.17
0.14
0.16
0.1
0.16
0.1
0.34
6.36
0.63
0.56
3.37
18.31
5.53
19.61
5.89
27.46
10.38
18.19
5.45
1.32
1.38
1.31
0.89
8.93
6.38
9.89
7.07
0.61
0.49
2.8
2.27
E
0.02
0.03
0.03
0.02
0.03
0.02
0.03
0.03
0.06
0.03
0.16
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.91
0.96
0.8
0.93
0.88
0.92
0.96
0.85
0.94
0.98
0.99
0.98
0.99
1
0.98
0.99
0.99
1
0.97
0.98
0.95
0.99
1
0.99
1
0.99
1
0.99
1
0.96
0.98
0.96
0.98
0.96
0.98
0.96
0.98
0.54
0.23
0.7
0.47
0.52
0.43
0.22
0.58
0.38
0.34
0.26
0.35
0.26
0.27
0.38
0.28
0.36
0.27
0.36
0.29
0.42
0.36
0.27
0.37
0.28
0.34
0.25
0.36
0.28
0.36
0.28
0.36
0.28
0.36
0.28
0.36
0.28
0.75
0.77
0.66
0.77
0.71
0.72
0.76
0.69
0.75
*
*
0.66
0.75
0.75
0.64
0.67
0.65
0.74
*
*
0.68
0.66
0.75
0.67
0.72
0.64
0.74
0.66
0.73
0.52
0.52
0.52
0.52
0.92
0.91
0.75
0.83
44
0.43
0.1
0.57
0.36
0.41
0.33
0.09
0.46
0.27
0.19
0.09
0.09
0.22
0.1
0.2
0.1
0.3
0.2
0.09
0.22
0.1
0.18
0.06
0.2
0.1
0.19
0.13
0.19
0.13
0.07
0.03
0.2
0.09
0.69
0.46
0.73
0.65
0.67
0.62
0.45
0.69
0.58
0.57
0.49
0.57
0.5
0.51
0.59
0.52
0.58
0.51
0.58
0.52
0.62
0.59
0.51
0.59
0.52
0.57
0.49
0.59
0.51
0.58
0.51
0.58
0.51
0.58
0.51
0.58
0.51
0.33
0.14
0.33
0.31
0.28
0.25
0.13
0.29
0.24
0.13
0.12
0.12
0.12
0.09
0.13
0.12
0.19
0.13
0.12
0.15
0.11
0.11
0.09
0.14
0.11
0.02
0.02
0.02
0.02
0.16
0.1
0.2
0.15
0.12
0.02
0.22
0.07
0.08
0.03
1.00E-02
0.07
0.02
0.34
0.32
2.56
0.99
0.43
0.94
0.55
0.83
0.55
0.15
0.15
0.75
0.82
0.41
0.99
0.44
0.5
0.3
0.82
0.4
0.36
0.28
0.37
0.17
0.95
0.41
1.72
0.42
50
51
52
53
100
100
100
100
6
6
6
6
100
100
100
100
0.93
0.95
0.95
0.98
0.55
0.25
0.37
0.3
0.82
0.82
0.55
0.56
0.46
0.15
0.2
0.15
0.71
0.47
0.58
0.53
0.4
0.2
0.05
0.05
0.54
0.17
1.55
1.27
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
Vars
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
Deg
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
N
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
AP
0.96
0.97
0.97
0.98
0.97
0.98
0.96
0.97
0.97
0.98
0.75
0.81
0.91
0.77
0.88
0.77
0.83
0.91
0.79
0.89
0.98
0.99
0.2
1
1
0.99
1
0.99
1
0.98
AR
0.65
0.59
0.64
0.58
0.64
0.58
0.65
0.59
0.64
0.58
0.95
0.92
0.83
0.94
0.88
0.92
0.87
0.7
0.9
0.77
0.62
0.56
0.13
0.61
0.62
0.68
0.62
0.67
0.62
0.64
AHP
0.64
0.65
0.6
0.59
0.77
0.8
0.92
0.94
0.93
0.94
0.67
0.72
0.82
0.68
0.79
0.69
0.74
0.81
0.7
0.79
*
*
0.16
0.76
0.75
0.79
0.75
0.79
0.75
*
AHR
0.4
0.37
0.37
0.33
0.48
0.44
0.41
0.32
0.41
0.31
0.82
0.8
0.72
0.81
0.77
0.79
0.75
0.6
0.78
0.67
0.1
0.42
0.42
0.49
0.42
0.48
0.43
-
McAdj
0.78
0.75
0.78
0.74
0.78
0.74
0.78
0.75
0.78
0.74
0.83
0.85
0.86
0.84
0.87
0.83
0.84
0.79
0.83
0.82
0.77
0.74
0.16
0.77
0.78
0.81
0.78
0.81
0.78
0.78
McArrow
0.2
0.2
0.14
0.12
0.37
0.38
0.46
0.39
0.46
0.39
0.51
0.54
0.59
0.51
0.59
0.5
0.52
0.49
0.51
0.52
0.08
0.32
0.31
0.39
0.31
0.39
0.32
-
E
0.24
0.16
0.41
0.15
0.54
0.3
0.98
0.43
1.07
0.48
1.37
0.88
0.41
0.74
0.28
1.03
0.55
0.16
0.55
0.16
1.78
1.23
14.46
18.05
3.75
34.76
4.37
18.05
3.96
0.7
45
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
Vars
100
100
100
100
100
100
100
100
100
100
100
100
100
Deg
6
6
6
6
6
6
6
6
6
6
6
6
6
N
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.98
0.97
0.99
1
0.99
1
1
1
0.99
1
0.97
0.98
0.97
0.98
0.97
0.98
0.97
0.98
0.9
0.95
0.96
0.97
AP
0.97
0.98
0.97
0.98
0.97
0.98
0.97
0.98
0.97
0.98
0.72
0.76
0.84
0.58
0.67
0.67
0.62
0.68
0.62
0.68
0.62
0.67
0.62
0.64
0.58
0.64
0.58
0.64
0.58
0.64
0.58
0.9
0.79
0.65
0.59
*
0.85
0.79
0.75
0.79
0.75
0.77
0.76
0.79
0.75
0.51
0.51
0.51
0.51
0.93
0.93
0.77
0.81
0.83
0.89
0.52
0.52
AR
0.72
0.68
0.71
0.67
0.71
0.67
0.72
0.68
0.71
0.67
0.97
0.96
0.91
46
AHP
0.64
0.65
0.6
0.59
0.81
0.82
0.93
0.94
0.94
0.94
0.64
0.68
0.75
0.59
0.49
0.42
0.49
0.43
0.48
0.43
0.49
0.42
0.33
0.3
0.33
0.3
0.41
0.31
0.49
0.45
0.81
0.71
0.35
0.31
AHR
0.46
0.43
0.42
0.38
0.56
0.53
0.53
0.45
0.52
0.45
0.84
0.83
0.8
0.74
0.8
0.81
0.78
0.81
0.78
0.82
0.78
0.81
0.78
0.78
0.74
0.78
0.74
0.78
0.74
0.78
0.74
0.89
0.86
0.78
0.75
0.51
0.39
0.31
0.39
0.31
0.37
0.33
0.39
0.31
0.02
0.02
0.02
0.02
0.46
0.39
0.37
0.38
0.66
0.65
0.03
0.03
McAdj McArrow
0.83
0.22
0.81
0.21
0.82
0.15
0.8
0.13
0.82
0.45
0.8
0.45
0.83
0.55
0.81
0.5
0.82
0.55
0.8
0.5
0.82
0.49
0.84
0.52
0.86
0.57
0.58
3.67
231.15
3.38
25.51
3.54
14.2
4.1
20.48
3.33
2.26
1.62
2.31
1.18
12.1
6.28
16.28
6.99
1.25
0.88
3.81
2.76
E
0.66
0.33
1.24
0.69
1.51
1.06
3.23
1.41
3.54
1.7
3.42
2.46
1.16
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.75
0.82
0.72
0.77
0.85
0.75
0.82
0.98
0.99
1
0.1
1
1
0.98
0.98
0.97
0.1
1
1
1
1
0.97
0.98
0.97
0.98
0.97
0.98
0.97
0.98
0.84
0.91
0.96
0.93
0.96
0.94
0.85
0.95
0.88
0.7
0.66
0.69
0.07
0.7
0.69
0.71
0.67
0.73
0.07
0.69
0.69
0.7
0.69
0.71
0.67
0.71
0.67
0.71
0.67
0.71
0.67
0.94
0.89
47
0.66
0.73
0.64
0.69
0.76
0.67
0.73
*
*
0.8
0.08
0.81
0.8
*
*
0.86
0.08
0.8
0.8
0.8
0.8
0.57
0.56
0.57
0.56
0.94
0.94
0.81
0.82
0.77
0.84
0.83
0.81
0.83
0.81
0.74
0.82
0.76
0.51
0.05
0.52
0.51
0.64
0.05
0.51
0.51
0.52
0.51
0.42
0.38
0.42
0.38
0.52
0.45
0.57
0.54
0.83
0.8
0.83
0.86
0.82
0.84
0.84
0.83
0.84
0.82
0.8
0.82
0.08
0.83
0.82
0.82
0.8
0.83
0.08
0.82
0.82
0.83
0.82
0.82
0.8
0.82
0.8
0.82
0.8
0.82
0.8
0.88
0.89
0.51
0.56
0.48
0.52
0.53
0.5
0.53
0.41
0.04
0.42
0.41
0.56
0.04
0.41
0.41
0.42
0.41
0.12
0.09
0.12
0.09
0.55
0.49
0.46
0.45
0.62
0.66
1.2
0.57
2.81
1.88
0.63
1.16
0.41
4.21
2.84
50.12
-9.90E-03
82.52
50.31
2.49
1.46
7.42
58.65
46.87
54.84
54.88
53.1
3.99
2.87
4.14
2.29
20.2
16.28
23.95
14.37
1.66
1.19
52
53
100
100
6
6
1000
1000
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
Vars
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Deg
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
N
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.97
0.98
AP
0.74
0.93
0.79
0.95
0.79
0.95
0.74
0.93
0.79
0.95
0.73
0.94
1
0.44
0.98
0.75
0.94
1
0.6
0.98
0.78
0.94
0.94
0.94
0.93
0.94
0.75
0.94
0.58
0.72
0.68
AR
0.66
0.55
0.65
0.54
0.65
0.54
0.66
0.55
0.65
0.54
0.72
0.62
0.34
0.8
0.56
0.71
0.62
0.33
0.75
0.56
0.62
0.53
0.54
0.54
0.53
0.54
0.65
0.54
0.68
0.57
0.57
0.43
0.4
0.83
0.81
AHP
0.42
0.52
0.43
0.5
0.63
0.84
0.96
0.99
0.97
0.99
0.6
0.85
0.96
0.35
0.89
0.62
0.85
0.95
0.48
0.89
*
*
0.87
0.88
0.88
0.88
*
*
0.39
AHR
0.29
0.18
0.27
0.17
0.3
0.19
0.12
0.05
0.12
0.05
0.49
0.29
0.05
0.61
0.23
0.47
0.29
0.05
0.56
0.23
0.11
0.11
0.1
0.11
0.47
McAdj
0.69
0.71
0.71
0.71
0.71
0.71
0.69
0.71
0.71
0.71
0.72
0.76
0.58
0.59
0.74
0.73
0.76
0.58
0.67
0.74
0.69
0.7
0.71
0.71
0.71
0.71
0.69
0.72
0.63
48
0.12
0.11
McArrow
-0.05
0.03
-0.04
7.10E-03
0.17
0.24
0.24
0.15
0.24
0.15
0.23
0.32
0.15
0.04
0.3
0.24
0.32
0.15
0.14
0.3
0.19
0.19
0.19
0.19
-9.76E-03
5.28
5.08
E
1.22
1.22
0.18
0.06
0.12
0.05
1.6
1.12
0.13
0.07
0.18
0.05
0.03
1.84
0.08
0.09
0.04
0.03
0.13
0.06
5.61
5.77
15.19
4.91
3.98
4.7
2.62
1.88
12.52
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.94
0.93
0.98
0.94
0.79
0.95
0.79
0.95
0.79
0.95
0.79
0.95
0.97
1
0.74
0.94
0.54
0.55
0.51
0.54
0.65
0.54
0.65
0.54
0.65
0.54
0.65
0.54
0.66
0.32
0.65
0.55
0.88
0.87
0.92
0.88
0.41
0.5
0.41
0.5
0.96
0.99
0.89
0.99
0.89
0.99
0.41
0.51
0.11
0.12
0.07
0.11
0.3
0.2
0.3
0.2
0.12
0.06
0.23
0.11
0.4
0.11
0.32
0.21
0.71
0.71
0.7
0.71
0.71
0.71
0.71
0.71
0.71
0.71
0.71
0.71
0.8
0.56
0.7
0.71
0.19
0.19
0.16
0.2
-0.07
3.60E-03
-0.07
3.60E-03
0.24
0.17
0.3
0.24
0.42
0.24
-0.07
0.02
5.03
6.06
3.76
5.11
3.44
3.26
3.37
3.08
4.15
5.3
4.52
4.15
3.33
1.75
29.44
30.23
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Vars
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Deg
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
N
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
AP
0.78
0.97
0.8
0.98
0.8
0.98
0.78
0.97
0.8
0.98
0.79
0.98
1
0.91
1
AR
0.95
0.91
0.95
0.91
0.95
0.91
0.95
0.91
0.95
0.91
0.98
0.95
0.84
0.96
0.89
AHP
0.53
0.71
0.53
0.68
0.7
0.92
0.99
0.99
0.99
0.99
0.73
0.95
0.98
0.87
0.97
AHR
0.55
0.52
0.54
0.5
0.68
0.63
0.55
0.45
0.55
0.45
0.79
0.7
0.57
0.74
0.63
McAdj
0.86
0.94
0.87
0.94
0.87
0.94
0.86
0.94
0.87
0.94
0.88
0.96
0.92
0.93
0.94
McArrow
0.16
0.33
0.15
0.28
0.44
0.61
0.61
0.53
0.61
0.53
0.55
0.68
0.62
0.65
0.66
E
1.56
1.23
0.39
0.18
0.46
0.25
1.66
1.53
0.73
0.29
0.58
0.16
0.11
0.22
0.17
49
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
0.83
0.98
1
0.91
1
0.81
0.97
0.97
0.97
0.97
0.97
0.79
0.97
0.54
0.97
0.97
0.99
0.97
0.8
0.98
0.8
0.98
0.8
0.98
0.8
0.98
1
1
0.78
0.97
0.94
0.84
0.96
0.89
0.95
0.91
0.91
0.91
0.91
0.91
0.95
0.91
0.97
0.91
0.91
0.9
0.91
0.95
0.91
0.95
0.91
0.95
0.91
0.95
0.91
0.95
0.84
0.95
0.77
0.95
0.98
0.88
0.97
*
*
0.87
0.86
0.86
0.86
*
*
0.43
0.86
0.86
0.86
0.86
0.51
0.65
0.51
0.65
0.99
1
0.93
0.99
0.98
0.99
0.52
50
0.78
0.7
0.57
0.74
0.63
0.6
0.6
0.59
0.6
0.78
0.6
0.6
0.58
0.6
0.58
0.54
0.58
0.54
0.55
0.46
0.65
0.57
0.7
0.59
0.59
0.9
0.96
0.92
0.94
0.94
0.88
0.94
0.94
0.94
0.94
0.94
0.87
0.94
0.72
0.94
0.94
0.94
0.94
0.87
0.94
0.87
0.94
0.87
0.94
0.87
0.94
0.97
0.92
0.86
0.58
0.68
0.62
0.66
0.66
0.53
0.53
0.52
0.53
0.22
0.53
0.53
0.52
0.53
0.13
0.25
0.13
0.25
0.61
0.54
0.63
0.63
0.72
0.64
0.15
0.16
0.13
0.1
0.17
0.15
7.37
6.33
44.89
12.23
8.48
11.66
2.73
2.19
43.54
12.68
13.02
15.67
13.27
4.93
3.6
4.24
3.57
9.9
8.05
10.6
7.09
4.65
3.77
31.43
53
500
2
500
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
Vars
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Deg
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
N
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.97
AP
0.77
0.98
0.78
0.98
0.78
0.98
0.77
0.98
0.78
0.98
0.78
0.97
1
0.95
1
0.82
0.98
1
0.95
1
0.8
0.98
0.98
0.98
0.98
0.78
0.98
0.55
-
0.91
0.68
AR
0.99
0.98
0.99
0.98
0.99
0.98
0.99
0.98
0.99
0.98
1
0.99
0.95
0.99
0.96
0.99
0.99
0.95
0.99
0.96
0.99
0.97
0.97
0.97
0.97
0.99
0.98
0.99
51
0.55
AHP AHR
0.57 0.62
0.8
0.62
0.56 0.6
0.76 0.6
0.71 0.73
0.95 0.69
0.99 0.66
1
0.6
0.99 0.66
1
0.6
0.73 0.82
0.96 0.74
0.99 0.68
0.93 0.75
0.99 0.69
0.77 0.81
0.96 0.74
0.99 0.68
0.93 0.75
0.99 0.69
*
*
0.94 0.68
0.94 0.67
0.94 0.68
*
*
0.44 0.81
-
0.94
McAdj
0.87
0.98
0.88
0.98
0.88
0.98
0.87
0.98
0.88
0.98
0.88
0.98
0.97
0.97
0.98
0.9
0.98
0.97
0.97
0.98
0.89
0.98
0.97
0.97
0.97
0.88
0.98
0.74
-
0.3
McArrow
0.25
0.48
0.22
0.43
0.48
0.68
0.69
0.65
0.7
0.65
0.57
0.73
0.71
0.71
0.72
0.61
0.73
0.71
0.72
0.72
0.66
0.66
0.66
0.25
-
28.73
E
2.02
1.41
0.78
0.37
0.89
0.53
2.36
1.89
1.58
0.71
0.97
0.29
0.2
0.31
0.26
0.29
0.22
0.2
0.27
0.26
8.78
7.86
18.73
11.53
17.49
3.04
2.71
166.5
-
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Vars
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Deg
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
N
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.98
0.97
0.99
0.97
0.78
0.98
0.78
0.98
0.78
0.98
0.78
0.98
1
1
0.77
0.98
AP
0.91
0.98
0.93
0.99
0.93
0.99
0.91
0.98
0.93
0.99
0.78
0.94
0.99
0.59
0.97
0.86
0.97 0.94
0.97 0.94
0.97 0.94
0.97 0.94
0.99 0.53
0.98 0.71
0.99 0.53
0.98 0.71
0.99 0.99
0.98 1
0.99 0.94
0.98 0.99
0.99 0.99
0.95 0.99
0.99 0.53
0.98 0.74
AR
0.5
0.4
0.49
0.38
0.49
0.38
0.5
0.4
0.49
0.38
0.73
0.56
0.26
0.78
0.48
0.62
AHP
0.54
0.56
0.5
0.52
0.76
0.85
0.97
0.98
0.98
0.98
0.69
0.84
0.93
0.53
0.88
0.76
52
0.68
0.68
0.67
0.68
0.63
0.61
0.63
0.61
0.66
0.6
0.71
0.67
0.73
0.68
0.64
0.63
0.97
0.97
0.98
0.97
0.88
0.98
0.88
0.98
0.88
0.98
0.88
0.98
1
0.97
0.87
0.98
AHR McAdj
0.24 0.67
0.16 0.62
0.22 0.67
0.15 0.62
0.3
0.67
0.17 0.62
0.13 0.67
0.05 0.62
0.13 0.67
0.05 0.62
0.62 0.75
0.43 0.73
0.07 0.51
0.68 0.68
0.33 0.68
0.51 0.73
0.66
0.66
0.66
0.66
0.19
0.37
0.19
0.37
0.69
0.66
0.69
0.71
0.75
0.71
0.21
0.41
McArrow
0.05
0.05
7.30E-03
0.02
0.26
0.23
0.25
0.16
0.25
0.15
0.4
0.41
0.17
0.27
0.36
0.39
18.93
19.45
36.99
18.76
5.2
4.02
4.37
3.99
12.05
7.72
11.86
12.8
4.92
4.44
31.72
28.56
E
1.17
1.36
0.15
0.1
0.18
0.11
1.14
1.2
0.19
0.11
1.64
0.35
0.04
3.95
0.16
0.16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.95
0.99
0.77
0.97
0.94
0.99
0.98
0.98
0.98
0.98
0.92
0.99
0.82
0.98
0.98
1
0.98
0.93
0.99
0.93
0.99
0.93
0.99
0.93
0.99
0.97
0.99
0.91
0.98
0.52
0.26
0.66
0.47
0.46
0.36
0.37
0.37
0.38
0.37
0.49
0.39
0.54
0.37
0.38
0.34
0.37
0.49
0.38
0.49
0.38
0.49
0.38
0.49
0.38
0.65
0.29
0.5
0.4
0.85
0.94
0.68
0.88
*
*
0.82
0.8
0.79
0.8
*
*
0.62
0.8
0.78
0.79
0.79
0.49
0.5
0.49
0.5
0.98
0.98
0.91
0.98
0.92
0.97
0.51
0.52
53
0.39
0.07
0.56
0.31
0.12
0.12
0.12
0.12
0.41
0.12
0.14
0.08
0.13
0.25
0.17
0.25
0.17
0.13
0.05
0.27
0.12
0.54
0.17
0.26
0.18
0.7
0.51
0.71
0.67
0.66
0.6
0.6
0.6
0.61
0.6
0.67
0.62
0.66
0.6
0.61
0.58
0.6
0.67
0.62
0.67
0.62
0.67
0.62
0.67
0.62
0.79
0.53
0.67
0.62
0.38
0.17
0.35
0.35
0.18
0.17
0.17
0.17
0.2
0.17
0.18
0.13
0.17
4.03E-03
1.29E-03
4.03E-03
1.29E-03
0.25
0.16
0.34
0.24
0.54
0.29
0.02
0.02
0.08
0.03
0.26
0.1
6.66
6.92
29.68
8.99
6.98
7.85
1.87
1.7
11.78
7.81
9.33
5.36
7.95
3.7
3.65
3.9
3.01
6.32
4.9
6.73
6.69
9.87
3.17
31.32
31.43
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
Vars
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Deg
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
N
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
AP
0.96
1
0.96
1
0.96
1
0.96
1
0.96
1
0.84
0.97
0.99
0.92
0.99
0.91
0.98
0.99
0.95
0.99
0.97
1
0.1
1
1
0.4
0.96
1
0.86
1
-
AR
0.87
0.81
0.87
0.8
0.87
0.8
0.87
0.81
0.87
0.8
0.98
0.95
0.81
0.97
0.87
0.95
0.92
0.78
0.94
0.84
0.86
0.79
0.08
0.79
0.79
0.31
0.87
0.8
0.9
0.79
-
AHP
0.68
0.69
0.63
0.63
0.88
0.93
0.99
0.99
0.99
0.99
0.81
0.95
0.98
0.9
0.98
0.89
0.96
0.98
0.93
0.98
*
*
0.09
0.84
0.85
0.34
*
*
0.78
0.84
54
AHR
0.58
0.51
0.53
0.47
0.72
0.66
0.6
0.48
0.6
0.48
0.91
0.85
0.7
0.89
0.77
0.88
0.82
0.67
0.86
0.74
0.06
0.61
0.61
0.24
0.81
0.61
-
McAdj
0.91
0.89
0.91
0.89
0.91
0.89
0.91
0.89
0.91
0.89
0.91
0.96
0.9
0.95
0.93
0.93
0.95
0.88
0.94
0.91
0.91
0.89
0.09
0.89
0.89
0.35
0.91
0.89
0.88
0.89
-
McArrow
0.32
0.29
0.23
0.2
0.64
0.64
0.65
0.56
0.65
0.56
0.73
0.82
0.72
0.79
0.77
0.77
0.8
0.69
0.8
0.75
0.05
0.52
0.52
0.21
0.6
0.52
-
E
2.16
1.76
1.46
0.63
2.33
1.41
5.91
3.02
5.15
2
3.6
1.32
0.64
1.34
0.58
0.66
0.5
0.24
0.48
0.36
12.2
10.12
24.44
38.28
24.05
15.19
3.88
3
34.09
38.23
-
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Vars
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Deg
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
N
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1
1
1
0.96
1
0.96
1
0.96
1
0.96
1
0.99
1
0.96
1
AP
0.97
1
0.97
1
0.97
1
0.97
1
0.97
1
0.84
0.97
0.99
0.95
0.99
0.91
0.98
0.99
0.79
0.79
0.79
0.87
0.8
0.87
0.8
0.87
0.8
0.87
0.8
0.96
0.86
0.87
0.81
0.85
0.83
0.84
0.58
0.59
0.58
0.59
0.99
0.99
0.93
0.97
0.98
0.99
0.61
0.63
AR
0.94
0.91
0.94
0.91
0.94
0.91
0.94
0.91
0.94
0.91
1
0.99
0.94
0.99
0.95
0.99
0.98
0.92
55
AHP
0.74
0.75
0.69
0.69
0.91
0.95
1
1
1
1
0.81
0.95
0.97
0.93
0.97
0.88
0.96
0.98
0.61
0.59
0.61
0.51
0.47
0.51
0.47
0.6
0.48
0.73
0.66
0.86
0.76
0.54
0.49
AHR
0.68
0.64
0.63
0.59
0.82
0.79
0.75
0.66
0.75
0.66
0.93
0.9
0.84
0.9
0.86
0.92
0.89
0.82
0.89
0.88
0.89
0.91
0.89
0.91
0.89
0.91
0.89
0.91
0.89
0.98
0.93
0.91
0.9
0.52
0.5
0.52
0.16
0.15
0.16
0.15
0.65
0.56
0.7
0.67
0.85
0.77
0.21
0.2
39.53
49.26
38.68
9.42
6.37
9.39
5.68
53.57
40.38
64.25
41.07
13.45
11.07
36.45
35.1
McAdj McArrow E
0.96
0.44
3.82
0.95
0.42
2.62
0.96
0.36
4.33
0.95
0.33
2.32
0.96
0.75
7.54
0.95
0.76
5.45
0.96
0.77
19.59
0.95
0.7
9.48
0.96
0.77
19.82
0.95
0.7
11.6
0.92
0.74
5.58
0.98
0.85
2.52
0.96
0.83
1.6
0.97
0.84
1.29
0.97
0.84
0.83
0.95
0.81
1.42
0.98
0.85
1
0.95
0.81
0.59
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
Alg
Vars
Deg
N
0.97
0.99
0.97
1
1
1
0.97
1
0.88
0.97
1
0.97
1
0.97
1
0.97
1
0.99
0.99
0.97
1
AP
0.98
0.94
0.94
0.9
0.9
0.9
0.94
0.91
0.96
0.94
0.91
0.94
0.91
0.94
0.91
0.94
0.91
0.99
0.95
0.94
0.91
AR
0.94
0.97
*
*
0.89
0.89
*
*
0.81
0.62
0.62
0.62
0.62
0.99
0.99
0.94
0.96
0.98
0.98
0.65
0.66
0.89
0.84
0.74
0.74
0.88
0.59
0.56
0.59
0.56
0.75
0.67
0.83
0.78
0.9
0.86
0.62
0.59
0.97
0.96
0.95
0.95
0.95
0.95
0.96
0.95
0.92
0.96
0.95
0.96
0.95
0.96
0.95
0.96
0.95
0.99
0.97
0.96
0.95
AHP
AHR
McAdj
56
0.85
0.83
0.66
0.67
0.69
0.24
0.22
0.24
0.22
0.77
0.7
0.78
0.77
0.88
0.85
0.29
0.29
McArrow
0.73
0.49
21.43
17.43
112.67
97.54
6.73
5.84
65.41
12.8
9.52
12.13
8.49
101.85
72.33
127.52
74.05
14.52
12.54
43.26
43.87
E
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.94
0.99
0.96
0.99
0.96
0.99
0.94
0.99
0.96
0.99
0.81
0.93
0.98
0.66
0.96
0.89
0.95
0.99
0.83
0.97
0.97
0.99
0.99
0.99
0.96
0.99
0.91
-
0.39
0.3
0.38
0.28
0.38
0.28
0.39
0.3
0.38
0.28
0.73
0.6
0.22
0.77
0.5
0.56
0.46
0.22
0.61
0.4
0.34
0.26
0.26
0.27
0.37
0.29
0.42
-
0.56
0.55
0.52
0.5
0.76
0.83
0.98
0.99
0.98
0.98
0.74
0.85
0.88
0.6
0.87
0.8
0.84
0.89
0.74
0.86
*
*
0.77
0.76
*
*
0.71
57
0.19
0.13
0.17
0.11
0.24
0.13
0.09
0.04
0.09
0.04
0.65
0.52
0.09
0.7
0.41
0.48
0.38
0.08
0.54
0.31
0.09
0.09
0.33
-
0.6
0.54
0.6
0.53
0.6
0.53
0.6
0.54
0.6
0.53
0.77
0.75
0.47
0.71
0.69
0.7
0.65
0.46
0.71
0.62
0.57
0.51
0.51
0.51
0.6
0.53
0.62
-
0.06
0.03
0.02
8.32E-04
0.22
0.19
0.21
0.14
0.21
0.13
0.47
0.47
0.17
0.36
0.42
0.4
0.37
0.17
0.39
0.34
0.13
0.13
0.24
-
1.22
1.24
0.26
0.16
0.28
0.14
1.33
1.35
0.32
0.12
4.26
2.4
0.07
9.46
1.29
0.31
0.17
0.06
0.69
0.15
6.78
6.08
11.9
9.87
2.12
1.65
13.55
-
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
0.96
0.99
0.96
0.99
0.96
0.99
0.96
0.99
0.96
0.99
0.94
0.99
0.38
0.28
0.38
0.28
0.38
0.28
0.38
0.28
0.64
0.26
0.39
0.3
0.5
0.49
0.5
0.49
0.97
0.98
0.87
0.96
0.91
0.94
0.52
0.51
0.19
0.13
0.19
0.13
0.09
0.04
0.22
0.09
0.58
0.17
0.2
0.14
0.6
0.53
0.6
0.53
0.6
0.53
0.6
0.53
0.78
0.5
0.6
0.54
8.02E-03 4.39
-4.94E-03 4.06
8.02E-03 5.01
-4.94E-03 3.01
0.21
8.23
0.14
4.56
0.28
11.5
0.21
4.86
0.57
25.03
0.28
4.64
0.02
32.08
0.01
32.29
Alg
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Vars
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Deg
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
N
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
AP
0.99
1
0.99
1
0.99
1
0.99
1
0.99
1
0.86
0.95
0.98
0.91
0.98
0.92
0.96
0.98
0.95
AR
0.75
0.68
0.74
0.67
0.74
0.67
0.75
0.68
0.74
0.67
0.98
0.96
0.87
0.98
0.91
0.93
0.89
0.73
0.92
AHP
0.65
0.65
0.6
0.59
0.88
0.9
0.99
0.99
0.99
0.99
0.84
0.93
0.96
0.89
0.96
0.9
0.94
0.96
0.92
AHR
0.48
0.42
0.43
0.38
0.64
0.58
0.52
0.4
0.52
0.4
0.94
0.91
0.82
0.93
0.86
0.88
0.84
0.68
0.87
McAdj
0.86
0.82
0.86
0.82
0.86
0.82
0.86
0.82
0.86
0.82
0.92
0.96
0.92
0.94
0.94
0.93
0.92
0.85
0.93
McArrow
0.24
0.21
0.15
0.13
0.57
0.55
0.58
0.49
0.58
0.49
0.78
0.85
0.79
0.83
0.83
0.79
0.79
0.68
0.8
58
E
3.86
2.4
4.57
1.87
7.39
3.92
16.45
6.27
16.86
6.01
14.27
7.67
5.18
4.67
2.78
4.53
1.86
0.66
1.84
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
Alg
1
Vars
500
Deg
6
N
1000
0.98
0.99
1
1
1
0.99
1
0.96
0.99
1
0.99
1
0.99
1
0.99
1
0.98
0.99
0.99
1
0.8
0.73
0.65
0.65
0.66
0.74
0.67
0.79
0.74
0.67
0.74
0.67
0.74
0.67
0.74
0.67
0.97
0.89
0.75
0.68
0.96
*
*
0.82
0.82
*
*
0.89
0.57
0.57
0.57
0.57
0.99
0.99
0.89
0.93
0.97
0.98
0.6
0.6
AP AR AHP
0.99 0.85 0.69
59
0.75
0.49
0.5
0.73
0.43
0.38
0.43
0.38
0.52
0.4
0.65
0.59
0.92
0.84
0.45
0.4
AHR
0.57
0.88
0.85
0.81
0.8
0.81
0.86
0.81
0.87
0.86
0.82
0.86
0.82
0.86
0.82
0.86
0.82
0.97
0.94
0.86
0.82
McAdj
0.91
0.73
0.42
0.43
0.65
0.12
0.1
0.12
0.1
0.58
0.49
0.59
0.59
0.89
0.83
0.16
0.14
McArrow
0.32
0.8
25.28
17.51
81.96
60.56
9.37
4.81
61.2
19.66
13.2
21.81
11.24
160.61
93.55
224.5
84.43
41.8
33.51
53.04
45.74
E
11.13
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
500
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1
0.99
1
0.99
1
0.99
1
0.99
1
0.86
0.94
0.97
0.93
0.97
0.91
0.95
0.97
0.95
0.97
0.99
1
0.1
0.99
1
0.97
-
0.8
0.84
0.79
0.84
0.79
0.85
0.8
0.84
0.79
0.99
0.99
0.95
0.99
0.97
0.98
0.96
0.87
0.96
0.9
0.83
0.78
0.08
0.84
0.79
0.87
60
0.68
0.64
0.62
0.9
0.91
0.99
0.99
0.99
0.99
0.84
0.92
0.95
0.91
0.95
0.89
0.93
0.95
0.92
0.95
*
*
0.09
*
*
0.92
-
0.53
0.53
0.48
0.74
0.7
0.68
0.58
0.67
0.58
0.95
0.94
0.9
0.94
0.92
0.93
0.91
0.82
0.91
0.85
0.06
0.82
-
0.89
0.91
0.89
0.91
0.89
0.91
0.89
0.91
0.89
0.93
0.96
0.96
0.96
0.97
0.94
0.95
0.92
0.95
0.93
0.91
0.88
0.09
0.91
0.89
0.92
-
0.29
0.23
0.2
0.67
0.65
0.7
0.63
0.7
0.63
0.79
0.86
0.86
0.86
0.87
0.83
0.84
0.78
0.84
0.81
0.05
0.76
-
5.79
15.7
6.37
31.96
17.17
81.91
30.85
82.92
29.51
17.82
11.68
9.05
4.95
3.49
6.64
4.46
2.13
2.85
1.39
71.41
44.48
-9.90E-03
24.8
13.24
161.08
-
40
41
42
43
44
45
46
47
48
49
50
51
52
53
500
500
500
500
500
500
500
500
500
500
500
500
500
500
6
6
6
6
6
6
6
6
6
6
6
6
6
6
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
1000
0.99
1
0.99
1
0.99
1
0.99
1
0.97
0.98
0.99
1
0.84
0.79
0.84
0.79
0.84
0.79
0.84
0.79
0.99
0.96
0.85
0.79
0.58
0.57
0.58
0.57
0.99
0.99
0.88
0.91
0.96
0.97
0.6
0.59
0.49
0.45
0.49
0.45
0.68
0.58
0.73
0.7
0.94
0.91
0.51
0.47
0.91
0.89
0.91
0.89
0.91
0.89
0.91
0.89
0.98
0.97
0.92
0.89
0.14
0.11
0.14
0.11
0.7
0.63
0.64
0.65
0.9
0.88
0.18
0.15
48.41
26.29
37.9
21.96
414.81
288.21
483.89
256.16
43.1
39.69
74.29
64.07
References
References
[1] Constantin F Aliferis, Alexander Statnikov, Ioannis Tsamardinos, Subramani Mani, and Xenofon D Koutsoukos. Local causal and markov
blanket induction for causal discovery and feature selection for classification part i: Algorithms and empirical evaluation. Journal of Machine
Learning Research, 11(Jan):171–234, 2010.
[2] David Maxwell Chickering. Optimal structure identification with greedy
search. Journal of machine learning research, 3(Nov):507–554, 2002.
[3] Diego Colombo and Marloes H Maathuis. Order-independent constraintbased causal structure learning. The Journal of Machine Learning Research, 15(1):3741–3782, 2014.
[4] Markus Kalisch, Martin Mächler, Diego Colombo, Marloes H Maathuis,
Peter Bühlmann, et al. Causal inference using graphical models with
the r package pcalg. Journal of Statistical Software, 47(11):1–26, 2012.
61
[5] Christopher Meek. Causal inference and causal explanation with background knowledge. In Proceedings of the Eleventh conference on Uncertainty in artificial intelligence, pages 403–410. Morgan Kaufmann
Publishers Inc., 1995.
[6] Kevin Murphy et al. The bayes net toolbox for matlab. Computing
science and statistics, 33(2):1024–1034, 2001.
[7] Preetam Nandy, Alain Hauser, and Marloes H Maathuis. Highdimensional consistency in score-based and hybrid structure learning.
arXiv preprint arXiv:1507.02608, 2015.
[8] Joseph Ramsey, Madelyn Glymour, Ruben Sanchez-Romero, and Clark
Glymour. A million variables and more: the fast greedy equivalence
search algorithm for learning high-dimensional graphical causal models,
with an application to functional magnetic resonance images. International journal of data science and analytics, 3(2):121–129, 2017.
[9] Joseph Ramsey, Jiji Zhang, and Peter L Spirtes. Adjacency-faithfulness
and conservative causal inference. arXiv preprint arXiv:1206.6843, 2012.
[10] Joseph D Ramsey and Daniel Malinsky. Comparing the performance
of graphical structure learning algorithms with tetrad. arXiv preprint
arXiv:1607.08110, 2016.
[11] Peter Spirtes, Clark Glymour, Richard Scheines, and Prediction Causation. Search. 1993. Lecture Notes in Statistics, 1993.
[12] Ben Tilly, Matthias Trautner Kromann, and Yuji SODE.
Statistics-distributions-js ymdf.
https://github.com/YujiSODE/
statistics-distributions-js_Ymdf, 2017.
[13] Ioannis Tsamardinos, Constantin F Aliferis, and Alexander Statnikov.
Time and sample efficient discovery of markov blankets and direct causal
relations. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 673–678. ACM,
2003.
[14] Ioannis Tsamardinos, Constantin F Aliferis, Alexander R Statnikov, and
Er Statnikov. Algorithms for large scale markov blanket discovery. In
FLAIRS conference, volume 2, pages 376–380, 2003.
62
[15] Ioannis Tsamardinos, Laura E Brown, and Constantin F Aliferis. The
max-min hill-climbing bayesian network structure learning algorithm.
Machine learning, 65(1):31–78, 2006.
63
| 2cs.AI
|
Static Contract Simplification
Technical Report
Matthias Keil
Peter Thiemann
arXiv:1703.10331v1 [cs.PL] 30 Mar 2017
University of Freiburg,
Freiburg, Germany
{keilr,thiemann}@informatik.uni-freiburg.de
Abstract
Existing contract systems for dynamically typed languages
(e.g., Racket’s contract framework (Flatt et al. 2014, Chapter 7),
Disney’s JavaScript contract system contracts.js (Disney 2013), or
TreatJS for JavaScript (Keil and Thiemann 2015b)) each suffer
from a considerable performance impact when extending a program with contracts.
In contrast, static contract checking (Xu et al. 2009) avoids any
run-time cost if it succeeds in proving and removing a contract at
compile time. Otherwise, it defers to plain contract checking. A
subsequent extension to hybrid contract checking (Xu 2012) is able
to exploit partial proofs and to extract a residual contract that still
needs to be checked at run time. Both of these works are phrased
as transformations in a typed intermediate language.
A similar motivation is behind static contract verification
(Nguyen et al. 2014). This work takes place in the context of an
untyped language. It applies symbolic execution as well as techniques from occurrence typing to prove contracts statically and
thus eliminate their run-time overhead.
Our work is complementary to these previous efforts. Our main
objective is to improve the efficiency of contract monitoring by using static techniques to evaluate as much of a contract as possible, to detect and remove redundant checking of contracts, and to
reassemble the fragments that cannot be discharged statically in
places where they are least likely to affect performance. Typically,
we move contract fragments to a surrounding module boundary and
compose them to a new contract.
Thus, our goal is not to detect contract violations at compile
time, but rather to optimize contracts based on static information.
We also do not rewrite the non-contract code of the underlying
program, but perform program transformations restricted to just
moving around contracts.
To demonstrate the essence of our approach, consider the following code snippet that contains a function with an intersection
contract applied to it.
Contracts and contract monitoring are a powerful mechanism for
specifying properties and guaranteeing them at run time. However,
run time monitoring of contracts imposes a significant overhead.
The execution time is impacted by the insertion of contract checks
as well as by the introduction of proxy objects that perform delayed
contract checks on demand.
Static contract simplification attacks this issue using program
transformation. It applies compile-time transformations to programs with contracts to reduce the overall run time while preserving the original behavior. Our key technique is to statically
propagate contracts through the program and to evaluate and merge
contracts where possible. The goal is to obtain residual contracts
that are collectively cheaper to check at run time.
We distinguish different levels of preservation of behavior,
which impose different limitations on the admissible transformations: Strong blame preservation, where the transformation is a
behavioral equivalence, and weak blame preservation, where the
transformed program is equivalent up to the particular violation
reported. Our transformations never increase the overall number of
contract checks.
Categories and Subject Descriptors D.3.1 [Formal Definitions
and Theory]: Semantics; D.2.4 [Software/Program Verification]:
Programming by contract
Keywords Higher-Order Contracts, Contract Simplification, Hybrid Contract Checking
1. Introduction
Software contracts (Meyer 1988) stipulate invariants for objects as
well as pre- and postconditions for functions a programmer regards
as essential for the correct execution of a program. Traditionally,
contracts are checked at run time by contract monitoring. This
approach has become a prominent mechanism to provide strong
guarantees for programs in dynamically typed languages while
preserving their flexibility and expressiveness.
Using formal and precise specifications in the form of contracts in this way sounds appealing, but it comes with a cost:
Contract monitoring degrades the execution time significantly
(Takikawa et al. 2016). This cost arises because every contract extends a program with additional code to check the contract assertions that needs to be executed on top of the program. Moreover,
human developers may inadvertently add contracts to frequently
used functions and objects so that contracts end up on hot paths in
a program. In particular, if contracts are naively applied to recursive functions, predicates will check the same values repeatedly,
which aggravates the performance degradation.
1
2
3
4
5
let addOne =
((λ plus (λ z ((plus 1) z)))
[(λ x (λ y (+ x y)))
@ ((Number? → (Number? → Number?)) ∩
(String? → (String? → String?)))])
The value addOne is a function that takes an argument z and returns
the result of applying function plus to the number 1 and to z. We
use the inline operator @ to indicate contract assertions that we
put in square brackets to improve readability. We further rely on
predefined flat contracts Number? and String? that are type checks
for numbers and strings, respectively.
If we assume that the + operator is overloaded as in JavaScript
or Java so that it works for number and string values (but possibly
fails or delivers meaningless results for other inputs), a programmer
We created an executable implementation of the semantics using
PLT Redex (Felleisen et al. 2009). The implementation is available
online at http://anonymous.
may add an intersection contract to plus. This contract has the effect
that the context of a use of plus (here function addOne) can choose
to use plus either with number or string arguments. Monitoring this
contract requires six predicate checks at every use of addOne.1
However, as the second argument of plus is already a number,
the right side of the intersection can never be fulfilled. So, it would
be sufficient to check if the first argument of plus and its return
value satisfy Number?. Moreover, the remaining predicates can
be lifted to an interface description on addOne, as the following
example demonstrates. Only two predicate checks remain.
6
7
8
Overview Section 2 provides a series of examples demonstrating
the essence of our contract simplification. Section 3 recalls contracts and contract satisfaction from the literature and introduces
an untyped, applied, call-by-value lambda calculus with contract
monitoring. Section 4 introduces Baseline Simplification, the first
tier in our simplification stack. Section 5 shows the second tier,
Subset Simplification, which also handles alternatives and reduces
contracts that are subsumed by other contracts. Section 6 states our
blame theorems and Section 7 briefly addresses run time improvements. Section 8 discusses related work and Section 9 concludes.
The appendix extends this paper with auxiliary functions and
definitions, further examples, and proofs of all theorems. In particular, it provides the semantics to enable picky-like evaluation semantics and the semantics of a third, approximating, simplification
level that provides the stringes simplification.
let addOne =
[((λ plus (λ z ((plus 1) z)))
((λ x (λ y (+ x y))))) @ (Number? → Number?)]
This simplification can be done without knowing whether addOne
is ever executed: we are not losing any blame due to the transformation. However, it would not be ok to lift argument contracts across
function boundaries. Such lifting would introduce checks that may
never happen in the original program.
Building on the formalization of Keil and Thiemann (2015a),
we present a transformation system for static contract simplification. Our system provides two stages of simplification with different consequences: The Baseline Optimization and the Subset Optimization.
The Baseline Optimization unfolds contracts and evaluates
predicates where possible while preserving the blame behavior
of a program. The Subset Optimization reorganizes contracts and
forms new contracts from the contract fragments, but it may reorder contract computations and thus change the blame (i.e., the
particular violation that is flagged at run time). Our simplification
rules respect three overall guidelines:
2. Contract Simplification
This section explains the main ideas of our transformation through
a series of examples. We start with a simple example and work up
to more complex ones.
2.1 Unrolling Delayed Contracts
The examples consider different contracts on the addOne example
from the previous section. In a first step we add a simple function
contract to plus.
1
2
Improvement Each transformation step improves the efficiency
of contract monitoring: it either reduces or maintains the total
number of predicate checks at run time. Simplification never
increases the number of run-time checks.
3
4
While Number? is a flat contract that is checked immediately,
function contracts are delayed contracts. A delayed contract cannot
be checked directly when asserted to a value. Delayed contracts
stay attached to a value and are checked when the value is used.
Thus, every call to addOne triggers three predicate checks.
Instead of asserting a contract that forces a number of delayed
contract checks, our Baseline Simplification unrolls a delayed contract to all uses of that value. This step pushes the delayed contract
inwards to all occurrences of plus in addOne.
Strong Blame-Preservation Each transformation step of the baseline simplification fully preserves the blame-behavior of a program. It maintains the oder of contract checks at run time and
thus the order in which violations are reported. A transformed
program produces the same outcome as the original program.
Weak Blame-Preservation Transformation steps of the subset
simplification may change the oder of predicate checks, so that
they may also change the order of observed violations. Thus,
a transformed program ends in a blame state if and only if
the original program would end in a blame state. However, the
transformed program may report a different violation.
5
6
7
8
An optional third level optimizes contracts across module boundaries. Even though it provides the strongest optimization, it may
over-approximate contract violations and it would violate all three
guideline from above.
• We present the semantics of a two-tier static contract simplifi-
• We show how to split alternatives into separated observation to
deal with intersection and union contracts.
• We give a specification of subcontracting that is used to simplify
contracts based on the knowledge of other asserted contracts.
1 Assuming the semantics of intersection and union contracts from the
literature (Keil and Thiemann 2015a).
let addOne =
((λ plus
(λ z ([plus @ (Number? → (Number? → Number?))] 1) z)))
(λ x (λ y (+ x y))))
This step performs no simplification, but it yields an equivalent
program that is better prepared for unfolding the function contract
to its domain and range because the contract is on an expression
in application position. The following snippet demonstrates the
(complete) unfolding of the contract, which combines a number
of primitive unfolding steps.
Contributions This work makes the following contributions:
cation that reorganizes and pre-evaluates contracts at compile
time while guaranteeing weak or strong blame preservation.
let addOne =
((λ plus (λ z ((plus 1) z)))
[(λ x (λ y (+ x y)))
@ (Number? → (Number? → Number?))])
9
10
11
12
let addOne =
((λ plus
(λ z [((plus [1 @ Number?]) [z @ Number?]) @ Number?]))
(λ x (λ y (+ x y))))
In the resulting expression, the function contracts are completely
decomposed into flat contract components which are spread all over
the expression.
Observe that, after unfolding, there are several flat contracts applied to values. Such contracts can be checked statically without
changing the outcome. For example, 1 satisfies Number?. All satisfied contracts can safely be removed from the program.
13
14
15
16
let addOne =
((λ plus
(λ z [((plus 1) [z @ Number?]) @ Number?]))
(λ x (λ y (+ x y))))
In a last step, we push the remaining contract fragments outwards
where possible. This transformation creates a new function contract
from that contract. The special contract ⊤ accepts every value.
17
18
19
20
report it at run time. The next snippet shows the result after processing satisfied and failing contracts and pushing the remaining
fragments outwards.
32
33
34
35
36
37
Afterwards, the outermost expression carries two function contracts
that check the return value of addOne; two contracts are left in the
body of addOne to check the argument.
let addOne =
[((λ plus
(λ z ((plus 1) [z @ Number?])))
(λ x (λ y (+ x y)))) @ (⊤ → Number?)]
2.3 Splitting Alternatives in Separated Observations
When remembering the last state in Section 2.2, we still check
String? even though we know that right side of the intersection
is never fulfilled. This is because the Baseline Simplification preserves the blame behaviour of a program and is thus not able to deal
with alternatives. However, the next Subset Simplification abandons
the strong blame preservation to further optimize contracts.
Instead of unfolding intersection and union contracts, we now
observe both alternatives in separation. After transforming both
alternatives, we join the remaining contracts or even discard a
whole branch if it definitely leads to a violation.
In our example, we see that 1 @ String? leads to a context
failure of the right operand of the intersection contract. Thus, it
is useless to keep the other fragments of the right operand which
definitely yields a context violation. The following code snippet
shows the result.
Afterwards, the outermost expression carries a function contract
that checks the return value of addOne and one contract is left in the
body of addOne to check the argument. Again, this step does not
simplify contracts, but it prepares for further simplification steps.
However, further simplification cannot be done while guaranteeing strong blame preservation. The Baseline Simplification is
not allowed to behave differently than the dynamics would do at
run time.
2.2 Treating Intersection and Union
Our next example considers the full addOne example from the
introduction, which we repeat for convenience.
21
22
23
24
25
let addOne =
((λ plus (λ z ((plus 1) z)))
[(λ x (λ y (+ x y)))
@ ((Number? → (Number? → Number?)) ∩
(String? → (String? → String?)))])
38
39
40
Intersections of function contracts are delayed contracts, they must
stay attached to a value until the value is used in an application.
Moreover, for an intersection, the context can choose to fulfill the
left or the right side of the intersection. Thus, every call to addOne
triggers six predicate checks.
Again, Baseline Simplification unrolls the intersection contract
to all uses of that value and unfolds the contract to its domain
and range when used in an application. The following snippet
demonstrates the unfolding of the intersection contract.
26
27
28
29
30
31
let addOne =
((λ plus
(λ z [[((plus
[[1 @2 String?] @1 Number?])
[[z @2 String?] @1 Number?]) @2 String?] @1 Number
?]))
(λ x (λ y (+ x y))))
In addition, the static contract monitor maintains additional information in the background (e.g., which contracts belong to the same
side of the intersection) to connect each contract with the enclosing
operation. In the example we indicate the originating operand of
the intersection by @1 and @2 .
After unfolding, the Baseline Simplification checks flat contracts applied to values. Here, 1 satisfies Number? but violates
String?. While Number? can now be removed, information about
a failing contract must remain in the source program. In general, it
is not possible to report failure statically because we do not know
if the code containing the failed contract (in this case, function
addOne) is ever executed.
However, we do not have to leave the original contract in the
expression. Instead, the transformation replaces it by a “false contract” ⊥, whose sole duty is to remember a contract violation and
let addOne =
[[((λ plus
(λ z ((plus
[1 @2 ⊥])
[[z @2 String?] @1 Number?])))
(λ x (λ y (+ x y)))) @1 (⊤ → Number?)] @2 (⊤ → String?)]
41
let addOne =
[((λ plus
(λ z ((plus [1 @2 ⊥]) [z @1 Number?])))
(λ x (λ y (+ x y)))) @1 (⊤ → Number?)]
It remains to check z and addOne’s return to be a number. This are
two remaining checks per use of addOne.
However, we still need to keep the information on the failing
alternative. This is required to throw an exception if also the other
alternative fails.
2.4 Lifting Contracts
As shown in the example above, contracts on the function’s body
are reassembled to a new function contract, whereas contracts on
function arguments must remain to preserve the order of predicate
checks. Continuing the example, our next step constructs a new
contract from the remaining contracts on arguments.
42
43
44
45
let addOne =
[[((λ plus
(λ z ((plus [1 @2 ⊥]) z)))
(λ x (λ y (+ x y)))) @1 (Number? → ⊤)] @1 (⊤ → Number?)]
The addOne function now contains another function contract that
restricts its domain to number values. As this step puts contracts
on arguments in front, it may change the order in which violations
arise. This lifting to module boundaries only reorganizes existing
contracts, but it helps to reduce contracts that are subsumed by
other contracts.
2.5 Condensing Contracts
Continuing the example in Section 2.4, the outermost expression
carries two function contracts that check the argument and the
return value of addOne. Moreover, both contracts arise from the
same source contract. These function contracts may be condensed
to a single function contract on addOne.
46
47
48
49
let addOne =
[((λ plus
(λ z ((plus [1 @2 ⊥]) z)))
(λ x (λ y (+ x y)))) @1 (Number? → Number?)]
L, M, N
K
U, V, W
E
This step is only possible if the argument contract, here Number
? → ⊤, is in negative position and if the return contract, ⊤ →
Number? is in positive position.
B ETA
OP
T RUE
FALSE
2.6 Contract Subsets
Splitting intersections and unions into separate observations has
another important advantage: We know that in every branch every contract must be fulfilled. Thus we can relate contracts to other
contracts and use knowledge about other contracts to reduce redundant checks. To demonstrate an example, we add a second contract
to our running example.
50
51
52
53
54
56
let addOne =
((λ plus [(λ z ((plus 1) z)) @ (Positive? → Positive?)])
[(λ x (λ y (+ x y)))
@ ((Number? → (Number? → Number?)) ∩
(String? → (String? → String?)))])
57
58
59
60
61
62
63
64
let addOne =
[[[((λ plus (λ z ((plus 1) z)))
(λ x (λ y (+ x y)))) @1 (Number? → ⊤)] @1 (⊤ → Number?)]
@ (Positive? → Positive?)]
k
let addOne =
[((λ plus (λ z ((plus [1 @2 ⊥]) z)))
(λ x (λ y (+ x y)))) @ (Positive? → Positive?)]
~ | if L M N
K | x | λx.M | M N | op(M)
false | true | 0 | 1 | . . .
K | λx.M
~ EM
~ ) | if E M N
| E N | V E | op(V
E Jλx.M V K
~ )K
E Jop(V
E Jif true M N K
E Jif false M N K
−→
−→
−→
−→
E JM [x 7→ V ]K
~ )K
E Jδop(V
E JM K
E JN K
~ ∈ dom(δop )
V
Figure 1. Syntax and semantics of λJ .
Here, Positive? is a flat contract that checks for positive numbers.
In addition to the contract on plus, addOne contains another function contract, requiring addOne to be called with positive numbers
and to return a positive number. After performing the same transformation steps as before, we obtain the following expressions.
55
::=
::=
::=
::=
b
κ
ς
C, D
I, J
Q, R
A
::=
::=
::=
::=
I |Q|C ∪D |I ∩C
flat(M )
C → D | x →A | Q∩R
Λx.C
B
L, M, N
U, V, W
E
::=
+=
+=
+=
+blame | -blame
M @♭ C || V @♭ι (eval M ) | B♭
|| V @♭ι Q
E @♭ C || V @♭ι (eval E )
::=
::=
::=
♭ || ι
b ◭W |b◭ ι|b◭ ι→ι |b◭ι∩ι|b◭ ι∪ι
·|κ:ς
Figure 2. Syntax extension for λCon .
65
(λ x (λ y (+ x y)))) @ (Positive? → Positive?)]
3. Contracts and Contract Satisfaction
The intersection is split into two parallel observations, as indicated by k. All contracts were pushed to the outermost boundary
and form a new module boundary. However, some of the contract
checks are redundant.
For example, the innermost contract in the first observation
requires addOne to be called with numbers, whereas the outermost
contract restricts the argument already to positive numbers. As
positive numbers are a proper subset of numbers, the inner contract
will never raise a violation if the outer contract is satisfied. Thus, it
can be removed without changing the blame behavior of a program.
The contract Positive? → Positive? is more restrictive than
any other contract on addOne, that is, Number? → ⊤ and ⊤ →
Number?. Here, a contract A is more restrictive than a contract B,
if the satisfaction of A implies the satisfaction of B.
However, if a more restrictive contract fails, that failure does
not imply failure of the less restrictive contract. But as every contract must be fulfilled, the program already stops with a contract
violation. Thus, there is no need to check a less restrictive contract
on that branch. Unfortunately, removing the less restrictive contract
might change the order of observed violations, as the less restrictive
contract might report its violation first.
To sum up, in the first observation the outer contract remains as
it subsumes any other contract, whereas in the second observation
both contracts remain.
Finally, after finishing all transformation steps, the simplifications joins the observation to final expression.
This section defines λCon , an untyped call-by-value lambda calculus with contracts which serves as a core calculus for contract monitoring. It first introduces the base calculus λJ , then it proceeds to
describe contracts and their semantics for the base calculus. Finally
it gives the semantics of contract assertion and blame propagation.
The core calculus is inspired by previous work from the literature (Keil and Thiemann 2015a).
let addOne =
[((λ plus
(λ z ((plus [1 @2 ⊥]) z)))
Figure 2 defines the syntax of λCon as an extension of λJ . It first
introduces constructs for contract monitoring in general and, second, it adds new terms specific to contract monitoring. Intermediate
3.1 The Base Language λJ
Figure 1 defines syntax and semantics of λJ , an applied call-byvalue lambda calculus. A term M is either a first-order constant,
a variable, a lambda abstraction, an application, a primitive operation, or a condition. Variables x, y, and z are drawn from a denumerable set of variables. Constants K range over a set of base type
values including booleans, numbers, and strings.
To define evaluation, values U , V , and W range over values and
E over evaluation contexts, which are defined as usual.
The small-step reduction relation −→ comprises beta-value reduction, built-in partial operations transforming a vector of values
into a value, and the reduction of conditionals. We write −→∗ for
its reflexive, transitive closure and 6−→∗ for its complement. That
is, M 6−→∗ N if, for all L such that M −→∗ L, it holds that
L 6= N . We also write M 6−→ to indicate that there is no N such
that M −→ N .
3.2 Contracts and Contracted λJ
terms that do not occur in source programs appear after double bars
“||”.
A contract C is either an immediate contact I, a delayed contract Q, a union between two contract (C ∪ D), or an intersection
(I ∩ C).
Without loss of generality we restrict top-level intersection contracts to intersections between an immediate contract and a rest
contract (cf. (Keil and Thiemann 2015a)). This requires that unions
and immediate parts are distributivity pulled out of intersections of
delayed contracts such that the immediate parts can be check right
away when asserted, whereas the delayed parts remain.
Immediate contracts I and J stand for flat contracts define by
a predicate M that can be evaluated right away when asserted to a
value.
A delayed contracts Q is either a function contract C → D
with domain contract D and range contract D, a dependent contract
x → A defined by contract abstraction A, or the finite intersection
of function (dependent) contracts. Delayed contracts stay with the
value until the value is subject of an application.
Consequently, values are extended with V @♭ι Q which represents a value wrapped in a delayed contract that is to be monitored
when the value is used. The extended set of values forces us to revisit the built-in operations. We posit that each partial function δop
first erases all contract monitoring from its arguments, then processes the underlying λJ -values, and finally returns a λJ -value.
The only new source term is contract assertion M @b C. The
adornments ♭ and ι are drawn from an unspecified denumerable set
of blame identifiers b, which comprises blame labels ♭ that occur
in source terms and blame variables ι that are introduced during
evaluation.
In the intermediate term V @♭ι (eval M ), the term M represents
the current evaluation state of the predicate of a flat contract. The
B♭ expression signal a contract violation at label ♭. Term B is either
a +blame (positive) or a -blame (negative) blame, representing
subject or context blame, respectively.
Evaluation contexts are extended in the obvious way: a contract
monitor is only applied to a value and a flat contract is checked
before its value is used.
In λCon , contract monitoring occurs via constraints κ imposed
on blame identifiers. There is an indirection constraint and one kind
of constraint for each kind of contract: flat, function, intersection,
and union. Constraints are collected in a list ς during reduction.
3.3 Semantics for λCon
Figure 3 specifies the small-step reduction semantics of λCon as a
relation ς, M −→ ς ′ , N on pairs of a constraint list and a term.
Instead of raising blame exceptions, the rewriting rules for contract enforcement generate constraints in ς: a failing contract must
not raise blame immediately, because it may be nested in an intersection or a union. The sequence of elements in the constraint list
reflects the temporal order in which the constraints were generated
during reduction. The latest, youngest constraints are always on top
of the list. Appendix A explains the semantics of the constraints.
The rule A SSERT introduces a fresh name for each new instantiation of a monitor in the source program. It is needed for technical
reasons. Flat contracts get evaluated immediately. Rule F LAT starts
evaluating a flat contract by evaluating the predicate M applied to
the unwrapped subject value W . Meta function ∇(V ) (Figure 4)
first erases all contract monitors from the subject value V . After
predicate evaluation, rule U NIT picks up the result and stores it in a
constraint.
The second group of rules U NION and I NTERSECTION implements contract checking of top-level union and intersection contracts. Both rules install a new constraint that combine the contract
satisfaction of the subcontracts to the satisfaction of the current
contract.
The rules D-F UNCTION, D-D EPENDENT, and D-I NTERSECTION
define the behavior of a contracted value under function application. Rule D-F UNCTION handles a call to a value with a function
contract. Different from previous work, the blame computation is
handled indirectly by creating new blame variables for the domain
and range part; a new constraint is added that transforms the outcome of both portions according to the specification of the function
contract.
The D-D EPENDENT rule defines the evaluation of a dependent
contract. It substitutes the function’s arguments for every occurrence of x in the body of C and proceeds with the assertion of the
resulting contract.
Rule D-I NTERSECTION embodies the evaluation of an intersection contract under function application. The generated constraint
serves to combine the results of the subcontracts. Unlike the union
contract, an intersection constraint occurs at each use of the contracted value, which implements the choice of the context (cf.
(Keil and Thiemann 2015a)).
Finally, λJ reductions are lifted to λCon reductions using rule
BASE. Built-in operations and conditions “see through” contracts
(rules D-O P and D-I F) and proceed on λJ -values without contracts.
This also implies that all base type operations only return values
that does not contain contracts.
3.4 Introducing Blame
The dynamics in Figure 3 use constraints to create a structure for
computing positive and negative blame according to the semantics
of subject and context satisfaction, respectively. To this end, each
blame identifier b is associated with two truth values, b.subject and
b.context. Intuitively, if b.subject is false, then the contract associated with b is not subject-satisfied and may lead to positive blame
for b. If b.context is false, then there is a context that does not respect contract b and may lead to negative blame for b. Appendix A
specifies the constraint satisfaction for λCon .
Computing a blame assignment boils down to computing an interpretation for a constraint list ς. To determine whether a constraint
list ς is a blame state (i.e., whether it should signal a contract violation), we check whether the state ς maps any source-level blame
label ♭ to false.
Definition 1. ς is a blame state for blame label ♭ iff
∃♭.ς(♭.subject ∧ ♭.context) ⊒ false.
ς is a blame state if there exists a blame label ♭ such that ς is a
blame state for this label.
To model reduction with blame, we define a new reduction
relation ς, M =⇒ ς ′ , N on configurations. It behaves like −→
unless ς is a blame state. In a blame state, it stops signaling the
violation. There are no reductions with B.
∃♭.ς(♭.subject) ⊒ false
ς, M =⇒ ς, +blame
♭
∃♭.ς(♭.context) ⊒ false
ς, M =⇒ ς, -blame♭
ς, M −→ ς ′ , N
6 ∃♭.ς(♭.subject ∧ ♭.context) ⊒ false
ς, M =⇒ ς ′ , N
3.5 Lax, Picky, and Indy Semantics
Contract monitoring distinguishes three blame semantics: Lax,
Picky, and Indy, initially introduced to handle correctness and
completeness of higher-order dependent contracts (Dimoulas et al.
A SSERT
ς,
E JV @♭ CK
−→
(♭ ◭ ι) : ς,
E JV @♭ι CK
F LAT
U NIT
ς,
ς,
E JV @♭ι flat(M )K
E JV @♭ι (eval W )K
−→
−→
ς,
(ι ◭ W ) : ς,
E JV @♭ι (eval (M U ))K
E JV K
U NION
I NTERSECTION
ς,
ς,
E JV @♭ι (C ∪ D)K
E JV @♭ι (I ∩ C)K
−→
−→
(ι ◭ ι1 ∪ ι2 ) : ς,
(ι ◭ ι1 ∩ ι2 ) : ς,
E J(V @♭ι1 C)@♭ι2 DK
E J(V @♭ι1 I)@♭ι2 CK
D-F UNCTION
D-D EPENDENT
D-I NTERSECTION
ς,
ς,
ς,
E J(V @♭ι (C → D)) W K
E J(V @♭ι (x → Λx.C)) W K
E J(V @♭ι (Q ∩ R)) W K
−→
−→
−→
(ι ◭ ι1 → ι2 ) : ς,
ς,
(ι ◭ ι1 ∩ ι2 ) : ς,
E J(V (W @♭ι1 C))@♭ι2 DK
E J(V W )@♭ι (C[x 7→ U ])K
E J((V @♭ι1 Q)@♭ι2 R) W K
BASE
D-O P
M −→ N
ς, M −→ ς, N
~ (V
ς, E Jop(U,
U = K | λx.M
~ )K −→ ς, E Jop(U,
~ V, W
~ )K
W
@♭ι Q),
ι 6∈ ς
U = ∇(V )
ι1 , ι2 ∈
6 ς
ι1 , ι2 ∈
6 ς
ι1 , ι2 6∈ ς
U = ∇(W )
ι1 , ι2 6∈ ς
D-I F
ς, E Jif (V @♭ι Q) M N K −→ ς, E Jif V M N K
Figure 3. Operational semantics of λCon .
∇(V ) =
(
∇(W ) V = W @♭ι Q
V
otherwise
Figure 4. Unwrap delayed contracts.
2011). The different styles address the point that a contract might
violate another contract.
The Lax semantics erase all contract monitors on values before
applying a value to a predicate. This is correct but not complete because it guarantees that a well-behaved program gets never blamed
for a violation taking place in a predicate. However, it omits to report violations in predicates.
The Picky semantics, in contrast, is complete but not correct. As
it retains contract monitors on the values it might wrongly blame
the program for violating a contract, but it guarantees that every
violation is reported.
The Indy semantics, an extension of Picky, introduces a third
party (the contract itself) in addition to the exiting parties context
and subject. Passing a value to a predicate reorganizes the monitor
such that wrong uses of that value blame the ill-behaved contract.
Our optimization considers the Lax style, as it strongly guarantees correctness and completeness in respect to source program. As
reorganizing contracts changes the order of predicate checks optimizing is simpler when using Lax semantics.
Appendix D discusses our simplification with respect to Picky
like evaluation semantics.
4. Baseline Simplification
This section presents the Baseline Simplification, the first tier in our
simplification stack. The baseline simplification encompass (i) the
unfolding of contract assertions and the creation of constraints
where possible, (ii) the compile-time evaluation of flat contracts
(predicates) on values, and (iii) the unrolling of delayed contracts
to all uses of the contracted values. To this end, we introduce a set
of canonical (non-transformable) terms T as a subset of λCon terms
M and extend syntax and semantics of λCon in some respects.
All baseline steps do exactly the same as the dynamics would
later on do. Thus, baseline simplification guarantees Strong Blamepreservation and Improvement for all transformation steps.
4.1 Syntax and Semantics Extension
Figure 5 shows some syntax extension for λCon .
C, D
L, M, N
κ
+=
+=
+=
⊤|⊥
M @♭ι C
b ◭ ¬ι
Figure 5. Syntax extension of λCon .
Contracts now contain ⊤ (true) and ⊥ (false), shortcuts for
predicates evaluating to true or false, respectively. ⊤ and ⊥ arise
during optimization and represent remaining knowledge about the
given proposition. Whereas ⊤ can be reduced statically, ⊥ must
remain in the source program to ensure correct blame propagation.
The shortcuts are more efficient because none of them must
be checked at run time. When reaching a term with ⊤ or ⊥ the
outcome can immediately be transfered into a constraint and must
not be checked like a predicate.
Source terms M now contain partially evaluated contracts on
source terms M @♭ι C and a new inversion constraint b ◭ ¬ι extends the set of constrains κ. The inversion constraint flips context
and subject responsibility and corresponds to the negative side of a
function constraint. This constraint only arises during optimization.
Figure 6 specified the reduction of the shortcuts. Rule R EDUCE /T RUE
and R EDUCE /FALSE creates a constraint for ⊤ and ⊥. Three special
rules, D-D OMAIN, D-R ANGE, and D-FALSE simplify the evaluation
of function contracts with ⊤ and ⊥ on its domain or range portion.
If one portion is true only the other portion must be checked. All
this function contrast arise only during optimization.
4.2 Canonical Terms and Contexts
Canonical (non-transformable) terms are the output of our transformation. A canonical terms is either a contract-free term or a term
with a contract that cannot further transformed at this level.
In general, canonical terms distinguish terms without a contract
on the outermost level (indicated by letter S) from terms with a
non-transformable contract (indicated by letter T ). Canonical terms
are a subset of λCon terms M . Appendix B.1 shows its definition.
A transformation context F, G, and H (Figure 7) is defined as
usual as a term with a hole and a value context V is a term with an
indefinite number of remaining ⊥ contracts. This terms result only
during transformation.
R EDUCE /T RUE
R EDUCE /FALSE
ς,
ς,
E JV @♭ι ⊤K
E JV @♭ι ⊥K
−→
−→
(ι ◭ true) : ς,
(ι ◭ false) : ς,
E JV K
E JV K
D-D OMAIN
D-R ANGE
D-FALSE
ς,
ς,
ς,
E J(V @♭ι (C → ⊤)) W K
E J(V @♭ι (⊤ → D)) W K
E J(V @♭ι (⊤ → ⊥)) W K
−→
−→
−→
(ι ◭ ¬ι1 ) : ς,
(ι ◭ ι1 ) : ς,
(ι ◭ false) : ς,
E JV (W @♭ι1 C)K
E J(V W )@♭ι1 DK
E JV W K
ι1 ∈
6 ς
ι1 ∈
6 ς
Figure 6. Operational semantics extension of λCon .
F, G, H
::=
|
~)
| λx.F | F M | T F | op(T~ F M
if F M N | if T F N | if T T F | F@♭ι C
V
::=
| V@♭ι ⊥
Figure 7. Transformation contexts.
4.3 Baseline Transformation
The baseline transformation ς, M 7−→B ς ′ , N ′ is defined on pairs
of a constrains list and terms. As before, we write 7−→∗B for its
reflexive, transitive closure.
Figure 8 specifies the baseline transformation rules for λCon
terms M . The first rules, U NFOLD /A SSERT, U NFOLD /U NION, and
U NFOLD /I NTERSECTION, unfold contracts and create constraints
equivalent to the dynamics in Subsection 3.5.
Rule U NFOLD /O P unpacks a delayed contract on a term used in a
base type operation. As this contract is never checked, it is true by
definition (cf. (Keil and Thiemann 2015a)). Here, terms TI and TQ
are canonical terms with a possible immediate or delayed contract.
Rules U NFOLD /D-F UNCTION and U NFOLD /D-I NTERSECTION unfold a delayed contract that appears left in an application corresponding to the dynamics. However, a function contract unfolds the
domain portion to the term on the right, which is not necessarily a
value.
Rule U NROLL unrolls a delayed contract on a term T right in
an application. This step removes the contract from T and grafts it
to all uses of T . Even though this steps duplicates the number of
contract assertions, we do not have more contract checks.
Rule L OWER create a new function contract by lowering the
contract on the function’s body. The new function contract contains
⊤ on it’s domain portion as it accepts every argument.
Rules P USH /I MMEDIATE and P USH /FALSE reverse the order of a
delayed contract nested in an assertion with an immediate contract
or ⊥. This step is required to evaluate immediate contracts on
values and to unfold or unroll function’s contracts.
Rule P USH /I F pushes a contract out of a condition if the contracts
appears on both branches of the condition.
Rule C ONVERT /T RUE reduces all occurrences of ⊤ and produces
a constraint that maps true to the corresponding blame variable ι.
Knowledge about satisfied contracts can safely be removed from
the source program because they do not compromise the blame
assignment.2
In contrast to ⊤, ⊥ (the knowledge about a failing contract)
must remain in the source program and cannot be translated into
a constraint statically. Doing this would over-approximate contract failures as the constrain might lead to a blame state, without knowing if the ill-behaved sources is ever executed. However,
2 Due to the specification of contract satisfaction it is also possible to
remove ⊤ from the source program without creating a constraint. A nonexisting constraint is handled as true because every non-violated contract
assertion is satisfied, either because it is fulfilled or never verified.
over-approximating violations might be intended, as Appendix G
demonstrates.
Finally, rules V ERIFY /T RUE and V ERIFY /T RUE evaluate flat contracts on values that are available at compile time and transform the
outcome of the evaluation into ⊤ or ⊥. Values might be nested in a
number of ⊥ contracts of previously checked predicates.
Lemma 1 (Progress). For all terms M it holds that either M is a
canonical term T or there exists a transformation step ς, M 7−→B
ς ′, N .
4.4 Strong Blame-preservation
To sum up, the Baseline Transformation makes exactly the same
transformation steps at compile time, than the dynamics would
later on do. Thus, 7−→B preserves the evaluation behavior of our
program.
Conjecture 1 (Strong Blame-preservation). For all transforma′
tions ςM , M 7−→B ςN , N it hold that either ςM , M =⇒ ςM
,V
′′
′
, V or that ςM , M =⇒ ςM
, B♭ and ςN , N =⇒
and ςN , N =⇒ ςN
′′
ςN
, B♭ .
5. Subset Simplification
The Subset Simplification is the second tier in our transformation
stack. Its objective is to reuse knowledge of previously checked
contracts to avoid redundant checks whenever possible and to propagate contract violations to the surrounding module boundary.
The transformation first branches every alternative given by
intersection and union contracts into an individual observation on
which every contract must be fulfilled. Later it reduces contracts
that are less restrictive and subsumed by other contracts. Finally, it
joins remaining fragments to a new contract where possible.
To this end we introduce a subcontract relation which is closely
related do already existing definitions of naive subtyping. Subcontracting claims a contract to be more restrictive than another contract.
5.1 Predicates
To achieve the best possible result it requires to have contracts
and predicates as fine grained as possible. Intersection and union
contracts enable to build complex contracts from different subcontracts. Distinct properties can be written in distinct predicates
(flat contracts) and combined using intersection and union.
For example, a contract that checks for positive even numbers
can be written as the intersection of one contract that checks for
positive numbers and one contract that checks for even numbers:
flat(λx.(x >= 0)) ∩ flat(λx.(x%2 = 0))
However, there is a twist. Fine-grained sub-contracts enable contracts be optimized efficiently. But, at evaluation time, they must be
handled in separation such that we result in more predicate checks.
For example, consider the following contract that checks for
natural numbers:
flat(λx.(x >= 0))
U NFOLD /A SSERT
ς,
FJT @♭ CK
7−→B
(♭ ◭ ι) : ς,
FJT @♭ι CK
U NFOLD /U NION
U NFOLD /I NTERSECTION
U NFOLD /O P
ς,
ς,
ς,
FJT @♭ι (C ∪ D)K
FJT @♭ι (I ∩ C)K
~I , (T @♭ι Q), T~Q )K
FJop(T
7 →B
−
7−→B
7−→B
(ι ◭ ι1 ∪ ι2 ) : ς,
(ι ◭ ι1 ∩ ι2 ) : ς,
(ι ◭ true) : ς,
FJ(T @♭ι1 C)@♭ι2 DK
FJ(T @♭ι1 I)@♭ι2 CK
~I , T, T~Q )K
FJop(T
ι1 , ι2 ∈
6 ς
ι1 , ι2 6∈ ς
U NFOLD /D-F UNCTION
U NFOLD /D-I NTERSECTION
ς,
ς,
FJ(T1 @♭ι (C → D)) T2 K
FJ(T1 @♭ι (C ∩ D)) T2 K
7 →B
−
7−→B
(ι ◭ ι1 → ι2 ) : ς,
(ι ◭ ι1 ∩ ι2 ) : ς,
FJ(T1 (T2 @♭ι1 C))@♭ι2 DK
FJ((T1 @♭ι1 C)@♭ι2 D) T2 )K
ι1 , ι2 ∈
6 ς
ι1 , ι2 6∈ ς
U NROLL
L OWER
ς,
ς,
FJVJλx.SK (T @♭ι Q)K
FJλx.(T @♭ι C)K
7−→B
7−→B
ς
ς
FJVJλx.S[x 7→ (x@♭ι Q)]K T K
FJ(λx.T )@♭ι (⊤ → C)K
P USH /I MMEDIATE
P USH /FALSE
P USH /I F
ς,
ς,
ς,
FJ(T @♭ι11 Q)@♭ι22 IK
FJ(T @♭ι11 Q)@♭ι22 ⊥K
FJ(if T0 (T1 @♭ι C) (T2 @♭ι C))K
7−→B
7−→B
7−→B
ς
ς
ς
FJ(T @♭ι22 I)@♭ι11 QK
FJ(T @♭ι22 ⊥)@♭ι11 QK
FJ(if T0 T1 T2 )@♭ι CK
C ONVERT /T RUE
ς,
FJT @♭ι ⊤K
7−→B
(ι ◭ true) : ς,
FJT K
V ERIFY /T RUE
·, M V −→ ς, W
ς, FJVJSVal K@♭ι flat(M )K
ι 6∈ ς
V ERIFY /FALSE
τ (W ) = true
7 →B ς, FJVJSVal K@♭ι ⊤K
−
·, M V −→ ς, W
ς, FJVJSVal K@♭ι flat(M )K
τ (W ) = false
7−→B ς, FJVJSVal K@♭ι ⊥K
Figure 8. Baseline transformation rules.
Obviously, one can write a similar contract like this:
flat(λx.(x > 0)) ∪ flat(λx.(x = 0))
The presence of intersection and union contracts enables developers to write fine grained contracts, which, in turn, enable us to identify redundant parts and to reduce already checked properties.
To avoid a run time impact cause by fine-grained predicates,
we can easily conjunct or disjunct predicates of the same blame
label after optimization. This is possible because the intersection
and union of flat contracts corresponds to the conjunction and
disjunction of predicates (cf. (Keil and Thiemann 2015a)).
5.2 Contract Subsets
When optimizing contracts, we use subcontracting to characterize when a contract is more restrictive than another contract,
i.e. it subsumes all of its obligations. The definition of subcontracting is closely related to naive subtyping from the literature
(Wadler and Findler 2009).
We write C ⊑∗ D if C is more restrictive than D, i.e.
E JM @♭ CK −→∗ V implies that E JM @♭ DK −→∗ V and
E JM @♭ DK −→∗ B♭ implies that E JM @♭ CK −→∗ B♭ .
As a contract specifies the interface between a subject and
its enclosing context, subcontracting must be covariant for both
parties: context and subject.
For example, let Positive? = flat(λx.x>0) and Natural? = flat(λ
x.x≥0). Than Positive? ⊑∗ Natural? as for all x it holds that x>0
implies x≥0.
For another example, Positive? → Positive? is a subcontract of
Natural? → Natural? as it is more restrictive on both portions.
Furthermore, Positive? → Natural? and Natural? → Positive? are
further subcontracts of Positive? → Positive?.
In order to specify our subcontracting judgement, we factor subcontracting into two subsidiary relations: subject subcontracting,
written C ⊑+ D, and context subcontracting, written C ⊑− D.
Relation C ⊑+ D indicates that the subject portion in C is more
restrictive than the subject portion in D, whereas C ⊑− D indicates that C restricts the context more than D.
In contrast to Wadler and Findler’s definition of positive and
negative subtyping (cf. (Wadler and Findler 2009)), which is related to ordinary subtyping, our definition of positive and negative
subcontracting is reversal because its is related to the naive subcontracting. We write C ⊑− D if C is more restrictive than D,
whereas Wadler and Findler write D ⊑− C.
Figure 9 shows both relations. The two judgement are defined
in terms of each other and track swapping of responsibilities on
function arguments. Both judgements are reflexive and transitive.
We further write M ≤ N if predicate M implies predicate N .
For example, (x>0) ≤ (x≥0) and (x=0) ≤ (x≥0).
However, our subcontract judgement stops at M ≤ N . Relation
≤ refers to an environment Γ that resolves relations on terms. Γ
might be given by a SAT solver or a set of predefined relations.
(M ≤ N ) ∈ Γ
M ≤N
It remains to defined subcontracting in terms of context and
subject subcontracting.
Definition 2. C is a naive subcontract of D, written C ⊑∗ D, iff
C ⊑− D
∧
C ⊑+ D.
We further define an ordinary subcontracting judgement, C ⊑
D, which corresponds to the usuals definition of subtyping. Ordinary subcontracting is contravariant for the context portion and
covariant for the subject portion.
Definition 3. C is an ordinary subcontract of D, written C ⊑ D,
iff
D ⊑− C ∧ C ⊑+ D.
In addition to the already mentiont notation, we sometimes use
⊏, ⊏∗ , ⊏+ , and ⊏− to indicate a proper subcontract, analogous to
already existing definitions on inequality.
Lemma 2. Relation ⊑∗ is a preorder, i.e. it is reflexive and transitive. For all C, C ′ , and D, we have that:
• C ⊑∗ C
Context Subcontract
+
C ⊑− C
CD ⊑ DD
I ⊑− J
−
CR ⊑ DR
CD → CR ⊑− DD → DR
CL ⊑− D
CD ⊑− DD
x → Λx.C ⊑− x → Λx.D
CR ⊑− D
CL ∩ CR ⊑ − D
C ⊑− DL
C ⊑− DR
CL ⊑− D
C ⊑− DL ∩ DR
C ⊑− DL ∩ DR
CL ∪ CR ⊑− D
CR ⊑− D
C ⊑− DL
CL ∪ CR ⊑− D
C ⊑− DR
C ⊑− DL ∪ DR
Subject Subcontract
C ⊑+ C
C ⊑+ ⊤
M ≤N
⊥ ⊑+ D
flat(M ) ⊑+ flat(N )
CR ⊏+ DR
CD ⊑− DD
+
CD → CR ⊑ DD → DR
CD ⊑+ DD
x → Λx.C ⊑+ x → Λx.D
CL ⊑+ D
CR ⊑+ D
CL ∩ CR ⊑+ D
CL ∩ CR ⊑ + D
C ⊑+ DL
C ⊑+ DR
C ⊑+ DL ∩ DR
CL ⊑+ D
CR ⊑+ D
CL ∪ CR ⊑+ D
C ⊑+ DL
C ⊑+ DR
C ⊑+ DL ∪ DR
C ⊑+ DL ∪ DR
Figure 9. Context and subject subcontracting.
T
A
B
::=
::=
::=
| (T k M ) | (T k T )
| A@♭ι C
~ ) | B@♭ι C | if B M N
| B M | T B | op(T~ B M
Figure 10. Parallel observations and contexts.
• if C ⊑∗ C ′ and C ′ ⊑∗ D then C ⊑∗ D
5.3 Canonical Terms, revisited
As for the Baseline Transformation, we first define a set of canonical terms (non-transformable) terms that serve as the output of this
level’s transformation. The new canonical terms are a subset of the
already defined terms in Section 4.2. Appendix B.2 shows its definition.
In addition, Figure 10 shows the definition of a new transformation contexts T that looks through parallel observation. An assertion context A is a number of contract assertions and a body
context B is a transformation context without lambda abstractions
and without conditions with a hole in a branch.
5.4 Subset Transformation
Figure 11 defines the subset transformation as an extension of the
baseline transformation in Figure 8.
A significant change is that the new rules F ORK /U NION and
F ORK /I NTERSECTION replace the already exiting rules with the name
U NFOLD /U NION and U NFOLD /D-I NTERSECTION. Both rules split the
current observation and handle both sides of an intersection or
union in separation. This step eliminates alternatives such that in
each branch every contract must be fulfilled.
Doing this enables us to remove a contracts whose properties
are subsumed by another contracts. Without splitting alternatives it
would not be possible to remove a contract on the basis of other
contracts: The other contracts might be nested in an alternative and
thus it might never apply (as there is a weaker alternative).
As for the baseline rules, unions are split immediately when
asserted to a value, whereas intersections must remain on the values
until the values is used in an application. This is because the context
can choose between both alternative on every use of the value.
Rule U NFOLD /I NTERSECTION (Figure 8) must not be redefined,
as the intersections with flat contracts is equivalent to a conjunction
of its constituents (cf. (Keil and Thiemann 2015a)).
Rule L IFT lifts an immediate contract on a variable bound in
a lambda abstraction and produces a function contract on that abstraction. However, at this level we are only allowed to lift contracts
on variables directly contained in the function body. Lifting from a
deeper abstraction would introduce checks that might never happen
at run time.
Second, it is only allowed to lift an immediate contract. This
is because a lifted delayed contract would remain on the argument
value and thus it would effect all uses of that value, which entirely
changes the meaning of the program.
Rule L OWER pushes a contract on a function body down and
creates a new function contract from that contract. But now we have
the restriction that the function’s return is not it’s argument.
Rule B LAME transforms a function body to a blame term B♭
in case a contract violation happens in that body. As we split
alternatives we know that every contract must be fulfilled, and thus
violating a contract immediately results in a contract violation for
that branch. However, we cannot transform the whole execution to
B♭ as we do not know if the body es ever executed (e.g. because
its is nested in a condition or the lambda term is never used in an
application).
Here, η(ι, ς) computes a blame term B♭ resulting from a predicate violation (ι ◭ false) in ι (cf. Appendix C.3).
As before, @♭ι ⊥ must remain on the B♭ term to remember the
violated contract. However, as the whole body transforms to B♭ , ⊥
may stay as a contract on a function body. In this case, rule L OWER
pushes the information of a failing contract down to the enclosing
body, which in turn is unrolled when used in an application and rule
B LAME transforms the enclosing context to B♭ . Appendix H shows
an example reduction.
The rules B LAME /I F /T RUE and B LAME /I F /FALSE do the same if a
blame appears in branch of a conditions and rule B LAME /G LOBAL
produces a global blame term if ⊥ appears on a top-level term.
Rules S UBSET /L EFT and S UBSET /R IGHT removes a contract from
a term that is less restrictive than another contract on that term. As
this step removes a contract it might change the order of arising
contract violations. The weaker contract might raise a contract violation before the stronger contract is checked. But, as the stronger
contract remains, we know that we definitely result in a blame state
until the stronger contract is satisfier. And, by definition, this implies that the weaker contract is also satisfied.
Rule P USH /I F pushes a contract out of an condition if the contract remains ob both branches of the condition.
Rule M ERGE merges remaining fragments of the same source
contract to a new function. However, this can only be done if
both contracts have the same responsibility, which is indicated by
precondition η(ι1 , ς) = η(ι2 , ς).
F ORK /U NION
F ORK /I NTERSECTION
ς,
ς,
FJT @♭ι (C ∪ D)K
FJ(T1 @♭ι (Q ∩ R)) T2 K
7 →′S
−
7−→′S
(ι ◭ ι1 ∪ ι2 ) : ς,
(ι ◭ ι1 ∩ ι2 ) : ς,
(FJT @♭ι1 CK k FJT @♭ι2 DK)
(FJ(T1 @♭ι1 Q) T2 )K k FJ(T1 @♭ι2 R) T2 )K)
L IFT
L OWER
ς,
ς,
FJλx.BJx@♭ι IKK
FJλx.(T @♭ι C)K
7−→′S
7−→′S
(ι ◭ ¬ι1 ) : ς,
ς
FJ(λx.BJxK)@♭ι1 (I → ⊤)K
FJ(λx.T )@♭ι (⊤ → C)K
B LAME
B LAME /I F /T RUE
B LAME /I F /FALSE
B LAME /G LOBAL
ς,
ς,
ς,
ς,
FJλx.BJT @♭ι ⊥KK
FJif T0 BJT1 @♭ι ⊥K N K
FJif T0 T1 BJT2 @♭ι ⊥KK
BJT @♭ι ⊥K
7−→′S
7−→′S
7−→′S
7−→′S
ς,
ς,
ς,
ς,
FJλx.(B♭ @♭ι ⊥K)
FJif T0 (B♭ @♭ι ⊥) N K
FJif T0 T1 (B♭ @♭ι ⊥)K
B♭
S UBSET /L EFT
S UBSET /R IGHT
ς,
ς,
FJAJT @♭ι1 CK@♭ι2 DK
FJAJT @♭ι1 CK@♭ι2 DK
7 →′S
−
7−→′S
ς,
ς,
FJAJT @♭ι1 CKK
FJAJT @♭ι2 DKK
P USH /I F
ς,
FJ(if T0 A1 JT1 @♭ι CK A2 JT2 @♭ι CK)K
7−→′S
ς
FJ(if T0 A1 JT1 K A2 JT2 K)@♭ι CK
M ERGE
Q=C→⊤
R=⊤→D
ς, FJAJT @♭ι1 QK@♭ι2 RK
7 →′S
−
η(ι1 , ς) = η(ι2 , ς)
ι1 ◭ ι3 : ι2 ◭ ι3 : ς
ι3 6∈ ς
FJAJT @♭ι3 C
→ DKK
ι1 , ι2 ∈
6 ς
ι1 , ι2 6∈ ς
ι1 6∈ ς
T 6= x
B♭
B♭
B♭
B♭
= η(ι,
= η(ι,
= η(ι,
= η(ι,
ς)
ς)
ς)
ς)
C ⊑∗ D
C ⊐∗ D
BASELINE
T RACE
ς, M 7−→B ς ′ , N
ς, M −
7 →′S ς ′ , N
ς, M −
7 →′S ς ′ , N
ς, T JM K 7−→S ς ′ , T JN K
Figure 11. Subset transformation.
M, N
L,
::=
|
::=
|
~ MN
~)
| λx.M | M N | M N | op(M
if M M N | if M M N | if M N M | L@♭ι C
~ MN
~)
λx.M | M N | M N | op(M
if M M N | if M M N | if M N M | L@♭ι C
Figure 12. Final terms and contexts.
We call terms that differ only in contract assertions and blame
terms as structurally equivalent. We write M ≃ N if term M is
structurally equivalent to term N and G ≃ H for the structural
equivalence of contexts.
Definition 4. Two terms are structural equal, written M ≃ N , if
they only differ in contract assertions and blame terms.
Definition 5. Two contexts are structural equal, written M ≃ N ,
if they only differ in contract assertions and blame terms.
Finally, the rule BASELINE lifts baseline transformations to subset transformations, and rule T RACE considers the transformation in
a parallel observation.
Lemma 3 (Progress). For all terms M it holds that either M is a
canonical term T or there exists a transformation step ς, M 7−→S
ς ′, N .
Apart from the subset transformation which uses native subcontracting to reduces contracts subsumed by other contract, we also
use ordinary subcontracting to simplifiy contracts contained in an
intersection or union contract. Appendix F demonstrates this.
5.5 Join Traces
The subset transformation splits intersection and union into separated observation, i.e. its outcome is a set of parallel observations.
Thus, after transforming all observations to canonical terms, we
need to join the remaining fragments to a valid source program.
Joining terms must be one of the last step as it is not possible to
apply further transformations afterwards. Subsequent transformations would mix up contracts from different alternatives and thus
they may change the meaning of a contract.
To this end, Figure 12 defines a context M, which is defined as
usual as a term with a hole. The only exception is that contexts M
did not have holes in a contract assertion.
As our transformation did not change the basic syntax of a
program (it only propagates contracts), terms remain equivalent
(excepting contract assertions) until they transform to a blame
term. It follows that the only difference between parallel terms are
contract assertions and blame terms.
Figure 13 shows the structural equivalence of terms and contexts. The equivalence relations looks through contract assertions
and compares only λJ terms.
Lemma 4. Relation ≃ on terms M and context M is an equivalence relation.
To join terms in parallel observations we only need to walk
through all terms and to synchronise contract assertions and blame
terms. Figure 14 presents the synchronisation of contracts and
terms, indicated by relation T 7−→J M on terms.
First, rule J OIN triggers the synchronization of two canonical
terms nested in a parallel observation and rule M ATCH dissolves a
parallel observation if both terms are identical.
The rules S YNCHRONIZE /L EFT and S YNCHRONIZE /R IGHT handle
the case that one side results in a blame state. In this case the
blame term is omitted and the operation proceeds with other term.
Remember, it is not required to maintain the blame term because
@♭ι ⊥ still remains in the source program.
Finally, rule S YNCHRONIZE /C ONTRACT looks for different contract assertions on a term S and merges the contract assertions.
Precondition G ≡ H requires that the enclosing contexts are structurally equivalent. Only the contract assertions must be different.
Here, Ar ⊔ Al computes the union of both contexts and synchronizing proceeds with the new context in place.
The easiest way to merge contexts is to place one context in
the hole of the other context. Even through this did not change the
blame behaviour of a program it might duplicate contract checks,
and thus it violates our ground rule not to introduce more checks.
Term Equivalence
K≃K
M ≃N
λx.M ≃ λx.N
x≃x
M0 ≃ N0
M1 ≃ N1
M0 M1 ≃ N0 N1
M ≃N
λx.M ≃ λx.N
~ ≃N
~
M
~
~)
op(M ) ≃ op(N
~0 ≃ N
~0
~1 ≃ N
~1
~2 ≃ N
~2
M
M
M
if M0 M1 M2 ≃ if N0 N1 N2
C ONDENSE
ς, MJAJM @♭ι00 CK@♭ι11 CK 7−→C ι1 ◭ ι0 : ς, MJAJM @♭ι00 CKK
Figure 15. Condense transformation.
Thus, merging assertion contexts only places contract assertions
in the hole that are not already contained in the context. Formally,
we specify the join of two assertion contexts A and A′ by
A ⊔ A′ = AJA′ \ AK
M ≃N
M @♭ι C
M ≃N
≃N
M ≃ B♭
M ≃ N @♭ι C
B♭ ≃ N
Context Equivalence
Here, A′ \ A computes a new context with assertions from A′ that
are not already contained in A. Appendix C.4 shows its computation.
5.6 Condense Remaining Contracts
≃
M≃N
M≃N
M@♭ι C ≃ N
M ≃ N @♭ι C
M≃N
M ≃N
MM ≃ N N
M≃N
λx.M ≃ λx.N
M≃N
M ≃N
MM≃NM
Finally, one last step remains to be done: After synchronizing
contracts it might happen that we have identical contracts on the
same term. As they belong to different alternatives we cannot
remove one of them, but we can condense them a single assertion.
Figure 15 shows the transformation. If a term has two assertions
of the same contract, Rule C ONDENSE removes one of them and
creates a new constraint that redirect the result from the other
assertion to the blame variable of the remove assertion.
5.7 Bringing all this together
~0 ≃ N
~0
~i ≃ N
~i
M≃N
M
M
~
~
~
~
op(M0 M Mi ) ≃ op(N0 N Ni )
M≃N
M1 ≃ N1
M2 ≃ N2
if M M1 M2 ≃ if N N1 N2
M≃N
M0 ≃ N0
M2 ≃ N2
if M0 M M2 ≃ if N1 N N2
To model transformation with parallel observations, we define a
new transformation Z=⇒ on terms M . It first applies the standard
transformation 7−→S until it reaches a canonical term T . Second,
it applies 7−→J to join terms and 7−→C to condense the remaining
contracts.
·, M 7−→∗S ς, T
T 7−→∗J L
·, M Z=⇒∗ ς, N
L 7−→∗C N
5.8 Weak Blame-preservation
M≃N
M0 ≃ N0
M1 ≃ N1
if M0 M1 M ≃ if N0 N1 N
Figure 13. Term and context equivalence.
While reorganizing contracts, the Subset Transformation did no
strictly preserve the blame behaviour of a program. An ill-behaved
program by lead to another contract violation first, whereas wellbehaved programs still result in the same output.
Conjecture 2 (Weak Blame-preservation). For all ςM , M Z=⇒∗
′
′
ςN , N it hold that either ςM , M =⇒ ςM
, V and ςN , N =⇒ ςN
,V
′′
♭
′′
♭′
or ςM , M =⇒ ςM , BM and ςN , N =⇒ ςN , BN .
J OIN
(M k N ) −
7 →′J (M ′ k N ′ )
T J(M k N )K 7−→J T J(M ′ k N ′ )K
M ATCH
(M k M ) 7−→′J M
S YNCHRONIZE /L EFT
M≃N
(MJAl JSKK k N JAr JB♭ KK) 7−→′J (MJAl JSKK k N JAr JSKK)
S YNCHRONIZE /R IGHT
M≃N
♭
(MJAl JB KK k N JAr JSKK)
7 →′J
−
(MJAl JSKK k N JAr JSKK)
S YNCHRONIZE /C ONTRACT
M≃N
Al 6= Ar
A = Ar ⊔ Al
(MJAl JSKK k N JAr JSKK) 7−→′J (MJAJSKK k N JAJSKK)
Figure 14. Join parallel observations.
6. Technical Results
In addition to weak and string blame preservation, our contract simplification grantees not to introduce more predicate checks at run
time. Even throut we reorganize and duplicate contract assertions,
the total number of predicate checks at run time remains the same
or decrease on every transformation step.
Definition 6 (Size). The size |L| of a closed term L is the number
of predicates checks during reduction =⇒ of L.
Theorem 1 (Improvement). For each term M with ·, M Z=⇒ ς, N
it holds that |N | ≤ |M |.
Finally, to prove soundness of our static contract simplification,
we need to show that our transformation terminates.
Theorem 2 (Termination). For each term M , there exists a canonical term T such that ςM , M Z=⇒∗ ςT , T .
Benchmark
Example1
Example2
Example3
Example4
Example5
Example6
Normal
time (sec)
39.4
87.1
66.5
114.5
148.3
295.7
Baseline
time (sec)
27.0
58.4
54.4
85.1
108.0
200.0
(-31.27 %)
(-33.00 %)
(-18.17 %)
(-25.63 %)
(-27.18 %)
(-32.36 %)
Subset
time (sec)
27.0
46.1
26.5
44.6
60.0
118.6
(-31.37 %)
(-47.12 %)
(-60.11 %)
(-61.01 %)
(-59.55 %)
(-59.90 %)
Figure 16. Results from running the TreatJS contract system.
Column Normal gives the baseline execution time of the unmodified program, whereas column Baseline and column Subset contain the execution time and the improvement (in percent) after applying the baseline or subset simplification, respectively.
7. Practical Evaluation
To give an insight into the run time improvements of our system
we apply the simplifications to different versions of the addOne
example from Section 2. Our testing procedure uses the addOne
function to increase a counter on each iteration in a while loop.
Each example program addresses a certain property and contains a different number of contracts. For example, Example2 corresponds the addOne function in Section 2.1 and Example4 corresponds to the function in Section 2.6. Appendix J shows the
JavaScript implementation of all benchmark programs.
To run the examples we use the TreatJS (Keil and Thiemann
2015b) contract system for JavaScript and the SpiderMonkey3
JavaScript engine. Lacking an implementation (beyond the PLT
Redex model) we applied the simplification steps manually.
Figure 16 shows the execution time required for a loop with
100000 iterations. The numbers indicate that the Baseline Simplification improves the run time by approximately 28%, whereas
the Subset Simplification makes an improvement up to 62%. Appendix J shows the full table of results.
In addition, there is another aspects not already addressed by
Keil and Thiemann (2015b): Adding contracts may also prevent a
program from being optimized. The SpiderMonkey engine, for example, uses two different optimizing compilers to speed up operation. When activating SpiderMonkey’s optimizing JIT, the run time
of the normal program without contracts decreases by 20x, whereas
the run time of the program with contracts only decreases by 2x.
This lack of optimization opportunities for the JIT is one important
reason for the big run time deterioration they reported.
8. Related Work
Higher-Order Contracts Software contracts were introduced
with Meyer’s Design by ContractTM methodology (Meyer 1988)
which stipulates the specification of Hoare-like pre- and postconditions for all components of a program and introduces the idea of
monitoring these contracts while the program executes.
Findler and Felleisen (2002) extend contracts and contract monitoring to higher-order functional languages. Their work has attracted a plethora of follow-up works that ranges from semantic investigations (Blume and McAllester 2006; Findler and Blume
2006) over studies on blame assignment (Dimoulas et al. 2011;
Wadler and Findler 2009) to extensions in various directions: intersection and union contracts (Keil and Thiemann 2015a), polymorphic contracts (Ahmed et al. 2011; Belo et al. 2011), behavioral and temporal contracts (Dimoulas et al. 2012; Disney et al.
2011), etc.
Contract Validation Contracts may be validated statically or dynamically. However, most frameworks perform run-time monitoring as proposed in Meyer’s work.
Dynamic contract checking, as for example performed by
Findler and Felleisen (2002); Keil and Thiemann (2015b), enable
programmers to specify properties of a component without restricting the flexibility of the underlying dynamic programming language. However, run time monitoring imposes a significant overhead as it extends the original program with contract checks.
Purely static frameworks (e.g., ESC/Java (Flanagan et al. 2002))
transform specifications and programs into verification conditions to be verified by a theorem prover. Others (Xu et al. 2009;
Tobin-Hochstadt and Van Horn 2012) rely on program transformation and symbolic execution to prove adherence to contracts.
Static contract checking avoids additional run-time checks, but
existing approaches are incomplete and limited to check a restricted
set of properties. They rely on theorem proving engines to discharge contracts that are written in the host language: the translation to logic may not always succeed.
Contract Simplification Nguyen et al. (2014) present a static
contract checker that has evolved from Tobin-Hochstadt and Van
Horn’s work on symbolic execution (Tobin-Hochstadt and Horn
2012). Their approach is to verify contracts by executing programs
on unknown abstract values that are refined by contracts as they
are applied. If a function meets its obligation, the corresponding
contract gets removed. However, their approach only applies to the
positive side of a function contract.
Propagating contracts is closely related to symbolic executions.
Both refine the value of a term based on contracts, whether by
moving the contract through the term or by using an abstract value.
Compared to our work, Nguyen et al. (2014) addresses the opposite direction. Where we unroll a contract to its enclosing context
and decompose a contract into its components, they verify the function’s obligations based on the given domain specification. However, in case the given contract cannot be verified, they must retain
the whole contract at its original position and they are not able to
simplify the domain portion of a function contract.
Furthermore, their symbolic verification is not able to handle
true alternatives in the style of intersection and union contracts.
We claim that both approaches are complementary to one another. Our contract simplification would benefit from a preceding verification that simplifies the function’s obligations before unrolling a contract to the enclosing context.
Xu (2012) also combines static and dynamic contract checking. Her approach translates contracts into verification conditions
that get verified statically. Whereas satisfied conditions will be removed, there may be conditions that cannot be proved: they remain
in the source program in the form of dynamic checks.
9. Conclusion
The goal of static contract simplification is to reduce the overhead
of run-time contract checking. To this end, we decompose contracts
and propagates their components through the program. In many
cases we are able to discharge parts of contracts statically. The
remaining components are simplified and reassembled to a new and
cheaper contract at a module boundary.
As a novel aspect, we can simplify contracts that cannot be
verified entirely at compile time. The remaining contract fragments
stay in the program and get lifted to the enclosing module boundary.
A case study with microbenchmarks shows some promise. Our
simplification decreases the total number of predicate checks at
run time and thus reduces the run-time impact caused by contracts
3 https://developer.mozilla.org/en-US/docs/Mozilla/Projects/SpiderMonkey
and contract monitoring. The study also shows that the degree of
improvement depends on the granularity of contracts. Fine-grained
contracts enable more improvement.
References
A. Ahmed, R. B. Findler, J. G. Siek, and P. Wadler. Blame for all. In
Ball and Sagiv (2011), pages 201–214.
T. Ball and M. Sagiv, editors. Proceedings of the 38th ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, POPL
2011, Austin, TX, USA, January 26-28, 2011, 2011. ACM.
J. F. Belo, M. Greenberg, A. Igarashi, and B. C. Pierce. Polymorphic contracts. In G. Barthe, editor, Proceedings of the 20th European Symposium on Programming, volume 6602 of Lecture Notes in Computer Science, pages 18–37, Saarbrücken, Germany, Mar. 2011. Springer-Verlag.
ISBN 978-3-642-19717-8.
M. Blume and D. McAllester. Sound and complete models of contracts.
Journal of Functional Programming, 16:375–414, July 2006. ISSN
0956-7968. doi: 10.1017/S0956796806005971.
C. Dimoulas, R. B. Findler, C. Flanagan, and M. Felleisen. Correct blame
for contracts: No more scapegoating. In Ball and Sagiv (2011), pages
215–226.
C. Dimoulas, S. Tobin-Hochstadt, and M. Felleisen. Complete monitors for
behavioral contracts. In European Symposium on Programming, volume
7211 of Lecture Notes in Computer Science, Tallinn, Estland, Apr. 2012.
Springer-Verlag.
T. Disney. contracts.js. https://github.com/disnet/contracts.js,
Apr. 2013.
T. Disney, C. Flanagan, and J. McCarthy. Temporal higher-order contracts.
In O. Danvy, editor, Proceedings International Conference on Functional Programming 2011, pages 176–188, Tokyo, Japan, Sept. 2011.
ACM Press, New York. ISBN 978-1-4503-0865-6.
M. Felleisen, R. B. Findler, and M. Flatt. Semantics Engineering with PLT
Redex. MIT Press, 2009.
R. B. Findler and M. Blume. Contracts as pairs of projections. In P. Wadler
and M. Hagiya, editors, Proceedings of the 8th International Symposium
on Functional and Logic Programming FLOPS 2006, pages 226–241,
Fuji Susono, Japan, Apr. 2006. Springer-Verlag.
R. B. Findler and M. Felleisen. Contracts for higher-order functions.
In S. Peyton-Jones, editor, Proceedings International Conference on
Functional Programming 2002, pages 48–59, Pittsburgh, PA, USA, Oct.
2002. ACM Press, New York.
C. Flanagan, K. R. M. Leino, M. Lillibridge, G. Nelson, J. B. Saxe, and
R. Stata. Extended static checking for Java. In Proceedings of the 2002
Conference on Programming Language Design and Implementation,
pages 234–245, Berlin, Germany, June 2002. ACM Press. ISBN 158113-463-0. doi: http://doi.acm.org/10.1145/512529.512558.
M.
Flatt, R. B. Findler, and PLT.
The racket guide.
http://docs.racket-lang.org/guide/index.html , Feb. 2014.
Version 6.0.
M. Keil and P. Thiemann. Blame assignment for higher-order contracts with
intersection and union. In K. Fisher and J. H. Reppy, editors, ICFP 2015,
pages 375–386, Vancouver, BC, Canada, Sept. 2015a. ACM.
M. Keil and P. Thiemann. Treatjs: Higher-order contracts for JavaScript.
In J. Boyland, editor, 29th European Conference on Object-Oriented
Programming, ECOOP 2015, July 5-10, 2015, Prague, Czech Republic,
volume 37 of LIPICS, pages 28–51, Prague, Czech Repulic, July 2015b.
Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik. ISBN 978-3939897-86-6.
G. T. Leavens and M. B. Dwyer, editors. Proceedings of the 27th Annual
ACM SIGPLAN Conference on Object-Oriented Programming, Systems,
Languages, and Applications, OOPSLA 2012, part of SPLASH 2012,
Tucson, AZ, USA, October 21-25, 2012, 2012. ACM.
B. Meyer. Object-Oriented Software Construction. Prentice-Hall, 1988.
P. C. Nguyen, S. Tobin-Hochstadt, and D. V. Horn. Soft contract verification. In J. Jeuring and M. M. T. Chakravarty, editors, ICFP’14, pages
139–152, Gothenburg, Sweden, Sept. 2014. ACM. ISBN 978-1-45032873-9. URL http://dl.acm.org/citation.cfm?id=2628136.
A. Takikawa, D. Feltey, B. Greenman, M. S. New, J. Vitek, and M. Felleisen.
Is sound gradual typing dead? In R. Bodı́k and R. Majumdar, editors, POPL 2016, pages 456–468, St. Petersburg, FL, USA, Jan. 2016.
ACM. ISBN 978-1-4503-3549-2. doi: 10.1145/2837614.2837630. URL
http://doi.acm.org/10.1145/2837614.2837630 .
S. Tobin-Hochstadt and D. V. Horn. Higher-order symbolic execution via
contracts. In Leavens and Dwyer (2012), pages 537–554.
S. Tobin-Hochstadt and D. Van Horn. Higher-order symbolic execution via
contracts. In Leavens and Dwyer (2012), pages 537–554.
P. Wadler and R. B. Findler. Well-typed programs can’t be blamed. In
G. Castagna, editor, Programming Languages and Systems, 18th European Symposium on Programming, ESOP 2009, Held as Part of
the Joint European Conferences on Theory and Practice of Software,
ETAPS 2009, York, UK, March 22-29, 2009. Proceedings, volume 5502
of Lecture Notes in Computer Science, pages 1–16. Springer, 2009.
ISBN 978-3-642-00589-3. doi: 10.1007/978-3-642-00590-9. URL
http://dx.doi.org/10.1007/978-3-642-00590-9 .
D. N. Xu. Hybrid contract checking via symbolic simplification. In O. Kiselyov and S. Thompson, editors, Proceedings of the ACM SIGPLAN 2012
Workshop on Partial Evaluation and Program Manipulation, PEPM
2012, Philadelphia, Pennsylvania, USA, January 23-24, 2012, pages
107–116. ACM, 2012. ISBN 978-1-4503-1118-2.
D. N. Xu, S. Peyton Jones, and K. Claessen. Static contract checking for
Haskell. In B. Pierce, editor, Proceedings 36th Annual ACM Symposium
on Principles of Programming Languages, pages 41–52, Savannah, GA,
USA, Jan. 2009. ACM Press. ISBN 978-1-60558-379-2.
V = false
false
τ (V ) = τ (W ) V = W @♭ι Q
true
otherwise
Figure 17. Mapping values to truth values.
CS-E MPTY
µ |= ·
CS-S TATE
µ |= κ
µ |= ς
µ |= κ : ς
CT-I NDIRECTION
µ(♭.subject) ⊒ µ(ι.subject)
µ(♭.context) ⊒ µ(ι.context)
µ |= ♭ ◭ ι
C T-F LAT
µ(b.subject) ⊒ τ (V )
µ(b.context) ⊒ true
µ |= b ◭ V
CT-F UNCTION
µ(b.subject) ⊒ µ(ι1 .context ∧ (ι1 .subject ⇒ ι2 .subject))
µ(b.context) ⊒ µ(ι1 .subject ∧ ι2 .context)
µ |= b ◭ ι1 → ι2
CT-I NTERSECTION
µ(b.subject) ⊒ µ(ι1 .subject ∧ ι2 .subject)
µ(b.context) ⊒ µ(ι1 .context ∨ ι2 .context)
µ |= b ◭ ι1 ∩ ι2
C T-U NION
µ(b.subject) ⊒ µ(ι1 .subject ∨ ι2 .subject)
µ(b.context) ⊒ µ(ι1 .context ∧ ι2 .context)
µ |= b ◭ ι1 ∪ ι2
CT-I NVERSION
µ(♭.subject) ⊒ µ(ι.context)
µ(♭.context) ⊒ µ(ι.subject)
µ |= ♭ ◭ ¬ι
Figure 18. Constraint satisfaction.
A. Constraint Satisfaction
The dynamics in Figure 3 use constraints to create a structure for
computing positive and negative blame according to the semantics
of subject and context satisfaction, respectively. To this end, each
blame identifier b is associated with two truth values, b.subject and
b.context. Intuitively, if b.subject is false, then the contract b is not
subject-satisfied and may lead to positive blame for b. If b.context
is false, then there is a context that does not respect contract b and
may lead to negative blame for b.
An interpretation µ of a constraint list ς is a mapping from
blame identifiers to records of elements of B = {true, false}, such
that all constraints are satisfied. We order truth values by true ⊏
false and write ⊑ for the reflexive closure of that ordering. This
ordering reflects gathering of informations with each execution
step.
Formally, we specify the mapping by
µ ∈ (LbM × {subject, context}) → B
where LbM ranges over metavariable b and constraint satisfaction by
a relation µ |= ς, which is specified in Figure 18.
In the premisses, the rules apply a constraint mapping µ to
boolean expressions over constraint variables. This application
stands for the obvious homomorphic extension of the mapping.
Every mapping satisfies the empty list of constraints (CS-E MPTY).
The concatenation of a constraint with a constraint list corresponds
to the intersection of sets of solutions (CS-S TATE). The indirection
constraint just forwards its referent (CT-I NDIRECTION).
In rule C T-F LAT, W is the outcome of the predicate of a flat
contract. The rule sets subject satisfaction to false if W = false
and otherwise to true, where the function τ (·) : LV M → B
translates values to truth values by stripping delayed contracts (see
Figure 17). A flat contract never blames its context so that b.context
is always true.
The rule CT-F UNCTION determines the blame assignment for a
function contract b from the blame assignment for the argument
and result contracts, which are available through ι1 and ι2 . Let’s
first consider the subject part. A function satisfies contract b if it
satisfies its obligations towards its argument ι1 .context and if the
argument satisfies its contract then the result satisfies its contract,
too. The first part arises if the function is a higher-order function,
which may pass illegal arguments to its function-arguments. The
second part is partial correctness of the function with respect to its
contract.
A function’s context (caller) satisfies the contract if it passes
an argument that satisfies contract ι1 .subject and uses the result
according to its contract ι2 .context. The second part becomes nontrivial with functions that return functions.
The rule CT-I NTERSECTION determines the blame assignment
for an intersection contract at b from its constituents at ι1 and
ι2 . A subject satisfies an intersection contract if it satisfies both
constituent contracts: ι1 .subject ∧ ι2 .subject. A context, however,
has the choice to fulfill one of the constituent contracts: ι1 .context∨
ι2 .context.
Dually, the rule C T-U NION determines the blame assignment
for a union contract at b from its constituents at ι1 and ι2 . A
subject satisfies a union contract if it satisfies one of the constituent
contracts: ι1 .subject ∨ ι2 .subject. A context, however, needs to
fulfill both constituent contracts: ι1 .context ∧ ι2 .context, because
it does not know which contract is satisfied by the subject.
Canonical Terms
SVal
::=
SNonVal ::=
|
|
|
Canonical Terms
K | λx.S
x | B♭
TI T | TConst T | TAbs TI | TAbs TVal
~ )
op(TNonQ
if T0 T1 T2 where not T1 = T1′ @♭ι C and
T2 = T2′ @♭ι C
SVal
SNonVal
::=
::=
|
|
|
K | λx.S where λx.S 6= λx.BJx@♭ι IK
x | B♭
TI TQ | K TQ | SVal TI | SVal S
~ )
op(TNonQ
if TQ AJTQ′ K AJTQ′′ K where not TQ′ = T ′ @♭ι C and
TQ′′ = T ′′ @♭ι C
S
::=
SVal | SNonVal
TI
TConst
TAbs
TVal
::=
::=
::=
K | TConst @♭ι ⊥
λx.S | TAbs @♭ι ⊥
SVal | TVal @♭ι ⊥
::=
|
SNonVal | SNonVal @♭ι I
AJTI @♭ι00 IK@♭ι11 J where I 6⊑∗ J, J 6⊑∗ I
TQ
::=
|
TI
TQ
TNonQ
::=
::=
::=
SNonVal | TI @♭ι I | TI @♭ι ⊥
TVal | TI | TQ @♭ι Q
TVal | TI
SVal | TI | SVal @♭ι Q | TI @♭ι Q
AJTQ @♭ι00 QK@♭ι11 R where Q 6⊑∗ R, Q 6⊑∗ R
and not Q 6= C → ⊤, R 6= ⊤ → D, ♭0 = ♭1
T
::=
TQ | (Tr k Tl ) | B♭ @♭ι ⊥
T
::=
TQ
Non-Canonical Terms
O, P
::=
|
|
|
|
|
|
~ ON
~)
λx.O | O N | M O | op(M
if O M N | if L O N | if L M O | O@♭ι C
(VJλx.M K) (N @♭ι Q) | (M @♭ι Q) N
VJV K@♭ι I | λx.(M @♭ι C)
~ (M @♭ι Q) N
~ ) | if L (M @♭ι C) (N @♭ι C)
op(L
♭
♭
M @ι ⊤ | M @ C | M @♭ι C ∪ D | M @♭ι I ∩ C
(M @♭ιq Q)@♭ιi I | (M @♭ιq Q)@♭ιi ⊥
Figure 19. Canonical and non-canonical terms of the Baseline
Transformation.
B. Canonical Terms
Canonical (non-transformable) terms are the output of our transformation. Each level has its own output grammar consisting of λCon
term M that cannot be further transformed at this level.
Canonical terms distinguish contract-free term (indicated by
latter S) from terms with a contract (indicated by latter T ).
B.1 Canonical Term of the Baseline Transformation
This sections considers the output terms of the Baseline Transformation (Section 4.3). Figure 19 defines the syntax of canonical
terms S and T as a subset of λCon term M .
A terms without a contract S is either a value term SVal , including first-order constants and lambda abstractions, or a non-value
term SNonVal , which is either a variable, a blame term, an application, a primitive operation, or a condition.
Applications are divided into different combinations. This excludes all reducible combinations and only the given terms remain.
For example, a function contract on the left side of an application
is reducible, whereas a function contract in the right side is only
reducible if the left side is a lambda term.
A canonical term T is a term TQ , which is either a term TVal
or a term TI with an indefinite number of delayed contracts on
the outermost position. Terms TVal are all values with remaining
⊥ contracts. Terms TI are all terms with immediate contracts or ⊥
contracts on the outermost position. As all immediate contracts on
values will be reduced, immediate contracts only remain on nonvalue terms SNonVal .
Non-Canonical Terms
O, P
+=
|
|
|
|
AJM @♭ι00 CK@♭ι11 D where C 6⊑∗ D, D 6⊑∗ C
AJM @♭ι0 CK@♭ι1 D where C ⊑
6 D, D 6⊑ C
λx.BJx@♭ι IK
if L (M @♭ι C) (N @♭ι C)
M @♭ι ⊥
Figure 20. Canonical and non-canonical terms of the Subset
Transformation.
In addition to the canonical terms, Figure 19 shows the syntax of
transformable (non-canonical) terms O, which is the complement
of the canonical terms T .
A non-canonical term O is either a term that contains a noncanonical term or one of the patterns addressed by our transformation rules in Section 4.3.
Lemma 5 (Terms). For all terms M it holds that M is either a
canonical term T or a transformable (non-canonical) term O.
B.2 Canonical Term of the Subset Transformation
Figure 20 shows the definition of canonical term T and source
terms S as a restriction of the already defined canonical terms.
For readability, we reuse the same symbols as in Section ??. The
repeatedly defined terms replace the already existing definition,
whereas every other definition remains valid.
A lambda abstractions λx.S now did not contain immediate
contracts on its argument x and terms only contain contracts that
are not subsumed by another contract on that term. Terms T now
contains a blame terms with a remaining ⊥ contract.
Lemma 6 (Terms). For all terms M it holds that M is either a
canonical term T or a transformable (non-canonical) term O.
K ::= | κ : K
Figure 21. Constraint context.
root-of(b, ς) =
♭,
root-of(b′ ,
root-of(b′ ,
root-of(b′ ,
root-of(b′ ,
root-of(b′ ,
root-of(b′ ,
root-of(b′ ,
root-of(b′ ,
b,
ς),
ς),
ς),
ς),
ς),
ς),
ς),
ς),
b=♭
b = ι, KJ(b′ ◭ ι) : ςK
b = ι0 , KJ(b′ ◭ ι0 → ι1 ) : ςK
b = ι1 , KJ(b′ ◭ ι0 → ι1 ) : ςK
b = ι0 , KJ(b′ ◭ ι0 ∩ ι1 ) : ςK
b = ι1 , KJ(b′ ◭ ι0 ∩ ι1 : ςK
b = ι0 , KJ(b′ ◭ ι0 ∪ ι1 ) : ςK
b = ι1 , KJ(b′ ◭ ι0 ∪ ι1 ) : ςK
b = ι, KJ(b′ ◭ ¬ι) : ςK
otherwise
Figure 22. Function root-of.
+blame,
sign-of(b′ , ς),
¬sign-of(b′ , ς),
sign-of(b′ , ς),
sign-of(b, ς) = sign-of(b′ , ς),
sign-of(b′ , ς),
sign-of(b′ , ς),
sign-of(b′ , ς),
¬sign-of(b′ , ς),
Figure 24. Function sign-of.
C.3 Computation of Blame Terms
To computes a blame term B♭ resulting from a predicate violation
(ι ◭ false) in ι we need to compute the sign of ι in ς and to annotate
the resulting blame term with the root blame identifier ♭ of ι in ς.
B = sign-of(b, ς)
¬B =
(
b=♭
b = ι, KJ(b′ ◭ ι) : ςK
b = ι0 , KJ(b′ ◭ ι0 → ι1 ) : ςK
b = ι1 , KJ(b′ ◭ ι0 → ι1 ) : ςK
b = ι0 , KJ(b′ ◭ ι0 ∩ ι1 ) : ςK
b = ι1 , KJ(b′ ◭ ι0 ∩ ι1 : ςK
b = ι0 , KJ(b′ ◭ ι0 ∪ ι1 ) : ςK
b = ι1 , KJ(b′ ◭ ι0 ∪ ι1 ) : ςK
b = ι, KJ(b′ ◭ ¬ι) : ςK
♭ = root-of(b, ς)
η(ι, ς) = B♭
-blame, B = +blame
+blame, B = -blame
Figure 23. Blame inversion.
C. Auxiliary Functions and Definitions
This section shows some auxiliary functions and definitions used
in the paper.
C.1 Function root-of(b, ς)
Each top-level contract assertion is rooted in one blame label ♭, i.e.
for every blame identifier b there exists a path from b to a blame
label ♭.
To computer the root label ♭ for an identifier b we need to walk
backwards through all constraints κ on a path until we reach the top
level blame label ♭.
To do so, Figure 21 introduces a constraint context K, which is
either a hole or constraint κ extending a constraint context K.
Figure 22 finally shows the computation of function root-of(b, ς).
C.2 Sign of a Blame Identifier
Following the definition in Section 3, constraints κ implement a
tree rooted in a blame label ♭. Within the tree every blame identifier
b can report positive or negative blame, depending on its constraint
when mapping one of its truth values (ι.subject or ι.context) to
false.
Thus, we define the sign of a blame identifier b as the responsibility to satisfy the subject part in this identifier b. We represent
signs by blame term B that would result from a subject blame in b.
Definition 7. B is the sign of a blame identifier b in ς if (b ◭
false) : ς leads to B♭ on a single path for any ♭.
In this case, every alternative must be treated as neutral, i.e. they
do not stop or change the arising blame. Thus we consider only a
single path from b to its root ♭.
To compute the sign of a blame identifier b, Figure 23 first defines an inversion of blame terms B that flips the responsibility. Figure 24 than shows the computation of the sign of blame identifier
b. It is defined by induction and tracks swapping of responsibilities
on function and inversion constraints.
C.4 Compute joined Assertion Context
We define the join of two assertion contexts A and A′ by
A ⊔ A′ = AJA′ \ AK
Term A′ \ A computes a new context with assertions from A′ that
are not already contained in A. Its computation is defined as:
′′
′
A = A′′ @♭ι C ∧ @♭ι C ∈ A′
A \ A ,
′
′′
′
♭
A \ A = (A \ A )@ι C, A = A′′ @♭ι C ∧ @♭ι C 6∈ A′
A,
otherwise
We write @♭ι C ∈ A to indicate that the assertion of C under
blame label ♭ and blame variable ι is contained in A, and @♭ι C 6∈
A for its negation. Contract entailment is defined by:
@♭ι C ∈ A@♭ι C
@♭ι C ∈ A
@♭ι C ∈ A@♭ι D
P USH /I MMEDIATE
ς, FJ(T @♭ι1 Q)@♭ι2 IK 7−→B ς FJ(T @♭ι2 (I ←֓ (@♭ι1 Q)))@♭ι1 QK
Figure 25. Changes for Picky Semantics.
E XTEND /P REDICATE
M = λx.N
ς, FJflat(M ) ←֓ (@♭ι Q)K 7−→B ς, FJλx.N [x 7→ (x@♭ι Q)]K
E XTEND /N ON -P REDICATE
M 6= λx.N
ς, FJflat(M ) ←֓ (@♭ι Q)K 7−→B ς, FJflat(M )K
Figure 26. Unroll delayed contract.
D. Picky Semantics
The semantics in Section 4 guarantee Strong Blame-preservation
with respect to Lax evaluation semantics (cf. Section 3.5). Lax unpacks contracts from values that flow to a predicate, i.e. a contract
cannot violate another contract.
The Baseline Transformation makes this intuitively when processing the rule P USH /I MMEDIATE. It pushes a delayed contract out
of an assertion with an immediate contract, such that the predicate
will not see the delayed contract. This step is entirely required for
following transformation steps.
When considering Picky (or even Indy) evaluation semantics,
a contract must remain on a value when the value is used in a
predicate. Consequently, we cannot pushes a delayed contract out
of an assertion with an immediate contract without changing the
blame behaviour of the underlying program.
To enables Picky (or Indy) semantics, rule P USH /I MMEDIATE
(Figure 8) must be replayed by another rule that unrolls a delayed
contract in an immediate contract when pushing the delayed contract. Figure 25 shows the new rule.
Here, term T @♭ι (I ←֓ (@♭ι′ Q)) is a new intermediate term
that unrolls a delayed contract in a predicate (cf. rule U NROLL).
Figure 26 shows its semantics.
Rule E XTEND /P REDICATE unrolls a delayed contract to all uses
of the value in a predicate and rule E XTEND /P REDICATE proceeds on
the unchanged predicate in case that the predicate is not a lambda
abstraction.
Indeed, it would also be possible to replace predicate M by
M @♭ι Q → ⊤ as it enforces Q when the predicate is used in
an application. However, unrolling the contract enables to apply
further simplification steps in the function’s body.
(2) From lemma 7, N and =⇒ it follows that ςN , (T1 (T2 @♭ι′ C))@♭ι′ D −→∗
E. Proofs of Theorems
1
E.1 Proof of Lemma 1
Proof of Lemma 1. Immediate from the definition of canonical and
non-canonical terms (Section B.1).
E.2 Proof of Lemma 4
Proof of Lemma 4. ≃ is an equivalence relation if and only if it is
reflexive, symmetric, and transitive. Proof by induction over ≃.
E.3 Proof of Conjecture 1
E.4 Congruence of States and Terms
Definition 8. Two configurations (ςM , M ) and (ςN , N ) are structural equivalent under renaming α, written ςM , M ∼
= α ςN , N
with α = ∅ | α[ι 7→ ι′ ] iff ςM , M ∼
=α′ (ςN , N )[ι 7→ ι′ ] with
α = α′ [ι 7→ ι′ ].
ι1
ι2
′′
Claim holds because (ι ◭ ι1 → ι2 ) : ςM
, (V1 (V2 @♭ι1 C))@♭ι2 D ∼
=α
′′
♭
♭
ςN , (V1 (V2 @ι′ C))@ι′ D with α = ∅[ι1 7→ ι′2 ][ι2 7→ ι′2 ].
1
2
Rule U NFOLD /D-I NTERSECTION: Equivalent to case U NFOLD /D-F UNCTION.
Rule U NROLL: Immediate from rule B ETA.
Rule L OWER: Immediate from rule D-F UNCTION.
Rule P USH /I MMEDIATE: Immediate from rule F LAT.
Rule P USH /FALSE: Immediate from rule F LAT.
Rule P USH /I F: Immediate from rules T RUE and FALSE.
Rule C ONVERT /T RUE: Immediate from rule U NIT.
E.6 Proof of Lemma 2
Lemma 7. For all configurations (ςM , M ) and (ςN , N ) with
′
ςM , M ∼
, M ′ and ςN , N −→
= ςN , N it holds that ςM , M −→ ςM
′
′
′
ςN
, N ′ with ςM
, M′ ∼
, N ′.
= ςN
Proof of Lemma 2. ⊑∗ is a preorder if and only if it is reflexive and
transitive. Proof by induction over ⊑∗ .
Proof of Theorem 7. Proof by induction over ς and M .
E.7 Proof of Lemma 3
E.5 Congruence of States and Terms
Proof of Lemma 3. Immediate from the definition of canonical and
non-canonical terms (Section B.2).
Definition 9. State ς ′ is an effect-free extension of state ς, written
ς 4 ς ′ , iff ς, M −→ ς ′′ , M ′ and ς ′ , M −→ ς ′′′ , M ′ with ς ′′ 4 ς ′′′ .
E.8 Proof of Conjecture 2
Lemma 8. For terms M , states ς, and transformations ς, M Z=⇒
ς ′ , N it hold that ς 4 ς ′ .
Proof of Theorem 7. Immediate by the definition of −→ and Z=⇒.
Every new constrains is composed of fresh blame variable ι. Proof
by induction over Z=⇒.
♭
Proof Sketch for Conjecture 1. Let reduction output O ::= V | B .
For all terms M and states ςM with ςM , M 7−→B ςN , N it hold
′
′
′ ∼
′
that ςM , M =⇒ ςM
, O and ςN , N =⇒ ςN
, O with ςM
.
= α ςN
As our transformation only propagates contract assertions and
did not modify the underlying program, we know that if ς, M −→∗
ς ′ , V and ς, N −→∗ ς ′′ , W then V = W .
To prove preservation, we show that the transformed term results in the same reduction sequence than the original term and that
the both, state and terms, are equivalent under variable renaming I.
Proof by induction over 7−→B .
Rule U NFOLD /A SSERT: M = T @♭ C, N = T @♭ι C, and ςN =
(♭ ◭ ι′ ) : ςM .
′
(1) From M and =⇒ it follows that ςM , T @♭ C −→∗ ςM
, V @♭ C
′
, V @♭ C −→ (♭ ◭
and by rule A SSERT it follows that ςM
′
ι) : ςM
, V @♭ι C.
′
(2) From lemma 7 it follows the ςN , T @♭ι′ C −→∗ ςN
, V @♭ι′ C
′
′
with ςM 4 ςN .
′
′
Claim holds because (♭ ◭ ι) : ςM
, V @♭ι C ∼
, V @♭ι′ C
= α ςN
with α = ∅[ι 7→ ι′ ].
Rule U NFOLD /U NION: Equivalent to case U NFOLD /A SSERT.
Rule U NFOLD /I NTERSECTION: Equivalent to case U NFOLD /A SSERT.
Rule U NFOLD /D-F UNCTION: M = (T1 @♭ι (C → D)) T2 , N =
(T1 (T2 @♭ι1 C))@♭ι2 D, and ςN = (ι ◭ ι′1 → ι′2 ) : ςM .
(1) From M and =⇒ it follows that ςM , (T1 @♭ι (C → D)), T2 −→∗
′
′
, (V1 @♭ι (C → D)), T2 and ςM
, (V1 @♭ι (C → D)), T2 −→∗
ςM
′′
♭
ςM , (V1 @ι (C → D)), V2 . By rule D-F UNCTION it follows
′′
that ςM
, (V1 @♭ι (C → D)), V2 −→ (ι ◭ ι1 → ι2 ) :
′′
, (V1 (V2 @♭ι1 C))@♭ι2 D.
ςM
2
′
′
, (V1 (T2 @♭ι′ C))@♭ι′ D and ςN
ςN
, (V1 (T2 @♭ι′ C))@♭ι′ D −→∗
1
2
1
2
′′
♭
♭
′′ ∼
′′
ςN , (V1 (V2 @ ′ C))@ ′ D with ςM =α ςN
.
Proof Sketch for Conjecture 2. Equivalent to proof of Conjecture
1. However, certain rules change the order of contract checks. So
wen need to consider two cases: one where the early contract is
satisfied, and one where it is violated.
E.9 Proof of Theorem 1
Proof of Theorem 1. Proof by induction over Z=⇒. It is immediate
from the definition that all rules, expect rule U NROLL, either decompose, eliminate, or graft a contract. It remains to shows that:
1. no rules duplicates contract checks, and
2. no rule pushes contract checks from one function body (or
branch in a condition) to the enclosing body.
Rule U NROLL . This rule grafts the assertion of a delayed contract
to all uses of the contracted value. From −→, we have that for
all λx.M (N @♭ι Q) first N −→∗ V and by rule B ETA it follows
that λx.M (V @♭ι Q) reduces to λx.M [x 7→ (V @♭ι Q)]. By rule
U NROLL we have that λx.M (N @♭ι Q) translates to λx.M [x 7→
(x@♭ι Q)] N . From −→, we now have that N −→∗ V and by
rule B ETA if follows that λx.M [x 7→ (x@♭ι Q)][x 7→ V ], which
is equivalent to the result of the original reduction.
Rule L IFT . This rule lifts an immediate contract nested in a function body to a new function contract. However, this contract
never applies if the function is not executed. Thus, it preserves
the number contract checks.
Rule L OWER . This rule lowers a contract to the function boundary.
But, this contract never applies if the function is not executed.
It also preserves the number contract checks.
Rule P USH /I F . This rule pushes an immediate contract out of a
condition, if the contract definitely applies to both branches.
Thus, it preserves the number contract checks.
Otherwise Every other rule either decomposes or eliminates a
contract, or reverses the order of contract checks. Thus they
reduces or preserves the total number contract checks at run
time.
Proof of Theorem 2. The transformation terminates with a canonical term because (1) there are no cycles and (2) there is no transformation step that introduces new contract check.
1. All rules, except rule U NROLL, decompose or eliminate a contract, or grafts a contract to an enclosing term. Only rule
U NROLL duplicates a contract to all uses of the contracted value.
However, as the transformation did not lift delayed contracts, it
is not a cycle.
2. Immediate from Theorem 1.
S IMPLIFY /U NION /1
C⊑D
ς, FJV @♭ (C ∩ D)K 7−→B ς, FJV @♭ DK
S IMPLIFY /U NION /2
D⊑C
♭
ς, FJV @ (C ∩ D)K 7−→B ς, FJV @♭ CK
S IMPLIFY /I NTERSECTION /1
C⊑D
ς, FJV @♭ (C ∩ D)K 7−→B ς, FJV @♭ CK
S IMPLIFY /I NTERSECTION /2
D⊑C
♭
ς, FJV @ (C ∩ D)K 7−→B ς, FJV @♭ DK
Figure 27. Contract simplification.
F. Simplify Intersection and Union
Apart from the subset optimization, which reduces a contract assertion based on more restrictive assertion, we can also use the subcontract relation to simplify contracts before unfolding.
For example, the intersection contract Positive? → Natural? ∩
Natural? → Positive? can be simplified to a single function contract Positive? → Positive?. This is because, the context can choose
to call the function with a natural or positive number. However, in
case the input is a natural number, but not a positive number, than
the range must be a positive number. And, in case the input is a positive number, than it is also a natural number (as positive numbers
are a proper subset of natural numbers) and thus the output needs
to satisfy both range contract, i.e. it must be a positive number.
So, we can simplify the intersection because Natural? →
Positive? is an ordinary subset of Positive? → Natural? (cf. Section
5.2). Its domain is less or equal restrictive and its range is more or
equal restrictive.
In general, we can simplify intersection and union contracts
if one operand is an ordinary subset of the other operand. Figure
27 shows some simplification rules that omits one branch of an
alternative if that alternative is subsumed by the other branch.
To keep the system deterministic, we must simplify contracts at
assertion time and thus we skip rule U NFOLD /A SSERT in this cases.
::=
::=
|
|
|
K | λx.S where λx.S 6= λx.BJx@♭ι CK
x | B♭
TI TQ | K TQ | SVal TI | SVal S
~I ) | op(S~Val )
op(T
TI
::=
|
SNonVal | SNonVal @♭ι I
AJTI @♭ι00 IK@♭ι11 J where I 6⊑∗ J, J 6⊑∗ I
TQ
::=
|
SVal | TI | SVal @♭ι Q | TI @♭ι Q
AJTQ @♭ι00 QK@♭ι11 R where Q 6⊑∗ R, Q 6⊑∗ R
and not Q 6= C → ⊤, R 6= ⊤ → D, ♭0 = ♭1
T
::=
TQ | (Tr k Tl )
SVal
SNonVal
if TQ AJTQ′ K AJTQ′′ K where not TQ′ = T ′ @♭ι C and
TQ′′ = T ′′ @♭ι C
Figure 28. Canonical terms and contexts.
L IFT
ι1 6∈ ς
ς,
FJλx.GJx@♭ι CKK
7 →′A
−
x ∈ (free? GJxK)
(ι ◭ ¬ι1 ) : ς, FJ(λx.GJxK)@♭ι1 C → ⊤K
B LAME
ς, FJT @♭ι ⊥K 7−→A (ι ◭ false) : ς, FJT K
S UBSET
T RACE
ς, M 7−→′S ς ′ , N
ς, M −
7 →′A ς ′ , N
ς, M −
7 →′A ς ′ , N
ς, T JM K 7−→A ς ′ , T JN K
Figure 29. Approximations rules.
G. Introducing new Success Contracts
Our core rules in Section 4 and Section 5 follows the overall guideline not to approximate contract violations. However, developers
are free to choose a higher degree of optimization, while overapproximating contract failures at compile time and introducing
new module boundaries.
This level’s rules create some kind of success contracts by overapproximating contracts and by lifting contracts to the upper-most
boundary. However, this is not a real optimization as it might lift
checks that never happen at run time.
Figure 28 defines the syntax of our canonical terms as a subset
of λCon term M , and Figure 29 shows possible transformation.
Rule L IFT lifts every contract on an argument x. This step overapproximates contract violations as it also lifts contracts nested in
conditions and function applications that might never apply.
Rule B LAME immediately updates the constraint state with information about a failing contract. This might result in a blame state,
even if the ill behaved term is never executed.
Finally, the rule S UBSET lifts to the Subset Reduction 7−→S and
rule T RACE unpacks parallel observation. Obviously, rule B LAME
replaces the rule with the same name in 7−→S and we are never
allowed to unroll a once lifted contract. Otherwise we would end
up in a cycle.
H. Propagating Blame
In the preceding example we have seen that all contracts of a failing
alternative can be removed and only one ⊥ must remain on the
first failing term. As we split alternatives in different observation
where every contract must be fulfilled, we can report a violation for
that branch immediately when observing a violation. Every other
contract can be ignored as this branch definitely runs into an error.
However, this is not completely true. We are only allowed to
remove contracts in the same context as the violation occurs. We
cannot treat a contract as violated if the violation only occurs in a
certain aspect of the program.
For example, consider the following source program.
66
67
68
69
70
let f =
((λ addOne (λ z (addOne z)))
((λ plus (λ z ((plus 0) z)))
[(λ x (λ y (+ x y)))
@ (Positive? → (Positive? → Positive?))]))
After applying some simplification steps, we obtain the following
intermediate term.
71
72
73
74
let f =
((λ addOne (λ z (addOne z)))
((λ plus (λ z ([(plus [0 @ ⊥]) @ (Positive? → Positive?)] z)))
(λ x (λ y (+ x y)))))
Even though we do not know that addOne violates the contract on
plus, we are not allowed treat the contract as violated because we do
not know if addOne is ever executed. But, we can replace addOne’s
body by a blame term that reports a violation whenever the function
is executed. In addition, a ⊥ remains on the functions body.
75
76
77
78
let f =
((λ addOne (λ z (addOne z)))
((λ plus (λ z ([blame @ ⊥])))
(λ x (λ y (+ x y)))))
Having a blame term and ⊥ seems to be twofold, but ⊥ can
be lowerd to a new function contract on addOne, which is than
unrolled to all places where addOne is used, as the following
example demonstrates.
79
80
81
let f =
((λ addOne (λ z ([addOne @ (⊤ → ⊥)] z)))
((λ plus (λ z blame)) (λ x (λ y (+ x y)))))
Now, the same transformation steps apply as for normal contract.
The function contract gets unrolled, whereby ⊥ gets asserted to another term. If it so happens the next function body gets transformed
to a blame term with a remaining ⊥.
Continuing in this way propagates the information of a failing
contact through applications and conditions until it reaches the
outermost boundary. Within this boundary we definitely run into
a contract violation. The following example shows the result.
82
83
84
let f =
[((λ addOne (λ z (addOne z)))
((λ plus (λ z blame)) (λ x (λ y (+ x y))))) @ (⊤ → ⊥)]
Benchmark
Richards
DeltaBlue
Crypto
RayTrace
EarleyBoyer
RegExp
Splay
SplayLatency
NavierStokes
pdf.js
Mandreel
MandreelLatency
Gameboy Emulator
Code loading
Box2DWeb
zlib
TypeScript
Full
Baseline
1 day, 18 hours, 21 min, 20 sec
2 days, 10 hours, 36 min, 49 sec
9 sec
23 hours, 12 min, 37 sec
1 min, 9 sec
9 sec
19 min, 19 sec
19 min, 19 sec
11 sec
6 sec
5 sec
5 sec
4 hours, 28 min, 28 sec
12 sec
5 hours, 19 min, 49 sec
11 sec
22 min, 46 sec
8 sec
4 sec
8 sec
4 sec
53 sec
8 sec
3 sec
3 sec
4 sec
6 sec
5 sec
5 sec
5 sec
9 sec
6 sec
11 sec
24 sec
Figure 30. Timings from running the Google Octane 2.0 Benchmark Suite. Column Full shows the time required to complete with
contract assertions. The last column Baseline gives the baseline
time without contract assertion.
Benchmark
Richards
DeltaBlue
Crypto
RayTrace
EarleyBoyer
RegExp
Splay
SplayLatency
NavierStokes
pdf.js
Mandreel
MandreelLatency
Gameboy Emulator
Code loading
Box2DWeb
zlib
TypeScript
Asssert
Internal
Predicate
24
54
1
42
3944
0
10
10
51
3
7
7
3206
5600
20075
0
4
1599377224
2319477672
5
687240082
89022
0
11620663
11620663
48334
15
57
57
141669753
34800
172755100
0
12673644
935751200
1340451212
3
509234422
68172
0
7067593
7067593
39109
9
28
28
97487985
18400
11266494
0
8449090
Figure 31. Statistic from running the Google Octane 2.0 Benchmark Suite. Column Assert shows the numbers of top-level contract assertions. Column Internal contains the numbers of internal
contract assertions whereby column Predicates lists the number of
predicate evaluations.
I. An Evaluation of Contract Monitoring
Dynamic contract checking impacts the execution time. Source of
this impact is (1) that every contract extents the original source with
additional checks that need to checked when the program executes
and (2) that the contract monitor itself causes some runtime overhead.
To demonstrate, we consider runtime values obtained from the
TreatJS contract framework for JavaScript (Keil and Thiemann
2015b). To the best of our knowledge, TreatJS is the contract
framework with the heaviest runtime deterioration. Reason for this
is its flexibility and its support for full intersection and union
contracts (Keil and Thiemann 2015a). The presence of intersection
and union contracts require that a failing contract must not signal
a violation immediately. As a violation depends on combinations
of failures in different sub-contracts the contract monitor muss
continue even if a first error occurs and connect the outcome of
each sub-contract with the enclosing operations.
To evaluate the runtime deterioration, they applied there framework to benchmark programs from the Google Octane 2.0 Benchmark Suite4 . The benchmark programs measure a JavaScript engine’s performance by running a selection of complex and demanding programs.
For testing, they inferred a function contract for every function
expression contained a program and assert the function contract
when the program executes.
Figure 30 contains the runtime values of all benchmark programs in two configurations, which are explained in the figure’s
caption. As expected, all run times increase when adding contracts.
The examples show that the run-time impact of contract monitor depends on the program and on the particular value that is
monitored. While some programs like Richards, DeltaBlue, RayTrace, and Splay are heavily affected, others are almost unaffected:
Crypto, NavierStokes, and Mandreel, for instance.
It follows that the impact of a contract depends on the frequency
of its application. A contract on a heavily used function (e.g.,
in Richards, DeltaBlue, or Splay) causes a significantly higher
deterioration of the run-time.
For better understanding, Figure 31 lists some numbers of internal counters. The numbers indicate that the heavily affected bench4 https://developers.google.com/octane/
marks (Richards, DeltaBlue, RayTrace, Splay) contain a very large
number of predicate checks. Other benchmarks are either not affected (RegExp, zlib) or only slightly affected (Crypto, pdf.js, Mandreel) by contracts.
For example, the Richards benchmark performs 24 top-level
contract assertions (this are all unique contracts in a source program), 1.6 billion internal contract assertions (including top-level
assertions, delayed contract checking, and predicate evaluation),
and 936 million predicate evaluation.
To sum up, we see that number of predicate checks is substantially responsible for the runtime deterioration of a contract system. Thus, reducing the number of predicate checks will entirely
improve the runtime.
J. Practical Evaluation, continued
To demonstrate the run time improvement we applied our contract
simplification to a number simple examples, as introduces in Section 2 and Section 7. The testing procedure is a simple while loop
that uses different versions of an addOne function to increase a
counter on every iteration.
This section shows the examples programs written in JavaScript
and the full table of results. To run the examples we use the
TreatJS5 (Keil and Thiemann 2015b) contract system for JavaScript
and the Mozilla SpiderMonkey6 engine. While there is actually no
implementation of our transformation system for JavaScript, we
applied all transformation manually and produces the simplified
contracts by hand.
In TreatJS, function assert asserts a contrat (given as second
argument) to a subject value (given as first argument). Constructor AFunction creates a function contract from a set of argument
contracts (given as first argument) and a return contract (given as
second argument). Constructor Intersection creates an intersection
contract from two other contracts. Number, Natural, Positive,
Negative, String are flat contrasts checking for number values,
natural numbers, positive numbers, negative numbers, and string
values, respectively.
J.1 The example programs
Example 1 (addOne1). In a first example we add a simple function
contract to plus restricting its domain and range to number values.
Every use of addOne1 triggers three predicates checks.
1
2
3
4
5
6
7
8
9
var addOne1 = (function () {
var plus = assert(function (x, y) {
return x + y;
}, AFunction([ Number, Number], Number))
var addOne = function (x) {
return plus (x, 1);
}
return addOne;
})();
22
23
24
25
26
27
28
29
30
Example 4 (addOne4). The next example merges the intersection
contract on plus (from Example 2) and the contract on addOne
(Example 3). Every call of addOne4 leads to a total number 8
predicate checks.
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
48
49
50
51
52
53
54
10
11
12
13
14
15
16
17
18
19
20
21
var addOne2 = (function () {
var plus = assert(function (x, y) {
return x + y;
}, Intersection(
AFunction([ Number, Number], Number),
AFunction([ String, String], String)
));
var addOne = function (x) {
return plus (x, 1);
}
return addOne;
})();
var addOne4 = (function () {
var plus = assert(function (x, y) {
return x + y;
}, Intersection(
AFunction([ Number, Number], Number),
AFunction([ String, String], String)
));
var addOne = assert(function (x) {
return plus (x, 1);
}, AFunction([ Natural], Positive));
return addOne;
})();
Example 5 (addOne5). In this example we also overload addOne
by making either string concatenation or addition, depending on
addOne’s input. While adding an intersection contract to addOne,
every use of addOne5 leads to 10 predicate checks.
47
Example 2 (addOne2). Our second example adds an intersection
contract to plus. As the native + operator is overload and works
for strings and numbers, our contract restricts the domain either to
string or number values and ensures that the function has to return
a string or a number value corresponding to the input. Every use
of addOne2 triggers six predicate checks.
var addOne3 = (function () {
var plus = assert(function (x, y) {
return x + y;
}, AFunction([ Number, Number], Number));
var addOne = assert(function (x) {
return plus (x, 1);
}, AFunction([ Natural], Positive));
return addOne;
})();
55
56
57
var addOne5 = (function () {
var plus = assert(function (x, y) {
return x + y;
}, Intersection(
AFunction([ Number, Number], Number),
AFunction([ String, String], String)
));
var addOne = assert(function (x) {
return (typeof x == ’string’) ? plus (x, ’1’) : plus (x, 1);
}, Intersection(
AFunction([ Natural], Positive),
AFunction([ String], String)
);
return addOne;
})();
Example 6 (addOne6). Our last example simulates the case that
we add fine-grained properties to plus and addOne. In this case
we state different properties in different function contracts and
use intersections to combine those properties. Before simplifying
the contract, every call of addOne6 leads to a total number 17
predicate checks.
var addOne6 = (function () {
var plus = assert(function (x, y) {
Example 3 (addOne3). The third example extends addOne1 by
60
return x + y;
adding another function contract to addOne. The contract restricts
61
}, Intersection(
addOne’s domain to natural numbers and requires a positive num62
Intersection(
ber as return. In combination, every use of addOne3 triggers 5
63
AFunction([ Number, Number], Number),
predicate checks.
64
AFunction([ String, String], String)),
65
Intersection(
5 https://github.com/keil/TreatJS
66
Intersection(
6 https://developer.mozilla.org/en-US/docs/Mozilla/Projects/SpiderMonkey
67
AFunction([ Natural, Positive], Positive),
58
59
68
69
70
71
72
73
74
AFunction([ Positive, Natural], Positive)),
AFunction([ Negative, Negative], Negative))));
var addOne = assert(function (x) {
return plus (x, 1);
}, AFunction([ Natural], Positive));
return addOne;
})();
J.2 The Run-Time Improvements
Figure 32 finally contains the execution time and the number of
predicate checks during execution of all examples programs in
different configuration, as explained in the figures caption.
In addition to the examples from Section J.1, it also contains
the run time of the addOne program without contracts, and the run
time of the addOne example where we only wrap the functions in
a proxy, but did not apply any contracts.
To sum up, the the Baseline Simplification improves the run time
by approximately 28%, whereas the Subset Simplification makes an
improvement by approximately 58.68%. Clearly, the improvement
strictly depends on the granularly of predicates and contracts.
The numbers also indicate that the run time improvements remain identical, whether we use an optimizing compiler or not. But,
deactivating the optimizing compiler increases the run time of the
program without contracts by factor 4.83 (No-Ion) and by factor
20.17 (No-Jit), whereas the version with contracts increases only
by factor 1.5 (No-Ion) and factor 2.09 (No-Jit).
The numbers also shows that adding contracts, even without any
functionality, increases the run time by factor 71.2 (Full), factor
13.7 (No-Ion), and factor 5 (No-Jit).
This illustrates that contracts prevent a program from being
optimized efficiently by an optimizing compiler.
Full
Benchmark
Normal
time (ms)
predicates
Without Contracts
Proxy only
2
142.33
0
0
Example 1
Example 3
Example 4
Example 4
Example 5
Example 6
39404
87143
66474
114468
148249
295682
300000
600000
500000
800000
1000000
1700000
Baseline
time (ms)
27081
58385
54396
85126
107956
200009
Subset
predicates
(- 31.27%)
(- 33.00%)
(- 18.17%)
(- 25.63%)
(- 27.18%)
(- 32.36%)
200000
400000
400000
600000
800000
1200000
(- 33.33%)
(- 33.33%)
(- 20.00%)
(- 25.00%)
(- 20.00%)
(- 29.41%)
time (ms)
27043
46085
26518
44633
59970
118579
predicates
(- 31.37%)
(- 47.12%)
(- 60.11%)
(- 61.01%)
(- 59.55%)
(- 59.90%)
200000
300000
200000
300000
500000
700000
(- 33.33%)
(- 50.00%)
(- 60.00%)
(- 62.50%)
(- 50.00%)
(- 58.82%)
No-Ion
Benchmark
Normal
time (ms)
predicates
Without Contracts
Proxy only
9.67
132.67
0
0
Example 1
Example 3
Example 4
Example 4
Example 5
Example 6
58906
131184
99776
173037
221619
441176
300000
600000
500000
800000
1000000
1700000
Baseline
time (ms)
40383
87544
81719
128754
160360
278601
Subset
predicates
(- 31.44%)
(- 33.26%)
(- 18.10%)
(- 25.59%)
(- 27.64%)
(- 36.85%)
200000
400000
400000
600000
800000
1200000
(- 33.33%)
(- 33.33%)
(- 20.00%)
(- 25.00%)
(- 20.00%)
(- 29.41%)
time (ms)
40271
69492
40218
66359
88369
163272
predicates
(- 31.64%)
(- 47.03%)
(- 59.69%)
(- 61.65%)
(- 60.13%)
(- 62.99%)
200000
300000
200000
300000
500000
700000
(- 33.33%)
(- 50.00%)
(- 60.00%)
(- 62.50%)
(- 50.00%)
(- 58.82%)
No-Jit
Benchmark
Normal
time (ms)
predicates
Without Contracts
Proxy only
40.33
200.67
0
0
Example 1
Example 3
Example 4
Example 4
Example 5
Example 6
81125
186069
136728
240724
315184
597406
300000
600000
500000
800000
1000000
1700000
Baseline
time (ms)
56439
124434
111596
179451
225316
404276
(- 30.43%)
(- 33.12%)
(- 18.38%)
(- 25.45%)
(- 28.51%)
(- 32.33%)
Subset
predicates
200000
400000
400000
600000
800000
1200000
(- 33.33%)
(- 33.33%)
(- 20.00%)
(- 25.00%)
(- 20.00%)
(- 29.41%)
time (ms)
54945
96271
55186
91034
123852
233124
(- 32.27%)
(- 48.26%)
(- 59.64%)
(- 62.18%)
(- 60.71%)
(- 60.98%)
predicates
200000
300000
200000
300000
500000
700000
(- 33.33%)
(- 50.00%)
(- 60.00%)
(- 62.50%)
(- 50.00%)
(- 58.82%)
Figure 32. Results from running the TreatJS contract system. Table Full shows the results of a run with both JIT compilers, whereas table
No-Ion shows the result of a run without IonMonkey (but the Baseline JIT remains enabled) and table No-JIT shows the result of a run
without any JIT compilation. Column Normal gives the baseline execution time and the total number of predicate checks of the unmodified
program. Column Baseline and column Subset contain the execution time and the total number of predicate checks after applying the baseline
simplification or the subset simplification, respectively. The value in parentheses indicates the improvement (in percent).
| 6cs.PL
|
MPAR: A Movement Pattern-Aware Optimal Routing for Social Delay Tolerant
Networks
Lei You1 , Jianbo Li2∗ , Changjiang Wei3 Lejuan Hu4
Information Engineering College, Qingdao University, Qingdao 266071, Shandong Province, China
[email protected],[email protected],[email protected], [email protected]
arXiv:1602.08455v2 [cs.IT] 21 Jun 2017
Abstract
Social Delay Tolerant Networks (SDTNs) are a special kind of Delay Tolerant Network (DTN) that consists of a number
of mobile devices with social characteristics. The current research achievements on routing algorithms tend to separately
evaluate the available profit for each prospective relay node and cannot achieve the global optimal performance in an
overall perspective. In this paper, we propose a Movement Pattern-Aware optimal Routing (MPAR) for SDTNs, by
choosing the optimal relay node(s) set for each message, which eventually based on running a search algorithm on a
hyper-cube solution space. Concretely, the movement pattern of a group of node(s) can be extracted from the movement
records of nodes. Then the set of commonly visited locations for the relay node(s) set and the destination node is
obtained, by which we can further evaluate the co-delivery probability of the relay node(s) set. Both local search scheme
and tabu-search scheme are utilized in finding the optimal set, and the tabu-search based routing Tabu-MPAR is proved
able to guide the relay node(s) set in evolving to the optimal one. We demonstrate how the MPAR algorithm significantly
outperforms the previous ones through extensive simulations, based on the synthetic SDTN mobility model.
Keywords:
Delay Tolerant Networks, Mobile Social Networks, Movement, Opportunistic routing
1. Introduction
In traditional data networks such as Internet, there are
usually some assumptions of the network model e.g. the
existence of at least one end-to-end path between sourcedestination pair [1]. Any arbitrary link connecting two
nodes is assumed to be bidirectional supporting symmetric data rates with low error probability and latency [2].
Messages are buffered in intermediate nodes (e.g. routers)
and further forwarded to the next-hop relay or successfully
received by the destination. In this case, each message
is not expected to occupy the buffer of nodes for a long
period of time. However, these all above usually fail in
the context of Delay Tolerant Networks (DTNs)[3]. Some
applications, e.g., email service, address the delivery success while having relatively flexible requirement of latency,
which is known as “delay-tolerant”. For further popularizing these kind of applications, we have to reconsider the
widely used network architecture so as to relax the assumption of the continuous end-to-end connectivity that
TCP/IP based [4].
Recently years there have seen the wide adoptions of
mobile devices such as smartphones, laptops and tablet
PCs. The mobile devices may form a network in an ad
hoc manner, to work as an auxiliary to cellular networks
for some services such as social information sharing. Many
researchers use the term “Social Delay Tolerant Networks
(SDTNs)” to describe this special kind of DTN, where
Preprint submitted to Elsevier Ad Hoc Networks Journal
mobile users move around and communicate with each
other via their carried short-distance wireless communication devices [5, 6, 7]. Since SDTNs experience intermittent connectivity incurred by the mobility of users, routing
is still the most challenging problem [8] for such networks.
The fundamental characteristics of STDN is that there exist potential social relationships behind the nodes in the
network, which in turn affects the movement pattern of
nodes. Many previous works have explored how to leverage
the social relationships so as to enhance the routing performance. For example, social-aware routing algorithms
based on social network analysis have been proposed, such
as Bubble Rap [9], SimBet [10]. Besides, algorithms that
based on some custom defined social metrics are proposed
in [11, 12, 13, 14, 15, 16].
In many real SDTNs, mobile users that have a common interest generally will visit some locations that are
related to this interest. Previous works observed that 50%
of mobile users in this network spent 74.0% of their time
at a single access point (AP) to show the characteristic
of frequently visiting a few locations [17]. This indicates
the feasibility to “connect” some nodes through their commonly visited locations where some cache devices are deployed for buffering the message, e.g. throwbox proposed
in [18]. Cache devices have several advantages over ordinary nodes. First, there would usually be some stable connections between nodes and the cache device [17], which
February 7, 2018
hardly exist between two ordinary nodes due to their mobility. Second, they are fixed as a kind of network infrastructures, and their capacity can be much more sufficient
than the ordinary nodes. From the above discussion, intuitively we can deploy cache devices in frequently visited
locations to improve the routing performance in SDTNs.
In this paper, we focus on the multi-copy unicast routing
problem in SDTNs, where there is one source-destination
nodes pair for each packet, and there can be multiple replicas of the packet in the network. Different from previous
works that based on classical social network analysis or
custom defined social metrics, our scheme makes use of
the movement records collected by nodes. The basic idea
is to view the set of nodes holding a packet or its replicas as an entirety, thus extracting its frequently visited
locations from the movement records. The frequently visited locations of the destination node is obtained in the
same way. Then we get the intersection between the two
locations sets and then derive the co-delivery probability
for the whole relay node set. Our goal is to calculate the
optimal nodes set that maximizes the co-delivery probability. In addition, since that each generated message is
only valid before reaching its deadline, we take the message time-to-live value into consideration when extracting
the movement pattern from the movement records. Under this model, we propose the Movement Pattern-Aware
optimal Routing (MPAR) for SDTNs. We adopt the optimal opportunistic routing scheme by maintaining an optimal node set for the destination node. To the best of our
knowledge, this is the first work to exploit the social-aware
routing algorithm that leverages the movement pattern of
a group of nodes. Our main contributions are summarized
as follows:
the proactive routing strategy. Tabu-MPAR is proved
able to guide the relay node(s) set in evolving to the
optimal one.
The rest of this paper is organized as follows. In section
2 we introduce the system model. In section 3 we analyze some key properties behind the routing, and model
the routing as an optimal search problem. We discuss
the computational hardness of this problem and propose
a heuristic method to approximately obtain the optimal
solution in section 4. The details about the routing algorithm are given in section 5. Section 6 analyses the simulation result. We conclude the paper in Section 7. The
proofs of Theorem 1 and 2 are presented in the Appendix.
2. Preliminaries
In this section, we give the system model and the assumptions behind our scheme.
2.1. Network Model
We consider a SDTN composed of n nodes N = {ni |ni ∈
N, 1 6 i 6 n} moving among m locations A = {aj |aj ∈
N, 1 6 j 6 m}. Each mobile node frequently visits some locations A ⊆ A. This model is derived from the case of real
mobile networks. A typical example is the Wi-Fi campus
network at Dartmouth College. Another SDTN that follows this characteristic is the Vehicular Ad-hoc NETworks
(VANETs), where lots of buses and taxies move among bus
stations and taxi stops. Like in the previous work [11, 19],
we assume that the behavior of each node ni visiting any
location aj ∈ A(ni ) follows the Poisson process. In other
words, the time interval that each node visits a location
follows an exponential distribution. Besides, each location
in our model is assumed to have a throwbox[18] to store
and transmit messages. And we assume that the capacity of the throwbox is sufficient enough to take custody of
messages.
We assume that each message has a time-to-live value
τl , indicating the remaining life time of the message. With
the time passing, τl gradually decreases, and when τl is
zero, the message is deleted from nodes’ buffer. This timeto-live field prevents a message from staying in the network for a very long time, and allows the application to
assign a deadline for the message. For applications such
as news dissemination or advertising publishing, the message is only valid before its deadline. When the message
reaches the deadline, there is no need to continue buffering
the message.
1. We present a periodical time-aware movement record
model and extract the movement pattern from the
movement record of nodes. Each node set is viewed
as an entirety during the whole routing process. Corresponding Movement pattern can be extracted from
the same set in different time intervals. The frequently visited locations are directly corresponding
to the movement pattern for each node(s) set.
2. Under our network model that cache devices are deployed in several positions, we analyze two key performance metrics, based on which the routing problem
is formally proved to be an NP − Hard combinatorial
optimization problem.
3. Two search algorithms are proposed to solve the optimization problem, which are respectively based on the
local search scheme and the tabu search scheme. The
key elements of the tabu search scheme are specifically
defined for our optimal set search problem.
4. Two respective movement pattern-aware routing
schemes are designed based on the local search algorithm and the tabu search algorithm, which is called
Local-MPAR and Tabu-MPAR and are respectively
corresponding to the reactive routing strategy and
2.2. Basic Definitions
As a large number of mobile devices are carried by
human-beings or relative to the society of people, we assume that the movement patterns of nodes are of periodicity. For example, Smith goes to work on Monday to Friday,
so that his frequently visited locations might be composed
of home, office, some bus-stations and a few nosheries. On
2
ni on the location aj by averaging rik,j for all k ∈ [1, h] in
column j, formally we have
,
k=h
X
i
rk,j h
(1)
Mi,j =
t0
a tim
th
a tim
e slo
t
t1
k=1
e int
erva
l
……
3. Overview of the Routing Problem
t2
Before discussing the details of our routing scheme, let
us overview the routing problem. We first discuss how
to extract the movement pattern of a set of node(s) in
Section 3.1. Then we analyze some key properties behind
the routing in Section 3.2. Finally, the routing problem is
formally defined as an optimal search problem in section
3.3.
a time point
…
t3
…
……
3.1. Movement Pattern
The movement pattern is extracted from the records of
node(s). We first define the function E that transforms a
movement record R to the corresponding movement pattern.
Figure 1: The period time T is divided into h time slots.
the weekends, he usually goes to a club or some coffee bars,
or other places different from those visited on workdays.
However, his actions usually repeat per week, i.e., there
exists some periodicity for his movement records. From
this point, we assume the period time is T . Normally, the
length of T will be several days. Then we divide T into h
T
. We represent
slots. The length of each slot is thereby h
the time points sequence as < t0 , t1 , t2 , . . . , th >, then any
two time points ts and te (ts < te ) form a time interval
[ts , te ], as shown in Figure 1. Thus we can extract the
movement pattern of a node for the time interval [ts , te ],
by referring to the movement records of nodes. The definition of the record of a node is shown as follows.
Definition 2. Function
E([x1 , x2 , . . . , xm ]) = [κ1 , κ2 , . . . , κm ]
where
κj =
1
0
Pi=m
δ
xj > m
i=1 xi
otherwise
(2)
where xi is the ith element of the vector processed by
and 0 < δ < 1 is a system parameter.
E
The basic utility of function E is to filter out the rarely
visited locations so as to leave the frequently visited ones
shown up. Following the above definition directly, we can
define the movement pattern.
Definition 1. Movement Record.
The movement record of a node ni is a h × m matrix R(i)
m locations
}|
{
z
1 i
/r1,1 1/ri1,2 . . . 1/ri1,m
1 i
/r2,1 1/ri2,2 . . . 1/ri2,m
R(i) = .
..
..
..
h time slots
.
.
.
..
1 i
i
i
1
1
/
r
/
r
.
.
.
/
r
h,1
h,2
h,m
Definition 3. Movement Pattern. A movement pattern
for a given node set N at the time interval [tp , tq ] is represented as P(V, [ts , te ]), where we have:
P(N, [ts , te ]) = E(
i=e
X X
R(x, ti ))
(3)
nx ∈N i=s
of which the kth row R(i, k) is the movement record during the time slot [tk−1 , tk ], and we have
Notice that in the above definition, the movement pattern P is extracted from the accumulation of several
records R of all nodes in set N with a specific constraint
on the time interval [ts , te ], by function E.
For example, assume that we have a network case, that
there are four nodes and two locations in total and n4 is
the destination node, and T is divided into two time slots
t1 and t2 , as shown in Figure 2.
The weight of each edge has shown in Table 1, which
stands for the average visit time interval between a given
”node-location” pair. Actually, these records are from the
real experiment of four students’ visiting interval (hours)
to the two Computer Labs in Qingdao University. We had
R(i, k) = [1/rik,1 , 1/rik,2 , . . . , 1/rik,m ]
where rik,j is the node ni ’s time interval on average to visit
the location aj , from the start to the end of the time slot
tk .
From the above definition, we know that 1/rik,j represents
ni ’s average meeting frequency on the location aj . Specifically, when ni has never arrived location aj during the
whole kth time slot, we have rik,j = ∞ and thus 1/rik,1 = 0.
We can also get the average meeting time interval Mi,j of
E.
3
Movement Records in time slot
t0
we get
[t0 , t1]
a1
i=m
0.95
δ X
× (1.414 + 1.974) ≈ 1.609
vi =
m
2
a2
i=1
t1
n1
n2
a1
t2
which is more than 1.414 and is less than 1.974, so that
we have
P({n1 , n2 , n3 }, [t0 , t2 ]) = [0, 1]
n4
n3
The movement pattern P indicates that the frequently visited location(s) set for the node set {n1 , n2 , n3 } during
the time interval [t0 , t2 ] is A({n1 , n2 , n3 }, [t0 , t2 ]) = {a2 }.
Note that there are totally 2(n−1) − 1 non-empty subsets
of N − nd , and each can be viewed as a entirety with some
internal cooperation to delivery the message to nd . We
extract all the movement patterns for all non-empty subsets of {n1 , n2 , n3 } and that of the node n4 . The results
are compiled as follows.
a2
Movement Records in time slot
[t1 , t2]
Figure 2: An example of the network instance
[t0 , t1 ]
[t1 , t2 ]
a1
2.08
4.4
8.05
2.6
6.02
4.0
6.05
3.5
n1
n2
n3
n4
n1
n2
n3
n4
a2
4.65
4.3
1.11
3.3
2.95
4.5
15.49
3.5
P({n1 }, [t0 , t2 ]) = [1, 0]
P({n2 }, [t0 , t2 ]) = [1, 1]
P({n3 }, [t0 , t2 ]) = [0, 1]
P({n4 }, [t0 , t2 ]) = [1, 1]
Respectively, the corresponding sets of locations are
listed as following.
Table 1: The weight of each edge in Figure 2
A({n1 }, [t0 , t2 ]) = {a1 }
A({n2 }, [t0 , t2 ]) = {a1 , a2 }
A({n3 }, [t0 , t2 ]) = {a2 }
A({n4 }, [t0 , t2 ]) = {a2 }
collected the data for about two weeks. Only the active 12
hours per day for students are counted. And we combine
the 2 weeks datas as a one week data set such that the time
period T is set to 14 × 12 = 168 hours. We set [t0 , t1 ] and
[t1 , t2 ] to be the first and the last 84 hours, respectively.
The record matrix for each student is as follow.
R(1) =
R(3) =
1/2.08
1/4.65
1/6.02
1/2.95
1/8.05
1/1.11
1/6.05
1/15.49
R(2) =
1/4.4
1/4.3
1/4.0
1/4.5
R(4) =
1/2.6
1/3.3
1/3.5
1/3.5
A({n1 , n2 }, [t0 , t2 ]) = {a1 }
A({n1 , n3 }, [t0 , t2 ]) = {a2 }
A({n2 , n3 }, [t0 , t2 ]) = {a2 }
A({n1 , n2 , n3 }, [t0 , t2 ]) = {a1 , a2 }
3.2. Analysis of Two Key Properties
3.2.1. Delivery Probability
Assume that the movement pattern of the node(s) set N
during [t0 , t0 + τl ] is P(N, [t0 , t0 + τl ]) and the movement
pattern of the destination node nd is P(nd , [t0 , t0 + τl ]),
then we denote the common visited location(s)
set by A,
T
where we have A = A(N, [t0 , t0 + τl ]) A((nd , [t0 , t0 +
τl ])). The basic idea is to view the relay node(s) set N
as an entirety to cooperatively deliver the message. And
the message can be delivered to the destination node nd
by the node set N if there exists an location aj , where
∃ni ∈ N and nd encounter. In addition, if we denote the
arrive time of ni and nd by the random variables Ti,j and
Td,j respectively, then they must satisfy Ti,j < Td,j < τl ,
i.e. ni comes to the location aj before nd and not exceeding the deadline τl . For convenience, we let the current
time point to be t0 = 0. Besides, as stated in section 2,
the time interval that each node visits an location follows
an exponential distribution. From equation 1, we can calculate the exponential distribution parameter of that node
ni visiting location aj by
The movement pattern of the node set {n1 , n2 , n3 } can
be extracted from the movement records by equation 3.
The accumulation of all the records is calculated as follows:
j=2
i=3 X
X
P({n1 , n2 }, [t0 , t2 ]) = [1, 0]
P({n1 , n3 }, [t0 , t2 ]) = [0, 1]
P({n2 , n3 }, [t0 , t2 ]) = [0, 1]
P({n1 , n2 , n3 }, [t0 , t2 ]) = [0, 1]
R(i, j) =
i=1 j=1
1
1
1
1
1
1
+
+
+
+
+
,
2.08 4.4 8.05 6.02 4.0 6.05
1
1
1
1
1
1
+
+
+
+
+
4.65 4.3 1.11 2.95 4.5 15.49
= [1.414, 1.974]
λi,j = 1/Mi,j
For the case that A is empty set, we set the related delivery probability as zero, and the expected delay to reach
We set δ = 0.95 in equation 2, and according to definition 2
4
the node set N as follows:
other location as ∞. For the case that A is non-empty
set, we can estimate the direct delivery probability of ni
to the destination node nd by equation 4. Then the message delivery probability for the set N can be estimated
by equation 5.
Y
Pi,d = 1 −
P{n1 ,n2 ,n3 },4 =
λ4,2
λ4,2
λ4,2
1−
λ1,2 + λ4,2
λ2,2 + λ4,2
λ3,2 + λ4,2
M2,2
M3,2
M1,2
=1−
M1,2 + M4,2
M2,2 + M4,2
M3,2 + M4,2
(1 − P(Ti,j < Td,j < τl ))
= 0.789
aj ∈A
Y
=1−
1−
Y
Z τZl τl
1−
Y
1−
Z τZl τl
Y
=1−
λd,j λi,j e−(λd,j td,j +λi,j ti,j ) dtd,j dti,j
λi,j 1 − e−τl (λd,j +λi,j )
λd,j + λi,j
1−
aj ∈A
Y
3.2.2. Expected Delay
Besides the delivery probability, another property to explore is the node’s expected delay before moving to other
frequently visited locations. Assume that the frequent visited location(s) set of ni is Ai , then we let the random
variable Di denote the minimum expected delay for ni
to arrive at another location aj ∈ Ai . Formally we have
Di = min{Ti,1 , Ti,2 , . . . , Ti,|Ai | }, and consequently we get
the probability density function of Di as
X
Y
fDi (t) =
λi,k
e−λi,h t
!
0 ti,j
aj ∈A
PN,d = 1 −
!
fTi,j (ti,j )fTd,j (td,j )dtd,j dti,j
0 ti,j
aj ∈A
= 1−
fTi,j ,Td,j (ti,j , td,j ) dtd,j dti,j
0 ti,j
aj ∈A
=1−
Z τZl τl
!
− eτl λi,j − 1 e−τl (λd,j +λi,j )
(4)
ak ∈Ai
(1 − Pi,d )
Thus we have
ni ∈N
=1−
Y Y
−τl (λd,j +λi,j )
λi,j 1 − e
λd,j + λi,j
1−
ni ∈N aj ∈A
− eτl λi,j − 1 e−τl (λd,j +λi,j )
E[Di ] =
=
aj ∈A
1−
−(λd,j td,j +λi,j ti,j )
λd,j λi,j e
dtd,j dti,j
0 ti,j
Y
1−
=1−
aj ∈A
λi,j
λd,j + λi,j
=
!
Y Y
1−
ni ∈N aj ∈A
λi,j
λd,j + λi,j
=1−
ni ∈N aj ∈A
λd,j
λd,j + λi,j
X
e−λi,h t dt
ah ∈Ai
λi,k
P
ak ∈Ai
λi,h
1
ak ∈Ai
2
λi,k
(8)
For the example shown in Figure 2, we can calculate
E[Di ] for node n1 , n2 and n3 respectively according to
equation 8.
Y Y
0
= P
(6)
Y
tλi,k
ak ∈Ai
and for PN,d we have
PN,d = 1 −
tfDi (t)dt
X Z∞
ak ∈Ai
Pi,d = 1−
Z ∞Z ∞
Z∞
−∞
(5)
Specially, when we set the time-to-live value for
each message to be infinity, i.e. τl = ∞, we have
Y
ah ∈Ai
E[D1 ] = 1/λ1,1 = M1,1 = 4.05
M2,1 M2,2
1
= M2,1
E[D2 ] = λ2,1 +λ
+M2,2 = 2.15
2,2
1
E[D3 ] = /λ3,2 = M3,2 = 8.3
(7)
For any pair of nodes ni and nj , if E[Di ] < E[Dj ], then
we can conclude that ni is expected to arrive at some
new locations sooner, which means more connectivity opportunities to other nodes. In the intermittent network
scenarios such as DTNs, when the multi-copy strategy is
employed, it is vital to quickly spread the message over the
network. Due to this reason, the expected delay E[D] is
a valuable metric for enhancing the routing performance.
In Local-MPAR, we use E[D] to evaluate the suitability of
a node to be the active node, which is the only node that
can change the state of the network for a certain message.
In Tabu-MPAR, E[D] is used as the utility to asymmetrically spraying a given number of replicas. In the example
We assume that the message time-to-live τl = ∞
and the destination node is n4 in the network instance shown in Figure 2.
The frequently visited
locations for N = {n1 , n2 , n3 } is A(N, [t0 , t3 ]) = {a2 }
and that for n4 is A(n4 , [t0 , t3 ]) = {a1 , a2 }, so
that we get the commonly visited location(s) set
A = {a2 }.
The average meeting interval time for
each node-location pair can be obtained by equation 1.
M1,1 = 4.05 M2,1 = 4.2 M3,1 = 7.05 M4,1 = 3.05
M1,2 = 3.8 M2,2 = 4.4 M3,2 = 8.3 M4,2 = 3.4
And then we can estimate the delivery probability for
5
The optimal set is the minimal node set that can achieve
the highest delivery probability by cooperatively relay the
message to the destination node, based on which we can
formally define the problem.
shown in Table 1, node n2 is the most suitable node to
work as the active node, since it is expected to take the
message to a new location more quickly than others.
3.3. Problem Formalization
Definition 5. Nopt Search Problem.
Let f(x) to be the objective function and g(x) the constraint
function, where x = {x1 , x2 , . . . , xn } is the decision variable. And we have
The routing objective in DTNs can be various. Basically, there are two primary metrics for routing performance evaluation, message delivery ratio and message
average delay. Though the end-to-end delay should be
viewed as a very important metric for optimization in some
applications emphasizing the timeliness of messages, the
routing performance is fundamentally based on an acceptable delivery ratio. From this point, a high delivery ratio
is the guarantee of a good routing performance for any scenarios of DTNs. Based on this, the goal of our routing is
to maximize the message delivery ratio for each generated
message. Specifically, the goal is to select a subset N ⊆ N
to make PN,d > PN ′ ,d hold for ∀N ′ ⊆ N, where nd is the
destination node of the message.
For the example in Figure 2, we can get the estimated
delivery probability of all non-empty subsets of {n1 , n2 , n3 }
as we do for it in section 3.2 by equation 7:
f(x) =
i=n
X
i=1
λi,j 1 − e−τl (λd,j +λi,j )
λd,j + λi,j
aj ∈A
− eτl λi,j − 1 e−τl (λd,j +λi,j )
xi log
Y
1−
g(x) =
i=n
X
xi
i=1
Thus the problem can be formalized as
min f(x)
s.t.∀x ′ , g(x ′ ) − g(x) > 0,
∀i, xi ∈ {0, 1}
P{n1 },4 = 0.430
P{n1 ,n2 },4 = 0.670
P{n2 },4 = 0.673
P{n2 ,n3 },4 = 0.600
P{n3 },4 = 0.291
P{n1 ,n3 },4 = 0.626
P{n1 ,n2 ,n3 } = 0.789
And when the optimal solution xopt is obtained, we have
The optimal node(s) set is {n1 , n2 , n3 } in this example.
Notice that the estimated probability is not in direct proportion to the number of nodes holding the message. The
delivery probability PN,d is focused on the node set N,
which is viewed as an entirety during the routing process.
It is possible to have that PN,d < PN ′ ,d where N ′ ⊂ N.
Actually, in the example shown in Figure 2, the delivery probability of the set {n2 , n3 } is lower than that of
the set {n2 }. If ignoring the constraint of buffer and energy resources, the routing strategy that broadcast each
message over the network achieves the best performance,
which means that a larger relay nodes may leads to a better (at least not worse) delivery performance. However,
this ideal situation hardly exist in real network scenarios.
So we tend to select the minimal node set that has a suitable movement pattern so as to maximize the estimated
co-delivery probability for this node set. We can see from
the example that if we treat the node set holding the same
message as an entirety, then the estimated delivery probability is influenced by both the scale of the selected set
and the movement pattern extracted from the records of
this node set. Before formally defining our problem, firstly
we give the definition of the optimal set Nopt .
Notice that in the above definition, there exists a oneto-one match between x and N. Actually we have f(x) =
log(1−PN,d ), where f(x) is negative correlation with PN,d .
In this way, the Nopt Search Problem is formalized as an
optimization problem.
Nopt = {ni |xi > 0, 1 6 i 6 n}
4. Analysis for Nopt Search Problem
In this section, we discuss the computational hardness of
this problem and propose a heuristic method to approximately obtain the optimal solution. We first show that
a newly added element in the input sequence can widely
change the value of the objective function. For the offline
algorithm, we put the analysis of the NP − Hardness. Finally we propose our algorithm based on the tabu search.
4.1. Computational Hardness
An online algorithm is one that can process its inputs
piece-by-piece, i.e. it processes the input data in the sequence that they are fed to the algorithm, instead of requiring the entire input data set at the beginning. The
input of our online algorithm is the movement records of
all nodes in the network. We simply represent the record
sequence as R1 , R2 , . . . , Rq (each element in the sequence is
a vector, e.g. Rj = R(x, ti ) for node nx in the time slot ti )
and we currently only know the 1—k part of the sequence.
Then what interests us is that how much the newly added
element Rk+1 can affect the solution, which is answered by
Theorem 1.
Definition 4. The optimal set Nopt
The optimal set Nopt is a set of nodes in the network,
where we have
Nopt = min{arg max PN,d }
N⊆N
6
Algorithm 1 Tabu Search Framework.
Require:
Theorem 1. For ∀k ∈ [1, z), even the 1—k part of the
sequence is known, the newly added element Rk+1 can still
change the movement pattern to any state with at least 1
non-zero element.
notation
p
N
S
E
Proof. See Appendix A.
Then we show that even all movement records are available for any node, there is not an efficient polynomial algorithm yet for solving the Nopt problem.
T
A
Step 1:
1: choose an initial feasible solution xnow in S
2: set the current best solution xbest ← xnow
3: set the tabu table T ← ∅
Step 2:
1: if E(xbest ) = true then
2:
return xbest
3: end if
4: get C by filtering N(xnow ) according to T and A
5: xnow ←arg max p(x)
Theorem 2. Even the global information is known in advance, i.e., the value of λi,j is available for any node ni ,
the Nopt Search Problem is still NP-Hard.
Proof. See Appendix B.
4.2. The Trap of Local Optimum
A direct approach is to employ a local search algorithm
to find the solution. However, this may lead us to the trap
of local optimum. For the example shown in Figure 2,
assume that node n2 generates a message for n4 , then
Table 2 shows the calculation process of a local search
algorithm.
x∈C
6:
7:
8:
9:
10:
Table 2: Local search process for the example in Figure 2
current solution
xnow
N
PN,4
Step 1
[0, 1, 0]
{n2 }
0.673
Stop
[0, 1, 0]
{n2 }
0.673
next options
xnext
N’
PN ′ ,4
[0, 1, 1] {n2 , n3 } 0.600
φ
0
[0, 0, 0]
[1, 1, 0] {n1 , n2 } 0.670
—
—
meaning
Evaluation function
Neighborhood of a solution
Solution space
Stop rules
ϕ Tabu element
tabu table
L
Tabu length
Aspiration criteria
if p(xnow ) < p(xbest ) then
xbest ← xnow
end if
update T
goto Step 2
the feasible solution step by step in the solution space,
meanwhile evaluating the value of function p and trying
to obtain the optimal one. Step 1 is the initial step of the
tabu search, where we choose an initial feasible solution in
the solution space S and represent it as xnow . Then we
set the current best solution xbest to be xnow . At this
moment, there is no tabu records in the tabu table T, as
shown in line 3 of step 1.
In step 2, if the stop rule E is satisfied, then the search
procedure ends and returns the current best solution xbest
as the final result. Otherwise, we choose a subset from
xnow ’s neighborhoods N(xnow ) by referring to the current tabu table T and the pre-defined aspiration criteria
A. Then the algorithm search in the subset C and choose
the solution textbfx with maximal value of function p. In
each iteration of step 2, we update the best record solution
xbest . Finally we update the corresponding tabu element
ϕ in the tabu table T using the pre-assigned tabu length
L(ϕ).
In the following context, we give the definitions of all
the notations listed in the table in algorithm 1.
—
In Table 2, only the node n2 holds the message at the
beginning, and the delivery probability of the set N = {n2 }
is 0.673, which is higher than any of the next possible
choice. So the algorithm mistakes xnow as the optimal
solution, while the real optimal one is [1, 1, 1] that has the
probability 0.789. In this case, we have fallen into the
trap of local optimum, and we cannot jump out of it by
the local search algorithm.
4.3. Heuristic Search in a Hypercube Space
Theorem 2 shows that there would not be any efficient
polynomial algorithm for solving the Nopt Search Problem
and section 4.2 shows that the local search algorithm may
lead us to the trap of local optimum, which drive us to
resort to some intelligent search schemes. In this paper,
we leverage the tabu search algorithm to deal with the
Nopt problem. First of all we show the framework of the
tabu search method in Algorithm 1.
The tabu search process can be mainly divided into two
steps, both of which require several basic rules and data
structures shown in the notation table in algorithm 1. The
formal definitions of these notations will come up in the
following context. The whole search process is to change
Definition 6. Evaluation function.
The evaluation function of the tabu search is represented
as p(x), where
p(x) = −f(x)
In definition 6, we set the evaluation function by taking
the opposite value of function f(x) in definition 5. Then
7
the goal of our tabu search is to obtain the maximum value
of the evaluation function. We give the definition about
the neighborhood of a feasible solution in the tabu search
in definition 7.
and
Definition 7. Neighborhood.
Assume that x = [x1 , x2 , . . . , xn ](∀i, xi ∈ 0, 1) is a feasible
solution, then the neighborhood of x in the tabu search is
defined as follows
n
X
′
N(x) = x’
xi − xj 6 1
(9)
In fact, this tabu element is a movement from the current vertex to any adjacent vertex in the solution space, as
shown in Figure ??, i.e. any movement from the current
blue vertex to a red vertex can be viewed as a tabu element. Then we can define the tabu length L(ϕ) for each
tabu element ϕ in definition 10.
Definition 7 indicates that there is only one different element in a neighborhood solution compared to the
original one. For example, if the current feasible solution x is [0, 1, 0, 1], then its neighborhood N(x) =
{[1, 1, 0, 1], [0, 0, 0, 1], [0, 1, 1, 1], [0, 1, 0, 0]}, where each of
the three vectors is a neighborhood solution for the current
feasible solution x with the underline element performed
“a bitwise NOT”.
The stop rule is one of the most important rules for
the tabu search, which determines when the tabu search
procedure stops. The stop rule for our tabu search scheme
is stated in definition 8.
L(ϕ) = ⌊T ⌋
x’ = (x1 , . . . , xi−1 , yi , xi+1 , . . . , xn ) (i ∈ [0, n], xi 6= yi )
Definition 10. Tabu length.
The tabu length is denoted by L(ϕ), and we have
i=1
where T is a random variable and T ∼ N(µ, σ2 ), and we
set
√
√n[1 + p(x’) − p(x)] p(x’) > p(x)
µ=
n
otherwise
In definition 10, the tabu length is set to be an rounded
down random variable obeying the normal distribution,
where the mean value µ guarantees that the variation of
solution leading to a higher evaluation function value statistically has a longer tabu length. The reason why we set
Definition 8. Stop rule.
the tabu length to be a random number instead of a conThe stop rule E is defined as following
stant one is to add some stochastic factors to our method,
best
true if p(x
) not improved after θ steps like what has been done in Sherwood Algorithm. The
E(xbest ) =
stochastic law can help us to jump out of some endless cyfalse otherwise
cle of solutions, or to deal with some deliberate inputs. We
establish the relationship between the mean value of T and
the variation of the evaluation function value |p(x)−p(x’)|.
As stated before, the value of evaluation function p is
There are two different cases for the variation of the evaluupdated in every iteration of the search process. If the
ation function value, p(x’) < p(x) that indicates the value
current xnow is corresponding to a better function value
might have reached a new “valley” and p(x’) > p(x) that
p(xnow ) > p(xbest ), then we update xbest to xnow .
indicates the value might have climbed on a higher mounHowever, if the best record value of the evaluation function
tain top. In the former case, we need a relatively large
p(xbest ) has not been improved after quite a few steps,
tabu length so as to jump out of the possible trap of local
then it might be difficult to get a better solution withoptimum. While in the later case, we should keep the tabu
out modifying the search algorithm itself. In this case, the
length smaller to avoid the value √
dropped
search procedure stops and returns the current best record
√ to a deep valley
again. Apparently, we have µ ∈ [ n, 2 n].
solution as the (approximate) optimal one.
From definition 9 and 10, the tabu table can be conseThe tabu table T is composed of two basic concepts, the
quently defined in definition 11.
tabu element ϕ and its tabu length L(ϕ). The tabu element is the object to tabu in the search algorithm, which
Definition 11. Tabu table.
can be a simple variation on the solution or the variation
The structure of the tabu table is as follows,
of the evaluation function value. The tabu element in our
tabu search scheme is defined in definition 9.
1
2 ··· ··· n − 1 n
T=
Definition 9. Tabu element.
t1 t2 · · · · · ·
tn−1 tn
A tabu element is a variation of the current solution vecAssume that the tabu element is ϕ : x− > x’, the update
tor, i.e. can be formally denoted by
rule is defined as
ϕ : x → x’
L(ϕ) if |xi − xi′ | = 1
0
if ti = 0
∀i, ti =
where we have
ti − 1 otherwise
x = (x1 , . . . , xi−1 , xi , xi+1 , . . . , xn ) (i ∈ [0, n])
8
The tabu table T is a 2 × n table, where the first row
stands for the variation position of the solution vector, and
the second row indicates the corresponding tabu length
for each variation position. For example, assume that the
variation of solution ϕ : [0, 1, 0] → [1, 1, 0] happens, then
we should update the current tabu table (t1 , t2 , t3 > 0) as
1
t1
2
t2
3
t3
−→
1
L(ϕ)
2
t2 − 1
and Tabu-MPAR, based on two search algorithms, the local search algorithm and the tabu search algorithm respectively. In Local-MPAR, the other node(s)’s λ value for any
one node is required when they encounter, while in TabuMPAR, we assume that each node knows the λ values of
all other node(s) in advance.
In this paper, we investigate the routing process for a
single message (it may have one or more than one replicas). To illustrate this further, we refer to the concept
in Epidemic routing protocol [20] that treats the message
as virus. Then a successful message replication operation
can be viewed as an infection process. The delivery operation is excluded in our discussion, i.e., when we say na
infects nb , nb is not the destination node. Start from this
point, we can classify all the nodes in the network into
three kinds.
3
t3 − 1
In each update operation, the corresponding variation position of the solution in this table is filled with the tabu
length L(ϕ), and all non-zero lengths are reduced by 1.
Any position in the solution vector with a non-zero length
is tabued, i.e. can only be modified after passing its tabu
length steps.
Another important rule for the tabu search algorithm is
the aspiration criterion that indicates the absolvable tabu
elements in the tabu table T. The aspiration for our tabu
search scheme is stated in definition 12
• Infected node. An infected node is a node holding the
message or any of its replicas. However an infected
node cannot infect other pure nodes.
Definition 12. Aspiration criterion.
The aspiration criterion is defined as following
• Pure node. A pure node is a node not holding the
message or any of its replicas. It can be infected by
any infectious node.
For ∀ solution x, if we have p(x) > p(xbest )
then the solution x can be chosen even if the
corresponding position is tabued in T.
• Infectious node. An infectious node is a special kind of
infected node that is able to infect other pure nodes.
An infectious state is able to move to an infected state,
which means that it still holds the message but unable to
add new ones to the network. A pure state can also be
transferred to an infected state or infectious state, which
means that it receives the message. We further explain
these states by referring to Epidemic [20] and SprayAndWait [21] protocols.
In Epidemic routing, there are only two kinds of nodes,
pure node and infectious node. The replicas distribution
process ends when all the pure nodes are infected and become infectious nodes. SprayAndWait routing employs all
the three kinds of nodes. In its source edition, there is only
one infectious node in the network, while in its binary edition the infectious nodes can be multiple but limited in a
fixed number. The replicas distribution process ends when
there is no infectious node in the network.
Actually, these three kinds of node exist in all kinds
of routing protocols in mobile delay tolerant networks.
Specifically, Epidemic and SprayAndWait are two most
representative zero-knowledge based routing protocols. All
other replicas distribution process in any routing protocol can more or less be viewed as the improvement of
these two basic schemes, by letting the node state transfers among pure, infected and infectious. Before introducing the Local-MPAR and Tabu-MPAR, we propose three
postulations about the routing.
When all the positions are tabued in T, the current solution vector can not move to any of its neighborhood.
However if there is a neighborhood solution x has the better evaluation function value than the current recorded
best one, we absolve the corresponding tabu element and
move the solution vector to x.
The tabu search process of the example in Figure 2 is
shown in Table 3. Like the local process illustrated in Table 1, the source node is n2 and thus the initial solution is
[0, 1, 0]. For simplification, the tabu length L constants at
3 rather than being randomly chosen. Besides, the variable
θ in definition 8 of stop rule is set to be 3.
In step 1, the tabu table T is empty, so all the next
options are choosable. According to line 5 in algorithm 1,
we choose the solution vector [1, 1, 0] that has the maximal
evaluation function value. In step 2, the first position of
the solution vector is tabued, so the next solution vector
can be chosen only between [0, 0, 0] and [1, 1, 1]. The search
procedure continues until it reaches step 5, where the best
recorded solution has not changed after θ = 3 steps, so the
tabu search procedure stops. Overview the whole process,
we can quickly go through the solution space and jump out
of the trap of local optimum, which cannot be achieved by
the local search algorithm in Table 2.
5. Movement Pattern-Aware Optimal Routing
Postulation 1. Any node na replicates the message to
nb only if nb currently does not hold any of its replicas.
In this section, we give the details about the routing process. We implement two routing protocols, Local-MPAR
9
Table 3: Tabu search process for the example in Figure 2
current solution
xnow
N
PN,4
current best solution
xbest
Nbest
PNbest ,4
Step 1
[0, 1, 0]
{n2 }
0.673
[0, 1, 0]
{n2 }
0.673
1
0
2
0
3
0
Step 2
[1, 1, 0]
{n1 , n2 }
0.670
[0, 1, 0]
{n2 }
0.673
1
3
2
0
3
0
Step 3
[1, 1, 1]
{n1 , n2 , n3 }
0.789
[1, 1, 1]
{n1 , n2 , n3 }
0.789
1
2
2
0
3
3
Step 4
[1, 0, 1]
{n1 , n3 }
0.626
[1, 1, 1]
{n1 , n2 , n3 }
0.789
1
1
2
3
3
2
Step 5
[1, 0, 1]
{n1 , n3 }
0.626
[1, 1, 1]
{n1 , n2 , n3 }
0.789
1
0
2
2
3
1
Stop
[1, 1, 1]
{n1 , n2 , n3 }
0.789
—
—
—
Postulation 2. The operation of cleaning messages out
of the buffer caused by time-to-live deadline or buffer size
constraint is called “drop”, while that caused by state transition is called “deletion”.
tabu table T
next options
N’
PN ′ ,4
{n1 , n2 }
0.670
φ
0
{n2 , n3 }
0.600
{n2 }
0.673
φ
0
{n1 , n2 , n3 } 0.789
{n2 , n3 }
0.600
{n1 , n3 }
0.626
{n1 , n2 }
0.670
{n3 }
0.291
{n1 , n2 , n3 } 0.789
{n1 }
0.430
{n3 }
0.291
{n1 , n2 , n3 } 0.789
{n1 }
0.430
xnext
[1, 1, 0]
[0, 0, 0]
[0, 1, 1]
[0, 1, 0]
[0, 0, 0]
[1, 1, 1]
[0, 1, 1]
[1, 0, 1]
[1, 1, 0]
[0, 0, 1]
[1, 1, 1]
[1, 0, 0]
[0, 0, 1]
[1, 1, 1]
[1, 0, 0]
—
—
—
discharge duplicator
Infectious
Infected
B
G
become duplicator
Postulation 3. The set of nodes holding the message or
any of its replicas is denoted by N, which is compose of all
infectious nodes and infected nodes.
status
choosable
choosable
choosable
tabu
choosable
choosable
tabu
choosable
tabu
tabu
tabu
tabu
choosable
tabu
tabu
—
—
delete message
Pure
W
get message
Figure 3: Node state transition in Local-MPAR.
Then we have the following theorem.
to be infectious. Correspondingly, N is initialized to be
composed of only ns . For the node set N, we have the
following postulation.
Theorem 3. The replication operation is corresponding
to the transition from W to G and deletion operation to
the transition from G to W.
Postulation 4. In Local-MPAR, the node set N is always
maintained by the active node.
Proof. From the definition of the three states, a node holding the message can be either in state B or G, and a node
not holding the message is accurately corresponding to
state W. Since that there is no direct transition between
B and W, a node in state B transfers to W has to pass the
state G, and vise versa. Then from postulation 1, the replication operation is corresponding to the transition from W
to B or W to T, i.e. W to G. From postulation 2, the deletion operation is corresponding to the transition from B
to W or G to W, i.e. G to W.
The whole routing process can be viewed as the dynamical adjustment of the node set N, as shown in the
second stage. The key operation of this stage is to let
each node finish the state transition according to the rules
shown inTable 4. For simplicity, in the following context
we name the table position in row B column G as BG, by
parity of reasoning. We can see from Table 4 that the state
transition happens only if either na or nb is in infectious
state, i.e. state B. There is nothing in BB because that it
cannot be true that both na and nb are in infectious state
in Local-MPAR.
5.1. Local Search Based Movement Pattern-Aware Routing
Lemma 1. The replication or deletion of message happens
only if one of the encountered two nodes is in infectious
state.
In Local-MPAR algorithm, each node could be of all
the three states. However, we only allow not more than
1 infectious node exists for each generated message. The
basic idea of Local-MPAR is to dynamically adjust the set
N that includes all nodes holding the message, so as to
maximize the co-delivery probability PN,d .
Algorithm 2 shows the Local-MPAR process. There are
three stages in Local-MPAR. In the initial stage, the source
node ns generates a message M, and then set its state
Proof. Directly obtained by Table 4.
Theorem 4. The routing stage of Local-MPAR is a local
search process in the solution space S.
Proof. From postulation 4, we know that the node set
N is updated only in infectious node. Lemma 1 directly
10
Algorithm 2 Local-MPAR Algorithm.
Initial Stage:
1: ns generates message M
2: ns .state ←B
3: N ← {ns }
Routing Stage:
1: for any pair of na and nb do
2:
if na and nb encounter then
3:
finish the state transition according to Table 4
4:
update N
5:
end if
6: end for
End Stage:
1: M is delivered
discharge duplicator
B
Infectious
become duplicator
get
get
message
message
Infected
G
delete
message
Pure
W
Figure 4: Node state transition in Tabu-MPAR.
Algorithm 3 Tabu-MPAR Algorithm.
Initial Stage:
1: ns generates message M
2: ns .state ←B
3: ns computes Nopt by tabu search and saves it in M
4: ns .tickets ← |Nopt |
Routing Stage:
1: for any pair of na and nb do
2:
if na and nb encounter then
b]
· |Nopt |
3:
a = E[DE[D
a ]+E[Db ]
shows that any happened replication or deletion operation
can be informed to the infectious node immediately, so
that N would be updated in time. Either a replication
or deletion operation just causes 1 replicas difference in
the network, thus only adding or removing one element in
the set N, which indeed transfers the correspondent solution vector to one of its neighborhoods N(N). We can
see from Table 4 that N would not vary if and only if
∀N ′ ∈ N(N), PN ′ ,d 6 PN,d , i.e., if and only if the solution
reaches the local optimum.
a]
b = E[DE[D
· |Nopt |
a ]+E[Db ]
5:
if a < 1 then
6:
na .tickets = ⌈a⌉
7:
nb .tickets = ⌊b⌋
8:
else
9:
na .tickets = ⌊a⌋
10:
nb .tickets = ⌈b⌉
11:
end if
12:
finish the state transition according to Table 5
13:
end if
14: end for
End Stage:
1: M is delivered
4:
There still remains one question. Since there is only one
infectious node in Local-MPAR, how should we choose the
infectious node? The basic principle is to choose the node
with the smallest expected delay to arrive at another frequently visited location, i.e. E[D] in equation 8. In Table 4
GB, when there is no need to change na from G to W, we
should consider that whether to change na to be the infectious node. If E[Da ] is less than E[Db ], then we deem
that na is more suitable to be the message duplicator, because it is expected to visit other location(s) sooner, which
means more transmission opportunities. Conversely, if the
duplicator responsibility moves to nb , i.e. G→B happens
in nb , we should symmetrically have B→G happens in na
to keep the uniqueness property of the infectious node, as
shown in Table 4 BG.
The routing process of the network instance of Figure 2
is shown in Figure 5. At time t1 , node n2 generates the
message of which the destination is n4 . From time t2 to t3 ,
since that the co-delivery probability of both {n1 , n2 } and
{n2 , n3 } is less than {n2 }. However, from the discussion
before, we know that the set Nopt = {n1 , n2 , n3 } could
reach the best delivery ratio. However in Local-MPAR
N is not able to evolve to Nopt due to the trap of local
optimum, as same as in Table 2.
Wait routing, we allow the infectious nodes to be multiple
but constrained in a finite number.
The Tabu-MPAR protocol is shown in algorithm 3.
There are three stages as same as in Local-MPAR. In the
initial stage, when the source node ns generates the message M, the state of ns is set to be B, as shown in line
1-2. The node set Nopt is calculated in the source node.
Besides, there is a ticket number for each generated message and is denoted by |Nopt |, which indicates that the
maximal number of replicas of this message is equal to the
size of Nopt . A node is in state B if and only if it has more
than one tickets for this message 1 . A node with only one
ticket for the message is of state G, and nodes not holding
this message are in state P.
5.2. Tabu Search Based Movement Pattern-Aware Routing
In Tabu-MPAR algorithm, each node also has the three
states as same as in Local-MPAR. Like that in SprayAnd-
1 To say strictly, this node is currently in B state for this message,
but for other message the state might be W or G
11
Table 4: Node state transition rules for Local-MPAR
state of nb
B
B
state of na
G
W
—
W if PN−{na },d > PN,d
B if PN−{na } 6 PN,d and E[Da ] < E[Db ]
G→
G else
G if PN S{nb },d > PN,d
W→
W else
G
G if G → B happens in nb
B else
B→
W
B
G
G
W
W
Table 5: Node state transition rules for Tabu-MPAR
state of nb
G
B
B
B
state of na
G
W
G → B|G
(depends on the
tickets left in na
after allocating to nb )
W → B|G
(depends on the
tickets left in na
after allocating to nb )
W
B → B|G
(depends on the
tickets left in na
after allocating to nb )
B → B|G
(depends on the tickets left in na
after allocating to nb )
G
W→
G if na ∈ Nopt and nb ∈
/ Nopt
W else
The second stage in algorithm 3 shows the routing process of Tabu-MPAR. There is two main differences compared to Local-MPAR. First, there is no need to update
the node set Nopt in other node, since that it had been
computed in the source node ns before and added to the
message head. Second, when any two nodes na and nb
encounter, we reallocate the tickets before start the state
transition process. The strategy of tickets distribution is
based on E[Da ] and E[Db ], as shown in line 3 and 4, which
ensure that the node with a smaller E[D] to be allocated
with more tickets. Line 5–10 guarantee that both a and
b are an integer between 1 and L − 1. Then we can define
the transition rule in Table 5. The state transition could
happen in all grids except for the diagonal. Besides, we
can see that the 3 × 3 table is a symmetrical matrix. Actually, when the transition in grid GB happens in na , the
transition in grid BG happens in nb , and vise versa. This
rule also applies for WB, BW and WG, GW. The possible
transitions among states are illustrated in Figure 4.
Now we prove that the routing stage in Tabu-MPAR
leads N to the optimal set Nopt .
G→
W
G
if na ∈
/ Nopt and nb ∈ Nopt
else
W
replicas only have 1 ticket. So the “only if” condition
works. The node set holding the message accords with
Nopt means that the number of nodes holding the message
is |Nopt |, which is the maximal number of replicas for this
message in the network. So the number of tickets in each
of these nodes is equal to 1. According to the definition
of node state in Tabu-MPAR, all of these nodes are in
infected state, and all other nodes are in pure state, which
indicates that currently there is no infectious node in the
network. So the “if” condition works.
Lemma 3. If the set N evolves to Nopt , then it would not
change any more.
Proof. From lemma 2, if the set N reaches Nopt , then all
the nodes in N are in infected (G) state, and there is no
non-pure node outside N. Since that there is no node in
B state, so the transition in grid GB could not happen
any more. In grid GW, the transition G → W happens
only if the current node is not belong to Nopt = N, which
conflicts the assumption. So all the states of all elements
in N no longer change.
Lemma 2. N = Nopt happens if and only if there is no
infectious nodes in the network, i.e. all nodes in N = Nopt
are in state G.
Theorem 5. The transition rule in Table 5 leads N to
Nopt .
Proof. Since that the number of all tickets in the network
is |Nopt |, if there exist an infectious node na , then its
tickets must be at least 2 and the tickets in all other nodes
are less than |Nopt | − 2. So the number of nodes holding
the message is at most |Nopt | − 1, and in this case na
has 2 tickets for the message, and all the other nodes with
Proof. The state transition between B and G does not
modify the set N. However, from line 4–10 in algorithm 3,
it is guaranteed that the distributed tickets are no less than
1, which means that any node in state B would transfer to
state G if there are enough encounter opportunities during the routing process. So we just need to focus on the
12
N={n2}
n4
N={n2}
N={n2}
n4
n4
n4
N={n 2 ,n}
4
Deliver !
n2
n1
n3
n1
n1
P {n2,n 3} < P {n2}
n2
n3
n 2 generates a message
n1
P {n1 ,n 2} < P {n2}
n3
n3
n2
n2
t1
(a) Time t1
t4
t3
t2
(b) Time t2
(c) Time t3
(d) Time t4
Figure 5: The routing process of Local-MPAR.
Nopt ={n1 ,n2 ,n3}
N={n2}
n4
Nopt ={n1 ,n2 ,n3}
Nopt ={n1 ,n2 ,n3}
N={n 2 ,n}
3
n4
N={n 1 ,nn
2
3}
n4
Nopt ={n1 ,n2 ,n3}
n1
n3
n 2’s tickets = 3
n 2 generates a message
n3
t1
n 2’s tickets = n
n 2’s tickets =
n 1 ’s tickets =
n3
n3
n 3’s tickets =
n 3’s tickets =
t3
t4
t2
(b) Time t2
Deliver !
n1
n2
n2
n2
(a) Time t1
n1
n 3’s tickets =
N={n 1 ,nn
3
4}
n2
n 1 ’s tickets =
n1
n4
(c) Time t3
(d) Time t4
Figure 6: The routing process of Tabu-MPAR.
transitions between W and G, as shown in grid WG and
GW, where the rule is to transfer the node belonging to
Nopt to G. Consequently our theorem comes directly from
lemma 2 and 3.
bus roads
home
office
Figure 6 shows the routing process of Tabu-MPAR, the
network situation is the same as Figure 5, where the source
n2 generates the message at time t1 . Since n2 computes
that Nopt = {n1 , n2 , n3 }, so the number of tickets for n2
is initialized to be |Nopt | = 3. At time t2 , when n2 and n3
encountered, we allocate the tickets by referring to E[D2 ]
and E[D3 ]. The node set N has evolved to Nopt at time
t3 , and there is no infectious nodes in the network, corresponding to lemma 2. Finally the routing process ends in
t4 due to the completion of the delivery.
meeting spot
Figure 7: Working day movement model in Manhattan blocks
modeling a random walk of a group along the streets. The
Manhattan community model confines the routing paths
of the mobile nodes to certain paths that reflect their real
moving pattern in addition to the colocation pattern. The
Manhattan model consists of grids in a matrix, in which
all nodes can only move on the sides of a grid, as shown
in Figure 7.
Concretely, the network is established as follows. We
set the number of pedestrians to be 200, 400, 600, 800.
The number of buses, offices and activity spots are correspondingly set to be 2%, 20%, 2.5% of the total number
of pedestrians. The working time for each simulation day
constants at 8 simulation hours for each person. Each
person’s probability to do evening activity is set to be 0.5.
Besides, we set the probability of owning a car for each person to be 0.2, so people might drive themselves to these
places instead of taking public vehicles. The walking speed
6. Simulation
In this section, we conduct extensive simulations to evaluate the performance of the MPAR algorithm by using the
ONE simulator [22]. We employ the Working Day Movement (WDM) model proposed in [23] and set the map
to be Manhattan blocks. This model incorporates some
sense of hierarchy and distinguishes between inter-building
and intra-building movements. Some sub-models such as
home, office, evening activities and different transports are
introduced in WDM, so as to capture the society characteristics of people. An office model reproduces a kind of
star-like trajectories around a desk of the person inside
the office building, while home model is just a sojourn in a
particular point of a home location. The evening activity
sub-model reflects a meeting with friends after work by
13
of pedestrians is set to be 0.8–1.4 m/s. As shown in Figure 7, there are three kinds of locations in the simulation,
i.e. home, office and meeting spot. There is a loop line
for buses in the center of the entire Manhattan square, so
the pedestrians can take a bus to reach their destinations.
We put a throw box in every home, office and meeting
spot. The capacity of each throwbox is set to be sufficient
enough to take custody of all received messages. All the
nodes communicate with the bluetooth interface, of which
the transmit diameter is 10m and the transmit speed is
250KB/s. Since that each spot (home, office and meeting
spot) is set to be a point in the map, if a node reach the
spot, it is ensured to encounter with the throwbox so as
to have its messages taken custody.
We compare the MPAR algorithm with the existing
social-aware algorithms: Delegation Forwarding (DF) [24]
and SimBet routing [10]. Four performance metrics commonly used are examined: delivery ratio, average latency
overhead ratio and average hop count. Simulation results
demonstrate that the MPAR algorithm can significantly
improve SDTNs routing performance.
nodes, the overloads for each node is released, so the overhead ratio becomes lower. Figure 11 shows the average
hop count performance. We can see that the two MPAR
algorithms have smaller average hop count than the other
two ones. With the number of nodes increasing, the average hop count increases. This is because the simulation
area is correspondingly magnified and it needs more nodes
to cooperate to delivery each message.
6.1.2. Varying Buffer Size
In the simulation of varying the buffer size, we set the
message time-to-live constant at 16 simulation hours. The
buffer size is set from 50 to 300 MB. The simulation results
are shown in Figure 12, 13, 14 and 15.
We can see from the results that the two MPAR algorithms significantly outperform DF and SimBet. Compared with DF, Tabu-MPAR and Local-MPAR increase
the delivery ratio by about 83.8% and 40.3%, and reduces
the average latency by about 83.8% and 57.8%, respectively. Compared with SimBet, Tabu-MPAR and LocalMPAR increase the delivery ratio by nearly 3 and 2.5
times, and reduces the average latency by about 80% and
60%. Regarding the results in Figure 14,Tabu-MPAR and
Local-MPAR have almost the same overhead ratio, which
is a little bit lower than SimBet and is much lower than
DF when the buffer size is smaller than 120M. When there
are only a few nodes in a relatively small area, the improvement of the overhead ratio of the social-aware routing
schemes is more obvious, as shown in Figure 14. The same
as the results in the simulation of varying TTL, we can
see from Figure 15 that the two MPAR algorithms have
smaller average hop count than the other two ones. With
the number of nodes increasing, the average hop count increases. The reason is as same as in section 6.3.1 that the
simulation area is correspondingly magnified and it needs
more nodes to cooperate to delivery each message.
6.1. Simulation in Working Day Movement Model
In the simulations, 30,000 messages are generated by
randomly selecting source and destination nodes among
all the pedestrians. The size of each message is between
500K and 1M. In simulations on evaluating the four metrics, we set |N|=200, 400, 600 and 800, respectively. The
simulation area is adjusted according to |N| to keep the
connectivity of the network. The entire simulation time is
set to be 12 days.
6.1.1. Varying Time-to-live
In the simulation of varying the time-to-live value, we set
the buffer size constant at 200M. The time-to-live value is
set from 10 to 30 hours. The simulation results are shown
in Figure 8, 9, 10 and 11.
These results show that the two MPAR algorithms significantly outperform DF and SimBet. Compared with
DF, Tabu-MPAR and Local-MPAR increase the delivery
ratio by about 71.1% and 37.8%, and reduces the average
latency by approximate 79.2% and 60.1%, respectively.
Compared with SimBet, Tabu-MPAR and Local-MPAR
increase the delivery ratio by about 95.2% and 55.3%, and
reduces the average latency by about 50% and 70%. Regarding the results in Figure 10, both Tabu-MPAR and
Local-MPAR have a comparably lower overhead ratio than
DF and SimBet. For the three social-aware routing algorithm Tabu-MPAR, Local-MPAR and SimBet, the overhead performance is much better than DF, especially when
the message TTL is set to be relatively large. When the
number of nodes is smaller, the improvement is a little
more apparent. Since that the MPAR routing algorithms
constrain the maximal number of replicas and utilize the
throwbox to deliver messages, the number of relay operations is very low. Moreover, along with the increasing of
7. Conclusion
In this paper, we proposed a movement pattern-aware
routing protocol MPAR for SDTNs. We present a periodical time-aware movement record model and extract the
movement pattern from the movement record of nodes,
and then each node set is viewed as an entirety during the
whole routing process. Two key properties for routing are
analyzed and consequently the routing problem is modeled to be a combinatorial optimal search problem. Two
search algorithms are proposed to solve the optimization
problem, which are respectively based on the local search
scheme and the tabu search scheme. Two respective movement pattern-aware routing schemes are designed based on
the local search algorithm and the tabu search algorithm,
which is called Local-MPAR and Tabu-MPAR and are respectively. In addition, we prove that the Tabu-MPAR can
guide the relay node(s) set in evolving to the optimal one.
We demonstrate how the MPAR algorithm significantly
14
delivery probability
à
à
à
æ
æ
à
à
æ
0.5
æ
à
0.4
à
0.3
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
à
æ
à
à
0.6
æ
0.7
delivery probability
æ
0.7
0.8
à
à
æ
æ
à
0.4
ò
ì
ò
ì
ò
ì
ò
ì
0.3
æ
à
æ
à
ò
ì
1.0
ò
ì
ò
ì
ò
ì
15
20
25
30
Tabu-MPAR
Local-MPAR
à
æ
à
0.6
æ
0.5æ
à
ò
ì
15
20
25
30
ò
ì
ò
ì
ò
ì
ò
ì
0.3
10
à
à
SimBet
ò
DF
Tabu-MPAR
æ
Local-MPAR
à
æ
0.7
15
0.5æ
à
25
30
ò
ì
ò
ì
ò
ì
ò
ì
10
15
SimBet
ò
DF
Tabu-MPAR
æ
Local-MPAR
à
ò
ì
ò
ì
ò
ì
ò
ì
20
message TTL HhourL
ì
à
ò
ì
ò
ì
à
à
0.4à
20
à
à
æ
ò
ì
10
æ
à
à
à
à
æ
0.6
0.3
message TTL HhourL
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
ò
ì
æ
0.8
à
à
à
æ
æ
à
à
à
æ
0.7
0.4à
message TTL HhourL
æ
æ
æ
0.9
0.8
ò
ì
1.0
æ
æ
à
ò
ì
æ
æ
æ
æ
æ
0.9
ò
ì
10
æ
à
à
æ
à
æ
à
æ
à
à
0.6
0.5
æ
æ
delivery probability
æ
delivery probability
æ
æ
æ
æ
æ
0.8
25
30
message TTL HhourL
ì
SimBet
ò
DF
Tabu-MPAR
æ
Local-MPAR
à
SimBet
ì
ò
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 8: Performance comparisons of delivery probability vs. time-to-live
30 000
ò
25 000
ò
15 000
ò
ì
ò
ì
10 000ò
ì
æ
10
ì
ì
ì
à
æ
à
æ
à
æ
5000à
ì
ì
ì
ò
20 000
ì
ò
ò
à
à
æ
æ
ò
ì
ò
ì
10 000ò
25
ì
30
à
æ
à
æ
à
æ
5000
20
10
Tabu-MPAR
Local-MPAR
à
ì
à
æ
à
à
æ
æ
15 000
ò
ò
ì
10 000
15
20
5000
25
30
ò
DF
Tabu-MPAR
æ
Local-MPAR
à
ì
ò
ì
ì
ì
10
à
æ
à
æ
à
æ
à
æ
à
æ
à
æ
à
æ
à
æ
à
æ
ò
ò
ì
à
æ
ì
ò
15 000
ò
20
25
ò
DF
Tabu-MPAR
æ
Local-MPAR
à
ì
ì
ì
ì
ò
10 000
ò
ì
ì
ì
à
æ
à
æ
ì
ò
ì
5000
30
à
æ
à
æ
10
à
æ
à
æ
15
à
æ
à
æ
à
æ
à
æ
20
message TTL HhourL
SimBet
ì
ì
ò
ò
à
æ
15
ò
ò
20 000
ì
ì
ì
message TTL HhourL
SimBet
ò
ì
ò
ò
ò
ì
ò
ì
æ
æ
æ
æ
à
à
à
à
ò
à
æ
message TTL HhourL
æ
ì
ì
ì
ì
ò
15 000
à
æ
15
20 000
ì
ò
ì
ò
20 000
ò
ò
ì
ò
25 000
ì
æ
æ
æ
æ
æ
à
à
à
à
à
25 000
ò
ò
ò
ò
average latency HsecondsL
35 000
average latency HsecondsL
average latency HsecondsL
ò
ì
ò
ò
30 000
average latency HsecondsL
35 000
25
30
message TTL HhourL
SimBet
ò
DF
Tabu-MPAR
æ
à
Local-MPAR
SimBet
ì
ò
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 9: Performance comparisons of average end-to-end latency vs. time-to-live
ò
ò
ò
ò
70
200
120
80
ò
ò
50
à
æ
à
0æ
10
ò
ì
ò
ì
ò
ì
ò
ì
à
æ
à
æ
15
ò
ò
ò
ò
ì
ì
à
æ
ì
à
æ
ì
à
æ
à
æ
80
ò
60
ì
à
æ
ì
à
æ
à
ì
æ
20ì
30
10
à
æ
à
æ
à
æ
à
æ
25
ò
ì
ò
ì
ò
ì
15
Tabu-MPAR
Local-MPAR
à
ò
ò
ò
ò
ì
ì
à
æ
ì
à
æ
ì
à
æ
à
æ
20
message TTL HhourL
æ
ò
40
40
ò
20
50
60
20ò
ì
æ
à
25
ò
ì
ì
à
æ
à
ì
æ
ì
à
æ
30
ò
ì
à
æ
à
æ
10
ò
ò
ò
ò
ò
ì
à
æ
ì
ì
à
æ
ì
à
æ
ì
à
æ
à
æ
15
20
message TTL HhourL
ì
overhead ratio
ò
100
ò
overhead ratio
overhead ratio
overhead ratio
ò
100
150
60
SimBet
ò
DF
æ
Tabu-MPAR
Local-MPAR
à
ò
40
30
20
ò
ò
ì
æ
à
à
ì
æ
ì
à
æ
25
10ì
ì
à
æ
à
æ
30
ò
ì
ò
ì
à
æ
10
à
æ
15
SimBet
ò
DF
æ
Tabu-MPAR
Local-MPAR
à
ò
ì
à
æ
ò
ò
ì
à
æ
ì
à
ì
æ
ì
à
æ
ì
à
æ
à
æ
20
message TTL HhourL
ì
ò
ì
à
æ
25
30
message TTL HhourL
ì
SimBet
ò
DF
æ
Tabu-MPAR
à
Local-MPAR
SimBet
ì
ò
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 10: Performance comparisons of overhead ratio vs. time-to-live
3.5
3.5
3.5
3.0
3.0
1.5
2.0
1.5
2.5
2.0
1.5
1.0
1.0
1.0
0.5
0.5
0.5
0.0
10
12
14
Tabu-MPAR
16
18
20
22
message TTL HhourL
Local-MPAR
24
26
SimBet
28
DF
30
0.0
3
3.0
2.5
10
12
14
Tabu-MPAR
16
18
20
22
message TTL HhourL
Local-MPAR
24
26
SimBet
28
0.0
30
DF
average hop count
2.0
average hop count
average hop count
average hop count
2.5
2
1
10
12
14
Tabu-MPAR
16
18
20
22
message TTL HhourL
Local-MPAR
24
26
SimBet
28
30
0
DF
10
12
14
Tabu-MPAR
16
18
20
22
message TTL HhourL
Local-MPAR
24
26
SimBet
28
30
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 11: Performance comparisons of average hop count vs. time-to-live
outperforms the previous ones through extensive simulations, based on the synthetic SDTN mobility model.
9. Acknowledgments
This research was supported in part by Natural Science Foundation of Shandong Province under Grant
No.ZR2013FQ022 and Foundation research project of
Qingdao Science and Technology Plan under Grant No.121-4-2-(14)-jch. Science and Technology Plan Project for
Colleges and Universities of Shandong Province under
Grant NO. J14LN85
8. Conflict of Interests
All authors do not have any possible conflict of interests.
15
æ
æ
æ
æ
æ
æ
à
à
à
à
à
à
à
à
à
æ
à
0.45
delivery probability
à
æ
0.55
æ
0.50
æ
0.60
0.40
æ
à
à
à
æ
æ
æ
à
à
à
à
à
æ
0.45
æ
æ
æ
æ
æ
0.65
0.6
æ
à
à
à
à
à
à
à
à
à
à
à
0.5
ì
ò
50
ò
ì
ò
ò
ì
ì
100
ò
ò
ò
ò
ì
ì
ì
150
ì
200
ò
ì
250
ò
ì
0.35
ì
ò
ì
50
ò
ì
ò
ò
ì
ì
100
ì
ò
ò
ì
ì
200
ò
ò
ì
ò
ì
250
0.4
ì
ì
ò
300
50
ò
ì
ò
ì
ò
ì
100
ì
SimBet
ò
DF
Tabu-MPAR
æ
æ
æ
æ
æ
æ
æ
à
à
à
à
à
à
à
à
ò
ì
ò
ì
ò
ì
ì
æ
0.60
æ
0.55
à
à
à
0.50
ò
ì
ì
ò
ò
ì
ì
200
ò
ì
ò
ì
250
0.40
ò
ì
ò
ì
ò
0.35ì
50
300
ò
ò
ì
ì
100
ì
SimBet
ò
DF
Tabu-MPAR
æ
ò
ì
200
ò
250
ò
ì
300
buffer size HMBL
Local-MPAR
à
ò
ì
150
buffer size HMBL
Local-MPAR
à
ò
150
buffer size HMBL
Local-MPAR
à
ò
150
buffer size HMBL
Tabu-MPAR
æ
0.45
ò
300
æ
æ
0.70
æ
à
à
æ
æ
0.7
0.50à
æ
æ
æ
æ
0.40
ò
ì
0.35
æ
æ
æ
delivery probability
0.55
delivery probability
æ
æ
delivery probability
æ
0.60
ì
SimBet
ò
DF
Tabu-MPAR
æ
Local-MPAR
à
ì
SimBet
ò
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 12: Performance comparisons of delivery probability vs. buffer size
ò
ò
ò
ò
ò
ò
ò
ò
16 000
ì
ì
ì
ì
ì
ì
ì
ì
ì
ì
ì
14 000
12 000
10 000
8000à
à
à
à
à
à
à
à
à
à
average latency HsecondsL
average latency HsecondsL
ò
ò
ò
ò
ò
ò
ò
14 000
ò
ì
ì
ì
ì
ì
ì
ì
ì
ì
ì
ì
14 000
12 000
10 000
ì
à
à
à
à
æ
æ
à
à
æ
æ
à
à
æ
æ
à
à
æ
æ
100
æ
æ
æ
æ
150
200
æ
æ
250
æ
50
æ
100
150
buffer size HMBL
Tabu-MPAR
à
Local-MPAR
200
æ
æ
à
æ
æ
250
SimBet
ò
DF
Tabu-MPAR
æ
à
Local-MPAR
ì
ì
ì
300
à
æ
50
à
à
æ
æ
100
à
æ
à
æ
à
æ
150
buffer size HMBL
ì
ì
ì
6000
à
4000
æ
300
ì
ò
ò
ì
à
æ
à
æ
200
à
æ
250
10 000
SimBet
ò
DF
Tabu-MPAR
æ
à
Local-MPAR
ò
ì
ì
ì
à
æ
à
æ
ò
ò
ì
ì
ì
ì
ì
8000
6000
4000
à
à
æ
à
æ
300
à
æ
50
à
æ
æ
100
à
æ
à
æ
à
æ
150
buffer size HMBL
ì
ò
ò
ì
ì
ì
ò
ò
ò
12 000
ò
8000
6000
æ
æ
50
ò
ò
ò
ì
ì
10 000ì
ò
ò
ò
ò
12 000
8000
à
ò
ò
ò
16 000
6000
æ
ò
ò
ò
18 000
average latency HsecondsL
ò
average latency HsecondsL
ò
ò
ò
18 000
à
æ
200
250
à
æ
300
buffer size HMBL
ì
SimBet
ò
DF
Tabu-MPAR
æ
à
Local-MPAR
ì
SimBet
ò
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 13: Performance comparisons of delivery probability vs. buffer size
250ò
400ò
ò
120
ò
ò
ò
150
100
ò
overhead ratio
200
100
ò
overhead ratio
ò
overhead ratio
overhead ratio
300
140ò
150
200
100
80
60
ò
ò
40
50
100
50
à
æ
ì
à
ì
æ
0
50
ì
à
æ
ò
ì
à
æ
100
ò
ì
à
æ
150
ò
ì
à
æ
ò
ì
à
æ
ò
ì
à
æ
200
ò
ì
à
æ
ò
ì
à
æ
250
ò
ì
à
æ
à
æ
ì
à
æ
ì
0
50
300
ì
à
æ
ò
ì
à
æ
100
buffer size HMBL
æ
Tabu-MPAR
à
Local-MPAR
ò
ì
à
æ
150
ò
ì
à
æ
ò
ì
à
æ
ò
ì
à
æ
200
ò
ì
à
æ
ò
ì
à
æ
à
æ
ì
ò
ì
à
æ
250
à
æ
ì
0
50
300
ì
à
æ
ò
ì
à
æ
100
buffer size HMBL
ì
SimBet
ò
DF
æ
Tabu-MPAR
à
Local-MPAR
ò
ì
à
æ
150
ò
ì
à
æ
ò
ì
à
æ
ò
ì
à
æ
200
ò
ì
à
æ
ò
ì
à
æ
250
20à
ò
ì
à
æ
à
æ
ì
æ
ì
0
50
300
ì
à
æ
ò
ì
à
æ
100
buffer size HMBL
ì
SimBet
ò
DF
æ
Tabu-MPAR
à
Local-MPAR
ò
ì
à
æ
150
ò
ì
à
æ
ò
ì
à
æ
ò
ì
à
æ
200
ò
ì
à
æ
ò
ì
à
æ
250
ò
ì
à
æ
300
buffer size HMBL
ì
SimBet
ò
DF
æ
Tabu-MPAR
à
Local-MPAR
ì
SimBet
ò
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 14: Performance comparisons of delivery probability vs. buffer size
4
4
3
3
4
1
0
2
1
50
75
100
125
Tabu-MPAR
150
175
200
buffer size HMBL
Local-MPAR
225
250
SimBet
275
DF
300
0
average hop count
2
average hop count
average hop count
average hop count
4
3
2
75
100
Tabu-MPAR
125
150
175
200
buffer size HMBL
Local-MPAR
225
250
SimBet
275
0
300
DF
2
1
1
50
3
50
75
100
Tabu-MPAR
125
150
175
200
buffer size HMBL
Local-MPAR
225
250
SimBet
275
300
0
DF
50
75
100
Tabu-MPAR
125
150
175
200
buffer size HMBL
Local-MPAR
225
250
SimBet
275
300
DF
(a) Number of nodes: |N| = 200 (b) Number of nodes: |N| = 400 (c) Number of nodes: |N| = 600 (d) Number of nodes: |N| = 800
Figure 15: Performance comparisons of delivery probability vs. buffer size
References
tions, technologies, architectures, and protocols for computer
communications - SIGCOMM ’03, ACM Press, New York, New
York, USA, 2003, pp. 27–34. doi:10.1145/863955.863960.
[4] J. Ott, 404 Not Found? A Quest For DTN Applications, in:
Proceedings of the third ACM international workshop on Mobile
Opportunistic Networks - MobiOpp ’12, ACM Press, New York,
New York, USA, 2012, p. 3. doi:10.1145/2159576.2159579.
[5] K. Wei, D. Zeng, S. Guo, K. Xu, On Social Delay-Tolerant
Networking: Aggregation, Tie Detection, and Routing, IEEE
Transactions on Parallel and Distributed Systems 25 (6) (2014)
1563–1573. doi:10.1109/TPDS.2013.264.
[1] Y. Cao, Z. Sun, Routing in Delay/Disruption Tolerant Networks: A Taxonomy, Survey and Challenges, IEEE Communications Surveys & Tutorials 15 (2) (2013) 654–677.
doi:10.1109/SURV.2012.042512.00053.
[2] N. Vastardis, Mobile Social Networks: Architectures, Social Properties, and Key Research Challenges, IEEE Communications Surveys & Tutorials 15 (3) (2013) 1355–1371.
doi:10.1109/SURV.2012.060912.00108.
[3] K. Fall, A Delay-Tolerant Network Architecture for Challenged
Internets, in: Proceedings of the 2003 conference on Applica-
16
doi:10.1145/1080139.1080143.
[6] G. J. Bigwood, T. N. H. Henderson, S. N. Bhatti, Social delay[22] A. Keränen, J. Ott, T. Kärkkäinen, The ONE simulator for
tolerant network routing, in: 2nd ICC Winter Workshop on
DTN protocol evaluation, in: Proceedings of the Second InterComplexity in Social Systems, 2009.
national ICST Conference on Simulation Tools and Techniques,
[7] A.-M. Nguyen, N. Cre, SOCIAL-DTN: Why Social NetworkICST, 2009, p. 55. doi:10.4108/ICST.SIMUTOOLS2009.5674.
ing Services is more fruitful to Mobile Delay-Tolerant Net[23] F. Ekman, A. Keränen, J. Karvo, J. Ott, Working day
works?, in: 2009 International Conference on Ultra Modmovement model, in: Proceeding of the 1st ACM SIGMOern Telecommunications & Workshops, IEEE, 2009, pp. 1–2.
BILE workshop on Mobility models - MobilityModels ’08,
doi:10.1109/ICUMT.2009.5345436.
ACM Press, New York, New York, USA, 2008, pp. 33–40.
[8] K. Wei, X. Liang, K. Xu, A Survey of Social-Aware Routing
doi:10.1145/1374688.1374695.
Protocols in Delay Tolerant Networks: Applications, Taxonomy
[24] V. Erramilli, A. Chaintreau, M. Crovella, C. Diot, Delegation
and Design-Related Issues, IEEE Communications Surveys &
forwarding, in: Proceedings of the 9th ACM international symTutorials (2013) 1–23doi:10.1109/SURV.2013.042313.00103.
posium on Mobile ad hoc networking and computing - Mobi[9] P. Hui, J. Crowcroft, E. Yoneki, BUBBLE Rap: Social-based
Hoc ’08, ACM Press, New York, New York, USA, 2008, p. 251.
Forwarding in Delay Tolerant Networks, in: Proceedings of the
doi:10.1145/1374618.1374653.
9th ACM international symposium on Mobile ad hoc networking
and computing - MobiHoc ’08, ACM Press, New York, New
York, USA, 2008, p. 241. doi:10.1145/1374618.1374652.
[10] E. M. Daly, M. Haahr, Social network analysis for routing in
Appendix A. Proof of Theorem 1
disconnected delay-tolerant MANETs, in: Proceedings of the
8th ACM international symposium on Mobile ad hoc networking
Theorem 1. For ∀k ∈ [1, q), even the 1—k part of the
and computing - MobiHoc ’07, ACM Press, New York, New
sequence
is known, the newly added element Rk+1 can still
York, USA, 2007, p. 32. doi:10.1145/1288107.1288113.
[11] M. Xiao, I. J. Wu, L. Huang, Community-Aware Opportunischange the movement pattern to any state with at least 1
tic Routing in Mobile Social Networks, IEEE Transactions on
non-zero element.
Computers (2013) 1–13doi:10.1109/TC.2013.55.
[12] Z. Li, H. Shen, SEDUM: Exploiting Social Networks in Utility–
Proof. Let us represent each vector Ri as
Based Distributed Routing for DTNs, IEEE Transactions on
Computers 62 (1) (2013) 83–97. doi:10.1109/TC.2011.232.
Ri = [ri1 , ri2 , . . . , rim ] ∀j ∈ [1, m], rij > 0
[13] T. Ning, Z. Yang, H. Wu, Z. Han, Self-Interest-Driven incentives for ad dissemination in autonomous mobile social
Pi=k
networks, 2013 Proceedings IEEE INFOCOM (2013) 2310–
Ri as
and the vector i=1
2318doi:10.1109/INFCOM.2013.6567035.
[14] J. Wu, Y. Wang, Social feature-based multi-path routi=k
X
ing in delay tolerant networks, in:
2012 ProceedRi = [Sk1 , Sk2 , . . . , Skm ]
ings IEEE INFOCOM, IEEE, 2012, pp. 1368–1376.
doi:10.1109/INFCOM.2012.6195500.
i=1
[15] W. Moreira, P. Mendes, S. Sargento, Opportunistic routing
based on daily routines, 2012 IEEE International Symposium on
Thus for ∀i ∈ [1, m] and k ∈ [1, z) we have
a World of Wireless, Mobile and Multimedia Networks (WoWMoM) (2012) 1–6doi:10.1109/WoWMoM.2012.6263749.
Sk+1
= Ski + rk+1
i
i
[16] R. I. Ciobanu, C. Dobre, V. Cristea, SPRINT: Social
prediction-based opportunistic routing, 2013 IEEE 14th InPi=k+1
Pi=k+1
Let P ′ = ( i=1
Ri ) and P = ( i=1
Ri ) be the old
ternational Symposium on ”A World of Wireless, Mobile and Multimedia Networks” (WoWMoM) (2013) 1–
and new movement pattern before and after adding Rk+1
7doi:10.1109/WoWMoM.2013.6583442.
respectively. And our objective is to prove that for any
[17] T. Henderson, D. Kotz, I. Abyzov, The changing usage of a
possible state of P, there exists a possible Rk+1 to transmature campus-wide wireless network, in: Proceedings of the
form P ′ to P. From this point, when P = [κ1 , κ2 , . . . , κm ]
10th annual international conference on Mobile computing and
networking - MobiCom ’04, ACM Press, New York, New York,
is known, it is obvious that there exists such an Rk+1 if
USA, 2004, p. 187. doi:10.1145/1023720.1023739.
and only if the following linear inequations have solution.
[18] M. Ibrahim, P. Nain, I. Carreras, Analysis of relay pro
tocols for throwbox-equipped DTNs, in: 2009 7th InternaP
Sk+1
tional Symposium on Modeling and Optimization in MoSk+1
> δ j=m
j=1 ( j /m) κi = 1
i
P
∀i
∈
[1,
m]
bile, Ad Hoc, and Wireless Networks, IEEE, 2009, pp. 1–9.
Sk+1
Sk+1
< δ j=m
j=1 ( j /m) κi = 0
i
doi:10.1109/WIOPT.2009.5291625.
[19] W. Gao, Q. Li, B. Zhao, G. Cao, Multicasting in delay tolerant
i.e.
networks, in: Proceedings of the tenth ACM international
symposium on Mobile ad hoc networking and computing P
P
j=m k+1
k
MobiHoc ’09, ACM Press, New York, New York, USA, 2009,
−m
rk+1
6m
Ski − j=m
κi = 1
j=1 rj
j=1 Sj
i
δ
δ
Pj=m k+1 m k+1 m k Pj=m
p. 299. doi:10.1145/1530748.1530790.
k
− δ ri > δ Si − j=1 Sj κi = 0
URL
http://dl.acm.org/citation.cfm?id=1530790
j=1 rj
papers2://publication/uuid/CA59F365-D544-4A4C-832F-1128140EF29C
Pj=m k
http://portal.acm.org/citation.cfm?doid=1530748.1530790
k
For ∀i, we denote m
j=1 Sj by wi , and represent
δ Si −
[20] A. Vahdat, D. Becker, Epidemic Routing for Partially Conw
for
κ
=
1
and
κ
=
0
as
w
i
i
i
i|κi =1 and wi|κi =0 respecnected ad hoc Networks, Duke University Technical Report Cs2000-06 (Tech. Rep.). doi:10.1.1.34.6151.
tively. We also represent rk+1
for κi = 1 and κi = 0 in
i
[21] T. Spyropoulos, K. Psounis, C. S. Raghavendra, Spray and
the similar way. Then we have
Wait: An Efficient Routing Scheme for Intermittently Connected Mobile Networks, in: Proceeding of the 2005 ACM SIGj=m
X
COMM workshop on Delay-tolerant networking - WDTN ’05,
m
rk+1
− rk+1
(A.1)
ACM Press, New York, New York, USA, 2005, pp. 252–259.
j
i|κi =0 > wi|κi =0
E
E
j=1
17
δ
and
j=m
X
rk+1
−
j
j=1
m k+1
r
6 wi|κi =1
δ i|κi =1
corresponding to the above mentioned set R. Then each
non-empty subset Ni of set N exist in a relationship of
one-to-one correspondence with subset Ri of set R (all the
subscripts are in one-to-one correspondence), and we denote any subset Nj as
(A.2)
Pj=m k+1
= max{wi } +ǫ (ǫ > 0),
We set rk+1
j=1 rj
i|κi =0 = 0 and
then the equation A.1 always holds for ∀κi = 0.
In the following context we prove that we can choose a
suitable ǫ value to make
A.2 hold. To achieve
P equation
k+1
m k+1
r
−
this, we only need that j=m
j=1 j
δ ri|κi =1 6 min{wi }
holds for ∀κi = 1, i.e.
max{wi } + ǫ −
Nj = {nj1 , nj2 , . . . , njk }
√
For ∀j, we let λi,j = 1, τl = 2, and for ∀ni ∈ N − d, we
let λi,j = Λi . Besides, we let all the nodes have the same
movement record R = [0, 0, . . . , 0]. Consequently we have
for all nodes that
m k+1
6 min{wi }
r
δ i|κi =1
P = [1, 1, . . . , 1]
From above discussion, we only need the following inequations hold simultaneously.
rk+1
i|κi =1
Then we have
δ
> (max{wi } − min{wi } + 1) (0 < δ < 1) (A.3)
m
j=m
X
rk+1
j|κj =1
= max{wi } + ǫ
PNj ,d = 1 −
We denote that κi = 1 − Λi e−Λi −1
(A.4)
v = log(gj1 gj2 · · · gjk ) ⇐⇒ PNj ,d = 1 − κj1 κj2 · · · κjn
To make equation A.3 and A.4 hold simultaneously, we
should have
Otherwise, if there is no solution for this SSP instance,
then there does not exist any solution for the corresponding Nopt instance either. This is because that if
the Nopt instance has a solution, then there exist a set
Rj = rj1 , rj2 , . . . , rjn keeping the following equation hold
rk+1
j|κj =1
j=1
i.e.
, and then we have
In this way, there must exist a solution instance of the
SSP corresponding to that of the Nopt problem, i.e.
δ
(max{wi } − min{wi } + ǫ) 6 δ(max{wi } − min{wi } + ǫ)
m
δ(max{wi } − min{wi } + ǫ) < max{wi } + ǫ =
n
PNj ,d = 1 − κj1 κj2 · · · κjk
Assume that the number of κi = 1 in P is z, since that
0 < z 6 m, we have
j=m
X
(1 − Λi e−Λi −1 )n
ni ∈N
j=1
z·
Y
P = PNj
δ
min{wi } − max{wi }
ǫ>
δ−1
which conflicts the assumption before.
Appendix B. Proof of Theorem 2
Theorem 2. Even the global information is known in advance, i.e., the value of λi,j is available for any node ni ,
the Nopt Search Problem is still NP-Hard.
Proof. We reduce the Nopt Search Problem as a Subset
Sum Problem (SSP). Assume that we know a solution for a
certain SSP instance of which all the elements are denoted
by R = r1 , r2 , . . . , rn and the target value is v. For clarity,
we transform R as follows:
R = {log 2r1 , log 2r2 , . . . , log 2rn }
= {log g1 , log g2 , . . . , log gn }
where we have gi = 2ri (i ∈ [1, n]). Let us denote all
the non-empty subsets as R1 , R2 . . . , R2|N|−1 , and then we
construct the corresponding Nopt instance, where the target value is P = v, and the node set is N = n1 , n2 , . . . , nn
18
| 7cs.IT
|
PARTITIONS OF SINGLE EXTERIOR TYPE
arXiv:1308.0220v1 [math.RT] 1 Aug 2013
WINFRIED BRUNS AND MATTEO VARBARO
A BSTRACT. We characterize the irreducible representations of the general linear group
GL(V ) that have multiplicity 1 in the direct sum of all Schur modules of a given exterior
power of V . These have come up in connection with the relations of the lower order
minors of a generic matrix. We show that the minimal relations conjectured by Bruns,
Conca and Varbaro are exactly those coming from partitions of single exterior type.
1. I NTRODUCTION
The main motivation for this note was the desire to provide further evidence for a
conjecture of Conca and the authors [BCV, Conjecture 2.12] on the polynomial relations
between the t-minors of a generic matrix. With the notation in [BCV], let X = (xi j )
denote an m × n matrix of indeterminates over a field K of characteristic 0, R = K[X ] the
polynomial ring over the variables xi j and At ⊆ R the K-subalgebra of R generated by the
t-minors of X . With respect to a choice of bases in K-vector spaces V and W of dimension
m and n, respectively, one has a natural action of the group G = GL(V ) × GL(W ) on R,
induced by
(A, B) · X = AX B−1
∀ A ∈ GL(V ), B ∈ GL(W ).
This action restricts to At , making At a G-algebra. Since the G-decomposition of At can
be deduced from the work of De Concini, Eisenbud and Procesi [DEP], it is natural to
exploit such an action. A presentation of At as a quotient of a polynomial ring is provided
by the natural projection
π : St → At ,
Vt
Vt ∗
where St = Sym( V ⊗ W ). Also St is a G-algebra, and the map π is G-equivariant.
Therefore the ideal of relations Jt = Ker(π ) is a G-module as well.
The conjecture [BCV, Conjecture 2.12] predicts a minimal list of irreducible G-modules
generating Jt , or, by Nakayama’s lemma, the decomposition of
Jt ⊗St K
where we identify K and the residue class field of R with respect to the irrelevant maximal
ideal generated by the indeterminates. In particular, the conjecture predicts that Jt is
generated in degrees 2 and 3.
In the assignment of partitions to Young diagrams and to an irreducible representation
of GL(V ) we follow Weyman [We]: a partition of nonnegative integers λ = (λ1 , . . . , λk ),
2000 Mathematics Subject Classification. 13A50, 14M12, 14L30.
Key words and phrases. Relations of minors; Determinantal varieties; Plethysms.
The authors thank the Mathematical Sciences Research Institute, Berkeley CA, where this work started
to take shape, for support and hospitality during Fall 2012.
1
2
WINFRIED BRUNS AND MATTEO VARBARO
λ1 ≥ · · · ≥ λk , is pictorially represented by k rows of boxes of lengths λ1 , .V. . , λk with
coordinates in the fourth quadrant, and a single row of length m represents m V . The
highest weight of the representation is then given by the transpose partition tλ in which
rows and columns are exchanged: (tλ )i = |{ j : λ j ≥ i}|. With this convention, we denote
the Schur module associated with the partition λ and the vector space V by Lλ V .
Because St is a quotient of
d t
d ^
t
d ^
t
t
MO
O
^
^
M O
∗
∗
W ,
V ⊗
=
V ⊗ W
Tt =
d≥0
d≥0
Pieri’s rule implies that the irreducible summands of Jt ⊗St K must be of the form
Lγ V ⊗ Lλ W ∗
where γ and λ are partitions satisfying the following conditions:
(i) γ , λ ⊢ dt,
(ii) both γ and λ have at most d rows.
We call such partitions (or bipartitions (γ |λ )) (t, d)-admissible (just t-admissible if we
do not need to emphasize the degree). In [BCV] a set A of (t, 2)-admissible bipartitions
(γ |λ ) and a set B of (t, 3)-admissible bipartitions (γ |λ ) were found such that
(1.1)
M
(γ |λ )∈A
Lγ V ⊗ Lλ W ∗ ⊕
M
(γ |λ )∈B
Lγ V ⊗ Lλ W ∗ ⊆ Jt ⊗St K.
Conjecture 2.12 in [BCV] states that the inclusion in Equation (1.1) is an equality. For
the convenience of the reader and since it is crucial for the following we recall how A and
B are defined.
(i) For u ∈ {0, . . . ,t} let:
τu = (t + u,t − u).
(ii) For u ∈ {1, . . . , ⌊t/2⌋} let
γu = (t + u, t + u, t − 2u)
and
λu = (t + 2u, t − u, t − u).
and
σu = (t + 2u − 1, t − u + 1, t − u).
(iii) For each u ∈ {2, . . ., ⌈t/2⌉} let
ρu = (t + u, t + u − 1, t − 2u + 1)
With this notation,
A = (τu |τv ) : 0 ≤ u, v ≤ t, u + v even, u 6= v ,
B = (γu |λu ), (λu|γu ) : 1 ≤ u ≤ ⌊t/2⌋
∪ ((ρv|σv ), (σv |ρv) : 2 ≤ v ≤ ⌈t/2⌉ .
Note that not all the partitions above are supported by the underlying vector spaces
if their dimensions are too small: a partition λ can only appear in a representation of
GL(V ) if λ1 ≤ dimV . For simplicity we have passed this point over since it is essentially
irrelevant. The reader is advised to remove all partitions from the statements that are too
large for the vector spaces under consideration.
PARTITIONS OF SINGLE EXTERIOR TYPE
3
The decomposition of St as a module over the “big” group
H = GL(E) × GL(F),
is well known by Cauchy’s rule:
St =
(1.2)
M
µ
E=
t
^
V,
F=
t
^
W,
Lµ E ⊗ Lµ F ∗
where µ is extended over all partitions. The GL(V )-decomposition of Lµ E is an essentially unsolved plethysm. However, the partitions in the definition of A and B play a very
special role in it, as was already observed in [BCV]:
Definition 1.1. Let λ ⊢ dt be t-admissible. Then λ is said to be ofVsingle t -type µ if
µ ⊢ d is the only partition such that Lλ V is a direct summand of Lµ ( t V ) and, moreover,
V
has multiplicity 1 in it. Without
specifying
µ , notice that λ is of single t -type if and
L
Vt
only if λ has multiplicity 1 in α ⊢d Lα ( V ).
V
In this note we will classify all partitions of single t -type (or simply single exterior
V
type) and show that the bi-partitions in the sets A and B are exactly those of single t -type
that occur in a minimal generating set of Jt . While this observation does certainly not
prove the conjecture in [BCV], it provides further evidence for it.
V
2. AUXILIARY
RESULTS ON PARTITIONS
In this section we discuss two transformations of partitions that preserve single exterior
type. It was already observed in [BCV] that trivial extensions in the following sense are
irrelevant: if a partition λ̃ arises from a t-admissible partition λ ⊢ dt by prefixing λ with
columns of length d, then λ̃ is called a trivial extension of λ . We quote [BCV, 1.16] (eλ
denotes the multiplicity of λ ):
Proposition 2.1. Let µ be a partition of d and consider partitions λ = (λ1 , . . . , λk ) ⊢ td
with k ≤ d and λ̃ = (λ1 + 1, . . . , λk + 1, 1, . . . , 1) ⊢ dt + d. If dimK V ≥ λ1 + 1, then
t
^
t+1
^
V
= eλ̃ Lµ
V .
eλ Lµ
In particular, λ is of single
Vt
-type µ if and only if λ̃ is of single
Vt+1
-type µ .
Next we want to show that aV similar result holds for dualization, in the sense that
V , n = dimV , is dual to t V (up to tensoring with the determinant). Let λ =
(λ1 , . . . , λk ) ⊢ td be t-admissible; then we set
Vn−t
λ ∗,n = (n − λd , . . . , n − λ1 ) ⊢ (n − t)d.
Evidently λ ∗,n is (n−t)-admissible. Note that λ and λ ∗,n rotated by 180◦ degrees complement each
other to a d ×n rectangle (representing the d-th tensor power of the determinant
V
detV = n V when n = dimV ).
Notice that λ ∗,n is a trivial extension of λ ∗,λ1 . In view of this we will denote λ ∗,λ1 just
with λ ∗ , calling it simply the dual of λ . Also, note that if k = d, so that λ is a trivial
extension of some γ , then λ ∗,n = γ ∗,n . Therefore, when speaking of dual partitions, we
will usually assume that n = λ1 and k < d.
4
WINFRIED BRUNS AND MATTEO VARBARO
Proposition 2.2. Let µ be a partition of d and consider a t-admissible partition λ =
(λ1 , . . . , λk ) ⊢ td. Suppose dimV = λ1 . Then
eλ Lµ
In particular, λ is of single
Vt
t
^
V = eλ ∗ Lµ (
λ^
1 −t
V .
-type if an only if λ ∗ is of single
Vλ1 −t
-type.
Proof. Set n = dimV = λ1 . Consider the GL(V )-equivariant multiplication
t
^
V⊗
n−t
^
V → detV.
It induces an equivariant isomorphism
t
n−t
n−t
n−t
^
^
^ ∗
^
V∼
V, detV =
V ⊗ detV =
V ∗ ⊗ detV.
= HomK
Next we can pass to the d-th tensor power on the right and the left, and apply the Young
symmetrizer Yµ (see Fulton and Harris [FH, p. 46] inverting rows and columns) to obtain
a GL(V )-equivariant isomorphism
Yµ
d ^
t
O
V∼
= Yµ
d n−t
^
O
V ∗ ⊗ detV .
Next we can go from Yµ d ( n−t V ∗ ⊗ detV ) to Yµ d n−t V ∗ , except that we have to
N V
N
subtract the weight of d detV from each weight in Yµ d ( n−t V ∗ ⊗ detV ). Finally,
if we replace GL(V ) by GL(V ∗ ) as the acting group, we see that every partition λ in
N V
N V
Yµ d t V goes with equal multiplicity to the partition λ ∗ in Yµ d n−t V ∗ . But the
multiplicities
depend only
on the dimension of the basic vector space, and therefore we
Vn−t ∗
Vn−t
can replace
V by
V.
N V
N V
Below we will use the obvious generalization of Proposition 2.2 to λ ∗,n that results
from Proposition 2.1.
3. PARTITIONS
OF SINGLE EXTERIOR TYPE
The characterization of partitions of single exterior type is based on a recursive criterion
established in [BCV]. For it and also for the characterization of the minimal relations of
single exterior type we need the same terminology.
Let λ be a (t, d)-admissible diagram. Given 1 ≤ e ≤ d, we say that α is a (t, e)predecessor of λ if and only if α is a (t, d − e)-admissible diagram such that tαi ≤ tλi ≤
tα + e for all i = 1, . . ., λ (we set tα = 0 if i > α ). In such a case we also say that λ is
i
1
i
1
a (t, e)-successor of α . If we just say that α is a t-predecessor of λ , we mean that α is a
(t, e)-predecessor of λ for some e, and analogously for λ being a t-successor of α . (This
terminology deviates slightly from [BCV] where a predecessor is necessarily a (t, 1)predecessor.) The Littlewood-Richardson rule implies at once that, for a (t, d)-admissible
diagram λ and a (t, d − e)-admissible diagram α the following are equivalent:
of λ .
(i) α is a (t, e)-predecessor
N V
(ii) Lλ V occurs in ( e t V ) ⊗ Lα V , where V is a K-vector space of dimension ≥ λ1 .
PARTITIONS OF SINGLE EXTERIOR TYPE
5
Now we can quote the following criterion for single t -type from [BCV, Proposition
1.22]. (Condition (iv) has been added here. It strengthens (iii), but follows from (iii) by
induction.)
V
Proposition 3.1. Let λ ⊢ dt and µ ⊢ d be partitions such that Lλ V occurs in Lµ ( t V ).
Then the following are equivalent:
V
(i) λ is of single t -type;
N V
(ii) the multiplicities of λ and of µ in d ( t V ) coincide;
V
(iii) every (t, 1)-predecessor λ ′ of λ is of single t -type µ ′ where µ ′ is a (1, 1)predecessor of µ , and no two distinct (t, 1)-predecessors of λ share the same
(1, 1)-predecessor µ ′ of µ ;
V
(iv) every t-predecessor λ ′ of λ is of single t -type µ ′ where µ ′ is a 1-predecessor of
µ , and no two distinct t-predecessors of λ share the same 1-predecessor µ ′ of µ .
V
As we will see in a moment, one class of single
Vt
-type partitions is given by the hooks.
Definition 3.2. A diagram λ = (λ1, . . . , λk ) with λ2 ≤ 1 is called a hook.
A hook can be always written like (a, 1b), where 1b means b ones.
Lemma 3.3. Let d > 0 and k ∈ {0, . . ., d − 1}. Then (td − k, 1k ) is of single
where:
(i) µ = (d − k, 1k ) if t is odd.
(ii) µ = (k + 1, 1d−k−1 ) if t is even.
Vt
-type µ
Proof. Let us fix t and use induction on d. For d = 2 the statement is very easy
to prove.
Vt
For d = 3 [BCV, Proposition 1.18] implies that L(3t−1,1)V occurs in L(2,1) ( V ), so we
are done in this case by Proposition 3.1V
(ii). Therefore assume d > 3.
N Vt
If t is odd, then L(dt)V occurs in L(d) ( t V ): In fact, L(dt)V has multiplicity 1 in
V,
Vt
d
so it can occur in Lµ ( V ) only if µ = (d) (the d-th exterior power) or µ = (1 ) (the dV
th symmetric power). Furthermore (2t) is a t-predecessor of (dt), and 2t V occurs in
V2 Vt
V V
( V ) (for instance see [BCV, Lemma 2.1]). Therefore L(dt)V occurs in 2 ( t V ) ⊗
N
V
V
( d−2 t V ). In particular, it cannot occur in L(1d ) ( t V ). In the same way, one sees that
V
L(dt)V occurs in L(1d ) ( t V ) whenever t is even.
From now on let us assume t odd; the even case is similar. If 0 < k < d − 1, then
(dt − k, 1k ) has two (t, 1)-predecessors, namely
((d − 1)t − k, 1k )
and
((d − 1)t − k + 1, 1k−1 ).
By induction, the respective Schur modules occur in
t
t
^
^
L(d−k−1,1k )
V
and
L(d−k,1k−1 )
V .
So, the Schur modules
corresponding to the
(t, 1)-successors of ((d − 1)t − k, 1k ) can
Vt
Vt
occur in L(d−k,1k ) ( V ) or in L(d−k+1,1k−1 ) ( V ), and the ones corresponding to the (t, 1)V
V
successors of ((d −1)t −k+1, 1k−1 ) can occur in L(d−k+1,1k−1 ) ( t V ) or in L(d−k,1k ) ( t V ).
By counting multiplicities and using
d > 3, one can check that the only possibility is that
Vt
L(dt−k,1k )V occurs in L(d−k,1k ) ( V ). Notice that the multiplicity of L(dt−k,1k )V is the
6
WINFRIED BRUNS AND MATTEO VARBARO
same of the one of L(d−k,1k ) (
conclude.
Vt
V ) in
N Vt
V , i.e.
d−1
k ,
so Proposition 3.1 (ii) lets us
We must pay particular attention to the duals of hooks: The dual of the hook (dt −
k, 1k ) ⊢ dt is the diagram ((dt − k)d−k−1 , (dt − k − 1)k ) ⊢ d(dt − k − t). Notice that is the
unique partition of d(dt − k − 1) with λd = 0 and λd−1 ≥ λ1 − 1.
Before stating the main theorem it is useful to remark the following:
Lemma 3.4. A diagram (a, b, c) ⊢ 3t (where c = 0 is not excluded) is of single
and only if
min{a − b, b − c} ≤ 1.
Vt
-type if
Since all partitions λ ⊢ 2t are of single t -type, one must find exactly those partitions
(a, b, c) ⊢ 3t that have no two predecessors in the second symmetric or second exterior
power. Since the latter are easily characterized (for example, see [BCV, Lemma 2.1]),
the proof of Lemma 3.4 is an easy exercise. Because of Proposition 2.1 one may assume
c = 0, and Proposition 2.2 helps to further reduce the number of
cases.
V
For theVproof of the next theorem we will abbreviate “single t -type” by “ST” and “not
of single t -type” by “NST”.
V
Theorem 3.5. A t-admissible diagram λ = (λ1 , . . . , λk ) ⊢ dt is of single
and only if it satisfies one (or more) of the following:
Vt
-type µ ⊢ d if
(i) λd ≥ t − 1, in which case µ = (λ1 − t + 1, . . . , λd − t + 1).
(ii) λ1 ≤ t + 1, in which case µ = λ ∗,t+1 .
(iii) λd ≥ λ2 − 1. If λ = (t d ), then µ = (1d ). Otherwise put k = max{i : λi > λd }:
according with t − λd being odd or even, µ = (d − k, 1k ) or µ = (k + 1, 1d−k−1 ).
(iv) λd−1 ≥ λ1 − 1. If λ = (t d ), then µ = (1d ). Otherwise put k = min{i : λi < λ1 }:
according with λ1 − t being odd or even, µ = (k, 1d−k ) or µ = (d − k + 1, 1k−1 ).
If λ is in one of the four classes above, then we know that it is of single t -type from
what done until now: (i) If λd ≥ t − 1, then it is a trivial extension of µ = (λ1 − t +
V
1, λ2 −t + 1, . . . , λd −t + 1) ⊢ d, that is obviously of single 1 -type; (ii) if λ1 ≤ t + 1, then
V
µ = λ ∗,t+1 ⊢ d is of single 1 -type, so Proposition 2.2 let us conclude; (iii) If λd ≥ λ2 −1,
then λ is a trivial extension of a hook. The shape of µ follows from Proposition 2.1 and
Lemma 3.3; (iv) if λd−1 ≥ λ1 − 1, then λ ∗ is a hook. From this, combining Lemma 3.3
and Proposition 2.2, we get the shape of µ .
As we have just seen, the four classes can described as follows: (i) consists of the trivial
extensions of 1-admissible partitions, (ii) is dual to (i) in the sense of Proposition 2.2, (iii)
contains the hooks and their trivial extensions, and (iv) is dual to (iii).
The classification in V
the theorem completely covers the cases d = 1 and d = 2, in which
all shapes are of single t -type, and also the case d = 3 done in Lemma 3.4. Therefore we
may assume that d ≥ 4. Then the theorem follows from the next lemma and Proposition
3.1. In its proof we will use the theorem inductively.
V
Lemma 3.6. If d ≥ 4 and λ is not of one of the types in the theorem, then it has an NST
(t, 1)-predecessor.
PARTITIONS OF SINGLE EXTERIOR TYPE
7
The lemma shows that the critical degree is d = 3 in which the condition that the predecessors of λ occur in pairwise different predecessors of µ must be used.
Proof. If t = 1 all partitions λ fall into the class (i) and are certainly ST. So we can assume
t ≥ 2.
Suppose first that λ is itself a trivial extension. Then we pass to its trivial reduction λ ′ .
It is enough to find an NST predecessor for λ ′ . It yields an NST predecessor of λ after
trivial extension. From now on we can assume that λ has at most d − 1 rows.
Suppose first that λ = (λ1 , . . . , λk ) is a successor of a hook. Let k′ = max{2, k − 1}. We
′
choose γ = ((d − 1)t − k′ , 2, 1k −2 ) ⊢ (d − 1)t. Then γ does not fall into one of the classes
(i)–(iv), provided γ1 ≥ t + 2. Using k′ ≤ d − 2, one derives this immediately from d ≥ 4
and t ≥ 2. The inequality γ1 ≥ t + 1 is sufficient to make γ a predecessor of λ .
Next suppose λ = (λ1 , λ2 , 1k−2 ) ⊢ dt. If λ has a hook predecessor, then we are done
by the previous case. Therefore we can assume that λ2 ≥ t + 2. If k = 2, we pass to
γ = (λ1 , λ2 − t), and if k ≥ 3, we choose γ = (λ1 , λ2 − (t − 1), 1k−3). Then γ is not of
types (i)–(iv). (We are dealing with this case separately since the duals will come up
below.)
In the remaining case we choose the predecessor γ of λ with the lexicographic smallest
set of indices for the columns in which γ and λ differ by 1. If γ is a hook, then we are
done as above. So we can assume that γ is not a hook.
Suppose that γ1 < λ1 . Then λ2 ≤ t − 1, and γ is not a trivial extension since the bottom
row of λ has been removed completely, and γ has at most d − 2 rows. On the other hand,
λ1 + (d − 2)λ2 ≥ dt implies λ1 ≥ 2t + 2, and so γ1 ≥ t + 2. It follows that γd−2 ≤ λd−2 <
γ1 − 1, and γ is not of type (i)–(iv).
The case γ1 = λ1 > t + 1 remains. We can assume that γ is ST. This is only possible if
(1) γd−2 ≥ γ1 − 1 or (2) γ is the trivial extension of a hook or (3) γd−1 ≥ t − 1.
(1) If γd−2 ≥ γ1 − 1, then λd−2 ≥ λ1 − 1, and λ ∗ is of the second type discussed. We
find an NST predecessor of λ ∗ and dualize back.
(2) If γ is a trivial extension of a hook, then γ2 ≤ γd−1 +1 and λd−1 ≥ t +1. In particular
γ2 = λ2 , and γd−1 = λd−1 − t ≤ λ2 − t = γ2 − t, which is a contradiction since t ≥ 2.
(3) In this case we must have λd−1 ≥ 2t − 1 since we remove min{λd−1 ,t} boxes from
row d − 1 of λ . This is evidently impossible (because t ≥ 2 and d ≥ 4).
4. M INIMAL
RELATIONS OF SINGLE EXTERIOR TYPE
In this last section we are going to prove the result which motivated us for producing
this note. We will adopt here the notation given in the introduction.
Let us first recall a result of [BCV]. As already mentioned, a decomposition of St =
Sym(E ⊗ F ∗ ) in irreducible H-representations is provided by the Cauchy formula (1.2),
namely
M
Lµ E ⊗ Lµ F ∗ ,
St =
µ
where µ ranges among all the partitions. So, because G is a subgroup of H whose action
is the restriction of that of H, the irreducible G-representation Lγ V ⊗ Lλ W ∗ occurs in the
G-decomposition of St if and only if there exists µ ⊢ d such that Lγ V occurs in the GL(V )V
V
decomposition of Lµ ( t V ) and Lλ W ∗ occurs in the GL(W )-decomposition of Lµ ( t W ∗ ).
8
WINFRIED BRUNS AND MATTEO VARBARO
Moreover, if such a µ ⊢ d exists and γ and λ are both of single t -type, then Lγ V ⊗ Lλ W ∗
is a direct summand of Jt ⊗St K if and only if γ 6= λ and the predecessors of γ and of λ
coincide [BCV, Proposition 1.21 and Theorem 1.23(iv)]. This is the fact on which the
proof of the next theorem is based.
V
Theorem 4.1. Let Lγ V ⊗ Lλ W ∗ be a direct summand of Jt ⊗St K such that both γ and λ
V
are diagrams of single t -type. Then (γ |λ ) ∈ A ∪ B.
Proof. For t = 1 there is nothing to prove because J1 = (0). So assume t ≥ 2.
From what said above γ and λ must be t-admissible
partitions of the same number dt.
L
∼
If d = 1 then γ = λ = (t); if d = 2, then (Jt )2 = (γ |λ )∈A Lγ V ⊗ Lλ W ∗ by [BCV, Lemma
2.1]; if d = 3 then [BCV, Proposition 3.16] does the job.
So from now on we will focus on d ≥ 4. Recall that in Theorem 3.5 have been identified
4 (not disjoint) sets of diagrams, say E1t = {diagrams as in (i)}, E2t = {diagrams as in (ii)}
and so on, such that:
{diagrams of single
We start by showing the following:
Vt
-type} = E1t ∪ E2t ∪ E3t ∪ E4t .
Lemma 4.2. Let Lγ V ⊗ Lλ W ∗ be a direct summand of Jt ⊗St K such that both γ and λ
V
are diagrams of single t -type belonging to the same Eit for some i ∈ {1, 2, 3, 4}. Then
(γ |λ ) ∈ A ∪ B.
Proof of Lemma 4.2. We know that γ and λ are different t-admissible partitions of dt
sharing the same µ ⊢ d. This excludes i ∈ {1, 2}, because in these cases Theorem 3.5
says that γ and λ cannot share the same µ if they are different.
Suppose i = 3. We must have γd 6= λd if λ and γ belong to the same µ . Assume λd > γd .
′
= γd . This cannot be a predecessor of λ ,
The diagram γ has a predecessor γ ′ with γd−1
and so γ and λ do not have the same predecessors.
So only the case i = 4 remains. Let s = max{γ1 , λ1 }. If s = t, then γ = λ = (t d ), so we
can assume s > t. If γ and λ share the same µ , by combining Propositions 2.2 and 2.1 γ ∗,s
and λ ∗,s share µ as well. Of course γ ∗,s and λ ∗,s belong to E3s−t , and they are different if
γ and λ are different. In this case, we know by the previous case that γ ∗,s and λ ∗,s have
different predecessors, and by dualizing we infer the same for γ and λ .
Let us go ahead with the proof of Theorem 4.1. Set γ = (γ1 , . . . , γh ) and λ = (λ1 , . . . , λk )
with h, k ≤ d. If h = k = d we can use induction on t since both γ and λ are trivial
extensions.
If h and k are both less than d, then neither γ nor λ belong to E1t . Assume that γ ∈ E3t .
Then µ is a hook. By the lemma, λ ∈
/ E3t . Since µ is a hook, k < d and λ ∈
/ E3t , it follows
t
t
t
that λ ∈ E4 (recall that E2 and E4 are not disjoint). Because γ is a t-admissible hook and
h < d, we get γ1 > dt − d + 1 ≥ 4t − 3. Then λ1 > 3t − 3, otherwise γ and λ would have
different predecessors. Therefore λ ⊢ dt > (d − 1)(3t − 3), which is impossible whenever
d ≥ 3 (recall that t ≥ 2). So, by symmetry, we can assume that neither γ nor λ is in E3t .
Therefore γ and λ belong to E2t ∪ E4t . However γ and λ share the same µ and, in such a
situation, µ is a hook if and only if γ and λ both belong to E4t , a case already excluded in
the lemma.
PARTITIONS OF SINGLE EXTERIOR TYPE
9
So, we can assume by symmetry that h < d and k = d. Notice that h = d − 1, because
all the predecessors of λ will have d − 1 rows. For the same reason we can even infer that
γd−1 > t, otherwise we could entirely remove γd−1 getting a predecessor of γ with d − 2
rows. Since d ≥ 4, we have γ2′ > t for all γ ′ predecessors of γ . So λ does not belong to
E3t , since in this case λ2 ≤ t. Since γd−1 > t, Theorem 3.5 tells us that γ ∈ E4t (once again,
recall that E2t and E4t are not disjoint): so µ must be a hook. If λ ∈ E1t , then γd−1 ≥ 2t − 1
′
< t − 1, that cannot be a predecessor
(otherwise γ would have a predecessor γ ′ with γd−1
of λ ). This is evidently impossible if d ≥ 4. So Theorem 3.5 implies that λ ∈ E4t , and the
lemma lets us conclude.
R EFERENCES
[BCV] W. Bruns, A. Conca, M. Varbaro, Relations among the Minors of a Generic Matrix. Adv. Math.
244, pp. 171-206, 2013.
[DEP] C. De Concini, D. Eisenbud, C. Procesi, Young Diagrams and Determinantal Varieties, Invent.
Math. 56, pp. 129-165, 1980.
[FH] W. Fulton, J. Harris, Representation Theory. A First Course, Graduate Texts in Mathematics 129,
1991.
[We] J. M. Weyman, Cohomology of vector bundles and syzygies, Cambridge Tracts in Mathematics 149,
2003.
U NIVERSIT ÄT O SNABR ÜCK , I NSTITUT F ÜR M ATHEMATIK , 49069 O SNABR ÜCK , G ERMANY
E-mail address: [email protected]
D IPARTIMENTO DI M ATEMATICA , U NIVERSIT À
E-mail address: [email protected]
DEGLI
S TUDI DI G ENOVA , I TALY
| 0math.AC
|
A DISCRETE-TIME PONTRYAGIN MAXIMUM PRINCIPLE ON
MATRIX LIE GROUPS
arXiv:1612.08022v3 [cs.SY] 6 Dec 2017
KARMVIR SINGH PHOGAT, DEBASISH CHATTERJEE, AND RAVI BANAVAR
Abstract. In this article we derive a Pontryagin maximum principle (PMP)
for discrete-time optimal control problems on matrix Lie groups. The PMP
provides first order necessary conditions for optimality; these necessary conditions typically yield two point boundary value problems, and these boundary
value problems can then solved to extract optimal control trajectories. Constrained optimal control problems for mechanical systems, in general, can only
be solved numerically, and this motivates the need to derive discrete-time models that are accurate and preserve the non-flat manifold structures of the underlying continuous-time controlled systems. The PMPs for discrete-time systems
evolving on Euclidean spaces are not readily applicable to discrete-time models
evolving on non-flat manifolds. In this article we bridge this lacuna and establish a discrete-time PMP on matrix Lie groups. Our discrete-time models are
derived via discrete mechanics, (a structure preserving discretization scheme,)
leading to the preservation of the underlying manifold over time, thereby resulting in greater numerical accuracy of our technique. This PMP caters to
a class of constrained optimal control problems that includes point-wise state
and control action constraints, and encompasses a large class of control problems that arise in various field of engineering and the applied sciences.
1. introduction
The Pontryagin maximum principle (PMP) provides first order necessary conditions for a broad class of optimal control problems. These necessary conditions
typically lead to two-point boundary value problems that characterize optimal control, and these problems may be solved to arrive at the optimal control functions.
This approach is widely applied to solve optimal control problems for controlled
dynamical systems that arise in various fields of engineering including robotics,
aerospace [2, 8, 21, 23], and quantum mechanics [6, 19].
Constrained optimal control problems for nonlinear continuous-time systems can,
in general, be solved only numerically, and two technical issues inevitably arise.
First, the accuracy guaranteed by a numerical technique largely depends on the
discretization of the continuous-time system underlying the problem. For control
systems evolving on complicated state spaces such as manifolds, preserving the
manifold structure of the state space under discretization is a nontrivial matter. For
controlled mechanical systems evolving on manifolds, discrete-time models should
preferably be derived via discrete mechanics since this procedure respects certain
system invariants such as momentum, kinetic energy, (unlike other discretization
schemes derived from Euler’s step,) resulting in greater numerical accuracy [25, 28,
29]. Second, classical versions of the PMP are applicable only to optimal control
Date: December 7, 2017.
Key words and phrases. Pontryagin maximum principle, Lie groups, optimal control, discrete
mechanics, state constraints.
The authors were supported partially by the grant 14ISROC007 from the Indian Space Research Organization.
1
2
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
problems in which the dynamics evolve on Euclidean spaces, and do not carry
over directly to systems evolving on more complicated manifolds. Of course, the
PMP, first established by Pontryagin and his students [15, 31] for continuous-time
controlled systems with smooth data, has, over the years, been greatly generalized,
see e.g., [2,3,10,11,13,17,26,27,37,40]. However, there is still no PMP that is readily
applicable to control systems with discrete-time dynamics evolving on manifolds.
As is evident from the preceding discussion, numerical solutions to optimal control
problems via digital computational means need a discrete-time PMP. The present
article contributes towards filling this lacuna: here we establish a PMP for a class
of discrete-time controlled systems evolving on matrix Lie groups.
Optimal control problems on Lie groups are of great interest due to their wide
applicability across the discipline of engineering: robotics [9], computer vision [39],
quantum dynamical systems [6, 19], aerospace systems such as attitude maneuvers
of a spacecraft [20,21,33]. Continuous-time optimal control problems on Lie groups
are solved by applying the PMP that leads to two point boundary value problems,
which are then solved numerically to obtain the optimal control functions. In case
of mechanical systems on Lie groups, to ensure greater numerical accuracy, discretetime models are obtained via discrete mechanics and optimal control problems are
posed in discrete-time for these models. Early results on optimal control problems
on Lie groups for discrete-time systems derived via discrete mechanics may be found
in [20, 21, 23, 23]. It is worth noting that simultaneous state and action constraints
have not been considered in any of these formulations. The inclusion of state and
action constraints in optimal control problems, while of crucial importance in all
real-world problems, makes constrained optimal control problems technically challenging, and, moreover, classical variational analysis techniques are not applicable
in deriving first order necessary conditions for such constrained problems [31, p. 3].
More precisely, the underlying assumption in calculus of variations that an extremal
trajectory admits a neighborhood in the set of admissible trajectories does not necessarily hold for such problems due to presence of the constraints. This article
addresses a class of optimal control problems in which the discrete-time controlled
system dynamics evolve on matrix Lie groups, and are subject to simultaneous
state and action constraints. We derive first order necessary conditions bypassing
techniques involving classical variational analysis. Discrete-time PMP for various
special cases are subsequently derived from the main result.
A discrete-time PMP is fundamentally different from a continuous-time PMP due
to intrinsic technical differences between continuous and discrete-time systems [7, p.
53]. While a significant research effort has been devoted to developing and extending
the PMP in the continuous-time setting, by far less attention has been given to
the discrete-time versions. A few versions of discrete-time PMP can be found in
[5,12,17].∗ In particular, Boltyanskii developed the theory of tents using the notion
of local convexity, and derived general discrete-time PMP’s that address a wide
class of optimal control problems in Euclidean spaces subject to simultaneous state
and action constraints [4]. This discrete-time PMP serves as a guiding principle in
the development of our discrete-time PMP on matrix Lie groups even though it is
not directly applicable in our problem; see Remark 3.2 ahead for details.
Our main result, a discrete-time PMP for controlled dynamical systems on matrix Lie groups, and its applications to various special cases are derived in §2. §3
provides a detailed proof of our main result, and the proofs of the other auxiliary
results and corollaries are collected in the Appendices.
∗
Some early attempts in establishing discrete-time PMP in Euclidean spaces have been mathematically incorrect [7, p. 53].
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
3
2. Background and Main Results
This section contains an introduction to Lie group variational integrators that
motivates a general form of discrete-time systems on Lie groups. Later in this
section we establish a discrete-time PMP for optimal control problems associated
with these discrete-time systems.
To illustrate the engineering motivation for our work, and ease understanding, we
first consider an aerospace application. Let us first consider an example of control
of spacecraft attitude dynamics in continuous time. The configuration space SO(3)
(the set of 3 × 3 orthonormal matrices with real entries and determinant 1) of a
spacecraft performing rotational maneuvers [21], is a matrix Lie group with matrix
multiplication as the group operation. Let R ∈ SO(3) be the rotation matrix that
relates coordinates in the spacecraft body frame to the inertial frame, (see Figure
1,) let ω ∈ R3 be the spacecraft momentum vector in the body frame, and let
u ∈ R3 be the torque applied to the spacecraft in the body frame. The attitude
dynamics in this setting is given in the spacecraft body frame in a standard way [21]
as:
(2.1)
Ṙ = Rω̂,
(2.2)
J ω̇ = ω̂Jω + u,
where ω̂ ∈ so(3), (the set of 3 × 3 skew-symmetric matrices with real entries) is the
Lie algebra [34] corresponding to the Lie group SO(3), J is the 3 × 3 moment of
inertia matrix of the spacecraft in the body frame. The first equation (2.1) describes
the kinematic evolution and the second equation (2.2) describes the dynamics.
n̂3
Inertial Frame
b̂3
n̂i = R b̂i
Body Frame
b̂2
n̂2
n̂1
b̂1
Figure 1. Rigid body orientation.
Let us, as a first step, uniformly discretize the continuous-time model (2.1)-(2.2)
to arrive at an approximate discrete-time model. Fixing a step length h > 0, we
have the discrete-time instances t ∈ {0} ∪ N corresponding to the continuous-time
instances th ∈ R in a standard way. Therefore, the system configurations at the
discrete-time instances defined above are given by
Rt := R(th),
ωt := ω(th)
for all
t ∈ {0} ∪ N.
4
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
If we assume that spacecraft body momentum is constant on the interval [th, (t+
1)h[, i.e., ω(s) = ω(th) for s ∈ [th, (t + 1)h[, then the corresponding kinematic
equations Ṙ(s) = R(s)ω̂t for s ∈ [th, (t + 1)h[ represent a linear system. This
linear system admits an analytical solution, and the discrete-time evolution of the
continuous-time kinematic equation (2.1) is approximated as
(2.3)
Rt+1 = Rt ehω̂t ,
where e : so(3) → SO(3) is the exponential map [1, p. 256] from the Lie algebra
so(3) to the Lie group SO(3).† Similarly, the discrete-time system corresponding to
the continuous-time dynamics (2.2) is approximated to be
(2.4)
Jωt+1 = (I3×3 + hω̂t )Jωt + hut .
It is worth noting here that the discrete integration step ehω̂t that describes the
discrete evolution of the kinematic equation (2.3) can, in general, be a function of
the configuration space variable Rt ∈ SO(3) along with the spacecraft momentum
vector ω̂t ∈ so(3). Similarly, the discrete-time evolution of the spacecraft momentum dynamics (2.4) can also, in general, depend on the orientation especially if the
spacecraft is subjected to internal actuations via reaction wheels.
The above considerations lead to the following general form of the state dynamics:
(
Rt+1 = Rt st (Rt , ωt ),
(2.5)
ωt+1 = ft (Rt , ωt , ut ),
where ft : SO(3)×R3 ×R3 → R3 , st : SO(3)×R3 → SO(3) are maps that define the
discrete evolution of the system. Note that the exponential map e : so(3) → SO(3)
is a diffeomorphism on a suitable neighborhood of 0 ∈ so(3). Let ât ∈ so(3) be
a vector; then there exists h > 0 such that so(3) 3 ât 7→ eât h ∈ SO(3) is a
diffeomorphism. The diffeomorphic property of the map so(3) 3 ât 7→ eât h ∈ SO(3)
restricted to a suitable open set is crucial for defining the local parametrization
of the Lie group SO(3) in terms of Lie algebraic elements, thus distilling a vector
space structure to the discrete-time optimal control problem defined on the Lie
group SO(3).
We are now in a position to generalize the idea of discretization brought forth in
the example of the attitude dynamics of a spacecraft to dynamical systems evolving
on matrix Lie groups. To this end, let N be a natural number; in the sequel N
will play the role of a planning or control horizon, and will be fixed throughout.
Inspired by (2.5), we consider the dynamics split into two parts, the first of which
occurs on a matrix Lie group G, and the second on a Euclidean space Rnx . The
discrete-time evolution of our control system on the configuration space G × Rnx
is defined by the recursion
(
qt+1 = qt st (qt , xt ) ,
(2.6)
t = 0, . . . , N − 1,
xt+1 = ft (qt , xt , ut ) ,
with the following data:
(2.6-a) qt ∈ G, xt ∈ Rnx are the states of the system,
(2.6-b) st : G × Rnx → G is a map depicting the dynamics on the matrix Lie group
G,
(2.6-c) ft : G × Rnx × Rnu → Rnx is a map capturing the dynamics on Rnx ,
†Let G be the Lie group with associated Lie algebra g. Then, for any X ∈ g, there exist a
d
eX (t) = X, and eX (t + s) = eX (s) eX (t),
map eX ( · ) : R → G such that: eX (0) = e ∈ G, dt
where e is the group identity.
t=0
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
5
(2.6-d) ut ∈ Ut ⊂ Rnu , where Ut is a nonempty set of feasible control actions at
time t.
A control action ut is applied to our system at the instant t to drive the sysN −1
tem states from (qt , xt ) to (qt+1 , xt+1 ) governed by (2.6). The sequence {ut }t=0
N
is known as the controller, the sequence {(qt , xt )}t=0 describing system states
N −1
is called the system trajectory
[4] under the controller {ut }t=0
, with the pair
N
N −1
{(qt , xt )}t=0 , {ut }t=0
referred to as a state-action trajectory.
We synthesize a controller for our system (2.6) by minimizing the performance
index
(2.7)
J (q, x, u) :=
N
−1
X
ct (qt , xt , ut ) + cN (qN , xN )
t=0
with the following data:
(2.7-a) ct : G × Rnx × Rnu → R is a map that accounts for the cost-per-stage for
each t = 0, . . . , N − 1,
(2.7-b) cN : G × Rnx → R is a map that accounts for the final cost.
In addition, we impose
(2.7-c) control constraints ut ∈ Ut for each t = 0, . . . , N − 1,
(2.7-d) state constraints gt (qt , xt ) ≤ 0, for each t = 1, . . . , N, where gt : G × Rnx →
Rnt is a given map,
(2.7-e) Initial conditions (q0 , x0 ) = (q̄0 , x̄0 ) , where (q̄0 , x̄0 ) is fixed.
The set
n
N
N −1
A :=
{(qt , xt )}t=0 , {ut }t=0
(q0 , x0 ) = (q̄0 , x̄0 ) , gt (qt , xt ) ≤ 0, and ut ∈ Ut
o
is termed as the set of admissible state-action trajectories.
Assumption 2.1. The following assumptions on the various maps in (2.6) and (2.7)
are enforced throughout this article:
(A-i) The maps st , ft , gt , ct , cN are smooth.
(A-ii) There exists an open set O ⊂ g such that:
(a) the exponential map e : O → e(O) ⊂ G is a diffeomorphism, and
(b) the integration step st ∈ e(O) for all t; see Figure 3.
(A-iii) The set of feasible control actions Ut is convex and compact for each t =
0, . . . , N − 1.
Assumption 2.1 is crucial, as we shall see, in order to ensure the existence of
multipliers that appear in the necessary optimality conditions for the optimal control problem. In particular, (A-i) ensures the existence of a convex approximation (known as a tent [4]) of the feasible region in a neighborhood of an optimal triple (q̊t , x̊t , ůt ). (A-ii) gives the local representation of admissible trajecN
tories {(qt , xt )}t=0 ∈ A in a Euclidean space. This assumption naturally holds
in situations in which the discrete-time dynamics are derived from an underlying
continuous-time system (2.6), thereby transforming our optimal control problem
to a Euclidean space; first order necessary conditions for optimality are thereafter
obtained using Boltyanskii’s method of tents [4]. These first order necessary conditions are interpreted in terms of the (global) configuration space variables. (A-iii)
leads to a pointwise non-positive condition of the gradient of the Hamiltonian over
the set of feasible control actions, which is explained in detail in §3.
6
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
Te∗ Φq
a
Tq∗ G
g∗
Te Φq
w
q
e
Tq G
Φq
g
G
Figure 2. Pictorial representation of the cotangent lift of an action Φ on G.
Before defining the optimal control problem (2.7) formally, let us discuss the
geometric notions that frequently arise in this article.
Definition 2.2 ( [24, p. 173]). Let Φ : G × G → G be a left action, so Φg : G → G
for all g ∈ G. The tangent lift of Φ, T Φ : G × T G → T G is the action
(g, (h, v)) 7→ T Φg (h, v) = (Φg (h), Th Φg (v)) ,
∗
∗
∗
(h, v) ∈ Th G.
The cotangent lift of Φ, T Φ : G × T G → T G is the action
(g, (h, a)) 7→ T ∗ Φg−1 (h, a) = Φg (h), TΦ∗g (h) Φg−1 (a) ,
a ∈ Th∗ G.
In particular, if we choose g = q −1 and h = q then for a ∈ Tq∗ G,
hTe∗ Φq (a), wi := ha, Te Φq (w)i ,
Note that the map
see Figure 2.
Te∗ Φq
:
Tq∗ G
∗
for all w ∈ g.
→ g is known as the cotangent left trivialization,
Definition 2.3 ( [24, p. 311]). The Adjoint action of G on g is
G × g 3 (g, β) 7→ Adg β :=
d
ds
s=0
g esβ g −1 ∈ g.
∗
The Co-Adjoint action of G on g is the inverse dual of the Adjoint action:
G × g∗ 3 (g, a) 7→ Ad∗g−1 a ∈ g∗ ,
where
Ad∗g−1 a, β = a, Adg−1 β
for all β ∈ g, a ∈ g∗ .
Definition 2.4 ( [24, p. 124]). Let f : M → R be a function defined on a manifold
M . The derivative of the function at a point q0 ∈ M is the map
Tq0 M 3 v 7→ Df (q0 )v :=
d
dt
t=0
f (g (t)) ∈ R,
where g(t) is a path in the manifold M with g(0) = q0 and
d
dt t=0
g(t) = v.
Hereinafter we let [N ] denote the natural numbers from zero to N in increasing
order. Now we shall proceed to define the optimal control problem (2.7) in a
mathematical form, and derive the first order necessary conditions for optimality
for the optimal control problem.
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
7
Collecting the definitions from above, our optimal control problem stands as:
minimize J (q, x, u) :=
N −1
{ut }t=0
N
−1
X
ct (qt , xt , ut ) + cN (qN , xN )
t=0
qt+1 = qt st (qt , xt )
xt+1 = ft (qt , xt , ut ) for each t = 0, . . . , N − 1,
ut ∈ Ut
g
t (qt , xt ) ≤ 0 for each t = 1, . . . , N,
(q0 , x0 ) = (q̄0 , x̄0 ) ,
Assumption 2.1.
(2.8)
subject to
Our main result is the following:
N −1
Theorem 2.5 (Discrete-time PMP). Let {ůt }t=0
be an optimal controller that
N
solves the problem (2.8) with {(q̊t , x̊t )}t=0 the corresponding state trajectory. Define
the Hamiltonian
∗
[N ] × g∗ × (Rnx ) × G × Rnx × Rnu 3 (τ, ζ, ξ, q, x, u) 7→
H ν (τ, ζ, ξ, q, x, u) := νcτ (q, x, u) + ζ, e−1 (sτ (q, x))
for ν ∈ R. There exist
N −1
g
+ hξ, fτ (q, x, u)i ∈ R,
∗
∗
• an adjoint trajectory {(ζ t , ξ t )}t=0 ⊂ g∗ × (Rnx ) , covectors µt ∈ (Rnt ) for
t = 1, . . . , N, and
• a scalar ν ∈ {−1, 0}
such that, with
γ̊t := t, ζ t , ξ t , q̊t , x̊t , ůt ,
the following hold:
(MP-i)
state and adjoint system dynamics
(
ν
q̊t+1 = q̊t eDζ H (γ̊t ) ,
state
x̊t+1 = Dξ H ν (γ̊t ) ,
(
ζ t−1 = Ad∗e−Dζ H ν (γ̊t ) ζ t + Te∗ Φq̊t Dq H ν (γ̊t ) + µt Dq gt (q̊t , x̊t ) ,
adjoint
ξ t−1 = Dx H ν (γ̊t ) + µt Dx gt (q̊t , x̊t ) ,
(MP-ii) transversality conditions
ζ N −1 = Te∗ Φq̊N νDq cN (q̊N , x̊N ) + µN Dq gN (q̊N , x̊N ) ,
ξ N −1 = νDx cN (q̊N , x̊N ) + µN Dx gN (q̊N , x̊N ) ,
(MP-iii) Hamiltonian non-positive gradient condition
Du H ν t, ζ t , ξ t , q̊t , x̊t , ůt , w − ůt ≤ 0 for all
(MP-iv) complementary slackness conditions
µtj gtj (q̊t , x̊t ) = 0
for all
j = 1, . . . , nt ,
and
w ∈ Ut ,
t = 1, . . . , N,
(MP-v) non-positivity condition
µt ≤ 0
for all
t = 1, . . . , N,
(MP-vi) non-triviality condition
N −1
N
adjoint variables {(ζ t , ξ t )}t=0 , covectors {µt }t=1 , and the scalar ν do not
simultaneously vanish.
8
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
We present a proof of Theorem 2.5 in §3. This discrete-time PMP on matrix
Lie groups is a generalization of the standard discrete-time PMP on Euclidean
spaces since the variable q in the combined state (q, x) evolves on the Lie group G;
consequently, the assertions of Theorem 2.5 appear different from the discrete-time
PMP on the Euclidean spaces [4]. Let us highlight some of its features:
• The adjoint system of the discrete-time PMP on matrix Lie groups corresponding
to the states q evolves on the dual of the Lie algebra g∗ (a vector space,) despite
the fact that the state dynamics (2.6) evolves on the Lie group G.
• The adjoint system is linear in the adjoint variables (ζ t , ξ t ) because
the maps
g∗ 3 ζ t 7→ Ad∗e−Dζ H ν (γ̊t ) (ζ t ) ∈ g∗ and Tq∗ G 3 Dq H ν (γ̊t ) 7→ Te∗ Φq̊t Dq H ν (γ̊t )
are linear for all t.
• The assertions the “Hamiltonian non-positive gradient condition”, the “complementary slackness condition”, the “non-positivity condition”, and the “nontriviality condition” are identical to the discrete-time PMP on Euclidean spaces.
Remark 2.6. (a) Assumption 2.1 is not the most general set of hypotheses for
which we can derive a discrete-time PMP on matrix Lie groups. Continuous
differentiability of the functions st , ft , gt , ct , cN suffices for Theorem 2.5 to hold.
(b) Due to certain fundamental differences between discrete-time and continuoustime optimal control problems, the standard Hamiltonian maximization condition in continuous time [35, Theorem MP on p. 14] does not carry over to
the discrete-time version. For a detailed discussion, see e.g., [32, p. 199]. We
do, however, get a weaker version contained in assertion (MP-iii) of Theorem
2.5. If the Hamiltonian is concave in control u over the set of feasible control
actions Ut for each t, in addition to (A-iii), the assertion can be strengthened
to
H ν t, ζ t , ξ t , q̊t , x̊t , ůt = max H ν t, ζ t , ξ t , q̊t , x̊t , w .
w∈Ut
In the rest of this section we apply Theorem 2.5 to a class of optimal control
problems on Lie groups that frequently arise in engineering applications, and derive
the corresponding first order necessary optimality conditions.
2.1. Problem 1. Consider the version of (2.8) in which the final conditions (qN , xN )
are constrained to take values in an immersed submanifold Mfin in G × Rnx . Let
us define the optimal control problem:
minimize
−1
{ut }N
t=0
J (q, x, u) :=
N
−1
X
ct (qt , xt , ut ) + cN (qN , xN )
t=0
(2.9)
subject to
qt+1 = qt st (qt , xt )
xt+1 = ft (qt , xt , ut ) for all t = 0, . . . , N − 1,
ut ∈ Ut
gt (qt , xt ) ≤ 0 for all t = 1, . . . , N − 1,
(q0 , x0 ) ∈ (q̄0 , x̄0 ) ,
(qN , xN ) ∈ Mfin ,
Assumption 2.1.
The first order necessary conditions for optimality for (2.9) are given by:
−1
Corollary 2.7. Let {ůt }N
t=0 be an optimal controller that solves the problem (2.9)
N
with {(q̊t , x̊t )}t=0 the corresponding state trajectory. Define the Hamiltonian
∗
[N ] × g∗ × (Rnx ) × G × Rnx × Rnu 3 (τ, ζ, ξ, q, x, u) 7→
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
H ν (τ, ζ, ξ, q, x, u) := νcτ (q, x, u) + ζ, e−1 (sτ (q, x))
for ν ∈ R. Then there exist
N −1
g
9
+ hξ, fτ (q, x, u)i ∈ R,
∗
∗
• an adjoint trajectory {(ζ t , ξ t )}t=0 ⊂ g∗ × (Rnx ) , covectors µt ∈ (Rnt ) for
t = 1, . . . , N − 1, and
• a scalar ν ∈ {−1, 0}
such that, with
γ̊t := t, ζ t , ξ t , q̊t , x̊t , ůt ,
the following hold:
(i) (MP-i) holds,
(ii) transversality conditions
Tq̊∗N Φq̊−1 ζ N −1 , ξ N −1 −νD(q,x) cN (q̊N , x̊N ) ⊥ T(q̊N ,x̊N ) Mfin ,
N
(iii) (MP-iii) holds,
(iv) complementary slackness conditions
µtj gtj (q̊t , x̊t ) = 0
for all
j = 1, . . . , nt ,
and
(v) non-positivity condition
µt ≤ 0
for all
t = 1, . . . , N − 1,
t = 1, . . . , N − 1,
(vi) non-triviality condition
N −1
N −1
adjoint variables {(ζ t , ξ t )}t=0 , covectors {µt }t=1 and the scalar ν do not simultaneously vanish.
2.2. Problem 2. Consider the version of (2.8) in which the boundary conditions
are given and fixed, and the final cost and the state inequality constraints are
absent. In other words, we have the control problem
minimize
−1
{ut }N
t=0
(2.10)
subject to
J (q, x, u) :=
N
−1
X
ct (qt , xt , ut )
t=0
qt+1 = qt st (qt , xt ) ,
xt+1 = ft (qt , xt , ut ) ,
ut ∈ Ut ,
(q0 , x0 ) = (q̄0 , x̄0 ) ,
(q
N , xN ) = (q̄N , x̄N ) ,
Assumption 2.1.
for all t = 0, . . . , N − 1,
The first order necessary conditions for optimality for (2.10) are given by:
−1
Corollary 2.8. Let {ůt }N
t=0 be an optimal controller that solves the problem (2.10)
N
with {(q̊t , x̊t )}t=0 the corresponding state trajectory. Define the Hamiltonian
∗
[N ] × g∗ × (Rnx ) × G × Rnx × Rnu 3 (τ, ζ, ξ, q, x, u) 7→
H ν (τ, ζ, ξ, q, x, u) := νcτ (q, x, u) + ζ, e−1 (sτ (q, x))
for ν ∈ R. There exist
N −1
∗
• an adjoint trajectory {(ζ t , ξ t )}t=0 ⊂ g∗ × (Rnx ) and
• a scalar ν ∈ {−1, 0}
such that, with
γ̊t := t, ζ t , ξ t , q̊t , x̊t , ůt ,
the following hold:
g
+ hξ, fτ (q, x, u)i ∈ R,
10
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
(i) state and adjoint system dynamics
(
ν
q̊t+1 = q̊t eDζ H (γ̊t ) ,
state
x̊t+1 = Dξ H ν (γ̊t ) ,
(
ζ t−1 = Ad∗e−Dζ H ν (γ̊t ) ζ t + Te∗ Φq̊t Dq H ν (γ̊t ) ,
adjoint
ξ t−1 = Dx H ν (γ̊t ) ,
(ii) (MP-iii) holds,
(iii) non-triviality condition
N −1
adjoint variables {(ζ t , ξ t )}t=0 , and the scalar ν do not simultaneously vanish.
2.3. Problem 3. Consider the version of (2.8) with fixed boundary conditions,
without state inequality constraints, and without the final cost. Let us consider the
case in which the integration step st ∈ G of the discrete-time evolution is related
to the states (qt , xt ) by an implicit equation vt (st , qt , xt ) = 0 such that the map
vt ( · , qt , xt ) : Oe → Rnq is a diffeomorphism for all admissible trajectories, i.e.,
{(qt , xt )}N
t=0 ∈ A, where Oe is a neighborhood of e in G. The optimal control
problem can be defined as follows:
minimize J (s, q, x, u) :=
N −1
{ut }t=0
(2.11)
subject to
N
−1
X
ct (st , qt , xt , ut )
t=0
qt+1 = qt st
v (s , q , x ) = 0
t
t t
t
x
=
f
(s
t+1
t
t , qt , xt , ut )
ut ∈ Ut
(q0 , x0 ) = (q̄0 , x̄0 ) ,
(qN , xN ) = (q̄N , x̄N ) ,
Assumption 2.1.
for all
t = 0, . . . , N − 1,
The first order necessary conditions for optimality for (2.11) are given by:
−1
Corollary 2.9. Let {ůt }N
t=0 be an optimal controller that solves the problem (2.11)
N −1
with {(q̊t , x̊t )}N
t=0 the corresponding state trajectory and {s̊t }t=0 ⊂ G such that
vt (s̊t , q̊t , x̊t ) = 0 for t = 0, . . . , N − 1. Define the Hamiltonian function
∗
[N ] × g∗ × (Rnx ) × G × G × Rnx × Rnu 3 (τ, ζ, ξ, s, q, x, u) 7→
H ν (τ, ζ, ξ, s, q, x, u) := νcτ (s, q, x, u) + ζ, e−1 (s)
g
+ hξ, fτ (s, q, x, u)i ∈ R
for ν ∈ R. There exist
N −1
∗
• an adjoint trajectory {(ζ t , ξ t )}t=0 ⊂ g∗ × (Rnx ) , and
• a scalar ν ∈ {−1, 0}
such that, with
γ̊t := t, ζ t , ξ t ,s̊t , q̊t , x̊t , ůt ,
the following hold:
(i) state and adjoint system dynamics
(
ν
q̊t+1 = q̊t eDζ H (γ̊t ) , 0 = vt (s̊t , q̊t , x̊t ) ,
state
x̊t+1 = Dξ H ν (γ̊t ) ,
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
t−1
=
ζ
adjoint
ξ t−1 =
11
Ad∗e−Dζ H ν (γ̊t ) ζ t + Te∗ Φq̊t Dq H ν (γ̊t )
−1
−Te∗ Φq̊t Ds H ν (γ̊t ) ◦ Ds vt (s̊t , q̊t , x̊t ) ◦ Dq vt (s̊t , q̊t , x̊t ) ,
−Ds H ν (γ̊t ) ◦ Ds vt (s̊t , q̊t , x̊t )
−1
◦ Dx vt (s̊t , q̊t , x̊t ) + Dx H ν (γ̊t ) ,
(ii) Hamiltonian non-positive gradient condition
Du H ν t, ζ t , ξ t ,s̊t , q̊t , x̊t , ůt , w − ůt ≤ 0
for all w ∈ Ut ,
(iii) non-triviality condition
N −1
adjoint variables {(ζ t , ξ t )}t=0 , and the scalar ν do not simultaneously vanish.
3. Proof of the Maximum Principle (Theorem 2.5)
Sketch of proof: We present our proof through the following steps:
Step (I) We prove the existence of a local parametrization of the Lie group G and
define the optimal control problem (2.8) in local coordinates.
Step (II) First order necessary conditions for the optimal control problem defined
in local coordinates are derived using the method of tents [5].
Step (III) The first order necessary conditions derived in Step (II) are represented
in configuration space variables.
Step (IV) We prove that the first order necessary conditions derived in Step (III)
are independent of the choice of the coordinate system.
N
N −1
Henceforth {(q̊t , x̊t )}t=0 , {ůt }t=0 denotes an optimal state-action trajectory.
3.1. Step (I). Local parametrization of the Lie group G: Let us define the
following local parametrization of the Lie group G induced by the exponential map.
−1
Fact 3.1. If (A-ii) holds, then the map φqt := (Φqt ◦ e) : Qt → O ⊂ g with Qt :=
{Φqt (s) | s ∈ e(O)} and Φqt (s) := qt s for all s ∈ G, provides a unique representation
of qt+1 ∈ G on the Lie algebra g for a given qt ∈ G for all t = 0, . . . , N − 1.
φqt (qt+1 )
σ
qt st
qt+1
Qt
qt
σ −1 (ηt+1 )
0
O
φqt (qt )
g
G
ηt+1
σ (O)
0
Rnq
Figure 3. Local parametrization of qt+1 given qt .
With the help of this local parametrization, we define the dynamics evolving on
the Lie group G in local coordinates. The Lie algebra g of the matrix Lie group G
is a finite dimensional vector space [22, Theorem 8.37]. Therefore, there exists a
linear homeomorphism
σ : g → Rnq ,
where nq is the dimension of the Lie algebra. To compress the notation we define
x := (x0 , . . . , xN ) ∈ R(N +1)nx ,
u := (u0 , . . . , uN −1 ) ∈ RN nu ,
12
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
and q := (q0 , . . . , qN ) ∈ G × · · · × G .
|
{z
}
(N +1) factors
Let us define the product manifold
(N +1) factors
M :=
G
· · G}
| ·{z
(N +1) factors
z
}|
{
nu
× Rnx · · · Rnx × R
·{z
· · Rnu}
|
N factors
such that the state-action trajectory is a point on M, i.e.,
(q, x, u) = (q0 , . . . , qN , x0 , . . . , xN , u0 , . . . , uN −1 ) ∈ M.
In order to translate the optimal control problem (2.8) to a Euclidean space, we
need to define a diffeomorphism from an open subset of a Euclidean space to an
open subset of the product manifold M such that the state-action trajectories lie
in the image of that diffeomorphism.
Let us define the map
Λ 3 (β0 , . . . , βN , x, u) 7→
(3.1)
Ψ (β0 , . . . , βN , x, u) := (ψ0 (β), . . . , ψN (β), x, u) ∈ Ψ (Λ) ⊂ M,
where
nq
Λ := σ(O) · · · σ(O) ×R(N +1)nx × RN nu ⊂ R
· · Rn}q ×R(N +1)nx × RN nu ,
| ·{z
|
{z
}
(N +1) factors
(N +1) factors
ψt (β) := q̄0 eσ
−1
(β0 )
· · · eσ
−1
(βt )
for t = 0, . . . , N, and q̄0 is a fixed element in G.
Observe that the map Ψ is a smooth bijection, and the inverse map is given by
(3.2)
Ψ (Λ) 3 (α0 , α1 , . . . , αN , x, u) 7→ Ψ−1 (α0 , α1 , . . . , αN , x, u)
−1
= σ ◦ e−1 q̄0−1 α0 , σ ◦ e−1 α0−1 α1 , . . . , σ ◦ e−1 αN
−1 αN , x, u ∈ Λ.
Since the inverse map Ψ−1 is also smooth, Ψ is a diffeomorphism. We claim that
state-action trajectories corresponding to all admissible control actions and starting
at q̄0 lie in the image of Ψ.
n
oN
0
0
0
−1
Let
qt , xt
, {ut }N
be a feasible state-action trajectory starting at
t=0
t=0
(q̄0 , x̄0 ) . If the state-action trajectory lies in the image of the diffeomorphism Ψ,
then it can be represented by
0 0 0
0 0 0
0
0
0
0
q , x , u := ψ0 (η ), . . . , ψN (η ), x , u ∈ Ψ (Λ) for some
η , x , u ∈ Λ,
0
0
0
0
0
0
where η := η0 , . . . , ηN . Observe that there exist a unique η , x , u ∈ Λ with
0
0
0
η0:t := η0 , . . . , ηt
for t = 0, . . . , N − 1,
0
:= σ ◦ e−1 ◦st ψ̃t η0:t , xt
for t = 0, . . . , N − 1,
0
η0 := 0,
0
ηt+1
and
0
ψ̃t η0:t := q̄0 e
such that
Notice that
0
σ −1 η0
···e
0
σ −1 ηt
for t = 0, . . . , N − 1,
0 0 0 0 0 0
Ψ η ,x ,u = q ,x ,u .
0
ηt ∈ σ (O)
for t = 0, . . . , N,
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
13
because of the assumption (A-ii), i.e., st (qt , xt ) ∈ e (O) for all feasible pairs (qt , xt ) .
This proves the claim.
We employ the diffeomorphism Ψ to translate the optimal control problem (2.8)
N +1
to the open subset Λ as: for η := (η0 , . . . , ηN ) ∈ (Rnq )
,
minimize J˜ (η, x, u) =
N −1
{ut }t=0
N
−1
X
ct (ψt (η) , xt , ut ) + cN (ψN (η) , xN )
t=0
(3.3)
subject to
−1
σ −1 (η0 )
· · · eσ (ηt ) for t = 0, . . . , N,
t (η) := q̄0 e
ψ
−1
ηt+1 = σ ◦ e ◦st (ψt (η) , xt )
x
for t = 0, . . . , N − 1,
t+1 = ft (ψt (η) , xt , ut )
ut ∈ Ut
g
(ψt (η) , xt ) ≤ 0 for t = 1, . . . , N,
t
(η0 , x0 ) = (0, x̄0 ) .
3.2. Step (II). Necessary optimality conditions in local coordinates. In
Step (I) we have distilled the optimal control problem (3.3) from (2.8) such that
(3.3) is defined on a Euclidean space. We apply first order necessary conditions for
optimality for constrained optimal control problems on Euclidean spaces derived
via method of tents in [4] to (3.3).
Remark 3.2. Note that even though (3.3) has been distilled from (2.8) and is a
well-defined problem on a Euclidean space, the standard discrete-time PMP does
not apply to (3.3) on account of the fact that the first constraint clearly shows that
the “system dynamics” has memory. In other words the system in (3.3) is in a
non-standard form. It turns out that first order necessary conditions do lead to a
PMP for (3.3) once we lift back the necessary conditions to the configuration space.
We convert the optimal control problem (3.3) into a relative extremum problem
in a higher-dimensional Euclidean space.
Let
>
>
>
>
z := (η, x, u) = η0> , . . . , ηN
, x>
0 , . . . , xN , u0 , . . . , uN −1
(3.4)
>
be the stacked vector of states and controls corresponding to (3.3), with
z ∈ Rm such that m = (nx + nq )(N + 1) + nu N.
Let us define the admissible action set Ut in terms of z ∈ Rm as
Ωtu
nq
:= R
|
(N +1) factors
(t+1)th factor
z
}|
{
·{z
· · R } × Rnx · · · Rnx × Rnu · · ·
|
z}|{
Ut
{z
nq
(N +1) factors
N factors
· · · Rnu ⊂ Rm
}
for t = 0, . . . , N − 1, and the set of initial conditions as
N factors
nq
ΩB := {0} × R
|
z
}|
{
nu
·{z
· · R } ×{x̄0 } × Rnx · · · Rnx × R
·{z
· · Rnu} ⊂ Rm .
|
nq
N factors
N factors
We define the cost function, the dynamics, and the state constraints in terms of
−1
−1
N +1
z as: for η := (η0 , . . . , ηN ) ∈ (Rnq )
, ψt (η) := q̄0 eσ (η0 ) · · · eσ (ηt ) for t =
0, . . . , N, and for z as in (3.4),
C(z) :=
N
−1
X
t=0
ct (ψt (η) , xt , ut ) + cN (ψN (η) , xN ) ,
14
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
St (z) := −ηt+1 + σ ◦ e−1 ◦ st (ψt (η) , xt ) ,
Ft (z) := −xt+1 + ft (ψt (η) , xt , ut ) ,
Gjt (z) := gtj (ψt (η) , xt )
for j = 1, . . . , nt .
The optimal control problem (3.3) can now be lifted to the augmented space Rm
as
minimize C(z)
z
j
St (z) = 0 for t = 0, . . . , N − 1, j = 1, . . . , nq ,
j
Ft (z) = 0 for t = 0, . . . , N − 1, j = 1, . . . , nx ,
(3.5)
subject to
Gjt (z) ≤ 0 for t = 1, . . . , N, j = 1, . . . , nt ,
N −1
T t
z ∈
Ωu ∩ ΩB .
t=0
Let us define
z̊ := Ψ(q̊, x̊, ů) ∈ Rm ,
where (q̊, x̊, ů) ∈ M is an optimal state-action trajectory. Then the point z̊ is a
solution of the optimization problem (3.5). The necessary conditions for optimality
for an optimization problem are defined in terms of dual cones. Before deriving the
necessary conditions for optimality for the optimization problem (3.5), we provide
a quick revision to cones and their dual cones.
Recall that the dual cone K + of a cone ‡ K in Rm is a convex cone such that
every vector in K + makes an obtuse angle with every element of K, i.e.,
K + := {ρ ∈ Rm | hρ, zi ≥ 0 for all z ∈ K} .
A set K (z̊) in Rm is a cone with apex at z̊ if for every z ∈ K (z̊) , z̊ + α (z − z̊) ∈
K (z̊) for all α ≥ 0. Subsequently, the dual cone K + (z̊) for the cone K (z̊) is defined
as [4, p. 8]
K + (z̊) := {a ∈ Rm |ha, z − z̊i ≥ 0, z ∈ K (z̊) } .
(3.6)
Clearly, K + (z̊) is a closed convex cone. Equipped with the notions of dual cones,
we state a theorem of Boltyanskii that provides necessary conditions for optimality
for an optimization problem. This theorem is a key result for deriving necessary
conditions for optimality for the optimization problem (3.5).
Theorem 3.3 ( [4, Theorem 18, p. 22]). Let Ω1 , . . . , Ωs be convex subsets of Rm ,
and let C be a real-valued smooth function whose domain contains the set
Σ = Ω1 ∩ . . . ∩ Ωs ∩ Ω∗ ,
where
Ω∗ := z ∈ Rm | F1 (z) = 0, . . . , Freq (z) = 0, f1 (z) ≤ 0, . . . , friq (z) ≤ 0 ,
and Fi , fj are real valued smooth functions for all i, j. Let z̊ ∈ Σ and let Ki (z̊) be
the support cone§ of Ωi with apex at z̊. If C attains its minimum at z̊ relative to Σ,
r
r
then there exist scalars ν, ηF1 , . . . , ηFeq , ηf1 , . . . , ηfiq , and vectors aj ∈ Kj+ (z̊) (dual
cone of Kj (z̊)) for j = 1, . . . , s, satisfying the following conditions:
‡A set K ⊂ Rm is a cone if for every z ∈ K, αz ∈ K for all α ≥ 0.
§The support cone K (z̊) of a convex set Ω ⊂ Rm with apex at z̊ ∈ Ω is defined as
!
K (z̊) := closure
[
α>0
For more details see [5, p. 29].
{z̊ + α (z − z̊) | z ∈ Ω}
.
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
r
15
r
(i) ν ≤ 0, and if ν = ηF1 = · · · = ηFeq = ηf1 = · · · = ηfiq = 0, then at least one of
the vectors aP
1 , . . . , as is not zero;
Priq j
req
(ii) νDz C(z̊) + i=1
ηFi Dz Fi (z̊) + j=1
ηf Dz fj (z̊) + a1 + . . . + as = 0;
(iii) for every j = 1, . . . , s we have
ηfj ≤ 0,
and
ηfj fj (z̊) = 0.
To apply Theorem (3.3) to the optimization problem (3.5), we define support
−1
cones of the convex sets Ω0u , . . . , ΩN
, ΩB at z̊. Let us derive the support cone
u
t
t
Ku (z̊) of the set Ωu with apex at z̊ ∈ Ωtu for t = 0, . . . , N − 1 as
!
[
Kut (z̊) := closure
z̊ + α (z − z̊) | z ∈ Ωtu
,
α>0
(N +1) factors
(N +1) factors
where Ut :=
(t+1)th factor
}|
{
z
· · Rn}q × Rnx · · · Rnx × Rnu · · ·
= closure |Rnq ·{z
|
z}|{
Ut
{z
N factors
· · · Rnu ,
}
S
{ůt + α (ut − ůt ) | ut ∈ Ut } is the collection of rays emanating
α>0
from ůt and passing through ut ∈ Ut . Consequently, the support cone Kut (z̊) is
defined by
(3.7)
Kut (z̊)
nq
=R
|
(N +1) factors
(t+1)th factor
z
}|
{
·{z
· · R } × Rnx · · · Rnx × Rnu · · ·
|
z }| {
Qtu (ůt )
{z
nq
(N +1) factors
N factors
where
· · · Rnu ⊂ Rm ,
}
!
Qtu
[
(ůt ) := closure
α>0
{ůt + α (ut − ůt ) | ut ∈ Ut }
is the support cone of the convex set Ut with apex at ůt . Note that the set ΩB is
a affine subspace passing through z̊, and therefore the support cone KB (z̊) of the
set ΩB with apex at z̊ is the set ΩB itself, i.e.,
KB (z̊) := ΩB .
(3.8)
Subsequently, the dual cone of the support cone KB (z̊) is given by
+
KB
(z̊) := {ρ ∈ Rm | hρ, z − z̊i ≥ 0 for all z ∈ KB (z̊)} ,
= {ρ ∈ Rm | hρ, vi ≥ 0 for all v ∈ KB } ,
where
N factors
z
}|
{
nx
nx
nu
nq
nq
KB := {0} × R
·
·
·
R
×{0}
×
R
·
·
·
R
×R
·{z
· · Rnu} .
|
{z
}
|
N factors
N factors
+
Note that KB is a subspace of Rm . Therefore, the dual cone KB
(z̊) is the orthogonal complement of the subspace KB , i.e.,
2N factors
+
KB
nq
(z̊) = R
nx
× {0} · · · {0} ×R
| {z }
z }| {
× {0} · · · {0} ⊂ Rm .
N factors
Now we are in the position to apply Theorem (3.3) to the optimization problem
(3.5).
By Theorem 3.3, if the function C attains its minimum at the point z̊ then there
exist scalars µtj for j = 1, . . . , nt and t = 1, . . . , N, λtj for j = 1, . . . , nq and t =
16
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
+
1, . . . , N − 1, ξjt for j = 1, . . . , nx and t = 1, . . . , N − 1, vectors b ∈ KB
(z̊) , at ∈
t +
(Ku ) (z̊) for t = 0, . . . , N −1, and a scalar ν ∈ R satisfying the following conditions:
(Opt-i) ν ≤ 0, and if ν = 0, µtj = 0 for j = 1, . . . , nt and t = 1, . . . , N, λtj = 0
for j = 1, . . . , nq and t = 1, . . . , N − 1, ξjt = 0 for j = 1, . . . , nx and
t = 1, . . . , N − 1, then at least one of the vectors a1 , . . . , aN −1 , b is not
zero;
PN −1 Pnx t
PN −1 Pnq t
j
j
(Opt-ii) νDz C(z̊) + t=0
j=1 λj Dz St (z̊) +
j=1 ξj Dz Ft (z̊)
t=0
PN Pnt t
+ t=1 j=1 µj Dz Gjt (z̊) + a0 + · · · + aN −1 + b = 0;
(Opt-iii) for every j = 1, . . . , nt we have
µtj Gjt (z̊) = 0
µtj ≤ 0,
for all t = 1, . . . , N.
Note that the family of cones Ku0 (z̊), · · · , KuN −1 (z̊), KB (z̊) are inseparable in Rm [4,
Theorem 7 on p. 10].¶ So, using the inseparability of this family, we arrive at more
strict conditions for optimality than (Opt-i)-(Opt-ii). To compress notation, let us
define covectors (row vectors) as
∗
for t = 1, . . . , N,
(Rnt ) 3 µt := µt1 , · · · , µtnt
t
t
nx ∗
t
for t = 0, . . . , N − 1,
(R ) 3 λ := λ0 , · · · , λnq
nq ∗
t
(R ) 3 ξ := ξ0t , · · · , ξnt x
for t = 0, . . . , N − 1.
Then the optimality conditions (Opt-i)-(Opt-ii) translates to the following:
(i) Suppose that the covectors µ1 , . . . , µN , λ0 , . . . , λN −1 , ξ 0 , . . . , ξ N −1 , and the scalar
ν are all zero. Then the optimality conditions (Opt-i)-(Opt-ii) leads to the following: there exist vectors a0 , . . . , aN −1 , b not all zero such that
a0 + · · · + aN −1 + b = 0.
That contradicts inseparability of the family of cones Ku0 (z̊), · · · , KuN −1 (z̊),
KB (z̊), and hence,
a0 = · · · = aN −1 = b = 0.
Therefore (Opt-i) leads to a stronger non-triviality condition:
(3.9)
ν ≤ 0, and if ν = 0, then at least one of the covectors
µ1 , . . . , µN , λ0 , . . . , λN −1 , ξ 0 , . . . , ξ N −1 is not zero.
(ii) Further, using the fact that
ha0 + · · · + aN −1 + b, z̃i ≥ 0 for all z̃ + z̊ ∈ Ku (z̊),
N −1
T t
where Ku (z̊) :=
Ku (z̊) ∩ KB (z̊); the (Opt-ii) is rewritten as
t=0
*
(3.10)
νDz C(z̊) +
N
X
i=0
i
λ Dz Si (z̊) +
N
X
k=0
k
ξ Dz Fk (z̊) +
N
X
t=1
+
t
µ Dz Gt (z̊), z̃
≤0
for all z̃ + z̊ ∈ Ku (z̊).
The variational inequality (3.10) gives rise to ‘state and adjoint system dynamics’
(MP-i), ‘transversality conditions’ (MP-ii) and ‘Hamiltonian non-positive gradient
condition’ (MP-iii). The optimality condition (Opt-iii) establishes ‘complementary
¶A family of convex cones is said to be separable if there exists a hyperplane that separates
one of them from the intersection of the others. If the family is not separable then it is called
inseparable. For more details see [18, p. 2575].
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
17
slackness conditions’ (MP-iv) and ‘non-positivity conditions’ (MP-v), and the nontriviality condition (MP-vi) comes from (3.9). Now we shall derive these conditions
(MP-i)-(MP-vi) in the configuration variables.
3.3. Step (III). Representation of the necessary conditions for optimality
in terms of the configuration variables: Let us define the Hamiltonian as
∗
[N ] × g∗ × (Rnx ) × G × Rnx × Rnu 3 (τ, ζ, ξ, q, x, u) 7→
H ν (τ, ζ, ξ, q, x, u) := νcτ (q, x, u) + ζ, e−1 (sτ (q, x))
g
+ hξ, fτ (q, x, u)i ∈ R,
with ν ∈ R.
• Adjoint system dynamics and transversality conditions (MP-i)-(MP-ii):
In order to derive the adjoint system dynamics and the transversality conditions
corresponding to the state variable x, we restrict the choice of the variable z̃ in
(3.10) to
(N +1) factors
(3.11)
Kxt (z̊)
z
:= {0} · · · {0} × {0} · · ·
| {z }
(N +1) factors
}|
nx
R
|{z}
(t+1)th factor
{
· · · {0} × {0} · · · {0} ⊂ Ku (z̊)
| {z }
N factors
for t = 1, . . . , N. For a fix t ∈ [N ], choosing the collection of z̃ in (3.10) such that
z̃ + z̊ ∈ Kxt (z̊), leads to the following set of conditions: for
ζ t := σ ∗ λt
and γ̊t := t, ζ t , ξ t , q̊t , x̊t , ůt ,
– adjoint equations corresponding to the state variables x :
ξ t−1 = Dx H ν (γ̊t ) + µt Dx gt (q̊t , x̊t )
for t = 1, . . . , N − 1,
– transversality conditions corresponding to the state variables x :
ξ N −1 = νDx cN (q̊N , x̊N ) + µN Dx gN (q̊N , x̊N ) .
Similarly, to derive the adjoint and transversality conditions corresponding to
the state variable q, we restrict the choice of the variable z̃ in (3.10) such that
z̃ + z̊ ∈ Kq (z̊) ,
where
(N +1) factors
(3.12)
nq
Kq (z̊) := {0} × R
|
z }| {
·{z
· · R } × {0} · · · {0} × {0} · · · {0} .
| {z }
nq
N factors
N factors
The choice of z̃ in (3.10) such that z̃+z̊ ∈ Kq (z̊) , leads to the following condition:
(3.13)
νDq cN (q̊N , x̊N ) Dη ψN (η̊) η̃ +
−
N
−1
X
t=0
λt η̃t+1 +
N
−1
X
t=0
N
X
t=1
µt Dq gt (q̊t , x̊t ) Dη ψt (η̊) η̃
Dq H ν (γ̊t ) Dη ψt (η̊) η̃ = 0
for all η̃ ∈ Kη ,
where
η̃ := (η̃0 , η̃1 , . . . , η̃N )
and Kη := {0} × Rnq × · · · × Rnq .
|
{z
}
N factors
In order to represent (3.13) in configuration variable, let us establish an association between η̃ and tangent vectors q̃t ∈ Tq̊t G for t = 0, . . . , N . It is evident
from the diffeomorphism (3.1) that
(3.14)
q̃t := Dη ψt (η̊) η̃ ∈ Tq̊t G for all t = 0, . . . , N.
18
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
Furthermore, we derive η̃ in terms of q̃ := (q̃0 , . . . , q̃N ) via (3.2) as follows
d
d
(3.15) η̃t :=
ηt (s) :=
σ ◦ e−1 qt−1 (s)−1 qt (s)
for t = 0, . . . , N,
ds s=0
ds s=0
where q−1 (s) ≡ q̄0 , ηt (0) = η̊t , and simplifies to
(3.16) η̃0 = σ (χ0 ) and η̃t = −σ Adq̊−1 q̊t−1 (χt−1 ) − χt
for
t
t = 1, . . . , N,
with χt := Tq̊t Φq̊−1 (q̃t ) ∈ g. Using the fact that
t
q̃t = Te Φq̊t (χt )
for t = 0, . . . , N,
and with (3.14) and (3.16), the necessary conditions (3.13) leads to the following
conditions on the Lie algebra g as
N
−1
−1
NX
X
(3.17)
σ ∗ λt−1 Adq̊−1 q̊t−1 (χt−1 ) − χt +
Te∗ Φq̊t Dq H ν (γ̊t ) χt
t
t=1
t=0
N
X
+ Te∗ Φq̊N νDq cN (q̊N , x̊N ) χN +
Te∗ Φq̊t µt Dq gt (q̊t , x̊t ) χt = 0
t=1
for all χ ∈ Kχ , where
χ := (χ0 , χ1 , . . . , χN )
and Kχ := {0} × g × · · · × g .
| {z }
N factors
Let us now restrict the choice of the variable χ in (3.17) to
(N +1) factors
(3.18)
Kχt
z
:= {0} · · ·
}|
g
|{z}
{
· · · {0} ⊂ Kχ
for t = 1, . . . , N,
(t+1)th factor
to derive the adjoint dynamics and the transversality conditions corresponding
to the states q ∈ G. For a fixed t ∈ [N ], choosing χt ∈ Kχt in (3.17) leads to the
following set of conditions:
– adjoint system corresponding to the states q :
(3.19)
ζ t−1 = Ad∗e−Dζ H ν (γ̊t ) ζ t + Te∗ Φq̊t Dq H ν (γ̊t ) + µt Dq gt (q̊t , x̊t ) ,
– transversality conditions corresponding to the states q :
(3.20)
ζ N −1 = Te∗ Φq̊N νDq cN (q̊N , x̊N ) + µN Dq gN (q̊N , x̊N ) .
• Hamiltonian non-positive gradient condition (MP-iii): Hamiltonian nonpositive gradient condition is derived via restricting the choice of the variable z̃
in (3.10) such that
z̃ + z̊ ∈ K̃ut (z̊) ,
where
(t+1)th factor
(N +1) factors
z }| {
(3.21) K̃ut (z̊) := {0} · · · {0} × {0} · · · {0} × {0} · · ·
| {z }
|
(N +1) factors
z }| {
Qtu (ůt )
{z
N factors
· · · {0} ⊂ Ku (z̊)
}
for t = 1, . . . , N − 1. For a fixed t ∈ [N ], choosing the collection of z̃ in (3.10)
such that z̃ + z̊ ∈ K̃ut (z̊), leads to the following set of conditions:
(3.22)
hDu H ν (γ̊t ) , ũt i ≤ 0
for all
ůt + ũt ∈ Qtu (ůt ),
for t = 0, . . . , N − 1. Since, Ut ⊂ Qtu (ůt ) , (3.22) simplifies to the following
hDu H ν (γ̊t ) , w − ůt i ≤ 0
for all
w ∈ Ut .
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
19
• Complementary slackness conditions (MP-iv): The complementary slackness conditions in the configuration variables follows from (Opt-iii) as:
µtj gtj (q̊t , x̊t ) = 0
for all j = 1, . . . , nt
and t = 1, . . . , N.
• Non-positivity condition (MP-v): The non-positivity condition follows from
(Opt-iii) as:
µt ≤ 0 for all t = 1, . . . , N.
• Non-triviality condition (MP-vi): The variable λt in the non-triviality condi∗
tion (Opt-i) is zero if and only if ζ t := σ ∗ (λt ) is zero because σ ∗ : (Rnq ) → g∗
is a vector space homeomorphism.
N −1
N
The adjoint variables {(ζ t , ξ t )}t=0 , covectors {µt }t=1 and the scalar ν do not
simultaneously vanish.
• State dynamics (MP-i): The system dynamics (2.6) in terms of the Hamiltonian is
q̊t+1 = q̊t eDζ H
(3.23)
ν
(γ̊t )
x̊t+1 = Dξ H ν (γ̊t ) .
,
3.4. Step (IV). The representation of the necessary conditions for optimality is coordinate free: Let us establish that the necessary conditions derived
in §3.3 are independent of the choice of the coordinate system. Suppose we have a
curve
I 3 s 7→ qt (s) ∈ G with qt (0) = q̊t ,
and the two different coordinate charts
X : DX ⊂ G → RX ⊂ Rnq ,
Y : DY ⊂ G → RY ⊂ Rnq .
such that our curve qt is contained in both domains, that is
qt (s) ∈ DX ∩ DY
1
for all s ∈ I.
Suppose our two coordinate charts are C related, in the following sense:
Definition 3.4 ( [36, Definition 2.3.1, p. 8]). Let G be a manifold, and let X , Y,
be two nq − dimensional charts on G. Let k be a nonnegative integer. We say that
X and Y are C k related if
(1) the images X DX ∩ DY , Y DX ∩ DY of the “overlap set” DX ∩ DY
under the coordinate maps X , Y, are open in RX , RY , respectively;
(2) the “change of coordinates” maps
Φ : X DX ∩ DY → Y DX ∩ DY ,
Ψ : Y DX ∩ DY → X DX ∩ DY ,
defined by the conditions
Φ(X) = Y
whenever
Ψ(Y ) = X whenever
are of class C k .
X = X (q), Y = Y(q)
X = X (q), Y = Y(q)
for some
for some
q ∈ DX ∩ DY ,
q ∈ DX ∩ DY ,
We know that X, Y are coordinate representations of the configuration q ∈ DX ∩
D and Ẋ, Ẏ are coordinate representations of the tangent vectors q̇ ∈ Tq G in
the coordinate charts X , Y respectively. Then we can express the tangent vector
Ẏ (or Ẋ) in one coordinate chart as a function of the configuration X(or Y ) and the
tangent vector Ẋ(or Ẏ ) in the other coordinate chart. The transformation rules
for tangent vectors is given by the following formulae:
∂Φ
∂Ψ
(3.24)
(Y )Ẏ , Ẏ =
(X)Ẋ,
Ẋ =
∂Y
∂X
Y
20
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
where
∂Ψ
∂Y
(Y )
ij
:=
∂Ψi
∂Yj (Y
) is the Jacobian matrix.
In order to show the coordinate invariance of the adjoint system and the transversality conditions, we need to establish
the transformation rules for the covectors
ζ t−1 , Ad∗e−Dζ H ν (γ̊t ) ζ t and Te∗ Φq̊t Dq H ν (γ̊t ) ∈ g∗ , and under that transformation
rule these covectors admit unique representations in the coordinate charts X , Y.
This task is accomplished in the following steps:
(1) We prove that the local representation Ẋ, Ẏ of q̇ ∈ Tq G in the coordinate
charts X , Y respectively, has a unique representation χX , χY ∈ g, the Lie
algebra of the Lie group G via the tangent lift of the left action. Further,
we establish the transformation rule for the vectors χX , χY that is derived
via the transformation rule of vectors Ẋ, Ẏ given by (3.24).
(2) Through the duality property, we derive the transformation rule for a local
representation ζX , ζY in the coordinate charts X , Y respectively, of the
covector ζ ∈ g∗ and establish the invariance of the adjoint equations and
transversality conditions.
In case of a Lie group G, the tangent space at a point q̊t ∈ G is characterized by
the left (right) invariant vector fields [34], i.e.,
Tq̊t G = Te Φq̊t (χ) χ ∈ g .
This representation of the tangent space Tq̊t G admits a linear homeomorphism with
the tangent space represented in coordinate charts X , Y. Let πX , πY be the linear
homeomorphisms such that
(3.25)
Te Φq̊t χX = πX Ẋ , Te Φq̊t χY = πY Ẏ .
Then the transformation rule for the vectors χX , χY ∈ g is derived by substituting
(3.25) in (3.24) as
(3.26)
χX = Γt χY , χY = ∆t χX
where
Γt :=
and
−1
(Te Φq̊t )
◦ πX ◦
∂Ψ
(Y ) ◦ πY−1 ◦ Te Φq̊t
∂Y
∂Φ
−1
(X) ◦ πX ◦ Te Φq̊t
∆t := (Te Φq̊t ) ◦ πY ◦
∂X
are linear transformations. These transformation rules of vectors in the Lie algebra
g induces the transformation rules for the corresponding covectors, i.e., for ζX , ζY ∈
g∗ , if
ζX , χX = ζY , ∆t χX
for all χX ∈ g,
then
−1
ζX = ∆∗t (ζY ) and ζY = Γ∗t (ζX ) .
(3.27)
Using the transformation rule (3.27), it follows that
(3.28)
∗
ν
t
∗
∗
ν
t
Te Φq̊t DqX H (γ̊t )+µ gt (q̊t , x̊t ) = ∆t Te Φq̊t DqY H (γ̊t )+µ gt (q̊t , x̊t )
,
where Te∗ Φq̊t DqX H ν (γ̊t )+µt gt (q̊t , x̊t ) and Te∗ Φq̊t DqY H ν (γ̊t )+µt gt (q̊t , x̊t )
are representations of the vector Te∗ Φq̊t Dq H ν (γ̊t ) + µt gt (q̊t , x̊t ) ∈ g∗ in charts
X and Y respectively. In the similar manner, it can be established that
t
(3.29)
Ad∗−DζX H ν (γ̊t ) ζX
= ∆∗t Ad∗−DζY H ν (γ̊t ) ζYt
e
e
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
21
where
E
E D
E D
D
t
t
Ad∗−DζX H ν (γ̊t ) ζX
, χX = Ad∗s−1 ζX
, χX = Ad∗−DζY H ν (γ̊t ) ζYt , χY .
e
e
t
t
Furthermore, we derive the transformation rules for the covectors ζYt , ζX
∈ g as
E
E D
D
t
ζX
, Ade−DζX H ν (γ̊t ) χX = ζYt , Adq̊−1 q̊t ∆t χX
t−1
E
D
t
= ζY , ∆t+1 Adq̊−1 q̊t χX
t−1
E
D
∗
t
for all χX ∈ g,
= ∆t+1 ζY , Ade−DζX H ν (γ̊t ) χX
which leads to
t
ζX
= ∆∗t+1 ζYt .
(3.30)
Using (3.28), (3.29) and (3.30), we conclude that the adjoint system (3.19) in coordinate charts X transform naturally to the adjoint system (3.19) in coordinate
charts Y, i.e.,
t−1
t
ζX
− Ad∗−DζX H ν (γ̊t ) ζX
− Te∗ Φq̊t DqX H ν (γ̊t ) + µt DqX gt (q̊t , x̊t )
e
∗
t−1
∗
t
∗
ν
t
= ∆t ζY − Ad −DζY H ν (γ̊t ) ζY − Te Φq̊t DqY H (γ̊t ) + µ DqY gt (q̊t , x̊t ) .
e
In other words,
t−1
t
− Te∗ Φq̊t DqX H ν (γ̊t ) + µt DqX gt (q̊t , x̊t ) = 0
ζX
− Ad∗−DζX H ν (γ̊t ) ζX
e
if and only if
ζYt−1 − Ad∗−DζY H ν (γ̊t ) ζYt − Te∗ Φq̊t DqY H ν (γ̊t ) + µt DqY gt (q̊t , x̊t ) = 0
e
∆∗t
is an invertible linear transformation. Similarly, we can conclude that
because
the transversality conditions (3.20) in coordinate charts X transform naturally to
the transversality conditions (3.20) in coordinate charts Y. This completes our
proof.
4. Example
For illustration of our results we pick an example of energy optimal single axis
maneuvers of a spacecraft. In effect, the state-space becomes R × SO(2), which
is isomorphic to R × S1 . This is a considerably elementary situation compared to
general rigid body dynamics on SO(3), but it is easier to visualize and represent
trajectories with figures. We adhere to this simpler setting in order not to blur the
message of this article while retaining the coordinate-free nature of the problem. Let
h > 0 be a step length, Rt , Ft ∈ SO(2),(the set of 2×2 orthonormal matrices,) be the
orientation of the spacecraft performing single axis maneuvers and the integration
step at the discrete-time instant t respectively. Let ωt ∈ R be the momentum
of the spacecraft about the axis of rotation and ut be the control applied to the
spacecraft about the rotation axis. Consider the discrete-time model of a spacecraft
performing single axis maneuvers as
(
Rt+1 = Rt F (ωt ),
(4.1)
ωt+1 = ωt + hut ,
where
√
F (ω) :=
1 − h2 ω 2
hω
√ −hω
.
1 − h2 ω 2
22
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
The optimal control problem is to maneuver a spacecraft from a fixed initial
configuration (Ri , ωi ) to a fixed final configuration (Rf , ωf ) via a minimum energy
path obeying state and action constraints simultaneously, i.e., |ωt | ≤ d, |ut | ≤ c for
all t. The optimal control problem in discrete-time can be defined as
minimize
ut ∈[−c,c]
(4.2)
subject to
J (u) :=
N
−1
X
t=0
u2t
2
Rt+1 = Rt F (ωt )
ωt+1 = ωt + hut for t = 0, . . . , N − 1,
|ut | ≤ c
1
2
2
≤ 0 for t = 1, . . . , N − 1,
2 ωt − d
(R0 , ω0 ) = (Ri , ωi ) ,
(RN , ωN ) = (Rf , ωf ) .
A set of first order necessary conditions for optimality for (4.2) is given by
Corollary 2.8 as follows:
4.1. First Order Necessary Conditions of optimality: Let ( ˆ· ) : R → so(2)∗
be a vector space homeomorphism. Define the Hamiltonian for the optimal control
problem (4.2) as
(4.3)
so(2)∗ × R × SO(2) × R × R 3 ζ̂, ξ, R, ω, u 7→
E
u2 D −1
H ν (ζ̂, ξ, R, ω, u) := ν
+ ζ̂, e (F (ω)) + ξ (ω + hu)
2
g
2
u
=ν
+ ζ sin−1 (hω) + ξ (ω + hu) ∈ R
2
−1
Let {ůt }N
control that solves the problem(4.2). Then there
t=0 be an optimal
n
oN −1 n
oN
exist a state-adjoint trajectory
ζ̂ t , ξ t
, R̊t , ω̊t
on the cotangent
t=0
t=0
N
bundle so(2)∗ × R∗ × SO(2) × R, covectors {µt }t=1 ⊂ R∗ and a scalar ν ∈ {−1, 0},
not all zero such that:
(i) State and adjoint system dynamics
>
ζ̂ t−1 = Ad∗−Dζ̂ H ν (ζ̂t ,ξt ,R˚t ,ω˚t ,u˚t ) ζ̂ t = F (ω̊t ) ζ̂ t F (ω̊t ) = ζ̂ t ,
e
t−1
ξ
= Dω H ν ζ̂ t , ξ t , R̊t , ω̊t , u˚t + µt Dω gt (ω̊t )
hζ t
=p
+ ξ t + µt ω̊t .
2
2
1 − h ω̊t
So, the state and adjoint system can be written as
ζ t−1 = ζ t ,
(4.4)
adjoint
hζ t
ξ t−1 = √ 2 2 + ξ t + µt ω̊t ,
1−h ω
˚t
(
R̊t+1 = R̊t F (ω̊t ) ,
(4.5)
state
ω̊t+1 = ω̊t + h˚
ut ,
(ii) complementary slackness conditions
(4.6)
µt ω̊t 2 − d2 = 0 for all t = 1, . . . , N − 1,
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
23
(iii) non-positivity condition
µt ≤ 0
for all
t = 1, . . . , N − 1,
(iv) Hamiltonian maximization pointwise in time
H ν ζ̂ t , ξ t , R̊t , ω̊t , u˚t := max H ν ζ̂ t , ξ t , R̊t , ω̊t , w .
w∈[−c,c]
Remark 4.1. Note that since, the Hamiltonian is concave in u the non-positive gradient condition, in this case, leads to the maximization of the Hamiltonian pointwise
in time.
It follows that
u˚t = arg max H ν ζ̂ t , ξ t , R̊t , ω̊t , w ,
w∈[−c,c]
= arg max ν
w∈[−c,c]
If ν = −1 then
(4.7)
c
u˚t = −c
t
hξ
w2
+ ξ t hw.
2
if hξ t ≥ c,
if hξ t ≤ −c,
elsewhere.
In case of an abnormal extremal, i.e., ν = 0,
(
c
if ξ t > 0,
(4.8)
u˚t =
−c
if ξ t < 0.
Remark 4.2. Note that the optimal control u˚t is the saturation function of the
co-state hξ t when ν = 1, and the control is bang-bang in the case of an abnormal
extremal, i.e., ν = 0.
The constrained boundary value problem (4.4)-(4.5) subject to boundary conditions (R0 , ω0 ) = (Ri , ωi ) , (RN , ωN ) = (Rf , ωf ) , the complementary slackness
conditions (4.6) and the state constraints |ωt | ≤ d for t = 1, . . . , N − 1, is solved
using multiple shooting methods [30].
Assume that a satellite has an inertia I = 800 kgm2 about the axis of rotation,
and is fitted with an actuation device capable of producing torque of the magnitude
20 Nm. The attitude maneuvers of the satellite are subject to a maximum permissible magnitude of the momentum of 70 Nms. The model of the satellite scaled to
unit inertia has been considered for the simulations with the following data:
•
•
•
•
sampling time (T ) = 0.05 s,
maximum torque or control bound (c) = 25 mNm,
maximum momentum (d) = 87.5 mNms,
time duration (tmax ) can range between 0 s and 150 s.
For ease of representation of the initial and final orientations in figures, we denotes
the initial and final orientations with the rotation angle θ such that
cos(θt ) sin(θt )
Rt :=
for θt ∈ [0, 2π[ .
− sin(θt ) cos(θt )
So, the initial and final configurations for the trajectories are defined by (θi , ωi ) and
(θf , ωf ) respectively. Three maneuvers with different initial and final conditions
have been simulated:
• T1 : (θi , ωi )= (0◦ , 0 Nms), (θf , ωf ) = (90◦ , 80 mNms), tf = 100 s,
24
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
• T2 : (θi , ωi )= (0◦ , 0 Nms), (θf , ωf ) = (75◦ , 0 Nms), tf = 19 s,
• T3 : (θi , ωi )= (90◦ , 0 Nms), (θf , ωf ) = (265◦ , 0 Nms), tf = 40 s.
ω = 0.0875
T1
θ = 90◦
θ = 0◦
ω=0
(a) Energy optimal trajectory
Orientation [θ]
ω
Control [mNms]
The distinguishing feature of this approach is that the system dynamics is defined
in the configuration space in contrast to the local representation that enables one to
find optimal trajectories which need more than one chart for local representation,
as shown in Figure 4(a). The trajectory T1 admits all orientations on SO(2), and
these can’t be represented on a single chart. So, the local representation of the
system dynamics cannot characterize such optimal trajectories.
5
0
0
20
40
60
80
100
0
20
40 60
Time [s]
80
100
400
300
200
100
0
(b) Optimal control profile
Figure 4. The state trajectory and the corresponding optimal
control for the maneuver T1 .
The trajectories T2 , T3 are shown in Figure 5(a)-6(a), and their corresponding
optimal control profiles are shown in Figure 5(b)-6(b). It is important to note that
whenever the state constraints are active i.e. |ωt | = d, where t ∈ [N ], the control
actions at such time instances will be zero. The optimal control corresponding to
the maneuver T2 saturates at the end points in order to achieve the maneuver in
a specified time, see Figure 5(b). On the other hand, the optimal control corresponding to trajectories T1 , T3 does not saturate because the time duration of these
maneuvers is higher then the minimum time needed for such maneuvers as shown
in Figure 6(b).
ω = 0.0875
T2
θ = 90◦
◦
θ=0
0
−20
0
5
0
5
10
15
20
10
15
Time [s]
20
80
60
40
20
0
ω=0
(a) Energy optimal trajectory
25
20
Orientation [θ]
ω
Control [mNms]
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
(b) Optimal control profile
ω = 0.0875
T3
◦
θ = 90
θ = 0◦
ω=0
(a) Energy optimal trajectory
Orientation [θ]
ω
Control [mNms]
Figure 5. The state trajectory and the corresponding optimal
control for the maneuver T2 .
20
0
−20
0
10
0
10
20
30
40
20
30
Time [s]
40
300
200
100
(b) Optimal control profile
Figure 6. The state trajectory and the corresponding optimal
control for the maneuver T3 .
Remark 4.3. The conjunction of discrete mechanics and optimal control (DMOC)
for solving constrained optimal control problems while preserving the geometric
properties of the system has been explored in [28]. The indirect geometric optimal
control technique employed in our present article differs from the aforementioned
DMOC technique on the account that ours is an indirect method; consequently [38],
the proposed technique is likely to provide more accurate solutions than the DMOC
technique. Another important feature of our PMP is that it can characterize abnormal extremals unlike DMOC and other direct methods. It would be interesting
to develop an indirect method for solving optimal control problems for higher-order
geometric integrators.
26
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
Acknowledgments
This work was partially supported by a research grant (14ISROC010) from the
Indian Space Research Organization. The authors acknowledge the fruitful discussions with Harish Joglekar, Scientist, of the Indian Space Research Organization.
Appendix A. Proofs of Corollaries 2.7 - 2.9.
A.1. Proof of Corollary 2.7.
Proof. In order to apply Theorem 2.5 to optimal control problem (2.9), let us define
the boundary constraint (qN , xN ) ∈ Mfin in a neighborhood of (q̊N , x̊N ) ∈ Mfin as
the zero level set of a smooth submersion. In other words, there exists an open set
O(q̊N ,x̊N ) ⊂ G × Rnx containing (q̊N , x̊N ) and a smooth submersion
O(q̊N ,x̊N ) 3 (qN , xN ) 7→ bfin (qN , xN ) ∈ RnG +nx −mf
such that
b−1
fin (0) = O(q̊N ,x̊N ) ∩ Mfin ,
where mf is the dimension of the manifold Mfin and the Lie group G is embedded
in RnG . So, the end point constraint is represented in the neighborhood O(q̊N ,x̊N )
of the optimal point (q̊N , x̊N ) as
(A.1)
bfin (qN , xN ) = 0.
Let us define the equality constraint (A.1) in the form of inequality constraints
gN (qN , xN ) ≤ 0,
(A.2)
where
O(q̊N ,x̊N ) 3 (qN , xN ) 7→ gN (qN , xN ) :=
−bfin (qN , xN )
bfin (qN , xN )
∈ R2(nG +nx −mf ) .
It is evident from A.2 that gN (qN , xN ) ≤ 0 if and only if bfin (qN , xN ) = 0. Now
applying Theorem 2.5 to optimal control problem (2.9) with end point conditions
defined by (A.2) leads to the following set of conditions:
(i) (MP-i) holds,
(ii) The transversality conditions (MP-ii) can then be rewritten as
(A.3)
Tq̊∗N Φq̊−1 ζ N −1 , ξ N −1 − νD(q,x) cN (q̊N , x̊N ) − µN D(q,x) gN (q̊N , x̊N ) = 0,
N
∗
where (RnN ) 3 µN ≤ 0, and nN := 2 (nG + nx − mf ) . Using (A.2), the
transversality conditions (A.3) further simplify to
(A.4)
N
Tq̊∗N Φq̊−1 ζ N −1 , ξ N −1 −νD(q,x) cN (q̊N , x̊N ) = µN
+ − µ− D(q,x) bN (q̊N , x̊N ) ,
N
where
N
N
N
N
N
:=
,
.
.
.
,
µ
µ− := µN
,
.
.
.
,
µ
and
µ
µ
1
+
(nG +nx −mf )+1
2(nG +nx −mf ) .
(nG +nx −mf )
Note that the submanifold Mfin is locally embedded in RnG × Rnx via the
embedding bfin . Hence, it follows that [16, p. 60]
D(q,x) bfin (q̊N , x̊N ) v = 0
for all v ∈ T(q̊N ,x̊N ) Mfin ,
nG +nx −mf ∗
and hence for a covector α ∈ (R
αD(q,x) bfin (q̊N , x̊N ) v = 0
) , we have
for all v ∈ T(q̊N ,x̊N ) Mfin ,
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
27
which is equivalent to the following:
(A.5)
∗
(G × Rnx ) 3 αD(q,x) bfin (q̊N , x̊N ) ⊥ T(q̊N ,x̊N ) Mfin .
T(q̊
N ,x̊N )
So, the transversality conditions (A.4) with (A.5) leads to
Tq̊∗N Φq̊−1 ζ N −1 , ξ N −1 −νD(q,x) cN (q̊N , x̊N ) ⊥ T(q̊N ,x̊N ) Mfin .
N
(iii) (MP-iii) holds,
(iv) complementary slackness conditions (MP-iv) for t = N are trivially satisfied,
i.e.,
j
µN
j gN (q̊N , x̊N ) = 0 for all j = 1, . . . , nN .
Therefore leads to the following set of complementary slackness conditions
µtj gtj (q̊t , x̊t ) = 0
for all j = 1, . . . , nt and t = 1, . . . , N − 1,
(v) non-positivity condition (MP-v) modifies to the following
µt ≤ 0
for all
t = 1, . . . , N − 1,
N −1
N −1
(vi) Suppose the adjoint variables {(ζ t , ξ t )}t=0 , covectors {µt }t=1 and the scalar
ν are all zero. Then (A.4) modifies to
N
µN
(A.6)
+ − µ− D(q,x) bN (q̊N , x̊N ) = 0.
Since the map bN is a smooth submersion at (q̊N , x̊N ), D(q,x) bN (q̊N , x̊N ) is full
rank, and therefore
N
(A.7)
µN
+ − µ− = 0.
N
In case µN
+ = µ− 6= 0, (A.6) leads to triviality, i.e., the half spaces
H − := v ∈ T(q̊N ,x̊N ) (G × Rnx ) | µN
+ D(q,x) bN (q̊N , x̊N ) , v ≤ 0
and
H + := v ∈ T(q̊N ,x̊N ) (G × Rnx ) | µN
+ D(q,x) bN (q̊N , x̊N ) , v ≥ 0
are separable. Otherwise, contradicts non-triviality condition (MP-vi). So, the
non-triviality condition translates to the following:
N −1
N −1
the adjoint variables {(ζ t , ξ t )}t=0 , covectors {µt }t=1 and the scalar ν are all
not zero,
which completes our proof.
A.2. Proof of Corollary 2.8.
Proof. Assume that the inequality constraints in the Corollary 2.7 are given by
(A.8)
G × Rnx 3 (q, x) 7→ gt (q, x) := −1 ∈ R
for t = 1, . . . , N − 1,
the final cost is given by
G × Rnx 3 (q, x) 7→ cN (q, x) := 0 ∈ R,
and the fixed boundary point is an immersed manifold, i.e.,
Mfin := (q̄N , x̄N )
where Mfin is an immersed submanifold of the configuration space G × Rnx . Now
applying the Corollary 2.7 to optimal control problem (2.10) leads to the following
set of conditions:
28
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
(i) Note that
D(q,x) gt (q̊t , x̊t ) = 0
for t = 0. . . . , N − 1,
which simplifies the state and adjoint dynamics to
(
ν
qt+1 = qt eDζ H (γ̊t ) ,
state
xt+1 = Dξ H ν (γ̊t ) ,
(
ζ t−1 = Ad∗e−Dζ H ν (γ̊t ) ζ t + Te∗ Φq̊t Dq H ν (γ̊t ) ,
adjoint
ξ t−1 = Dx H ν (γ̊t ) ,
(ii) The manifold Mfin is a singleton. Hence
T(q̊N ,x̊N ) Mfin = 0,
∗
∗
and therefore the annihilator of T(q̊N ,x̊N ) Mfin in T(q̊
(G × Rnx ) is T(q̊
(G × Rnx ).
N ,x̊N )
N ,x̊N )
So, the transversality condition of Corollary 2.7 is trivially satisfied.
(iii) Hamiltonian non-positive gradient condition:
Du H ν t, ζ t , ξ t , q̊t , x̊t , ůt , w − ůt ≤ 0 for all w ∈ Ut .
(iv) In view of (A.8) and the complementary slackness conditions of the Corollary
2.7, we obtain
µt = 0 for all t = 1, . . . , N − 1.
The non-positivity condition of the Corollary 2.7 is trivially satisfied, and the
non-triviality condition translates to the following:
N −1
adjoint variables {ζ t , ξ t }t=0 and the scalar ν do not simultaneously vanish.
A.3. Proof of Corollary 2.9.
Proof. We first prove that there exist a unique representation of st in the neighborhood of e ∈ G satisfying vt (st , qt , xt ) = 0. Then the optimal control problem is
posed in the standard form (2.10) and necessary conditions are derived by applying
the Corollary 2.8.
Lemma A.1. If the map vt ( · , qt , xt ) : Oe ⊂ G → Rnq is a diffeomorphism for all
feasible pairs
n
o
N
At := (qt , xt ) ∈ G × Rnx | {(qt , xt )}t=0 satisfying (2.6) ,
where Oe is an open set in G containing e ∈ G, then for a fixed ((q̄t , x̄t , ) , s̄t ) ∈
At × Oe , there exist open neighborhoods Nt of (q̄t , x̄t , ) in G × Rnx , Rt of s̄t in G,
and a continuously differentiable map κt : Nt → Rt such that
st = κt (qt , xt ) ,
satisfies
vt (κt (qt , xt ) , qt , xt ) = 0.
Furthermore, the derivative Dκt (q̄t , x̄t ) : Tq̄t G × Rnx → R is given by
Dq κt (q̄t , x̄t ) = −Ds vt (s̄t , q̄t , x̄t )
−1
◦ Dq vt (s̄t , q̄t , x̄t )
and
−1
Dx κt (q̄t , x̄t ) = −Ds vt (s̄t , q̄t , x̄t ) ◦ Dx vt (s̄t , q̄t , x̄t ) .
N
N −1
N −1
Let {q̊t , x̊t }t=0 , {ůt }t=0 be an optimal state-action trajectory with {s̊t }t=0 ⊂
Oe such that vt (s̊t , q̊t , x̊t ) = 0 for t = 0, . . . , N − 1. Using Lemma A.1: there exist
open neighborhoods Nt ⊂ G × Rnx of (q̊t , x̊t ) , Rt ⊂ G of s̊t ∈ Oe , and the maps
κt : Nt → Rt such that
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
st = κt (qt , xt ) ,
satisfies
29
vt (κt (qt , xt ) , qt , xt ) = 0
for t = 0, . . . , N − 1. Let us define
N := N0 × · · · × NN ,
where NN ∈ G × Rnx is a neighborhood containing (q̊N , x̊N ) . Note that the system
N −1
dynamics in (2.11) is smooth in the control variables {ut }t=0
. Therefore there
exists a number r > 0 such that state trajectories starting at (q̄0 , x̄0 ) remains in N
under all control actions
(u0 , . . . , uN −1 ) ∈ Ũ := Ũ0 × · · · × ŨN −1 ,
where Ũt := (Ut ∩ Br (ůt )) ⊂ Rnu , and Br (ůt ) is an open neighborhood of radius r
containing ůt .
This leads to the representation of the optimal control problem (2.11) in the
neighborhood N × Ũ of the optimal state-action trajectory as :
(A.9)
N
−1
X
minimize J (q, x, u) =
ct (κt (qt , xt ) , qt , xt , ut )
N −1
{ut }t=0
subject to
t=0
qt+1 = qt κt (qt , xt )
xt+1 = ft (κt (qt , xt ) , qt , xt , ut )
ut ∈ Ũt
(q0 , x0 ) = (q̄0 , x̄0 ) ,
(qN , xN ) = (q̄N , x̄N ) ,
Assumption 2.1 in place.
for all t = 0, . . . , N − 1,
Let us define the Hamiltonian function as
∗
[N ] × g∗ × (Rnx ) × G × G × Rnx × Rnu 3 (τ, ζ, ξ, s, q, x, u) 7→
H ν (τ, ζ, ξ, s, q, x, u) := νcτ (s, q, x, u) + ζ, e−1 (s)
g
+ hξ, fτ (s, q, x, u)i ,
for ν ∈ R. Then applying Corollary 2.8 to (A.9) will immediately leads to the
following conditions:
(i) State and adjoint system dynamics:
Dζ H ν (γ̊t )
,
q̊t+1 = q̊t e
state vt (s̊t , q̊t , x̊t ) = 0,
x̊t+1 = Dξ H ν (γ̊t ) ,
−1
ζ t−1 = −Te∗ Φq̊t Ds H ν (γ̊t ) ◦ Ds vt (s̊t , q̊t , x̊t ) ◦ Dq vt (s̊t , q̊t , x̊t )
+ Te∗ Φq̊t (Dq H ν (γ̊t )) + Ad∗e−Dζ H ν (γ̊t ) ζ t ,
adjoint
−1
ξ t−1 = −Ds H ν (γ̊t ) ◦ Ds vt (s̊t , q̊t , x̊t ) ◦ Dx vt (s̊t , q̊t , x̊t )
+ Dx H ν (γ̊t ) .
(ii) Hamiltonian non-positive gradient condition:
Du H ν t, ζ t , ξ t ,s̊t , q̊t , x̊t , ůt , w̃ − ůt ≤ 0
for all
w̃ ∈ Ũt .
Since the set Ut is convex,
w̃α := ůt + α (w − ůt ) ∈ Ut
for any
α ∈ [0, 1]
and w ∈ Ut .
30
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
In particular, choosing
n
o
(
r
min 2kw−ů
,
1
tk
ᾱ :=
0
for
w 6= ůt ,
otherwise,
ensures that w̃ᾱ ∈ Ũt , and consequently,
Du H ν t, ζ t , ξ t ,s̊t , q̊t , x̊t , ůt , ᾱ (w − ůt ) ≤ 0
for all
w ∈ Ut .
Since ᾱ ≥ 0, it follows that
Du H ν t, ζ t , ξ t ,s̊t , q̊t , x̊t , ůt , w − ůt ≤ 0
for all
w ∈ Ut ,
and this proves the assertion.
(iii) Non-triviality condition:
N −1
adjoint variables {ζ t , ξ t }t=0 and the scalar ν do not simultaneously vanish.
Appendix B. Proofs
Lemma A.1. If the map vt ( · , qt , xt ) : Oe ⊂ G → Rnq is a diffeomorphism for all
feasible pairs
n
o
N
At := (qt , xt ) ∈ G × Rnx | {(qt , xt )}t=0 satisfying (2.6) ,
where Oe is an open set in G containing e ∈ G, then for a fixed ((q̄t , x̄t , ) , s̄t ) ∈
At × Oe , there exist open neighborhoods Nt of (q̄t , x̄t , ) in G × Rnx , Rt of s̄t in G,
and a continuously differentiable map κt : Nt → Rt such that
st = κt (qt , xt ) ,
satisfies
vt (κt (qt , xt ) , qt , xt ) = 0.
Furthermore, the derivative Dκt (q̄t , x̄t ) : Tq̄t G × Rnx → R is given by
Dq κt (q̄t , x̄t ) = −Ds vt (s̄t , q̄t , x̄t )
−1
◦ Dq vt (s̄t , q̄t , x̄t )
and
Dx κt (q̄t , x̄t ) = −Ds vt (s̄t , q̄t , x̄t )
−1
◦ Dx vt (s̄t , q̄t , x̄t ) .
Proof. Given a fixed feasible pair (q̄t , x̄t ) ∈ At , the map
(B.1)
vt ( · , q̄t , x̄t ) : Oe → vt (Oe , q̄t , x̄t )
is a local diffeomorphism. Let (U, α) and (V, β) be the local charts of the Lie
group G at q̄t ∈ U and s̄t ∈ V respectively, and subsequently define
U × V × Rnx 3 (s, q, x) 7→ ϕt (s, q, x) := (α(s), β(q), x) ∈ α (U ) × β (V ) × Rnx
So, the local representation of the map (B.1)
α(U ) 3 (ς) 7→ ṽt (ς) := vt ◦ ϕ−1
(ς, β(q̄t ), x̄t ) ∈ Rnq
t
is a diffeomorphism onto its image, and therefore Dṽt (α(s̄t )) is invertible with
ṽt (α(s̄t )) = 0.
Using the Implicit Function Theorem [14, p. 100]: there exist neighborhoods Ñt
of (β(q̄t ), x̄t ) ∈ Rnq × Rnx , R̃t of α(s̄t ) ∈ Rnq , and a C 1 differentiable map
κ̃t : Ñt → R̃t
such that
s̃t = κ̃t (β(qt ), xt ) and
vt ◦ ϕ−1
(s̃t , β(qt ), xt ) = 0.
t
DISCRETE-TIME PMP ON MATRIX LIE GROUPS
31
The sets Nt := δ −1 Ñt ⊂ G × Rnx , Rt := α−1 R̃t ⊂ G are open, and the
map
Nt 3 (q, x) 7→ κt (q, x) := α−1 ◦ κ̃t ◦ δ (q, x) := s ∈ Rt ,
where δ(q, x) := (α(q), x), represent s in terms of (q, x) uniquely. In other words,
st = κt (qt , xt ) and vt (κt (qt , xt ) , qt , xt ) = 0.
Taking the derivative of the equation vt (κt (qt , xt ) , qt , xt ) = 0 with respect to x
and q gives the following results:
Dq κt (qt , xt ) = −Ds vt (κt (qt , xt ) , qt , xt )
−1
◦ Dq vt (κt (qt , xt ) , qt , xt )
and
Dx κt (qt , xt ) = −Ds vt (κt (qt , xt ) , qt , xt )
−1
◦ Dx vt (κt (qt , xt ) , qt , xt ) .
References
[1] R. Abraham and J. Marsden, Foundations of Mechanics, AMS Chelsea Publishing, 1978.
[2] A. A. Agrachev and Y. L. Sachkov, Control Theory from the Geometric Viewpoint,
Springer-Verlag New York, 2003.
[3] M. Barbero-Liñán and M. C. Muñoz-Lecanda, Geometric approach to Pontryagin maximum principle, Acta Applicandae Mathematicae, 108 (2009), pp. 429–485.
[4] V. G. Boltyanski, The method of tents in the theory of extremal problems, Russian Math
Surveys, 30 (1975), pp. 1–54.
[5] V. G. Boltyanski, H. Martini, and V. Soltan, Geometric Methods and Optimization
Problems, Kluwer Academic Publishers, 1998.
[6] B. Bonnard and D. Sugny, Optimal control with applications in space and quantum dynamics, American Institute of Mathematical Sciences, Springfield, USA, 2012.
[7] L. Bourdin and E. Trélat, Optimal sampled-data control, and generalizations on time
scales, Mathematical Control and Related Fields, 6 (2016), pp. 53–94.
[8] R. W. Brockett, Lie theory and control systems defined on spheres, SIAM Journal on
Applied Mathematics, 25 (1973), pp. 213–225.
[9] F. Bullo and K. M. Lynch, Kinematic controllability for decoupled trajectory planning
in underactuated mechanical systems, IEEE Transactions on Robotics and Automation, 17
(2001), pp. 402–412.
[10] F. Clarke, Functional Analysis, Calculus of Variations and Optimal Control, vol. 264,
Springer Science & Business Media, 2013.
[11] F. H. Clarke, The maximum principle under minimal hypotheses, SIAM Journal on Control
and Optimization, 14 (1976), pp. 1078–1091.
[12] A. Y. Dubovitskii, The discrete principle of the maximum, Avtomatika i Telemekhanika,
(1978), pp. 55–71.
[13] A. Y. Dubovitskii and A. A. Milyutin, Necessary conditions for a weak extremum in optimal control problems with mixed constraints of the inequality, USSR Computational Mathematics and Mathematical Physics, 8 (1968), pp. 24–98.
[14] J. J. Duistermaat and J. A. C. Kolk, Multidimensional Real Analysis I: Differentiation,
Cambridge University Press, 2004.
[15] R. V. Gamkrelidze, Discovery of the maximum principle, Journal of Dynamical and Control
Systems, 5 (1999), pp. 437–451.
[16] D. Holm, T. Schmah, and C. Stoica, Geometric Mechanics and Symmetry: From Finite
to Infinite Dimensions, OUP Oxford, 2009.
[17] J. M. Holtzman, Convexity and the maximum principle for discrete systems, IEEE Transactions on Automatic Control, 11 (1966), pp. 30–35.
[18] V. Janković, Tent theory, Journal of Mathematical Sciences, 110 (2002), pp. 2572–2582.
[19] N. Khaneja, R. Brockett, and S. J. Glaser, Time optimal control in spin systems, Physical Review A, 63 (2001), p. 032308.
[20] M. B. Kobilarov and J. E. Marsden, Discrete geometric optimal control on Lie groups,
IEEE Transactions on Robotics, 27 (2011), pp. 641–655.
[21] T. L., M. Leok, and N. McClamroch, Time optimal attitude control for a rigid body,
American Control Conference, (2008), pp. 5210–5215.
[22] J. M. Lee, Introduction to Smooth Manifolds, Springer Science & Business Media, 2013.
32
K. S. PHOGAT, D. CHATTERJEE, AND R. BANAVAR
[23] T. Lee, M. Leok, and N. McClamroch, Optimal attitude control of a rigid body using
geometrically exact computations on SO(3), Journal of Dynamical and Control Systems,
Springer, 14 (2008), pp. 465–487.
[24] J. E. Marsden and T. S. Ratiu, Introduction to Mechanics and Symmetry: A Basic Exposition of Classical Mechanical Systems, Springer-Verlag New York, 2010.
[25] J. E. Marsden and M. West, Discrete mechanics and variational integrators, Acta Numerica, 10 (2001), pp. 357–514.
[26] A. A. Milyutin and N. Osmolovskii, Calculus of Variations and Optimal Control, American
Mathematical Society, 1998.
[27] B. S. Mordukhovich, Maximum principle in the problem of time optimal response with
nonsmooth constraints, Journal of Applied Mathematics and Mechanics, 40 (1976), pp. 960–
969.
[28] S. Ober-Blöbaum, Discrete Mechanics and Optimal Control, PhD thesis, University of
Paderborn, Paderborn, 2008.
[29] S. Ober-Blöbaum, O. Junge, and J. E. Marsden, Discrete mechanics and optimal control:
An analysis, ESAIM: Control, Optimisation and Calculus of Variations, 17 (2011), pp. 322–
352.
[30] K. S. Phogat, D. Chatterjee, and R. N. Banavar, Multiple shooting technique for optimal attitude control of a spacecraft with momentum and control constraints, arXiv.org,
abs/1509.04521 (2015).
[31] L. S. Pontryagin, Mathematical Theory of Optimal Processes, CRC Press, 1987.
[32] B. N. Pshenichnyi, Necessary Conditions for an Extremum, Marcel Dekker Inc, New York,
1971.
[33] A. Saccon, J. Hauser, and A. P. Aguiar, Optimal control on Lie groups: The projection
operator approach, IEEE Transactions on Automatic Control, 58 (2013), pp. 2230–2245.
[34] Y. L. Sachkov, Control theory on Lie groups, Journal of Mathematical Sciences, 156 (2009),
pp. 381–439.
[35] H. J. Sussmann, An introduction to the coordinate-free maximum principle, In Geometry of
Feedback and Optimal Control, B. Jakubczyk and W. Respondek, pp. 463–557.
[36] H. J. Sussmann, The Euler-Lagrange equation and its invariance properties. https://
math.la.asu.edu/~kawski/classes/apm582/16sprg/various/weizman-HJS-2000.pdf, 2000.
Lecture Notes.
[37] H. J. Sussmann, A Pontryagin maximum principle for systems of flows, Recent Advances in
Learning and Control (Lecture Notes in Control and Information Sciences), (2008), pp. 219–
232.
[38] E. Trélat, Optimal control and applications to aerospace: some results and challenges,
Journal of Optimization Theory and Applications, 154 (2012), pp. 713–758.
[39] R. Vemulapalli, F. Arrate, and R. Chellappa, Human action recognition by representing
3d skeletons as points in a Lie group, IEEE Conference on Computer Vision and Pattern
Recognition, (2014), pp. 588–595.
[40] J. Warga, Optimal Control of Differential and Functional Equations, Academic press, 1972.
Systems & Control Engineering, IIT Bombay, Mumbai 400076, India
E-mail address, K. S. Phogat: [email protected]
URL, K. S. Phogat: http://www.sc.iitb.ac.in/~karmvir.p
E-mail address, D. Chatterjee: [email protected]
URL, D. Chatterjee: http://www.sc.iitb.ac.in/~chatterjee
E-mail address, R. Banavar: [email protected]
URL, R. Banavar: http://www.sc.iitb.ac.in/~banavar
| 3cs.SY
|
Joint Blind Motion Deblurring and Depth Estimation of Light Field
Dongwoo Lee†
Haesol Park†
In Kyu Park‡
Kyoung Mu Lee†
†
Department of ECE, ASRI, Seoul National University
‡
Department of Information and Communication Engineering, Inha University
arXiv:1711.10918v1 [cs.CV] 29 Nov 2017
[email protected]
[email protected]
[email protected]
[email protected]
Abstract
Removing camera motion blur from a single light field is
a challenging task since it is highly ill-posed inverse problem. The problem becomes even worse when blur kernel
varies spatially due to scene depth variation and high-order
camera motion. In this paper, we propose a novel algorithm
to estimate all blur model variables jointly, including latent sub-aperture image, camera motion, and scene depth
from the blurred 4D light field. Exploiting multi-view nature of a light field relieves the inverse property of the optimization by utilizing strong depth cues and multi-view blur
observation. The proposed joint estimation achieves high
quality light field deblurring and depth estimation simultaneously under arbitrary 6-DOF camera motion and unconstrained scene depth. Intensive experiment on real and
synthetic blurred light field confirms that the proposed algorithm outperforms the state-of-the-art light field deblurring
and depth estimation methods.
(a)
(b)
(c)
(d)
1. Introduction
Figure 1: The proposed algorithm jointly estimates latent
image, depth map, and camera motion from a single light
field. (a) Center-view of blurred light field sub-aperture image. (b) Deblurred image of (a). (c) Estimated depth map.
(d) Camera motion path and orientation (6-DOF).
For the last decade, motion deblurring has been an active
research topic in computer vision. Motion blur is produced
by relative motion between camera and scene during the exposure where blur kernel, i.e. point spread function (PSF),
is spatially non-uniform. In blind non-uniform deblurring
problem, pixel-wise blur kernels and corresponding sharp
image are estimated simultaneously.
Early works on motion deblurring [4, 7, 11, 26, 35] focus on removing spatially uniform blur in the image. However, the assumption of uniform motion blur is often broken in real world due to nonhomogeneous scene depth and
rolling motion of camera. Recently, a number of methods [8, 12, 13, 14, 16, 18, 29, 32, 37] have been proposed for
non-uniform deblurring. However, they still can not completely handle non-uniform blur caused by scene depth variation. The main challenge lies in the difficulty of estimating the scene depth with only single observation, which is
highly ill-posed.
A light field camera ameliorates the ill-posedness of
single-shot deblurring problem of the conventional camera.
4D light field is equivalent to multi-view images with narrow baseline, i.e. sub-aperture images, taken with an identical exposure [22]. Consequently, motion deblurring using light field can be leveraged by its multi-dimensional nature of captured information. First, strong depth cue is obtained by employing multi-view stereo matching between
sub-aperture images. In addition, different blurs in the subaperture images can help the optimization converge more
fast and precise.
In this paper, we propose an efficient algorithm to jointly
estimate latent image, sharp depth map, and 6-DOF camera motion from a blurred single 4D light field as shown
in Figure 1. In the proposed light field blur model, latent sub-aperture images are formulated by 3D warping of
the center-view sharp image using the depth map and the
6-DOF camera motion. Then, motion blur is modeled as
4321
Input
Blurred Light Field
Center-view Image
Depth Initialization
Joint Optimization
Initial Depth Map
Initial Camera Motion
Latent Image
Outputs
Sharp Image
Sharp Light Field
Sharp Depth Map
Camera Motion
Camera Motion
Depth Map
Motion Initialization
Figure 2: Overview of the proposed algorithm. The proposed algorithm jointly estimates the latent image, depth map, and
camera motion from a single light field.
the integral of latent sub-aperture images during the shutter
open. Note that the proposed center-view parameterization
reduces light field deblurring problem in lower dimension
comparable to a single image deblurring. The joint optimization is performed in an alternating manner, in which
the deblurred image, depth map, and camera motion are refined during iteration. The overview of the proposed algorithm is shown in Figure 2. In overall, the contribution of
this paper is summarized as follows.
scene depth. Kim and Lee [18] jointly estimated the spatially varying motion flow and the latent image. Sun et
al. [29] adopted a learning method based on convolutional
neural network (CNN) and assumed that the motion was locally linear. However, the locally linear blur assumption
does not hold in large motion.
Video and Multi-View Deblurring Xu and Jia [36] decomposed the image region according to the depth map obtained from a stereo camera and recombined them after independent deblurring. Recently, several methods [9, 19, 23,
25, 34] have addressed the motion blur problem in video
sequences. Video deblurring shows good performance, because it exploits optical flow as a strong guide for motion
estimation.
Light Field Deblurring In last a few years, several approaches [5, 10, 17, 27, 28] have been proposed to perform
motion deblurring on the light field. Jin et al. [17] quantized
the depth map into two layers and removed the motion blur
in each layer. Their method assumed that the camera motion
is in-plane translation and utilized depth value as a scale factor of translational motion. Dansereau et al. [10] applied the
Richardson-Lucy deblurring algorithm to the light field with
non-blind 6-DOF motion blur. Although their method dealt
with 6-DOF motion blur, it was assumed that the ground
truth camera motion was known. Unlike [10], in this paper, we address the problem of blind deblurring which is a
more highly ill-posed problem. Srinivasan et al. [28] solved
the light field deblurring under 3D camera motion path and
showed visually pleasing result. However, their methods do
not consider 3D orientation change of the camera.
In contrast to the previous works of light field deblurring, the proposed method completely handles 6-DOF motion blur and unconstrained scene depth variation.
• We propose a joint method which simultaneously
solves deblurring, depth estimation, and camera motion estimation problems from a single light field.
• Unlike the previous state-of-the-art algorithm, the proposed method handles blind light field motion deblurring under 6-DOF camera motion.
• Practical and extensible blur formulation that can be
extended to any multi-view camera system.
2. Related Works
Single Image Deblurring Conventional method to remove the spatially-varying blur in a single image is to find
the motion density function (MDF) first and generate the
pixel-wise kernel [12, 13, 14]. Gupta et al. [12] and Hu
and Yang [14] modeled the camera motion in a discrete 3D
motion space. They deblurred the image by iteratively optimizing the MDF and the latent image that best describe the
blurred image. They parameterize the spatially-varying blur
kernel in low dimension. However, modeling the motion
blur using only MDF often fails when the motion blur is affected by both camera motion and scene depth. In [13], the
image was segmented by employing a matting algorithm,
and the MDF and layered depth were found through the
expectation-maximization algorithm.
A few methods [18, 29] estimated linear blur kernels locally, and they showed acceptable results for the arbitrary
3. Motion Blur Formulation in Light Field
A pixel in a 4D light field has four coordinates, i.e. (x, y)
for spatial and (u, v) for angular coordinates. A light field
4322
Pctm = exp(
can be interpreted as a set of u × v multi-view images with
narrow baseline, which are often called sub-aperture images [21]. Throughout this paper, a sub-aperture image is
represented as I(x, u) where x = (x, y) and u = (u, v).
For each sub-aperture image, the blurred image B(x, u) is
the average of the sharp images It (x, u) during the shutter
open over [t0 , t1 ] as follows:
Z
It (x, u)dt.
(1)
t0
Following the blur model of [23, 25], we approximate all
the blurred sub-aperture images by projecting a single latent
image with 3D rigid motion. We choose the center-view (c)
of sub-aperture images and the middle of the shutter time
(tr ) as the reference angular position and the time stamp
of the latent image. With above notations, the pixel correspondence from each sub-aperture image to the latent image
Itr (x, c) is expressed as follows:
It (x, u) = Itr (wt (x, u), c),
4. Joint Estimation of Latent Image, Camera
Motion, and Depth Map
In order to estimate all blur variables in the proposed
light field blur model, we need to recover the latent variables, i.e. Itr (x, c), Dtr (x, c), and Pct0 . We model an energy function as follows:
(2)
E=
where
wt (x, u) =
−1
−1
Πc (Pctr (Pu
t ) Πu (x, Dt (x, u))).
(3)
(4)
M −1
1 X
It (wtm (x, u), c).
(Ψ ◦ I)(x, u) =
M m=0 r
(5)
u
x
+ λL
X
λu k(Ψ ◦ I)(x, u) − B(x, u)k1
k∇Itr (x, c)k2 + λD
X
k∇Dtr (x, c)k2 .
x
(8)
The data term imposes the brightness consistency between
the input blurred light field and the restored light field.
Notice that the L1-norm is employed in our approach as
in [18], where it effectively removes the ringing artifact
around object boundary and provides more robust deblurring results on large depth change. The last two terms are
the total variation (TV) regularizers [1] for the latent image
and the depth map, respectively.
In our energy model, Dtr (x, c) and Pct0 are implicitly
included in the warping function (5). Simultaneously optimizing the three variables is complicated because the warping function (5) has severe nonlinearity. Therefore, our
strategy is to optimize three latent variables in an alternating manner. We minimize one variable while the others are
fixed. The optimization (8) is carried out in turn for the
three variables. The L1 optimization is approximated using iterative reweighted least square (IRLS) [24]. The optimization procedure converges in small number of iterations (< 10). An example of the iterative optimization is
illustrated in Figure 3 in which the intermediate latent variables are shown.
where
In (5), tm is mth uniformly sampled time stamp during the
interval [t0 , t1 ].
Our goal is to formulate (Ψ ◦ I)(x, u) with only centerview variables, i.e. Itr (x, c), Dtr (x, c), and Pct0 . Pu
tm and
Dtm (x, u) are variables related to u in the warping function
(5). Therefore, we parameterize Pu
tm and Dtm (x, u) by employing center-view variables. Because the relative camera
c
pose Pc→u is fixed over time, Pu
tm is expressed by Pt0 and
c
Pt1 as follows:
c→u c
Pu
Ptm ,
tm = P
XX
x
wt (x, u) computes the warped pixel position from u to c,
and from t to tr . Πc , Π−1
u are the projection and backprojection function between the image coordinate and the
3D homogeneous coordinate using the camera intrinsic parameters. Matrices Pctr and Pu
t ∈ SE(3) denote the 6-DOF
camera pose at the corresponding angular position and the
time stamp. Dt (x, u) is the depth map at the time stamp t.
In the proposed model, the blur operator Ψ(·) is defined
by approximating the integral in (1) as a finite sum as follows:
B(x, u) ≈ (Ψ ◦ I)(x, u),
(7)
where exp and log denote the exponential and logarithmic maps between Lie group SE(3) and Lie algebra se(3)
space [2]. To minimize the viewpoint shift of the latent image, we assume Pct1 = (Pct0 )−1 which makes Pctm an identity matrix when tm = tr . Note that we use the camera
path model used in [23, 25]. However, the Bézier camera
path model used in [28] can be directly applied to (7) as
well. Dtm (x, u) is also represented by Dtr (x, c) by forward warping and interpolation.
t1
B(x, u) =
m
−1
log(Pct1 (Pct0 ) ))Pct0 ,
M
4.1. Update of the Latent Image
The proposed algorithm first updates the latent image
Itr (x, c). In our data term, the blur operator (5) is simplified as the linear matrix multiplication, if Dtr (x, c) and
Pct0 remain fixed. Updating the latent image is equivalent to
(6)
4323
Initial Iter. 1 Iter. 3 Iter. 5
(a)
(b)
Initial Iter. 1 Iter. 3 Iter. 5
(c)
(d)
Figure 3: Example of the iterative joint estimation. The proposed method converges in small number of iteration. (a)∼(b)
Input blurred image and deblurring results by iteration. (c)∼(d) Initial blurred depth map and depth estimation results by
iteration.
4.3. Initialization
minimizing (8) as follows:
X
min
kK u Itcr − B u k1 + λL k∇Itcr k2 .
c
It
Since deblurring is a highly ill-posed problem and the
optimization is done in a greedy and iterative fashion, it is
important to start with good initial values. First, we initialize the depth map using the input sub-aperture images of the
light field. It is assumed that the camera is not moving and
(8) is minimized to obtain the initial Dtr (x, c). Minimizing
(8) becomes a simple multi-view stereo matching problem.
Figure 3(c) shows the initial depth map which exhibits fattened object boundary.
Camera motion Pct0 is initialized from the local linear
blur kernels and initial scene depth. We first estimate the
local linear blur kernel of B(x, c) using [29]. Then, we fit
the pixel coordinates moved by the linear kernel and the reprojected coordinates by the warping function as follows:
(9)
u
Itcr , B u ∈ Rn are vectorized images and K u ∈ Rn×n is the
blur operator in square matrix form, where n is the number
of pixels in the center-view sub-aperture image. TV regularization serves as a prior to the latent image with clear
boundary while eliminating the ringing artifacts.
4.2. Update of the Camera Pose and Depth Map
Since (5) is a non-linear function of Dtr (x, c) and Pct0 ,
it is necessary to approximate it in a linear form for efficient
computation. In our approach, the blur operation (5) is approximated as a first-order expansion. Let D0 (x, c) and Pc0
denote the initial variables, then (5) is approximated as follow:
(Ψ ◦ I)(x, u)
= B0 (x, u) +
∂B0
∂f
∂f ( ∂Dtr (x,c) ∆Dtr (x, c)
min
c
Pt
∂f
∂εt0 εt0 ),
+
(10)
0
=Pc0 ,
(11)
Note that f is motion flow generated by warping function,
and εt0 denotes six-dimensional vector on se(3). The partial
derivatives related to Dtr (x, c) and εt0 are given in [2].
Once it is approximated using ∆Dtr (x, c) and εt0 , (8)
can be optimized using IRLS. The resulting ∆Dtr (x, c) and
εt0 are incremental values for the current Dtr (x, c) and Pct0 ,
respectively. They are updated as follows:
Dtr (x, c) = Dtr (x, c) + ∆Dtr (x, c),
Pct0 = exp(εt0 )Pct0 ,
kwt0 (xi , c) − l(xi )k22 ,
(13)
i=1
where xi is the sampled pixel position and l(xi ) is the point
that xi is moved by the end point of the linear kernel. Pct0 is
obtained by fitting xi moved by wt0 (·, c) and l(·). Pct0 is the
only variable of wt0 (·, c) since the scene depth is fixed to
the initial depth map. In our implementation, RANSAC is
used to find the camera motion that best describes the pixelwise linear kernels. N is the number of random samples,
which is fixed to 4.
Figure 4 shows an example of camera motion initialization. It is shown that [29] underestimates the size of the
motion (upper blue rectangle) and produces noisy motion
where the texture is insufficient (lower blue rectangle).
where
B0 (x, u) = (Ψ ◦ I)(x, u)|Dtr (x,c)=D0 (x,c),Pct
0
N
X
(12)
5. Experimental Results
where Pct0 is updated through the exponential mapping of
the motion vector εt0 .
Figure 3 shows the initial latent variables and final outputs. After joint estimation, both the latent image and the
depth map become clean and sharp.
The proposed algorithm is implemented using Matlab on
an Intel i7 7770K @ 4.2GHz with 16GB RAM and is evaluated for both synthetic and real light fields. Synthetic light
field is generated using Blender [3] for qualitative as well as
4324
5.1. Light Field Deblurring
(a)
(b)
(c)
Figure 5 and Figure 6 show the light field deblurring
results for blurred real light field with spatially varying blur
kernels. In Figure 5, the result is compared with the existing
motion deblurring methods [18, 29] which utilize motion
flow estimation. It is shown that the proposed algorithm
reconstructs sharper latent image better than others. Note
that [18, 29] show satisfactory performance only for small
blur kernels.
Figure 6 shows the comparison results with the deblurring method based on the global camera motion model [13,
28]. In comparison with [28], we deblur only cropped regions shown in the yellow boxes of Figure 6(c) due to GPU
memory overflow (>12GB) for larger spatial resolution.
[13] assumes the scene depth is piecewisely planar.
Therefore, it cannot be generalized to arbitrary scene, yielding unsatisfactory deblurring result. [28] estimates the reasonably correct camera motion of the blurred light field
while their output is less deblurred. Note that [28] can not
handle the rotational camera motion which produces completely different blur kernels from translational motion. On
the other hand, the proposed algorithm fully utilizes the 6DOF camera motion and the scene depth, yielding outperforming results for the arbitrary scene.
The performance of the proposed algorithm is evaluated
using synthetic light field dataset, as shown in Figure 7
and Table 1. In Figure 7, we visualize and compare the
deblurring performance with existing motion flow methods [18, 29] and a camera motion method [13]. In all examples, the proposed algorithm produces sharper deblurred
images than others as shown clearly in the cropped boxes.
Table 1 shows the quantitative comparison of deblurring
performance by measuring PSNR and SSIM to the ground
truth. It shows that the proposed joint estimation algorithm significantly outperforms the others. Sun et al. [29]
achieves comparable performance to the proposed algorithm in which CNN is trained with MSE loss. Other algorithms achieve minor improvement from the input image
because the assumed blur models are simple and inconsistent with the ground truth blur.
For the comparison with [28], we crop the each light field
to 200 × 200 because of the GPU memory overflow. Note
that we use the original setting of [28]. [28] shows lower
performance than the input blurred light field due to the
spatial viewpoint shift as in the output of [28]. Since the
original point exists at the end point of the camera motion
path in [28], the viewpoint shift occurs when the estimated
3D motion is large. It is observed that this is an additional
cause to decrease PSNR and SSIM when the estimated 3D
motion is different from the ground truth. The proposed
algorithm estimates the latent image with ignorable viewpoint shift because the origin is located in the middle of the
camera motion path.
(d)
Figure 4: Example of camera motion initialization on a synthetic light field. (a) Blurred input light field. (b) Ground
truth motion flow. (c) Sun et al. [29] (EPE = 3.05), (d) Proposed initial motion (EPE = 0.95). In (b) and (d), the linear
blur kernels are approximated only using the end points of
camera motion.
quantitative evaluation. It includes 6 types of camera motion for 3 different scenes in which each light field has 7 × 7
angular structure of 480×360 sub-aperture images. Synthetic blur is simulated by moving the camera array over a
sequence of frames (≥ 40) and then by averaging the individual frames. On the other hand, real light field data is
captured using Lytro Illum camera which generates 7 × 7
angular structure of 552×383 sub-aperture images. Light
fields are blurred by moving camera quickly under arbitrary
motion, while the scene remains static. In our implementation, we fixed most of the parameters except λD such that
λu = 15, λc = 1, λL = 5. λD is set to a larger value
for a real light field (λD = 400) than for synthetic data
(λD = 20).
For quantitative evaluation of deblurring, we use both
peak signal to noise ratio (PSNR) and structural similarity (SSIM). Note that PSNR and SSIM are measured by
the maximum (best) ones among individual PSNR and
SSIM values computed between the deblurred image and
the ground truth images (along the motion trajectory) as
adopted in [20]. For comparison with light field depth estimation methods, we use the relative mean absolute error (L1-rel) defined as
L1-rel(D, D̂) =
1 X |Di − D̂i |
,
n i
D̂i
(14)
which computes the relative error of the estimated depth D̂
to the ground truth depth D. The accuracy of camera motion
estimation is measured by the average end point error (EPE)
to the end point of ground truth blur kernels. In our evaluation, we compute the EPE by generating an end point of blur
kernel using the estimated camera motion and ground truth
depth. We compare the performance of the proposed algorithm to linear blur kernel methods that directly computes
the EPE between the ground truth and their pixel-wise blur
kernel.
4325
(a)
(b)
(c)
(d)
Figure 5: Deblurring result for real light field dataset with comparison to local linear blur kernel deblurring methods. (a)
Blurred input image. (b) Result of Kim and Lee [18]. (c) Sun et al. [29]. (d) Proposed algorithm.
(a)
(b)
(c)
(d)
Figure 6: Deblurring result for real light field dataset with comparison to global camera motion estimation methods. (a)
Blurred input image. (b) Result of Hu et al. [13]. (c) Srinivasan et al. [28]. (d) Proposed algorithm.
4326
(a)
(b)
(c)
(d)
(e)
Figure 7: Deblurring result for synthetic light field. (a) Blurred input light field. (b) Result of Hu et al. [13]. (c) Kim and
Lee [18]. (d) Sun et al. [29]. (e) Proposed algorithm.
Table 1: Quantitative evaluation of deblurring on synthetic light field dataset (in PSNR and SSIM).
Forward
PSNR SSIM
Rotation
PSNR SSIM
Translation
PSNR SSIM
Forward+Rot.
PSNR SSIM
Forward+Tran.
PSNR SSIM
Rot.+Tran.
PSNR SSIM
Input
Hu et al. [13]
Kim and Lee [18]
Sun et al. [29]
Proposed Method
20.72
20.00
20.06
27.69
29.24
0.740
0.716
0.721
0.896
0.915
20.37
19.42
19.78
27.68
29.14
0.731
0.704
0.714
0.881
0.913
21.82
21.42
21.42
25.41
26.99
0.758
0.745
0.749
0.856
0.876
19.84
19.18
19.32
27.40
28.92
0.723
0.701
0.706
0.874
0.905
20.30
19.43
19.65
27.23
28.91
0.731
0.699
0.708
0.899
0.922
19.79
19.24
19.34
26.42
27.85
0.728
0.711
0.714
0.868
0.893
Input (cropped)
Srinivasan et al. [28]
Proposed Method
21.01
17.15
27.15
0.758
0.730
0.871
21.19
19.02
27.32
0.746
0.652
0.870
19.39
16.28
25.30
0.698
0.620
0.836
21.73
19.17
28.83
0.758
0.660
0.904
21.67
16.20
28.01
0.782
0.726
0.901
20.50
16.38
25.88
0.745
0.626
0.867
Methods
Table 2: Comparison of depth estimation (in average L1-rel
error).
5.2. Light Field Depth Estimation
To show the performance of light field depth estimation,
we compare the proposed method with several state-of-theart methods [6, 15, 30, 31, 33]. For comparison, all blurred
sub-aperture images are independently deblurred using [29]
before running their own depth estimation algorithms.
Figure 8 shows the visual comparison of estimated depth
map generated by different methods, which confirms that
the proposed algorithm produces significantly better depth
map in terms of accuracy and completeness. Since independent deblurring of all sub-aperture images does not
consider correlation between them, conventional correspondence and defocus cue do not produce reliable matching,
yielding noisy depth map. Only the proposed joint estimation algorithm results in sharp and unfattened object boundary, and produces the closest result to the ground truth.
Quantitative performance comparison of depth map estimation is shown in Table 2. For three synthetic scenes
with three different motion for each scene, the average L1rel error of the estimated depth map is computed and compared. The comparison clearly shows that the proposed
Methods
Forward
Rotation
Trans.
Overall
Chen et al. [6]
Tao et al. [30]
Wang et al. [31]
Jeon et al. [15]
Williem and Park [33]
Proposed Method
0.0251
0.0251
0.0312
0.0835
0.0615
0.0198
0.0326
0.0359
0.0377
0.0916
0.0895
0.0150
0.0331
0.0371
0.0400
0.0921
0.0966
0.0243
0.0303
0.0327
0.0363
0.0891
0.0825
0.0197
method produces the lowest error in all types of camera motion. Note that the second best result is achieved by Chen et
al. [6], which is relatively robust in the presence of motion
blur because bilateral edge preserving filtering is employed
for cost computation.
The depth estimation experiment demonstrates that consistent deblurring between sub-aperture images is essential. The cues used in conventional depth estimation are
seriously hampered in the independently deblurred image.
However, the proposed joint estimation algorithm results in
4327
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 8: Depth estimation results on blurred light field. (a) Blurred center sub-aperture image. (b) Ground truth depth. (c)
Result of Jeon et al. [15]. (d) Williem and Park [33]. (e) Tao et al. [30]. (f) Wang et al. [31]. (g) Chen et al. [6]. (h) Proposed
algorithm.
Table 3: Comparison of motion estimation (in EPE).
Methods
Forward
Rotation
Translation
Kim and Lee [18]
Sun et al. [29]
Proposed Method
2.153
1.492
0.325
3.317
2.557
0.171
1.989
1.810
0.590
5.3. Camera Motion Estimation
(a)
(b)
(c)
(d)
Table 3 shows the EPE of the estimated motion on
synthetic light field dataset. Compared with other methods [18, 29], the proposed method improves the accuracy
of the estimated motion significantly. In particular, a large
gain is obtained in the rotational motion, which indicates
that the rotational motion cannot be modeled accurately as
a linear blur kernel used in [18, 29].
Figure 9 shows the motion estimation results compared
to the ground truth motion. Since the camera orientation
changes while the camera is moving, the 6-DOF camera
motion can not be recovered properly by [28]. As shown in
Figure 9(b) and (c), the deblurring results are similar to the
input, because the motion can not converge to the ground
truth. In contrast, the proposed algorithm converges to the
ground truth 6-DOF motion and also produces the sharp deblurring result.
Figure 9: Deblurring and camera motion estimation result
for synthetic light field with comparison to [28]. (a) Input
light field and ground truth camera motion. (b) Result of
Srinivasan et al. [28] (quadratic). (b) Srinivasan et al. [28]
(cubic). (d) Proposed algorithm.
6. Conclusion
In this paper, we presented the novel light field deblurring algorithm that estimated latent image, sharp depth map,
and camera motion jointly. Firstly, we modeled all the
blurred sub-aperture images by center-view latent image using 3D warping function. Then, we developed the algorithm
to initialize the 6-DOF camera motion from the local linear blur kernel and scene depth. In the iterative joint opti-
robust and accurate depth estimation, providing the justification of solving deblurring and depth estimation in a joint
manner.
4328
mization, the nonlinear energy minimization was solved efficiently using IRLS. The evaluation on both synthetic and
real light field data showed that the proposed model and algorithm worked well with general camera motion and scene
depth variation.
[16] H. Ji and K. Wang. A two-stage approach to blind spatiallyvarying motion deblurring. In Proc. of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 73–80,
2012.
[17] M. Jin, P. Chandramouli, and P. Favaro. Bilayer blind deconvolution with the light field camera. In Proc. of the
IEEE International Conference on Computer Vision Workshop, pages 10–18, 2015.
[18] T. H. Kim and K. M. Lee. Segmentation-free dynamic scene
deblurring. In Proc. of IEEE Conference on Computer Vision
and Pattern Recognition, pages 2766–2773, 2014.
[19] T. H. Kim and K. M. Lee. Generalized video deblurring for
dynamic scenes. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5426–5434,
2015.
[20] R. Köhler, M. Hirsch, B. Mohler, B. Schölkopf, and
S. Harmeling.
Recording and playback of camera
shake: Benchmarking blind deconvolution with a real-world
database. pages 27–40, 2012.
[21] R. Ng. Digital light field photography. PhD thesis, stanford
university, 2006.
[22] R. Ng, M. Levoy, M. Brédif, G. Duval, M. Horowitz, and
P. Hanrahan. Light field photography with a hand-held
plenoptic camera. Computer Science Technical Report,
2(11):1–11, 2005.
[23] H. Park and K. M. Lee. Joint estimation of camera pose,
depth, deblurring, and super-resolution from a blurred image
sequence. In Proc. of the IEEE International Conference on
Computer Vision, 2017.
[24] J. A. Scales and A. Gersztenkorn. Robust methods in inverse
theory. Inverse problems, 4(4):1071, 1988.
[25] A. Sellent, C. Rother, and S. Roth. Stereo video deblurring.
In Proc. of the European Conference on Computer Vision,
pages 558–575, 2016.
[26] Q. Shan, J. Jia, and A. Agarwala. High-quality motion deblurring from a single image. In ACM Trans. on Graphics,
volume 27, page 73, 2008.
[27] A. Snoswell and S. Singh. Light field de-blurring for robotics
applications. In Australasian Conference on Robotics and
Automation, 2014.
[28] P. P. Srinivasan, R. Ng, and R. Ramamoorthi. Light field
blind motion deblurring. Proc. of the IEEE Conference on
Computer Vision and Pattern Recognition, 2017.
[29] J. Sun, W. Cao, Z. Xu, and J. Ponce. Learning a convolutional neural network for non-uniform motion blur removal.
In Proc. of IEEE Conference on Computer Vision and Pattern Recognition, pages 769–777, 2015.
[30] M. W. Tao, P. P. Srinivasan, J. Malik, S. Rusinkiewicz, and
R. Ramamoorthi. Depth from shading, defocus, and correspondence using light-field angular coherence. In Proc.
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 1940–1948, 2015.
[31] T.-C. Wang, A. A. Efros, and R. Ramamoorthi. Occlusionaware depth estimation using light-field cameras. In Proc.
of the IEEE International Conference on Computer Vision,
pages 3487–3495, 2015.
References
[1] A. Beck and M. Teboulle. Fast gradient-based algorithms for
constrained total variation image denoising and deblurring
problems. IEEE Trans. on Image Processing, 18(11):2419–
2434, 2009.
[2] J.-L. Blanco. A tutorial on se (3) transformation parameterizations and on-manifold optimization. University of Malaga,
Tech. Rep, 3, 2010.
[3] Blender Online Community. Blender - A 3D modelling and
rendering package. Blender Foundation, Blender Institute,
Amsterdam,
[4] T. F. Chan and C.-K. Wong. Total variation blind deconvolution. IEEE Trans. on Image Processing, 7(3):370–375, 1998.
[5] P. Chandramouli, D. Perrone, and P. Favaro. Light field blind
deconvolution. CoRR, abs/1408.3686, 2014.
[6] C. Chen, H. Lin, Z. Yu, S. Bing Kang, and J. Yu. Light field
stereo matching using bilateral statistics of surface cameras.
In Proc. of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 1518–1525, 2014.
[7] S. Cho and S. Lee. Fast motion deblurring. In Proc. of the
ACM Trans. on Graphics, volume 28, page 145, 2009.
[8] S. Cho, Y. Matsushita, and S. Lee. Removing non-uniform
motion blur from images. In Proc. of the IEEE International
Conference on Computer Vision, pages 1–8, 2007.
[9] S. Cho, J. Wang, and S. Lee. Video deblurring for handheld cameras using patch-based synthesis. ACM Trans. on
Graphics, 31(4):64, 2012.
[10] D. G. Dansereau, A. Eriksson, and J. Leitner. Richardsonlucy deblurring for moving light field cameras. Proc. of the
IEEE Conference on Computer Vision and Pattern Recognition Workshop, 2017.
[11] R. Fergus, B. Singh, A. Hertzmann, S. T. Roweis, and W. T.
Freeman. Removing camera shake from a single photograph.
In Proc. of the ACM Trans. on Graphics, volume 25, pages
787–794, 2006.
[12] A. Gupta, N. Joshi, C. L. Zitnick, M. Cohen, and B. Curless. Single image deblurring using motion density functions. In Proc. of the European Conference on Computer
Vision, pages 171–184, 2010.
[13] Z. Hu, L. Xu, and M.-H. Yang. Joint depth estimation and
camera shake removal from single blurry image. In Proc.
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 2893–2900, 2014.
[14] Z. Hu and M.-H. Yang. Fast non-uniform deblurring using
constrained camera pose subspace. In Proc. of the British
Machine Vision Conference, volume 2, page 4, 2012.
[15] H.-G. Jeon, J. Park, G. Choe, J. Park, Y. Bok, Y.-W. Tai,
and I. S. Kweon. Accurate depth map estimation from a
lenslet light field camera. In Proc. of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 1547–
1555, 2015.
4329
[32] O. Whyte, J. Sivic, A. Zisserman, and J. Ponce. Non-uniform
deblurring for shaken images. International Journal of Computer Vision, 98(2):168–186, 2012.
[33] Williem and I. K. Park. Robust light field depth estimation
for noisy scene with occlusion. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, pages
4396–4404, 2016.
[34] J. Wulff and M. J. Black. Modeling blurred video with layers.
In Proc. of the European Conference on Computer Vision,
pages 236–252, 2014.
[35] L. Xu and J. Jia. Two-phase kernel estimation for robust
motion deblurring. In Proc. of the European Conference on
Computer Vision, pages 157–170, 2010.
[36] L. Xu and J. Jia. Depth-aware motion deblurring. In Proc. of
IEEE International Conference on Computational Photography, pages 1–8, 2012.
[37] S. Zheng, L. Xu, and J. Jia. Forward motion deblurring. In
Proc. of the IEEE International Conference on Computer Vision, pages 1465–1472, 2013.
4330
| 1cs.CV
|
arXiv:1709.06253v1 [math.AC] 19 Sep 2017
ASYMPTOTIC ASSOCIATE PRIMES
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
Abstract. We investigate three cases regarding asymptotic associate primes.
First, assume (A, m) is a Cohen-Macaulay (CM), non-regular local ring, and
M = SyzA
1 (L) for some maximal CM A-module L which is free on the punctured spectrum. Let I be a normal ideal. In this case, we examine when
m∈
/ Ass(M/I n M ) for all n ≫ 0. We give sufficient evidence to show that this
occurs rarely. Next, assume that (A, m) is Gorenstein, non-regular isolated
singularity, and M is a CM A-module with projdimA (M ) = ∞ and dim(M ) =
dim(A) − 1. Let I be a normal ideal with analytic spread l(I) < dim(A). In
n
this case, we investigate when m ∈
/ Ass TorA
1 (M, A/I ) for all n ≫ 0. We
give sufficient evidence to show that this also occurs rarely. Finally, suppose
A is a local complete intersection ring. For finitely generated A-modules M
and N , we show that if TorA
i (M, N ) 6= 0 for some i > dim(A), then there
exists a non-empty finite subset A of Spec(A) such that for every
p ∈ A, at
least one of the following holds true: (i) p ∈ Ass TorA
(M, N ) for all i ≫ 0;
2i
(ii) p ∈ Ass TorA
2i+1 (M, N ) for all i ≫ 0. We also analyze the asymptotic
A
behaviour of Tori (M, A/I n ) for i, n ≫ 0 in the case when I is principal or I
has a principal reduction generated by a regular element.
1. Introduction
In this paper, we investigate three cases regarding asymptotic associate primes.
We will introduce it one by one.
I: Let (A, m) be a Noetherian local ring of dimension d, and let M be a finitely
generated A-module. By a result of Brodmann [Bro79], there exists n0 such that
the set of associate primes AssA (M/I n M ) = AssA (M/I n0 M ) for all n > n0 . We
denote this eventual constant set by Ass∞
I (M ).
A natural question is when does m ∈ Ass∞
/ Ass∞
I (M ), or the opposite m ∈
I (M )?
In general, this question is hopeless to resolve. So well make try to make a few
assumptions which are quite general but still amenable to answer our question.
(I.1) We first assume that (A, m) is an excellent Cohen-Macaulay local ring with
infinite residue field. We note that this assumption is quite general.
(I.2) We also assume that M is maximal Cohen-Macaulay (MCM). In fact in
the study of modules over Cohen-Macaulay rings the class of MCM modules is the
most natural class to investigate. To keep things interesting we also assume M is
not free. In particular we are assuming A is also not regular.
We note that in general the answer to the question on when does m ∈ Ass∞
I (A)
is not known. However, by results of Ratliff, McAdam a positive answer is known
Date: September 19, 2017.
2010 Mathematics Subject Classification. Primary 13A17, 13A30, 13D07; Secondary 13A15,
13H10.
Key words and phrases. Asymptotic associate primes; Asymptotic grade; Associated graded
rings and modules; Local cohomology; Tor; Complete intersections.
1
2
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
when I is normal, i.e., I n is integrally closed for all n > 1. In this case, it is known
that m ∈ Ass∞
I (A) if and only if l(I), the analytic spread of I is equal to d = dim A;
see [McA83, 4.1]. So our third assumption is
(I.3) I is a normal ideal of height > 2.
Before proceeding further, we want to remark that in analytically unramified
local rings (i.e., its completion is reduced), there exist plenty of normal ideals. In
fact, for any ideal I, it is not terribly difficult to prove that for all n ≫ 0, the ideal
I n is normal (where J denotes the integral closure of J).
Finally, we note that as we are only interested in the question on whether m ∈
Ass∞
I (M ) or not, it is convenient to assume
(I.4) M is free on the punctured spectrum, i.e., MP is free for every prime P 6= m.
We note that (4) is automatic if A is an isolated singularity, i.e., AP is regular
for every prime P 6= m. In general (even when A is not an isolated singularity), any
sufficiently high syzygy of a finite length module (of infinite projective dimension)
will be free on the punctured spectrum.
Remark 1.1. As discussed above, our hypotheses are satisfied by a large class of
rings, modules and ideals.
stating our results, we need to introduce a few notation. Let GI (A) =
L Before
n
n+1
I
/I
be the associated graded ring of A with respect to I. Let GI (A)+ =
Ln>0 n n+1
I /I
be its irrelevant ideal. If M is an A-module, then GI (M ) =
Ln>1 n
n+1
I
M/I
M is the associated graded module of M with respect to I (and
n>0
considered as an GI (A)-module).
Our first result is
Theorem 1.2. With assumptions as in I.1, I.2, I.3 and I.4, suppose that d > 3,
and that M = SyzA
1 (L) for some MCM A-module L. We have that
if m ∈
/ Ass∞
I (M ), then grade(GI n (A)+ , GI n (M )) > 2 for all n ≫ 0.
We note that the assumption M = SyzA
1 (L) for an MCM A-module L is automatically satisfied if A is Gorenstein.
We now describe the significance of Theorem 1.2. The third author of this paper
has worked extensively of associated graded rings and modules. He feels that the
condition grade(GI n (A)+ , GI n (M )) > 2 for n ≫ 0 is quite special. In ‘most cases’
we will have only grade(GI n (A)+ , GI n (M )) = 1 for all n ≫ 0. So Theorem 1.2
implies that in ‘most cases’ we should have m ∈ Ass∞
I (M ).
By a result of Melkersson and Schenzel [MS93,
Theorem
1], it is known that for
n
(E,
A/I
)
is
constant
for n ≫ 0. We denote
an A-module E, the set AssA TorA
1
this stable value by T1∞ (I, E). We note that if L is a (non-free) MCM A-module
n
which is free on the punctured spectrum of A, then TorA
1 (L, A/I ) has finite length
A
∞
for all n > 1. In this case, m ∈
/ T1 (I, L) if and only if Tor1 (L, A/I n ) = 0 for all
∞
n ≫ 0. If m ∈
/ AssI (A) (this holds if I is normal and l(I) < d), then it is easy to
A
n
see that TorA
/ Ass∞
1 (L, A/I ) = 0 for all n ≫ 0 if and only if m ∈
I (Syz1 (L)). If the
latter holds, then by Theorem 1.2, we will have grade(GI n (A)+ , GI n (SyzA
1 (L)) > 2.
Thus another significance of Theorem 1.2 is that it suggests that in ‘most cases’ if
I is a normal ideal with height > 2 and l(I) < d, then for a non-free MCM module,
n
we should have TorA
1 (L, A/I ) 6= 0 for all n ≫ 0.
II: In the previous subsection, we consider T1∞ (I, M ) when M is MCM and
locally free on the punctured spectrum. In this subsection, we consider the case
ASYMPTOTIC ASSOCIATE PRIMES
3
when M is Cohen-Macaulay of dimension d − 1. A trivial case to consider is
when projdimA (M ) is finite. Then it is easy to see that if m ∈
/ Ass∞
I (A), then
∞
m∈
/ T1 (I, M ).
If projective dimension of M is infinite, then we are unable to analyze T1∞ (I, M )
for arbitrary Cohen-Macaulay rings. However, we have made progress in this question when A is an isolated Gorenstein singularity.
In general, when a local ring (R, n) is Gorenstein, and D is a finitely generated
R-module, there exists an MCM approximation of D, i.e., an exact sequence s : 0 →
Y → X → D → 0, where X is an MCM R-module and projdimR (Y ) < ∞. It is
known that if s′ : 0 → Y ′ → X ′ → D → 0 is another MCM approximation of
D, then X and X ′ are stably isomorphic, i.e., there exist finitely generated free
R-modules F, G such that X ⊕ F ∼
= X ′ ⊕ G. It is clear that X is free if and only if
projdimR (D) < ∞. Thus if projdimR (D) = ∞, then SyzR
1 (X) is an invariant of D.
Our second result is
Theorem 1.3 (= 4.6). Let (A, m) be an excellent Gorenstein local ring of dimension d > 3. Suppose A has isolated singularity. Let I be a normal ideal of A with
height(I) > 2 and l(I) < d. Let M be a Cohen-Macaulay A-module of dimension
d − 1 and projdimA (M ) = ∞. Let s : 0 → Y → X → M → 0 be an MCM approximation of M . Set N := SyzA
1 (X). Then the following statements are equivalent:
n
/ AssA (TorA
(i) m ∈
/ T1∞ (I, M ) (i.e., m ∈
1 (M, A/I )) for all n ≫ 0).
∞
n
(ii) m ∈
/ AssI (N ) (equivalently, depth(N/I N ) > 1 for all n ≫ 0).
Furthermore, if this holds true, then grade(GI n (A)+ , GI n (N )) > 2 for all n ≫ 0.
As per our discussion after Theorem 1.2, it follows that in ‘most cases’ we should
n
have m ∈ AssA (TorA
1 (M, A/I )) for all n ≫ 0.
IIIα: Let (A, m) be a local complete intersection ring of codimension c. Let M
and N be finitely generated A-modules. Set
M
M
E(M, N ) :=
ExtiA (M, N ), and T (M, N ) :=
TorA
i (M, N ).
i>0
i>0
b are modules over a ring
b and T (M, N ) ⊗A A
It is well known that E(M, N ) ⊗A A
b 1 , . . . , ξc ], where A
b is the m-adic completion of
of cohomology operators S := A[ξ
b
A. Moreover, E(M, N ) ⊗A A is a finitely generated graded S-module. But the
b is very rarely finitely generated. However, by a result of
S-module T (M, N ) ⊗A A
finite length for all i ≫ 0 (say
Gulliksen [Gul74, Theorem 3.1], if TorA
i (M, N ) has
L
from i > i0 ), then the S-submodule T>i0 (M, N ) := i>i0 TorA
i (M, N ) is *Artinian.
By standard arguments, it follows that for l = 0, 1, the set AssA (Ext2i+l
A (M, N ))
is constant for i ≫ 0. However, we do not have a similar result for Tor. By a result
of Avramov and Buchweitz (Theorem 5.5), the case when TorA
i (M, N ) = 0 for all
i ≫ 0 is well-understood. Our third result is that
Theorem 1.4 (= 5.10). Let A be a local complete intersection ring. Let M and
N be finitely generated A-modules. Assume that TorA
i (M, N ) 6= 0 for some i >
dim(A). Then there exists a non-empty finite subset A of Spec(A) such that for
every p ∈ A, at least one of the following statements holds true:
all i ≫ 0;
(i) p ∈ AssA TorA
2i (M, N ) for
(ii) p ∈ AssA TorA
(M,
N
)
for
all i ≫ 0.
2i+1
4
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
IIIβ: In [GP16, Corollary 4.3], the first and the third author proved that if
(A, m) is a local complete intersection ring, I is an ideal of A, and M, N are finitely
n
generated A-modules, then for every l = 0, 1, the set AssA (Ext2i+l
A (M, N/I N ))
is constant for all i, n ≫ 0. We do not have a similar result for Tor. It follows
from [GP16, Theorem 6.1] that complexity of N/I n N is stable for n ≫ 0. Thus,
n
by results of Avramov and Buchweitz, the case when TorA
i (M, N/I N ) = 0 for all
i, n ≫ 0 is well-understood. Our final result is
Theorem 1.5 (= 5.15). Let A be a local complete intersection ring. Let M be a
finitely generated A-module, and I be an ideal of A. Suppose either I is principal
or I has a principal reduction generated by an A-regular element. Then there exist
n
i0 and n0 such that either TorA
i (M, N/I N ) = 0 for all i > i0 and n > n0 , or there
is a non-empty finite subset A of Spec(A) such that for every p ∈ A, at least one
of the following statements holds true:
n
all i > i0 and n > n0 ;
(i) p ∈ AssA TorA
2i (M, A/I ) for
A
n
(ii) p ∈ AssA Tor2i+1 (M, A/I ) for all i > i0 and n > n0 .
Techniques used to prove our results: To prove Theorems 1.4 and 1.5, we use
the well-known technique of Eisenbud operators over resolutions of modules over
complete complete-intersection rings. We also use results of Gulliksen, Avramov
and Buchweitz stated above.
However, to prove Theorems 1.2 and 1.3, we use a new technique in the study
of asymptotic primes, i.e., we investigate the function
I
ξM
(n) := grade(GI n (A)+ , GI n (M )).
Note that by a result of Elias [Eli04, Proposition 2.2], we have that depth(GI n (A))
is constant for all n ≫ 0 (and a similar argument works for modules). However, the
I
function ξM
(when dim(A/I) > 0) has not been investigated before, neither in the
study of blow-up algebra’s or with the connection with associate primes. Regarding
I
ξM
, we prove two results: The first is
Theorem 1.6 (= 3.9). Let (A, m) be a Noetherian local ring. Let I be an ideal of A,
and M be a finitely generated A-module such that grade(I, M ) = g > 0. Then either
grade(GI n (A)+ , GI n (M )) = 1 for all n ≫ 0, or grade(GI n (A)+ , GI n (M )) > 2 for
all n ≫ 0.
Our second result in this direction is
Theorem 1.7 (= 3.10). Let (A, m) be a Cohen-Macaulay local ring, and I be an
ideal of A such that height(I) > dim(A) − 2. Let M be an MCM A-module. Then
grade(GI n (A)+ , GI n (M )) is constant for all n ≫ 0.
I
Although he does not have an example, the third author feels that ξM
may not
be constant for n ≫ 0 if dim(A/I) > 3.
Remark 1.8. In [HM99, Theorem 3.4], Huckaba and Marley proved that if A is
Cohen-Macaulay of dimension > 2, and if I is a normal ideal with grade(I) > 1,
then depth(GI n (A)) > 2 for all n ≫ 0. A crucial ingredient for the proofs of our
results is to compute grade(GI n (A)+ , GI n (A)) for all n ≫ 0 when I is normal. So
we prove the following result.
Theorem 1.9 (= 3.4). Let I be a normal ideal of A such that grade(I) > 2. Then
grade(GI n (A)+ , GI n (A)) > 2 for all n ≫ 0.
ASYMPTOTIC ASSOCIATE PRIMES
5
We now describe in brief the contents of this paper. In Section 2, we discuss a
few preliminaries on grade and local cohomology that we need. In Section 3, we inI
(n) := grade(GI n (A)+ , GI n (M )) and prove Theorems 1.6,
vestigate the function ξM
1.7 and 1.9. In Section 4, we prove Theorems 1.2 and 1.3. Finally, in Section 5, we
prove Theorems 1.4 and 1.5.
2. Preliminaries on grade and local cohomology
Throughout this article, all rings are assumed to be commutative Noetherian
rings with identity. Throughout, let (A, m) be a local ring of dimension d with
infinite residue field, and M be a finitely generated A-module. Let I be an ideal
of A (which need not be m-primary). If p ∈ M is non-zero, and j is the largest
integer such that p ∈ I j M , then p∗ denotes the image of p in I j M/I j+1 M , and
L
L
let 0∗ = 0. Set R(I) := n>0 I n tn , the Rees ring, and R̂(I) := n∈Z I n tn is the
n
extended Rees
L ring nof An with respect to I, where I = A forLevery nn 6n0. Set
R(I, M ) := n>0 I M t , the Rees module, and R̂(I, M ) := n∈Z I M t is the
L
extended Rees Module of M with respect to I. Let GI (A) := n>0 I n /I n+1 be the
L
associated graded ring of A with respect to I, and GI (M ) := n>0 I n M/I n+1 M
be the associated graded
of M with respect to I. Throughout
article,
L this
L module
j
j+1
n n
of
we denote the ideal
n>1 I /I
n>1 I t of R(I) by R+ , and the ideal
GI (A) by G+ .
L
2.1. Set LI (M ) := n>0 M/I n+1 M . The A-module LI (M ) can be given an R(I)module structure as follows. The Rees ring R(I) is a subring of R̂(I), and R̂(I)
is
of S = A[t, t−1 ]. So S is an R(I)-module. Therefore M [t, t−1 ] =
La subring
n
n∈Z M t = M ⊗A S is an R(I)-module. The exact sequence
(2.1.1)
0 −→ R̂(I, M ) −→ M [t, t−1 ] −→ LI (M )(−1) −→ 0
defines an R(I)-module structure on LI (M )(−1), and hence on LI (M ).
The following result is well known and easy to prove.
L
Lemma 2.2. Let R =
i>0 Ri be a graded ring. Let E be a graded R-module
(need not be finitely generated). Then the following statements hold true:
(i) If En = 0 for all n ≫ 0, and there is an injective graded homomorphism
E(−1) ֒→ E, then E = 0.
(ii) If En = 0 for all n ≪ 0, and there is an injective graded homomorphism
E ֒→ E(−1), then E = 0.
L
2.3. Let R = i>0 Ri be a graded ring. Let E be a graded R-module. Let l be a
L
positive integer. The lth Veronese subring of R is defined by R<l> := n>0 Rnl ,
L
and the lth Veronese submodule of E is defined to be E <l> := n∈Z Enl . It can
be observed that E <l> is a graded R<l> -module.
Remark 2.4.
(i) LI (M )(−1) behaves well with respect to the Veronese functor. It can be easily
checked that
l
LI (M )(−1)<l> = LI (M )(−1).
(ii) [HM99, Proposition 2.5] Veronese functor commutes with local cohomology:
Let J be a homogeneous ideal of R. Then, for every i > 0, we have
<l>
∼
HJi (E)
= HJi <l> (E <l> ) as graded R<l> -modules.
6
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
Although, in general, LI (M ) is not finitely generated as an R(I)-module, but it
has the following vanishing property.
Lemma 2.5. Suppose grade(I, M ) = g > 0. Then, for every 0 6 i 6 g − 1,
i
HR+
(LI (M ))n = 0 for all n ≫ 0.
Proof. Since grade(I, M ) = g > 0, there exists an M -regular sequence x1 , . . . , xg
in I. It can be observed that x1 t, . . . , xg t ∈ R(I)1 becomes an M [t, t−1 ]-regular
i
sequence. So HR+
(M [t, t−1 ]) = 0 for 0 6 i 6 g − 1. Therefore, in view of the short
exact sequence (2.1.1) and using the corresponding long exact sequence in local
cohomology, we get that
(2.5.1)
(2.5.2)
i+1
i
HR+
(LI (M )(−1)) ∼
(R̂(I, M )) for 0 6 i 6 g − 2, and
= HR+
g−1
g
HR+
(LI (M )(−1)) ⊆ HR+
(R̂(I, M )).
L
0
i
(U ) = U and HR
(U ) = 0
Set U := n<0 M tn . Since U is R+ torsion, we have HR+
+
for all i > 1. Considering the short exact sequence of R(I)-modules
M
0 −→ R(I, M ) −→ R̂(I, M ) −→
M tn −→ 0,
n<0
the corresponding long exact sequence in local cohomology yields the exact sequence
(2.5.3)
1
1
(R(I, M )) −→ HR+
(R̂(I, M )) −→ 0,
0 −→ U −→ HR
+
(2.5.4)
i
i
(R̂(I, M )) for i > 2.
(R(I, M )) ∼
and HR
= HR
+
+
i
It is well known that for each i > 0, HR
(R(I, M ))n = 0 for all n ≫ 0. Therefore,
+
i
in view of (2.5.3) and (2.5.4), for each i > 0, HR
(R̂(I, M ))n = 0 for all n ≫ 0.
+
Hence the lemma follows from (2.5.1) and (2.5.2).
2.6. The Ratliff-Rush closure of M with respect to I is defined to be
[
g :=
IM
(I m+1 M :M I m ).
m>1
nM =
It is shown in [Nag05, Proposition 2.2.(iv)] that if grade(I, M ) > 0, then I]
I n M for all n ≫ 0. This motivates the following definition:
i M = I i M for all i > n}.
ρI (M ) := min{n : Ig
We call ρI (M ) the Ratliff-Rush number of M with respect to I.
2.7. Let I = (x1 , . . . , xm ). L
Set S := A[X1 , . . . , Xm ] with degA = 0 and degXi = 1
for i = 1, . . . , m. Then S = n>0 Sn , where Sn is the collection of all homogeneous
L
polynomials of degree n. So A = S0 . We denote the ideal n>1 Sn of S by S+ . We
have a surjective homogeneous homomorphism of A-algebras, namely ϕ : S → R(I),
where ϕ(Xi ) = xi t. We also have the natural map ψ : R(I) → GI (A). Note that
ϕ(S+ ) = R+ ,
ψ(R+ ) = G+
and ψ ◦ ϕ(S+ ) = G+ .
By graded independence theorem ([BS98, 13.1.6]), it does not matter which ring we
use to compute local cohomology. So now onwards, we simply use H i (−) instead
i
i
of HR
(−) or HG
(−).
+
+
ASYMPTOTIC ASSOCIATE PRIMES
7
2.8. The natural map 0 → I n M /I n+1 M → M/I n+1 M → M/I n M → 0 induces
the first fundamental exact sequence (as in [Put07, (5)]) of R(I)-modules:
0 −→ GI (M ) −→ LI (M ) −→ LI (M )(−1) −→ 0.
(2.8.1)
2.9. Let x be an M -superficial element with respect to I. Set N = M/xM . For
every n > 1, we have an exact sequence of A-modules:
M
M ψn
N
(I n+1 M :M x)
−→ n+1
−→ n
−→ n+1 −→ 0,
I nM
I M
I
M
I
N
where ψn (m + I n M ) = xm + I n+1 M for m ∈ M . These sequences induce the
second fundamental exact sequence (as in [Put07, 6.2]) of R(I)-modules:
0 −→
(2.9.1)
ρ
Ψ
xt
LI (M ) −→ LI (N ) −→ 0,
0 −→ B I (x, M ) −→ LI (M )(−1) −→
where Ψxt is multiplication by xt ∈ R(I)1 , and
M
B I (x, M ) :=
(I n+1 M :M x)/I n M .
n>0
2.10. It is shown in [Put07, Proposition 4.7] that if grade(I, M ) > 0, then
ρI (M)−1
0
(LI (M ))
HR+
∼
=
M
i=0
i+1 M
I^
.
i+1
I M
2.11. Let x ∈ I rI 2 . If x∗ is GI (M )-regular, then GI (M )/x∗ GI (M ) ∼
= GI (M/xM )
(the proof in [Put03, Theorem 7] generalizes in this context).
We now show that grade(G+ , GI (M )) is always bounded by grade(I, M ).
Lemma 2.12. We have that grade(G+ , GI (M )) 6 grade(I, M ).
Proof. We prove the result by induction on g := grade(I, M ). Let us first consider
the case g = 0. If possible, suppose grade(G+ , GI (M )) > 1. Then there is a
GI (M )-regular element u = x + I 2 ∈ G1 for some x ∈ I. Since grade(I, M ) = 0,
x cannot be M -regular, i.e., there exists a 6= 0 in M such that xa = 0. By Krull’s
Intersection Theorem, there exists c > 0 such that a ∈ I c M r I c+1 M . Then
a∗ 6= 0 in I c M/I c+1 M , but ua∗ = xa + I c+2 M = 0 yields that a∗ = 0, which is a
contradiction. Therefore grade(G+ , GI (M )) = 0.
We assume the result for g = l − 1, and prove it for g = l (> 1). If possible, suppose the result is not true for g = l, i.e, grade(I, M ) = l and grade(G+ , GI (M )) >
l + 1. Then there exists a GI (M )-regular sequence u1 , . . . , ul+1 ∈ G1 , where
ui = xi + I 2 for some xi ∈ I, 1 6 i 6 l + 1. By applying a similar procedure as
above, one obtains that x1 is M -regular. We note that grade(I, M/x1 M ) = l − 1,
but u2 , . . . , ul+1 is regular on GI (M )/x∗1 GI (M ) ∼
= GI (M/x1 M ); see 2.11. This
contradicts our induction hypothesis.
The result below gives a relationship between the first few local cohomologies of
LI (M ) and that of GI (M ).
Theorem 2.13. Suppose grade(I, M ) = g > 0. Then, for s 6 g − 1, we have
H i (LI (M )) = 0 for all 0 6 i 6 s if and only if H i (GI (M )) = 0 for all 0 6 i 6 s.
Proof. In view of the short exact sequence (2.8.1) and the corresponding long exact
sequence in local cohomology, it follows that if H i (LI (M )) = 0 for i = 0, . . . , s,
then H i (GI (M )) = 0 for i = 0, . . . , s. We now prove the converse part by using
8
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
induction on s. For s = 0, let us assume that H 0 (GI (M )) = 0. Then (2.8.1)
yields an injective graded homomorphism H 0 (LI (M )) ֒→ H 0 (LI (M ))(−1). Hence,
in view of Lemma 2.2.(ii), we obtain that H 0 (LI (M )) = 0.
We now assume the result for s = l − 1, and prove it for s = l, where l > 1.
Let H i (GI (M )) = 0 for 0 6 i 6 l. So grade(G+ , GI (M )) > l + 1. Then there
is x ∈ I r I 2 such that x∗ is GI (M )-regular. Hence it can be easily shown that
(I n+1 M :M x) = I n M for all n > 0. In particular, we have B I (x, M ) = 0 and
x is M -superficial. Set N := M/xM . Note that GI (M )/x∗ GI (M ) ∼
= GI (N ) (see
2.11). So grade(G+ , GI (N )) > l, and hence H i (GI (N )) = 0 for 0 6 i 6 l − 1.
Therefore, by induction hypothesis, we have H j (LI (N )) = 0 for 0 6 j 6 l − 1.
Since B I (x, M ) = 0, the short exact sequence (2.9.1) and the corresponding long
exact sequence in local cohomology provide us the exact sequences:
(2.13.1)
0 −→ H i (LI (M ))(−1) −→ H i (LI (M ))
for 0 6 i 6 l.
In view of Lemma 2.12, grade(I, M ) > grade(G+ , GI (M )) > l + 1. Hence, by
Lemma 2.5, for every 0 6 i 6 l, H i (LI (M ))n = 0 for all n ≫ 0. Therefore it
follows from (2.13.1) and Lemma 2.2(i) that H i (LI (M )) = 0 for all 0 6 i 6 l.
As a consequence of Theorem 2.13, we obtain the following characterization of
grade(G+ , GI (M )) in terms of local cohomology of LI (M ).
Corollary 2.14. Suppose grade(I, M ) = g > 0. Then
grade(G+ , GI (M )) = min{i : H i (LI (M )) 6= 0, where 0 6 i 6 g}.
Proof. It is well known that
i
(GI (M )) 6= 0}.
grade(G+ , GI (M )) = min{i : HG
+
By Lemma 2.12, we have grade(G+ , GI (M )) 6 g. Set
α := min{i : H i (LI (M )) 6= 0, where 0 6 i 6 g}.
By considering (2.8.1), it can be easily observed that H i (LI (M )) 6= 0 for some i
with 0 6 i 6 grade(G+ , GI (M )) (6 g). So α 6 grade(G+ , GI (M )). Hence, by
virtue of Theorem 2.13, it follows that α = grade(G+ , GI (M )).
3. Asymptotic grade for associated graded modules
In the present section, we explore the asymptotic behaviour of the associated
graded modules for powers of an ideal. We particularly study its grade with respect
to the irrelevant ideals of associated graded rings.
Throughout this section, we work with the following hypothesis, but we do not
need Cohen-Macaulay assumption everywhere.
Hypothesis 3.1. Let (A, m) be a Cohen-Macaulay local ring with infinite residue
field, and M be a Cohen-Macaulay A-module. Let I be an ideal of A such that
grade(I, M ) = g > 0.
3.2 (A few invariants). In our study, we use the following invariants.
(i) ξI (M ) := min{g, i : H i (LI (M ))−1 6= 0, or H i (LI (M ))j 6= 0 for infinitely
many j < 0, where i varies in 0 6 i 6 g − 1}. Note that 1 6 ξI (M ) 6 g.
(ii) The amplitude of M with respect to I is defined to be
ampI (M ) := max{|n| : H i (LI (M ))n−1 6= 0 for some 0 6 i 6 ξI (M ) − 1}.
It follows from (i) and Lemma 2.5 that ampI (M ) < ∞.
ASYMPTOTIC ASSOCIATE PRIMES
9
(iii) Let N be a graded module (not necessarily finitely generated). Define
end(N ) := sup{n ∈ Z : Nn 6= 0}.
i
(iv) By Lemma 2.5, for every 0 6 i 6 g − 1, we have HR+
(LI (M ))n = 0 for all
n ≫ 0. So we set
i
bIi (M ) := end HR+
(LI (M )) for every 0 6 i 6 g − 1.
We start by showing a special property of the first local cohomology of LI (M ).
Lemma 3.3. For a fixed integer c < 0, the following conditions are equivalent:
(i) H 1 (LI (M ))c = 0.
(ii) H 1 (LI (M ))j = 0 for all j 6 c.
Proof. We only need to prove {(i) ⇒ (ii)}. Suppose H 1 (LI (M ))c = 0. Let x be
an M -superficial element with respect to I. Then (I n+1 M :M x) = I n M for every
n ≫ 0, i.e, B I (x, M ) is G+ torsion. Therefore H 0 (B I (x, M )) = B I (x, M ), and
H i (B I (x, M )) = 0 for all i > 1. Hence, by splitting (2.9.1) into two short exact
sequences, and considering the corresponding long exact sequences, one obtains the
following exact sequence:
(3.3.1)
0 → B I (x, M ) −→H 0 (LI (M ))(−1) −→ H 0 (LI (M )) −→ H 0 (LI (N ))
−→H 1 (LI (M ))(−1) −→ H 1 (LI (M )) −→ H 1 (LI (N )),
where N = M/xM . Therefore, for every n < 0, since H 0 (LI (N ))n = 0, we have
the following exact sequence:
0 −→ H 1 (LI (M ))n−1 −→ H 1 (LI (M ))n .
Hence, since H 1 (LI (M ))c = 0, it follows that H 1 (LI (M ))j = 0 for all j 6 c.
In [HM99, Theorem 3.4], Huckaba and Marley proved that if A is Cohen-Macaulay
with dim(A) > 2, and I is a normal ideal with grade(I) > 1, then depth(GI n (A)) >
2 for all n ≫ 0. Here we prove a similar result for grade(GI n (A)+ , GI n (A)).
Theorem 3.4. Let I be a normal ideal of A with grade(I) > 2. Also assume that
A is excellent. Then
grade(GI n (A)+ , GI n (A)) > 2 for all n ≫ 0.
i
Proof. Set u := max{bI0 (A), bI1 (A)} + 2. Let l > u. We write HR
(LI (A)) =
+
L
i
i
n∈Z Vn as it is a graded R(I)-module. It can be observed that Vnl−1 = 0 for all
n > 1 and i = 0, 1. We note that
l
l
I
I
i
i
∼
L
(A)(−1)
L
(A)
(−1)
H
HR(I
=
l
l)
R(I )+
+
<l>
i
∼
(3.4.1)
[by Remark 2.4.(i)]
LI (A)(−1)
= H(R
<l>
+)
<l>
i
∼
LI (A)(−1)
[by Remark 2.4.(ii)]
= HR
+
M
i
∼
Vnl−1
.
=
n∈Z
i
Therefore, for every i ∈ {0, 1}, since Vnl−1
= 0 for all n > 1, we have
l
I
i
i
LK (A) n = 0 for all n > 0,
(A))n = 0, i.e., HR(K)
(3.4.2)
HR(I
l ) (L
+
+
10
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
0
(LK (A)) = 0. We now show
where K := I l . In particular, it follows that HR(K)
+
1
(LK (A)) = 0. Note that K is integrally closed. Therefore, by virtue of
that HR(K)
+
[HU14, Theorem 2.1], after a flat extension, there exists a superficial element x ∈ K
such that the ideal J := K/(x) is integrally closed in B := A/(x). In view of a
0
(LK (B))n = 0 for
sequence like (3.3.1), by applying (3.4.2), we obtain that HR(K)
+
e = 0 as J is integrally
(LK (B)) ∼
all n > 1. Hence, by 2.10, we have H 0
= J/J
R(K)+
closed; see [RR78, 2.3.3]. Therefore, for every n, a sequence like (3.3.1) yields the
following exact sequence:
(3.4.3)
1
1
(LK (A))n .
(LK (A))n−1 −→ HR(K)
0 −→ HR(K)
+
+
1
(LK (A))n = 0 for all n > 0, it can be proved by repeatedly applying
Since HR(K)
+
1
1
(LK (A)) = 0. Thus,
(LK (A))n = 0 for all n, and hence HR(K)
(3.4.3) that HR(K)
+
+
by virtue of Corollary 2.14, we have that grade(GI l (A)+ , GI l (A)) > 2, and this
holds true for every l > u, which completes the proof of the theorem.
Remark 3.5. We have used [HU14, Theorem 2.1] crucially in the proof above.
This is the only place where we need that the ring is excellent.
Since grade(I n ) = grade(I) for every n > 1, as an immediate consequence of
Theorem 3.4 and Lemma 2.12, one obtains the following result.
Corollary 3.6. Let I be a normal ideal of A with grade(I) = 2. Also assume that
A is excellent. Then
grade(GI n (A)+ , GI n (A)) = 2 for all n ≫ 0.
The following theorem gives an asymptotic lower bound of grade of associated
graded modules for powers of an ideal.
Theorem 3.7. For each l > ampI (M ), we have grade(GI l (A)+ , GI l (M )) > ξI (M ).
Proof. Set E i := H i (LI (M )(−1)), and u := ξI (M ). Fix an arbitrary l > ampI (M ).
i
Also fix i with 0 6 i 6 u − 1. Then, for n 6= 0, we have Enl
= H i (LI (M ))nl−1 = 0
i
i
I
as |n|l > l > ampL
I (M ). Also E0 = H (L (M ))−1 = 0 since 0 6 i 6 ξI (M ) − 1.
i
= 0. So, by Remark 2.4 (as in (3.4.1)), it follows that
Hence (E i )<l> = n∈Z Enl
l
<l>
I
i
I
i
L
(M
)(−1)
=
H
L
(M
)(−1)
= 0.
(3.7.1)
HR(I
l)
R+
+
l
I
i
L
(M
)
= 0 for all 0 6 i 6 u − 1. Hence, by virtue of
Therefore HR(I
l)
+
Corollary 2.14, grade(GI l (A)+ , GI l (M )) > u = ξI (M ) for all l > ampI (M ).
The following corollary shows that how the vanishing of a single component of
certain local cohomology plays a crucial role in the study of grade of asymptotic
associated graded modules.
Corollary 3.8. The following conditions are equivalent:
(i) H 1 (LI (M ))−1 = 0.
(ii) grade(GI l (A)+ , GI l (M )) > 2 for all l > ampI (M ).
(iii) grade(GI l (A)+ , GI l (M )) > 2 for some l > 1.
ASYMPTOTIC ASSOCIATE PRIMES
11
Proof. (i) ⇒ (ii): Let H 1 (LI (M ))−1 = 0. So, by Lemma 3.3, H 1 (LI (M ))j = 0 for
all j 6 −1. Therefore, since ξI (M ) > 1 (always), it follows that ξI (M ) > 2. Hence,
in view of Theorem 3.7, grade(GI l (A)+ , GI l (M )) > 2 for all l > ampI (M ).
(ii) ⇒ (iii): It holds trivially.
(iii) ⇒ (i): Suppose grade(GI l (A)+ , GI l (M )) > 2 for some l > 1. Then it
l
1
LI (M ) = 0. Therefore, as in (3.7.1), we
follows from Corollary 2.14 that HR(I
l)
+
<l>
i
LI (M )(−1)
= 0, and hence its 0th component provides us
obtain that HR
+
H 1 (LI (M ))−1 = 0.
As a consequence, we obtain the following asymptotic behaviour of associated
graded modules for powers of an ideal.
Corollary 3.9. Exactly one of the following alternative statements must hold true:
(i) grade(GI n (A)+ , GI n (M )) = 1 for all n > ampI (M ).
(ii) grade(GI n (A)+ , GI n (M )) > 2 for all n > ampI (M ).
Proof. Since ξI (M ) > 1, by virtue of Theorem 3.7, grade(GI n (A)+ , GI n (M )) > 1
for all n > ampI (M ). Hence the result follows from Corollary 3.8.
Here we prove our main result of this section.
Theorem 3.10. With Hypothesis 3.1, assume that height(I) > dim(A) − 2. Then
grade(GI l (A)+ , GI l (M )) = ξI (M ) for every l > ampI (M ).
Proof. Set u := ξI (M ). By virtue of Theorem 3.7, grade(GI l (A)+ , GI l (M )) > u
for every l > ampI (M ). If possible, suppose that grade(GI l (A)+ , GI l (M )) > u for
l
u
LI (M ) = 0. Thus,
some l > ampI (M ). Then, in view of Corollary 2.14, HR(I
l)
+
<l>
u
LI (M )(−1)
= 0, and hence
as in (3.7.1), we obtain that HR
+
(3.10.1)
H u (LI (M ))nl−1 = 0 for all n ∈ Z.
We note that u < grade(GI l (A)+ , GI l (M )) 6 g; see Lemma 2.12.
The long exact sequence corresponding to (2.8.1) provides an exact sequence:
(3.10.2)
H u−1 LI (M ) (−1) → H u (GI (M )) → H u (LI (M )) → H u LI (M ) (−1).
Since u = ξI (M ), it follows from the definition of ξI (M ) that H u−1 (LI (M ))n = 0
for all n ≪ 0. Therefore (3.10.1) and (3.10.2) yield that H u (GI (M ))nl−1 = 0
for all n ≪ 0. Hence, since H u (GI (M )) is tame (due to [BFL04, Lemma 4.3]),
there exists some c < 0 such that H u (GI (M ))j = 0 for all j 6 c. (Note that
dim(A/I) 6 dim(A) − height(I) 6 2, and GI (M ) is a finitely generated graded
GI (A)-module).
Since H u (GI (M ))j = 0 for all j 6 c, (3.10.2) produces an exact sequence
(3.10.3)
0 −→ H u (LI (M ))j −→ H u (LI (M ))j−1 for every j 6 c.
Therefore, if m, n 6 c are integers such that m 6 n, then H u (LI (M ))n can be considered as a submodule of H u (LI (M ))m . Using this fact and (3.10.1), one can prove
that H u (LI (M ))j = 0 for all j 6 n′ l, where n′ is a fixed integer such that n′ l 6 c.
Thus we have H u (LI (M ))j = 0 for all j ≪ 0, and H u (LI (M ))−1 = 0 by (3.10.1).
This contradicts that u = ξI (M ) < g. Therefore grade(GI l (A)+ , GI l (M )) = ξI (M )
for every l > ampI (M ).
12
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
∞
4. On the sets Ass∞
I (M ) and T1 (I, M )
Let (A, m) be a local ring. Let I be an ideal of A, and M be a finitely generated A-module. By a result of Brodmann [Bro79], there exists n0 such that
AssA (M/I n M ) = AssA (M/I n0 M ) for all n > n0 . The eventual constant set (i.e.,
AssA (M/I n0 M )) is denoted by Ass∞
I (M ). In [MS93, Theorem 1], Melkersson and
Schenzel generalized Brodmann’s
result
by proving that for every fixed i > 0, the
n
(M,
A/I
)
is
constant
for
all n ≫ 0. We denote this stable value
set AssA TorA
i
(M
)
is
nothing
but T0∞ (I, M ). In this section, we
by Ti∞ (I, M ). Note that Ass∞
I
∞
mainly study the question that when does m ∈ Ass∞
I (M ) (resp. T1 (I, M ))? Our
first result in this direction is regarding the set Ass∞
(M
).
I
Theorem 4.1. Let (A, m) be a Cohen-Macaulay local, non-regular ring, and L
be an MCM A-module. Suppose M = SyzA
1 (L) (6= 0), and MP is free for every
P ∈ Spec(A) r {m}. Let I be an ideal of A such that grade(GI n (A)+ , GI n (A)) > 2
for all n ≫ 0.
If m ∈
/ Ass∞
I (M ), then grade(GI n (A)+ , GI n (M )) > 2 for every n > ampI (M ).
Proof. Since L is an MCM A-module, every A-regular element is L-regular. By
virtue of Lemma 2.12, from the given hypotheses, it follows that grade(I, A) > 0,
and hence grade(I, L) > 0. So, by Corollary 3.9, grade(GI n (A)+ , GI n (L)) > 1 for
all n ≫ 0. Therefore, in view of Corollary 2.14, we obtain that
n
I
0
L
(L)
= 0 for all n ≫ 0.
(4.1.1)
HR(I
n)
+
Note that M is an MCM A-module. So as above grade(I, M ) > 0.
We have a short exact sequence
(4.1.2)
0 −→ M −→ F −→ L −→ 0,
where F is a free A-module. For every n, by applying (A/I n ) ⊗A − on (4.1.2), we
obtain an exact sequence:
(4.1.3)
n
n
n
n
0 −→ TorA
1 (A/I , L) −→ M/I M −→ F/I F −→ L/I L −→ 0.
For every P ∈ Spec(A) r {m}, since MP is free,
we get that LP is free, and hence
A
n
n
TorA
1 (A/I , L)P = 0. So AssA Tor1 (A/I , L) ⊆ {m} for every n. Therefore,since
A
n
m∈
/ Ass∞
I (M ), in view of (4.1.3), it can be deduced that AssA Tor1 (A/I , L) = φ
A
(empty set) for all n ≫ 0, and hence there is c′ > 1 such that Tor1 (A/I n , L) = 0
for every n > c′ . Thus (4.1.3) yields an exact sequence:
(4.1.4)
0 −→ M/I n M −→ F/I n F −→ L/I n L −→ 0
for every n > c′ . In particular, for every n > c′ , we have short exact sequences:
0 −→ M/I nk M −→ F/I nk F −→ L/I nk L −→ 0
for all k > 1, which induce an exact sequence of R(I)-modules:
(4.1.5)
n
n
n
0 −→ LI (M )(−1) −→ LI (F )(−1) −→ LI (L)(−1) −→ 0.
The corresponding long exact sequence of local cohomology modules yields
n
n
n
I
I
0
I
0
0
(4.1.6) 0 −→HR(I
L
(L)
L
(F
)
−→
H
L
(M
)
−→
H
n
n
n)
R(I )+
R(I )+
+
n
n
I
I
1
1
−→HR(I n )+ L (M ) −→ HR(I n )+ L (F ) .
ASYMPTOTIC ASSOCIATE PRIMES
13
Since grade(GI n (A)+ , GI n(A)) > 2 for all n ≫ 0, by virtue of Corollary 2.14, we
n
i
get that HR(I
LI (A) = 0 for i = 0, 1, and for all n ≫ 0. Therefore
n)
+
n
n
I
I
1
0
L
(F
)
for all n ≫ 0.
L
(F
)
=
0
=
H
(4.1.7)
HR(I
n
n)
R(I )+
+
It follows from (4.1.1), (4.1.6) and (4.1.7) that
n
n
I
I
1
0
L
(M
)
L
(M
)
=
0
=
H
HR(I
n
n)
R(I )+
+
for all n ≫ 0.
Hence, in view of Corollaries 2.14 and 3.8, grade(GI n (A)+ , GI n (M )) > 2 for every
n > ampI (M ), which completes the proof of the theorem.
We now give
Proof of Theorem 1.2. This follows from Theorem 4.1 and Theorem 3.4.
The following result gives a variation of Theorem 4.1.
Theorem 4.2. Let (A, m) be a Cohen-Macaulay local ring of dimension d > 3.
Set M := SyzA
1 (L) for some MCM A-module L. Let I be a locally complete intersection ideal of A with height(I) = d − 1, the analytic spread l(I) = d, and
grade(GI n (A)+ , GI n (A)) > 2 for all n ≫ 0. We have
if m ∈
/ Ass∞
I (M ), then grade(GI n (A)+ , GI n (M )) > 2 for every n > ampI (M ).
Remark 4.3. See [HH01, 2.2] for cases when the hypotheses on the ideal I are
satisfied.
n
Proof of Theorem 4.2. We claim that TorA
1 (A/I , L) has finite length for all n ≫ 0.
n
To show this, consider P ∈ Spec(A) r {m}. If LP is free, then TorA
1 (A/I , L)P = 0.
A
So we may assume that LP is not free. If I * P , then also Tor1 (A/I n , L)P = 0
for every n > 1. So we assume that I ⊆ P . Since height(I) = d − 1 and P 6= m,
we have height(P ) = d − 1, and hence P is minimal over I. (So there are finitely
many such prime ideals). Note that IP is a P AP -primary ideal of AP . Since I is
locally complete intersection, IP is generated by an AP -regular sequence of length
n
d − 1. Therefore, in view of [Put03, Remark 20], we obtain that TorA
1 (A/I , L)P =
A
n
0 for all n ≫ 0. Hence Tor1 (A/I , L) has finite length for all n ≫ 0. Now,
along with the same arguments as in the proof of Theorem 4.1, it follows that
grade(GI n (A)+ , GI n (M )) > 2 for every n > ampI (M ).
4.4. (MCM approximations and an invariant of modules). Let (A, m) be a Gorenstein local ring. Consider a finitely generated A-module M . By virtue of [AB89,
Theorem A], there is an MCM approximation of M , i.e., a short exact sequence
s : 0 → Y → X → M → 0 of A-modules, where X is MCM and Y has finite injective
dimension (equivalently, Y has finite projective dimension since A is Gorenstein).
We say that X is an MCM approximation of M . In view of [AB89, Theorem B], if
s′ : 0 → Y ′ → X ′ → M → 0 is another MCM approximation of M , then X and X ′
are stably isomorphic, i.e., there exist finitely generated free A-modules F and G
A
A
′
∼
such that X ⊕ F ∼
= X ′ ⊕ G, and hence SyzA
1 (X) = Syz1 (X ). Thus Syz1 (X) is an
A
invariant of M . Note that Syz1 (X) = 0 if and only if projdimA (M ) is finite.
We use the following lemma to prove our result on T1∞ (I, M ).
14
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
Lemma 4.5. Let (A, m) be a Gorenstein local ring of dimension d. Suppose A has
isolated singularity. Let I be a normal ideal of A such that l(I) < d. Let M be a
Cohen-Macaulay A-module of dimension d − 1, and projdimA (M ) = ∞. Let XM
be an MCM approximation of M . Then the following statements are equivalent:
n
(i) m ∈
/ T1∞ (I, M ) (i.e., m ∈
/ AssA (TorA
1 (M, A/I )) for all n ≫ 0).
A
n
(ii) Tor1 (XM , A/I ) = 0 for all n ≫ 0.
Proof. Since XM is an MCM approximation of M , and depth(M ) > d − 1, there is
a short exact sequence 0 → F → XM → M → 0, where F is a free A-module. The
corresponding long exact sequences of Tor-modules yield an exact sequence
(4.5.1)
A
n
n
0 −→ TorA
1 (XM , A/I ) −→ Tor1 (M, A/I ) −→
F/I n F −→ XM /I n XM −→ M/I n M −→ 0 for every n > 1.
(i) ⇒ (ii): Since A has isolated singularity, it follows that (XM )P is free AP n
module for every P ∈ Spec(A) r {m}. So TorA
1 (XM , A/I ) has finite length, and
A
n
hence AssA Tor1 (XM , A/I ) ⊆ {m} for every n > 1. Therefore, since m ∈
/
n
(X
,
A/I
)
=
φ
for
all
T1∞ (I, M ), in view of (4.5.1), we obtain that AssA TorA
M
1
n
n ≫ 0, which implies that TorA
1 (XM , A/I ) = 0 for all n ≫ 0.
(ii) ⇒ (i): Since I is normal, and l(I) < d, by virtue of [McA83, Proposition 4.1],
A
n
we have m ∈
/ Ass∞
I (A). In view of (4.5.1), since Tor1 (XM , A/I ) = 0 for all n ≫ 0,
∞
∞
∞
/ T1 (I, M ).
it follows that T1 (I, M ) ⊆ AssI (A), and hence m ∈
The following theorem provides us a necessary and sufficient condition for ‘m ∈
T1∞ (I, M )’ on certain class of ideals and modules over a Gorenstein local ring.
Theorem 4.6. Let (A, m) be an excellent Gorenstein local ring of dimension d.
Suppose A has isolated singularity. Let I be a normal ideal of A with height(I) >
2 and l(I) < d. Let M be a Cohen-Macaulay A-module of dimension d − 1
and projdimA (M ) = ∞. Let XM be an MCM approximation of M . Set N :=
SyzA
1 (XM ). Then the following statements are equivalent:
n
/ AssA (TorA
(i) m ∈
/ T1∞ (I, M ) (i.e., m ∈
1 (M, A/I )) for all n ≫ 0).
∞
n
(ii) m ∈
/ AssI (N ) (equivalently, depth(N/I N ) > 1 for all n ≫ 0).
Furthermore, if this holds true, then grade(GI n (A)+ , GI n (N )) > 2 for every n >
ampI (N ).
Proof. Note that N = SyzA
1 (XM ) is a non-zero module. We have an exact sequence
0 → N → G → XM → 0, where G is a free A-module. The corresponding long
exact sequences of Tor-modules yield an exact sequence (for every n > 1):
(4.6.1)
n
n
n
n
0 −→ TorA
1 (XM , A/I ) −→ N/I N −→ G/I G −→ XM /I XM −→ 0.
n
(i) ⇒ (ii): Since m ∈
/ T1∞ (I, M ), by virtue of Lemma 4.5, TorA
1 (XM , A/I ) =
∞
∞
∞
0 for all n ≫ 0. Hence (4.6.1) yields that AssI (N ) ⊆ AssI (G) = AssI (A).
Therefore, since m ∈
/ Ass∞
I (A) (due to [McA83, Proposition 4.1]), we obtain that
∞
m∈
/ AssI (N ).
(ii) ⇒ (i): Since A has isolated
singularity, as in the proof of Lemma 4.5, it follows
n
(X
,
A/I
)
⊆
{m}
for every n > 1. Thus, since m ∈
/ Ass∞
that AssA TorA
M
1
I (N ), in
A
n
view of (4.6.1), it can be observed that AssA Tor1 (XM , A/I ) = φ for all n ≫ 0.
n
Equivalently, TorA
1 (XM , A/I ) = 0 for all n ≫ 0. Hence the implication follows
from Lemma 4.5.
ASYMPTOTIC ASSOCIATE PRIMES
15
Since height(I) > 2, we have grade(GI n (A)+ , GI n (A)) > 2 for all n ≫ 0; see
Theorem 3.4. So the last assertion follows from Theorem 4.1.
5. Asymptotic prime divisors over complete intersections
Let A be a local complete intersection, and M be a finitely generated A-module.
Suppose either I is a principal ideal or I has a principal reduction generated by an
A-regular element. In this section, we analyze the asymptotic stability of certain
n
associated prime ideals of Tor-modules TorA
i (M, A/I ) if both i and n tend to ∞.
5.1. Module structure on Tor. We first discuss the graded module structure on
direct sum of Tor-modules which we are going to use in order to prove our main
results on asymptotic prime divisors of Tor-modules.
Let Q be a ring, and f = f1 , . . . , fc be a Q-regular sequence. Set A := Q/(f ).
Let M and N be finitely generated A-modules.
5.1. Let F : · · · → Fn → · · · → F1 → F0 → 0 be a free resolution of M by finitely
generated free A-modules. Let
t′j : F −→ F(−2),
16j6c
be the Eisenbud operators defined by f = f1 , . . . , fc ; see [Eis80, Section 1]. In
view of [Eis80, Corollary 1.4], the chain maps t′j are determined uniquely up to
homotopy. In particular, they induce well-defined maps
A
tj : TorA
i (M, N ) −→ Tori−2 (M, N )
(for all i ∈ Z and j = 1, . . . , c) on the homology of F⊗A N . In [Eis80, Corollary 1.5],
it is shown that the chain maps t′j (1 6 j 6 c) commute up to homotopy. Thus
M
TorA
TorA
⋆ (M, N ) :=
−i (M, N )
i∈Z
turns into a Z-graded S := A[t1 , . . . , tc ]-module, where S is the N-graded polynomial ring over A in the operators tj defined by f with deg(tj ) = 2 for all j = 1, . . . , c.
A
Here note that for every i ∈ Z, the ith component of TorA
⋆ (M, N ) is Tor−i (M, N ).
This structure depends only on f , are natural in both module arguments and commute with the connecting maps induced by short exact sequences.
5.2. Stability of primes in AssA TorA
i (M, N ) . Here we study the asymptotic
stability of certain associated prime ideals of Tor-modules TorA
i (M, N ), (i > 0),
where M and N are finitely generated modules over a local complete intersection
ring A (see Corollary 5.10).
We denote the collection of all minimal prime ideals in the support of M by
MinA (M ) (or simply by Min(M )). It is well known that Min(M ) ⊆ AssA (M ) ⊆
Supp(M ). Recall that a local ring (A, m) is called a complete intersection ring if
b = Q/(f ), where Q is a complete regular local ring and
its m-adic completion A
f = f1 , . . . , fc is a Q-regular sequence. To prove our results, we may assume that
A is complete because of the following well-known fact on associate primes:
Lemma 5.2. For an A-module M , we have
o
n
b .
AssA (M ) = q ∩ A : q ∈ AssAb M ⊗A A
It is now enough to prove our result with the following hypothesis:
16
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
Hypothesis 5.3. Let A = Q/(f1 , . . . , fc ), where Q is a regular local ring, and
f1 , . . . , fc is a Q-regular sequence. Let M and N be finitely generated A-modules.
We show our result with the following more general hypothesis:
Hypothesis 5.4. Let A = Q/a, where Q is a regular ring of finite Krull dimension,
and a ⊆ Q is an ideal such that aq ⊆ Qq is generated by a Qq -regular sequence for
every q ∈ Var(a). Let M and N be finitely generated A-modules.
Note that a ring A satisfies Hypothesis 5.3 implies that A satisfies Hypothesis 5.4.
With the Hypothesis 5.4, we have the following well-known bounds for complete
intersection dimension and complexity:
CI-dimA (M ) = max{CI-dimAm (Mm ) : m ∈ Max(A)}
6 max{dim(Am ) : m ∈ Max(A)}
[by definition]
[see, e.g., [AB00, 4.1.5]]
= dim(A);
cxA (M ) = max{cxAm (Mm ) : m ∈ Max(A)}
6 max{codim(Am ) : m ∈ Max(A)}
[by definition]
[see, e.g., [AB00, 1.4]]
6 dim(Q).
Therefore, by [AB00, Theorem 4.9], we have the following result:
Theorem 5.5. With the Hypothesis 5.4, the following statements are equivalent:
(1) TorA
i (M, N ) = 0 for dim(Q) + 1 consecutive values of i > dim(A);
(2) TorA
i (M, N ) = 0 for all i ≫ 0;
A
(3) Tori (M, N ) = 0 for all i > dim(A).
Let us recall the following asymptotic behaviour of Tor-modules.
Lemma 5.6. [Gul74, Theorem 3.1]With the Hypothesis 5.3, if λA TorA
i (M, N )
is finite for all sufficiently large integer i (where λA (−) is the length function), then
M
TorA
−i (M, N ) is a *Artinian graded A[t1 , . . . , tc ]-module,
i≪0
where deg(tj ) = 2 for all j = 1, . . . , c.
As a consequence of this lemma, we obtain the following result:
Proposition 5.7. Let A be a local complete intersection
ring. Let M and N be
(M,
N
)
is
finite
for all sufficiently large
finitely generated A-modules. If λA TorA
i
integer i, then we have that
and λA TorA
λA TorA
2i+1 (M, N )
2i (M, N )
are given by polynomials in i over Q for all sufficiently large integer i.
Proof. Without loss of generality, we may assume that A is complete. Then the
proposition follows from Lemmas 5.6 and 5.8.
The following result is a consequence of the graded version of Matlis duality and
the Hilbert-Serre Theorem. It might be known for the experts. But we give a proof
here for the reader’s convenience.
ASYMPTOTIC ASSOCIATE PRIMES
17
L
Lemma 5.8. Let (A, m) be a complete local ring. Let L = i∈Z Li be a *Artinian
graded A[t1 , . . . , tc ]-module, where deg(tj ) = 2 for all 1 6 j 6 c, and λA (Li ) is
finite for all i ≪ 0. Then λA (L−2i ) and λA (L−2i−1 ) are given by polynomials in i
over Q for all sufficiently large i.
Proof. We use the graded Matlis duality. Let us recall the following definitions:
*complete from [BH98, p. 142]; *local from [BH98, p. 139]; and * Hom(−, −) from
[BH98, p. 33]. Note that A[t1 , . . . , tc ] is a Noetherian *complete *local ring. We
set E := EA (A/m), the injective hull of A/m. Also set L∨ := * Hom(L, E). Notice
that (L∨ )i = HomA (L−i , E) for all i ∈ Z.
Since A[t1 , . . . , tc ] is a Noetherian *complete *local ring, by virtue of Matlis duality for graded modules ([BH98, 3.6.17]), we obtain that L∨ is a finitely generated
graded A[t1 , . . . , tc ]-module. Let i0 be such that λA (L−i ) is finite for all i > i0 .
Hence
λA ((L∨ )i ) = λA (HomA (L−i , E)) = λA (L−i )
L
∨
is finite for all i > i0 ; see, e.g., [BH98, 3.2.12]. Since
i>i0 (L )i is a graded
L
∨
∨
A[t1 , . . . , tc ]-submodule of L , we have that
i>i0 (L )i is a finitely generated
graded module over A[t1 , . . . , tc ]. Therefore, by the Hilbert-Serre Theorem, we
obtain that λA ((L∨ )2i ) and λA ((L∨ )2i+1 ) are given by polynomials in i over Q for
all i ≫ 0, and hence the lemma follows from (5.8.1).
(5.8.1)
We are now in a position to prove our main result of this section.
Theorem 5.9. With the Hypothesis 5.4, exactly one of the following alternatives
must hold:
(1) TorA
i (M, N ) = 0 for all i > dim(A);
(2) There exists a non-empty finite subset A of Spec(A) such that for every p ∈ A,
at least one of the following statements holds true:
(i) p ∈ Min TorA
all i ≫ 0;
2i (M, N ) for
(ii) p ∈ Min TorA
(M,
N
)
for
all i ≫ 0.
2i+1
Proof. We set
[
Supp TorA
B :=
i (M, N ) : dim(A) < i 6 dim(A) + dim(Q) + 1 .
If B = φ (empty set), then TorA
i (M, N ) = 0 for all dim(A) < i 6 dim(A) +
dim(Q) + 1, and hence, by virtue of Theorem 5.5, we get that TorA
i (M, N ) = 0 for
all i > dim(A). So we may assume that B 6= φ. In this case, we prove that the
statement (2) holds true. We denote the collection of minimal primes in B by A,
i.e.,
(5.9.1)
A := {p ∈ B : q ∈ Spec(A) and q ( p ⇒ q ∈
/ B} .
Clearly, A is a non-empty finite subset of Spec(A). We claim that A satisfies
statement (2) in the theorem. To prove this claim, let us fix an arbitrary p ∈ A.
A
If q ∈ Spec(A) be such that q ( p, then q ∈
/ B, i.e., Tori q (Mq , Nq ) = 0 for all
dim(A) < i 6 dim(A) + dim(Q) + 1, and hence, in view of Theorem 5.5, we obtain
A
that Tori q (Mq , Nq ) = 0 for all i > dim(A). Therefore
A
(5.9.2)
Supp Tori p (Mp , Np ) ⊆ {pAp } for all i > dim(A),
18
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
which gives
(5.9.3)
A
λAp Tori p (Mp , Np )
is finite for all i > dim(A).
Since Ap satisfies Hypothesis 5.3, by virtue of Proposition 5.7, there are polynomials
P1 (z) and P2 (z) in z over Q such that
A
(5.9.4)
λAp Tor2ip (Mp , Np ) = P1 (i) for all i ≫ 0;
Ap
(5.9.5)
(Mp , Np ) = P2 (i) for all i ≫ 0.
λAp Tor2i+1
We now show that both P1 (z) and P2 (z) cannot be zero polynomials. If this is
not the case, then we have
A
Tori p (Mp , Np ) = 0
for all i ≫ 0,
which yields (by Theorem 5.5) that
A
Tori p (Mp , Np ) = 0 for all i > dim(A),
i.e., p ∈
/ B, and hence p ∈
/ A, which is a contradiction. Therefore at least one of P1
and P2 must be a non-zero polynomial.
Assume that P1 is a non-zero polynomial. Then P1 may have only finitely many
A
roots. Therefore P1 (i) 6= 0 for all i ≫ 0, which yields Tor2ip (Mp , Np ) 6= 0 for all
i ≫ 0. So, in view of (5.9.2), we obtain that
A
Supp Tori p (Mp , Np ) = {pAp } for all i ≫ 0,
which implies that p ∈ Min TorA
2i (M, N ) for all i ≫ 0.
Similarly, if P2 is a non-zero polynomial, then we have that
p ∈ Min TorA
for all i ≫ 0.
2i+1 (M, N )
This completes the proof of the theorem.
As a corollary of this theorem, we obtain the following result on associate primes.
Corollary 5.10. Let A be a local complete intersection ring. Let M and N be
finitely generated A-modules. Then exactly one of the following alternatives must
hold:
(1) TorA
i (M, N ) = 0 for all i > dim(A);
(2) There exists a non-empty finite subset A of Spec(A) such that for every p ∈ A,
at least one of the following statements holds true:
all i ≫ 0;
(i) p ∈ AssA TorA
2i (M, N ) for
(M,
N
)
for
all i ≫ 0.
(ii) p ∈ AssA TorA
2i+1
Proof. For every finitely generated A-module D, we have MinA (D) ⊆ AssA (D).
Therefore, if A is complete, then the corollary follows from Theorem 5.9. Now the
general case can be deduced by using Lemma 5.2.
Here we give an example which shows that both statements (i) and (ii) in the
assertion (2) of Corollary 5.10 might not hold together.
ASYMPTOTIC ASSOCIATE PRIMES
19
Example 5.11. Let Q = k[[u, x]] be a ring of formal power series in variables u
and x over a field k. We set A := Q/(ux) and M = N := Q/(u). Clearly, A is a
local complete intersection ring, and M , N are A-modules. Then, for every i > 1,
A
∼
we have that TorA
2i (M, N ) = 0 and Tor2i−1 (M, N ) = k; see [AB00, Example 4.3].
So, for all i > 1, we obtain that
AssA TorA
2i (M, N ) = φ and
AssA TorA
2i−1 (M, N ) = AssA (k) = {(u, x)/(ux)}.
n
5.3. Stability of primes in AssA TorA
i (M, A/I ) . We now study the asympn
totic stability of certain associated prime ideals of Tor-modules TorA
i (M, A/I ),
(i, n > 0), where M is a finitely generated module over a local complete intersection ring A, and either I is a principal ideal or I has a principal reduction generated
by an A-regular element (see Corollary 5.15). We start with the following lemma
which we use in order to prove our result when I is a principal ideal.
Lemma 5.12. Let A be a ring, and M be an A-module. Fix an element a ∈ A.
Then there exist an ideal J of A and a positive integer n0 such that
A
n
∼
TorA
i+2 M, A/(a ) = Tori (M, J) for all i > 1 and n > n0 .
Proof. For every positive integer n, we consider the following short exact sequence:
0 → (0 :A an ) → A → (an ) → 0,
which yields the following isomorphisms:
A
n
n
∼
(5.12.1)
TorA
i+1 M, (a ) = Tori M, (0 :A a )
for all i > 1.
For every n > 1, the short exact sequence 0 → (an ) → A → A/(an ) → 0 gives
A
n
n
∼
for all i > 1.
(5.12.2)
TorA
i+1 M, A/(a ) = Tori M, (a )
Thus (5.12.1) and (5.12.2) together yield
A
n
n
∼
(5.12.3)
TorA
i+2 M, A/(a ) = Tori M, (0 :A a )
for all i > 1.
Since A is a Noetherian ring, the ascending chain of ideals
0 : A a ⊆ 0 :A a 2 ⊆ 0 : A a 3 ⊆ · · ·
will stabilize somewhere, i.e., there exists a positive integer n0 such that
for all n > n0 .
(5.12.4)
0 :A an = 0 :A an0
Then the lemma follows from (5.12.3) and (5.12.4) by setting J := (0 :A an0 ).
Here is another lemma which we use in order to prove our result when I has a
principal reduction generated by an A-regular element.
Lemma 5.13. Let A be a ring. Let I be an ideal of A having a principal reduction
generated by an A-regular element. Then there exist an ideal J of A and a positive
integer n0 such that
n ∼
TorA
= TorA
i (M, A/J) for all i > 2 and n > n0 .
i M, A/I
20
DIPANKAR GHOSH, PROVANJAN MALLICK, AND TONY J. PUTHENPURAKAL
Proof. Since I has a principal reduction generated by an A-regular element, there
exist an A-regular element y and a positive integer n0 such that
I n+1 = yI n for all n > n0 .
Then it can be shown that for every n > n0 , we obtain a short exact sequence:
y·
0 −→ A/I n −→ A/I n+1 −→ A/(y) −→ 0.
Now note that TorA
i M, A/(y) = 0 for all i > 2. Therefore the short exact
sequence (5.13.1) yields
n+1
n ∼
for all i > 2 and n > n0 .
TorA
= TorA
i M, A/I
i M, A/I
(5.13.1)
Hence the lemma follows by setting J := I n0 .
Now we can achieve one of the main goals of this article.
Theorem 5.14. Let A be as in Hypothesis 5.4. Let M be a finitely generated Amodule, and I be an ideal of A. Suppose either I is principal or I has a principal
reduction generated by an A-regular element. Then there exist positive integers i0
and n0 such that exactly one of the following alternatives must hold:
n
(1) TorA
i (M, A/I ) = 0 for all i > i0 and n > n0 ;
(2) There exists a non-empty finite subset A of Spec(A) such that for every p ∈ A,
at least one of the following statements
holds true:
n
(i) p ∈ Min TorA
(M,
A/I
)
for
all
i
> i0 and n > n0 ;
2i
A
n
(ii) p ∈ Min Tor2i+1 (M, A/I ) for all i > i0 and n > n0 .
Proof. If I is a principal ideal, then the result follows from Theorem 5.9 and
Lemma 5.12. In another case, i.e., when I has a principal reduction generated
by an A-regular element, then we use Theorem 5.9 and Lemma 5.13 to get the
desired result of the theorem.
As an immediate corollary of this theorem, we obtain the following:
Corollary 5.15. Let A be a local complete intersection ring. Let M be a finitely
generated A-module, and I be an ideal of A. Suppose either I is principal or I has
a principal reduction generated by an A-regular element. Then there exist positive
integers i0 and n0 such that exactly one of the following alternatives must hold:
n
(1) TorA
i (M, A/I ) = 0 for all i > i0 and n > n0 ;
(2) There exists a non-empty finite subset A of Spec(A) such that for every p ∈ A,
at least one of the following statements
holds true:
n
(M,
A/I
)
for
all
i > i0 and n > n0 ;
(i) p ∈ AssA TorA
2i
A
n
(ii) p ∈ AssA Tor2i+1 (M, A/I ) for all i > i0 and n > n0 .
Proof. Since MinA (D) ⊆ AssA (D) for every finitely generated A-module D, the
corollary follows from Theorem 5.14 when A is complete.
b Also note that if I
Now note that if I is principal, then so is its completion I.
has a principal reduction generated by an A-regular element, then Ib has a principal
b
reduction generated by an A-regular
element. It is well known that
n
b
A
n
∼
c
b
b
b
M
,
A/(
I)
for all i, n > 0.
Tor
A
(M,
A/I
)
⊗
TorA
=
A
i
i
Therefore the general case can be easily deduced by using Lemma 5.2.
ASYMPTOTIC ASSOCIATE PRIMES
21
Acknowledgements
Ghosh was supported by DST, Govt. of India under the DST-INSPIRE Faculty
Scheme. Mallick was supported by UGC, MHRD, Govt. of India.
References
[AB89]
[AB00]
[Bro79]
[BFL04]
[BS98]
[BH98]
[Eis80]
[Eli04]
[GP16]
[Gul74]
[HU14]
[HM99]
[HH01]
[McA83]
[MS93]
[Nag05]
[Put03]
[Put07]
[RR78]
M. Auslander and R.-O. Buchweitz, The homological theory of maximal Cohen-Macaulay
approximations, Mém. Soc. Math. France (N.S.) 38 (1989), 5–37. 13
L. L. Avramov and R.-O. Buchweitz. Support varieties and cohomology over complete
intersections, Invent. Math. 142 (2000), 285–318. 16, 19
M. Brodmann, Asymptotic stability of Ass(M/I n M ), Proc. Amer. Math. Soc. 74 (1979),
16–18. 1, 12
M. Brodmann, S. Fumasoli and C. S. Lim, Low-codimensional associated primes of
graded components of local cohomology modules, J. Algebra 275 (2004), 867–882. 11
M. P. Brodmann and R. Y. Sharp, Local cohomology: an algebraic introduction with
geometric applications, Cambridge Studies in Advanced Mathematics 60, Cambridge
University Press, Cambridge, 1998. 6
W. Bruns and J. Herzog, Cohen-Macaulay Rings, Cambridge Studies in Advanced Mathematics 39, Revised Edition, Cambridge University Press, Cambridge, 1998. 17
D. Eisenbud, Homological algebra on a complete intersection, with an application to
group representations, Trans. Amer. Math. Soc. 260 (1980), 35–64. 15
J. Elias, Depth of higher associated graded rings, J. London Math. Soc. 70 (2004), 41–58.
4
D. Ghosh and T. J. Puthenpurakal, Asymptotic prime divisors over complete intersection
rings, Math. Proc. Cambridge Philos. Soc. 160 (2016), 423–436. Corrigendum, Math.
Proc. Cambridge Philos. Soc. 163 (2017), 381–384. 4
T. H. Gulliksen, A change of ring theorem with applications to Poincaré series and
intersection multiplicity, Math. Scand. 34 (1974), 167–183. 3, 16
J. Hong and B. Ulrich, Specialization and integral closure J. Lond. Math. Soc. 90 (2014),
861–878. 10
S. Huckaba and T. Marley, On associated graded rings of normal ideals, J. Algebra 222
(1999), 146–163. 4, 5, 9
R. Hübl and C. Huneke, Fiber cones and the integral closure of ideals, Collect. Math.
52 (2001), 85–10 13
S. McAdam, Asymptotic prime divisors, Lecture Notes in Mathematics 1023, SpringerVerlag, Berlin, 1983. 2, 14
L. Melkersson and P. Schenzel, Asymptotic prime ideals related to derived functors, Proc.
Amer. Math. Soc. 117 (1993), 935–938. 2, 12
R. Naghipour, Ratliff-Rush closures of ideals with respect to a Noetherian module, J.
Pure Appl. Algebra 195 (2005), 167–172. 6
T. J. Puthenpurakal, Hilbert coefficients of a Cohen-Macaulay module, J. Algebra 264
(2003), 82–97. 7, 13
T. J. Puthenpurakal, Ratliff-Rush filtration, regularity and depth of higher associated
graded modules. I, J. Pure Appl. Algebra 208 (2007), 159–176. 7
L. J. Ratliff, Jr. and D. E. Rush, Two notes on reductions of ideals Indiana Univ. Math.
J. 27 (1978), 929–934. 10
Chennai Mathematical Institute, H1, SIPCOT IT Park, Siruseri, Kelambakkam, Chennai 603103, Tamil Nadu, India
E-mail address: [email protected]
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai
400076, India
E-mail address: [email protected]
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai
400076, India
E-mail address: [email protected]
| 0math.AC
|
Computing the Rectilinear Center of Uncertain
Points in the Plane
arXiv:1509.05377v1 [cs.CG] 17 Sep 2015
Haitao Wang and Jingru Zhang
Department of Computer Science
Utah State University, Logan, UT 84322, USA
[email protected],[email protected]
Abstract. In this paper, we consider the rectilinear one-center problem
on uncertain points in the plane. In this problem, we are given a set P
of n (weighted) uncertain points in the plane and each uncertain point
has m possible locations each associated with a probability for the point
appearing at that location. The goal is to find a point q ∗ in the plane
which minimizes the maximum expected rectilinear distance from q ∗ to
all uncertain points of P, and q ∗ is called a rectilinear center. We present
an algorithm that solves the problem in O(mn) time. Since the input
size of the problem is Θ(mn), our algorithm is optimal.
1
Introduction
In the real world, data is inherently uncertain due to many facts, such as
the measurement inaccuracy, sampling discrepancy, resource limitation, and so
on. A large amount of work has recently been done on uncertain data, e.g.,
[1,2,3,9,12,13,17,18]. In this paper, we study the one-center problem on uncertain points in the plane with respect to the rectilinear distance.
Let P = {P1 , P2 , . . . , Pn } be a set of n uncertain points in the plane, where
each uncertain point Pi ∈ P has m possible locations pi1 , pi2 , · · · , pim and for
each 1 ≤ j ≤ m, pij is associated with a probability fij ≥ 0 for Pi being at pij
(which is independent of other locations).
For any (deterministic) point p in the plane, we use xp and yp to denote the xand y-coordinates of p, respectively. For any two points p and q, we use d(p, q) to
denote the rectilinear distance between p and q, i.e., d(p, q) = |xp −xq |+|yp −yq |.
Consider a point q in the plane. For any uncertain point Pi ∈ P, the expected
rectilinear distance between q and Pi is defined as
Ed(Pi , q) =
m
X
fij · d(pij , q).
j=1
Let Edmax (q) = maxPi ∈P Ed(Pi , q). A point q ∗ is called a rectilinear center
of P if it minimizes the value Edmax (q ∗ ) among all points in the plane. Our goal
is to compute q ∗ . Note that such a point q ∗ may not be unique, in which case
we let q ∗ denote an arbitrary such point.
We assume that for each uncertain point Pi of P, its m locations are given
in two sorted lists, one by x-coordinates and the other by y-coordinates. To the
best of our knowledge, this problem has not been studied before. In this paper,
we present an O(mn) time algorithm. Since the input size of the problem is
Θ(nm), our algorithm essentially runs in linear time, which is optimal.
Further, our algorithm is applicable to the weighted version of this problem
in which each Pi ∈ P has a weight wi ≥ 0 and the weighted expected distance,
i.e., wi ·Ed(Pi , q), is considered. To solve the weighted version, we can first reduce
it to the unweighted version by changing each fij to wi · fij for all 1 ≤ i ≤ n
and 1 ≤ j ≤ m, and then apply our algorithm for the unweighted version. The
running time is still O(mn).
1.1
Related Work
The problem of finding one-center among uncertain points on a line has been
considered in our previous work [21], where an O(nm) time algorithm was given.
An algorithm for computing k centers for general k was also given in [21] with
the running time O(mn log mn + n log n log k). In fact, in [21] we considered the
k-center problem under a more general uncertain model where each uncertain
point can appear in m intervals. We also studied the one-center problem for
uncertain points on tree networks in [20], where a linear-time algorithm was
proposed.
There is also a lot of other work on facility location problems for uncertain
data. For instances, Cormode and McGregor [7] proved that the k-center problem
on uncertain points each associated with multiple locations in high-dimension
space is NP-hard and gave approximation algorithms for different problem models. Foul [10] considered the Euclidean one-center problem on uncertain points
each of which has a uniform distribution in a given rectangle in the plane. de
Berg. et al. [8] studied the Euclidean 2-center problem for a set of moving points
in the plane (the moving points can be considered uncertain).
The k-center problems on deterministic points are classical problems and have
been studied extensively. When all points are in the plane, the problems on most
distance metrics are NP-hard [16]. However, some special cases can be solved in
polynomial time, e.g., the one-center problem [15], the two-center problem [6],
the rectilinear three-center problem [11], the line-constrained k-center problems
(where all centers are restricted to be on a given line in the plane) [5,14,19].
1.2
Our Techniques
Consider any uncertain point Pi ∈ P and any (deterministic) point q in the
plane R2 . We first show that Ed(Pi , q) is a convex piecewise linear function with
respect to q ∈ R2 . More specifically, if we extend a horizontal line and a vertical
line from each location of Pi , these lines partition the plane into a grid Gi of
(m + 1) × (m + 1) cells. Then, Ed(Pi , q) is a linear function (in both the x- and
y-coordinates of q) in each cell of Gi . In other words, Ed(Pi , q) defines a plane
surface patch in 3D on each cell of Gi . Then, finding q ∗ ∈ R2 is equivalent to
2
finding a lowest point p∗ in the upper envelope of the n graphs in 3D defined by
Ed(Pi , q) for all Pi ∈ P (specifically, q ∗ is the projection of p∗ onto the xy-plane).
The problem of finding p∗ , which may be interesting in its own right, can
be solved in O(nm2 ) time by the linear-time algorithm for the 3D linear programming (LP) problem [15]. Indeed, for a plane surface patch, we call the plane
containing it the supporting plane. Let H be the set of the supporting planes of
the surface patches of the functions Ed(Pi , q) for all Pi ∈ P. Since each function
Ed(Pi , q) is convex, p∗ is also a lowest point in the upper envelope of the planes
of H. Thus, finding p∗ is a LP problem in R3 and can be solved in O(|H|) time
[15]. Note that |H| = Θ(nm2 ) since each grid Gi has (m + 1)2 cells.
We give an O(mn) time algorithm without computing the functions Ed(Pi , q)
explicitly. We use a prune-and-search technique that can be considered as an
extension of Megiddo’s technique for the 3D LP problem [15]. In each recursive
step, we prune at least n/32 uncertain points from P in linear time. In this way,
q ∗ can be found after O(log n) recursive steps.
Unlike Megiddo’s algorithm [15], each recursive step of our algorithm itself
is a recursive algorithm of O(log m) recursive steps. Therefore, our algorithm
has O(log n) “outer” recursive steps and each outer recursive step has O(log m)
“inner” recursive steps. In each outer recursive step, we maintain a rectangle R
that always contains q ∗ in the xy-plane. Initially, R is the entire plane. Each
inner recursive step shrinks R with the help of a decision algorithm. The key
idea is that after O(log m) steps, R is so small that there is a set P ∗ of at least
n/2 uncertain points such that R is contained inside a single cell of the grid Gi
of each uncertain point Pi of P ∗ (i.e., R does not intersect the extension lines
from the locations of Pi ). At this point, with the help of our decision algorithm,
we can use a pruning procedure similar to Megiddo’s algorithm [15] to prune
at least |P ∗ |/16 ≥ n/32 uncertain points of P ∗ . Each outer recursive step is
carefully implemented so that it takes only linear time.
In particular, our decision algorithm is for the following decision problem. Let
R be a rectangle in the plane and R contains q ∗ (but the exact location of q ∗ is
unknown). Given an arbitrary line l that intersects R, the decision problem is to
determine which side of l contains q ∗ . Megiddo’s technique [15] gave an algorithm
that can solve our decision problem in O(m2 n) time. We give a decision algorithm
of O(mn) time. In fact, in order to achieve the overall O(mn) time for computing
q ∗ , our decision algorithm has the following performance. For each 1 ≤ i ≤ n,
let ai and bi be the number of columns and rows
Pnof the grid Gi intersecting R,
respectively. Our decision algorithm runs in O( i=1 (ai + bi )) time.
The rest of the paper is organized as follows. In Section 2, we introduce some
observations. In Section 3, we present our decision algorithm. Section 4 gives the
overall algorithm for computing the rectilinear center q ∗ . Section 5 concludes.
2
Observations
Let p be a point in the plane R2 . The vertical line and the horizontal line through
p partition the plane into four (unbounded) rectangles. Consider another point
3
q ∈ R2 . We consider d(p, q) as a function of q ∈ R2 . For each of the above
rectangle R, d(p, q) on q ∈ R is a linear function in both the x- and y-coordinates
of q, and thus d(p, q) on q ∈ R defines a plane surface patch in R3 . Further, d(p, q)
on q ∈ R2 is a convex piecewise linear function.
For ease of exposition, we make a general position assumption that no two
locations of the uncertain points of P have the same x- or y-coordinate.
Consider an uncertain point Pi of P. We extend a horizontal line and a
vertical line through each location of Pi to obtain a grid, denoted by Gi , which
has (m + 1) × (m + 1) cells (and each cell is a rectangle). According to the
above discussion, for each location pij of P, the function d(pij , q) of q in each
cell of Gi is linear and defines a plane surface patch in R3 . Therefore, if we
consider Ed(Pi , q) as a function of q, since Ed(Pi , q) is the sum of fij · d(pij , q)
for all 1 ≤ j ≤ m, Ed(Pi , q) of q in each cell of Gi is also linear and defines a
plane surface patch in R3 . Further, since each d(pij , q) for q ∈ R2 is convex, the
function Ed(Pi , q), as the sum of convex functions, is also convex.
In the following, since Ed(Pi , q) is normally considered as function of q, for
convenience, we will use Edi (x, y) to denote it for q = (x, y) ∈ R2 .
The above discussion leads to the following observation.
Observation 1 For each uncertain point Pi ∈ P, the function Edi (x, y) is convex piecewise linear. More specifically, Edi (x, y) on each cell of the grid Gi is
linear and defines a plane surface patch in R3 (e.g., see Fig. 1).
Consider the function Edi (x, y) of any Pi ∈ P. Clearly, the complexity of
Edi (x, y) is Θ(m2 ). However, since Edi (x, y) on each cell C of Gi is a plane
surface patch in R3 , Edi (x, y) on C is of constant complexity. We use Edi (x, y, C)
to denote the linear function of Edi (x, y) on C. Note that Edi (x, y, C) is also the
function of the supporting plane of the surface patch of Edi (x, y) on C.
As discussed in Section 1.2, our algorithm will not compute the function
Edi (x, y) explicitly. Instead, we will compute it implicitly. More specifically,
we will do some preprocessing such that given any cell C of Gi , the function
Edi (x, y, C) can be determined efficiently. We first introduce some notation.
Let Xi = {xi1 , xi2 , · · · , xim } be the set of the x-coordinates of all locations
of Pi sorted in ascending order. Let Yi = {yi1 , yi2 , · · · , yim } be the set of their
y-coordinates in ascending order. Note that Xi and Yi can be obtained in O(m)
time from the input (recall that the locations of Pi are given in two sorted lists
in the input). For convenience of discussion, we let xi0 = −∞, and let Xi also
include xi0 . Similarly, let yi0 = −∞, and let Yi also include yi0 . Note that due
to our general position assumption, the values in Xi (resp., Yi ) are distinct.
For any value z, we refer to the largest value in Xi that is smaller or equal
to z the predecessor of z in Xi , and we use Iz (Xi ) to denote the index of the
predecessor. Similarly, Iz (Yi ) is the index of the predecessor of z in Yi .
Consider any point q in the plane. The predecessor of the x-coordinate of q
in Xi is also called the predecessor of q in Xi . Similarly, the predecessor of the
y-coordinate of q in Yi is also called the predecessor of q in Yi . We use Iq (Xi )
and Iq (Yi ) to denote their indices, respectively.
4
Figure 1. Illustrating the function Edi (x, y) of an uncertain point Pi with m = 4.
Consider any cell C of the grid Gi . For convenience of discussion, we assume
C contains its left and bottom sides, but does not contain its top and right sides.
In this way, any point in the plane is contained in one and only one cell of Gi .
Further, all points of C have the same predecessor in Xi and also have the same
predecessor in Yi . This allows us to define the predecessor of C in Xi as the
predecessor of any point in Xi , and we use IC (Xi ) to denote the index of the
predecessor. We define IC (Yi ) similarly. We have the following lemma.
Lemma 1. For any uncertain point Pi ∈ P, after O(m) time preprocessing, for
any cell C of the grid Gi , if IC (Xi ) and IC (Yi ) are known, then the function
Edi (x, y, C) can be computed in constant time.
Proof. For each location p ∈ Pi , let xp and yq be the x- and y-coordinates of p,
respectively, and let fp be the probability associated with p.
For any P
point q = (x, y) inP
R2 , recall that the expected distance function
Edi (x, y) = p∈PP
fp · d(p, q) = p∈PP
fp · (|xp − x| + |yp − y|). Therefore, we can
i
i
following, we first
write Edi (x, y) = p∈Pi fP
p ·|xp − x|+
p∈Pi fp ·|yp − y|. In theP
discuss how to compute p∈Pi fp · |xp − x| and the case for p∈Pi fp · |yp − y|
is very similar.
Let S1 denote the set of all locations of Pi whose x-coordinates are smaller
than or equal to x, i.e., the x-coordinate of q. Let S2 = Pi \ S1 . Then, we have
the following:
X
p∈Pi
fp · |xp − x| =
X
fp · (x − xp ) +
p∈S1
=x·
X
fp · (xp − x)
p∈S2
X
fp −
p∈S1
=x· 2·
X
p∈S1
X
X
X
fp −
fp · xp +
fp · xp
p∈S2
fp −
X
p∈Pi
5
p∈S1
p∈S2
X
X
fp − 2
fp · xp +
fp · xp .
p∈S1
p∈Pi
(1)
P
Thus,
to compute
Pit is sufficient to know the four
P p∈Pi fp · |xp − x|,
P in orderP
values p∈S1 fp , p∈Pi fp , p∈S1 fp · xp , and p∈Pi fp · xp . To this end, we do
the following preprocessing.
P
P
First, we compute p∈Pi fp and p∈Pi fp · xp , which can be done in O(m)
time. Second, recall that Xi = {xi0 , xi1 , . . . , xim } maintains the x-coordinates
of the locations of Pi sorted in ascending order. Note that given any index j
with 1 ≤ j ≤ m, we can access the information of the location of Pi whose
x-coordinate is xij in constant time, and this can be done by linking each xij
to the corresponding location of Pi when we create the list Xi from the input.
For each j with 1 ≤ j ≤ n, we let f (j) be the probability associated with the
location of Pi whose x-coordinate is xij .
In the preprocessing, we compute two arrays A[0 · · · m] and B[0 · · · m]. SpecifP
P
ically, for each 1 ≤ j ≤ n, A[j] = jk=1 f (k) and B[j] = jk=1 f (k) · xik . For
j = 0, we let A[0] = B[0] = 0. As discussed above, since we can access f (j) in
constant time for any 1 ≤ j ≤ m, the two arrays A and B can be computed in
O(m) time.
Let t = Iq (Xi ), i.e.,
P the index of the predecessor of q in Xi .PNote that
t ∈ [0, m]. To compute P
p∈Pi fp ·|xp −x|, an easy observation is that
p∈S1 fp is
with the
exactly equal to A[t] and p∈S1 fp ·xp is exactly equal to B[t]. Therefore,
P
above preprocessing, if t is known, according to Equation (1), p∈Pi fp · |xp − x|
can be computed in O(1) time.
The above shows that
P with O(m) time preprocessing, given Iq (Xi ), we can
compute the function p∈Pi fp · |xp − x| of x at q = (x, y) in constant time.
In a similar
Pway, with O(m) time preprocessing, given Iq (Yi ), we can compute
the function p∈Pi fp · |yp − y| of y at p = (x, y) in constant time.
Let q be any point in the cell C. Hence, Iq (Xi ) = IC (Xi ) and Iq (Yi ) =
IC (Yi ). Further, the function Edi (x, y) on q = (x, y) ∈ C is exactly the function
Edi (x, y, C). Therefore, with O(m) time preprocessing, given IC (Xi ) and IC (Yi ),
we can compute the function Edi (x, y, C) in constant time.
The lemma thus follows.
⊓
⊔
Due to Lemma 1, we have the following corollary.
Corollary 1. For each uncertain point Pi ∈ P, after O(m) time preprocessing,
given any point q in the plane, the expected distance Ed(Pi , q) can be computed
in O(log m) time.
Proof. Given any point q ∈ R2 , we can compute Iq (Xi ) in O(log m) time by
doing binary search on Xi . Similarly, we can compute Iq (Yi ) in O(log m) time.
Let C be the cell containing q. Recall that IC (Xi ) = Iq (Xi ) and IC (Yi ) = Iq (Yi ).
Hence, by Lemma 1, we can compute the function Edi (x, y, C) in constant time.
Then, Ed(Pi , q) is equal to Edi (qx , qy , C), where qx and qy are the x- and ycoordinates of q, respectively. Thus, after Edi (x, y, C) is known, Ed(Pi , q) can be
computed in constant time. The corollary thus follows.
⊓
⊔
Recall that Edmax (q) = maxPi ∈P Ed(Pi , q) for any point q in the plane. For
convenience, we use Edmax (x, y) to represent Edmax (q) as a function of q =
6
(x, y) ∈ R2 . Note that Edmax (x, y) is the upper envelope of the functions Edi (x, y)
for all i = 1, 2, . . . , n. Since each Edi (x, y) is convex on R2 , Edmax (x, y) is also
convex on R2 . Further, the rectilinear center q ∗ corresponds to a lowest point p∗
on Edmax (x, y). Specifically, q ∗ is the projection of p∗ on the xy-plane. Therefore,
computing q ∗ is equivalent to computing a lowest point in the upper envelope
of all functions Edi (x, y) for all i = 1, 2, . . . , n.
For each 1 ≤ i ≤ n, let Hi denote the set of supporting planes of all surface
patches of the function Edi (x, y). Let H = ∪ni=1 Hi . Since Edi (x, y) is convex,
Edi (x, y) is essentially the upper envelope of the planes in Hi . Hence, Edmax (x, y)
is also the upper envelope of all planes in H. Therefore, as discussed in Section
1.2, finding p∗ is essentially a 3D LP problem on H, which can be solved in
O(|H|) time by Megiddo’s technique [15]. Since the size of each Hi is Θ(m2 ),
|H| = Θ(nm2 ). Therefore, applying the algorithm in [15] directly can solve the
problem in O(nm2 ) time. In the following, we give an O(nm) time algorithm.
In the following paper, we assume we have done the preprocessing of Lemma
1 for each Pi ∈ P, which takes O(mn) time in total.
3
The Decision Algorithm
In this section, we present a decision algorithm that solves a decision problem,
which is needed later in Section 4. We first introduce the decision problem.
Let R = [x1 , x2 ; y1 , y2 ] be an axis-parallel rectangle in the plane, where x1
and x2 are the x-coordinates of the left and right sides of R, respectively, and
y1 and y2 are the y-coordinates of the bottom and top sides of R, respectively.
Suppose it is known that q ∗ is in R (but the exact location of q ∗ is not known).
Let L be an arbitrary line that intersects the interior of R. The decision problem
asks whether q ∗ is on L, and if not, which side of L contains q ∗ . We assume the
two predecessor indices Ix1 (Xi ) and Iy1 (Yi ) are already known.
For each 1 ≤ i ≤ n, let ai = Ix2 (Xi )−Ix1 (Xi )+1 and bi = Iy2 (Yi )−Iy1 (Yi )+1.
In fact, ai and bi are the numbers of columns and rows of Gi that intersect R,
respectively.
Below, we give a decision algorithm
Pn that solves the decision problem
Pn
in O( i=1 (ai + bi )) time. Note that 2n ≤ i=1 (ai + bi ) ≤ 2(mP+ 1)n.
n
We first show that the decision problem can be solved in O( i=1 ai · bi ) time
by using the decision algorithm
for the 3D LP problem [15]. Later we will reduce
Pn
the running time to O( i=1 (ai + bi )) time.
Recall that p∗ is a lowest point in the upper envelope of the functions Edi (x, y)
for i = 1, 2, . . . , n. Since q ∗ is in R and each function Edi (x, y) is convex, an easy
observation is that p∗ is also a lowest point in the upper envelope of Edi (x, y)
for i = 1, 2, . . . , n restricted on (x, y) ∈ R. This implies that we only need to
consider each function Edi (x, y) restricted on R.
For each 1 ≤ i ≤ n, let Gi (R) be the set of cells of Gi that intersect R,
and let Hi (R) be the set of supporting planes of the surface patches of Ed(Pi , q)
defined on the cells of Gi (R). Let H(R) = ∪ni=1 Hi (R). By our above analysis,
p∗ is a lowest point of the upper envelope of all planes
Pin H(R). Note that
|Hi (R)| = ai · bi for each 1 ≤ i ≤ n. Thus, |H(R)| = ni=1 ai · bi . Then, we
7
can apply the decision algorithm in [15] (Section 5.2) on H(R) to determine
which side of L contains q ∗ in O(|H(R)|) time. In order to explain our improved
algorithm later, we sketch this algorithm below.
We consider each plane of H(R) as a function of the points q on the xy-plane
R2 . In the first step, the algorithm finds a point q ′ on L that minimizes the
maximum value of all functions in H(R) restricted on the line q ∈ L. This is
essentially a 2D LP problem because each function of H(R) restricted on L is
a line, and thus the problem can be solved in O(|H(R)|) time [15]. Let Φq′ be
the set of functions of H(R) whose values at q ′ are equal to the above maximum
q ′ is computed. This
value. The set Φq′ can be found in O(|H(R)|) time
Pafter
n
finishes the first step, which takes O(|H(R)|) = O( i=1 ai · bi ) time.
The second step solves another two instances of the 2D LP problem on
thePplanes of Φq′ , which takes O(|Φq′ |) time. An easy upper bound for |Φq′ |
n
is i=1 ai · bi . A close analysis can show that |Φq′ | = O(n). Indeed, for each
1 ≤ i ≤ n, since the function Edi (x, y) is convex, among all ai · bi planes in
Hi (R), at most four of them are in Φq′ . Therefore, |Φq′ | = O(n). Hence, the second step runs in O(n) time. Since in our problem there always exists a solution,
according to [15], the second step will either conclude that q ′ is q ∗ or tell which
side
L contains q ∗ , which solves the decision problem. The algorithm takes
Pof
n
O( i=1 ai · bi ) time in total, which is dominated by the first step.
In the sequel, we reducePthe running time of the above algorithm, in parn
ticular, the first step, to O( i=1 (ai + bi )). Our goal is to compute q ′ and Φq′ .
′
By the definition, q is a lowest point in the upper envelope of all functions
of H(R) restricted on the line L. Consider any uncertain point Pi ∈ P. Let
Hi (R, L) be the set of supporting planes of the surface patches defined on the
cells of Gi (R) intersecting L. Observe that since Edi (x, y) is convex, the upper
envelope of all the functions of Hi (R) restricted on L is exactly the upper envelope of the functions of Hi (R, L) restricted on L. Therefore, q ′ is also a lowest
point in the upper envelope of the functions of H(R, L) restricted on L, where
H(R, L) = ∪ni=1 Hi (R, L). In other words, among all planes in H(R), only the
planes of H(R, L) are relevant for determining q ′ . Thus, suppose H(R, L) has
been computed; then q ′ can be computed based on the planes of H(R, L) in
O(|H(R, L)|) time by the 2D LP algorithm [15]. After q ′ is computed, the set
Φq′ can also be determined inPO(|H(R, L)|) time.
n
Note that |H(R, L)| = O( i=1 (ai + bi )), since for each 1 ≤ i ≤ n, |Hi (R, L)|,
which is equal to the number of cells of Gi (R) intersecting L, is O(ai + bi ).
It remains to compute H(R, L), i.e., compute Hi (R, L) for each 1 ≤ i ≤ n.
Recall that R = [x1 , x2 ; y1 , y2 ] and the two predecessor indices Ix1 (Xi ) and
Iy1 (Yi ) for each 1 ≤ i ≤ n are already known. The following lemma gives an
O(ai + bi ) algorithm to compute Hi (R, L).
Lemma 2. For each 1 ≤ i ≤ n, Hi (R, L) can be computed in O(ai + bi ) time.
Proof. Let Gi (R, L) be the set of cells of Gi (R) intersecting L. To compute
the planes in Hi (R, L), it is sufficient to determine the plane surface patches
of Edi (x, y) defined on the cells of Gi (R, L). By Lemma 1, this amounts to
8
determine the indices of the predecessors of these cells in Xi and Yi , respectively.
In the following, we give an algorithm to compute the cells of Gi (R, L) and
determine their predecessor indices in Xi and Yi , respectively, and the algorithm
runs in O(ai + bi ) time.
The main idea is that we first pick a particular point p on L ∩ R and locate
the cell of Gi (R) containing p (clearly this cell belongs to Gi (R, L)), and then
starting from p, we traverse on L and Gi (R) simultaneously to trace other cells
of Gi (R, L) until we move out of R. The details are given below.
We focus on the case where L has a positive slope. The other cases can be
handled similarly. Recall that L intersects the interior of R. Let p be the leftmost
intersection of L with the boundary of R. Hence, p is either on the left side or
the bottom side of R.
Let C be the cell of Gi that contains p. We first determine the two indices
Ip (Xi ) and Ip (Yi ). Note that IC (Xi ) = Ip (Xi ) and IC (Yi ) = Ip (Yi ).
Since p ∈ R, the index Ip (Xi ) can be found in O(ai ) time by scanning the
list Xi from the index Ix1 (Xi ). Similarly, Ip (Yi ) can be found in O(bi ) time by
scanning the list Yi from the index Iy1 (Yi ). After IC (Xi ) = Ip (Xi ) and IC (Yi ) =
Ip (Yi ) are computed, by Lemma 1, the function Edi (x, y, C) can be computed in
constant time, and we add the function to Hi (R, L).
Next, we move p on L rightwards. We will show that when p crosses the
boundary of C, we can determine the new cell containing p and update the two
indices Ip (Xi ) and Ip (Yi ) in constant time. This process continues until p moves
out of R. Specifically, when p moves on L rightwards, p will cross the boundary
of C either from the top side or the right side.
First, we determine whether p will move out of R before p crosses the boundary of C. If yes, then we terminate the algorithm. Otherwise, we determine
whether p moves out of C from its right side or left side. All above can be easily
done in constant time. Depending on whether p crosses the boundary of C from
its top side, right side, or from both sides simultaneously, there are three cases.
1. If p crosses the boundary of C from the top side and p does not cross the right
side of C, then p enters into a new cell that is on top of C. We update C to the
new cell. We increase the index Ip (Yi ) by one, but keep Ip (Xi ) unchanged.
Clearly, the above two indices are correctly updated and IC (Xi ) = Ip (Xi )
and IC (Yi ) = Ip (Yi ) for the new cell C. Again, by Lemma 1, the function
Edi (x, y, C) for the new cell C can be computed in constant time. We add
the new function to Hi (R, L).
2. If p crosses the boundary of C from the right side and p does not cross the top
side of C, then p enters into a new cell that is on right of C. The algorithm
in this case is similar to the above case and we omit the discussions.
3. The remaining case is when p crosses the boundary of C through the top
right corner of C. In this case, p enters into the northeast neighboring cell of
C. We first add to Hi (R, L) the supporting planes of the surface patches of
Edi (x, y) defined on the top neighboring cell and the right neighboring cell
of C, which can be computed in constant time as the above two cases. Then,
we update C to the new cell p is entering. We increase each of Ip (Xi ) and
9
Ip (Yi ) by one. Again, the two indices are correctly updated for the new cell
C. Finally, we compute the new function Edi (x, y, C) and add it to Hi (R, L).
When the algorithm stops, Hi (R, L) is computed. In general, during the
procedure of moving p on L, we spend constant time on finding each supporting
plane of Hi (R, L). Therefore, the total running time of the entire algorithm is
O(ai + bi ). The lemma thus follows.
⊓
⊔
With the preceding lemma, we have the following result.
Pn
Theorem 1. The decision problem can be solved in O( i=1 (ai + bi )) time.
4
Computing the Rectilinear Center
In this section, with the help of our decision algorithm in Section 3, we compute
the rectilinear center q ∗ in O(mn) time.
As discussed in Section 1.2, our algorithm is a prune-and-search algorithm
that has O(log n) “outer” recursive steps each of which has O(log m) “inner” recursive steps. In each outer recursive step, the algorithm prunes at least |P|/32
uncertain points of P such that these uncertain points are not relevant for computing q ∗ . After O(log n) outer recursive steps, there will be only a constant
number of uncertain points remaining in P. Each outer recursive step runs in
O(m|P|) time, where |P| is the number of uncertain points remaining in P. In
this way, the total running time of the algorithm is O(mn).
Each outer recursive step is another recursive prune-and-search algorithm,
which consists of 2 + log m inner recursive steps. Let X = ∪ni=1 Xi and Y =
∪ni=1 Yi . Hence, |X | = |Y| = mn. We maintain a rectangle R = [x1 , x2 ; y1 , y2 ]
that contains q ∗ . Initially, R is the entire plane. In each inner recursive step, we
shrink R such that the x-range [x1 , x2 ] (resp., y-range [y1 , y2 ]) of the new R only
contains half of the values of X (resp., Y) in the x-range (resp., y-range) of the
previous R. In this way, after log m + 2 inner recursive steps, the x-range (resp.,
y-range) of R only contains at most n/4 values of X (resp., Y). At this moment,
a key observation is that there is a subset P ∗ of at least n/2 uncertain points,
such that for each Pi ∈ P ∗ , R is contained in the interior of a cell of the grid Gi ,
i.e., the x-range (resp., y-range) of R does not contain any value of Xi (resp.,
Yi ). Due to the observation, we can use a pruning procedure similar to that in
[15] to prune at least |P ∗ |/16 ≥ n/32 uncertain points.
In the following, in Section 4.1, we give our algorithm on pruning the values
of X and Y to obtain P ∗ . In Section 4.2, we prune uncertain points of P ∗ .
4.1
Pruning the Coordinate Values of X and Y
Consider a general step of the algorithm where we are about to perform the j-th
inner recursive step for 1 ≤ j ≤ log m+ 2. Our algorithm maintains the following
j−1 j−1 j−1
algorithm invariants. (1) We have a rectangle Rj−1 = [xj−1
; y1 , y2 ]
1 , x2
∗
that contains q . (2) For each 1 ≤ i ≤ n, the index Ixj−1 (Xi ) of the predecessor
1
10
of xj−1
in Xi is known, and so is the index Iyj−1 (Yi ). (3) We have a sublist Xij−1
1
1
of Xi that consists of all values of Xi in [xj−1
, x2j−1 ] and a sublist Yij−1 of Yi
1
j−1 j−1
that consists of all values of Yi in [y1 , y2 ]. Note that these sublists can be
empty. (4) |X j−1 | ≤ mn/2j−1 and |Y j−1 | ≤ mn/2j−1 , where X j−1 = ∪ni=1 Xij−1
and Y j−1 = ∪ni=1 Yij−1 .
Initially, we set R0 = [−∞, +∞; −∞, +∞], Xi0 = Xi and Yi0 = Yi for each
1 ≤ i ≤ n, with X 0 = X and Y 0 = Y. It is easy to see that before we start the
first inner recursive step for j = 1, all the algorithm invariants hold.
In the sequel, we give the details of the j-th inner recursive step. We will
show that its running time is O(mn/2j + n) and all algorithm invariants are still
maintained after the step.
Let xm be the median of X j−1 and ym be the median of Y j−1 . Both xm and
ym can be found in O(|X j−1 | + |Y j−1 |) time.
For each 1 ≤ i ≤ n, let aj−1
= I(xj−1
) − I(xj−1
) + 1 and bj−1
= I(y2j−1 ) −
2
1
i
i
j−1
j−1
j−1
j−1
j−1
I(y1 ) + 1. Observe that ai = |Xi | + 1 and bi = |Yi | + 1.
Let x∗ and y ∗ be the x- and y-coordinates of q ∗ , respectively.
We first determine whether x∗ > xm , x∗ < xm , or x∗ = xm . This can
be done by applying our decision algorithm on Rj−1 and L with L being the
vertical
. By Theorem 1, the running time of our decision algorithm
Pnline x = xmj−1
is O( i=1 (aj−1
+ bi )), which is O(n + |X j−1 | + |Y j−1 |).
i
Note that if x∗ = xm , then according to our decision algorithm, q ∗ will be
found by the decision algorithm and we can terminate the entire algorithm. Otherwise, without loss of generality, we assume x∗ > xm . We proceed to determine
whether y ∗ > ym or y ∗ < ym , or y ∗ = ym by applying our decision algorithm on
Rj−1 and L with L being the horizontal line y = ym . Similarly, if y ∗ = ym , then
the decision algorithm will find q ∗ and we are done. Otherwise, without loss of
generality we assume y ∗ > ym . The above calls our decision algorithm twice,
which takes O(n + |X j−1 | + |Y j−1 |) time in total.
j−1
Now we know that q ∗ is in the rectangle [xm , xj−1
]. We let Rj =
2 ; ym , y2
j
j
j
j
j
j
j−1
[x1 , x2 ; y1 , y2 ] be the above rectangle, i.e., x1 = xm , x2 = x2 , y1j = ym , and
y2j = y2j−1 . Clearly, the first algorithm invariant is maintained.
We further proceed as follows to maintain the other three invariants.
For each 1 ≤ i ≤ n, by scanning the sorted list Xij−1 , we compute the
index Ixj (Xi ) of the predecessor of xj1 in Xi (each element of Xij−1 maintains its
1
original index in Xi ), and similarly, by scanning the sorted list Yij−1 , we compute
the index Iyj (Yi ). Computing these indices in all Xi and Yi for i = 1, 2, . . . , n
1
can be done in O(|X j−1 | + |Y j−1 |) time. This maintains the second algorithm
invariant.
Next, for each 1 ≤ i ≤ n, we scan Xij−1 to compute a sublist Xij , which
consists of all values of Xij−1 in [xj1 , xj2 ], and similarly, we scan Yij−1 to compute a
sublist Yij , which consists of all values of Yij−1 in [y1j , y2j ]. Computing the lists Xij
and Yij for all i = 1, 2, . . . , n as above can be done in overall O(|X j−1 | + |Y j−1 |)
time. This maintains the third algorithm invariant.
11
Pn
Pn
j
j
j
Let X j =
=
i=1 Xi and Y
i=1 Yi . According to our above algoj
j−1
j
j−1
rithm, |X | ≤ |X
|/2 and |Y | ≤ |Y
|/2. Since |X j−1 | ≤ nm/2j−1 and
j−1
j−1
j
|Y
| ≤ nm/2 , we obtain |X | ≤ nm/2j and |Y j | ≤ nm/2j . Hence, the
fourth algorithm invariant is maintained.
In summary, after the j-th inner recursive step, all four algorithm invariants
are maintained. Our above analysis also shows that the total running time is
O(n + |X j−1 | + |Y j−1 |), which is O(nm/2j + n).
We stop the algorithm after the t-thPinner recursive step, for t = 2 + log m.
t
The total time for all t steps is thus O( j=1 (n + mn/2j )) = O(mn).
After the t-th step, by our algorithm invariants, the rectangle Rt contains
∗
q , and |X t | ≤ mn/2t = n/4 and |Y t | ≤ mn/2t = n/4.
We say that an uncertain point Pi is prunable if both Xit and Yit are empty
(and thus Rt is contained in the interior of a cell of Gi ). Let P ∗ denote the
set of all prunable uncertain points of P. The following is an easy but crucial
observation.
Observation 2 |P ∗ | ≥ n/2.
Proof. Since X t ≤ n/4, among the n sets Xit for i = 1, 2, . . . , n, at most n/4
of them are non-empty. Similarly, since Y t ≤ n/4, among the n sets Yit for
i = 1, 2, . . . , n, at most n/4 of them are non-empty. Therefore, there are at most
n/2 uncertain points Pi ∈ P such that either Xit or Yit is non-empty. This implies
that there are at least n/2 prunable uncertain points in P.
⊓
⊔
After the t-th inner recursive step, the set P ∗ can be obtained in O(n) time
by checking all sets Xit and Yit for i = 1, 2, . . . , n and see whether they are empty.
The reason we are interested in prunable uncertain points is that for each
prunable uncertain point Pi of P ∗ , since Rt contains q ∗ and Rt is contained in
a cell Ci of Gi , there is only one surface patch of Edi (x, y) (i.e., the one defined
on Ci ) that is relevant for computing q ∗ . Let hi denote the supporting plane of
the above surface patch. We call hi the relevant plane of Pi . Note that we can
obtain hi in constant time. Indeed, observe that the predecessor index ICi (Xi )
is exactly Ixt1 (Xi ), which is known by our algorithm invariants. Similarly, the
index ICi (Yi ) is also known. By Lemma 1, the function Edi (x, y, Ci ), which is
also the function of hi , can be obtained in constant time. Hence, the relevant
planes of all prunable uncertain points of P ∗ can be obtained in O(n) time.
Remark. One may wonder why we did not perform the inner recursive steps
for t = log mn times (instead of t = 2 + log m time) so that X t and Y t would
each have a constant number of values in the range of R. The reason is that
based on our analysis, that would take O(mn + n log nm) time, which may not
be bounded by O(mn) (e.g., when m = o(log n)). In fact, performing the inner
recursive steps for t = 2 + log m times such that X t and Y t each have at most n4
values in the range of R is an interesting and crucial ingredient of our techniques.
12
4.2
Pruning Uncertain Points from P ∗
Consider a prunable uncertain point Pi of P ∗ . Recall that Hi is the set of supporting planes of all surface patches of Edi (x, y). The above analysis shows that
among all planes in Hi , only the relevant plane hi is useful for determining q ∗ .
In other words, the point p∗ , as a lowest point of all planes in H = ∪ni=1 Hi , is
also a lowest point of the planes in the union of ∪Pi ∈P ∗ hi and ∪Pi ∈P\P ∗ Hi . This
will allow us to prune at least |P ∗ |/16 uncertain points from P ∗ . The idea is
similar to Megiddo’s pruning scheme for the 3D LP algorithm in [15].
For each Pi ∈ P ∗ , its relevant plane hi is also considered as a function in
the xy-plane. Arrange all uncertain points of P ∗ into |P ∗ |/2 disjoint pairs. Let
D(P ∗ ) denote the set of all these pairs. For each pair (Pi , Pj ) ∈ D(P ∗ ), if the
value of the function hi at any point of Rt is greater than or equal to that of hj ,
then Pj can be pruned immediately; otherwise, we project the intersection of hi
and hj on the xy-plane to obtain a line Lij dividing Rt into two parts, such that
hi ≥ hj on one part and hi ≤ hj on the other.
Let L denote the set of the dividing lines Lij for all pairs of D(P ∗ ). Let Lm
be the line whose slope has the median value among the lines of L. We transform
the coordinate system by rotating the x-axis to be parallel to Lm (the y-axis does
not change). For ease of discussion, we assume no other lines of L are parallel to
Lm (the assumption can be easily lifted; see [15]). In the new coordinate system,
half the lines of L have negative slopes and the other half have positive slopes.
We now arrange all lines of L into disjoint pairs such that each pair has a line
of a negative slope and a line of positive slope. Let D(L) denote the set of all
these line pairs.
For each pair (Li , Lj ) ∈ D(L), we define yij as the y-coordinate of the intersection of Li and Lj . We find the median ym of the values yij for all pairs in
D(L). Let x∗ and y ∗ respectively be the x- and y-coordinate of q ∗ in the new
coordinate system. We determine in O(mn) time whether y ∗ > ym , y ∗ < ym or
y ∗ = ym by using our decision algorithm (here an O(mn) time decision algorithm
is sufficient for our purpose). If y ∗ = ym , then our decision algorithm will find
q ∗ and we can terminate the algorithm. Otherwise, without of loss generality,
we assume y ∗ < ym .
Let D′ (L) denote the set of all pairs (Li , Lj ) of D(L) such that yij ≥ ym .
Note that |D′ (L)| ≥ |D(L)|/2. For each pair (Li , Lj ) ∈ D′ (L), let xij be the
x-coordinate of the intersection of Li and Lj . We find the median xm of all such
xij ’s. By using our decision algorithm, we can determine in O(mn) time whether
x∗ > xm , x∗ < xm , or x∗ = xm . If x∗ = xm , our decision algorithm will find q ∗
and we are done. Otherwise, without loss of generality, we assume x∗ < xm .
Now for each pair (Li , Lj ) of D′ (L) with xij ≥ xm and yij ≥ ym (there are
at least |D′ (L)|/2 such pairs), we can prune either Pi or Pj , as follows. Indeed,
one of the lines in such a pair (Li , Lj ), say Li , has a negative slope and does not
intersect the region R = {(x, y) | x < xm , y < ym } (e.g., see Fig. 2). Suppose Li
is the dividing line of two relevant planes hk1 and hk2 of two uncertain points
Pk1 and Pk2 of P ∗ . It follows that either hk1 ≥ hk2 or hk1 ≤ hk2 holds on the
region R. Since q ∗ ∈ R, one of Pk1 and Pk2 can be pruned.
13
Li
x = xm
Lj
y = ym
q∗
Figure 2. The intersection of Li and Lj is in the first quarter of the intersection of
x = xm and y = ym while q ∗ is in the interior of the third quarter.
As a summary, the above pruning algorithm prunes at least |P ∗ |/16 ≥ n/32
uncertain points and the total time is O(mn).
4.3
Wrapping Things Up
The algorithm in the above two subsections either computes q ∗ or prunes at least
n/32 uncertain points from P in O(mn) time. We assume the latter case happens. Then we apply the same algorithm recursively on the remaining uncertain
points for O(log n) steps, after which only a constant number of uncertain points
remain. The total running time can be described by the following recurrence:
T (m, n) = T (m, 31·n
32 ) + O(mn). Solving the recurrence gives T (m, n) = O(mn).
Let P ′ be the set of the remaining uncertain points, with |P ′ | = O(1). Hence,
the rectilinear center q ∗ is determined by P ′ . In other words, q ∗ is also a rectilinear center of P ′ . In fact, like other standard prune-and-search algorithms, the
way we prune uncertain points of P guarantees that any rectilinear center of P
is also a rectilinear center of P ′ , and vice versa. By using an approach similar to
that in Section 4.1, Lemma 3 finally computes q ∗ based on P ′ in O(m) time.
Lemma 3. The rectilinear center q ∗ can be computed in O(m) time.
Proof. Let c = |P ′ |, which is a constant. Let X ′ = ∪Pi ∈P ′ Xi and Y ′ = ∪Pi ∈P ′ Yi .
We apply the same recursive algorithm in Section 4.1 on X ′ and Y ′ for O(log m)
steps, after which we will obtain a rectangle R such that R contains q ∗ and for
each Pi ∈ P ′ , the x-range (resp., y-range) of R only contains a constant number
of values of Xi (resp., Yi ), and thus R intersects a set Gi (R) of only a constant
number of cells of Gi . Therefore, for each Pi ∈ P ′ , only the surface patches
of Edi (x, y) defined on the cells of Gi (R) are relevant for computing q ∗ . The
supporting planes of these surface patches can be determined immediately after
the above O(log m) recursive steps. By the same analysis as in Section 4.1, all
above can be done in O(c · m) time.
The above found O(c) “relevant” supporting planes such that q ∗ corresponds
to a lowest point in the upper envelope of them. Consequently, q ∗ can be found
in O(c) time by applying the linear-time algorithm for the 3D LP problem [15]
on these O(c) relevant supporting planes.
⊓
⊔
This finishes our algorithm for computing q ∗ , which runs in O(mn) time.
14
Theorem 2. A rectilinear center q ∗ of the uncertain points of P in the plane
can be computed in O(mn) time.
5
Concluding Remarks
In this paper, we refine the prune-and-search technique [15] to solve in linear
time the rectilinear one-center problem on uncertain points in the plane. Note
that the problem can also be considered as the one-center problem on uncertain
points in the plane under the L1 distance metric. Since the L∞ and L1 metrics
are closely related to each other (by rotating the coordinate axes by 45◦ ), the
same problem under the L∞ metric can be solved in linear time as well.
The Euclidean version of the problem seems more natural. Unfortunately,
even if P contains only one uncertain point P1 and all locations of P1 have the
same probability, finding a center for P1 is essentially the 1-median problem in
the plane, which is known as the Weber problem and no exact algorithm exists
for it due to the computation challenge [4].
References
1. P.K. Agarwal, S.-W. Cheng, Y. Tao, and K. Yi. Indexing uncertain data. In Proc.
of the 28th Symposium on Principles of Database Systems (PODS), pages 137–146,
2009.
2. P.K. Agarwal, A. Efrat, S. Sankararaman, and W. Zhang. Nearest-neighbor searching under uncertainty. In Proc. of the 31st Symposium on Principles of Database
Systems (PODS), pages 225–236, 2012.
3. P.K. Agarwal, S. Har-Peled, S. Suri, H. Yıldız, and W. Zhang. Convex hulls under
uncertainty. In Proc. of the 22nd Annual European Symposium on Algorithms
(ESA), pages 37–48, 2014.
4. C. Bajaj. The algebraic degree of geometric optimization problems. Discrete and
Computational Geometry, 3:177–191, 1988.
5. P. Brass, C. Knauer, H.-S. Na, C.-S. Shin, and A. Vigneron. The aligned k-center
problem. International Journal of Computational Geometry and Applications,
21:157–178, 2011.
6. T.M. Chan. More planar two-center algorithms. Computational Geometry: Theory
and Applications, 13(3):189–198, 1999.
7. G. Cormode and A. McGregor. Approximation algorithms for clustering uncertain
data. In Proc. of the 27th Symposium on Principles of Database Systems (PODS),
pages 191–200, 2008.
8. M. de Berg, M. Roeloffzen, and B. Speckmann. Kinetic 2-centers in the blackbox model. In Proc. of the 29th Annual Symposium on Computational Geometry
(SoCG), pages 145–154, 2013.
9. X. Dong, A.Y. Halevy, and C. Yu. Data integration with uncertainty. In Proceedings
of the 33rd International Conference on Very Large Data Bases, pages 687–698,
2007.
10. A. Foul. A 1-center problem on the plane with uniformly distributed demand
points. Operations Research Letters, 34(3):264–268, 2006.
15
11. M. Hoffmann. A simple linear algorithm for computing rectilinear 3-centers. Computational Geometry, 31(3):150–165, 2005.
12. P. Kamousi, T.M. Chan, and S. Suri. Closest pair and the post office problem for
stochastic points. In Proc. of the 12nd International Workshop on Algorithms and
Data Structures (WADS), pages 548–559, 2011.
13. P. Kamousi, T.M. Chan, and S. Suri. Stochastic minimum spanning trees in Euclidean spaces. In Proc. of the 27th Annual Symposium on Computational Geometry (SoCG), pages 65–74, 2011.
14. A. Karmakar, S. Das, S.C. Nandy, and B.K. Bhattacharya. Some variations on
constrained minimum enclosing circle problem. Journal of Combinatorial Optimization, 25(2):176–190, 2013.
15. N. Megiddo. Linear-time algorithms for linear programming in R3 and related
problems. SIAM Journal on Computing, 12(4):759–776, 1983.
16. N. Megiddo and K.J. Supowit. On the complexity of some common geometric
location problems. SIAM Journal on Comuting, 13:182–196, 1984.
17. S. Suri and K. Verbeek. On the most likely voronoi diagram and nearest neighbor searching. In Proc. of the 25th International Symposium on Algorithms and
Computation (ISAAC), pages 338–350, 2014.
18. S. Suri, K. Verbeek, and H. Yıldız. On the most likely convex hull of uncertain
points. In Proc. of the 21st European Symposium on Algorithms (ESA), pages
791–802, 2013.
19. H. Wang and J. Zhang. Line-constrained k-median, k-means, and k-center problems in the plane. In Proc. of the 25th International Symposium on Algorithms
and Computation (ISAAC), pages 104–115, 2014.
20. H. Wang and J. Zhang. Computing the center of uncertain points on tree networks.
In Proc. of the 14th Algorithms and Data Structures Symposium (WADS), pages
606–618, 2015.
21. H. Wang and J. Zhang. One-dimensional k-center on uncertain data. Theoretical
Computer Science, online first, 2015.
16
| 8cs.DS
|
A Note on Easy and Efficient Computation
of Full Abelian Periods of a Word1
G. Ficia , T. Lecroqb, A. Lefebvreb , É. Prieur-Gastonb, W. F. Smythc
a Dipartimento
di Matematica e Informatica, Università di Palermo, Italy
EA4108, Université de Rouen, 76821 Mont-Saint-Aignan Cedex, France
c Dept of Computing and Software, McMaster University, Hamilton ON L8S 4K1, Canada and Faculty of Engineering &
Information Technology, Murdoch University, Murdoch WA 6150, Australia
arXiv:1510.00634v1 [cs.DS] 2 Oct 2015
b LITIS
Abstract
Constantinescu and Ilie (Bulletin of the EATCS 89, 167–170, 2006) introduced the idea of an Abelian
period with head and tail of a finite word. An Abelian period is called full if both the head and the tail are
empty. We present a simple and easy-to-implement O(n log log n)-time algorithm for computing all the full
Abelian periods of a word of length n over a constant-size alphabet. Experiments show that our algorithm
significantly outperforms the O(n) algorithm proposed by Kociumaka et al. (Proc. of STACS, 245–256, 2013)
for the same problem.
Keywords: Abelian period; Abelian power; weak repetition; design of algorithms; text algorithms;
combinatorics on words.
1. Introduction
The study of repetitions in words is a classical topic in Stringology. A word is called an (integer) power
if it can be written as the concatenation of two or more copies of another word, like barbar. However, any
word can be written as a fractional power; that is, given a word w, one can always find a word u such that
w = un u′ , where u′ is a (possible empty) prefix of u and n is an integer greater than or equal to one. In
this case, the length of u is called a period of the word w. A word w can have different periods, the least
of which is usually called the period of w.
Recently, a natural extension of this setting has been considered involving the notion of commutative
equivalence. Two words are called commutatively equivalent if they have the same number of occurrences of
each letter; that is, if one is an anagram of the other. An Abelian power (also called a weak repetition [2])
is a word that can be written as the concatenation of two or more words that are commutatively equivalent,
like iceddice.
Recall that the Parikh vector Pw of a word w is the vector whose ith entry is the number of occurrences
of the ith letter of the alphabet in w. For example, given the (ordered) alphabet Σ = {a, b, c}, the Parikh
vector of the word w = aaba is Pw = (3, 1, 0). Two words are therefore commutatively equivalent if and
only if they have the same Parikh vector.
Constantinescu and Ilie [3] introduced the definition of an Abelian period with head and tail of a word
w over a finite ordered alphabet Σ = {a1 , a2 , . . . , aσ }: An integer p > 0 is an Abelian period of w if one
can write
Pσ w = u0 u1 · · · uk−1 uk where for 0 < i < k all the factors ui ’s have the same Parikh vector P such
that j=1 P[j] = p and the Parikh vectors of u0 and uk are “contained” in P, in the sense that they are
1 The
results in this note have been presented in preliminary form in [1].
Email addresses: [email protected] (G. Fici), [email protected] (T. Lecroq),
[email protected] (A. Lefebvre), [email protected] (É. Prieur-Gaston), [email protected] (W.
F. Smyth)
Published in Discrete Applied Mathematics
October 5, 2015
proper sub-Parikh vectors of P (see next section for the formal definition of “contained”). In this case, u0
and uk are called the head and the tail of the Abelian period p, respectively. This definition of an Abelian
period matches that of an Abelian power when u0 and uk are both empty and k > 2.
As an example, the word w = abaababa over the alphabet Σ = {a, b} can be written as w = u0 u1 u2 u3 ,
where u0 = ab, u1 = aab, u2 = aba, u3 = ε, with ε the empty word, so that 3 is an Abelian period of w with
Parikh vector (2, 1) (the Parikh vector of u0 is (1, 1) and that of u3 is (0, 0) which are both “contained” in
(2, 1)). Notice that w has also Abelian period 2, since it can be written as w = u0 u1 u2 u3 u4 , with u0 = a,
u1 = ba, u2 = ab, u3 = ab, u4 = a.
This example shows that a word can have different Abelian periods. Moreover, a word can have the
same Abelian period p corresponding to different factorizations; that is, with different heads. Actually, a
word of length n can have Θ(n2 ) many different Abelian periods [4], if these are represented in the form
(h, p), where h is the length of the head—the length of the tail is uniquely determined by h and p.
Recently [5, 4] we described algorithms for computing all the Abelian periods of a word of length n in
time O(n2 × σ). This was improved to time O(n2 ) in [6]. In [7] the authors derived an efficient algorithm
for computing the Abelian periods based on prior computation of the Abelian squares.
An Abelian period is called full if both the head and the tail are empty. Clearly, a full Abelian period
is a divisor of the length of the word.
A preliminary version of the present paper appeared in [1] where we presented brute force algorithms to
compute full Abelian periods and Abelian periods without head and with tail in O(n2 ) time and a quasilinear time algorithm QLFAP for computing all the full Abelian periods of a word. In [8] Kociumaka et
al. gave a linear time algorithm LFAP for the same problem. Here we first briefly outline LFAP, followed
by a description of QLFAP. Then, extending the presentation in [1], we add an experimental section to
demonstrate that our algorithm significantly outperforms LFAP in practice, both on pseudo-randomly generated and genomic data. Our method has the additional advantage of being conceptually simple and easy
to implement.
2. Notation
Let Σ = {a1 , a2 , . . . , aσ } be a finite ordered alphabet of cardinality σ and Σ∗ the set of finite words over
Σ. We let |w| denote the length of the word w. Given a word w = w[0 . . n − 1] of length n > 0, we write
w[i] for the (i + 1)-th symbol of w and, for 0 6 i 6 j < n, we write w[i . . j] for the factor of w from the
(i + 1)-th symbol to the (j + 1)-th symbol, both included. We let |w|a denote the number of occurrences of
the symbol a ∈ Σ in the word w.
The Parikh vector of w, denoted by Pw , counts the occurrences of each letter of Σ in w, that is,
Pw = (|w|a1 , . . . , |w|aσ ). Notice that two words have the same Parikh vector if and only if one word is a
permutation of the other (in other words, an anagram).
Given the Parikh vector Pw of a word w, we let Pw [i] denote its i-th component and |Pw | its norm,
defined P
as the sum of its components. Thus, for w ∈ Σ∗ and 1 6 i 6 σ, we have Pw [i] = |w|ai and
σ
|Pw | = i=1 Pw [i] = |w|.
Finally, given two Parikh vectors P, Q, we write P ⊂ Q if P[i] 6 Q[i] for every 1 6 i 6 σ and |P| < |Q|.
This makes precise the notion of “contained” used in the Introduction.
Definition 1 (Abelian period [3]). A word w has an Abelian period (h, p) if w = u0 u1 · · · uk−1 uk such
that:
• Pu0 ⊂ Pu1 = · · · = Puk−1 ⊃ Puk ,
• |u0 | = h, |u1 | = p.
We call u0 and uk respectively the head and the tail of the Abelian period. Notice that the length
t = |uk | of the tail is uniquely determined by h, p and |w|, namely t = (|w| − h) mod p.
The following lemma gives a bound on the maximum number of Abelian periods of a word.
2
Lemma 1 ([5]). The maximum number of different Abelian periods (h, p) for a word of length n over an
alphabet of size σ is Θ(n2 ).
Proof. The word (a1 a2 · · · aσ )n/σ has Abelian period (h, p) for any p ≡ 0 mod σ and every h such that
0 6 h 6 min(p − 1, n − p).
An Abelian period is called full if it has head and tail both empty. We are interested in computing all
the full Abelian periods of a word. Notice that a full Abelian period of a word of length n is a divisor of n.
In the remainder of this note, we will therefore write that a word w has an Abelian period p if and only if
it has full Abelian period (0, p).
3. Previous work
We now outline the linear algorithm LFAP given in [8].
Let w be a word of length n. Let Pwi = Pw[0..i] . Two positions i, j ∈ {1, . . . , n} are called proportional,
which is denoted by i ∼ j, if Pwi [k] = c × Pwj [k] for each k, where c is a real number independent of k.
An integer k is called a candidate (as a potential Abelian period) if i ∼ k for each i ∈ Mult(k, n) where
Mult(k, n) is the set of multiples of k not exceeding n, or equivalently k ∼ 2k ∼ 3k ∼ · · · .
A positive integer q ≤ n is a full Abelian period of w if and only if q | n and q is a candidate.
The algorithm first computes the set A = {k : k ∼ n} and then the set B = {d ∈ Div (n) : Mult(d, n) ⊆
A}.
Let ℓ be the rank in the alphabet of a least frequent letter a of w. Let q0 be the position of the first
occurrence of a in w. For i ∈ {q0 , . . . , n} let γi = Pwi /Pwi [ℓ]. Vectors γi are used in order to deal with
vector equality instead of vector proportionality.
If i, j ∈ {q0 , . . . , n} then i ∼ j is equivalent to γi = γj . The problem of computing the set A reduces to
check if δi = γn − γi is equal to (0, . . . , 0) for q0 ≤ i ≤ n − 1. This is done in linear time by using what the
authors called a diff-representation of the γi ’s (see [8] for details).
The authors show that the set B can be computed in linear time. This can be done by showing that
after an O(n) time preprocessing, the value gcd(k, ℓ) for any k, ℓ ∈ {1, . . . , n} can be computed in constant
time (see [8] for details).
4. The new algorithm
In this section, we describe our algorithm QLFAP for computing all the full Abelian periods of a word
w of length n over a constant-size alphabet Σ = {a1 , a2 , . . . , aσ }.
In a linear-time preprocessing phase, we compute Pw [j] for j = 1, 2, . . . , σ, i.e., the components of the
Parikh vector of the word w. Let g denote the greatest common divisor of the elements of the Parikh vector
of w, computable in O(σ + log n/σ) time [9]. Then we compute s = n/g. We can suppose σ ≥ 2 and g > 1,
√
otherwise the solution is trivial. In O( g) time we compute a stack D of all divisors d ≥ 1 of g in increasing
order.
Observation 1. The only possible full Abelian periods of w are of the form p = d × s, where d is an entry
in D. Therefore, the smallest possible full Abelian period of w is s.
Definition 2 (scaled). A factor u of a word w is scaled if and only if there exists k > 0 such that
Pu = k × Pw /s, where s = n/g is the smallest possible Abelian period of w.
Observation 2. If w = uv and u is scaled, then v is scaled.
A scaled factor is called irreducible if it cannot itself be factored into scaled factors. According to
Definition 2 and Observation 2, every word w can be factored uniquely into irreducible scaled factors by
computing the shortest scaled prefix u of w = uv and then factoring the suffix v recursively, until this
suffix is empty. This factorization is computed by calling procedure ComputeL given in Figure 1, which
3
ComputeL(w, s, g, Pw )
1 (i, T, L) ← (0, 0, 0|w| )
2 while i 6 |w| − s do
3
j←0
4
t←0
5
count ← 0σ
6
while j < σ do
7
t←t+1
8
for k ← 1 to s do
9
count[w[i]] ← count[w[i]] + 1
10
i←i+1
11
j←0
12
while j < σ and ⌊count[j]/t⌋ = Pw [j]/g do
13
j ←j+1
14
L[i − t × s] ← 1
15
T ← max(T, t)
16 return (L, T )
Figure 1: Algorithm computing array L such that L[i] = 1 iff i is the starting position of a scaled suffix of
w.
computes the scaled prefixes of w (or analogously, by the previous observation, the scaled suffixes of w). It
returns a boolean array L of length n defined by: L[i] = 1 if and only if i is the starting position of a scaled
suffix of w. It also returns the value T such that the longest scaled factor computed is of length s × T .
Observation 3. Consider two positions i, j, with 0 6 i < j < |w|, such that L[i] = L[j] = 1. Then
w[i . . j − 1] is scaled.
Observation 4. A full Abelian period of w must be greater than or equal to s × T .
Proposition 2. The algorithm ComputeL computes the boolean array L of a word w of length n over an
alphabet of size σ in time O(n).
Proof. The internal loop in line 6 is performed until a scaled factor of length t × s is found: the loop in
line 8 computes the Parikh vector of the factor w[i − t × s. . i] and the loop in line 12 verifies that this factor
is scaled: its length is equal to t × s and it should hold for every letter j that count[j]/(t × s) = Pw [j]/n.
This can be rewritten as count[j]/t = Pw [j]/g since n = s × g. Since, by definition of g, Pw [j]/g is an
integer the test can be reduced to ⌊count[j]/t⌋ = Pw [j]/g since it should hold for every letter. This avoids
to use real numbers. Furthermore the Pw [j]/g are constant for a given w and can thus be precomputed.
When such a scaled factor is found, its starting position is i − t × s and L[i − t × s] is set to 1. Since the
algorithm starts from position i = 0, therefore, according to Observation 2, array L is filled correctly.
The algorithm visits each position i in w exactly once, and corresponding to each i performs a constanttime processing. The internal loop in line 12 is performed at most σ times, every s positions. It is performed
|w|/s times and since σ 6 s, each iteration costs O(s) time. Thus the algorithm runs in time O(|w|).
Example 1. w = abaababbbabaabbabbaaabbababbaa, Pw = (15, 15).
i
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
w[i] a b a a b a b b b a b a a b b a b b a a a b b a b a b b a a
L[i] 1 0 1 0 0 0 0 0 1 0 1 0 1 0 1 0 1 0 0 0 1 0 1 0 1 0 1 0 0 0
4
The word w is factored into irreducible scaled factors in the following way: w = ab · aababb · ba · ba · ab ·
ba · bbaa · ab · ba · ba · bbaa. Since the length of the longest word in this factorization is 6 and s = 2, the
value of T is 3.
Lemma 3. The word w has full Abelian period p if and only if for every k = 0, 1, . . . , |w|/p − 1, k × p is
the starting position of a scaled suffix of w.
Proof. Suppose that p is an Abelian period of w. According to Observation 3, we have that for every
k = 0, 1, . . . , |w|/p − 1 the position k × p is the starting position of a scaled factor and then of a scaled suffix
of w.
Conversely, if k × p is the starting position of a scaled suffix of w for every k = 0, 1, . . . , |w|/p − 1, it
follows from Observation 3 that w[k × p . . (k + 1) × p − 1] is scaled and consequently that p is an Abelian
period of w.
The algorithm QuasiLinearFullAbelianPeriods(w), given in Figure 2, computes all the full Abelian
periods of a word w. It first computes the greatest common divisor g of the elements of Pw (line 1). Then it
computes s = n/g (line 2). This is followed by the computation of the array L giving the starting positions
of the scaled suffixes and of the threshold T (line 3). Then it computes the stack D of all the divisors of g
greater than or equal to T (line 4). Then it checks for every divisor d of g in D, whether all the multiples
of d are starting positions of scaled suffixes (lines 6–12).
Example 2. w = abaababbbabaabbabbaaabbababbaa. We have n = 30, Pw [1] = Pw [2] = 15, g = 15,
s = 2, D = (1, 3, 5, 15). When d = 1, p = 2 and 4 = 2 × p is not a starting position of a scaled suffix. When
d = 3, p = 6 and 6 is not a starting position of a scaled suffix. When d = 5, p = 10 and 10, 20 are starting
positions of scaled suffixes, hence 10 is a full Abelian period of w. The case where d = 15 is trivial since it
corresponds to Abelian period n. Thus the algorithm returns {10, 30}, which is the set of full Abelian periods
of w.
Theorem 4. The algorithm QuasiLinearFullAbelianPeriods(w) computes all the full Abelian periods
of a word w of length n over an alphabet of size σ in time O(n log log n) and space O(n).
Proof. The correctness comes from Observation 1 and from Lemma 3. The algorithm QuasiLinearFullAbelianPeriods(w) scans all the multiples of the divisors d ∈ D, whose number is equal to the sum of
the divisors of s, which in turn is O(n log log n) [10]. The algorithm requires O(n) space for L and D.
5. Experimental results
5.1. Implementation
We implemented both QuasiLinearFullAbelianPeriods (QLFAP) and the algorithm LinearFullAbelianPeriods (LFAP) of [8] in a homogeneous way in C.
LFAP requires seven integer arrays of space O(n) to compute the diff-representation plus five more for
the preprocessing of the set B (see Sect. 3 for terminology). On the other hand, QLFAP requires only two
integer arrays L and R of length n, together with the Parikh vector of length σ and the stack D (implemented
as an integer array of length g).
In both algorithms, the memory space for the different arrays has been allocated one by one.
All the experiments have been performed on a computer with a 1.3 GHz Intel Core i5 processor and 4
Go 1600 MHz DDR3 RAM.
5
QuasiLinearFullAbelianPeriods(w = [w0 . . wn−1 ])
1 g ← greatest common divisor of elements of Pw
2 s ← n/g
3 (L, T ) ← ComputeL(w, s, g, Pw )
4 D ← divisors of g greater or equal than T
5 R←∅
6 while d < g do
7
k←1
8
while k × d × s < n and L[k × d × s] = 1 do
9
k ←k+1
10
if n 6 k × d × s then
11
R ← R ∪ {i}
12
d ← Pop(D)
13 R ← R ∪ {n}
14 return R
Figure 2: Algorithm computing the set of Abelian periods of word w.
5.2. Results
Figures 3 and 4 plots the raw execution times versus the word length (in logarithmic scale) while Figures 5–7 shows execution time ratios for LFAP over QLFAP. Experiments were performed on alphabet sizes
2, 5, 10 and 20, and input words of lengths 1000, 2000, . . . , 10000. For each alphabet size and word length,
1000 words were randomly generated.
In order to verify the correct execution of both algorithms, words have been generated in such a way
that they have at least one non trivial Abelian period. Both algorithms have been compiled with the same
options and run on the same machine under the same conditions.
The plots show that, on the range of tested values, the two algorithms have a linear behaviour. The
results show that our algorithm is significantly faster than LinearFullAbelianPeriods algorithm. One
can also observe that the ratio increases both with word length and with the length of the Abelian period (see
Figure 7). This may be due to the larger number of arrays needed by the LinearFullAbelianPeriods
algorithm.
Similar results were obtained with the E.coli file from the Canterbury Large Corpus (Figures 8 and 9).
The source codes of LFAP and QLFAP algorithms can be found at bioinfo.univ-rouen.fr/qlfap.
6. Conclusions and perspectives
In this note we presented an algorithm for computing all the full Abelian periods of a word of length n
in time O(n log log n) and space O(n). We showed that this algorithm is very efficient in practice, even if
its theoretical worst-case behavior is superlinear.
Our experiments show that it is between 6 to 19 times faster than the linear time algorithm of [8] and
we can expect that the ratio will be even larger for longer strings. Our method requires less space. In
addition the latter algorithm needs to compare real numbers, which can be tricky in practice, while our
method avoids this difficulty. Furthermore, our algorithm is conceptually simple, and stems from the notion
of scaled factorization of a word, which could be of some interest in other related problems on abelian
combinatorics on words.
A possible direction of further investigation could be that of extending the same approach to the computation of all the abelian periods with empty head but non-empty tail, or even to compute the smallest of
such abelian periods.
6
(a)
(b)
250
250
QFLAP
FLAP
QFLAP
FLAP
150
150
t
200
t
200
100
100
50
50
0
1000
10000
0
1000
1x106
100000
10000
n
1x106
100000
n
(c)
(d)
250
300
QFLAP
FLAP
QFLAP
FLAP
250
200
200
t
t
150
150
100
100
50
50
0
1000
10000
100000
0
1000
1x106
n
10000
100000
1x106
n
Figure 3: Running times of the LFAP and QLFAP algorithms for (a) alphabet of size 2 and at least one non
trivial Abelian period 5; (b) alphabet of size 5 and at least one non trivial Abelian period 5; (c) alphabet
of size 10 and at least one non trivial Abelian period 5; (d) alphabet of size 20 and at least one non trivial
Abelian period 5.
7
(a)
(b)
250
250
QFLAP
FLAP
QFLAP
FLAP
150
150
t
200
t
200
100
100
50
50
0
1000
10000
0
1000
1x106
100000
10000
n
1x106
100000
n
(c)
(d)
250
300
QFLAP
FLAP
QFLAP
FLAP
250
200
200
t
t
150
150
100
100
50
50
0
1000
10000
0
1000
1x106
100000
10000
n
1x106
100000
n
Figure 4: Running times of the LFAP and QLFAP algorithms for (a) alphabet of size 2 and at least one non
trivial Abelian period 20; (b) alphabet of size 5 and at least one non trivial Abelian period 20; (c) alphabet
of size 10 and at least one non trivial Abelian period 20; (d) alphabet of size 20 and at least one non trivial
Abelian period 20.
Word length
7.90
7.38
7.04
6.13
9.64 10.51 11.50 11.45 11.42 12.02 11.64 11.38 11.62
9.03 9.50 10.32 10.26 10.32 10.54 10.38 10.34 10.46
8.28 8.67 9.37 9.03 9.13 9.55 9.47 9.33 9.47
6.66 6.93 7.08 7.14 7.16 7.31 7.30 7.18 7.20
Word length
2
5
10
20
2 8.88 10.72 11.06
5 11.33 13.98 15.47
10 10.93 13.30 14.45
20 8.81 11.72 12.46
1000
(B) period 20
3000
Alphabet
2000
Alphabet
(A) period 5
1000
2000
4000
3000
5000
6000
7000
8000
9000
10000
4000
5000
6000
7000
8000
9000
10000
11.88
15.99
14.95
12.74
11.67
15.91
14.69
12.46
11.87
16.70
15.57
13.06
12.32
16.71
15.62
13.07
12.48
16.98
15.90
13.17
12.32
16.57
15.44
13.35
12.90
16.77
16.03
13.59
Figure 5: Execution time ratios for LFAP over QLFAP algorithms. (A) Ratios computed on random words
with at least one non trivial Abelian period 5. (B) Ratios computed on random words with at least one non
trivial Abelian period 20.
8
Word length
(B) period 20
Alphabet
(A) period 5
Alphabet
10000
2
5
10
20
90000
100000
11.62 12.38 12.32 15.82 15.86 15.84 16.27 16.65 15.94
10.46 10.85 11.25 13.67 13.84 13.90 14.34 14.31 13.98
9.47 10.07 10.30 11.86 11.86 11.86 12.13 12.08 11.86
7.20 7.65 7.71 8.54 8.57 8.59 8.71 9.24 8.75
Word length
16.72
14.46
12.20
8.92
2
5
10
20
20000
30000
40000
50000
60000
70000
80000
10000
20000
30000
40000
50000
60000
70000
80000
90000
100000
12.90
16.77
16.03
13.59
13.43
17.80
16.50
13.73
13.44
18.45
17.17
14.37
17.00
24.27
21.80
17.41
17.23
24.58
22.46
18.27
17.46
24.69
22.85
18.49
17.63
26.55
24.47
19.43
17.21
27.07
25.31
19.07
18.37
27.20
25.09
19.02
18.10
28.42
25.45
19.33
Figure 6: Execution time ratios for LFAP over QLFAP algorithms. (A) Ratios computed on random words
with at least one non trivial Abelian period 5. (B) Ratios computed on random words with at least one non
trivial Abelian period 20.
(B) period 20
Alphabet
(A) period 5
Alphabet
Word length
50000
100000
200000
2 15.86
5 13.84
10 11.86
20 8.54
16.72
14.42
12.20
8.92
17.87 19.29
15.44 16.24
12.88 13.89
9.53
9.34
Word length
2
5
10
20
400000
800000
1600000
19.50
16.99
14.44
10.07
20.35
17.66
15.08
10.57
50000
100000
200000
400000
800000
1600000
17.23
24.58
22.46
18.27
18.10
29.42
25.45
19.33
20.04
30.21
27.84
21.17
20.82
30.76
29.42
22.05
21.49
31.16
28.33
22.42
23.39
34.13
31.23
24.45
Figure 7: Execution time ratios for LFAP over QLFAP algorithms. (A) Ratios computed on random words
with at least one non trivial Abelian period 5. (B) Ratios computed on random words with at least one non
trivial Abelian period 20.
(a)
300
QFLAP
FLAP
250
t
200
150
100
50
0
1000
10000
100000
n
Figure 8: Running times of the LFAP and QLFAP algorithms for the E.coli file from the Canterbury
Large Corpus.
9
Word length 1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
Ratio
19.72 24.79 26.04 26.95 25.23 28.31 28.79 31.93 30.24 31.51
Word length 20000 30000 40000 50000
60000
70000
80000
90000
100000
Ratio
34.75 37.10 95.35 101.73 102.57 107.82 108.00 116.06 114.54
Figure 9: Execution time ratio between LFAP and QLFAP algorithms on words taken from the E.coli file
from the Canterbury Large Corpus.
Acknowledgements
The authors thank the anonymous reviewers that greatly improved a first version of this paper.
References
[1] G. Fici, T. Lecroq, A. Lefebvre, E. Prieur-Gaston, W. F. Smyth, Quasi-linear time computation of the Abelian periods
of a word, in: J. Holub, J. Zdárek (Eds.), Prague Stringology Conference, PSC 2012. Prague, Czech Republic, August
27–28, 2012. Proceedings, Czech Technical University in Prague, 2012, pp. 103–110.
[2] L. J. Cummings, W. F. Smyth, Weak repetitions in strings, Journal of Combinatorial Mathematics and Combinatorial
Computing 24 (1997) 33–48.
[3] S. Constantinescu, L. Ilie, Fine and Wilf’s theorem for abelian periods, Bulletin of the European Association for Theoretical
Computer Science 89 (2006) 167–170.
[4] G. Fici, T. Lecroq, A. Lefebvre, É. Prieur-Gaston, Algorithms for computing Abelian periods of words, Discrete Applied
Mathematics 163 (2014) 287–297.
[5] G. Fici, T. Lecroq, A. Lefebvre, É. Prieur-Gaston, Computing Abelian periods in words, in: J. Holub, J. Žďárek (Eds.),
Proceedings of the Prague Stringology Conference 2011, Czech Technical University in Prague, Czech Republic, 2011, pp.
184–196.
[6] M. Christou, M. Crochemore, C. S. Iliopoulos, Identifying all abelian periods of a string in quadratic time and relevant
problems, International Journal of Foundations of Computer Science 23 (6) (2012) 1371–1384.
[7] M. Crochemore, C. Iliopoulos, T. Kociumaka, M. Kubica, J. Pachocki, J. Radoszewski, W. Rytter, W. Tyczyński, T. Waleń,
A note on efficient computation of all abelian periods in a string, Information Processing Letters 113 (3) (2013) 74–77.
[8] T. Kociumaka, J. Radoszewski, W. Rytter, Fast algorithms for abelian periods in words and greatest common divisor
queries, in: Proc. of STACS, 2013, pp. 245–256, available from http://stacs-conf.org.
[9] G. Bradley, Algorithm and bound for the greatest common divisor of n integers, Comm. ACM 13 (7) (1970) 433–436.
[10] T. H. Gronwall, Some asymptotic expressions in the theory of numbers, Transactions of the American Mathematical
Society 14 (1913) 113–122.
10
| 8cs.DS
|
ALGEBRAIC PROPERTIES OF THE BINOMIAL EDGE IDEAL OF A
COMPLETE BIPARTITE GRAPH
arXiv:1301.0789v1 [math.AC] 4 Jan 2013
PETER SCHENZEL, SOHAIL ZAFAR
Abstract. Let JG denote the binomial edge ideal of a connected undirected graph on
n vertices. This is the ideal generated by the binomials xi yj − xj yi , 1 ≤ i < j ≤ n, in
the polynomial ring S = K[x1 , . . . , xn , y1 , . . . , yn ] where {i, j} is an edge of G. We study
the arithmetic properties of S/JG for G, the complete bipartite graph. In particular we
compute dimensions, depths, Castelnuovo-Mumford regularities, Hilbert functions and
multiplicities of them. As main results we give an explicit description of the modules of
deficiencies, the duals of local cohomology modules, and prove the purity of the minimal
free resolution of S/JG .
1. Introduction
The main intention of the present paper is the study of the binomial edge ideal of the
complete bipartite graph. Let G denote a connected undirected graph on n vertices labeled
by [n] = {1, 2, . . . , n}. For an arbitrary field K let S = K[x1 , . . . , xn , y1 , . . . , yn ] denote the
polynomial ring in 2n variables. To the graph G one can associate the binomial edge ideal
JG ⊂ S generated by binomials xi yj −xj yi , i < j, such that {i, j} is an edge of G. This is an
extension of the edge ideal (generated by the monomials) as it was studied for instance in
[12]. The binomial edge ideals of a graph G might have applications in algebraic statistics
(see [6]). By the work of Herzog et al. (see [6]) the minimal primary decomposition of
JG is known. Besides of that not so much is known about the arithmetic properties of
S/JG . If G denotes the complete graph on n vertices, then S/JG is the coordinate ring
of the Segre embedding P1K × PnK . This is a variety of minimal degree. Therefore S/JG
is a Cohen-Macaulay ring with a linear resolution. In the paper Ene, Herzog and Hibi
(see [4]) they studied Cohen-Macaulayness property for some special classes of graphs.
By view of the primary decomposition of JG (see [6]) it follows that S/JG is not so often
a Cohen-Macaulay ring. As a certain generalization of the Cohen-Macaulay property the
second author has studied approximately Cohen-Macaulay rings (see [13]). In the present
paper we investigate the binomial edge ideal of another important class of graphs, namely
the complete bipartite graph G = Km,n (see the definitions in Section 3).
As the main result of our investigations we prove (among others) the following results:
Theorem 1.1. With the previous notation let JG ⊂ S denote the binomial edge ideal
associated to the complete bipartite graph G = Km,n .
2010 Mathematics Subject Classification. 05E40, 13H10, 13D45.
Key words and phrases. binomail edge ideal, complete bipartite graph, pure resolution.
This research was partially supported by the Higher Education Commission, Pakistan.
1
2
PETER SCHENZEL, SOHAIL ZAFAR
(a) dim S/JG = max{n + m + 1, 2m} and
m + 2, if n = 1,
depth S/JG =
n + 2, if m ≥ n > 1.
(b) There is an explicit expression of the Hilbert series and the multiplicity equals
(
1,
if m > n + 1 or n = 1 and m > 2,
e(S/JG ) =
2m, otherwise.
(c) The Castelnuovo-Mumford regularity is reg S/JG = 2 and S/JG admits a pure
minimal free resolution.
(d) There are at most 5 non-vanishing local cohomology modules HSi + (S/JG ). The
modules of deficiencies ω i (S/JG ) = HomK (HSi + (S/JG ), K) are either CohenMacaulay modules or the direct sum of two Cohen-Macaulay modules.
(e) S/JG is a Cohen-Macaulay canonically ring in the sense of [10].
For the details on the modules of deficiencies we refer to Section 4 of the paper. This
is - at least for us - the first time in the literature that there is a complete description of
the structure of the modules of deficiencies besides of sequentially Cohen-Macaulay rings
or Buchsbaum rings. Our analysis is based on the primary decomposition of JG as shown
in [6].
In Section 2 we start with preliminary and auxiliary results needed in the rest of the
paper. In particular we give a short overview on the modules of deficiencies. In Section
3 we study some of the properties of the the binomial edge ideal JG ⊂ S associated to
a complete bipartite graph. In Section 4 we give a complete list of all the modules of
deficiencies of the complete bipartite graphs. In the final Section 5 we prove the purity of
the minimal free resolution of S/JG . This is the heart of our investigations. It gives in a
natural way some non-Cohen-Macaulay rings with pure resolutions. We might relate our
investigations as a better understanding of general binomial edge ideals.
2. Preliminaries and auxiliary results
First of all we will introduce the notation used in the sequel. Moreover we summarize
a few auxiliary results that we need.
We denote by G a connected undirected graph on n vertices labeled by [n] =
{1, 2, . . . , n}. For an arbitrary field K let S = K[x1 , . . . , xn , y1, . . . , yn ] denote the polynomial ring in the 2n variables x1 , . . . , xn , y1 , . . . , yn . To the graph G one can associate
an ideal JG ⊂ S generated by all binomials xi yj − xj yi for all 1 ≤ i < j ≤ n such that
{i, j} is an edge of G. This construction was invented by Herzog et al. in [6] and [4]. At
first let us recall some of their definitions.
Definition 2.1. Fix the previous notation. For a set T ⊂ [n] let G̃T denote the complete
graph on the vertex set T . Moreover let G[n]\T denote the graph obtained by deleting all
vertices of G that belong to T .
Let c = c(T ) denote the number of connected components of G[n]\T . Let G1 , . . . , Gc
denote the connected components of G[n]\T . Then define
PT (G) = (∪i∈T {xi , yi }, JG̃1 , . . . , JG̃C(T ) ),
BINOMIAL EDGE IDEAL OF A COMPLETE BIPARTITE GRAPH
3
where G̃i , i = 1, . . . , c, denotes the complete graph on the vertex set of the connected
component Gi , i = 1, . . . , c.
The following result is important for the understanding of the binomial edge ideal of
G.
Lemma 2.2. With the previous notation the following holds:
(a) PT (G) ⊂ S is a prime ideal of height n − c + |T |, where |T | denotes the number of
elements of T .
(b) JG = ∩T ⊆[n] PT (G).
(c) JG ⊂ PT (G) is a minimal prime if and only if either T = ∅ or T 6= ∅ and
c(T \ {i}) < c(T ) each i ∈ T .
Proof. For the proof we refer to [6].
Therefore JG is the intersection of prime ideals. That is, S/JG is a reduced ring.
Moreover, we remark that JG is an ideal generated by quadrics and therefore homogeneous,
so that S/JG is a graded ring with natural grading induced by the N-grading of S. As a
technical tool we shall need the following result.
Proposition 2.3. Let I ⊂ S denote an ideal. Let f = f1 , . . . , fr denote an S/I-regular
sequence. Then f S ∩ I = f I.
Proof. It is easy to see that TorS1 (S/f S, S/I) ∼
= f S ∩ I/f I. Moreover
TorS1 (S/f S, S/I) ∼
= H1 (f ; S/I),
where Hi (f ; S/I) denotes the Koszul homology of f with respect to S/I. But these
homology modules vanish for i > 0.
Let M denote a finitely generated graded S-module. In the sequel we shall use also the
local cohomology modules of M with respect to S+ , denoted by H i (M), i ∈ Z. Note that
they are graded Artinian S-modules. We refer to the textbook of Brodmann and Sharp
(see [1]) for the basics on it. In particular the Castelnuovo-Mumford regularity reg M of
M is defined as
reg(M) := max{e(H i (M)) + i| depth(M) ≤ i ≤ dim(M)},
where e(H i (M)) is the least integer m such that, for all k > m, the degree k part of the
i-th local cohomology module of M is zero. For our investigations we need the following
definition.
Definition 2.4. Let M denote a finitely generated graded S-module and d = dim M.
For an integer i ∈ Z put
ω i (M) = ExtS2n−i (M, S(−2n))
and call it the i-th module of deficiency. Moreover we define ω(M) = ω d (M) the canonical
module of M. We write also ω2× (M) = ω(ω(M)). These modules have been introduced
and studied in [8].
4
PETER SCHENZEL, SOHAIL ZAFAR
Note that by the graded version of Local Duality (see e.g. [1]) there is the natural
graded isomorphism ω i (M) ∼
= HomK (H i (M), K) for all i ∈ Z. For a finitely generated
graded S-module M and an integer i ∈ N we set
(Ass M)i = {p ∈ Ass M| dim S/p = i}.
In the following we shall summarize a few properties on the modules of deficiencies.
Proposition 2.5. Let M denote a finitely generated graded S-module and d = dim M.
(a) dim ω i (M) ≤ i and dim ω d (M) = d.
(b) (Ass ω i (M))i = (Ass M)i for all 0 ≤ i ≤ d.
(c) M satisfies the Serre condition S2 if and only if dim ω i (M) ≤ i−2 for all 0 ≤ i < d.
(d) There is a natural homomorphism M → ω d(ω d (M)). It is an isomorphism if and
only if M satisfies the Serre condition S2 .
(e) For a homogeneous ideal I ⊂ S there is a natural isomorphism ω d (ω d (S/I)) ∼
=
HomS (ω d(S/I), ω d (S/I)), d = dim S/I, and it admits the structure of a commutative Noetherian ring, the S2 -fication of S/I.
(f) The natural map S/I → HomS (ω d (S/I), ω d(S/I)), d = dim S/I, sends the unit
element to the identity map. Therefore it is a ring homomorphism.
Proof. The results are shown in [8] and [10]. The proofs in the graded case follow the
same line of arguments.
A decreasing sequence {Mi }0≤i≤d of a d-dimensional S-module M is called dimension
filtration of M, if Mi /Mi−1 is either zero or of dimension i for all i = 0, . . . , d, where
M−1 = 0. It was shown (see [9]) that the dimension filtration exists and is uniquely
determined.
Definition 2.6. An S-module M is called sequentially Cohen-Macaulay if the dimension
filtration {Mi }0≤i≤d has the property that Mi /Mi−1 is either zero or an i-dimensional
Cohen-Macaulay module for all i = 0, . . . , d, (see [9]). Note that in [9] this notion was
originally called Cohen-Macaulay filtered.
Note that a sequentially Cohen-Macaulay S-module M with depth M ≥ dim M − 1
was studied by Gôto (see [5]) under the name approximately Cohen-Macaulay. For our
purposes here we need the following characterization of sequentially Cohen-Macaulay
modules.
Theorem 2.7. Let M be a finitely generated graded S-module with d = dim M. Then the
following conditions are equivalent:
(i) M is a sequentially Cohen-Macaulay.
(ii) For all 0 ≤ i < d the module of deficiency ω i (M) is either zero or an i-dimensional
Cohen-Macaulay module.
(iii) For all 0 ≤ i ≤ d the modules ω i (M) are either zero or i-dimensional CohenMacaulay modules.
Proof. In the case of a local ring admitting a dualizing complex this result was shown in
[9, Theorem 5.5]. Similar arguments work also in the case of a finitely generated graded
S-module M. Note that the equivalence of (i) and (iii) was announced (without proof)
in [11].
BINOMIAL EDGE IDEAL OF A COMPLETE BIPARTITE GRAPH
5
3. Complete bipartite graphs
A bipartite graph is a graph whose vertices can be divided into two disjoint sets V1 and
V2 such every edge of G connects a vertex in V1 to one in V2 . Now the complete bipartite
graph is a bipartite graph G such that for any two vertices, v1 ∈ V1 and v2 ∈ V2 , v1 v2 is
an edge in G. If |V1 | = n and |V2 | = m then it is usually denoted by Kn,m . To simplifying
notations we denote it often by G.
Definition 3.1. For a sequence of integers 1 ≤ i1 < i2 < . . . < ik ≤ n + m let
I(i1 , i2 , . . . , ik ) denote the ideal generated by the 2 × 2 minors of the matrix
xi1 xi2 · · · xik
.
y i1 y i2 · · · y ik
Note that I(i1 , i2 , . . . , ik ) is the ideal of the complete graph on the vertex set {i1 , i2 , . . . , ik }.
Let JG be the binomial edge ideal of complete bipartite graph on [n + m] vertices and
JG̃ be the binomial edge ideal of complete graph on [n + m] vertices. We begin with a
lemma concerning the dimension of S/JG .
Lemma 3.2. Let G = Km,n , m ≥ n, denote the complete bipartite graph. Let G̃ denote
the complete graph on [n + m]. Let An = (x1 , . . . , xn , y1, . . . , yn ) for n ≥ 1 and Bm =
(xn+1 , . . . , xn+m , yn+1, . . . , yn+m) for n ≥ 2 and Bm = S for n = 1.
(a) JG = JG̃ ∩ An ∩ Bm is the minimal primary decomposition of JG .
(b) dim S/JG = max{n + m + 1, 2m}.
(c) (JG̃ ∩ An , JG̃ ∩ Bm ) = JG̃ .
Proof. We start with the proof of (a). We use the statement proved in Lemma 2.2. At first
consider the case m > n = 1. By view of Lemma 2.2 we have to find all ∅ =
6 T ⊆ [1 + m]
such that c(T \{i}) < c(T ). Clearly T0 = {1} satisfy the condition because c(T0 ) = m > 1.
Let T denote T ⊂ [1 + m] a subset different of T0 . Then If 1 ∈ T then c(T ) = m + 1 − |T |
and c(T \ {i}) = m + 2 − |T | for i 6= 1 and if 1 6∈ T then c(T ) = 1 and c(T \ {i}) = 1 for
all i ∈ T . Hence we have the above primary decomposition.
Now consider the case of m ≥ n ≥ 2. As above we have to find all ∅ 6= T ⊆ [n + m]
such that c(T \ {i}) < c(T ) for all i ∈ T . T1 = {1, 2, . . . , n} satisfy the above condition
because c(T ) = m and c(T \ {i}) = 1 for all i ∈ T . Similarly T2 = {n+ 1, n+ 2, . . . , n+ m}
also satisfies the above condition.
Our claim is that no other T ⊆ [n + m] satisfies this condition. If T1 6⊆ T and T2 6⊆ T
then c(T ) = 1 so in this case T does not satisfy the above condition. Now suppose that
T1 ( T then c(T ) = m − |T \ T1 | and c(T \ {i}) = m + 1 − |T \ T1 | if i ∈ T \ T1 . The same
argument works if T2 ( T . Hence we have JG = JG̃ ∩ An ∩ Bm .
Then the statement on the dimension in (b) is a consequence of the reduced primary
decomposition shown in (a). To this end recall that dim S/An = 2m, dim S/Bm = 2n and
dim S/JG̃ = n + m + 1.
For the proof of (c) we use the notation of the Definition 3.1. Then it follows that
JG̃ ∩ An = (I(1, . . . , n, n + i), i = 1, . . . , m, I(n + 1, . . . , n + m) ∩ An ).
Now An consists of an S/I(n + 1, . . . , n + m)-regular sequence and
I(n + 1, . . . , n + m) ∩ An = An I(n + 1, . . . , n + m)
6
PETER SCHENZEL, SOHAIL ZAFAR
by Proposition 2.3. Therefore we get
JG̃ ∩ An = (I(1, . . . , n, n + i), i = 1, . . . , m, An I(n + 1, . . . , n + m))
and similarly
JG̃ ∩ Bm = (I(j, n + 1, . . . , n + m), j = 1, . . . , n, Bm I(1, . . . , n)).
But this clearly implies that (JG̃ ∩ An , JG̃ ∩ Bm ) = JG̃ which proves the statement in
(c).
For the further computations we use the previous Lemma 3.2. In particular we use
three exact sequences shown in the next statement.
Corollary 3.3. With the previous notation we have the following three exact sequences.
(1) 0 → S/JG → S/JG̃ ∩ An ⊕ S/JG̃ ∩ Bm → S/JG̃ → 0.
(2) 0 → S/JG̃ ∩ An → S/JG̃ ⊕ S/An → S/(JG̃ , An ) → 0.
(3) 0 → S/JG̃ ∩ Bm → S/JG̃ ⊕ S/Bm → S/(JG̃ , Bm ) → 0.
Proof. The proof is an easy consequence of the primary decomposition as shown in Lemma
3.2. We omit the details.
Note that in case of n = 1 we have Bm = S therefore it is enough to consider the exact
sequence (2) as (1) and (3) gives no information.
Corollary 3.4. With the previous notation we have that
m + 2, if n = 1 ;
depth S/JG =
n + 2, if m ≥ n > 1
and reg S/JG ≤ 2.
Proof. The statement is an easy consequence of the short exact sequences shown in Corollary 3.3. To this end note that S/JG̃ , S/(JG̃ , An ) and S/(JG̃ , Bm ) are Cohen-Macaulay
rings of dimension n + m + 1, m + 1 and n + 1 respectively. Moreover reg S/JG̃ =
reg S/(JG̃ , An ) = reg S/(JG̃ , Bm ) = 1. By using the exact sequences it provides the
statement on the regularity. For the behaviour of the depth respectively the regularity in
short exact sequences see [2, Proposition 1.2.9] respectively [3, Corollary 20.19].
4. The modules of deficiency
The goal of this section is to describe all the local cohomology modules H i (S/JG ) of the
binomial edge ideal of a complete bipartite graph G. We do this by describing their Matlis
duals which by Local Duality are the modules of deficiencies. Moreover,
Pfor a homogeneous
ideal J ⊂ S let H(S/J, t) denote the Hilbert series, i.e. H(S/J, t) = i≥0 (dimK [S/J]i )ti .
We start our investigations with the so-called star graph. That is complete bipartite
graph Km,n with n = 1. For m ≤ 2 the ideal JG is a complete intersection generated by
one respectively two quadrics so let us assume that m > 2.
Theorem 4.1. Let G denote the star graph Km,1 . Then the binomial edge ideal JG ⊂ S
has the following properties:
(a) reg S/JG = 2.
(b) ω i (S/JG ) = 0 if and only if i 6∈ {m + 2, 2m}.
BINOMIAL EDGE IDEAL OF A COMPLETE BIPARTITE GRAPH
7
(c) ω 2m (S/JG ) ∼
= S/A1 (−2m)
(d) ω m+2 (S/JG ) is a (m + 2)-dimensional Cohen-Macaulay module and there is an
isomorphism ω m+2 (ω m+2 (S/JG )) ∼
= (JG̃ , A1 )/JG̃ .
Proof. We use the short exact sequence of Corollary 3.3 (2). It induces a short exact
sequence
0 → H m+1 (S/(JG̃ , A1 )) → H m+2 (S/JG ) → H m+2 (S/JG̃ ) → 0
and an isomorphism H 2m (S/JG ) ∼
= H 2m (S/A1 ). Moreover the Cohen-Macaulayness of
S/JG̃ , S/A1 and S/(JG̃ , A1 ) of dimensions m + 2, 2m and m + 1 respectively imply that
H i (S/JG ) = 0 if i 6∈ {m + 2, 2m}.
The short exact sequence on local cohomology induces the following exact sequence
0 → ω m+2 (S/JG̃ ) → ω m+2 (S/JG ) → ω m+1 (S/(JG̃ , A1 )) → 0
by Local Duality. Now we apply again local cohomology and take into account that both
ω m+2 (S/JG̃ ) and ω m+1 (S/(JG̃ , A1 )) are Cohen-Macaulay modules of dimension m + 2 and
m + 1 respectively. Then depth ω m+2 (S/JG ) ≥ m + 1. By applying local cohomology and
dualizing again it induces the following exact sequence
f
0 → ω m+2 (ω m+2 (S/JG )) → S/JG̃ → S/(JG̃ , A1 ) → ω m+1 (ω m+2 (S/JG )) → 0.
The homomorphism f is induced by the commutative diagram
S/JG̃
→
S/(JG̃ , A1 )
↓
↓
ω2× (S/JG̃ ) → ω2× (S/JG̃ , A1 ).
Note that the vertical maps are isomorphisms (see Proposition 2.5). Since the upper horizontal map is surjective the lower horizontal map is surjective too. Therefore
ω m+1 (ω m+2 (S/JG )) = 0. That is depth ω m+2 (S/JG ) = m + 2 and it is a Cohen-Macaulay
module. Moreover ω m+2 (ω m+2 (S/JG )) ∼
= (JG̃ , A1 )/JG̃ . This finally proves the statements
in (b), (c) and (d).
It is well known that reg S/JG = reg S/(JG̃ , A1 ) = 1 and reg S/A1 = 0. Then an
inspection with the short exact sequence of Corollary 3.3 shows that reg S/JG = 2.
In the next result we will consider the modules of deficiencies of the complete bipartite
graph G = Km,n , n ≥ 2.
Theorem 4.2. Let m ≥ n > 1 and assume that the pair (m, n) is different from (n + 1, n)
and (2n − 2, n). Then
(a) reg S/JG = 2.
(b) ω i (S/JG ) = 0 if and only if i 6∈ {n + 2, m + 2, 2n, m + n + 1, 2m} and there are the
following isomorphisms and integers
i
ω i (S/JG )
depth ω i (S/JG ) dim ω i (S/JG )
n+2
ω n+1 (S/(JG̃ , Bm ))
n+1
n+1
m+1
ω
(S/(JG̃ , An ))
m+1
m+1
m+2
2n
S/Bm (−2n)
2n
2n
m+n+1
n+m+1
ω
(S/JG̃ )
n+m+1
n+m+1
2m
S/An (−2m)
2m
2m
8
PETER SCHENZEL, SOHAIL ZAFAR
Proof. Under the assumption of n+1 < m < 2n−2 it follows that 2m > m+n+1 > 2n >
m + 2 > n + 2. Then the short exact sequences (see Corollary 3.3) induce the following
isomorphisms:
(1) H n+2(S/JG ) ∼
= H n+1 (S/(JG̃ , Bm )),
= H n+2 (S/JG̃ ∩ Bm ) ∼
m+2
m+2
∼
(S/JG̃ ∩ An ) ∼
(2) H
(S/JG ) = H
= H m+1 (S/(JG̃ , An )),
(3) H 2n (S/JG ) ∼
= H 2n (S/Bm ) and
2m
(4) H (S/JG ) ∼
= H 2m (S/An ).
Moreover there is the following short exact sequence
0 → H m+n+1 (S/JG ) → H m+n+1 (S/JG̃ ∩An )⊕H m+n+1 (S/JG̃ ∩Bm ) → H m+n+1 (S/JG̃ ) → 0
and H i (S/JG ) = 0 if i 6∈ {n + 2, m + 2, 2n, m + n + 1, 2m}.
Because of the short exact sequences in Corollary 3.3 there are isomorphisms
H m+n+1 (S/J ∩ Bm ) ∼
= H m+n+1 (S/J ) ∼
= H m+n+1 (S/J ∩ An ).
G̃
G̃
G̃
So by Local Duality we get a short exact sequence
0 → ω(S/JG̃ ) → ω(S/JG̃ ) ⊕ ω(S/JG̃ ) → ω m+n+1 (S/JG ) → 0.
This implies depth ω(S/JG ) ≥ n + m. Moreover by applying local cohomology and again
the Local Duality we get the following commutative diagram with exact rows
0→
S/JG
↓
→ S/JG̃ ∩ An ⊕ S/JG̃ ∩ Bm → S/JG̃ →
↓
k
0 → Ω(S/JG ) →
S/JG̃ ⊕ S/JG̃
0
f
→ S/JG̃ → ω m+n (ω m+n+1 (S/JG )) → 0,
where Ω(S/JG ) = ω m+n+1(ω m+n+1 (S/JG )). Now we show that ω m+n (ω m+n+1(S/JG )) = 0.
This follows since f is easily seen to be surjective. That is, ω m+n+1 (S/JG ) is a (m+n+1)dimensional Cohen-Macaulay module. Moreover f is a split homomorphism and therefore
Ω(S/JG ) ≃ S/JG̃ By duality this implies that ω m+n+1 (S/JG ) ∼
= ω(S/JG̃ ). This completes
the proof of the statements in (b). By similar arguments the other cases for (m, n) different
of (n + 1, n) and (2n − 2, n) can be proved. We omit the details. Clearly reg S/JG = 2 as
follows by (b).
As a next sample of our considerations let us consider the case of the complete bipartite
graph Km,n with (m, n) = (n + 1, n).
Theorem 4.3. Let m = n + 1 and n > 3. Then:
(a) reg S/JG = 2.
(b) ω i (S/JG ) = 0 if and only if i 6∈ {n+ 2, n+ 3, 2n, 2n+ 2} and there are the following
isomorphisms and integers
i
ω i (S/JG )
depth ω i (S/JG ) dim ω i (S/JG )
n+2
ω n+1(S/(JG̃ , Bn+1 ))
n+1
n+1
ω n+2 (S/(JG̃ , An ))
n+2
n+2
n+3
2n
S/Bn+1 (−2n)
2n
2n
2n + 2 ω(S/JG̃ ) ⊕ S/An (−2n − 2)
2n + 2
2n + 2.
Proof. By applying the local cohomology functors H · (−) to the exact sequence (2) in
Corollary 3.3 we get the following:
BINOMIAL EDGE IDEAL OF A COMPLETE BIPARTITE GRAPH
9
(1) H n+3(S/JG̃ ∩ An ) ∼
= H n+2 (S/(JG̃ , An )),
2n+2
(2) H
(S/JG̃ ∩ An ) ∼
= H 2n+2 (S/JG̃ ) ⊕ H 2n+2 (S/An ) and
(3) H i (S/JG̃ ∩ An ) = 0 for i 6= n + 2, 2n + 2.
Similarly, if we apply H · (−) to the exact sequence (3) in Corollary 3.3 we get
(4) H n+2(S/JG̃ ∩ Bn+1 ) ∼
= H n+1 (S/(JG̃ , Bn+1 )).
2n
(5) H (S/JG̃ ∩ Bn+1 ) ∼
= H 2n (S/Bn+1 ).
(6) H 2n+2 (S/JG̃ ∩ Bn+1 ) ∼
= H 2n+2 (S/JG̃ ).
i
(7) H (S/JG̃ ∩ Bn+1 ) = 0 for i 6= n + 2, 2n, 2n + 2.
With these results in mind the short exact sequence (1) of Corollary 3.3 provides (by
applying the local cohomology functor) the vanishing H i (S/JG ) = 0 for all i 6= n + 2, n +
3, 2n, 2n + 2. Moreover it induces isomorphisms
H n+2(S/JG ) ∼
= H 2n (S/Bn+1 )
= H n+1 (S/(J , Bn+1 )) and H 2n (S/JG ) ∼
G̃
and as n > 3 so 2n > n + 3 the isomorphism H n+3(S/JG ) ∼
= H n+2 (S/(JG̃ , An )). Moreover
we obtain the following short exact sequence
0 → H 2n+2 (S/JG ) → H 2n+2 (S/JG̃ ) ⊕ H 2n+2 (S/An ) ⊕ H 2n+2 (S/JG̃ ) → H 2n+2 (S/JG̃ ) → 0.
By Local Duality this proves the first three rows in the table of statement (b). By Local
Duality we get also the the following short exact sequence
0 → ω(S/JG̃ ) → ω(S/JG̃ ) ⊕ ω(S/An ) ⊕ ω(S/JG̃ ) → ω(S/JG ) → 0.
Note that we may write ω instead of ω 2n+2 because all modules above are canonical
modules. First of all the short exact sequence provides that depth ω(S/JG ) ≥ 2n + 1. By
applying local cohomology and dualizing again we get the following exact sequence
f
0 → ω2× (S/JG ) → S/JG̃ ⊕ S/An ⊕ S/JG̃ → S/JG̃ → ω 2n+1(ω(S/JG )) → 0.
As in the proof of Theorem 4.1 we see that f is surjective. Therefore ω 2n+1(ω(S/JG )) = 0
and depth ω(S/JG ) = 2n+2. Whence ω(S/JG ) is a (2n+2)-dimensional Cohen-Macaulay
module. Then f is a split surjection and ω2× (S/JG ) ∼
= S/JG̃ ⊕ S/An . This implies the
∼
isomorphism ω(S/JG ) = ω(S/JG̃ ) ⊕ ω(S/An ) and this finishes the proof of (b). Clearly
reg S/JG = 2.
Theorem 4.4. Let m > 3 and n = 2. Then:
(a) reg S/JG = 2.
(b) ω i (S/JG ) = 0 if and only if i 6∈ {4, m + 2, m + 3, 2m} and there are the following
isomorphisms and integers
i
ω i (S/JG )
depth ω i (S/JG ) dim ω i (S/JG )
4
ω 4 ((JG̃ , Bm )/Bm )
4
4
m+1
m+1
m + 2 ω m+1 (S/(JG̃ , A2 ))
m+3
ω(S/JG̃ )
m+3
m+3
2m
S/A2 (−2m)
2m
2m.
Moreover there is an isomorphism ω 4 (ω 4 (S/JG ) ∼
= (J , Bm )/Bm .
G̃
Proof. By applying the local cohomology functors H · (−) to the exact sequence (2) in
Corollary 3.3 we get the following:
10
PETER SCHENZEL, SOHAIL ZAFAR
(1) H m+2 (S/JG̃ ∩ A2 ) ∼
= H m+1 (S/(JG̃ , A2 )),
m+3
∼
(2) H
(S/JG̃ ∩ A2 ) = H m+3 (S/JG̃ ),
(3) H 2m (S/JG̃ ∩ A2 ) ∼
= H 2m (S/A2 ) and
i
(4) H (S/JG̃ ∩ A2 ) = 0 for i 6= m + 2, m + 3, 2m.
Similarly, if we apply H · (−) to the exact sequence (3) in Corollary 3.3 we get the isomorphism H m+3 (S/JG̃ ∩ Bm ) ∼
= H m+3 (S/JG̃ ) and the exact sequence
0 → H 3 (S/(JG̃ , Bm )) → H 4 (S/JG̃ ∩ Bm ) → H 4 (S/Bm ) → 0.
The short exact sequence on local cohomology induces the following exact sequence
0 → ω 4 (S/Bm ) → ω 4 (S/JG̃ ∩ Bm ) → ω 3 (S/(JG̃ , Bm )) → 0
by Local Duality. Now we apply again local cohomology and taking into account that
both ω 4(S/Bm ) and ω 3(S/(JG̃ , Bm )) are Cohen-Macaulay modules of dimension 4 and
3 respectively. Then depth ω 4 (S/JG̃ ∩ Bm )) ≥ 3. By applying local cohomology and
dualizing again it induces the following exact sequence
f
0 → ω 4 (ω 4(S/JG̃ ∩ Bm )) → S/Bm → S/(JG̃ , Bm ) → ω 3 (ω 4(S/JG̃ ∩ Bm )) → 0.
Now the homomorphism f is an epimorphism. ω 3 (ω 4(S/JG̃ ∩ Bm )) = 0. That is
depth ω 4 (S/JG̃ ∩ Bm ) = 4 and it is a Cohen-Macaulay module. Moreover ω 4(ω 4 (S/JG̃ ∩
Bm )) ∼
= (JG̃ , Bm )/Bm . With these results in mind the short exact sequence (1) of Corollary 3.3 provides (by applying the local cohomology functor) the vanishing H i (S/JG ) = 0
for all i 6= 4, m + 2, m + 3, 2m. Moreover it induces isomorphisms
H 4 (S/JG ) ∼
= H m+1 (S/(J , A2 ))
= H 4 (S/J ∩ Bm ), H m+2 (S/JG ) ∼
G̃
G̃
and H 2m (S/JG ) ∼
= H 2m (S/A2 ). Moreover we obtain the following short exact sequence
0 → H m+3 (S/JG ) → H m+3 (S/JG̃ ) ⊕ H m+3 (S/JG̃ ) → H m+3 (S/JG̃ ) → 0.
This implies the isomorphism ω m+3 (S/JG ) ∼
= ω(S/JG̃ ).
As a final step we shall consider the case of the complete bipartite graph Km,n with
2n = m + 2. In all of the previous examples we have the phenomenon that ω i (S/JG ) is
either zero or a Cohen-Macaulay module with i − 1 ≤ dim ω i (S/JG ) ≤ i for all i ∈ Z.
and the canonical module ω(S/JG ) = ω d (S/JG ), d = dim S/JG , is a d-dimensional CohenMacaulay module. For 2n = m + 2 this is no longer true.
Theorem 4.5. Let m + 2 = 2n and m > n + 1. Then:
(a) reg S/JG = 2.
(b) ω i (S/JG ) = 0 if and only if i 6∈ {n + 2, m + 2 = 2n, m + n + 1, 2m} and there are
the following isomorphisms and integers
i
ω i (S/JG )
depth ω i (S/JG ) dim ω i (S/JG )
n+2
ω n+1 (S/(JG̃ , Bm ))
n+1
n+1
m+1
m+2
ω
(S/(JG̃ , An )) ⊕ S/Bm (−2n)
m+1
m+2
m+n+1
ω(S/JG̃ )
m+n+1
m+n+1
2m
S/An (−2m)
2m
2m.
BINOMIAL EDGE IDEAL OF A COMPLETE BIPARTITE GRAPH
11
Proof. It is easily seen that n + 2 < 2n = m + 2 < m + n + 1 < 2m. Then the short
exact sequences of Corollary 3.3 provide that H i (S/JG ) = 0 for all i 6= n + 2, m + 2 =
2n, m + n + 1, 2m. Moreover, it induces the following isomorphisms
(1) H n+2(S/JG ) ∼
= H n+1 (S/(JG̃ , Bm )),
m+2
(2) H
(S/JG ) ∼
= H m+1 (S/(JG̃ , An ) ⊕ H m+2 (S/Bm ),
(3) H m+n+1 (S/JG ) ∼
= H m+n+1 (S/JG̃ ) and
2m
(4) H (S/JG ) ∼
= H 2m (S/An ).
This easily yields the statements in (a) and (b).
The difference of the case handled in Theorem 4.5 is the fact the ω m+2 (S/JG ) is not
a Cohen-Macaulay module. It is the direct sum of two Cohen-Macaulay modules of
dimensions m + 2 and m + 1 respectively. In [10] a finitely generated S-module M is
called canonically Cohen-Macaulay module whenever ω(M) is a Cohen-Macaulay module.
Note that if M is a Cohen-Macaulay module, then it is also a Cohen-Macaulay canonical
module. The converse is not true.
Now we prove some corollaries about the Cohen-Macaulayness and related properties.
Corollary 4.6. Let JG ⊂ S denote the binomial edge ideal of a complete bipartite graph.
(a) S/JG is a Cohen-Macaulay canonical ring and depth ω i (S/JG ) ≥ i − 1 for all
depth S/JG ≤ i ≤ dim S/JG . Moreover S/JG is a Cohen-Macaulay ring if and
only if (m, n) ∈ {(2, 1)(1, 1)}.
(b) S/JG is sequentially Cohen-Macaulay and not Cohen-Macaulay if and only if n = 1
and m > 2 or n = m = 2.
Proof. By view of Theorems 4.1, 4.2, 4.3, 4.4 and 4.5 we get the statements on the
Cohen-Macaulayness of ω(S/JG ) and the estimates of of the depth of ω i (S/JG ) for all
possible bipartite graphs G. By Lemma 3.2 and Corollary 3.4 the claim on the CohenMacaulayness of S/JG is easily seen. Similar arguments work for the sequentially CohenMacaulay property as it is easily seen by the definition.
5. On the purity of the free resolution
In the following let Jr ⊂ S, r ≤ m + n, denote the binomial edge ideal corresponding
to the complete graph on r vertices. As a technical tool for our further investigations we
need the following Lemma.
Lemma 5.1. (a) Let M denote a finitely generated graded S-module. Let f = f1 , . . . , fl
denote an M-regular sequence of forms of degree 1. Then
TorSi (K, M/f M) ∼
= ⊕lj=0 TorSi−j (K, M)(j) (−j).
l
(b) TorSi (K, S/Jr ) ∼
= K bi (r) (−i − 1) for i = 1, . . . , r − 1, where bi (r) = i
r
i+1
.
Proof. For the proof of the statement in (a) let l = 1 and f = f1 . Then the short exact
f
sequence 0 → M(−1) → M → M/f M → 0 provides an isomorphism
TorSi (K, M/f M) ∼
= TorSi (K, M) ⊕ TorSi−1 (K, M)(−1)
12
PETER SCHENZEL, SOHAIL ZAFAR
for all i ∈ Z. By an easy induction argument this yields the isomorphisms in (a). The
statement in (b) is well-known since S/Jr has a linear resolution (see e.g. [3, Exercise
A2.19]).
For a certain technical reason we need the following Lemma that describes the ideals
JG̃ ∩ An respectively JG̃ ∩ Bm as binomial edge ideals.
Lemma 5.2. The ideal JG̃ ∩ An is the binomial edge ideal of the graph G obtained by
deleting all edges {i, j} of the complete graph on [n + m] vertices such that n < i < j ≤
m + n. Similarly JG̃ ∩ Bm is the binomial edge ideal of the graph where all the edges {i, j}
of the complete graph on [n + m] vertices such that 1 ≤ i < j ≤ n are deleted.
Proof. Let us consider the ideal JG̃ ∩ An . Look at the primary decomposition of the graph
G. We have to find all ∅ =
6 T ⊂ [n + m] such that c(T \ {i}) < c(T ) for all i ∈ T . If
T = {1, . . . , n}, then c(T ) = m > 1 and c(T \ {i}) = 1 for all i. Let T ⊂ [n + m] denote a
subset with T 6= {1, . . . , n}. Then it is easy to see that the condition c(T \ {i}) < c(T ) for
all i ∈ T can not be satisfied. So the claim follows by Lemma 2.2. A similar consideration
proves the case of JG̃ ∩ Bm .
As usual we define βi,j (M) = dimK TorSi (K, M)i+j , i, j ∈ Z, the graded Betti numbers
of M, a finitely generated S-module. Then reg M = max{j ∈ Z|βi,j (M) 6= 0}. In the
following we shall prove that S/JG has a pure resolution. Note that all the βi,j (S/JG )
outside of the Betti table are zero.
Theorem 5.3. Let S/JG denote the binomial edge ideal of the complete bipartite graph
Km,n . Then the Betti diagram has the following form
0 1
2
0 1 0
0
1 0 mn 0
2 0 0 β2,2
··· p
··· 0
··· 0
· · · βp,2
where
p=
m,
if n = 1 ;
2m + n − 2, if m ≥ n > 1.
Proof. Because of the regularity and depth of S/JG , the non-vanishing part of the Betti
table is concentrated in the frame of the one given in the statement. Clearly β0,0 = 1 and
βi,0 = 0 for all i > 0. Furthermore β1,0 = β2,0 = 0. Since JG is minimally generated by
mn binomials we get that β1,1 = mn and β1,2 = 0.
In order to prove the statement we have to show that β2,1 = 0 because this implies that
βi,1 = 0 for all i ≥ 2 as a consequence of the minimality of the free resolution. Here we
have two cases:
Case(a): Let m ≥ n > 1. We take the short exact sequence (1) of Corollary 3.3. It
induces a graded homomorphism of degree zero
TorS2 (K, S/JG̃ ) = K b2 (m+n) (−3) → TorS1 (K, S/JG ) = K mn (−2).
Therefore it is the zero homomorphism. On the other side it induces a homomorphism
TorS3 (K, S/JG̃ ) = K b3 (m+n) (−4) → TorS2 (K, S/JG ),
BINOMIAL EDGE IDEAL OF A COMPLETE BIPARTITE GRAPH
13
which is the zero homomorphism when restricted to degree 3 since reg S/JG = 2. Therefore
there is a short exact sequence of K-vector spaces
0 → TorS2 (K, S/JG )3 → TorS2 (K, S/JG̃ ∩ An )3 ⊕ TorS2 (K, S/JG̃ ∩ Bm )3 → K b2 (m+n) → 0.
That is β2,1 (S/JG ) = β2,1 (S/JG̃ ∩ An ) + β2,1 (S/JG̃ ∩ Bm ) − b2 (m + n).
In the next step we shall compute β2,1 (S/JG̃ ∩ An ) and β2,1 (S/JG̃ ∩ Bm ). We start with
the first of them. To this end we use the short exact sequence (2) of Corollary 3.3. At
first we note that β1,2 (S/JG̃ ∩ An ) = 0 which is true since JG̃ ∩ An is minimally generated
by quadrics as follows by Lemma 5.2. Because of
β3,0 (S/JG̃ ∩ An ) = β3,0 (S/JG̃ ) = β2,1 (S/An ) = 0
we get the following exact sequence of K-vector spaces.
0 → TorS3 (K, S/An )3 → TorS3 (K, S/(JG̃ , An ))3 → TorS2 (K, S/JG̃ ∩ An )3 →
TorS2 (K, S/JG̃ )3 → TorS2 (K, S/(JG̃ , An ))3 → 0.
By counting vector space dimensions this provides that
β2,1 (S/JG̃ ∩ An ) = b2 (m + n) + β3,0 (S/(JG̃ , An )) − β2,1 (S/(JG̃ , An )) −
2n
3
.
Since (JG̃ , An ) = (Jm , An ), where Jm = I(n + 1, . . . , n + m) we might use Lemma 5.1 for
(2n)
the calculation of these dimensions. Therefore TorS3 (K, S/(JG̃ , An ))3 ∼
= TorS0 (K, S/Jm )0 3
and
(2n)
TorS2 (K, S/(JG̃ , An ))3 ∼
= TorS2 (K, S/Jm )3 ⊕ TorS1 (K, S/Jm )2 1 .
Therefore β3,0 (S/(JG̃ , An )) = 2n
and β2,1 (S/(JG̃ , An )) = b2 (m) + 2nb1 (m). Putting these
3
integers together it follows that
β2,1 (S/JG̃ ∩ An ) = b2 (m + n) − b2 (m) − 2nb1 (m).
Interchanging the rôles of m and n we derive a corresponding formula for β2,1 (S/JG̃ ∩Bm )),
namely
β2,1 (S/JG̃ ∩ Bm ) = b2 (m + n) − b2 (n) − 2mb1 (n)
Finally we use both expressions in the above formula in order to confirm that β2,1 (S/JG )
vanishes.
Case(b): Let m > n = 1. Because JG̃ ∩ A1 = JG we might use exact sequence (2) of
Corollary 3.3. Since (JG̃ , A1 ) = (I(2, . . . , n + m), A1 ) the statements in Lemma 5.1 imply
that TorS3 (K, S/(JG̃ , A1 ))3 = 0. Whence there is an exact sequence of K-vector spaces
0 → TorS2 (K, S/JG )3 → TorS2 (K, S/JG̃ )3 ⊕ TorS2 (K, S/A1)3 → TorS2 (K, S/(JG̃ , A1 ))3 → 0.
Therefore β2,1 (S/JG ) = b2 (m + 1) − β2,1 (S/(JG̃ , A1 )). Again by the statement of Lemma
5.1 (a) it follows that β2,1 (S/(JG̃ , A1 )) = b2 (m) + 2b1 (m). Finally
β2,1 (S/JG ) = b2 (m + 1) − b2 (m) − 2b1 (m) = 0,
as required.
As a final feature of the investigations we will describe the explicit values of the Betti
numbers β2,i (S/JG ), 2 ≤ i ≤ p, as they are indicated in Theorem 5.3.
Theorem 5.4. Let G = Km,n denote the complete bipartite graph with m ≥ n ≥ 1.
14
PETER SCHENZEL, SOHAIL ZAFAR
(a) The Hilbert function of S/JG is given by
H(S/JG , t) =
1
(1
(1−t)m+n+1
+ (m + n − 1)t) +
− (1−t)1m+1 (1 + (m − 1)t) −
1
(1
(1−t)n+1
1
(1−t)2m
+
1
(1−t)2n
+ (n − 1)t).
(b) For the multiplicity e(S/JG ) it follows
(
1,
if m > n + 1 or n = 1 and m > 2,
e(S/JG ) =
2m, otherwise.
m
(c) Let n = 1. Then βi,2 (S/JG ) = m mi − i+1
− m+1
for all 2 ≤ i ≤ p = m. Let
i+1
m ≥ n > 1. Then
2n
2m
m+2n−1
2m+n−1
βi,2 (S/JG ) = m+n
+
+
+
m
+
n
i+2
i+2
i+2
i+1
i+1
m+2n
2m+n
m+n−1
− i+2 − i+2 − (m + n) i+1
for all 2 ≤ i ≤ p = 2m + n − 2.
Proof. In order to prove (a) we use the short exact sequences of Corollary 3.3. By the
additivity of the Hilbert series we get the following equalities:
H(S/JG , t) = H(S/JG̃ ∩ An , t) + H(S/JG̃ ∩ Bn , t) − H(S/JG̃ , t),
H(S/JG̃ ∩ An , t) = H(S/JG̃ , t) + H(S/An , t) − H(S/(JG̃ An ), t), and
H(S/JG̃ ∩ Bm , t) = H(S/JG̃ , t) + H(S/Bm , t) − H(S/(JG̃ Bm ), t).
Substituting the Hilbert series of the complete graphs S/JG̃ , S/(JG̃ , An ) and S/(JG̃ , Bm )
as well as the Hilbert series of the polynomial rings S/An , S/Bm we get the desired formula
in (a). Then (b) is an easy consequence of (a).
For the proof of (c) we note at first the structure of the finite free resolution of S/JG
0 → S βp (−p − 2) → · · · → S β3 (−5) → S β2 (−4) → S β1 (−2) → S
with β1 = mn βi = βi,2 (S/JG ), 2 ≤ i ≤ p, as shown in Theorem 5.3. By the additivity of
the Hilbert series this provides the following expression
p
X
1
2
H(S/JG , t) =
(−1)i βi ti+2 )
(1
−
β
t
+
1
(1 − t)2m+2n
i=2
(see also [3, Exercise 19.14]). Now we use the expression of the Hilbert series H(S/JG , t)
as shown in (a) and compare it with the one of the minimal free resolution. By some nasty
calculations we derive the formulas for the Betti numbers as given in the statement.
Acknowledgement. The authors are grateful to Prof. Jürgen Herzog for drawing
their attention to binomial edge ideals and suggesting their investigation.
References
[1]
[2]
[3]
M. Brodmann, R. Sharp: Local Cohomology. An Algebraic Introduction with Geometric Applications. Cambr. Stud. in Advanced Math., No. 60. Cambridge University Press, (1998).
W. Bruns, J. Herzog: Cohen-Macaulay Rings, Cambridge University Press, 1993.
D. Eisenbud: Commutative Algebra (with a View Toward Algebraic Geometry). Springer-Verlag,
1995.
BINOMIAL EDGE IDEAL OF A COMPLETE BIPARTITE GRAPH
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
15
V. Ene, J. Herzog and T. Hibi: Cohen Macaulay Binomial edge ideals. Nagoya Math. J. 204
(2011) 57-68.
S. Goto: Approximately Cohen-Macaulay Rings. J. Algebra 76, 214-225 (1982)
J. Herzog, T. Hibi, F. Hreinsdotir, T. Kahle, J, Rauh: Binomial edge ideals and conditional
independence statements. Adv. Appl. Math. 45 (2010) 317-333.
H. Matsumura: Commutative Ring Theory, Cambridge Studies in Advanced Mathematics 8,
Cambridge University Press, Cambridge (1986).
P. Schenzel: On The Use of Local Cohomology in Algebra and Geometry. In: Six Lectures in
Commutative Algebra, Proceed. Summer School on Commutative Algebra at Centre de Recerca
Matemàtica, (Ed.: J. Elias, J. M. Giral, R. M. Miró-Roig, S. Zarzuela), Progr. Math. 166, pp.
241-292, Birkhäuser, 1998.
P. Schenzel: On the dimension filtration and Cohen-Macaulay filtered modules. Van Oystaeyen,
Freddy (ed.), Commutative algebra and algebraic geometry. Proceedings of the Ferrara meeting in
honor of Mario Fiorentini on the occasion of his retirement, Ferrara, Italy. New York, NY: Marcel
Dekker. Lect. Notes Pure Appl. Math. 206, 245-264 (1999).
P. Schenzel: On birational Macaulayfications and Cohen-Macaulay canonical modules. J. Algebra
275 (2004), 751-770.
R. P. Stanley: Combinatorics and commutative algebra. 2nd ed. Progress in Mathematics
(Boston, Mass.). 41. Basel: Birkhuser (1996).
R. H. Villarreal: Monomial Algebras, New York: Marcel Dekker Inc. (2001).
S. Zafar: On approximately Cohen-Macaulay binomial edge ideal, to appear in Bull. Math. Soc.
Sci. Math. Roumanie.
Martin-Luther-Universität Halle-Wittenberg, Institut für Informatik, D — 06 099
Halle (Saale), Germany
E-mail address: [email protected]
Abdus Salam School of Mathematical Sciences, GCU, Lahore Pakistan
E-mail address: [email protected]
| 0math.AC
|
Fractality of Massive Graphs: Scalable Analysis
with Sketch-Based Box-Covering Algorithm∗
Takuya Akiba† , Kenko Nakamura‡ , Taro Takaguchi§
†
§
Preferred Networks, Inc. ‡ Recruit Communications Co., Ltd.
National Institute of Information and Communications Technology
arXiv:1609.07994v1 [cs.DS] 26 Sep 2016
[email protected], [email protected], [email protected]
Abstract—Analysis and modeling of networked objects are
fundamental pieces of modern data mining. Most real-world
networks, from biological to social ones, are known to have
common structural properties. These properties allow us to
model the growth processes of networks and to develop useful
algorithms. One remarkable example is the fractality of networks,
which suggests the self-similar organization of global network
structure. To determine the fractality of a network, we need
to solve the so-called box-covering problem, where preceding
algorithms are not feasible for large-scale networks. The lack
of an efficient algorithm prevents us from investigating the
fractal nature of large-scale networks. To overcome this issue,
we propose a new box-covering algorithm based on recently
emerging sketching techniques. We theoretically show that it
works in near-linear time with a guarantee of solution accuracy.
In experiments, we have confirmed that the algorithm enables us
to study the fractality of million-scale networks for the first time.
We have observed that its outputs are sufficiently accurate and
that its time and space requirements are orders of magnitude
smaller than those of previous algorithms.
I. I NTRODUCTION
Graph representation of real-world systems, such as social relationship, biological reactions, and hyperlink structure,
gives us a strong tool to analyze and control these complex
objects [21]. For the last two decades, we have witnessed
the spark of network science that unveils common structural
properties across a variety of real networks. We can exploit
these frequently observed properties to model the generation
processes of real networked systems [20] and to develop
graph algorithms that are applicable to various objects [19].
A notable example of such properties is the scale-free property [3], [7], which manifests a power-law scaling in the vertex
degree distribution and existence of well-connected vertices
(often called hubs). The scale-free property, existence of hubs
especially, underlies efficient performance of practical graph
algorithms on realistic networks [1], [17].
Although the scale-free property inspires us to design better
network models and algorithms, it is purely based on the
local property of networks, i.e., the vertex degree. Real-world
networks should possess other common properties beyond the
local level. As a remarkable example of such non-local properties, the fractality of complex networks was found in network
science [27], [14]. The fractality of a network suggests that
∗ This work is done while all authors were at National Institute of
Informatics. A shorter version of this paper appeared in the proceedings of
ICDM 2016 [2].
the network shows a self-similar structure; if we replace
groups of adjacent vertices with supervertices, the resultant
network holds a similar structure to the original network
(see Section II-C for its formal definition). The fractality of
networks gives us unique insights into modeling of growth
processes of real-world networks [28]. In addition, fractal and
non-fractal networks, even with the same degree distribution,
indicate striking differences in facility of spreading [25] and
vulnerability against failure [16]. Aside from theoretical studies, the fractality provides us with useful information about
network topology. Examples include the backbone structure of
networks [15] and the hierarchical organization of functional
modules in the Internet [8], metabolic [28] and brain [13]
networks, to name a few.
Determination of the fractality of a network is based on the
so-called box-covering problem [27] (also see Section II-C).
We locally cover a group of adjacent vertices with a box such
that all vertices in a box are within a given distance from
each other, and then we count the number of boxes we use to
cover the whole network. In principle, we have to minimize
the number of boxes that cover the network, which is known
to be an NP-hard problem (see [26] and references therein).
Although different heuristic algorithms are proposed in the
previous work (e.g., [26], [24]), they are still not so efficient
as to be able to process networks with millions of vertices. This
limitation leaves the fractal nature of large-scale networks far
from our understanding.
Contributions: The main contribution of the present study
is to propose a new type of box-covering algorithm that is
much more scalable than previous algorithms. In general,
previous algorithms first explicitly instantiate all boxes and
then reduce the box cover problem to the famous set cover
problem. This approach requires quadratic Θ(n2 ) space for
representing neighbor sets and is obviously infeasible for
large-scale networks with millions of vertices. In contrast, the
central idea underlying the proposed method is to solve the
problem in the sketch space. That is, we do not explicitly
instantiate neighbor sets; instead, we construct and use the
bottom-k min-hash sketch representation [9], [10] of boxes.
Technically, we introduce several new concepts and algorithms. First, to make the sketch-based approach feasible,
we introduce a slightly relaxed problem called (1 − )B OX C OVER. We also define a key subproblem called the
(1 − )-S ET C OVER problem. The proposed box-cover algorithm consists of two parts. First, we generate min-hash
sketches of all boxes to reduce the (1−)-B OX C OVER problem
to the (1 − )-S ET C OVER problem. Our sketch generation
algorithm does not require explicit instantiation of actual boxes
and is efficient in terms of both time and space. Second, we
apply our efficient sketch-space set-cover algorithm to obtain
the final result. Our sketch-space set-cover algorithm is based
on a greedy approach, but is carefully designed with eventdriven data structure operations to achieve near-linear time
complexity.
We theoretically guarantee both the scalability and the
solution quality of the proposed box-cover algorithm. Specifically, for a given trade-off parameter k and radius parameter
`, it works in O((n + m)k log k min {`, log n}) time and
O(nk +m) space. The produced result is a solution of (1−)B OX C OVER within a factor
p 1 + 2 ln n of the optimum for
B OX C OVER for ≥ 2 5(ln n)/k, with a high probability
that asymptotically approaches 1.
In experiments, we have confirmed the practicability of
the proposed method. First, we observed that its outputs are
quite close to those of previous algorithms and are sufficiently
accurate to recognize networks with ground-truth fractality.
Second, the time and space requirements are orders of magnitude smaller than those of previous algorithms, resulting in
the capability of handling large-scale networks with tens of
millions of vertices and edges. Finally, we applied our algorithm to a real-world million-scale network and accomplish its
fractality analysis for the first time.
TABLE I
F REQUENTLY USED NOTATIONS .
Notation
Description
(In
G = (V, E)
n, m
Nδ (v)
{Sp }p∈P
n
k
ri
Se
e S)
e
C(
the context of the box cover problem)
The graph.
The numbers of vertices and edges in G.
The vertices with distance at most δ from v.
(In the context of the set cover problem)
The set family.
The number of elements and collections.
(Bottom-k min-hash sketch)
The trade-off parameter of min-hash sketches.
The rank of an item i.
The min-hash sketch of set S.
The estimated cardinality of set S.
set of items. We first assign a random rank value ri ∼ U (0, 1)
to each item i ∈ X, where U (0, 1) is the uniform distribution
on (0, 1). Let S be a subset of X. For an integer k ≥ 1,
e where
the bottom-k min-hash sketch of S is defined as S,
e
i ∈ S ⇐⇒ ri ≤ k-th {rj | j ∈ S} . In other words, Se is
the set of vertices with the k smallest rank values. We define
Se = S if |S| < k.
For a set S ⊆ X, the threshold rank τ (S) is defined as
follows. If |S| ≥ k, τ (S) = k-th {ri | i ∈ S}. Otherwise,
e Using sketch S,
e
τ (S) = (k − 1)/ |S|. Note that τ (S) = τ (S).
e
e
we estimate the cardinality |S| as C(S) = (k − 1)/τ (S). Its
relative error is theoretically bounded as follows.
Organization: The remainder of this paper is organized as
follows. We describe the definitions and notations in Section II.
In Section III, we present our algorithm for sketch-space
S ET C OVER. We explain our sketch construction algorithm to
complete the proposed method for B OX C OVER in Section IV.
In Section V, we present a few empirical techniques to further
improve the proposed method. We explain the experimental
evaluation of the proposed method in Section VI. We conclude
in Section VII.
Lemma 1 (Bottom-k cardinality estimator [9], [10]): The
e
cardinality estimation C(S)
is an unbiased estimator√of |S|,
and has a coefficient of variation (CV)1 of at most 1/ k − 2.
II. P RELIMINARIES
In addition, our algorithms heavily rely on the mergeability
of min-hash sketches. Suppose S1 , S2 ⊆ X and S3 = S1 ∪ S2 .
Then, since Se3 ⊆ Se1 ∪ Se2 , Se3 can be obtained only from Se1
and Se2 . We denote this procedure as Merge-and-Purify (e.g.,
Se3 = Merge-and-Purify(Se1 , Se2 )).
For simplicity, we assume that ri is unique for i ∈ X,
and sometimes identify i with ri . In particular, we use the
comparison between elements such as i < j for i, j ∈ X,
where we actually compare ri and rj . We also define k-th(S)
as the element with the k-th smallest rank in S ⊆ X.
A. Notations
We focus on networks that are modeled as undirected
unweighted graphs. Let G = (V, E) be a graph, where V
and E are the vertex set and edge set, respectively. We use
n and m to denote |V | and |E|, respectively. For d ≥ 0 and
v ∈ V , we define Nd (v) as the set of vertices with distance at
most d from v. We call Nd (v) the d-neighbor. When d = 1,
we sometimes omit the subscript, i.e., N (v) =SN1 (v). We also
define Nd (S) for a set S ⊆ V as Nd (S) = v∈S Nd (v). In
other words, Nd (S) represents the set of vertices with distance
at most d from at least one vertex in S. The notations we will
frequently use hereafter are summarized in Table I.
B. Bottom-k Min-Hash Sketch
In this subsection, we review the bottom-k min-hash sketch
and its cardinality estimator [9], [10]. Let X denote the ground
The following corollary can be obtained by applying Chernoff bounds [11].
Corollary 2: For > 0 and c > 1, by setting k ≥ (2 +
c)−2 ln |X|, the probability of the estimation having a relative
c
error larger than is at most 1/ |X| .
C. Problem Definition
1) Graph Fractality: The fractality of a network [27] is a
generalization of the fractality of a geometric object in Euclidean space [12]. A standard way to determine the fractality
of a geometric object is to use the so-called box-counting
1 The
CV is the ratio of the standard deviation to the mean.
106
10
∝
∝
5
b ( `)
104
2) Box Cover: As we described in the previous section, the
fractality of graphs is analyzed by solving the box covering
problem. The problem has two slightly different versions: the
diameter version [27] and the radius version [26]. It has been
empirically shown that these two versions yield negligible
difference in the results. In this study, we focus on the radius
version, which is defined as follows.
`−1.6
e−0.3 `
103
102
101
100 0
10
101
102
`
103
104
Fig. 1. Fitting of b(`) to the power-law and exponential functions for a
synthetic fractal network.
method; we tile the object with cubes of a fixed length and
count the number of cubes needed. If the number of cubes
follows a power-law function of the cube length, the object is
said to be fractal. A fractal object holds a self-similar property
so that we observe similar structure in it when we zoom in
and out to it.
The idea of the box counting method is generalized to
analyze the fractality of networks [27]. The box-covering
method for a network works by covering the network with
boxes of finite length `, which refers to a subset of vertices in
which all vertices are within distance `. For example, a box
with ` = 1 is a set of nodes all adjacent to each other. If
the number of boxes of length ` needed to cover the whole
network, denoted by b(`), follows a power-law function of `:
b(`) ∝ `−d , the network is said to be fractal. The exponent d is
called the fractal dimension. As can be noticed, b(`) crucially
depend on how we put the boxes. Theoretically, we have to
put boxes such that b(`) is minimized to assess its precise
scaling. However, this box-covering problem is NP-hard and
that is why we propose our new approximation algorithm of
this problem in the rest of this paper.
After computing b(`) for a network, we want to decide
whether the network is fractal or not. A typical indicator of
a non-fractal network is an exponential form: b(`) ∝ exp(c`),
where c is a constant factor [29]. Therefore, comparison of
the fitting of the obtained b(`) to power-law and exponential functions enables us to determine the fractality of the
network. Figure 1 illustrates the comparison for (3, 3, 7)flower, a network model with ground-truth fractality [23] (see
Section VI-A). Since b(`) is closer to the power-law than to the
exponential function in this case, the fitting procedure correctly
indicates the fractality of this network model.
It should be noted that the fractality of a network suggests
its self-similarity. Let us aggregate the vertices in a box into a
supervertex and then aggregate the edges spanning two boxes
into a superedge. Then we obtain a coarse-grained version
of the original network. If the original network is fractal,
the vertex degree distributions of the original and coarsegrained networks are (statistically) the same [27]. Note that
the fractality and self-similarity of a scale-free network are
not equivalent, and a non-fractal scale-free network can be
self-similar under certain conditions [18].
Problem 1 (B OX C OVER): In the B OX C OVER problem, given
a graph G and a radius limit ` > 0, the objective is to find a
set S ⊆ V of the minimum size such that N` (S) = V .
The size of set S is equal to b(`) discussed in Section II-C1.
In this study, we consider a slightly relaxed variant of the
B OX C OVER problem, named (1−)-B OX C OVER. The (1−)B OX C OVER problem is defined as follows.
Problem 2 ((1 − )-B OX C OVER): In the (1 − )-B OX C OVER
problem, we are given a graph G, a radius limit ` > 0 and
an error tolerance parameter > 0. The objective is to find a
set S ⊆ V of the minimum size such that |N` (S)| ≥ (1 − )n.
3) Set Cover: The B OX C OVER problem is a special case
of the S ET C OVER problem, which is defined as follows.
Problem 3 (S ET C OVER): In the S ET C OVER problem, we are
given a set family {Sp }p∈P . The objective
is to S
find a set
S
R ⊆ P of the minimum size such that p∈R Sp = p∈P Sp .
The proposed box-covering algorithm deals with a slightly
different version of S ET C OVER, named (1 − )-S ET C OVER
with sketched input as a key subproblem, which is defined as
follows.
Problem 4 ((1 − )-S ET C OVER with sketched input): In the
sketched input version of the (1 − )-S ET C OVER problem, we
are given the min-hash sketches of a set family {Sep }p∈P and
an error tolerance parameter > 0. The objective
S is to find
a set R S⊆ P of the minimum size such that | p∈R Sp | ≥
(1 − )| p∈P Sp |.
We first design an efficient approximation algorithm for
(1 − )-S ET C OVER (Section III). We then propose a new boxcovering algorithm using it (Section IV).
III. S ET C OVER IN S KETCH S PACE
In this section, we design an efficient approximation algorithm for the sketched-input version of (1 − )-S ET C OVER
(Problem 4). We call each p ∈ P a collection and i ∈ Sp
an element. Because of the connection to the B OX C OVER
problems, we assume that the numbers of collections and
elements
are equal. We denote them by n, that is,
S
S |P | =
| p∈P Sp | = n. For R ⊆ P , we define SR = p∈R Sp .
e R ) by C(R),
e
Moreover, for simplicity, we denote C(S
which
can be calculated from merged min-hash sketch SeR .
We first explain the basic greedy algorithm that runs in
O(n2 k) time, and then present its theoretical solution guarantee. Finally, we propose an efficient greedy algorithm, which
runs in O(nk log n) time and produces the exact same solution
as the basic algorithm.
Algorithm 1 Select-Greedily-Naive({Sep }p∈P )
At the i-th iteration, there is some collection p such that
1
n − C(Ri ∪ {p}) ≤ 1 − ∗ (n − C(Ri )) ,
(1)
|R |
eR ← {}
1: R ← {} , S
e
2: while R 6= P and C(R)
< (1 − /2)n do
e ∪ {p}) | p ∈ P \ R}
3:
p ← argmax{C(R
4:
R ← R ∪ {p} , SeR ← Merge-and-Purify(SeR , Sep )
5: return R
A. Basic Greedy Algorithm
Our basic greedy algorithm Select-Greedily-Naive is described as Algorithm 1. We start with an empty set R = {}. In
e ∪ {p}) for every p ∈ P \ R,
each iteration, we calculate C(R
and select p that maximizes the estimated cardinality, and add
e
it to R. We repeat this until C(R)
gets at least (1 − /2)n,
and the resulting R is the solution.
e ∪ {p}), together with R, we manage the
To calculate C(R
merged min-hash sketch SeR , so that SeR always corresponds
to the min-hash sketch of SR . To this end, we use the merger
operation of min-hash sketch. Let us assume that the items in
a min-hash sketch are stored in the ascending order of their
ranks. Then, merging two min-hash sketches can be done in
O(k) time like in the merge sort algorithm; we just need to
pick the top-k distinct items with the lowest ranks in the two
min-hash sketches. The complexity analysis of the algorithm
is as follows.
Lemma 3: Algorithm Select-Greedily-Naive
O(n2 k) time and O(nk) space.
runs
in
Proof Sketch. This algorithm always terminates as, even in the
worst case, after the n-th iteration, R gets P . Each iteration
takes O(nk) time, and the number of iterations is at most n.
Therefore, the time complexity is O(n2 k) time.
B. Theoretical Solution Guarantee
We can guarantee the quality of the solution produced by
the above algorithm as follows.
p
Lemma 4: For ≥ 2 5(ln n)/k, algorithm SelectGreedily-Naive produces a solution of (1 − )-S ET C OVER
within a factor 1 + 2 ln n of the optimum for S ET C OVER with
a probability of at least 1 − 1/n.
In other words, with a high probability that asymptotically
approaches 1, R is the solution of (1 − )-S ET C OVER and
|R| ≤ (1 + 2 ln n)|R∗ |, where R is the output of algorithm
Select-Greedily-Naive and R∗ is the optimum solution of
S ET C OVER (with the same set family as the input).
Proof. Let R be the output of algorithm Select-GreedilyNaive, and R∗ be the optimum solution to S ET C OVER. Let
R0 = ∅ and Ri ⊆ P be the currently selected sets after
the i-th iteration
of algorithm Select-Greedily-Naive. Let
S
C(R) = | p∈R Sp |. From Corollary 2, for set R ⊆ P , the
e
probability of C(R)
having a relative error larger than /2 is
3
at most 1/n .
since otherwise there would not be any solution R∗ of that
size to S ET C OVER. During the i-th iteration, there are at
most n new sets to be examined, and thus the union bound
e is at most /2
implies that the relative error between C and C
2
with a probability of at least 1 − 1/n . Therefore, with that
probability,
e i+1 ) ≥ C(R
e i ∪ {p}) ≥ 1 − C(Ri ∪ {p}).
(2)
C(R
2
From Inequalities 1 and 2, with some calculation, we have
1
e
e i) ,
1−
n−C(Ri+1 ) ≤ 1 − ∗
n − C(R
1−
2
|R |
2
with a probability of at least 1 − 1/n2 . As the number of
iterations is at most n, by applying the union bound over all
iterations, we obtain
i
i
1
e
1−
n − C(Ri ) ≤ 1 − ∗
n < e− |R∗ | n,
2
|R |
with a probability of at least 1−1/n. If i is at least 2|R∗ | ln n,
it becomes strictly less than 1/n, which is smaller than the
e Therefore, the number of iterations is at most
resolution of C.
∗
d2|R | ln ne, and thus |R| ≤ d2|R∗ | ln ne ≤ (1 + 2 ln n)|R∗ |.
e
Moreover, C(R) ≥ (1 − /2)C(R)
≥ (1 − )n, and thus R is
the solution to the (1 − )-S ET C OVER problem.
C. Near-Linear Time Greedy Algorithm
Algorithm Select-Greedily-Naive takes quadratic time,
which is unacceptable for large-scale set families. Therefore,
we then design an efficient greedy algorithm Select-GreedilyFast, which produces the exact same output as algorithm
Select-Greedily-Naive but runs in O(nk log n) time. As the
input size is O(nk), this algorithm is near-linear time.
The behavior of Select-Greedily-Fast at a high level is
the same as that of Select-Greedily-Naive. That is, we start
with an empty set R = {}, and, at each iteration, it adds
e to R. The central
p ∈ P \ R with the maximum gain on C
idea underlying the speed-up is to classify the state of each
p ∈ P at each iteration into two types and manage differently
to reduce the reevaluation of the gain. To this end, we closely
look at the relation between sketches Sep and SeR .
Types and Variables: Let us assume that we are in the main
loop of the greedy algorithm. Here, we have a currently
incomplete solution R ⊂ P . Let p ∈ P \ R. We define
that p belongs to type A if the k-th element of SeR∪p is in
Sp , i.e., k-th(SeR∪p ) ∈ Sp . Otherwise, p is type B. Note that
SeR∪p = Merge-and-Purify(SeR , Sep ).
We define ap = |SeR∪p ∩ Sep |, bp = |SeR∪p ∩ SeR |, and
cp = |SeR ∩ Sep |. Please note that, if p is a type-A collection,
e
C(R
∪ {p}) is determined by k-th(SeR∪p ) = ap -th(Sep ).
e
Similarly, if p is a type-B collection, C(R∪{p})
is determined
e
e
by k-th(SR∪p ) = bp -th(SR ).
Algorithm 2 Select-Greedily-Fast({Sep }p∈P )
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
// Initialization
R ← {} , SeR ← {}
QA ← an empty min-queue (key: ranks, value: collections)
QB ← an empty min-queue (key: integers, value: collections)
for j = 1, 2, . . . , k do
T (j) ← a binary search tree (key: ranks, value: collections)
Ii ← {p ∈ P | i ∈ Sep }
for all p ∈ P do
Insert (p, k-th{ri | i ∈ Sep }) to QA and T (1)
ap ← k, bp ← 0
// Main loop
e
while R 6= P and C(R)
< (1 − /2)n do
// Selection
e ∪ {p}) | p is at the top of QA or QB }
p ← argmax{C(R
R ← R ∪ {p}
Remove p from QA , QB and T
0
SeR
← Merge-and-Purify(SeR , Sep )
0
0
∆ ← SeR
\ SeR , SeR ← SeR
// Notifying events of type 3
for all i ∈ ∆ do
for all p ∈ Ii do
if p ∈ QB then
Move p from T (bp ) to T (bp + 1)
else if ap -th(Sep ) = i then
while ap -th(Sep ) = i do ap ← ap − 1
Update p’s key in T (bp + 1) to ap -th(Sep )
Remove p from QA and Insert (p, bp + 1) to QB
else
Move p from T (bp + 1) to T (bp + 2)
bp ← bp + 1
// Notifying events of types 1 and 2-1
for j = 1, 2, . . . , k do
P 0 ← Retrieve those with keys ≥ j-th(SeR ) from T (j)
for all p ∈ P 0 do
r ← j-th(SeR )
Remove p from QA , QB and T (j)
if p ∈ QA then
ap ← ap − 1, bp ← bp + 1
if ap -th(Sep ) ∈ SeR then
bp ← bp − 1, r ← (j − 1)-th(SeR )
while ap -th(Sep ) ∈ SeR do ap ← ap − 1
if ap -th(Sep ) > r then
Insert (p, ap -th(Sep )) to QA and T (bp + 1)
else
Insert (p, bp ) to QB
Insert (p, ap -th(Sep )) to T (bp )
return R
Events to be Captured: Suppose that we have decided to
adopt a new collection and R is about to be updated to R0
(i.e., R0 = R ∪ {p0 } for some collection p0 ∈ P ). Let us first
assume that a single element appeared in the merged sketch,
i.e., SeR0 \ SeR = {i}. Let p ∈ P \ R0 . In the following, we
examine and classify the events where the evaluation of p is
e ∪ {p}) 6= C(R
e 0 ∪ {p}) (types 1 and 2), or
updated, i.e., C(R
cp is updated (type 3).
Type 1: We assume that i 6∈ Sep and p is type A. From the
definition, τ (SeR∪p ) = ap -th(Sep ), and, bp -th(SeR ) ≤ ap -th(Sep )
< (bp + 1)-th(SeR ). Therefore, τ (SeR∪p ) 6= τ (SeR0 ∪p ) if and
only if (bp +1)-th(SeR0 ) 6= (bp +1)-th(SeR ) and (bp +1)-th(SeR0 )
< ap -th(Sep ). We define that a type-1 event happens to p when
this condition holds.
Type 2: Similarly, we assume that i 6∈ Sep and p is
type B. From the definition, τ (SeR∪p ) = bp -th(SeR ) and
ap -th(Sep ) < bp -th(SeR ). Thus τ (SeR∪p ) 6= τ (SeR0 ∪p ), if
and only if bp -th(SeR ) 6= bp -th(SeR0 ). There are two cases:
bp -th(SeR0 ) ≤ ap -th(Sep ) (type 2-1), after which p becomes
type A, or bp -th(SeR0 ) > ap -th(Sep ) (type 2-2), after which p
still belongs to type B.
Type 3: If i ∈ Sep , then cp will be incremented.
The following lemma is the key to the efficiency of our
algorithm.
Lemma 5: For each p ∈ P , throughout the algorithm execution, events of type 1, type 2-1, or type 3 occur at most 3k
times in total.
Proof Sketch. We use the progress indicator Φ = k − ap +
bp + cp . Initially, ap = k and bp = cp = 0; hence Φ = 0. For
each event occurrence, Φ increases by at least one. As ap ≥ 0
and bp , cp ≤ k, Φ ≤ 3k.
Please note that events of type 2-2 are not considered in
the above lemma, and, indeed, they happen Θ(n) times in
the worst case for each collection. Therefore, we design the
algorithm so that we do not need to capture type-2-2 events.
Finding the Maximum Gain: To adopt a new collection in
each iteration, we need to efficiently find the collection that
gives the maximum gain. We clarify the ordering relation in
each type.
e ∪
Type A: Let p, q ∈ P \ R be type-A collections, then C(R
e
e
e
{p}) ≥ C(R ∪ {q}) if and only if ap -th(Sp ) ≤ aq -th(Sq ).
e ∪
Type B: Let p, q ∈ P \ R be type-B collections, then C(R
e ∪ {q}) if and only if bp -th(SeR ) ≤ bq -th(SeR ),
{p}) ≥ C(R
which is equivalent to bp ≤ bq .
Data Structures: We use the following data structures to
notify the collections about an event occurrence.
Type 1: For each type-A collection p, as we observed above,
p wants to be notified about a type-1 event when (bp +
1)-th(SeR ) becomes smaller than ap -th(Sep ). Therefore, for
j = 1, 2, . . . , k, we prepare a binary search tree T (j), where
values are collections and keys are ranks (i.e., collections are
managed in the ascending order of ranks in each tree). For
each type-A collection p, we put p in T (bp + 1) with key
ap -th(Sep ). Then, when j-th(SeR ) is updated to a new value,
from T (j), we retrieve collections with keys larger than or
equal to the new value and notify them about an event.
Type 2-1: Similarly, for each type-B collection p, p wants to
be notified about a type-2-1 event when bp -th(SeR ) becomes
smaller than or equal to ap -th(Sep ). Thus, we store p in T (bp )
and set its key to ap -th(Sep ). Then, when j-th(SeR ) is updated
to a new value, we retrieve those in T (j) with keys larger than
or equal to the new value and notify them about an event.
Type 3: To capture type-3 events, the use of an inverted index
suffices. That is, for each i ∈ X, we precompute Ii = {p ∈
P | i ∈ Sep }. When i comes to SeR , we notify the collections
in Ii .
Moreover, we also need data structures to find the collections with the maximum gain as follows.
Type A: Type-A collections are managed in a minimumoriented priority queue, where the key of a collection p is
ap -th(Sep ).
Type B: Type-B collections are managed in another minimumoriented priority queue, where the key of a collection p is bp .
Overall Set-Cover Algorithm: The overall algorithm of
Select-Greedily-Fast is described as Algorithm 2. In each
iteration, we adopt the new collection with maximum gain,
which can be identified by comparing the top elements of
the two priority queues. Then, we process events to update
variables and data structures. At the beginning of Section III-C,
we assumed that a single element appears in the new sketch.
When more than one elements come to the new sketch, we
basically process each of them separately. See Algorithm 2 for
the details of the update procedure. The algorithm complexity
and solution quality are guaranteed as follows.
Lemma 6: Algorithm Select-Greedily-Fast
O(nk log n) time and O(nk) space.
runs
in
Proof Sketch. Each data structure operation takes O(log n)
time, which, from Lemma 5, happens at most 3k times for
each collection.
Lemma 7: Algorithm Select-Greedily-Fast produces the
same solution as algorithm Select-Greedily-Naive.
Algorithm 3 Build-Sketches(G, `)
1: Xv ← {v} for all v ∈ V .
2: for ` times do
3:
for all v ∈ V in the increasing order of rv do
4:
Av ← {u ∈ V | v is added to Xu in the last iteration}
5:
for all w ∈ N (Av ) do
6:
Xw ← Merge-and-Purify(Xw , {v})
7:
if Xv was not modified for any v ∈ V then
8:
break
9: return {Xv }v∈V
In each iteration, for each vertex, we essentially merge the
sketches of its neighbors into its sketch in a message-passinglike manner. Two speed-up techniques are employed here to
avoid an unnecessary insertion check. For v ∈ V , let Av
be the vertices in whose sketches v is added to in the last
iteration. First, for each v ∈ V , we try to insert v only into
the sketches of the vertices that are neighbors of Av , as v
cannot be inserted into other vertices. Second, we conduct the
procedure above in the increasing order of ranks, since this
decreases the unnecessary insertion. We prove its correctness
and complexity as follows.
Lemma 8: In algorithm Build-Sketches, after the i-th iterei (v) for all v ∈ V .
ation, Xv = N
Proof Sketch. We prove the lemma by mathematical induction
e0 (v), it is true for i = 0. Now we assume
on i. Since {v} = N
it holds for i and prove it also holds for i + S
1. Let B =
{u ∈ V | (v, u) ∈ E}. Since Ni+1 (v) = {v} ∪ u∈B Ni (u),
ei+1 (v) can be obtained by
and {v} ∈ Ni (v) ⊆ Ni+1 (v), N
ei (u) for all u ∈ B ∪ {v}.
merging N
Proof Sketch. Both algorithms choose the collection with the
maximum gain in each iteration.
e` (v) for
Corollary 9: Algorithm Build-Sketches computes N
all v ∈ V .
IV. S KETCH -BASED B OX C OVERING
Lemma 10: Algorithm Build-Sketches runs in O((n + m)k
log k min {`, log n}) expected time and O(nk + m) space.
In this section, we complete our sketch-based box-covering
algorithm for the (1 − )-B OX C OVER problem (Problem 2).
We first propose an efficient algorithm to construct min-hash
sketches representing the `-neighbors, and then present and
analyze the overall box-covering algorithm,
A. Sketch Generation
For v ∈ V , we denote the min-hash sketch of N` (v) as
e` (v). Here, we construct N
e` (v) for all vertices v ∈ V
N
to reduce the (1 − )-B OX C OVER problem to the (1 − )S ET C OVER problem (Problem 4). Our sketch construction
algorithm Build-Sketches is described as Algorithm 3.
It receives a graph G and a radius parameter `. Each vertex
v manages a tentative min-hash sketch Xv . Initially, Xv only
includes the vertex itself, i.e., Xv = {v}, which corresponds to
e0 (v). Then, we repeat the following procedure for ` times so
N
ei (v). This algorithm has
that, after the i-th iteration, Xv = N
a similar flavor to algorithms for approximated neighborhood
functions and all-distances sketches [22], [5], [10].
Proof Sketch. In addition to the graph, the algorithm stores a
sketch of size k for each vertex, and hence it works in O(nk +
m) space. Each insertion trial takes O(log k) time (Line 6).
Therefore, it suffices to prove that the number of traversed
edges is O((n + m)k`) and O((n + m)k log n). The former
bound is easier, since, in each iteration, the number of last
inserted elements in each sketch is at most k, and thus we
traverse each edge at most k times.
For the latter bound, we count the expected number of
vertices that are inserted once into Xv for a vertex v ∈ V .
The vertex that is i-th to arrive at v is inserted into Xv with
a probability of min {1, k/i}, and thus it is at most
n
X
k
min 1,
= k + k(H(n) − H(k)) = O(k log n),
i
i=1
where H(i) is the i-th Harmonic number. Therefore, each edge
is traversed at most O(k log n) times in total.
B. Overall Box-Cover Algorithm
The overall box-covering algorithm Sketch-Box-Cover is
as follows. We first construct the min-hash sketches using algorithm Build-Sketches and then solve the set cover problem
in the sketch space using algorithm Select-Greedily-Fast.
The guarantees on performance and accuracy are immediate
from the previous lemmas and corollaries as below.
Theorem 11 (Scalability guarantee): Algorithm Sketch-BoxCover works in O((n + m)k log k min {`, log n}) time and
O(nk + m) space.
Theorem 12 (Solution accuracy guarantee):
p With a probability of at least 1 − 1/n, for ≥ 2 5(ln n)/k, algorithm Sketch-Box-Cover produces a solution to the (1 − )B OX C OVER problem within a factor 1 + 2 ln n of the optimum
for the B OX C OVER problem.
Assuming k is a constant, the time and space complexities
are near-linear. Similarly, given a constant , the time and
space complexities are still near-linear, since it suffices to set
k = d20−2 ln ne. In practice, as seen in our experiments,
the algorithm produces solutions that are much closer to the
optimum than what is expected from the above approximation
ratio with much smaller k.
V. P RACTICAL I MPROVEMENT
In this section, we propose techniques to improve the
practicality of the proposed method.
Exact Coverage Management: For the termination condition
in the greedy selection algorithm (i.e., Line 12 in Algorithm 2),
when applied to the box cover problem, we propose to use
the exact coverage C(R) instead of the estimated coverage
e
C(R).
This technique makes the results more stable. We can
efficiently manage the exact coverage as follows.
First, we prepare an array δ, and initialize it as δ[v] = ∞
for all v ∈ V . After selecting a vertex v in each iteration, we
conduct a pruned breadth-first search (BFS) from v. Suppose
we are visiting vertex u with distance d in this BFS. If
δ[u] ≤ d, then we prune this BFS, i.e., we do not traverse the
edges from u. Otherwise, we set δ[u] = d and continue the
search. We do not visit vertices with a distance larger than `.
The number of covered vertices is the number of non-infinity
values in array δ. Since the value of δ[u] changes at most `+1
times, each vertex or edge is visited O(`) times. Therefore, the
total time consumption of this process throughout all iterations
is O((n + m)`).
Multi-Pass Execution: On the basis of the above exact coverage management technique, we sometimes detect that, even
e
e )),
while the estimated coverage is saturated (i.e., C(R)
= C(P
the actual coverage is below the specified threshold. In that
case, to choose more vertices, we propose to repeat the
algorithm from sketch construction until the actual coverage
becomes higher than the threshold.
In the i-th pass, we only care for vertices that are not
covered by the previous passes. This can be easily realized
by modifying the algorithm Build-Sketches so that, at Line
1, we set Xv = ∅ for already covered vertices. For accurate
results, node ranks should be reassigned for each pass.
Exact Neighborhood: To further improve the accuracy, we
propose to combine our sketch-based algorithm with a nonsketch-based algorithm. For a very small radius parameter
`, neighborhood N` (v) is sometimes much smaller than k.
Moreover, even for a larger `, when using the above multi-pass
execution technique, the remaining neighbors may become
small in later passes. In these cases, the sketching approach has
little advantage. Therefore, we detect such circumstances and
switch to a non-sketch-based greedy algorithm. Interestingly,
e` (v)| ≤ k, then
this switching can be done seamlessly. If |N
e` (v) = N` (v). Therefore, under such circumstances, the
N
output of algorithm Build-Sketches can be immediately given
to the non-sketch-based greedy algorithm.
The proposed overall procedure is as follows. We specify
a parameter α. We start by constructing the “sketches” with
algorithm Build-Sketches, but, at first, we apply the algorithm
as if k = ∞, i.e., we do not conduct purification on the minhash sketches. During the construction, if the total number of
elements in all min-hash sketches exceeds αnk at some point,
then we conduct purification on all the min-hash sketches,
continue the construction with the actual k value, and pass
the resulting sketches to the sketch-based greedy algorithm.
Otherwise, we apply the non-sketch-based greedy algorithm
to the resulting “sketches,” which are actually exact neighborhood sets. Assuming parameter α is a constant, the total time
and space complexity remain the same.
Exact Box Covering: Together with the preceding three techniques, to further make the results reliable, we propose to use
our algorithm for solving the original B OX C OVER problem
(Problem 1) rather than (1 − )-B OX C OVER problem (Problem 2). In other words, we recommend setting = 0 to ensure
that all vertices are completely covered. As we will see in
the experiments, even with this seemingly extreme threshold,
thanks to the above techniques, both the running time and the
solution quality are reasonable.
VI. E XPERIMENTS
In this section, we present our experimental results to verify
the performance of the algorithm. Specifically, we compare our
algorithm with other preceding algorithms in terms of accuracy
and computation time.
We mainly focus on model networks, instead of empirical
ones, in order to validate the results of our algorithm with
ground-truth theoretical solutions and to investigate the scalability of the algorithm for various sizes of networks. However,
we also demonstrate the practicalness of our algorithm by
applying it to a real million-scale web graph. On the basis
of the result, we reveal the fractality of such large-scale real
graph for the first time.
A. Setup
Environment: Experiments were conducted on a Linux server
with Intel Xeon X5650 (2.67 GHz) and 96GB of main
approximation ratio
(3, 0, 6)-SHM
(3, 3, 7)-flower
(2, 0, 8)-SHM
24 25 26 27 28 29 210
k
1.8
1.6
1.4
1.2
1.0
approximation ratio
1.8
1.6
1.4
1.2
1.0
(3, 0, 6)-SHM
(3, 3, 7)-flower
(2, 0, 8)-SHM
2-3 2-2 2-1 20 21 22 23
Fig. 2. Average approximation ratio to the theoretical solutions for various
k and α.
memory. Algorithms were implemented in C++ and compiled
by gcc 4.8.4 with -O3 option.
Algorithms: For comparison, we used a naive algorithms
named greedy coloring (GC) and three advanced and popular algorithms, named maximum excluded mass burning
(MEMB), minimal value burning (MVB), and compact box
burning (CBB). GC, MEMB, and CBB were introduced in
[26] and MVB was in [24].
Network Models: We used two network models with groundtruth fractality: the (u, v)-flower [23] and the Song-HavlinMakse (SHM) [28] model. These models have power-law degree distributions, the representative characteristic of complex
networks. Both models can be either fractal or non-fractal,
depending on the structural parameter values. We refer to them
as the (u, v, g)-flower and (c, e, g)-SHM model to indicate the
parameter settings. The common parameter g (g = 1, 2, 3, . . . )
determines the network size n: n = (w − 2/w − 1)wg +
w/w − 1, where w ≡ u + v for the (u, v, g)-flower, and
n = (2c+1)g n0 for the (c, e, g)-SHM model. In addition to the
flower and SHM models, we considered the Barabási-Albert
(BA) network model [3] as one of the most famous models of
complex networks. The BA model is not fractal [27]. We refer
to this model as (c, t)-BA, where c is the number of edges that
a new node has and t sets the network size as n = 125 × 2t .
Fractality Decision Procedure: After the computation of the
box-covering algorithms, we determined whether the obtained
b(`) indicates the fractality or not. This task was done by
fitting the b(`) curve with a power-law function (i.e., fractal) and an exponential function (i.e., non-fractal) by using optimize.leastsq function in SciPy package of
Python. We used the parameters estimated by fitting the curves
to linearized models as the initial values for the nonlinear
fitting. The key quantity was the ratio between the residual
error of fitting to a power-law function and that to an exponential function, denoted by rfit . If − log10 rfit is postive (i.e.,
rfit < 1), the network was supposed to be fractal. Otherwise,
it was supposed to be non-fractal. This procedure of fitting
and comparison follows that used in [29].
B. Parameter Settings
First of all, we have to decide the parameter values of our
algorithm: (error tolerance), k (sketch size), and α (exact
neighborhood switch threshold). In principle, the accuracy
of results as well as running time increases with k and α,
and it decreases with . As we discussed in Section V, we
fixed = 0. To choose k and α, we plotted the average
approximation ratio of our results to the theoretical solutions
for several fractal network models as a function of k and α
in Figure 2. The average approximation ratio is defined by
ρ ≡ hbsketch (`)/btheory (`)i` , where h·i` is the average over
`. To compute bsketch (`), we executed the algorithm for ten
times and took the average of the resulting b(`) over the ten
runs.
In the left panel of Figure 2, we varied 24 ≤ k ≤ 210
while fixing α = 1. The ρ values were affected slightly by
k for the SHM models and tended to decrease with k for
the flower network. On the basis of the results, we decided
to use k = 27 throughout the following experiments. In the
right panel of Figure 2, we varied 2−3 ≤ α ≤ 23 while fixing
k = 27 . The ρ values were almost constant regardless of the
α values for all of the three networks considered. Therefore,
taking into account the running time, we decided to use α = 1
throughout the following experiments. It is worth noting that
Figure 2 also demonstrates the high accuracy and robustness
of our algorithm over a broad range of parameter values.
C. Accuracy and Scalability
Table II summarizes the main results of this paper and shows
the comparison of our algorithm (Sketch) with other preceding
algorithms for fractal and non-fractal network models with
various sizes. We evaluated the performance of algorithms by
two measures. The first was the accuracy given by − log10 rfit
(Section VI-A). If this measure took a positive (negative) value
for a fractal (non-fractal) network, the algorithm correctly
distinguished the fractality of the network. The second was
computation time in seconds.
Discrimination Ability: As we can see in Table II, the sketch
algorithm perfectly distinguishes between the fractal and nonfractal networks as the other algorithms do (except for CBB for
(3, 4, 5)-flower). The proposed algorithm shows its advantage
in computation time: the algorithm is generally faster than
other algorithms and is able to handle large networks that
other algorithms do not terminate. Although MEMB is faster
than Sketch for some relatively small network models, this
result is expected because actual neighborhood sets are not
significantly larger than sketch sizes in these networks. As
a summary, (i) the sketch algorithm correctly detected the
fractality of network models with around ten times smaller
computation time than the fastest previous algorithm. In addition, (ii) the algorithm was able to deal with networks with
millions of nodes with acceptable computation time (within
1 day), whereas other algorithms could not in our machine
environment.
Time and Memory Consumption: The proposed algorithm
is scalable for not only for computation time but also for
memory usage. In Figure 3, computation time (seconds) and
memory usage (KB) of the five algorithms were plotted as
a function of the number of vertices. We use (2, 2, g)-flower
(3 ≤ g ≤ 11) and (2, t)-BA (0 ≤ t ≤ 15) networks as the
example of a fractal and a non-fractal network, respectively.
TABLE II
RUNNING TIME IN SECONDS (Time) AND THE RELATIVE ERROR RATIO OF A POWER - LAW FUNCTION , − log10 rfit (Fit). DNF MEANS THAT IT DID NOT
FINISH IN ONE DAY OR RAN OUT OF MEMORY.
Graph
|V |
Model
|E|
Sketch
Time
Fit
MEMB [26]
Time
Fit
GC [26]
Time
Fit
MVB [24]
Time
Fit
CBB [26]
Time
Fit
5 Networks with ground-truth fractality (“Fit” values are expected to be positive.)
(2,
(2,
(2,
(2,
(2,
(2,
(2,
(2,
(2,
(3,
(3,
(3,
(3,
(2,
(2,
(2,
(3,
2,
2,
2,
2,
3,
3,
3,
4,
4,
3,
3,
4,
4,
0,
0,
0,
0,
4)-flower
7)-flower
10)-flower
11)-flower
6)-flower
7)-flower
8)-flower
6)-flower
7)-flower
6)-flower
7)-flower
5)-flower
7)-flower
6)-SHM
7)-SHM
8)-SHM
6)-SHM
172
10,924
699,052
2,796,204
11,720
58,595
292,970
37,326
223,950
37,326
223,950
14,007
686,287
12,501
62,501
312,501
67,229
256
16,384
1,048,576
4,194,304
15,625
78,125
390,625
46,656
279,936
46,656
279,936
16,807
823,543
12,500
62,500
312,500
67,228
0
15
8,628
62,138
26
286
2,913
138
1,526
148
1,779
34
8,873
33
224
2,728
207
0.8
2.5
3.5
4.0
1.2
1.1
1.0
1.1
1.0
1.1
1.2
0.7
0.8
1.2
1.2
1.1
1.0
0
10
DNF
DNF
14
377
DNF
121
DNF
101
DNF
16
DNF
8
206
DNF
190
1.0
3.4
—
—
1.1
1.0
—
1.0
—
1.3
—
0.9
—
1.1
1.1
—
0.9
0
228
DNF
DNF
146
8,538
DNF
2,422
DNF
10,751
DNF
560
DNF
872
48,116
DNF
21,726
28.7
27.7
—
—
0.1
0.1
—
1.7
—
2.1
—
0.1
—
1.1
1.1
—
0.9
199
DNF
DNF
DNF
DNF
DNF
DNF
DNF
DNF
DNF
DNF
DNF
DNF
32
1,126
DNF
628
1.0
—
—
—
—
—
—
—
—
—
—
—
—
1.1
1.1
—
0.9
0
122
DNF
DNF
5,593
DNF
DNF
2,559
DNF
1,284
61,562
3,380
DNF
325
6,579
DNF
4,623
28.0
27.4
—
—
0.5
—
—
0.6
—
1.4
1.5
-0.4
—
0.7
0.9
—
0.9
-5.4
-6.2
-7.0
-4.8
-6.0
—
-1.8
-1.8
-1.9
-0.4
-0.4
-0.3
-0.3
-0.6
-0.6
-0.6
-0.6
—
—
364
3,610
DNF
DNF
DNF
DNF
DNF
DNF
DNF
126
8,615
25
2,070
54
DNF
DNF
DNF
DNF
DNF
-2.2
-2.6
—
—
—
—
—
—
—
-3.5
-4.9
-2.6
-4.2
-0.5
—
—
—
—
—
21,833
DNF
DNF
1,862
61,953
DNF
3,781
DNF
DNF
7,129
DNF
1,224
DNF
0
404
DNF
DNF
DNF
DNF
-2.6
—
—
-3.3
-4.4
—
-1.9
—
—
-2.5
—
-2.9
—
-0.3
-0.1
—
—
—
—
5 Networks with ground-truth non-fractality (“Fit” values are expected to be negative.)
102
102
Sketch
MEMB
CBB
GC
MVB
100
10-4 2
10
108
107
106
105
104
103 2
10
time (sec)
104
104
103
number of vertices
Sketch
MEMB
CBB
GC
MVB
104
103
number of vertices
100
-2
105
108
466
1,774
20
123
4,195
23
223
1,678
31
390
12
210
0
1
17
377
6,474
36,125
-2.9
-3.8
-4.6
-3.0
-4.7
-6.0
-1.0
-0.8
-0.7
-3.5
-4.9
-2.6
-4.1
-0.9
-2.7
-1.3
-1.5
-1.4
-1.4
Sketch
MEMB
CBB
GC
MVB
10
102 103 104 105 106 107
number of vertices
108
Sketch
MEMB
107
CBB
106
GC
MVB
5
10
104
103 2
10 103 104 105 106 107
number of vertices
105
Fig. 3. Scalability of computation time (top) and memory usage (bottom) for
(2, 2, g)-flower (left) and (2, t)-BA (right) networks.
The symbols corresponding to an algorithm were not shown
if the algorithm did not stop within 24 hours or could not
197
1,641
DNF
16
280
DNF
20
548
DNF
32
1,397
8
580
0
0
76
3,535
DNF
DNF
-2.9
-3.8
—
-2.7
-3.2
—
-0.6
-0.7
—
-3.5
-4.9
-2.6
-4.2
-0.9
-2.0
-1.3
-1.5
—
—
105
104
103
102
101
100 0
10
286
2,999
38,278
44
826
DNF
53
1,598
67,866
433
17,703
97
9,504
0
2
154
12,457
DNF
DNF
theory
0.20
0.15
0.10
CV
59,049
177,147
531,441
16,384
65,536
1,048,576
15,625
78,125
390,625
31,104
186,624
16,384
131,072
497
3,997
31,997
255,997
2,047,997
8,191,997
104
10-2
memory (KB)
29,526
88,575
265,722
10,924
43,692
699,052
11,720
58,595
292,970
24,885
149,301
14,045
112,349
250
2,000
16,000
128,000
1,024,000
4,096,000
b ( `)
2, 10)-flower
2, 11)-flower
2, 12)-flower
3, 7)-flower
3, 8)-flower
3, 9)-flower
4, 6)-flower
4, 7)-flower
4, 8)-flower
1, 6)-SHM
1, 7)-SHM
1, 5)-SHM
1, 6)-SHM
1)-BA
4)-BA
7)-BA
10)-BA
13)-BA
15)-BA
memory (KB)
time (sec)
(1,
(1,
(1,
(1,
(1,
(1,
(1,
(1,
(1,
(2,
(2,
(3,
(3,
(2,
(2,
(2,
(2,
(2,
(2,
0.05
101
102
`
103
104
0.00
100
101
102
`
103
104
Fig. 4. Results of different runs for (3, 3, 7)-flower. (Left) b(`) and (Right)
CV of them as a function of `.
execute owing to memory shortage. The performance of the
proposed algorithm is comparable to or worse than some other
algorithms when the network is relatively small (i.e., n < 104 ).
However, the algorithm is orders of magnitude faster than
other algorithms for large networks. Also, it achieves such
high a high speed with incomparably smaller memory usage
than MEMB, the second fastest algorithm.
Robustness over Randomness: The sketch algorithm accurately recovers b(`) of theoretical prediction for fractal network
b ( `)
106
10
5
10
4
∝
∝
VII. C ONCLUSIONS
`−3.4
e−0.3 `
103
102
101
100 0
10
101
`
102
Fig. 5. Results for a real web graph.
models, and the results are robust over different execution runs.
The left panel of Figure 4 shows b(`) of ten different runs
of the proposed algorithm on (3, 3, 7)-flower. The b(`) values
follow well the theoretical solution, which is indicated by the
solid line. As we can clearly observe, the fluctuation in the b(`)
values due to the randomness is very small. The consistency
over different runs is captured by the CV of b(`) (i.e., the ratio
of the standard deviation of b(`) to its average over ten runs)
as a function of ` (right panel of Figure 4). The CV values
tend to increase with `. This tendency can be explained by
the following two factors. First, the b(`) value takes a positive
integer value and monotonically decreases with ` by definition.
Thus, even a change of ±1 in b(`) might cause a large CV
value if ` is large. Second, our algorithm intrinsically fluctuates
more when ` is larger. This could be because the sizes of the
solutions become smaller for larger `, and hence the algorithm
gets a little more sensitive to estimation errors. Nevertheless,
it should be noted that the variance of our algorithm was
considerably small even for large ` (i.e., CV ∼ 0.19 at most).
This magnitude of variance would have little impact on the
estimation of fractality.
D. Application to Real Large Network
In closing this section, we applied the sketch algorithm to a
large-scale real graph to show the scalability of the proposed
algorithm with an empirical instance. The results also gave
us some insight on the fractality of large-scale real-world
networks, which is beyond the reach of previous algorithms.
As a representative instance of a real-world large graph, we
considered the in-2004 network [6], [4], which is a crawled
web graph of 1, 382, 908 vertices and 16, 917, 053 edges. We
discarded the direction of the edges (i.e., hyperlinks) to make
the network undirected. The algorithm took 11.7 hours in total.
The resulting b(`) of the sketch algorithm and the fitting
curves are shown in Figure 5. We omitted the three points
with the smallest ` values from the fitting because empirical
networks would not show a perfect fractality, contrary to welldesigned network models. A large part of the points fall on the
line of the fitted power-law function, and indeed, our fractality
decision procedure yielded − log10 rfit = 0.79, which suggests
the fractality of the in-2004 network. It is worth mentioning
that the fractality of this network was unveiled for the first time
for the sake of our algorithm.
Fractality is an interesting property that appears in some
classes of real networks. In the present study, we designed a
new box-covering algorithm, which is useful for analyzing the
fractality of large-scale networks. In theory, we have shown
desirable guarantees on scalability and solution quality. In the
experiments, we confirmed that the algorithm’s outputs are
sufficiently accurate and that it can handle large networks with
millions of vertices and edges. We hope that our method enables further exploration of graph fractality and its applications
such as graph coarsening.
Repeatability: Our implementation of the proposed and previous box-cover algorithms is available at http://git.io/fractality.
It also contains the generators of the synthetic network models,
and thus the results in this paper can be perfectly replicated.
We hope that our public software will enable further exploration of graph fractality and its applications.
Acknowledgment: This work was supported by JSPS KAKENHI (No. 15H06828), JST, ERATO, Kawarabayashi Large
Graph Project, and JST, PRESTO. Web graph data was downloaded from http://law.di.unimi.it/datasets.php. T.T. thanks to
K. Takemoto for valuable discussions.
R EFERENCES
[1] T. Akiba, Y. Iwata, and Y. Yoshida. Fast exact shortest-path distance
queries on large networks by pruned landmark labeling. In SIGMOD,
page 349, 2013.
[2] T. Akiba, K. Nakamura, and T. Takaguchi. Fractality of massive graphs:
Scalable analysis with sketch-based box-covering algorithm. In ICDM,
2016. to appear.
[3] A.-L. Barabási and R. Albert. Emergence of Scaling in Random
Networks. Science, 286(5439):509–512, 1999.
[4] P. Boldi, M. Rosa, M. Santini, and S. Vigna. Layered label propagation: A multiresolution coordinate-free ordering for compressing social
networks. In WWW, pages 587–596, 2011.
[5] P. Boldi, M. Rosa, and S. Vigna. HyperANF: Approximating the
neighbourhood function of very large graphs on a budget. In WWW,
pages 625–634, 2011.
[6] P. Boldi and S. Vigna. The WebGraph framework I: Compression
techniques. In WWW, pages 595–601, 2004.
[7] G. Caldarelli. Scale-Free Networks. Oxford University Press, 2007.
[8] S. Carmi, S. Havlin, S. Kirkpatrick, Y. Shavitt, and E. Shir. A model
of Internet topology using k-shell decomposition. Proc. Natl. Acad. Sci.
USA, 104(27):11150–11154, 2007.
[9] E. Cohen. Size-estimation framework with applications to transitive
closure and reachability. J. Comput. Syst. Sci., 55(3):441–453, 1997.
[10] E. Cohen. All-distances sketches, revisited: HIP estimators for massive
graphs analysis. IEEE TKDE, 27(9):2320–2334, 2015.
[11] E. Cohen, D. Delling, T. Pajor, and R. F. Werneck. Sketch-based
influence maximization and computation: Scaling up with guarantees.
In CIKM, pages 629–638, 2014.
[12] K. Falconer. Fractal Geometry: Mathematical Foundations and Applications, Second Edition. Wiley-Blackwell, 2003.
[13] L. K. Gallos, H. A. Makse, and M. Sigman. A small world of weak ties
provides optimal global integration of self-similar modules in functional
brain networks. Proc. Natl. Acad. Sci. USA, 109(8):2825–2830, 2012.
[14] L. K. Gallos, C. Song, and H. A. Makse. A review of fractality and
self-similarity in complex networks. Physica A, 386:686–691, 2007.
[15] K.-I. Goh, G. Salvi, B. Kahng, and D. Kim. Skeleton and Fractal Scaling
in Complex Networks. Phys. Rev. Lett., 96(1):018701, 2006.
[16] T. Hasegawa and K. Nemoto. Hierarchical scale-free network is fragile
against random failure. Phys. Rev. E, 88(6):062807, 2013.
[17] D. Kempe, J. Kleinberg, and É. Tardos. Maximizing the spread of
influence through a social network. In KDD, pages 137–146, 2003.
[18] J. S. Kim, K.-I. Goh, B. Kahng, and D. Kim. Fractality and selfsimilarity in scale-free networks. New J. Phys., 9(6):177, 2007.
[19] J. Kleinberg. The small-world phenomenon. In STOC, pages 163–170,
2000.
[20] J. Leskovec, J. Kleinberg, and C. Faloutsos. Graphs over Time :
Densification Laws, Shrinking Diameters and Possible Explanations. In
KDD, pages 177–187, 2005.
[21] M. E. J. Newman. Networks: an Introduction. Oxford University Press,
2010.
[22] C. R. Palmer, P. B. Gibbons, and C. Faloutsos. ANF: A fast and scalable
tool for data mining in massive graphs. In KDD, pages 81–90, 2002.
[23] H. D. Rozenfeld, S. Havlin, and D. Ben-Avraham. Fractal and transfractal recursive scale-free nets. New J. Phys., 9(6):175, 2007.
[24] C. M. Schneider, T. A. Kesselring, J. S. Andrade, and H. J. Herrmann.
Box-covering algorithm for fractal dimension of complex networks.
Phys. Rev. E, 86(1):016707, 2012.
[25] M. Á. Serrano, D. Krioukov, and M. Boguñá. Percolation in Self-Similar
Networks. Phys. Rev. Lett., 106(4):048701, 2011.
[26] C. Song, L. K. Gallos, S. Havlin, and H. A. Makse. How to calculate
the fractal dimension of a complex network: the box covering algorithm.
J. Stat. Mech., 2007(03):P03006, 2007.
[27] C. Song, S. Havlin, and H. A. Makse. Self-similarity of complex
networks. Nature, 433(7024):392–395, 2005.
[28] C. Song, S. Havlin, and H. A. Makse. Origins of fractality in the growth
of complex networks. Nat. Phys., 2(4):275–281, 2006.
[29] K. Takemoto. Metabolic networks are almost nonfractal: A comprehensive evaluation. Phys. Rev. E, 90(2):022802, 2014.
| 8cs.DS
|
Geometric Controllability of The Purcell’s
Swimmer and its Symmetrized Cousin
Sudin Kadam ∗ Ravi N. Banavar ∗
arXiv:1611.06249v1 [cs.SY] 17 Nov 2016
∗
Systems and Control Engineering Department, Indian Institute of
Technology Bombay, Mumbai, India 400076 (e-mail:
[email protected], [email protected]).
Abstract: We analyse weak and strong controllability notions for the locomotion of the 3-link
Purcell’s swimmer, the simplest possible swimmer at low Reynolds number from a geometric
framework. After revisiting a purely kinematic form of the equations, we apply an extension of
Chow’s theorem to analyze controllability in the strong and weak sense. Further, the connection
form for the symmetric version of the Purcells’ swimmer is derived, based on which, the
controllability analysis utilizing the Abelian nature of the structure group is presented. The
novelty in our approach is the usage of geometry and the principal fiber bundle structure of the
configuration manifold of the system to arrive at strong and weak controllability notions.
Keywords: Purcell’s swimmer, Locomotion, Principal Fiber Bundle, Nonlinear Controllability.
1. INTRODUCTION
Swimming at micro scales is a topic of growing interest. A
vast majority of living organisms are found to perform motion at microscopic scales. There has been a lot of research
and a growing interest in exploring new and efficient ways
to generate propulsion at these scales, see [Becker et al.
(2003)], [Dreyfus et al. (2005)]. A better understanding of
the mechanism of swimming can lead to many useful applications in several fields such as medicine, micro-machining
or micro and nano technology. Microbial motion occurs in
a fluid medium with very low Reynolds number, which
is the ratio of the inertial to viscous forces acting on the
swimmer’s body. The Reynold’s number in such regimes is
of the order of 10−4 [Najafi and Golestanian (2004)]. The
environmental interactions experienced by such microorganisms are essentially different from those experienced
by larger animals, which have prominent inertial effects.
On the contrary, the viscous forces strongly dominate the
motion at low Reynold’s number.
E. M. Purcell, in his lecture on Life at Low Reynolds
Number [Purcell (1977)], presented a three-link swimmer
that can propel itself at low Reynolds numbers. This
swimmer can be considered as a simplified flagellum made
of three slender rods articulated at two hinges. This
proposition gave rise to a lot of research in modelling,
control, optimal gait design etc. of this swimmer, see
[Tam and Hosoi (2007)], [Passov and Or (2012)], [Burton
(2013)], [Avron and Raz (2008)], [Melli et al. (2006)], and
the references therein.
[Bloch (2003)], [Holm et al. (2009)], [Ostrowski and Burdick (1998)] indicate that geometric mechanics and control
theory play a crucial role in the analysis of robotic and animal locomotion . For a large class of locomotion systems,
including underwater vehicles, fishlike swimming, flapping
winged vehicles, spacecraft with rotors and wheeled or
legged robots, it is possible to model the motion using the
mathematical structure of a connection on a principal bundle, see [Cabrera (2008)], [Kelly et al. (2012)], [Ostrowski
and Burdick (1998)].
Although it seems relevant to apply geometric mechanics
and control theoretic concepts to the Purcell’s swimmer,
the literature shows few geometric-oriented approaches.
One of the recent works on the Purcell’s swimmer by
[Hatton and Choset (2013)] studies the problem in such
a geometric framework by introducing the connection
form for modelling it as a purely kinematic system. We
shall base our work on this model for the basic Purcell’s
swimmer, and will proceed to controllability analysis in
the geometric framework. Furthermore, we derive the
connection form followed by the controllability analysis
for the symmetrized version of the swimmer, also referred
to as the symmetrized cousin of the Purcell’s swimmer in
[Avron and Raz (2008)].
1.1 Contribution:
The main contributions of this paper consist of highlighting the particular geometry of a trivial principal fiber
bundle that the configuration space of this swimmer follows. [Giraldi et al. (2013)] analyzes the controlability of
the swimmer using Chow’s theorem by modelling it as a
control affine, driftless, non-linear system. In our work,
the recognition of the geometric structure is utilized to
bring the controllability results in the light of strong and
weak notions, which gives a more complete description to
controllability results for a locomotion problem. Moreover,
the derivation of the connection form for the symmetrized
version of the Purcell’s swimmer, followed by identification
of its structure group’s Abelian nature to generate the
local controllability results are the novel contributions this
paper achieves.
1.2 Organization of paper:
In the next section we show the geometry that typical
locomotion systems have, followed by the example of the
Purcell’s swimmer. In section 3, we review the kinematic
model of this swimmer and define the ideas of strong
and weak controllability. In section 4, the expression for
the connection form for the symmetrized version of the
Purcell’s swimmer is derived. We then identify 3 sets on it’s
base space at which system shows different controllability
properties.
2. GEOMETRY OF CONFIGURATION SPACE OF
THE PURCELL’S SWIMMER
While studying problem of locomotion using internal
shape change, the geometry of the configuration space,
which is generally a differential manifold requires attention
for an elegant and insightful solutions. The configuration
space is written as the product of two manifolds, either
locally or globally. One part, the base manifold M , describes the configuration of the internal shape variables of
the mechanism. The other part depicts the macro-position
of the locomoting body, a Lie group G, representing displacement of the body coordinate frame with respect to the
reference frame. The total configuration space of the robot
Q is defined by both G and M . Such systems follow the
topology of a trivial principal fiber bundle, see [Kobayashi
and Nomizu (1963)]. Figure 1 shows a schematic of a fiber
bundle. With such a separation of the configruation space,
locomotion is readily seen as the means by which changes
in shape affect the macro position. We refer to [Bloch
et al. (1996)], [Kelly and Murray (1995)] for more on the
topology of locomoting systems.
Fig. 2. Purcell’s swimmer
of mechanism is parametrized by the two joint angles
(α1 , α2 ) ∈ S1 × S1 . Hence, the configuration space is
Q = SE(2) × S1 × S1
(1)
Definition : For Q a configuration manifold and G a Lie
group, a trivial principal fiber bundle with base M and
structure group G is a manifold Q = M × G with a free
left action of G on Q given by left translation in the group
variable: φh (x, g) = (x, hg) for x ∈ M and g ∈ G. [Kelly
and Murray (1995)]
The structure group in our case is G = SE(2), which is
a matrix Lie group. The shape space M is parametrized
by (α1 , α2 ). Since all the points q ∈ Q are represented by
(α1 , α2 , g) with g ∈ G, Q has global product structure of
the form M × G. Moreover, SE(2) acts via left action as
a matrix multiplication, and has a single identity element,
which is a 3×3 identity matrix. Hence left action of group,
defined by Φh : (x, g) ∈ Q −→ (x, hg) is free, for x ∈ M
and h, g ∈ G.
Thus, according to the definition, the configuration space
of the basic Purcell’s swimmer satisfies the trivial principal
fiber bundle structure.
3. BASIC PURCELL’S SWIMMER
3.1 Reconstruction equation and local connection form
Fig. 1. Fiber Bundle [Rowland (2016)]
The Purcell’s swimmer is a 3 link mechanism moving in a
fluid with low Reynold’s number. Each link of the swimmer
is modelled as a rigid slender body of length 2L. The
original form of Purcell’s swimmer, shown in fig. 2 has
three links always in a common plane, the outer 2 links
are actuated through respective rotary joints with base
link. It is to be noted that we follow the notations similar
to those in [Hatton and Choset (2013)].
We represent the orientation of the outer links with respect
to the base link through shape variables. The position of
the three-link system in Fig. 2 is defined by location of
the midpoint of the base link and its orientation. This is
represented by g, which belongs to the Special Eucledean
group SE(2), parametrized by (x, y, θ). The shape space
In this section we briefly revisit the model of the Purcell’s
swimmer explained in [Hatton and Choset (2013)]. It uses
Resistive Force Theory, according to which hydrodynamic
forces exerted on the swimmer can be approximated with
local drag forces, which depend linearly on the velocity of
each point [Friedrich et al. (2010)]. The resulting model is
in a purely kinematic form, equation (2). See [Shammas
et al. (2007)] for details on purely kinematic systems.
ξ = −A(x)ẋ
(2)
where A(x) is the local connection form defined at each
x ∈ M . For the Purcell’s swimmer local connection form
A(x) is a 3×3 matrix which appears in the form of ω1−1 ω2 .
The matrices ω1 and ω2 are of size 3 × 3 and 3 × 2,
respectively, and they depend on the lengths of the limbs,
viscous drag coefficient k and the shape of mechanism
(α1 , α2 ). We refer to the [Hatton and Choset (2013)] for
their explicit form. We recall that for the shape manifold
M , its tangent space at a point x ∈ M is denoted by Tx M ,
and the shape velocity ẋ = (α̇1 , α̇2 ) ∈ Tx M . The local
connection form is thus defined as A(x) : Tx M −→ ξ.
In our example the Lie group is the Special Eucledean
group SE(2), and ξ = [ξx , ξy , ξθ ]T belongs to its tangent
space at the identity, with ξx , ξy being the translational
velocity of the base link and ξθ is its rotational component.
The connection form and the other notions mentioned
here have roots in geometric mechanics, see [Bloch (2003)],
[Holm et al. (2009)] for details.
3.2 Controllabitlity analysis
We now proceed to analyze the controllability of the
Purcell’s swimmer in geometric setting. Since control in
our case is the shape velocity itself, (2) can be written in
driftless control affine form as
ẋ
I
=
u
ξ
A(x)
(3)
where, the negative sign in equation (2) is absorbed in
A(x), I is a 2×2 identity matrix and u = [α̇1 , α̇2 ]T ∈ Tx M
is the control input.
3.3 Strong and weak controllability
Since our configuration space is naturally split into shape
and structure group, we write a point in configuration
space as q = (x, g) ∈ M × G = Q. We recall that for
a curve x(t) ∈ M , the horizontal lift x∗ (t) ∈ Q is a curve
which projects to x(t) under the projection map defining
principal fiber bundle and components of its tangent
vectors ẋ∗ (t) ∈ Tq Q satisfy the reconstruction equation
(2). We refer to [Kelly and Murray (1995)] for more details
on this and also for the following 2 controllability notions.
• A locomotion system is said to be strongly controllable
if, for any initial q0 = (x0 , g0 ) and final qf = (xf , gf ),
there exists a time T > 0 and a curve passing through
q0 satisfying x∗ (0) = q0 and x∗ (T ) = qf .
• A locomotion system is said to be weakly controllable
if, for any initial position g0 ∈ G, and final position
gf ∈ G, and initial shape x0 ∈ M , there exists a time
T > 0 and a base space curve x(t) satisfying x(0) =
x0 such that the horizontal lift of x(t) passing through
(x0 , g0 ) satisfies x∗ (0) = q0 and x∗ (T ) = (x(T ), gf ).
As mentioned in the introduction, [Giraldi et al. (2013)]
analyzes controllability of the Purcell’s swimmer by applying Chow’s theorem (see [Bullo and Lewis (2004)] to
the system in equation [3] by treating it as a driftless,
control-affine system. It does so by checking span of the
space formed by the successive Lie brackets of the vector
fields corresponding to the two control inputs α̇1 and
α̇2 . In our approach, we rather utilize the principal fiber
bundle structure, which gives rise to these strong and weak
controllability notions. These define controllability ideas
in more detail, and are of practical relevance since many
times just reaching the desired group component without
strict requirement on shape of the system is sufficient. We
define following vector spaces, referred from [Kelly and
Murray (1995)].
h1 = span{A(X) : X ∈ Tx M },
h2 = span{DA(X, Y ) : X, Y ∈ Tx M },
h3 = span{LZ DA(X, Y ) − [A(Z), DA(X, Y )],
[DA(X, Y ), DA(W, Z)] : W, X, Y, Z ∈ T M }
..
.
hk = span{LX ξ − [A(X), ξ], [η, ξ] : X ∈ Tx M, ξ ∈ hk−1 ,
η ∈ h2 ⊕ · · · ⊕ hk−1 }
Then a system defined on a trivial principal bundle Q is
locally weakly controllable near q ∈ Q if and only if the
space of Lie algebra (g) of structure group is spanned by
the vector fields h1 , h2 , · · · as follows
g = h1 ⊕ h2 ⊕ · · ·
(4)
Whereas, the system is locally strongly controllable if and
only if
g = h2 ⊕ h3 ⊕ · · ·
(5)
The term corresponding to h1 is just the space spanned by
the columns of the local connection form A(x), which in
our case is a 3×2 matrix. h2 is the column corresponding to
the curvature DA, which is a differential 2-form over Tx M .
The explicit calculation is done using DA = dA − [A, A],
where d(∗) is the exterior derivative, and [∗, ∗] is the Lie
bracket of the columns of A(x). The terms from h3 onwards
are the spaces spanned by the bracketing operation of
terms from connection form and its curvature’s, along with
LX ξ term, which is a Lie derivative with respect to the
vector is in the tangent of the shape space.
For the Purcell’s swimmer we explicitly calculated A(x),
followed by terms h1 , h2 , h3 the Purcell’s swimmer. The
term LX ξ is evaluated using Cartan’s magic formula
[Bloch (2003)]. We found that the rank of h2 ⊕ h3 is always
3, hence satisfying the strong controllability conditions
at all the points. Since all of these terms have a large
expressions and are unwieldy to mention in the paper, we
did a numerical calculation at few points in shape space
M and have shown the results in the Appendix A.
4. SYMMETRIC PURCELL’S SWIMMER
Now we turn our attention to the symmetrized version
of the Purcell’s swimmer, discussed in [Avron and Raz
(2008)]. Fig. 3 shows the schematic of the swimmer. As
compared to the original Purcell’s swimmer, this has an
additional limb at each of the 2 joints. Moreover, the limbs
located at the same point rotate symmetrically about the
base link (α3 = −α1 , α4 = −α2 ). Thus the shape space of
this swimmer is 2 dimensional, parametrized by (α1 , α2 ).
In the following section we derive the connection form
for the symmetric Purcell’s swimmer to show that it can
perform motion in only one dimension, which is along the
length of the base link.
4.1 Connection form
In order to calculate the connection form, we extend the
approach used for the basic Purcell’s swimmer in [Hatton
and Choset (2013)]. Each of the limbs moves in the Special
Eucledean group SE(2). Hence we represent the velocity
of i’th link by ξi = [ξi,x , ξi,y , ξi,θ ]T . As a consequence
#
cos α1
sin α1
0
cos α1
0 ,
T1 = − sin α1
L sin α1 −L(cos α1 + 1) 1
#
"
cos α2
− sin α2
0
cos α2
0
T2 = sin α2
L sin α2 L(cos α2 + 1) 1
"
Due to symmetry, these force transformation matrices for
links 3 and 4 are obtained by replacing α1 by −α3 and α2
by −α4 , respectively. Again, since the link 0 is the reference
link, we get T0 as the identity transformation. The total
force is summation of the forces acting on the individual
links, transformed to the coordinate frame of the base link.
Ftotal =
4
X
"
Ti HBi
i=0
Fig. 3. Symmetric Purcell’s swimmer
of modelling of the limbs according to Cox theory [Cox
(1970)] and the resistive force theory [Tam and Hosoi
(2007)] at low Reynold’s number, for k as the differential
viscous drag, we get the force acting on each limb through
linear relationship,
kL 0
Fi = Hξi , H = 0 2kL
0 0
0
0
2
3
kL
3
(6)
The velocity of the i’th link ξi depends on the velocity of
the base link and shape velocities (α̇1 , α̇2 ). This relationship is given by the following transformation " #
"
#
ξ
ξ0,x
ξi = Bi α̇1 , ξ = ξ0,y
(7)
α̇2
ξ0,θ
where,
cos α1 − sin α1
L sin α1
B1 = sin α1 cos α1 −L(cos α1 + 1)
0
0
1
"
cos α2 sin α2
L sin α2
B2 = − sin α2 cos α2 −L(cos α2 + 1)
0
0
1
"
#
0 0
L 0 ,
−1 0
#
0 0
L0
1 0
B3 and B4 are obtained by replacing α1 by −α3 and α2
by −α4 in B1 and B2 respectively, and multiplying the
last 2 columns of B1 and B2 by −1, since the direction of
motion of the symmetric limb would always be equal in
magnitude and opposite in direction. Since the link 0 is
the base link, the coordinate frame attached to which is
the reference body frame, we get
"
#
10000
B0 = 0 1 0 0 0
(8)
00100
We transform the force acting on the each link to the
coordinate frame attached to the base link. The force
transformation matrices for link 1 and 2 are given by
ξ
α̇1
α̇2
#
(9)
The consequence of being at low Reynolds number is
that the net forces and moments on an isolated system
is zero [Hatton and Choset (2013)]. Moreover, in order to
bring the system of equations in a pure kinematic form,
[Shammas et al. (2007)], we write the terms Ti HBi in the
block matrix form to separate those columns which are
being multiplied by group velocity term ξ and those being
multiplied by shape velocity term [α̇1 , α̇2 ]T . This yields
" #
4
ξ
X
0=
[[Ti HBi ]3×3 [Ti HBi ]3×2 ] α̇1
(10)
α̇2
i=0
Thus we write the system in a pure kinematic form, like
in equation 2, as
4
4
X
X
α̇
[Ti HBi ]3×3 ξ =
[Ti HBi ]2×2 1
(11)
α̇2
i=0
i=0
with local connection form as
4
4
X
X
[Ti HBi ]3×2
A = [ [Ti HBi ]3×3 ]−1
i=0
(12)
i=0
On explicit calculation, for k = 1, L = 1, we get connection
form as a 2 × 2 matrix whose columns h1 and h2 are given
as
#
"
#
"
4 sin α1
−4 sin α2
h1 =
2 sin2 α1 +2 sin2 α2 +5
0
0
, h2 =
2 sin2 α1 +2 sin2 α2 +5
0
0
(13)
The rows of the connection form corresponding to velocity
directions ξy and ξθ are identically zero, which implies that
the motion of the swimmer is always along the length of
the base link. Thus the configuration space of the swimmer
is obtained as Q = R1 × S1 × S1 .
4.2 Controllability of the symmetric Purcell’s swimmer
Since the structure group of the symmetric Purcell’s swimmer is just the real line R, which is an Abelian group, in
order to prove the controllability we use the special case
of the Ambrose-Singer theorem, referred from [Kelly and
Murray (1995)]. It says that a locomotion system on an
Abelian principal bundle is strongly controllable at x ∈ M ,
if and only if
span{dA(X, Y ) : X, Y ∈ Tx M, x ∈ M } = g
(14)
A similar argument as that for the basic Purcell’s swimmer
shows that the symmetric Purcell’s swimmer also follows
a trivial principal fiber bundle structure. Hence, we can
apply this test for its controllability. From equation (13),
by taking the exterior derivative, we get the expression for
dA, a differential 2-form as
16 sin α1 sin α2 (cos α1 − cos α2 )
dA =
dα1 ∧ dα2 (15)
(2 sin2 α1 + 2 sin2 α2 + 5)2
studies controllability and optimal gait design for planar
swimming alone. It seems pertinent to explore the applicability of the underlying theory of low Reynold’s number for
slender members to 3-dimensional motion. This extension
can open up many avenues for replicating more realistic
microbial motion, followed by subsequent mechanistic and
control theoretic analysis.
We define following 3 sets
S1 = (α1 , α2 ) ∈ M \ {(0, ∗), (∗, 0), (π, ∗), (∗, π),
(c, c), ∀c ∈ (0, π)}
S2 = (α1 , α2 ) ∈ {(0, 0), (π, π)}
S3 = (α1 , α2 ) ∈ {M \ S1} \ S2
Avron, J. and Raz, O. (2008). A geometric theory
of swimming: Purcell’s swimmer and its symmetrized
cousin. New Journal of Physics, 10(6), 063016.
Becker, L.E., Koehler, S.A., and Stone, H.A. (2003). On
self-propulsion of micro-machines at low reynolds number: Purcell’s three-link swimmer. Journal of fluid mechanics, 490, 15–35.
Bloch, A. (2003). Nonholonomic mechanics. In Nonholonomic Mechanics and Control, 207–276. Springer.
Bloch, A.M., Krishnaprasad, P., Marsden, J.E., and Murray, R.M. (1996). Nonholonomic mechanical systems
with symmetry. Archive for Rational Mechanics and
Analysis, 136(1), 21–99.
Bullo, F. and Lewis, A.D. (2004). Geometric control of
mechanical systems: modeling, analysis, and design for
simple mechanical control systems, volume 49. Springer
Science & Business Media.
Burton, L.J. (2013). The dynamics and kinematics of
bio-in swimming systems. Ph.D. thesis, Massachusetts
Institute of Technology.
Cabrera, A. (2008). Base-controlled mechanical systems
and geometric phases. Journal of Geometry and Physics,
58(3), 334–367.
Cox, R. (1970). The motion of long slender bodies in a
viscous fluid part 1. general theory. Journal of Fluid
mechanics, 44(04), 791–810.
Dreyfus, R., Baudry, J., Roper, M.L., Fermigier, M., Stone,
H.A., and Bibette, J. (2005). Microscopic artificial
swimmers. Nature, 437(7060), 862–865.
Friedrich, B., Riedel-Kruse, I., Howard, J., and Jülicher, F.
(2010). High-precision tracking of sperm swimming fine
structure provides strong test of resistive force theory.
The Journal of experimental biology, 213(8), 1226–1234.
Giraldi, L., Martinon, P., and Zoppello, M. (2013). Controllability and optimal strokes for n-link microswimmer. In Decision and Control (CDC), 2013 IEEE 52nd
Annual Conference on, 3870–3875. IEEE.
Hatton, R.L. and Choset, H. (2013). Geometric swimming
at low and high reynolds numbers. Robotics, IEEE
Transactions on, 29(3), 615–624.
Holm, D.D., Schmah, T., Stoica, C., and Ellis, D.C.
(2009). Geometric mechanics and symmetry: from finite
to infinite dimensions. Oxford University Press London.
Kelly, S., Fairchild, M.J., Hassing, P.M., and Tallapragada,
P. (2012). Proportional heading control for planar
navigation. In 2012 American Control Conference, ACC
2012.
Kelly, S.D. and Murray, R.M. (1995). Geometric phases
and robotic locomotion. Journal of Robotic Systems,
12(6), 417–431.
Kobayashi, S. and Nomizu, K. (1963). Foundations of
differential geometry, volume 1. New York.
Melli, J.B., Rowley, C.W., and Rufat, D.S. (2006). Motion
planning for an articulated body in a perfect planar
Thus we conclude the weak and strong controllability for
the symmetric version of the swimmer as below
• The system is strongly controllable when
span{dA} = 1 ∀(α1 , α2 ) ∈ S1
• The system is not controllable when
span{h1 ⊕ h2 } = span{dA} = 0, ∀(α1 , α2 ) ∈ S2
• The system is weakly controllable when
span{h1 ⊕ h2 } = 1, span{dA} = 0 ∀(α1 , α2 ) ∈ S3
These 3 sets are represented in figure 4 below.
Fig. 4. Controllable sets of symmetric Purcell’s swimmer
5. CONCLUSION AND FUTURE WORK
In this paper we explored the geometry of the Purcell’s
swimmer’s configuration and showed its strong controllability by using an extension of Chow’s theorem for systems
on a principal fiber bundle. We then used the resistive
force theory to derive the kinematic form of the symmetric
version of the Purcell’s swimmer. We showed that its
structure group is Abelian and used the Ambrose-Singer
theorem to characterize the points in the configuration
space at which the system is locally strongly and weakly
controllable and uncontrollable.
This work and many other contributions in the Purcell’s
swimmers modelling, control and controllability revolve
around the basic planar swimmer. Even the N-link extension of the Purcell’s swimmer [Giraldi et al. (2013)]
REFERENCES
fluid. SIAM Journal on applied dynamical systems, 5(4),
650–669.
Najafi, A. and Golestanian, R. (2004). Simple swimmer
at low reynolds number: Three linked spheres. Physical
Review E, 69(6), 062901.
Ostrowski, J. and Burdick, J. (1998). The geometric
mechanics of undulatory robotic locomotion. The international journal of robotics research, 17(7), 683–701.
Passov, E. and Or, Y. (2012). Dynamics of purcells threelink microswimmer with a passive elastic tail. The
European Physical Journal E, 35(8), 1–9.
Purcell, E.M. (1977). Life at low reynolds number. Am.
J. Phys, 45(1), 3–11.
Rowland, T. (2016).
Fiber bundle. From
MathWorld—A Wolfram Web Resource.
URL
http://mathworld.wolfram.com/FiberBundle.html.
Last visited on 29/1/2016.
Shammas, E.A., Choset, H., and Rizzi, A.A. (2007). Geometric motion planning analysis for two classes of underactuated mechanical systems. The International Journal
of Robotics Research, 26(10), 1043–1073.
Tam, D. and Hosoi, A.E. (2007). Optimal stroke patterns
for purcells three-link swimmer. Physical Review Letters, 98(6), 068105.
Appendix A. BASIC PURCELL’S
CONTROLLABILITY CONDITION
We take a few points from the configuration space of the
basic Purcell’s swimmer, and tabulate numerical values
of the terms h1 , h2 , h3 , explained in section 2. It is to
be noted that differential viscous drag k and half of the
limb lengths L are both taken to be unity for simplicity;
the controllability results still hold the same for any other
realistic value. We note that the dimension of the space
spanned by these terms is 3 at all these points.
(α1 , α2 )
h1
"
(0, 0)
"
(0, π/4)
(0, π/2)
(0, 3π/4)
(π, 0)
,
,
# "
0
{ 0.333
−0.333
"
#
,
# "
0
{ −0.629
0.185
,
"
}
#
}
#
−0.416
0.714
0.343
#
0
5.66
2.33
#
#
#
"
}
#
0
−0.037
0.481
#
−2.888
0
{
0
"
}
}
#
0.419
0
{
0
}
,
,
#
0
{ 34.481
12.481
"
,
3
}
3
}
3
}
3
#
−0.418
2.828
0.325
"
,
}
#
#
0
77.629
10.518
#
0
{ −0.059
−0.372
3
−0.339
0.812
−0.301
"
,
}
#
−0.341
0.778
−0.348
# "
−0.033
{ 1.261
0.227
"
,
# "
−0.024
{ 0.308
−0.167
"
}
#
0
0.821
−0.329
# "
−0.087
{ 0.233
−0.237
"
}
−0.289
{ −0.317
0.037
# "
0
{ 0.228
−0.242
"
}
−0.223
{ −0.321
−0.064
"
}
"
}
−0.389
{ −0.185
−0.058
"
span{h2 ⊕ h3 }
h3
#
−0.469
0
{
0
}
−0.395
0.283
0.209
# "
−0.127
{ 0.222
−0.366
"
(0, π)
,
"
−0.226
0.306
0.243
# "
−0.16
{ 0.271
−0.320
"
,
#
0
0.333
0.259
# "
−0.106
{ 0.317
−0.277
"
h2
# "
0
{ 0.333
−0.259
#
0
−0.166
−0.232
}
3
| 3cs.SY
|
1
Spatio-Temporal Backpropagation for Training
High-performance Spiking Neural Networks
arXiv:1706.02609v3 [cs.NE] 12 Sep 2017
Yujie Wu † , Lei Deng † , Guoqi Li, Jun Zhu and Luping Shi
Abstract—Compared with artificial neural networks (ANNs),
spiking neural networks (SNNs) are promising to explore the
brain-like behaviors since the spikes could encode more spatiotemporal information. Although existing schemes including pretraining from ANN or direct training based on backpropagation
(BP) make the supervised training of SNNs possible, these
methods only exploit the networks’ spatial domain information
which leads to the performance bottleneck and requires many
complicated training techniques. Another fundamental issue is
that the spike activity is naturally non-differentiable which
causes great difficulties in training SNNs. To this end, we build
an iterative LIF model that is friendlier for gradient descent
training. By simultaneously considering the layer-by-layer spatial
domain (SD) and the timing-dependent temporal domain (TD)
in the training phase, as well as an approximated derivative for
the spike activity, we propose a spatio-temporal backpropagation
(STBP) training framework without using any complicated skill.
We design the corresponding fully connected and convolution
architecture and evaluate our framework on the static MNIST
and a custom object detection dataset, as well as the dynamic
N-MNIST. Results show that our approach achieves the best
accuracy compared with existing state-of-the-art algorithms on
spiking networks. This work provides a new perspective to explore the high-performance SNNs for future brain-like computing
paradigm with rich spatio-temporal dynamics.
I. I NTRODUCTION
Deep neural networks (DNNs) have achieved outstanding
performance in diverse areas [1]–[5], while it seems that
the brain uses another network architecture, spiking neural
networks, to realize various complicated cognitive functions
[6]–[8]. Compared with the existing DNNs, SNNs mainly have
two superiorities: 1) the spike pattern flowing through SNNs
fundamentally codes more spatio-temporal information, while
most DNNs lack timing dynamics, especially the widely used
feedforward DNNs; and 2) event-driven paradigm of SNNs
can make it more hardware friendly, and be adopted by many
neuromorphic platforms [9]–[14].
However, it remains challenging in training SNNs because
of the quite complicated dynamics and non-differentiable
nature of the spike activity. In summary, there exist three
kinds of training methods for SNNs: 1) unsupervised learning;
2) indirect supervised learning; 3) direct supervised learning.
The first one origins from the biological synaptic plasticity for
† The
authors contribute equally.
Yujie Wu, Lei Deng, Guoqi Li and Luping Shi are with Center
for Brain-Inspired Computing Research (CBICR),Department of
Precision Instrument,Tsinghua University, 100084 Beijing, China.
(email:[email protected];)
Jun Zhu is with state Key Lab of Intelligence Technology and System,Tsinghua National Lab for Information Science and Technology,Tsinghua
University, 100084 Beijing, China. (email:[email protected])
weight modification, such as spike timing dependent plasticity
(STDP) [15]–[17]. Because it only considers the local neuronal
activities, it is difficult to achieve high performance. The
second one firstly trains an ANN, and then transforms it into
its SNN version with the same network structure where the
spiking rate of SNN neurons acts as the analog activity of
ANN neurons [18]–[21]. This is not a bio-plausible way to
explore the learning nature of SNNs. The most promising
method to obtain high-performance training is the recent direct
supervised learning based on the gradient descent theory with
error backpropagation. However, such a method only considers
the layer-by-layer spatial domain and ignores the dynamics
in temporal domain [22], [23]. Therefore many complicated
training skills are required to improve performance [19], [23],
[24], such as fixed-amount-proportional reset, lateral inhibition, error normalization, weight/threshold regularization, etc.
Thus, a more general dynamic model and learning framework
on SNNs are highly required.
In this paper, we propose a direct supervised learning
framework for SNNs which combines both the SD and TD in
the training phase. Firstly, we build an iterative LIF model with
SNNs dynamics but it is friendly for gradient descent training.
Then we consider both the spatial direction and temporal
direction during the error backpropagation procedure, i.e,
spatio-temporal backpropagation (STBP), which significantly
improves the network accuracy. Furthermore, we introduce
an approximated derivative to address the non-differentiable
issue of the spike activity. We test our SNNs framework
by using the fully connected and convolution architecture on
the static MNIST and a custom object detection dataset, as
well as the dynamic N-MNIST. Many complicated training
skills which are generally required by existing schemes, can
be avoided due to the fact that our proposed method can
make full use of STD information that captures the nature of
SNNs. Experimental results show that our proposed method
could achieve the best accuracy on either static or dynamic
dataset, compared with existing state-of-the-art algorithms.
The influence of TD dynamics and different methods for
the derivative approximation are systematically analyzed. This
work shall open a way to explore the high-performance SNNs
for future brain-like computing paradigms with rich STD
dynamics.
II. M ETHOD AND M ATERIAL
A. Iterative Leaky Integrate-and-Fire Model in Spiking Neural
Networks
Compared with existing deep neural networks, spiking
neural networks fundamentally code more spatio-temporal
2
Fig. 1: Illustration of the spatio-temporal characteristic of SNNs. Besides the layer-by-layer spatial dataflow like ANNs,
SNNs are famous for the rich temporal dynamics and non-volatile potential integration. However, the existing training algorithms
only consider either the spatial domain such as the supervised ones via backpropagation, or the temporal domain such as the
unsupervised ones via timing-based plasticity, which causes the performance bottleneck. Therefore, how to build an learning
framework making full use of the spatio-temporal domain (STD) is fundamentally required for high-performance SNNs that
forms the main motivation of this work.
information due to two facts that i) SNNs can also have
deep architectures like DNNs, and ii) each neuron has its own
neuronal dynamic properties. The former one grants SNNs rich
spatial domain information while the later one offers SNNs the
power of encoding temporal domain information. However,
currently there is no unified framework that allows the effective training of SNNs just as implementing backpropagation
(BP) in DNNs by considering the spatio-temporal dynamics.
This has challenged the extensive use of SNNs in various applications. In this work, we will present a framework based on
iterative leaky integrate-and-fire (LIF) model that enables us
to apply spatio-temporal backpropagation for training spiking
neural networks.
It is known that LIF is the most widely applied model to
describe the neuronal dynamics in SNNs, and it can be simply
governed by
τ
du(t)
= −u(t) + I(t)
dt
(1)
where u(t) is the neuronal membrane potential at time t, τ
is a time constant and I(t) denotes the pre-synaptic input
which is determined by the pre-neuronal activities or external
injections and the synaptic weights. When the membrane
potential u exceeds a given threshold Vth , the neuron fires
a spike and resets its potential to ureset . As shown in Figure
1, the forward dataflow of the SNN propagates in the layerby-layer SD like DNNs, and the self-feedback injection at
each neuron node generates non-volatile integration in the
TD. In this way, the whole SNN runs with complex STD
dynamics and codes spatio-temporal information into the spike
pattern. The existing training algorithms only consider either
the SD such as the supervised ones via backpropagation, or the
TD such as the unsupervised ones via timing-based plasticity,
which causes the performance bottleneck. Therefore, how to
build an learning framework making full use of the STD is
fundamentally required for high-performance SNNs that forms
the main motivation of this work.
However, obtaining the analytic solution of LIF model in
(1) directly makes it inconvenient/obscure to train SNNs based
on backpropagation. This is because the whole network shall
present complex dynamics in both SD and TD. To address this
issue, the following event-driven iterative updating rule
u(t) = u(ti−1 )e
ti−1 −t
τ
+ I(t)
(2)
can be well used to approximate the neuronal potential u(t) in
(1) based on the last spiking moment ti−1 and the pre-synaptic
input I(t). The membrane potential exponentially decays until
the neuron receives pre-synaptic inputs, and a new update
round will start once the neuron fires a spike. That is to say, the
neuronal states are co-determined by the spatial accumulations
of I(t) and the leaky temporal memory of u(ti−1 ).
As we know, the efficiency of error backpropagation for
training DNNs greatly benefits from the iterative representation of gradient descent which yields the chain rule for layerby-layer error propagation in the SD backward pass. This
motivates us to propose a iterative LIF based SNN in which
the iterations occur in both the SD and TD as follows:
l(n−1)
xt+1,n
i
=
ut+1,n
=
i
ot+1,n
i
=
X
n t+1,n−1
wij
oj
j=1
t,n
ut,n
i f (oi )
g(ut+1,n
)
i
+ xt+1,n
+ bni
i
(3)
(4)
(5)
where
x
f (x) = τ e− τ
(
g(x) =
1,
0,
x ≥ Vth
x < Vth
(6)
(7)
3
In above formulas, the upper index t denotes the moment at
time t, and n and l(n) denote the nth layer and the number
of neurons in the nth layer, respectively. wij is the synaptic
weight from the jth neuron in pre-synaptic layer to the ith
neuron in the post-synaptic layer, and oj ∈ {0, 1} is the
neuronal output of the jth neuron where oj = 1 denotes a
spike activity and oj = 0 denotes nothing occurs. xi is a
simplified representation of the pre-synaptic inputs of the ith
neuron, similar to the I in the original LIF model. ui is the
neuronal membrane potential of the ith neuron and bi is a bias
parameter related the threshold Vth .
Actually, formulas (4)-(5) are also inspired from the LSTM
model [25]–[27] by using a forget gate f (.) to control the TD
memory and an output gate g(.) to fire a spike. The forget
gate f (.) controls the leaky extent of the potential memory in
the TD, the output gate g(.) generates a spike activity when
it is activated. Specifically, for a small positive time constant
τ , f (.) can be approximated as
(
τ, ot,n
=0
t,n
i
f (oi ) ≈
(8)
0, ot,n
=1
i
direction, thus the states in the TD at different time steps
can be distinguished that enables the chain rule for iterative
propagation. Similar idea can be found in the BPTT algorithm
for training RNNs in [28].
Now, we discuss how to obtain the complete gradient
descent based on the following four cases. Firstly, we denote
that
∂L
(10)
δit,n = t,n
∂oi
Case 1: t = T at the output layer n = N .
In this case, the derivative ∂o∂L
T ,N can be directly obtained since
i
it depends on the loss function in Eq.(9) of the output layer.
We could have
∂L
∂oT,N
i
=−
T
1
1 X k,N
(yi −
oi ).
TS
T
(11)
k=1
The derivation with respect to uT,N
is generated based on
i
oT,N
i
1
since τ e− τ ≈ 0. In this way, the original LIF model could
be transformed to an iterative version where the recursive
relationship in both the SD and TD is clearly describe, which
is friendly for the following gradient descent training in the
STD.
B. Spatio-Temporal Backpropagation Training
In order to present STBP training methodology, we define
the following loss function L in which the mean square error
for all samples under a given time windows T is to be
minimized
S
T
1 X
1 X t,N 2
L=
k ys −
o
k2
2S s=1
T t=1 s
∂L
∂uT,N
i
=
∂L ∂oT,N
i
= δiT,N
∂oT,N
∂uT,N
i
i
∂oT,N
i
∂uT,N
i
.
(12)
Case 2: t = T at the layers n < N .
In this case, the derivative ∂o∂L
T ,n iteratively depends on the
i
error propagation in the SD at time T as the typical BP
algorithm. We have
l(n+1)
∂L
∂oT,n
i
=
∂oT,n+1
j
δjT,n+1
∂oT,n
i
j=1
X
l(n+1)
=
X
δjT,n+1
j=1
∂g
∂uT,n
i
wji .
(13)
(9)
where ys and os denote the label vector of the sth training
sample and the neuronal output vector of the last layer N ,
respectively.
By combining equations (3)-(9) together it can be seen that
L is a function of W and b. Thus, to obtain the derivative of
L with respect to W and b is required for the STBP algorithm
based on gradient descent. Assume that we have obtained
∂L
∂L
and ∂u
at each layer n at time t, which
derivative of ∂o
i
i
∂L
is an essential step to obtain the final ∂W
and ∂L
∂b . Figure2
describes the error propagation (dependent on the derivation)
in both the SD and TD at the single-neuron level (figure2.a)
and the network level (figure2.b). At the single-neuron level,
the propagation is decomposed into a vertical path of SD and
a horizontal path of TD. The dataflow of error propagation in
the SD is similar to the typical BP for DNNs, i.e. each neuron
accumulates the weighted error signals from the upper layer
and iteratively updates the parameters in different layers; while
the dataflow in the TD shares the same neuronal states, which
makes it quite complicated to directly obtain the analytical
solution. To solve this problem, we use the proposed iterative
LIF model to unfold the state space in both the SD and TD
Similarly, the derivative
∂L
=
∂uT,n
i
∂L
,n
∂uT
i
∂L
∂uTi +1,n
∂uTi +1,n
∂uT,n
i
yields
=
∂L
∂uTi +1,n
f (oTi +1,n ).
(14)
Case 3: t < T at the output layer n = N .
In this case, the derivative ∂o∂L
t,N depends on the error propi
agation in the TD direction. With the help of the proposed
iterative LIF model in Eq.(3)-(5) by unfolding the state space
in the TD, we acquire the required derivative based on the
chain rule in the TD as follows
∂L
∂ot,N
i
= δit+1,N
∂ut,N
i
∂L
,N
∂oT
i
=
∂ot,N
i
+
∂L
∂oT,N
i
∂g
∂L
t+1,N
t,N ∂f
= δi
ui
+ T,N ,
t,N
∂ut+1,N
∂o
∂o
i
j
i
∂L
where
∂ot+1,N
i
∂L ∂ot,N
i
∂ot,N
i
∂ut,N
i
= − T1S (yi −
1
T
= δit,N
PT
k=1
∂g
∂ut,N
i
,
(15)
(16)
(17)
ok,N
) as in Eq.(11).
i
4
Fig. 2: Error propagation in the STD. (a) At the single-neuron level, the vertical path and horizontal path represent the error
propagation in the SD and TD, respectively. (b) Similar propagation occurs at the network level, where the error in the SD
requires the multiply-accumulate operation like the feedforward computation.
Case 4: t < T at the layers n < N .
In this case, the derivative ∂o∂L
t,n depends on the error propagai
tion in both SD and TD. On one side, each neuron accumulates
the weighted error signals from the upper layer in the SD
like Case 2; on the other side, each neuron also receives the
propagated error from self-feedback dynamics in the TD by
iteratively unfolding the state space based on the chain rule
like Case 3. So we have
l(n+1)
X t,n+1 ∂ot,n+1
∂L
∂L ∂oit+1,n
j
=
δ
+
j
t,n
∂ot,n
∂ot,n
∂oit+1,n ∂oi
i
i
j=1
l(n+1)
=
X
j=1
δjt,n+1
(18)
T
T
X ∂L ∂ut,n
X ∂L
∂L
=
=
,
∂bn
∂ut,n bn
∂ut,n
t=1
t=1
T
(22)
T
X ∂L ∂ut,n
X ∂L
∂L
T
=
=
ot,n−1 ,
n
t,n
n
t,n
∂W
∂u
W
∂u
t=1
t=1
(23)
where ∂u∂L
t,n can be obtained from in Eq.(11)-(21). Given
the W and b according to the STBP, we can use gradient
descent optimization algorithms to effectively train SNNs for
achieving high performance.
∂g
t+1,n ∂g
t,n ∂f
,
t,n wji + δi
t,n ui
∂ui
∂ui
∂ot,n
i
(19)
∂L ∂ot,n
∂L ∂uit+1,n
∂L
i
=
+
t,n
t,n
∂ut,n
∂ot,n
∂uit+1,n ∂ui
i
i ∂ui
∂g
∂L
= δit,n t,n +
f (oit+1,n ).
∂ui
∂uit+1,n
(20)
(21)
Based on the four cases, the error propagation procedure
(depending on the above derivatives) is shown in Figure2.
At the single-neuron level (Figure2.a), the propagation is
decomposed into the vertical path of SD and the horizontal
path of TD. At the network level (Figure2.b), the dataflow of
error propagation in the SD is similar to the typical BP for
DNNs, i.e. each neuron accumulates the weighted error signals
from the upper layer and iteratively updates the parameters in
different layers; and in the TD the neuronal states are unfolded
iteratively in the timing direction that enables the chain-rule
propagation. Finally, we obtain the derivatives with respect to
W and b as follows
C. Derivative Approximation of the Non-differentiable Spike
Activity
In the previous sections, we have presented how to obtain
the gradient information based on STBP, but the issue of nondifferentiable points at each spiking time is yet to be addressed.
Actually, the derivative of output gate g(u) is required for the
STBP training of Eq.(11)-(22). Theoretically, g(u) is a nondifferentiable Dirac function of δ(u) which greatly challenges
the effective learning of SNNs [23]. g(u) has zero value
everywhere except an infinity value at zero, which causes the
gradient vanishing or exploding issue that disables the error
propagation. One of existing method viewed the discontinuous
points of the potential at spiking times as noise and claimed
it is beneficial for the model robustness [23], [29], while it
did not directly address the non-differentiability of the spike
activity. To this end, we introduce four curves to approximate
the derivative of spike activity denoted by h1 , h2 , h3 and h4
in Figure3.b:
5
B. Dataset Experiments
1
a1
h1 (u) = sign(|u − Vth | < ),
(24)
a1
2
√
a2
a2
2
− |u − Vth |)sign( √ − |u − Vth |),
h2 (u) = (
2
4
a2
(25)
Vth −u
1
e a3
,
h3 (u) =
a3 (1 + e Vtha3−u )2
(26)
(u−Vth )2
1
e− 2a4 ,
(27)
2πa4
where ai (i = 1, 2, 3, 4) determines the curve shape and
steep degree. In fact, h1 , h2 , h3 and h4 are the derivative of the
rectangular function, polynomial function, sigmoid function
and Gaussian cumulative distribution function, respectively.
To be consistent with the Dirac function δ(u), we introduce
the coefficient ai to ensure the integral of each function is
1. Obviously, it can be proven that all the above candidates
satisfy that
h4 (u) = √
lim hi (u) =
ai →0+
Thus,
∂g
∂u
dg
, i = 1, 2, 3, 4.
du
(28)
in Eq.(11)-(22) for STBP can be approximated by
dg
dhi
≈
.
(29)
du
du
In section III-C, we will analyze the influence on the SNNs
performance with different curves and different values of ai .
III. R ESULTS
A. Parameter Initialization
The initialization of parameters, such as the weights, thresholds and other parameters, is crucial for stabilizing the firing
activities of the whole network. We should simultaneously
ensure timely response of pre-synaptic stimulus but avoid too
much spikes that reduces the neuronal selectivity. As it is
known that the multiply-accumulate operations of the prespikes and weights, and the threshold comparison are two key
steps for the computation in the forward pass. This indicates
the relative magnitude between the weights and thresholds
determines the effectiveness of parameter initialization. In this
paper, we fix the threshold to be constant in each neuron
for simplification, and only adjust the weights to control the
activity balance. Firstly, we initial all the weight parameters
sampling from the standard uniform distribution
W ∼ U [−1, 1]
Then, we normalize these parameters by
n
wij
n
wij
= qP
, i = 1, .., l(n)
l(n−1) n 2
wij
j=1
(30)
(31)
The set of other parameters is presented in TableI.
Furthermore, throughout all the simulations in our work, any
complex skill as in [19], [23] is no longer required, such
as the fixed-amount-proportional reset, error normalization,
weight/threshold regularization, etc.
We test our SNNs model and the STBP training method
on various datasets, including the static MNIST and a custom
object detection dataset, as well as the dynamic N-MNIST
dataset. The input of the first layer should be a spike train,
which requires us to convert the samples from the static
datasets into spike events. To this end, the Bernoulli sampling
from original pixel intensity to the spike rate is used in this
paper.
1) Spatio-temporal fully connected neural network: Static
Dataset. The MNIST dataset of handwritten digits [30] (figure4.b) and a custom dataset for object detection [14] (figure4.a) are chosen to test our method. MNIST is comprised
of a training set with 60,000 labelled hand-written digits, and a
testing set of other 10,000 labelled digits, which are generated
from the postal codes of 0-9. Each digit sample is a 28×28
grayscale image. The object detection dataset is a two-category
image dataset created by our lab for pedestrian detection.
It includes 1509 training samples and 631 testing samples
of 28×28 grayscale image. By detecting whether there is a
pedestrian, an image sample is labelled by 0 or 1, as illustrated
in Figure4.a. The upper and lower sub-figures in Figure4.c
are the spike pattern of 25 input neurons converted from
the center patch of 5×5 pixels of a sample example on the
object detection dataset and MNIST, respectively. Figure4.d
illustrates an example for the spike pattern of output layer
within 15ms before and after the STBP training over the
stimulus of digit 9. At the beginning, neurons in the output
layer randomly fires, while after the training the 10th neuron
coding digit 9 fires most intensively that indicates correct
inference is achieved.
TableII compares our method with several other advanced
results that use the similar MLP architecture on MNIST.
Although we do not use any complex skill, the proposed STBP
training method also outperforms all the reported results. We
can achieve 98.89% testing accuracy which performs the best.
TableIII compares our model with the typical MLP on the object detection dataset. The contrast model is one of the typical
artificial neural networks (ANNs), i.e. not SNNs, and in the
following we use ’non-spiking network’ to distinguish them.
It can be seen that our model achieves better performance
than the non-spiking MLP. Note that the overall firing rate of
the input spike train from the object detection dataset is higher
than the one from MNIST dataset, so we increase its threshold
to 2.0 in the simulation experiments.
Dynamic Dataset. Compared with the static dataset,
dynamic dataset, such as the N-MNIST [32], contains richer
temporal features, and therefore it is more suitable to exploit
SNN’s potential ability. We use the N-MNIST database as
an example to evaluate the capability of our STBP method
on dynamic dataset. N-MNIST converts the mentioned static
MNIST dataset into its dynamic version of spike train by
using the dynamic vision sensor (DVS) [33]. For each original
sample from MNIST, the work [32] controls the DVS to
move in the direction of three sides of the isosceles triangle in
turn (figure5.b) and collects the generated spike train which
6
Fig. 3: Derivative approximation of the non-differentiable spike activity. (a) Step activation function of the spike activity and its original derivative
function which is a typical Diract function δ(u) with infinite value at u = 0 and zero value at other points. This non-differentiable property disables the error
propagation. (b)Several typical curves to approximate the derivative of spike activity.
TABLE I: Parameters set in our experiments
Network parameter
T
Vth
τ
a1 , a2 , a3 , a4
dt
r
β1 , β 2 , λ
Description
Time window
Threshold (MNIST/object detection dataset/N-MNIST)
Decay factor (MNIST/object detection dataset/N-MNIST)
Derivative approximation parameters(Figure3)
Simulation time step
Learning rate (SGD)
Adam parameters
Value
30ms
1.5, 2.0, 0.2
0.1ms, 0.15ms, 0.2ms
1.0
1ms
0.5
0.9, 0.999, 1-10−8
Fig. 4: Static dataset experiments. (a) A custom dataset for object detection. This dataset is a two-category image set built
by our lab for pedestrian detection. By detecting whether there is a pedestrian, an image sample is labelled by 0 or 1. The
images in the yellow boxes are labelled as 1, and the rest ones are marked as 0. (b)MNIST dataset. (c) Raster plot of the spike
pattern of 49 input neurons converted from the center patch of 5×5 pixels of a sample example on the object detection dataset
(up) and MNIST (down). (d) Raster plot presents the comparison of output spike pattern before and after the STBP training
over a digit 9 on MNIST dataset.
is triggered by the intensity change at each pixel. Figure5.a
records the saccade results on digit 0. Each sub-graph records
the spike train within 10ms and each 100ms represents one
saccade period. Due to the two possible change directions
of each pixel intensity (brighter or darker), DVS could
capture the corresponding two kinds of spike events, denoted
by on-event and off-event, respectively (figure5.c). Since
N-MNIST allows the relative shift of images during the
saccade process, it produces 34×34 pixel range. And from
the spatio-temporal representation in figure5.c, we can see
7
TABLE II: Comparison with the state-of-the-art spiking networks with similar architecture on MNIST.
Model
Spiking RBM (STDP) [31]
Spiking RBM(pre-training*) [20]
Spiking MLP(pre-training*) [19]
Spiking MLP(BP) [22]
Spiking MLP(STDP) [15]
Network structure
784-500-40
784-500-500-10
784-1200-1200-10
784-200-200-10
784-6400
Spiking MLP(BP) [23]
784-800-10
Spiking MLP(STBP)
784-800-10
Training skills
None
None
Weight normalization
None
None
Error normalization/
parameter regularization
None
Accuracy
93.16%
97.48%
98.64%
97.66%
95.00%
98.71%
98.89%
We mainly compare with these methods that have the similar network architecture, and * means that their model is based on pre-trained ANN models.
TABLE III: Comparison with the typical MLP over object detection dataset.
Model
Non-spiking MLP(BP)
Spiking MLP(STBP)
Network structure
784-400-10
784-400-10
Mean
98.31%
98.34%
Accuracy
Interval∗
[97.62%, 98.57%]
[97.94%, 98.57%]
* results with epochs [201,210].
that the on-events and off-events are so different that we use
two channel to distinguish it. Therefore, the network structure
is 34×34×2-400-400-10.
Fig. 5: Dynamic dataset of N-MNIST. (a) Each sub-picture
shows a 10ms-width spike train during the saccades. (b) Spike
train is generated by moving the dynamic vision sensor (DVS)
in turn towards the direction of 1, 2 and 3. (c) Spatio-temporal
representation of the spike train from digit 0 [32]where the
upper one and lower one denote the on-events and off-events,
respectively.
TableIV compares our STBP method with some state-ofthe-art results on N-MNIST dataset. The upper 5 results are
based on ANNs, and lower 4 results including our method
uses SNNs. The ANNs methods usually adopt a frame-based
method, which collects the spike events in a time interval
(50ms ∼ 300ms) to form a frame of image, and use the
conventional algorithms for image classification to train the
networks. Since the transformed images are often blurred, the
frame-based preprocessing is harmful for model performance
and abandons the hardware friendly event-driven paradigm. As
can be seen from TableIV, the models of ANN are generally
worsen than the models of SNNs. In contrast, SNNs could
naturally handle event stream patterns, and by better use of
spatio-temporal feature of event streams, our proposed STBP
method achieves best accuracy of 98.78% when compared all
the reported ANNs and SNNs methods. The greatest advantage
of our method is that we did not use any complex training
skills, which is beneficial for future hardware implementation.
2) Spatio-temporal convolution neural network: Extending
our framework to convolution neural network structure allows
the network going deeper and grants network more powerful
SD information. Here we use our framework to establish
the spatio-temporal convolution neural network. Compared
with our spatio-temporal fully connected network, the main
difference is the processing of the input image, where we
use the convolution in place of the weighted summation.
Specifically, in the convolution layer, each convolution neuron
receives the convoluted input and updates its state according
to the LIF model. In the pooling layer, because the binary
coding of SNNs is inappropriate for standard max pooling,
we use the average pooling instead.
Our spiking CNN model are also tested on the MNIST
dataset as well as the object detection dataset . In the MNIST,
our network contains one convolution layers with kernel
size of 5 × 5 and two average pooling layers alternatively,
followed by one hidden layer. And like traditional CNN, we
use the elastic distortion [36] to preprocess dataset. TableV
records the state-of-the-art performance spiking convolution
neural networks over MNIST dataset. Our proposed spiking CNN model obtain 98.42% accuracy, which outperforms
other reported spiking networks with slightly lighter structure.
Furthermore, we configure the same network structure on a
custom object detection database to evaluate the proposed
model performance. The testing accuracy is reported after
training 200 epochs. TableVI indicates our spiking CNN model
could achieve a competitive performance with the non-spiking
8
TABLE IV: Comparison with state-of-the-art networks over N-MNIST.
Model
Non-spiking CNN(BP) [24]
Non-spiking CNN(BP) [34]
Non-spiking MLP(BP) [23]
LSTM(BPTT) [24]
Phased-LSTM(BPTT) [24]
Spiking CNN(pre-training*) [34]
Network structure
34 × 34 × 2-800-10
-
Spiking MLP(BP) [23]
34 × 34 × 2-800-10
Spiking MLP(BP) [35]
Spiking MLP(STBP)
34 × 34 × 2-10000-10
34 × 34 × 2-800-10
Training skills
None
None
None
Batch normalization
None
None
Error normalization/
parameter regularization
None
None
Accuracy
95.30%
98.30%
97.80%
97.05%
97.38%
95.72%
98.74%
92.87%
98.78%
We only show the network structure based on MLP, and the other network structure refers to the above references. *means that their model is based on
pre-trained ANN models.
TABLE V: Comparison with other spiking CNN over MNIST.
Model
Spiking CNN (pre-training∗ ) [13]
Spiking CNN(BP) [23]
Spiking CNN (STBP)
Network structure
28×28×1-12C5-P2-64C5-P2-10
28×28×1-20C5-P2-50C5-P2-200-10
28×28×1-15C5-P2-40C5-P2-300-10
Accuracy
99.12%
99.31%
99.42%
We mainly compare with these methods that have the similar network architecture, and * means that their model is based on pre-trained ANN models.
TABLE VI: Comparison with the typical CNN over object detection dataset.
Model
Non-spiking CNN(BP)
Spiking CNN(STBP)
Network structure
28 × 28 × 1-6C3-300-10
28 × 28 × 1-6C3-300-10
Mean
98.57%
98.59%
Accuracy
Interval∗
[98.57%, 98.57%]
[98.26%, 98.89%]
* results with epochs [201,210].
CNN.
C. Performance Analysis
1) The Impact of Derivative Approximation Curves: In
section II-B, we introduce different curves to approximate the
ideal derivative of the spike activity. Here we try to analyze
the influence of different approximation curves on the testing
accuracy. The experiments are also conducted on the MNIST
dataset, and the network structure is 784 − 400 − 10. The
testing accuracy is reported after training 200 epochs. Firstly,
we compare the impact of different curve shapes on model
performance. In our simulation we use the mentioned h1 ,
h2 , h3 and h4 shown in Figure3.b. Figure6.a illustrates the
results of approximations of different shapes. We observe that
different nonlinear curves, such as h1 , h2 , h3 and h4 , only
present small variations on the performance.
Furthermore, we use the rectangular approximation
as an example to explore the impact of width on the
experiment results. We set a1 = 0.1, 1.0, 2.5, 5.0, 7.5, 10 and
corresponding results are plotted in figure6.b. Different colors
denote different a1 values. Both too large and too small a1
value would cause worse performance and in our simulation,
a1 = 2.5 achieves the highest testing accuracy, which implies
the width and steepness of rectangle influence the model
performance. Combining figure 6.a and figure 6.b, it indicates
that the key point for approximating the derivation of the
spike activity is to capture the nonlinear nature, while the
specific shape is not so critical.
2) The Impact of Temporal Domain: A major contribution
of this work is introducing the temporal domain into the existing spatial domain based BP training method, which makes
full use of the spatio-temporal dynamics of SNNs and enables
the high-performance training. Now we quantitatively analyze
the impact of the TD item. The experiment configurations keep
the same with the previous section (784 − 400 − 10) and we
also report the testing results after training 200 epochs. Here
the existing BP in the SD is termed as SDBP.
TableVII records the simulation results. The testing accuracy of SDBP is lower than the accuracy of the STBP
on different dataset, which shows the time information is
beneficial for model performance. Specifically, compared to
the STBP, the SDBP has a 1.21% loss of accuracy on the
objective tracking dataset, which is 5 times larger than the loss
on the MNIST. And results also imply that the performance
of SDBP is not stable enough. In addition to the interference
of the dataset itself, the reason for this variation may be the
unstability of SNNs training. Actually, the training of SNNs
relies heavily on the parameter initialization, which is also
a great challenge for SNNs applications. In many reported
works, researchers usually leverage some special skills or
mechanisms to improve the training performance, such as
the lateral inhibition, regularization, normalization, etc. In
contrast, by using our STBP training method, much higher performance can be achieved on the same network. Specifically,
the testing accuracy of STBP reaches 98.48% on MNIST and
98.32% on the object detection dataset. Note that the STBP
can achieve high accuracy without using any complex training
skills. This stability and robustness indicate that the dynamics
in the TD fundamentally includes great potential for the SNNs
computing and this work indeed provides a new idea.
9
Fig. 6: Comparisons of different derivation approximation curves. (a) The impact of different approximations. (b) The
impact of different widths of regular approximation.
TABLE VII: Comparison for the SDBP model and the STBP model on different datasets.
Model
Spiking MLP
(SDBP)
Spiking MLP
(STBP)
Dataset
Objective tracking
MNIST
Objective tracking
MNIST
Network structure
784-400-10
784-400-10
784-400-10
784-400-10
Training skills
None
None
None
None
Mean
97.11%
98.29%
98.32%
98.48%
Accuracy
Interval∗
[96.04%,97.78%]
[98.23%, 98.39%]
[97.94%, 98.57%]
[98.42%, 98.51%]
* results with epochs [201,210].
IV. C ONCLUSION
In this work, a unified framework that allows supervised
training spiking neural networks just like implementing backpropagation in deep neural networks (DNNs) has been built
by exploiting the spatio-temporal information in the networks.
Our major contributions are summarized as follows:
1) We have presented a framework based on an iterative
leaky integrate-and-fire model, which enables us to
implement spatio-temporal backpropagation on SNNs.
Unlike previous methods primarily focused on its spatial
domain features, our framework further combines and
exploits the features of SNNs in both the spatial domain
and temporal domain;
2) We have designed the STBP training algorithm and
implemented it on both MLP and CNN architectures.
The STBP has been verified on both static and dynamic
datasets. Results have shown that our model is superior
to the state-of-the-art SNNs on relatively small-scale
networks of spiking MLP and CNNs, and outperforms
DNNs with the same network size on dynamic NMNIST dataset. An attractive advantage of our algorithm
is that it doesn’t need extra training techniques which
generally required by existing schemes, and is easier
to be implemented in large-scale networks. Results also
have revealed that the use of spatio-temporal complexity
to solve problems could fulfill the potential of SNNs
better;
3) We have introduced an approximated derivative to
address the non-differentiable issue of the spike activity.
Controlled experiment indicates that the steepness and
width of approximation curve would affect the model’s
performance and the key point for approximations is to
capture the nonlinear nature, while the specific shape is
not so critical.
Because the brain combines complexity in the temporal and
spatial domains to handle input information, we also would
like to claim that implementing STBP on SNNs is more bioplausible than applying BP on DNNs. The property of STBP
that doesn’t rely on too many training skills makes it more
hardware-friendly and useful for the design of neuromorphic
chip with online learning ability. Regarding the future research
topics, two issues we believe are quite necessary and very
important. One is to apply our framework to tackle more
problems with the timing characteristics, such as dynamic data
processing, video stream identification and speech recognition.
The other is how to accelerate the supervised training of large
scale SNNs based on GPUs/CPUs or neuromorphic chips. The
former aims to further exploit the rich spatio-temporal features
of SNNs to deal with dynamic problems, and the later may
greatly prompt the applications of large scale of SNNs in real
life scenarios.
R EFERENCES
[1] P. Chaudhari and H. Agarwal, Progressive Review Towards Deep Learning Techniques. Springer Singapore, 2017.
[2] L. Deng and D. Yu, “Deep learning: Methods and applications,” Foundations and Trends in Signal Processing, vol. 7, no. 3, pp. 197–387,
2014.
[3] Jia, Yangqing, Shelhamer, Evan, Donahue, Jeff, Karayev, Sergey, Long,
and Jonathan, “Caffe: Convolutional architecture for fast feature embedding,” Eprint Arxiv, pp. 675–678, 2014.
[4] G. Hinton, L. Deng, D. Yu, and G. E. Dahl, “Deep neural networks
for acoustic modeling in speech recognition: The shared views of four
research groups,” IEEE Signal Processing Magazine, vol. 29, no. 6, pp.
82–97, 2012.
10
[5] K. He, X. Zhang, S. Ren, and J. Sun, Spatial Pyramid Pooling in Deep
Convolutional Networks for Visual Recognition. Springer International
Publishing, 2014.
[6] X. Zhang, Z. Xu, C. Henriquez, and S. Ferrari, “Spike-based indirect
training of a spiking neural network-controlled virtual insect,” in Decision and Control (CDC), 2013 IEEE 52nd Annual Conference on.
IEEE, 2013, pp. 6798–6805.
[7] J. N. Allen, H. S. Abdel-Aty-Zohdy, and R. L. Ewing, “Cognitive
processing using spiking neural networks,” in IEEE 2009 National
Aerospace and Electronics Conference, 2009, pp. 56–64.
[8] N. Kasabov and E. Capecci, “Spiking neural network methodology for
modelling, classification and understanding of eeg spatio-temporal data
measuring cognitive processes,” Information Sciences, vol. 294, no. C,
pp. 565–575, 2015.
[9] B. V. Benjamin, P. Gao, E. Mcquinn, S. Choudhary, A. R. Chandrasekaran, J. M. Bussat, R. Alvarez-Icaza, J. V. Arthur, P. A. Merolla,
and K. Boahen, “Neurogrid: A mixed-analog-digital multichip system
for large-scale neural simulations,” Proceedings of the IEEE, vol. 102,
no. 5, pp. 699–716, 2014.
[10] P. A. Merolla, J. V. Arthur, R. Alvarezicaza, A. S. Cassidy, J. Sawada,
F. Akopyan, B. L. Jackson, N. Imam, C. Guo, and Y. Nakamura,
“Artificial brains. a million spiking-neuron integrated circuit with a
scalable communication network and interface.” Science, vol. 345, no.
6197, pp. 668–73, 2014.
[11] S. B. Furber, F. Galluppi, S. Temple, and L. A. Plana, “The spinnaker
project,” Proceedings of the IEEE, vol. 102, no. 5, pp. 652–665, 2014.
[12] T. Hwu, J. Isbell, N. Oros, and J. Krichmar, “A self-driving robot
using deep convolutional neural networks on neuromorphic hardware,”
arXiv.org, 2016.
[13] S. K. Esser, P. A. Merolla, J. V. Arthur, A. S. Cassidy, R. Appuswamy,
A. Andreopoulos, D. J. Berg, J. L. Mckinstry, T. Melano, and D. R.
Barch, “Convolutional networks for fast, energy-efficient neuromorphic
computing,” Proceedings of the National Academy of Sciences of the
United States of America, vol. 113, no. 41, p. 11441, 2016.
[14] S. S. Zhang, L.P. Shi, “Creating more intelligent robots through braininspired computing,” Science(suppl), vol. 354, 2016.
[15] P. U. Diehl and M. Cook, “Unsupervised learning of digit recognition
using spike-timing-dependent plasticity,” Frontiers in Computational
Neuroscience, vol. 9, p. 99, 2015.
[16] D. Querlioz, O. Bichler, P. Dollfus, and C. Gamrat, “Immunity to device
variations in a spiking neural network with memristive nanodevices,”
IEEE Transactions on Nanotechnology, vol. 12, no. 3, pp. 288–295,
2013.
[17] S. R. Kheradpisheh, M. Ganjtabesh, and T. Masquelier, “Bio-inspired
unsupervised learning of visual features leads to robust invariant object
recognition,” Neurocomputing, vol. 205, no. C, pp. 382–392, 2016.
[18] J. A. Perezcarrasco, B. Zhao, C. Serrano, B. Acha, T. Serranogotarredona, S. Chen, and B. Linaresbarranco, “Mapping from framedriven to frame-free event-driven vision systems by low-rate rate-coding
and coincidence processing. application to feed forward convnets.” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 35,
no. 11, pp. 2706–19, 2013.
[19] P. U. Diehl, D. Neil, J. Binas, and M. Cook, “Fast-classifying, highaccuracy spiking deep networks through weight and threshold balancing,” in International Joint Conference on Neural Networks, 2015, pp.
1–8.
[20] O. Peter, N. Daniel, S. C. Liu, D. Tobi, and P. Michael, “Real-time
classification and sensor fusion with a spiking deep belief network,”
Frontiers in Neuroscience, vol. 7, p. 178, 2013.
[21] E. Hunsberger and C. Eliasmith, “Spiking deep networks with lif
neurons,” Computer Science, 2015.
[22] P. O’Connor and M. Welling, “Deep spiking networks,” arXiv.org, 2016.
[23] J. H. Lee, T. Delbruck, and M. Pfeiffer, “Training deep spiking neural
networks using backpropagation,” Frontiers in Neuroscience, vol. 10,
2016.
[24] D. Neil, M. Pfeiffer, and S. C. Liu, “Phased lstm: Accelerating recurrent
network training for long or event-based sequences,” arXiv.org, 2016.
[25] F. A. Gers, J. Schmidhuber, and F. Cummins, “Learning to forget:
continual prediction with lstm,” Neural Computation, vol. 12, no. 10, p.
2451, 1999.
[26] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[27] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Gated feedback
recurrent neural networks,” Computer Science, pp. 2067–2075, 2015.
[28] P. J. Werbos, “Backpropagation through time: what it does and how to
do it,” Proceedings of the IEEE, vol. 78, no. 10, pp. 1550–1560, 1990.
[29] Y. Bengio, T. Mesnard, A. Fischer, S. Zhang, and Y. Wu, “An objective
function for stdp,” Computer Science, 2015.
[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998.
[31] E. Neftci, S. Das, B. Pedroni, K. Kreutzdelgado, and G. Cauwenberghs,
“Event-driven contrastive divergence for spiking neuromorphic systems,”
Frontiers in Neuroscience, vol. 7, p. 272, 2013.
[32] G. Orchard, A. Jayawant, G. K. Cohen, and N. Thakor, “Converting
static image datasets to spiking neuromorphic datasets using saccades,”
Frontiers in Neuroscience, vol. 9, 2015.
[33] P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128x128 120db 15us
latency asynchronous temporal contrast vision sensor,” IEEE Journal of
Solid-State Circuits, vol. 43, no. 2, pp. 566–576, 2007.
[34] D. Neil and S. C. Liu, “Effective sensor fusion with event-based sensors
and deep network architectures,” in IEEE Int. Symposium on Circuits and
Systems, 2016.
[35] G. K. Cohen, G. Orchard, S. H. Leng, J. Tapson, R. B. Benosman, and
A. V. Schaik, “Skimming digits: Neuromorphic classification of spikeencoded images,” Frontiers in Neuroscience, vol. 10, no. 184, 2016.
[36] P. Y. Simard, D. Steinkraus, and J. C. Platt, “Best practices for
convolutional neural networks applied to visual document analysis,” in
International Conference on Document Analysis and Recognition, 2003,
p. 958.
| 9cs.NE
|
1
Millimeter-Wave Beamformed Full-dimensional
MIMO Channel Estimation Based on Atomic
arXiv:1709.03843v1 [cs.SY] 12 Sep 2017
Norm Minimization
Yingming Tsai, Member, IEEE, Le Zheng, Member, IEEE,
and Xiaodong Wang, Fellow, IEEE
Electrical Engineering Dept., Columbia Univ., New York, NY 10027
Abstract
The millimeter-wave (mmWave) full-dimensional (FD) MIMO system employs planar arrays at both the
base station and user equipment and can simultaneously support both azimuth and elevation beamforming. In
this paper, we propose atomic-norm-based methods for mmWave FD-MIMO channel estimation under both
uniform planar arrays (UPA) and non-uniform planar arrays (NUPA). Unlike existing algorithms such as
compressive sensing (CS) or subspace methods, the atomic-norm-based algorithms do not require to discretize
the angle spaces of the angle of arrival (AoA) and angle of departure (AoD) into grids, thus provide much
better accuracy in estimation. In the UPA case, to reduce the computational complexity, the original largescale 4D atomic norm minimization problem is approximately reformulated as a semi-definite program (SDP)
containing two decoupled two-level Toeplitz matrices. The SDP is then solved via the alternating direction
method of multipliers (ADMM) where each iteration involves only closed-form computations. In the NUPA
case, the atomic-norm-based formulation for channel estimation becomes nonconvex and a gradient-decentbased algorithm is proposed to solve the problem. Simulation results show that the proposed algorithms
achieve better performance than the CS-based and subspace-based algorithms.
Keywords: Full-dimensional (FD) MIMO, uniform planar array (UPA), non-uniform planar array
(NUPA), atomic norm, channel estimation, millimeter-wave, alternating direction method of multipliers (ADMM), gradient descent.
2
I. I NTRODUCTION
Millimeter wave (mmWave) communications have been proposed as an important physical-layer
technology for the 5th generation (5G) mobile networks to provide multi-gigabit services [1]. Two
prominent features of the mmWave spectrum are the massive bandwidth available and the tiny
wavelengths compared to conventional microwave bands, thus enabling dozens or even hundreds of
antenna elements to be accommodated at communication link ends within a reasonable physical form
factor. This suggests that massive MIMO and mmWave technologies should be considered jointly to
provide higher data rates and spectrum efficiency. In particular, the mmWave full-dimensional MIMO
(FD-MIMO) systems [2],[3] employ uniform or non-uniform planar arrays at both the basestation
(BS) and user equipment (UE) and provide an extra degree of freedom in the elevation-angle domain.
Users can now be distinguished not only by their AoAs in the azimuth domain but also by their AoDs
in the elevation domain [4]. In this paper, we consider channel estimation for mmWave FD-MIMO
systems that simultaneously support both azimuth and elevation beamforming.
The mmWave band channel is significantly different from those in sub-6GHz bands. The key
challenge in designing new radio access technologies for mmWave is how to overcome the much
larger path-loss and reduce blockage probability. To that end, beamforming is essential in combating
the serve path-loss for wireless system operating in mmWave bands [5]. However, to estimate the
full channel state information (CSI) under beamformed FD-MIMO is somehow challenging because
the receiver typically only obtains the beamformed CSI instead of full CSI. To address this issue,
fast beam scanning and searching techniques have been extensively studied [3], [6]. The objective
of beam scanning is to search for the best beamformer-combiner pair by letting the transmitter and
receiver scan the adaptive sounding beams chosen from pre-determined sounding beam codebooks.
However, the exhaustive search may be hampered by the high training overhead in practice and suffer
from low spectral efficiency. Another approach is to estimate the mmWave channel or its associated
parameters, by exploiting the sparse scattering nature of the mmWave channels [7],[8],[9], that is,
mmWave channel estimation can be formulated as a sparse signal recovery problem [10], [11],
[12], [13] and solved using the compressive sensing (CS)-based approach [14]. In the CS-based
approach, a sensing matrix needs to be constructed first, by dividing certain parameter space into a
3
finite number of grids and thus the channel estimation performance is limited by the grid resolution.
On the other hand, in [15], a subspace-based mmWave MIMO channel estimation method that makes
use of the MUSIC algorithm is proposed. A 2D-MUSIC algorithm for beamformed mmWave MIMO
channel estimation is proposed in [12] to further enhance the channel estimation performance. The
MUSIC algorithm is able to identify multiple paths with high resolution but it is sensitive to antenna
position, gain, and phase errors.
Recently, the atomic norm minimization [16] has been applied to many signal processing problems
such as super-resolution frequency estimation [17], [18], spectral estimation [19], AoA estimation,
[20], [21], uplink multiuser MIMO channel estimation [22] and linear system identification [23].
Under certain conditions, atomic norm minimization can achieve exact frequency localization, avoiding the effects of basis mismatch which can plague grid-based CS techniques. Different from the
prior works such as CS-based and subspace-based channel estimation methods mentioned above,
we formulate the mmWave FD-MIMO channel estimation as an atomic norm minimization problem. Unlike [22] that considers uplink multiuser MIMO channel estimation, in which the uniform
linear array is assumed and only the AoA parameter is estimated, in this paper, we consider the
mmWave beamformed FD-MIMO channel, which involves the estimation of both AoA and AoD.
Hence, instead of one-dimensional (1D) atomic norm minimization, our problem is formulated as
a four-dimensional (4D) atomic norm minimization problem. The 4D atomic norm minimization
can be transformed into semi-definite program (SDP) which is of high dimensional and involves
block Toeplitz matrices, leading to very high computational complexity. Therefore, we introduce
a 4D atomic norm approximation method to reduce the computational complexity and an efficient
algorithm based on the alternating direction method of multipliers (ADMM) is derived.
Recently, non-uniform planar array (NUPA) has attracted more interest due to its ability in
reducing sidelobes and antenna correlation [24], [25]. NUPA can potentially increase the achievable
multiplexing gain of mmWave FD-MIMO beamforming. However, the corresponding atomic norm
minization problem cannot be transformed into an SDP when the antennas are not uniformly placed
[16]. Hence, we propose a gradient descent method for mmWave FD-MIMO channel estimation
with NUPA.
4
The remainder of the paper is organized as follows. In Section II, the mmWave beamformed
FD-MIMO channel model is introduced. In Section III, we formulate the mmWave FD-MIMO
channel estimation as an atomic norm minimization problem for the case of UPA. In Section IV, we
develop efficient algorithms for implementing the proposed atomic-norm-based channel estimator.
In Section V, we consider the case of NUPA and provide the formulation and algorithm for the
atomic-norm-based channel estimator. In Section VI, simulation results are provided. Finally, Section
VII concludes the paper.
II. S YSTEM D ESCRIPTIONS AND BACKGROUND
A. System and Channel Models
We consider a mmWave FD-MIMO system with M receive antennas and N transmit antennas that
simultaneously supports elevation and azimuth beamforming. The channel matrix can be expressed
in terms of transmit and receive array responses [9]:
H
H = BΣA =
L
X
σl b(fl )a(gl )H ,
(1)
l=1
where (·)H denotes the Hermitian transpose; the matrix Σ = diag(σ) = diag [σ1 σ2 . . . σL ]T is
a diagonal matrix with each σl ∈ C denoting the l-th multipath gain; L denotes the number of
paths; the matrices B = [b(f1 ) . . . b(fL )] and A = [a(g1 ) . . . a(gL )] denote the steering responses
of the receive and transmit arrays, respectively. For a linear array with half-wavelength separation
of adjacent antenna elements, the array response is in the form of a uniformly sampled complex
sinusoid with frequency x ∈ [− 21 ,
1
):
2
T
1
cn (x) = √ 1 ej2πx · · · ej2π(n−1)x ∈ Cn×1 .
n
(2)
We assume that both the transmitter (Tx) and receiver (Rx) are equipped with uniformly spaced
planar antenna arrays (UPA)s [26], [27], each with half-wavelength antenna element separations
along the elevation-and-azimuth-axis. Then the Tx and Rx array responses can be expressed as [27]
a(gl ) = cN1 (gl,1 ) ⊗ cN2 (gl,2 ) ,
(3)
b (fl ) = cM1 (fl,1 ) ⊗ cM2 (fl,2 ) ,
(4)
5
with
gl
fl
1
1
=
gl,1 = sin (ϑl ) cos (ϕl ) , gl,2 = cos (ϑl ) ,
2
2
1
1
=
fl,1 = sin (θl ) cos (φl ) , fl,2 = cos (θl ) ,
2
2
(5)
(6)
where ⊗ denotes the Kronecker product; ϑl , ϕl denote elevation and azimuth angles of the angle
of departure (AoD) of the l-th path, respectively; and θl , φl denote elevation and azimuth angles of
the angle of arrival (AoA), respectively. Here, N1 , N2 denote the numbers of elevation and azimuth
transmit antennas, respectively, and the total number of transmit antennas is N = N1 N2 . Similarly,
M1 , M2 denote the numbers of elevation and azimuth receive antennas, respectively, and the total
number of receive antennas is M = M1 M2 . For the UPA configuration, it can resolve the AoA
and AoD in 360◦ range, thereby ϑl , θl , ϕl , φl ∈ [−π, π] and gl,1 =
gl,2 = 12 cos (ϑl ) ∈ [− 12 ,
1
),
2
1
2
sin (ϑl ) cos (ϕl ) ∈ [− 12 ,
1
),
2
fl,1 = 12 sin (θl ) cos (φl ) ∈ [− 21 , 12 ), fl,2 = 12 cos (θl ) ∈ [− 12 , 21 ).
To estimate the channel matrix, the transmitter transmits P distinct beams during P successive
time slots. i.e., in the p-th time slot, the beamforming vector pp ∈ CN ×1 is selected from a set
of unitary vectors in the form of Kronecker-product-based codebook, e.g., pp = pp,1 ⊗ pp,2 where
pp,1 ∈ CN1 and pp,2 ∈ CN2 are selected from two DFT codebooks of dimensions N1 and N2 ,
respectively [28]. The p-th received signal vector can be expressed as
yp = Hpp sp + wp ,
(7)
where wp ∼ CN (0, σw2 IM ) is the additive white Gaussian noise (AWGN) with IM denoting the
M × M identity matrix, and sp denotes the pilot symbol in the p-th time slot. The receiver collects
yp ∈ CM ×1 for p = 1, . . . P and concatenates them to obtain the signal matrix
Y = [y1 y2 . . . yP ] = HPS + W = BΣAH PS + W,
(8)
where P = [p1 p2 . . . pP ] ∈ CN ×P , W = [w1 w2 . . . wP ] ∈ CM ×P and S = diag ([s1 s2 . . . sP ]) ∈
√
CP ×P . For simplicity, we assume that S = Pt IP , where Pt is the power of the pilot symbol. Then
we have
Y=
p
Pt HP + W =
p
Pt BΣAH P + W.
(9)
6
Our goal is to estimate the channel matrix H ∈ CM ×N from the measurements Y ∈ CM ×P . Note
that the number of pilots is usually smaller than the number of transmit antennas, i.e., P < N .
Hence, we need to exploit the sparsity of H for its estimation, which will be discussed in the next
section.
B. Existing mmWave Channel Estimation Methods
Before describing our proposed mmWave channel estimator, we briefly discuss some existing
mmWave channel estimation methods [10], [29], [13], [12] which can be divided into two categories.
1) CS-based mmWave channel estimators: The mmWave channel is usually composed of small
number of propagation paths and CS-based algorithms have been developed [10], [29], [13] for
channel estimation. First the dictionary matrices AD ∈ CN ×NG and BD ∈ CM ×NG are constructed
based on quantized AoD angle of the transmitter and AoA angle of the receiver. The AoDs and
AoAs are assumed to be taken from a uniform grid of NG points with NG L. The resulting
dictionary matrix is expressed (take the transmitter AD for example, the receiver dictionary matrix
BD is similar.)
AD = [a(ḡ1 ) a(ḡ2 ) . . . a(ḡNG )] ,
where ḡi = {ḡi,1 , ḡi,2 } = 21 sin ϑ̄i cos(ϕ̄i ), 21 cos(ϑ̄i ) with ϑ̄i =
(10)
(i−1)2π
NG
− π, ϕ̄i =
(i−1)2π
NG
−π
denotes the transmit array response vector for the grid point ϑ̄i and ϕ̄i for i = 1, 2, . . . , NG . The
size NG of the angle grids can be set according to the desired angular resolution. On this basis, the
received signal Y in (9) can be vectorized as [10]
p
y = vec (Y) = Pt PT ⊗ IM vec (H) + w
p
p
=
Pt PT ⊗ IM (A∗D BD )x + w = Pt Gx + w,
where
(11)
(12)
denotes the matrix Khatri-Rao products, (·)T denotes the transpose operation, (·)∗ denotes
4
the complex conjugate, x ∈ CNG is a sparse vector that has non-zero elements in the locations
associated with the dominant paths. Note that the angle spaces of interest are discretized into a large
number of grids, and the actual AoA and AoD angles may not exactly reside on the predefined
grids. Those off-grid angles can lead to mismatches in the channel model and degrade the estimation
performance.
7
2) Subspace-based mmWave channel estimators: Another existing approach to mmWave channel
estimation is based on the subspace methods such as the MUSIC algorithm [12]. The MUSIC
algorithm firstly calculates the covariance matrix of the received signal Y and then finds the signal
and noise subspaces via eigendecomposition. It then estimates each channel path’s array response,
i.e., ĝl and f̂l for l = 1, 2, . . . , L̂, where L̂ is the estimated number of paths, by exploiting the
orthogonality between the signal and noise subspaces. Finally, each channel path’s coefficient, i.e.,
σ̂l can be estimated via the least-squares (LS) method. The MUSIC algorithm has been popular for
its good resolution and accuracy in AoD/AoA estimation [30], [31]. However, it is also reported that
the off-grid CS method [16] can outperform the MUSIC algorithm in terms of estimation accuracy
in noisy environments [19], [32].
III. C HANNEL E STIMATION VIA ATOMIC N ORM M INIMIZATION
As explained in the previous section, the performance of the mmWave channel estimators based
on on-grid methods such as CS can be degraded due to grid mismatch. In this section, we propose a
new mmWave channel estimator based on an off-grid CS method, i.e., the atomic norm minimization
method.
A. Background on Multi-dimensional Atomic Norm
First we briefly introduce the concept of multi-dimensional atomic norm [33]. A d-dimensional
(d-dim) atom is defined as
qd (x1 , . . . , xd ) = cn1 (x1 ) ⊗ . . . ⊗ cnd (xd ) ,
(13)
where ni is the length of the normalized vector cni (xi ) defined in (2) and xi ∈ − 21 , 12 for i =
1, 2, . . . , d. The d-dim atomic set is then given by
1 1
A = qd (x1 , . . . , xd ) : xi ∈ − ,
, i = 1, . . . , d .
2 2
(14)
8
For any vector td of the form td =
P
αl qd (xl,1 , xl,2 , . . . , xl,d ), its d-dim atomic norm with respect
l
to A is defined as
ktd kA = inf {t : td ∈ tconv (A)} ,
(
)
X
X
=
inf
|αl | td =
αl qd (xl,1 , xl,2 , . . . , xl,d ) ,
1 1
xl,1 ,xl,2 ,...,xl,d ∈[− 2 , 2 )
αl ∈C
l
(15)
l
where conv (A) is the convex hull of A. The d-dim atomic norm of td has following equivalent
form [33]:
1
1
Tr
(T
(U
))
+
t
d
d
2n1 n2 ...nd
2
,
ktd kA =
inf
Ud ∈C(2nd −1)×(2nd−1 −1)×...×(2n1 −1) ,t∈R
T
(U
)
t
d
d
d
0
s.t.
H
t
t
(16)
d
where Tr (·) is the trace of the input matrix, Ud ∈ C(2nd −1)×(2nd−1 −1)×...×(2n1 −1) is a d-way tensor
and Td (Ud ) is a d-level block Toeplitz, which is defined recursively as follows. Denote nd =
(nd , nd−1 , . . . , n1 ) and Ud−1 (i) = Ud (i, :, :, ..., :) for i = −nd + 1, −nd + 2, ..., nd − 1. For d = 1,
n1 = (n1 ) and T1 (u1 ) = Toep(u1 ) with u1 ∈ C(2n1 −1)×1 , i.e.,
u1 (0)
u1 (1)
u1 (−1)
u1 (0)
T1 (u1 ) = Toep(u1 ) =
..
..
.
.
u1 (1 − n1 ) u1 (2 − n1 )
For d ≥ 2, we have
Td (Ud ) =
Td−1 (Ud−1 (0))
Td−1 (Ud−1 (1))
· · · u1 (n1 − 1)
· · · u1 (n1 − 2)
.
..
..
.
.
···
u1 (0)
. . . Td−1 (Ud−1 (nd − 1))
(17)
Td−1 (Ud−1 (−1))
Td−1 (Ud−1 (0))
. . . Td−1 (Ud−1 (nd − 2))
.
..
..
..
...
.
.
.
Td−1 (Ud−1 (1 − nd )) Td−1 (Ud−1 (2 − nd )) . . .
Td−1 (Ud−1 (0))
(18)
9
B. Atomic Norm Minimization Formulation
In this subsection, we formulate the atomic norm minimization problem for channel estimation.
First, we vectorize the mmWave FD-MIMO channel matrix H in (1) as
L
X
h = vec(H) =
σl a(gl )∗ ⊗ b(fl )
l=1
=
L
X
∗
σl cN1 (gl,1 ) ⊗ cN2 (gl,2 ) ⊗ cM1 (fl,1 ) ⊗ cM2 (fl,2 )
l=1
=
L
X
σl c∗N1 (gl,1 ) ⊗ c∗N2 (gl,2 ) ⊗ cM1 (fl,1 ) ⊗ cM2 (fl,2 ).
(19)
l=1
Comparing (15) and (19), for the mmWave FD-MIMO channel with UPA configuration, the atom
has the form of
q4 (g, f ) = c∗N1 (g1 ) ⊗ c∗N2 (g2 ) ⊗ cM1 (f1 ) ⊗ cM2 (f2 ),
(20)
and the set of atoms is defined as the collection of all normalized 4D complex sinusoids: A =
q4 (g, f ) : f ∈ [− 12 , 12 ) × [− 21 , 12 ), g ∈ [− 12 , 12 ) × [− 21 , 12 ) [34], [35]. The 4D atomic norm for
any h defined in (19) can be written as [34]:
(
khkA =
fl ∈[− 21 ,
gl ∈[− 12 ,
)
X
inf
1
)×[− 12 ,
2
1
)×[− 21 ,
2
σl ∈C
1
),
2
1
),
2
l
|σl | h =
X
σl q4 (gl , fl ) .
(21)
l
The atomic norm can enforce sparsity in the atom set A. On this basis, an optimization problem will
be formulated for the estimation of the path frequencies {fl , gl }. For the convenience of calculation,
we will use the equivalent form of the atomic norm given by (16), i.e.,
1
1
Tr
(T
(U
))
+
t
4
4
2M N
2
khkA =
inf
,
U4 ∈C(2N1 −1)×(2N2 −1)×(2M1 −1)×(2M2 −1) ,
T
(U
)
h
4
4
t∈R
0
s.t.
H
h
t
(22)
where T4 (U4 ) is a 4-level Toeplitz matrix defined in (18). Define the minimum frequency separations
as
∆min,fi = min0 min{|fl,i − fl0 ,i |, 1 − |fl,i − fl0 ,i |},
(23)
∆min,gi = min0 min{|gl,i − gl0 ,i |, 1 − |gl,i − gl0 ,i |},
(24)
l6=l
l6=l
10
for i = 1, 2. To show the connection between the atomic norm and the channel matrix, we obtain
the following theorem via extending Theorem 1.2 in [36] for 1D atomic norm to 4D atomic norm.
Theorem 1. If the path component frequencies are sufficiently separated, i.e.,
for i = 1, 2, then we have khkA =
1
,
b(Mi − 1)/4c
1
≥
,
b(Ni − 1)/4c
∆min,fi ≥
(25)
∆min,gi
(26)
P
l
|σl |, so the component atoms of h can be uniquely located
via computing its atomic norm.
The proof follows the same line as that in Theorem 1.2 [36], with the dual polynomial constructed
by interpolation with a 4D kernel. The theorem holds because all bounds in the proof of [Theorem
1.2, 34] hold by leveraging the 1D results.
To estimate the mmWave FD-MIMO channel H in (1) based on the signal Y in (9), we then
formulate the following optimization problem:
p
2
1
y − Pt PT ⊗ IM h ,
(27)
2
2
h∈CM N
p
where y = vec(Y) is given by (11) and µ ∝ σw M N log (M N ) is a weight factor [37]. Using
ĥ = min µkhkA +
(22), (27) can be equivalently formulated as a semi-definite program (SDP):
p
µ
1
µ
min
Tr (T4 (U4 )) + t +
y − Pt PT ⊗ IM h
2M N
2
2
U4 ∈C(2N1 −1)×(2N2 −1)×(2M1 −1)×(2M2 −1) ,
2
2
h∈CM N , t∈R
s.t
T4 (U4 ) h
h
H
t
0.
(28)
The above problem is convex, and can be solved by using a standard convex solver. Suppose the
−1
solution to (28) is ĥ. Then the estimated channel matrix is given by Ĥ = vec
ĥ where vec−1 (·)
is the inverse operation of vec(·).
IV. E FFICIENT A LGORITHM FOR C HANNEL E STIMATION UNDER UPA
A. A Formulation Based on 2D MMV Atomic Norm
Note that the dimension of the positive semidefinite matrix in (28) is (M N + 1) × (M N + 1),
and the 4D atomic norm minimization formulation is of high computational complexity and has
11
large memory requirements. To reduce the complexity, we can treat Y as 2D multiple measurement
vectors (MMV) [32] in transmit and receive dimensions.
Unlike the 4D atomic norm that is calculated with input vector h, the MMV atomic norm is
calculated with the matrix input H. Specifically, we define the atom Q̄ (f , ā) = b (f ) āH with
f ∈ [− 12 ,
1
)
2
and ā ∈ CN ×1 with kāk2 = 1. Correspondingly, the atom set is defined as
1 1
1 1
(29)
AMMV = Q̄ (f , ā) : f ∈ [− , ) × [− , ), kāk2 = 1 .
2 2
2 2
× [− 21 ,
1
),
2
It is worth noting that ā is not restricted by the structural constraint in (3). With (29), we extend
the 1D MMV atomic norm [32] to the 2D MMV atomic norm of H defined by
)
(
X
X
kHkAMMV =
inf 1 1
σl Q̄ (fl , āl ) , kāk2 = 1 .
|σl | H =
1 1
fl ∈[− 2 , 2 )×[− 2 , 2 ),
āl ∈CN ×1 , σl ∈C
(30)
l
l
This atomic norm is equivalent to the solution of the following SDP [32]:
1
1
Tr (T2 (U2 )) + 2N Tr (X)
2M
,
kHkAMMV =
inf
U2 ∈C(2M2 −1)×(2M1 −1) ,X∈CN ×N
T2 (U2 ) H
0
s.t.
H
H
X
(31)
where X is constrained to be a Hermitian matrix. Then using (9), we can formulate the following
optimization problem for channel estimation:
Ĥ =
min µkHkAMMV +
H∈CM ×N
1 p
Pt HP − Y
2
2
,
(32)
F
where k·kF denotes matrix Frobenius norm. Plugging (31) into (32), the size of the positive semidefinite matrix in the constraint is (M + N ) × (M + N ), resulting in considerably lower computational
complexity and memory requirement than (28).
B. An Approximation to 4D Atomic Norm Minimization
Next we propose an approximation to the 4D atomic norm to reduce the computational complexity.
In [38], the authors explore the approximation of 2D atomic norm to improve the efficiency. Here,
we extend the results from 2D atomic norm to 4D atomic norm case. Similar to the 2D MMV
atomic norm, the proposed approximation is calculated with input H. From (1), H is the sum of
12
σl b(fl )a(gl )H , in which both a(gl ) and b(fl ) are Fourier bases. Different from the vectorized atomic
norm, we introduce the matrix atom Q (f , g) = b(f )a(g)H and the matrix atom set
1 1
1 1
1 1
1 1
H
AM = Q (f , g) = b(f )a(g) : f ∈ [− , ) × [− , ), g ∈ [− , ) × [− , ) .
2 2
2 2
2 2
2 2
(33)
The matrix atomic norm is then given by
(
kHkAM =
fl ∈[− 21 ,
gl ∈[− 12 ,
inf
1
)×[− 12 ,
2
1
)×[− 12 ,
2
σl ∈C
)
X
1
),
2
1
),
2
|σl | H =
l
X
σl Q (fl , gl ) .
(34)
l
The matrix atom set is composed of rank-one matrices, and hence it amounts to atomic norm of
low rank matrices. Since the operator vec(·) is a one-to-one mapping and the mapping AM → A
is also one-to-one, it is straightforward to conclude that kHkAM = khkA . Hence, if the component
frequencies satisfy the sufficient separation condition given by (25) and (26), we have kHkAM =
P
l |σl | by Theorem 1.
Finding the harmonic components via atomic norm is an infinite programming problem over all
feasible f and g, which is difficult. For better efficiency, we use SDP(H) in the following Lemma
to approximate kHkAM .
Lemma 1. For H given by (1), we have kHkAM ≥ SDP(H) ≥ kHkAMMV , where
1
1
Tr (T2 (U2 )) + 2N Tr (T2 (V2 ))
2M
,
SDP(H) ,
inf
U2 ∈C(2M2 −1)×(2M1 −1) ,V2 ∈C(2N2 −1)×(2N1 −1)
T2 (U2 )
H
0
s.t.
H
H
T (V )
2
(35)
2
with T2 (U2 ) and T2 (V2 ) being 2-level Toeplitz matrices defined in (18).
Proof. The relation SDP (H) ≥ kHkAMMV can be directly obtained from the definitions in (31) and
(35). It remains to show kHkAM ≥ SDP(H). Denote
1
ã(gl , ωl ) = √ ej2πωl c∗N1 (gl,1 ) ⊗ c∗N2 (gl,2 ),
N
1 j2πχl
b̃(fl , χl ) = √ e
cM1 (fl,1 ) ⊗ cM2 (fl,2 ),
M
13
with ωl ∈ [0, 2π] and χl ∈ [0, 2π] such that σl = |σl | ej2π(ωl +χl ) . For any H =
P
l
σl b (fl ) a (gl )H , if
we set
U2 = [u1 (−M1 + 1), u1 (−M1 + 2), ..., u1 (M1 − 1)],
(36)
V2 = [v1 (−N1 + 1), v1 (−N1 + 2), ..., v1 (N1 − 1)],
(37)
where
1 X
(38)
u1 (i) = √
|σl |c̃M2 (fl,2 )ej2π(i−1)fl,1 ,
M l
1 X
(39)
v1 (i) = √
|σl |c̃∗N2 (gl,2 )e−j2π(i−1)gl,1 ,
N l
T
with c̃n (x) = √1n ej2π(1−n)x , ej2π(2−n)x , · · · , ej2π(n−1)x ∈ C2n×1 , then the 2-level Toeplitz matrices
T2 (U2 ) and T2 (V2 ) satisfy
T2 (U2 ) =
=
X
l
X
|σl |b (fl ) b (fl )H
|σl |b̃ (fl , χl ) b̃ (fl , χl )H ,
(40)
l
T2 (V2 ) =
=
X
l
X
|σl |a (gl ) a (gl )H
|σl |ã (gl , ωl ) ã (gl , ωl )H .
(41)
l
Moreover, the matrix
M=
T2 (U2 )
H
HH
T2 (V2 )
=
X
l
H
b̃ (fl , χl )
b̃ (fl , χl )
|σl |
ã (gl , ωl ) ã (gl , ωl )
(42)
is positive semidefinite, indicating that the constraints in (35) are satisfied. Note that SDP(H) ≤
P
1
1
Tr (T2 (U2 )) + 2N
Tr (T2 (V2 )) = l |σl | according to the definition in (35). Since this holds for
2M
any decomposition of H, we obtain SDP (H) ≤ kHkAM .
The above lemma shows that SDP(H) is a lower bound of the matrix atomic norm. Moreover, the
following lemma states that if the component frequencies are sufficiently separated, then SDP(H)
is equivalent to kHkAM .
Lemma 2. If (25)-(26) hold, then kHkAM = SDP(H).
14
Figure 1: The approximation errors kHkAM − SDP(H) and kHkAM − kHkAMMV when the
separations satisfy ∆min,fi ≥ δ(Mi − 1), ∆min,gi ≥ δ/(Ni − 1), Ni = Mi = 16, for i = 1, 2.
The simulations are run 100 times for each δ.
Proof. First it follows from Theorem 4 in [32] that if (25)-(26) hold, then we have kHkAMMV =
P
l |σl |. Using Theorem 1 and the fact that khkA = kHkAM , we have kHkAM = kHkAMMV . Finally
by Lemma 1 we have kHkAM = kHkAMMV = SDP(H).
When the sufficient separation condition given by (25) and (26) is not satisfied, SDP(H) may
not be the same as kHkAM . However, it is found via simulations that SDP(H) still provides a good
approximation to kHkAM and usually results in good performance in channel estimation. Moreover,
as shown by Lemma 1, SDP(H) is a lower bound of the atomic norm kHkAM (or khkA equivalently),
i.e., khkA = kHkAM ≥ SDP(H) in general.
To show the approximation performances of both kHkAMMV and SDP(H) to khkA , we perform a
series of Monte Carlo trials for parameters M1 = M2 = 16, N1 = N2 = 16 with L = 2. fl and gl take
random values from [− 12 ,
1
)
2
× [− 12 ,
1
)
2
such that the separations satisfy ∆min,fi ≥ δ/ b(Ni − 1)c,
∆min,gi ≥ δ/ b(Ni − 1)c with 1 ≤ δ ≤ 6. In Fig. 1, we plot the approximation error against δ and
the bars show 95% confidence interval. As δ decreases, both approximation errors become larger.
However, SDP(H) provides a more accurate approximation than kHkAMMV . When δ ≥ 4, both
approximation errors become zero.
15
Therefore, instead of solving the original 4D atomic norm minimization in (28), we can solve the
following SDP
Ĥ =
µ
µ
1 p
Tr (T2 (U2 )) +
Tr (T2 (V2 )) +
Pt HP − Y
2M
2N
2
min
H∈CM ×N ,
2
(43)
F
U2 ∈C(2M2 −1)×(2M1 −1) ,
V2 ∈C(2N2 −1)×(2N1 −1)
s.t.
M=
T2 (U2 )
H
H
H
T2 (V2 )
0.
The size of the positive semidefinite matrix in the constraint is (M + N ) × (M + N ), resulting in
considerably lower computational complexity and memory requirement than (28).
C. ADMM for Approximate 4D Atomic Norm Minimization
To meet the requirement of real-time signal processing, we next derive an iterative algorithm for
solving the SDP in (43), based on the alternating direction method of multipliers (ADMM) [39]. To
put our problem in an appropriate form for ADMM, rewrite (43) as
arg min
H∈CM ×N ,
U2 ∈C(2M2 −1)×(2M1 −1) ,
V2 ∈C(2N2 −1)×(2N1 −1)
γ
γ
1
kHP − Yk2F +
Tr (T2 (U2 )) +
Tr (T2 (V2 )) + I∞ (M 0),(44)
2
2M
2N
where I∞ (z) is an indicator function that is 0 if z is true, and infinity otherwise. Dualize the equality
constraint via an augmented Lagrangian, we have
Lρ (U2 , V2 , H, Υ, M) =
γ
γ
1
Tr (T2 (U2 )) +
Tr (T2 (V2 )) + kHP − Yk2F + I∞ (M 0)
2M
2N
+ 2
*
T2 (U2 )
H
+ Υ, M −
H
H
T2 (V2 )
2
T2 (U2 )
H
ρ
,
+ M−
(45)
2
HH
T (V )
2
2
F
16
where Υ is the dual variable, hΥ, Mi , Re Tr(MH Υ) , ρ > 0 is the penalty parameter. The
ADMM consists of the following update steps:
l+1
l+1
(Ul+1
) = arg
2 , V2 , H
Lρ (U2 , V2 , H, Υl , Ml ),
min
H∈CM ×N ,
(46)
U2 ∈C(2M2 −1)×(2M1 −1) ,
V2 ∈C(2N2 −1)×(2N1 −1)
l+1
l+1
, Υl , M),
Lρ (Ul+1
2 , V2 , H
l+1
l+1
T2 (U2 )
H
.
= Υl + ρ Ml+1 −
l+1
l+1 H
(H )
T2 (V2 )
Ml+1 = arg
min
M∈C(M +N )×(M +N ) 0
Υl+1
(47)
(48)
Now we derive the updates of (46) and (47) in detail. For convenience, the following partitions
are introduced:
Ml =
Υl =
Ml0
Ml2
(Ml2 )H
Ml1
Υl0
Υl2
(Υl2 )H
Υl1
,
(49)
,
(50)
where Ml0 and Υl0 are M × M matrices, Ml2 and Υl2 are M × N matrices, Ml1 and Υl1 are N × N
matrices. Computing the derivative of Lρ (U2 , V2 , H, Υ, M) with respect to H, U2 and V2 , we have
∇H Lρ = (HP − Y)PH − 2Υl2 + 2ρ(H − Ml2 ),
(51)
γ2 + M1 ρU2 (i, k) − Tr(ρMl0 + Υl0 ), i = k = 0,
∇U2 (i,k) Lρ =
(52)
M2P
−i−1
(1)
(M1 − i)(M2 − k)ρU2 (i, k) −
Trk Si,k (ρMl0 + Υl0 ) , i 6= 0 or k 6= 0,
m=0
γ
2 + N1 ρV2 (i, k) − Tr(ρMl1 + Υl1 ), i = k = 0,
∇V2 (i,k) Lρ =
(53)
N2P
−i−1
(2)
(N1 − i)(N2 − k)ρV2 (i, k) −
Trk Si,k (ρMl1 + Υl1 ) , i 6= 0 or k 6= 0,
m=0
where U2 (i, k) and V2 (i, k) are the (i, k)-th elements of U2 and V2 , respectively. For X ∈ CM ×M ,
(1)
(2)
Si,k (X) returns the (i, k)-th M1 × M1 submatrix Xi,k . For X ∈ CN ×N , Si,k (X) returns the (i, k)-th
N1 × N1 submatrix Xi,k . Trk (·) outputs the trace of the k-th sub-diagnal of the input matrix. Tr0 (·)
outputs the trace of the input matrix.
17
By setting the derivatives to 0, Hl+1 , Ul+1
and V2l+1 can be updated by:
2
Hl+1 = (YPH + 2ρMl2 + 2Υl2 )(PPH + 2ρIN )−1 ,
γ
= T∗2 (Ml0 + Υl0 /ρ) −
Ul+1
e1 ,
2
2M ρ
γ
e1 ,
V2l+1 = T∗2 (Ml1 + Υl1 /ρ) −
2N ρ
(54)
(55)
(56)
where e1 = [1, 0, 0, ..., 0]T , T∗2 (·) denotes the adjoints of the map T2 (·). Specifically, suppose Z =
T∗2 (X) where Z = [z−M2 +1 , z−M2 +2 , ..., zM2 −1 ] with zi = [zi (−M1 +1), zi (−M1 +2), ..., zi (M1 −1)]T
when X ∈ CM ×M . Then we have
MX
1 −i−1
1
(1)
zi (k) =
Trk (Si,m (X)),
(M1 − i)(M2 − k) m=0
(57)
for i = −M2 + 1, −M2 + 2, ..., M2 − 1 and k = −M1 + 1, −M1 + 2, ..., M1 − 1.
The update of M is given by
Ml+1 = arg
M − M̃l+1
min
2
,
(58)
− Υl+1 /ρ.
(59)
M∈C(M +N )×(M +N ) 0
F
where
M̃l+1 =
T2 (Ul+1
2 )
Hl+1
(Hl+1 )H
T2 (V2l+1 )
It is equivalent to projecting M̃l+1 onto the positive semidefinite cone. Specifically, the projection is
accomplished by setting all negative eigenvalues of M̃l+1 to zero. Note that in ADMM the update
of variables H, U2 , V2 and M are in closed-form. Compared to the off-the-shelf solvers such as
SeDuMi [40] and SDPT3 [41], whose computational complexity is O ((M + N )6 ) in each iteration,
the complexity of ADMM is O ((M + N )3 ) in each iteration, so it runs much faster.
V. T HE G ENERAL P LANAR A RRAY C ASE
So far we have focused on the uniform planar array (UPA). For mmWave beamformed FD-MIMO,
because of the larger average inter-antenna element spacing, non-uniform planar array (NUPA)
requires fewer elements than UPA, whereby reducing the weight and cost of the system in large
array applications. Also, the irregular spacing allows the antenna grid spacing to become larger
than a half wavelength so it can effectively reduce the channel correlation and enhance multiplexing
18
gain [42]. Furthermore, there is a fundamental limitation of UPA, namely, the lower resolution of
elevation AoA, which limits the UPA performance [24].
In this section we consider the beamformed mmWave FD-MIMO channel estimation for NUPA.
Define dt =
2
λ
[(dt,1 (1), dt,2 (1)) . . . (dt,1 (N ), dt,2 (N ))] as the normalized transmit antenna locations,
where (dt,1 (i), dt,2 (i)) is the i-th transmit antenna coordinate in a 2D planar surface. Similarly,
dr =
2
λ
[(dr,1 (1) , dr,2 (1)) . . . (dr,1 (M ) , dr,2 (M ))] is the normalized receive antenna locations where
(dr,1 (i), dr,2 (i)) is the i-th receive antenna coordinate in a 2D planar surface. Then the steering
responses of the transmit and receive arrays for the l-th path can be respectively written as [43]
"
#T
2dt,2 (1)
2dt,1 (N )
2dt,2 (N )
2dt,1 (1)
g
+
g
j2π
g
+
g
j2π
1
l,1
l,2
l,1
l,2
λ
λ
λ
λ
adt (gl ) = √
···e
e
,
(60)
N
"
#T
2dr,1 (1)
2dr,2 (1)
2dr,2 (M )
2dr,1 (M )
j2π
f
+
f
f
+
f
j2π
1
l,1
l,2
l,1
l,2
λ
λ
λ
λ
e
bdr (fl ) = √
···e
.
(61)
M
With (60) and (61), the channel matrix H of NUPA is given by (1) with array responses a (gl ) and
b (fl ) replaced by adt (gl ) and bdr (fl ), respectively.
The atom for NUPA is then defined as
qNU (g, f ) = a∗dt (g) ⊗ bdr (f ) .
(62)
And the atom set for NUPA is given by
1 1
−1 1
−1 1
−1 1
ANU , qNU (g, f ) , g ∈ [ , ) × [− , ), f ∈ [ , ) × [ , ) .
2 2
2 2
2 2
2 2
The atomic norm khkANU for any h = vec (H) is then given by
(
)
X
X
khkANU =
inf 1 1
|σl | h =
σl qNU (gl , fl ) .
1 1
fl ∈[− 2 ,
gl ∈[− 21 ,
)×[− ,
),
2
2 2
1
)×[− 12 , 21 ),
2
σl ∈C
l
(63)
(64)
l
To estimate the channel, we propose to solve the following optimization problem
min µ khkANU +
h
p
1
y − Pt PT ⊗ IM h
2
2
.
(65)
2
Note that the atom defined in (62) is not based on uniform sampling, and consequently the atomic
norm in (64) does not have the equivalent SDP form as in (28) or (43). Hence, (65) cannot be solved
19
via convex optimization. According to Corollary 2.1 of [44], (65) shares the same optimum as the
following optimization problem
L
fl ∈[− 12 ,
gl ∈[− 12 ,
min
1
p
X
1
T
P
Γ
({g
y
−
qNU (gl , fl ) σl
t P ⊗ IM
l , fl , σl }) = µ kσk1 +
1
2
),
2
l=1
)×[− 21 ,
2
1
)×[− 12 , 21 ),
2
σl ∈C
2
.
(66)
2
Since the problem given by (66) is nonconvex, we will employ a gradient-descent algorithm to
obtain its local optimum. In practice, L is unknown, so we initialize q (gl , fl ) on L̃0 grid points
such that L ≤ L̃0 ≤ M P , where P is the number of training beams defined in (8). For example, let
0
0
each gl and fl be taken from a uniform grid of NG points with L̃0 = NG4 ≤ M P , i.e., gl,i
and fl,i
are uniformly taken from [−1/2, 1/2) for i = 1, 2 and 1 ≤ l ≤ NG4 , where the supercript 0 indicates
T
iteration 0, i.e., initialization. Let Ω0 = (gl0 , fl0 )1≤l≤L̃ . The initial value of σ 0 = σ10 . . . σL̃0 can
then be obtained by the least-squares (LS) estimate
σ0 =
where
†
PT ⊗ IM
†
qNU g10 , f10 . . . qNU gL̃0 , fL̃0
y,
(67)
indicates the pseudo inverse of the matrix. Then the gradient descent method is used to
find the local optimum. We use superscript k to denote the quantities in the k-th iteration. Then the
gradient descent search proceeds as follows
k
1
gl,i − κk ∇gl,i Γ {glk , flk , σlk } −2 1 ,
2
k
21
k
k k
k
= fl,i − κ ∇fl,i Γ {gl , fl , σl } − 1 ,
2
k
k
k k
k
= σl − κ ∇σl Γ {gl , fl , σl } ,
k+1
gl,i
=
(68)
k+1
fl,i
(69)
σlk+1
(70)
for l = 1, . . . , L̃k and i = 1, 2, where κk is the step size that can be obtained via Armijo line search
[45] and [x]ab defines the operator that outputs x =
mod (x, a) when x < b, and outputs x =
mod (x, b) when x > a, mod (a, b) defines the modulo operator. Specifically, in the k-th iteration,
κk is initialized as κk = κ̄. If Γ {glk+1 , flk+1 , σlk+1 } ≥ Γ {glk , flk , σlk } , then κk is updated by
20
multiplication with a constant 0 < α < 1, i.e., κk ← ακk . The gradients are calculated
H
L̃
X
∇gl,i Γ ({gl , fl , σl }) = R σl P̄
qNU (gl , fl ) σl − y P̄ a∗dt,i (gl ) ⊗ bdr (fl )
, (71)
l=1
H
L̃
X
∇fl,i Γ ({gl , fl , σl }) = R σl P̄
qNU (gl , fl ) σl − y P̄ a∗dt (gl ) ⊗ bdr,i (fl )
, (72)
l=1
T
L̃
X
∗
1
σl
∇σl Γ ({gl , fl , σl }) = µ
+ P̄
qNU (gl , fl ) σl − y P̄qNU (gl , fl ) ,
(73)
2|σl | 2
l=1
where R {·} returns the real part of the input,
p
Pt PT ⊗ IM ,
P̄ =
j2π
T
[dt,i (1), . . . , dt,i (N )]
◦ adt (gl ) ,
adt,i (gl ) =
λ
j2π
T
[dr,i (1), . . . , dr,i (M )]
◦ bdr (fl ) ,
bdr,i (fl ) =
λ
(74)
(75)
(76)
and ◦ denotes Hadamard product. The derivations of (71) - (73) are given in the Appendix. To
accelerate the convergence, we introduce a pruning step to remove the atoms whose coefficients are
smaller than a threshold during each iteration. Specifically, at the k-th iteration, if |σlk | < η k where
η k is a given threshold at the k-th iteration, then l-th path are removed from the set and number
of estimated paths is decreased by one, i.e., Ωk ← Ωk \ glk , flk and L̃k ← L̃k − 1 at the k-th
P k
iteration. The algorithm stops when hk+1 − hk < ε, where hk = L̃l=1 qNU glk , flk σlk denotes
the channel estimation at the k-th iteration.
VI. S IMULATION R ESULTS
A. Simulation Setup
In this section, we evaluate the performance of the proposed channel estimators for mmWave FDMIMO links with UPA or NUPA. We compare the channel estimation performance of the proposed
algorithm with some existing algorithms including the 4D-MUSIC [12] and the orthogonal matching
pursuit (OMP) [46]. The simulation parameters are set as follows.
1, The numbers of transmit and receive antenna are N = 16 and M = 16, respectively. For UPA,
we set N1 = 4, N2 = 4, M1 = 4 and M2 = 4.
21
2, In the UPA case, the DFT codebooks at the transmitter for elevation and azimuth are given by
P1 = [cN1 (ψ1,0 ) cN1 (ψ1,1 ) · · · cN1 (ψ1,P1 −1 )] ∈ CN1 ×P1 ,
P2 = [cN2 (ψ2,0 ) cN2 (ψ2,1 ) · · · cN2 (ψ2,P2 −1 )] ∈ CN2 ×P2 ,
where P1 and P2 are the sizes of elevation and azimuth codebooks, respectively. The DFT angles
are ψ1,i =
i
P1
for i = 0, . . . , P1 − 1 and ψ2,i =
i
P2
for i = 0, . . . , P2 − 1. We take the Kronecker
product of P1 and P2 to form the product codebook P = P1 ⊗ P2 with size P = P1 P2 . Each
beamforming vector has a unit norm, i.e., kpp k = 1 for p = 1, . . . , P and rank (P) = P .
p
3, The weight factor in (28) and (43) is set as µ = σw M N log (M N ). The weight for the
augmented Lagrangian in (45) is set as ρ = 0.05.
4, gl and fl for each path are assumed to uniformly take values in [− 21 , 12 ) × [− 21 , 12 ). The number
of paths L = 3.
5, The signal power is controlled by the signal-to-noise ratio (SNR) which is defined as SNR =
Pt
2
σw
with σw2 = 1.
6, For NUPA, we use circular arrays for both transmitter and receiver with N and M antenna
elements located on the 2D plane, respectively. Specifically, the n-th transmit antenna location is
set as dt,1 (n) = Rt cos (χn ) , dt,2 (n) = Rt sin (χn ) , n = 1, 2, . . . , N , where χn = 2π Nn is the
angular position of the n-th element and Rt is the radius of the transmit array. Similarly, the mth receive antenna location is dr,1 (m) = Rr cos (χm ) , dr,2 (m) = Rr sin (χm ) , m = 1, 2, . . . , M ,
m
is the angular position of the m-th element and Rr is the radius of the
where χm = 2π M
receive array.
7, For the gradient descent algorithm, we set L̃0 = M P as the initial value in both UPA and NUPA
cases. The pruning threshold in the k-th step is set as η k = 0.7 max1≤l≤L̃k σlk .
8, For the OMP and 4D-MUSIC algorithms, the AoD and AoA grid points are set as ϑ̄i =
ϕ̄i =
(i−1)2π
NG
− π and θ̄i =
(i−1)2π
NG
− π, φ̄i =
(i−1)2π
NG
(i−1)2π
−π,
NG
− π, respectively, for i = 1, . . . NG .
9, In the simulation, we use the CVX package [47] to compute the 4D atomic norm-based estimator.
22
B. Performance Evaluation
We use the normalized mean square error (NMSE), i.e., NMSE = E
2
kĤ−HkF
kHk2F
as the channel
estimation performance metric. The NMSE statistics across different SNRs with different test setups
are evaluated. Each curve is obtained by averaging over 100 realizations. First we compare the
channel estimation performance of different algorithms under the UPA setting. Then we show the
channel estimation performance for NUPA with the proposed gradient descent estimator and compare
it with the 4D-MUSIC and OMP algorithms.
The computational complexity of the proposed approximate 4D atomic-norm-based channel estimator is O((M + N )3 ) per-iteration. The computational complexity of the MUSIC estimator is
O (N M )3 + NG4 (N M )2 where O (N M )3 is for eigen decomposition and O NG4 (N M )2 is for
grid search. The complexity of the OMP estimator is O NG4 (N M )2 per iteration. The complexity
of proposed gradient descent estimator is O (M (N + P )) per iteration.
1) Convergence Behavior of the Proposed Channel Estimators: We illustrate the convergence of
the proposed ADMM implementation of the approximate 4D atomic-norm-based channel estimator
through a simulation example. We compare the NMSE of the ADMM channel estimator with that
of the CVX solver [47] that directly solves (43). As can be seen from Fig. 2, the proposed ADMM
channel estimator converges to the solution given by the CVX after 300-400 iterations for different
SNR. It is worth noting that the ADMM runs much faster than the CVX solver because the calculation
in each iteration is in closed-form. We then show the convergence behavior and the number of
estimated paths of the proposed gradient-descent-based channel estimator for UPA and NUPA in
Fig. 3. It is seen that the algorithm converges within 1500-2000 iterations for different SNR. The
estimated number of paths is more accurate at higher SNR when the algorithm converges, as more
spurious frequencies arise when the noise is stronger. It is also worth noting that the computational
complexity of the gradient descent method is lower than that of the ADMM, but the overall running
time is higher because it takes more iterations.
2) Comparison of On-grid and Off-grid Algorithms: We compare the proposed off-grid channel
estimator with two existing on-grid approaches including OMP and MUSIC. For the on-grid algorithms, the continuous AoA and AoD parameter spaces are discretized into a finite set of grids
23
(a) SNR= 4dB
(b) SNR= 10dB
Figure 2: Convergence of proposed ADMM channel estimator with different SNR.
covering [−π, π], and the estimation performance improves with higher grid resolution (i.e., larger
NG ). However, higher grid resolution leads to higher computational complexity.
In Fig. 4, the NMSE and running time of different channel estimators are plotted against NG . In
this simulation, we use CVX solver to compute the 4D atomic-norm-based estimator and the ADMM
algorithm to compute the approximate 4D atomic-norm-based estimator. It is worth noting that the
proposed approximate 4D atomic-norm-based estimator has the smallest complexity while its NMSE
is much smaller than those of the on-grid algorithms. As the algorithm does not require the grids,
its computational complexity does not change with NG . In addition, its NMSE performance is only
slightly worse than the 4D atomic-norm-based channel estimator, indicating that the performance
loss caused by the approximation of kHkAM by SDP(H) is small.
24
(a)
(b)
(c)
(d)
Figure 3: Convergence and the number of estimated paths of the proposed gradient descent algorithm
for (a)(c) UPA and (b)(d) NUPA.
(a)
(b)
Figure 4: Comparison of channel estimation performance and running time against grid size, SNR=
10 dB. (a) NMSE performance; (b) running time.
25
3) Channel Estimation Performance: Fig. 5 plots the NMSE curves as a function of SNR for
different channel estimators under UPA. The number of grid points are set as NG = 90, 180 for
4D-MUSIC-based and OMP-based channel estimators.
It is seen that 4D atomic-norm-based and approximate 4D atomic-norm-based estimators outperform the 4D-MUSIC-based and OMP-based estimators. Meanwhile, the 4D atomic-norm-based
channel estimator achieves better performance than the approximate 4D atomic-norm-based channel
estimator by about from 0.5 - 0.8 dB. And the approximate 4D atomic-norm-based channel estimator
outperforms the gradient-descent-based algorithm by more than 1.0 dB.
Figure 5: The NMSE performance as a function of SNR for UPA.
In Fig. 6, we plot the NMSE curves as a function of SNR for different channel estimators under
NUPA. It is seen that the proposed gradient-descent-based channel estimator outperforms the 4DMUSIC and OMP-based channel estimators across the range of SNRs from 2 to 10 dB. This is
because the proposed gradient-descent-based channel estimator optimizes the frequency basis in
each iteration, so it outperforms the on-grid algorithms.
26
Figure 6: The NMSE performance as a function of SNR for NUPA.
VII. C ONCLUSIONS
In this paper, we have proposed new channel estimation schemes for mmWave beamformed FDMIMO systems based on atomic norm minimization under both UPA and NUPA settings. For the
UPA case, we approximate the original large-scale 4D atomic norm minimization problem using a
semi-definite program (SDP) containing two decoupled two-level Toeplitz matrices which is then
solved by an ADMM-based fast algorithm. For the NUPA case, a gradient descent-based algorithm
is provided to obtain a suboptimal solution. Simulation results show that the proposed atomic norm
based mmWave FD-MIMO channel estimators provide better performance compared to the existing
methods based on compressed sensing and MUSIC algorithms.
A PPENDIX
A. Derivation for (71) and (72)
√
For clarity, define P̄ = Pt PT ⊗ IM . Then the gradient with respect to gl,i can be calculated
by
H
H ∂ P̄h
y − P̄h
1 ∂ y − P̄h
∇gl,i Γ ({gl , fl , σl }) =
= R P̄h − y
,
2
∂gl,i
∂gl,i
(77)
27
where
P
∂ P̄ Ll=1 qNU (gl , fl ) σl
∂ P̄h
∂qNU (gl , fl )
=
= σl P̄
,
∂gl,i
∂gl,i
∂gl,i
∂a∗dt (gl ) ⊗ bdr (fl )
∂a∗dt (gl )
∂qNU (gl , fl )
=
=
⊗ bdr (fl ) ,
∂gl,i
∂gl,i
∂gl,i
∂a∗dt (gl )
−j2π
=
[dt,i (1), . . . , dt,i (N )]T ◦ a∗dt (gl ) .
∂gl,i
λ
(78)
(79)
(80)
By plugging (80) into (77), we have (71). Similarly we can obtain (72).
B. Derivation for (73)
The gradient with respect to σl can be calculated by
2
∂ µ kσk1 + 12 y − P̄h 2
∇σl Γ ({gl , fl , σl }) =
∂σl∗
∂kσk1 1 ∂yH P̄h 1 ∂hH P̄H y 1 ∂hH P̄H P̄h
=
−
−
+
,
∂σl∗
2 ∂σl∗
2 ∂σl∗
2
∂σl∗
(81)
where
∂yH P̄h
∂σl ∗
∂hH P̄H y
∂σl∗
∂hH P̄H P̄h
∂σl∗
∂kσk1
∂σl∗
∂σl
= 0,
∂σl∗
H T ∂σl∗
H T
= P̄qNU (gl , fl ) y
=
P̄q
(g
,
f
)
y ,
NU
l
l
∂σl∗
T
H
= P̄qNU (g, f ) P̄qNU (gl , fl ) σl ,
P
∂ l |σl |
∂|σl |
1
∂|σl |
∂|σl |
=
=
=
+i
∂σl∗
∂σl∗
2 ∂R{σl }
∂I{σl }
!
p
p
∂ R2 {σl } + I 2 {σl }
1 ∂ R2 {σl } + I 2 {σl }
=
+i
2
∂R{σl }
∂I{σl }
σl
=
,
2|σl |
= yH P̄qNU (gl , fl )
(82)
(83)
(84)
(85)
where I {·} returns the imaginary part of the input. Plugging (82)-(85) into (81), we obtain (73).
R EFERENCES
[1] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez,
“Millimeter wave mobile communications for 5g cellular: It will work!” IEEE Access, vol. 1, pp. 335–349, 2013.
28
[2] X. Cheng, B. Yu, L. Yang, J. Zhang, G. Liu, Y. Wu, and L. Wan, “Communicating in the real world: 3d mimo,” IEEE Wireless
Communications, vol. 21, no. 4, pp. 136–144, August 2014.
[3] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and A. Ghosh, “Millimeter wave beamforming for wireless backhaul
and access in small cell networks,” IEEE Trans. Commun., vol. 61, no. 10, pp. 4391–4403, October 2013.
[4] Y. H. Nam, B. L. Ng, K. Sayana, Y. Li, J. Zhang, Y. Kim, and J. Lee, “Full-dimension mimo (fd-mimo) for next generation
cellular technology,” IEEE Communications Magazine, vol. 51, no. 6, pp. 172–179, June 2013.
[5] S. Kutty and D. Sen, “Beamforming for millimeter wave communications: An inclusive survey,” IEEE Communications Surveys
Tutorials, vol. 18, no. 2, pp. 949–973, Secondquarter 2016.
[6] J. Wang, “Beam codebook based beamforming protocol for multi-gbps millimeter-wave wpan systems,” IEEE Journal on
Selected Areas in Communications, vol. 27, no. 8, pp. 1390–1399, October 2009.
[7] M. K. Samimi and T. S. Rappaport, “3-d millimeter-wave statistical channel model for 5g wireless system design,” IEEE
Transactions on Microwave Theory and Techniques, vol. 64, no. 7, pp. 2207–2225, July 2016.
[8] T. A. Thomas, H. C. Nguyen, G. R. MacCartney, and T. S. Rappaport, “3d mmwave channel model proposal,” in 2014 IEEE
80th Vehicular Technology Conference (VTC2014-Fall), Sept 2014, pp. 1–6.
[9] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, “Spatially sparse precoding in millimeter wave mimo systems,”
IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, March 2014.
[10] A. Alkhateeb, O. E. Ayach, G. Leus, and R. W. Heath, “Channel estimation and hybrid precoding for millimeter wave cellular
systems,” IEEE Journal of Selected Topics in Signal Processing, vol. 8, no. 5, pp. 831–846, Oct 2014.
[11] A. Alkhateeby, G. Leusz, and R. W. Heath, “Compressed sensing based multi-user millimeter wave systems: How many
measurements are needed?” in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
April 2015, pp. 2909–2913.
[12] Z. Guo, X. Wang, and W. Heng, “Millimeter-wave channel estimation based on two-dimensional beamspace music method,”
IEEE Transactions on Wireless Communications, vol. PP, no. 99, 2017.
[13] S. Sun and T. S. Rappaport, “Millimeter wave mimo channel estimation based on adaptive compressed sensing,” in 2017 IEEE
International Conference on Communications Workshops (ICC Workshops), May 2017, pp. 47–53.
[14] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inform. Theory, vol. 52, no. 4, pp. 1289–1306, April 2006.
[15] T. S. Rappaport, R. W. Heath, R. C. Daniels, and J. N. Murdock, Millimeter Wave Wireless Communication.
Prentice-Hall,
2014.
[16] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed sensing off the grid,” IEEE Trans. Inform. Theory, vol. 59, no. 11,
pp. 7465–7490, Nov 2013.
[17] B. N. Bhaskar, G. Tang, and B. Recht, “Atomic norm denoising with applications to line spectral estimation,” IEEE Trans. Sig.
Proc., vol. 61, no. 23, pp. 5987–5999, 2013.
[18] J. Steinwandt, F. Roemer, C. Steffens, M. Haardt, and M. Pesavento, “Gridless super-resolution direction finding for strictly noncircular sources based on atomic norm minimization,” in 2016 50th Asilomar Conference on Signals, Systems and Computers,
Nov 2016, pp. 1518–1522.
[19] B. N. Bhaskar, G. Tang, and B. Recht, “Atomic norm denoising with applications to line spectral estimation,” IEEE Trans. Sig.
Proc., vol. 61, no. 23, pp. 5987–5999, Dec 2013.
29
[20] Y. Zhang, G. Zhang, and X. Wang, “Array covariance matrix-based atomic norm minimization for off-grid coherent directionof-arrival estimation,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), March
2017, pp. 3196–3200.
[21] Z. Tian, Z. Zhang, and Y. Wang, “Low-complexity optimization for two-dimensional direction-of-arrival estimation via decoupled
atomic norm minimization,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
March 2017, pp. 3071–3075.
[22] P. Zhang, L. Gan, S. Sun, and C. Ling, “Atomic norm denoising-based channel estimation for massive multiuser mimo systems,”
in 2015 IEEE International Conference on Communications (ICC), June 2015, pp. 4564–4569.
[23] P. Shah, B. N. Bhaskar, G. Tang, and B. Recht, “Linear system identification via atomic norm regularization,” in 2012 IEEE
51st IEEE Conference on Decision and Control (CDC), Dec 2012, pp. 6265–6270.
[24] W. Liu, Z. Wang, C. Sun, S. Chen, and L. Hanzo, “Structured non-uniformly spaced rectangular antenna array design for
fd-mimo systems,” IEEE Trans. Wireless Commun., vol. 16, no. 5, pp. 3252–3266, May 2017.
[25] P. Wang, Y. Li, Y. Peng, S. C. Liew, and B. Vucetic, “Non-uniform linear antenna array design and optimization for millimeterwave communications,” IEEE Transactions on Wireless Communications, vol. 15, no. 11, pp. 7343–7356, Nov 2016.
[26] J. Choi, K. Lee, D. J. Love, T. Kim, and R. W. Heath, “Advanced limited feedback designs for fd-mimo using uniform planar
arrays,” in 2015 IEEE Global Communications Conference (GLOBECOM), Dec 2015, pp. 1–6.
[27] Y. Han, S. Jin, X. Li, Y. Huang, L. Jiang, and G. Wang, “Design of double codebook based on 3d dual-polarized channel for
multiuser mimo system,” EURASIP Journal on Advances in Signal Processing, vol. 2014, no. 1, p. 111, 2014.
[28] D. Yang, L. L. Yang, and L. Hanzo, “Dft-based beamforming weight-vector codebook design for spatially correlated channels
in the unitary precoding aided multiuser downlink,” in 2010 IEEE International Conference on Communications, May 2010,
pp. 1–5.
[29] J. Lee, G. T. Gil, and Y. H. Lee, “Exploiting spatial sparsity for estimating channels of hybrid mimo systems in millimeter
wave communications,” in 2014 IEEE Global Communications Conference, Dec 2014, pp. 3326–3331.
[30] P. Gupta and S. P. Kar, “Music and improved music algorithm to estimate direction of arrival,” in 2015 International Conference
on Communications and Signal Processing (ICCSP), April 2015, pp. 0757–0761.
[31] W. Zhang, W. Liu, J. Wang, and S. Wu, “Joint transmission and reception diversity smoothing for direction finding of coherent
targets in mimo radar,” IEEE Journal of Selected Topics in Signal Processing, vol. 8, no. 1, pp. 115–124, Feb 2014.
[32] Z. Yang and L. Xie, “Exact joint sparse frequency recovery via optimization methods,” IEEE Trans. Sig. Proc., vol. 64, no. 19,
pp. 5145–5157, Oct 2016.
[33] Z. Yang, L. Xie, and P. Stoica, “Vandermonde decomposition of multilevel toeplitz matrices with application to multidimensional
super-resolution,” IEEE Trans. Inform. Theory, vol. 62, no. 6, pp. 3685–3701, June 2016.
[34] Y. Chi and Y. Chen, “Compressive recovery of 2-d off-grid frequencies,” in 2013 Asilomar Conference on Signals, Systems and
Computers, Nov 2013, pp. 687–691.
[35] ——, “Compressive two-dimensional harmonic retrieval via atomic norm minimization,” IEEE Trans. Sig. Proc., vol. 63, no. 4,
pp. 1030–1042, 2015.
[36] E. J. Candès and C. Fernandez-Granda, “Towards a mathematical theory of super-resolution,” Communications on Pure and
Applied Mathematics, vol. 67, no. 6, pp. 906–956, 2014.
30
[37] L. Zheng and X. Wang, “Super-resolution delay-doppler estimation for ofdm passive radar,” IEEE Trans. Sig. Proc., vol. 65,
no. 9, pp. 2197–2210, May 2017.
[38] Z. Tian, Z. Zhang, and Y. Wang, “Low-complexity optimization for two-dimensional direction-of-arrival estimation via decoupled
atomic norm minimization,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
March 2017, pp. 3071–3075.
[39] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating
direction method of multipliers,” Foundations and Trends R in Machine Learning, vol. 3, no. 1, pp. 1–122, 2011.
[40] J. F. Sturm, “Using sedumi 1.02, a matlab toolbox for optimization over symmetric cones,” Optimization methods and software,
vol. 11, no. 1-4, pp. 625–653, 1999.
[41] K.-C. Toh, M. J. Todd, and R. H. Tutuncu, “Sdpt3-a matlab software package for semidefinite programming, version 1.3,”
Optimization methods and software, vol. 11, no. 1-4, pp. 545–581, 1999.
[42] E. Torkildson, C. Sheldon, U. Madhow, and M. Rodwell, “Nonuniform array design for robust millimeter-wave mimo links,”
in GLOBECOM 2009 - 2009 IEEE Global Telecommunications Conference, Nov 2009, pp. 1–7.
[43] S. E. Nai, W. Ser, Z. L. Yu, and H. Chen, “Beampattern synthesis for linear and planar arrays with antenna selection by convex
optimization,” IEEE Transactions on Antennas and Propagation, vol. 58, no. 12, pp. 3923–3930, 2010.
[44] Q. Li and G. Tang, “Approximate support recovery of atomic line spectral estimation: A tale of resolution and precision,” in
Signal and Information Processing (GlobalSIP), 2016 IEEE Global Conference on.
IEEE, 2016, pp. 153–156.
[45] N. Boumal and P.-A. Absil, “Low-rank matrix completion via preconditioned optimization on the grassmann manifold,” Linear
Algebra and its Applications, vol. 475, pp. 200–239, 2015.
[46] T. T. Cai and L. Wang, “Orthogonal matching pursuit for sparse signal recovery with noise,” IEEE Trans. Inform. Theory,
vol. 57, no. 7, pp. 4680–4688, July 2011.
[47] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex programming, version 2.1,” http://cvxr.com/cvx, Mar.
2014.
| 3cs.SY
|
arXiv:1510.07710v3 [math.GR] 11 Sep 2016
Invariant random subgroups of groups acting on hyperbolic
spaces
D. Osin∗
Abstract
Suppose that a group G acts non-elementarily on a hyperbolic space S and does not
fix any point of ∂S. A subgroup H ≤ G is geometrically dense in G if the limit sets
of H and G on ∂S coincide and H does not fix any point of ∂S. We prove that every
invariant random subgroup of G is either geometrically dense or contained in the elliptic
radical (i.e., the maximal normal elliptic subgroup of G). In particular, every ergodic
measure preserving action of an acylindrically hyperbolic group on a Borel probability
space (X, µ) either has finite stabilizers µ-almost surely or otherwise the stabilizers are
very large (in particular, acylindrically hyperbolic) µ-almost surely.
1
Introduction
To every discrete group G, one can associate a topological dynamical system as follows. Let
S(G) denote the space of all subgroups of G with topology inherited from the Tichonoff
product topology of {0, 1}G . Then S(G) is Hausdorff and compact; if G is countable, S(G)
is second countable and hence metrizable. The group G acts on S(G) by conjugation,
written g.H = gHg−1 for g ∈ G and H ∈ S(G). It is easy to check that conjugations are
homeomorphisms of S(G).
Definition 1.1. An invariant random subgroup of a group G (abbreviated IRS ) is a Borel
G-invariant probability measure on S(G).
The term “invariant random subgroup” was suggested in [2], but this notion has been
studied for quite some time. For a brief history of the subject and a survey of some recent
developments we refer to [2, 10].
One can think of IRSs as generalizations of normal subgroups. Indeed every N ⊳ G
gives rise to an atomic IRS, the Dirac measure concentrated at N . The study of IRSs is
motivated by the following.
Example 1.2. Let G be a group acting by probability measure preserving (abbreviated
p.m.p.) transformations on a Borel space (X, ν). For a point x ∈ X, let
Gx = {g ∈ G | gx = x}
∗
This work was supported by the NSF grant DMS-1308961.
1
denote the stabilizer of x. Auslander and Moore [3] proved that the map f : X → S(G)
given by x → Gx is measurable. Thus for every Borel subset A ⊆ S(G) we can define
µ(A) = ν(f −1 (A)).
(1)
It is clear that µ is a probability measure on S(G) and is G-invariant since gGx g−1 = Ggx
and the action of G on X preserves ν. Thus µ is an IRS of G.
Abert, Glasner and Virag [2] proved that the this example is universal in the following
sense: For every IRS µ of a countable group G, there exists a p.m.p. action of G on a Borel
space (X, ν) such that µ and ν are related by (1). Thus studying IRSs of G is the same as
studying stabilizers of p.m.p. actions of G.
The main goal of this paper is to study IRSs of groups acting on hyperbolic spaces. We
briefly recall necessary terminology here and refer to Section 3 for more details. All group
actions on metric spaces considered in this paper are supposed to be isometric by default.
Given an action of a group G on a hyperbolic space S, we denote by Λ(G) the limit set of
G on the Gromov boundary ∂S. The action of G on S is non-elementary if Λ(G) is infinite
and of general type if it is non-elementary and G does not fix any point of ∂S. A subgroup
H ≤ G is called elliptic (with respect to the given action of G) if it has bounded orbits.
The following result is likely known, but we could not find it in the literature.
Proposition 1.3 (Prop. 3.4). Suppose that a group G admits a non-elementary action on
a hyperbolic space. Then there exists a maximal normal elliptic subgroup E(G) of G with
respect to this action.
Note that if the action of G is elementary, such a maximal subgroup may not exist (see
Example 3.5). We call E(G) the elliptic radical of G. Further we say that a subgroup H ≤ G
is geometrically dense with respect to a given general type action of G on a hyperbolic space
S if H does not fix any point of ∂S and Λ(H) = Λ(G).
Example 1.4. Let G be a non-elementary hyperbolic group acting on its Cayley graph with
respect to some finite generating set. Then E(G) is finite and
(a) every infinite normal subgroup of G is geometrically dense;
(b) a quasi-convex subgroup H ≤ G is geometrically dense iff it is of finite index.
Recall that an IRS µ of G is ergodic if so is the action of G on (S(G), µ). That is, for
every G-invariant subset A ⊆ S(G), we have µ(A) = 0 or µ(A) = 1. We also say that an
IRS µ has some property P if µ-almost every subgroup of G has P . Our main goal is record
the following fairly elementary (but seemingly useful) fact.
Theorem 1.5. Let G be a countable group acting on a hyperbolic space S. Assume that
the action of G on S is of general type and let µ be an ergodic IRS of G. Then either µ is
geometrically dense or µ is contained in E(G).
The case of a closed non-elementary subgroup G of a simple Lie group of R-rank 1 was
previously considered in [1, Proposition 11.3]; the result obtained there is stated in a slightly
2
different way, but is easily seen to be equivalent to a particular case of Theorem 1.5. A
similar theorem for groups acting on CAT (0) spaces was also obtained in [7].
We mention one particular application of Theorem 1.5. Recall that a group is acylindrically hyperbolic if it admits a non-elementary acylindrical action on a hyperbolic space.
(For more details, we refer to Section 4.2.) Examples of such groups include non-elementary
hyperbolic and relatively hyperbolic groups, infinite mapping class groups of punctured surfaces, Out(Fn ) for n ≥ 2, most 3-manifold groups, finitely presented groups of deficiency
at least 2, and many other examples of interest, see [8, 16, 17, 18] and references therein.
Bowen, Grigorchuk, and Kravchenko [5] proved that every acylindrically hyperbolic group
has continuously many non-atomic ergodic IRSs.
We say that a subgroup H of an acylindrically hyperbolic group G is totally geometrically
dense if it is geometrically dense with respect to every non-elementary acylindrical action
of G on a hyperbolic space. Every acylindrically hyperbolic group G contains a unique
maximal normal finite subgroup, denoted K(G) [8, Theorem 2.24].
Corollary 1.6. Let µ be an ergodic IRS of a countable acylindrically hyperbolic group.
Then either µ is totally geometrically dense or µ is supported on subgroups of K(G); in
particular, µ is acylindrically hyperbolic in the former case and finite in the latter case.
A particular case of this corollary can be reformulated in terms of actions as follows.
Corollary 1.7. Let G be a countable acylindrically hyperbolic group. Suppose that G acts
ergodically by measure preserving transformations of a Borel probability space (X, ν). Then
either the stabilizer of ν-a.e. point of X is acylindrically hyperbolic or the stabilizer of ν-a.e.
point of X is finite. If, in addition, K(G) = 1, the action is essentially free in the latter
case.
In particular, this corollary implies that if G is an acylindrically hyperbolic group and
K(G) = 1, then every p.m.p. action of G with “small” stabilizers (e.g., amenable, or, more
generally, without non-cyclic free subgroups) is essentially free. This statement can be made
precise using the notion of a small subgroup introduced in [15].
The proof of our main theorem consists of two steps. First we observe that every IRS
of a countable group satisfies a purely algebraic normality-like condition (Proposition 2.5),
and then show that this condition implies the dichotomy as in Theorem 1.5. In fact, the
first step can be easily extracted from [9] and the author is grateful to Yair Glasner for
explaining details of [9]. The author is also grateful to the referee for careful reading of the
manuscript and useful comments.
2
Recurrent subgroups and dynamical systems
The following notion plays the central role in our paper.
Definition 2.1. We say that a subgroup H of a group G is recurrent if for every g ∈ G, H is
a recurrent point of the dynamical system (S(G), g). That is, for every open neighborhood
U ⊆ S(G) of H, there are infinitely many n ∈ N such that g n .H = gn Hg−n ∈ U .
3
It is useful to reformulate this definition in a purely algebraic form.
Lemma 2.2. A subgroup H of a group G is recurrent if and only if for every g ∈ G and
every finite subset F ⊆ G, there exists at least one (equivalently, infinitely many) n ∈ N
such that H ∩ F = g n Hg−n ∩ F.
Proof. By the definition of the topology of S(G), the base of neighborhoods of H is formed
by the subsets
UF (H) = {K ≤ G | K ∩ F = H ∩ F},
where F ranges in the set of all finite subsets of G. Clearly H is recurrent if and only if
Definition 2.1 holds for neighborhoods from this base. It remains to note that the condition
g.H = gn Hg−n ∈ UF (H) is equivalent to g n Hg−n ∩ F = H ∩ F.
Throughout this paper we write xy for y −1 xy.
Corollary 2.3. Let H be a recurrent subgroup of a group G. Then for every g ∈ G and
n
every h ∈ H, there exist infinitely many n ∈ N such that hg ∈ H.
Proof. Apply Lemma 2.2 to F = {h}.
Example 2.4. (a) If G is a torsion group, then every subgroup of G is recurrent. To make
this notion more useful in the torsion case, one can state Definition 2.1 using arbitrary
infinite subsemigroups of G instead of cyclic subgroups hgi. Proposition 2.5 proved
below holds for this definition as well. However we do not need this in our paper.
(b) Suppose that the normalizer of H has finite index in G. Then H is recurrent. The
converse is false in general, but is true, for example, if G is polycyclic. We leave this
as an exercise for the reader.
The main result of this section is the following. As we already mentioned in the introduction, it can be extracted from [9] although it is not stated explicitly in this form
there.
Proposition 2.5. Let G be a countable group and let µ be an IRS of G. Then µ-a.e.
subgroup H ≤ G is recurrent.
Proof. Let µ be an IRS on G. For an element g ∈ G, let Rg denote the set of recurrent
points of the dynamical system (S(G), g). Since the space S(G) is second countable, we
can apply the topological
version of the Poincaré recurrence theorem, which states that
T
µ(Rg ) = 1. Let R = g∈G Rg . Clearly R is the set of all recurrent subgroups of G. Since
G is countable and µ is countably additive, we have µ(R) = 1.
4
3
Elliptic radical of groups acting on hyperbolic spaces
3.1. Hyperbolic spaces and group actions. All actions of groups on metric spaces are
supposed to be isometric by default. We first recall necessary definitions and properties of
groups acting on hyperbolic spaces. Although many existent proofs of the results mentioned
below assume properness of the space, they also hold in the general case; we refer to Sections
8.1-8.2 in [12] or to [14] for complete proofs in a more general context.
Let (S, d) be a (not necessarily proper) hyperbolic space. The Gromov product of two
points x, y ∈ S with respect to a point z ∈ S is defined by
1
(x, y)z = (d(x, z) + d(y, z) − d(x, y)).
2
The next lemma is well-known (see [12] or Proposition 21 in [13, Chapter 2]).
Lemma 3.1. Let S be a δ-hyperbolic space. Then for any x, y, z, t ∈ S, we have
(x, z)t ≥ min{(x, y)t , (y, z)t } − 8δ.
By ∂S we denote the Gromov boundary of S, which is defined as the set of equivalence
classes of sequences converging at infinity. More precisely, a sequence (xn ) of elements of
S converges at infinity if (xi , xj )s → ∞ as i, j → ∞; this is independent of the choice of
the base point s ∈ S. Two such sequences (xi ) and (yi ) are equivalent if (xi , yj )s → ∞ as
i, j → ∞. If x is the equivalence class of (xi ), we say that the sequence xi converges to x.
This naturally defines a topology on Sb = S ∪ ∂S extending the topology on S so that S is
b
dense in S.
Remark 3.2. It is useful to remember the following elementary fact: if (xi ), (yi ) are two
sequences of points in S converging to infinity such that supi d(xi , yi ) < ∞, then (xi ) and
(yi ) converge to the same point of ∂S.
We denote by Λ(G) the set of limit points of G on ∂S. That is,
Λ(G) = Gs ∩ ∂S,
b Using Remark 3.2 it is easy to
where s ∈ S and Gs is the closure of the G-orbit of s in S.
show that this definition is independent on the choice of a particular orbit.
It is also easy to check that if a sequence (xi ) converges to a point x ∈ ∂S, then for
every g ∈ G, the sequence (gxi ) also converges to a point of ∂S, which only depends on x
and g. This allows one to define an action of G on ∂S, which turns out to be continuous.
An element g ∈ G is called elliptic if it fixes (setwise) a bounded subset of S. An element
g ∈ G is loxodromic if the map Z → S defined by n 7→ gn s is a quasi-isometric embedding
for every s ∈ S (here we assume that Z is equipped with the standard metric). Equivalently,
g is loxodromic if it is not elliptic and fixes exactly two distinct points g + and g − on ∂S.
For every loxodromic element g ∈ G, we have
lim gn s = g+ ∀s ∈ Sb \ {g − }
n→∞
and
5
lim g−n s = g− ∀s ∈ Sb \ {g + }
n→∞
as n → ∞. This is called the north-south dynamics of the action of a loxodromic element.
In particular, we have Λ(hgi) = {g+ , g− }.
Two loxodromic elements g, h ∈ G are called independent if {g+ , g− } ∩ {h+ , h− } = ∅.
We denote by L(G) the set of all loxodromic elements of G and let
H(G) = {g± | g ∈ L(G)}.
We also denote by F ix(G) the set of fixed points of G on ∂S.
Possible actions of groups on hyperbolic spaces can be classified as follows according to
the cardinality of Λ(G) (see [12, Sections 8.1-8.2], [14], or [6, Section 3a])
1) (elliptic action) |Λ(G)| = 0. Equivalently, G has bounded orbits.
2) (parabolic action) |Λ(G)| = 1. Equivalently, G has unbounded orbits and contains no
loxodromic elements. In this case F ix(G) = Λ(G).
3) (lineal action) |Λ(G)| = 2. Equivalently, G contains a loxodromic element and any
two loxodromic elements have the same limit points on ∂S. In this case F ix(G) ⊆
Λ(G).
4) |Λ(G)| = ∞. Then G always contains loxodromic elements. In turn, this case breaks
into two subcases.
a) (quasi-parabolic action) F ix(G) 6= ∅. Then G fixes a unique point of ∂S.
b) (general type action) F ix(G) = ∅. Equivalently, G contains at least 2 (or infinitely many) independent loxodromic elements.
The action of G is called elementary in cases 1)–3) and non-elementary in case 4).
The following can be found in [12, Section 8.2]; alternatively, see [14, Theorems 2.6, 2.9].
Lemma 3.3. Suppose that a group G acts non-elementarily on a hyperbolic space S. Then
H(G) is dense in Λ(G). Moreover, if the action is of general type, then for any two points
x, y ∈ Λ(G) and any open neighborhoods A, B ⊆ ∂S of x and y, respectively, there exists a
loxodromic element g ∈ G such that g + ∈ A and g − ∈ B.
3.2. Elliptic radical. Given an action of a group G on a hyperbolic space S, we define
the elliptic radical of G with respect to this action by
E(G) = {g ∈ G | gx = x ∀ x ∈ Λ(G)}.
(2)
Note that this definition makes sense for arbitrary actions on hyperbolic spaces. For
non-elementary actions, the following proposition provides an equivalent characterization.
Proposition 3.4. Suppose that a group G admits an action of general type on a hyperbolic
space S. Then E(G) is the unique maximal elliptic normal subgroup of G.
6
Proof. Obviously E(G) is normal in G. Further, let N be a normal elliptic subgroup of G.
Fix some s ∈ S. Since N is elliptic, the diameter D = diam (N s) is finite. For every a ∈ N
and every sequence (gi ) ⊆ G such that lim gi s = x ∈ ∂S for some s ∈ S, we have
i→∞
d(agi s, gi s) = d(gi−1 agi s, s) ≤ D
since gi−1 agi ∈ N . By Remark 3.2, this implies that ax = x and thus N ≤ E(G).
Finally we note that by the classification of group actions on hyperbolic spaces, only
elliptic subgroups can fix more that 2 points of ∂S. Since the action of G is non-elementary
we have |F ix(E(G))| ≥ |Λ(G)| = ∞ and thus E(G) is elliptic.
Example 3.5. In general, the proposition may fail in various ways for elementary (lineal and
parabolic) actions. Indeed, for any lineal action, E(G) as defined by (2) contains loxodromic
elements and hence is not elliptic. Furthermore, the maximal normal elliptic subgroup of
G may not exist for parabolic actions. Indeed, let
i
W = ha, t | a2 = 1, [at , a] = 1, ∀i ∈ Zi = Z2 ≀ Z.
It is easy to see that W is an ascending HNN-extension of the locally finite subgroup
i
A = hai , i ∈ Ni, where ai = at , associated to a monomorphism A → A given by ai 7→ ai+1 .
Let T be the corresponding Bass-Serre tree and let G = hai , i ∈ Zi. Every finitely generated
subgroup of G is normal (since G is abelian) and finite (hence elliptic with respect to the
action on T ). However G has an unbounded orbit in T and hence its action is parabolic.
Thus the maximal normal elliptic subgroup of G does not exist.
4
A dichotomy for IRSs of groups acting on hyperbolic
spaces
Let G be a group acting on a hyperbolic space (S, d). Throughout this section we assume
that G is countable and the action is of general type.
4.1. Recurrent subgroups of groups acting on hyperbolic spaces We will need
the following result from [18, Lemma 3.5].
Lemma 4.1. Suppose that for some a, b ∈ G there exist x, y ∈ S and C > 0 such that
max{d(x, ax), d(y, by)} ≤ C
(3)
min{d(x, bx), d(y, ay)} ≥ d(x, y) + 3C.
(4)
and
Then ab is a loxodromic isometry.
Recall that for a point a ∈ ∂S, Ga denotes the stabilizer of a.
7
Lemma 4.2. Suppose that a group G acts on a hyperbolic space S. Let us fix some s ∈ S
and let g ∈ G be a loxodromic element. Then for every f ∈ G \ Gg+ , there exists a constant
D such that
|d(f g n s, gn s) − 2d(gn s, s)| ≤ D
(5)
for all n ∈ N.
Proof. Since f ∈
/ Gg+ , the sequences (gn s) and (f gn s) converge to distinct points of ∂S. It
follows that the Gromov products (f gn s, gn s)s are uniformly bounded by some constant C.
Therefore, we obtain
d(f gn s, gn s) = d(f gn s, s) + d(gn s, s) − 2(f gn s, gn s)s ≥
d(f gn s, f s) − d(f s, s) + d(gn s, s) − 2C =
(6)
2d(gn s, s) − d(f s, s) − 2C.
On the other hand, we obviously have
d(f g n s, gn s) ≤ d(f gn s, f s) + d(f s, s) + d(s, gn s) ≤ 2d(gn s, s) + d(f s, s).
(7)
Combining (6) and (7), we obtain (5) for D = d(f s, s) + 2C.
Lemma 4.3. Let g ∈ L(G) and let f ∈ G be an element such that f g + 6= g + and f g− 6= g− .
Then for all sufficiently large n ∈ N, the commutator [f, g n ] = f g n f −1 g−n is loxodromic.
Proof. Let us fix some x ∈ S and let C = d(f x, x). Let also y = g n x. By Lemma 4.1, it
suffices to show that the inequalities (3) and (4) are satisfied for the elements a = f and
b = gn f −1 g −n .
Note that d(y, by) = d(g n x, g n f −1 x) = d(x, f −1 x) = C and thus the inequality (3) holds.
Since f g+ 6= g+ , we can apply Lemma 4.2, which tells us that d(y, ay) = d(g n x, f g n x)
grows at least as 2d(x, y) minus a constant as n → ∞ (note the multiple 2 here). Since g
is loxodromic, we have d(x, y) → ∞ as n → ∞. In particular, if n is sufficiently large, the
inequality d(y, ay) ≥ d(x, y) + 3C holds. Similarly we show that d(x, bx) ≥ d(x, y) + 3C
whenever n is sufficiently large; we need the assumption f g− 6= g − here.
Lemma 4.4. Let H ≤ G be a recurrent subgroup. If H acts non-trivially on Λ(G), then H
contains a loxodromic element.
Proof. Assume that some f ∈ H acts nontrivially on Λ(G). Let x ∈ Λ(G) be a point
such that f x 6= x. Let y = f x. We fix two disjoint open neighborhoods U and V of x
and y, respectively, such that f U ⊆ V . By Lemma 3.3, we can find a loxodromic element
g ∈ G such that g± ⊆ U . Then f g± ⊆ f U ⊆ V . In particular, we have f g± 6= g ± . By
Lemma 4.3, the commutator [f, gn ] is loxodromic for all sufficiently large n ∈ N. Since H
is recurrent, we conclude that there exist arbitrarily large n such that gn f −1 g −n ∈ H and
hence [f, gn ] ∈ H.
8
We are now ready to prove the main result of this section.
Theorem 4.5. Suppose that a countable group G admits a general type action on a hyperbolic space S. Then every recurrent subgroup H of G is either geometrically dense in G
with respect to this action or belongs to E(G). In particular, in the former case the action
of H on S is of general type and in the latter case H is elliptic.
Proof. If H acts trivially on Λ(G), then H ≤ E(G). Thus we can assume that the action
of H on Λ(G) is nontrivial. We first prove that Λ(H) = Λ(G) in this case.
Since limit sets are closed, it suffices to show that every open neighborhood U ⊆ Sb
of every x ∈ Λ(G) contains an element of Λ(H). By Lemma 4.4, there exists loxodromic
h ∈ H. If h+ ∈ U or h− ∈ U , we are done. Thus we can assume that {h+ , h− } ∩ U = ∅.
By Lemma 3.3, there exists a loxodromic element g ∈ G such that g+ ∈ U . In particular,
we have h ∈
/ Gg+ . Let us fix some s ∈ S. We claim that
n
lim hg s = g− .
(8)
n→∞
Indeed using Lemma 4.2 we obtain
n
2(hg s, g−n s)s = d(g−n hg n s, s) + d(g−n s, s) − d(g−n hg n s, g−n s) =
d(hg n s, gn s) + d(s, gn s) − d(g n s, h−1 s) ≥
(9)
3d(s, gn s)
−D
− (d(gn s, s) +
d(s, h−1 s))
=
2d(s, gn s) − D − d(hs, s) → ∞
as n → ∞. Since the sequence (g −n s) is convergent, we also have
(g−n s, g−m s)s → ∞
(10)
n
as m, n → ∞. Combining (9), (10), and Lemma 3.1, we obtain that (hg s) converges
to infinity and is equivalent to (g−n s). Hence we obtain (8). It remains to note that
n
by Corollary 2.3, (hg ) has an infinite subsequence consisting of elements of H. Hence
g− ∈ Λ(H) and thus U ∩ Λ(H) is non-empty.
Finally we note that H cannot fix a point a ∈ ∂S. Indeed since the action of G is of
general type, we can find a loxodromic element g ∈ G such that g ± ∈
/ {a, h+ , h− }. Using the
n
±
north-south dynamics of the action of g, we conclude that g h 6= a for sufficiently large n.
On the other hand, the recurrency condition implies that there are arbitrarily large n such
that gn hg −n ∈ H. It is clear that gn hg −n is loxodromic and its fixed points are exactly
gn h± . Thus H contains an element that does not fix a.
Proof of Theorem 1.5. The theorem follows from Proposition 2.5 and Theorem 4.5 immediately.
9
4.2. An application to acylindrically hyperbolic groups. Recall that an isometric
action of a group G on a metric space (S, d) is acylindrical if for every ε > 0 there exist
R, N > 0 such that for every two points x, y with d(x, y) ≥ R, there are at most N elements
g ∈ G satisfying
d(x, gx) ≤ ε and d(y, gy) ≤ ε.
The notion of acylindricity goes back to Sela’s paper [19], where it was studied for groups
acting on trees. In the context of general metric spaces, this concept is due to Bowditch
[4]. Further, a group G is called acylindrically hyperbolic if it admits a non-elementary
acylindrical action on a hyperbolic space. For details and recent developments in the study
of acylindrically hyperbolic groups we refer to [18].
Proof of Corollary 1.6. Let G be a group acting non-elementarily and acylindrically on a
hyperbolic space S. By [8, Lemma 6.15] and part (a) of Proposition 3.4, E(G) = K(G) is
finite in this case. Thus for every IRG µ of G, µ-almost every subgroup of G is geometrically
dense with respect to the action on S or belongs to K(G). It is clear that the set of subgroups
of K(G) is G-invariant, and thus if µ is ergodic it must have measure 0 or 1. This implies
the first claim of the corollary. The second part follows immediately from the fact that the
action of a geometrically dense subgroup is always non-elementary and the fact that K(G)
is finite.
Corollary 1.7 is an obvious reformulation of Corollary 1.6 (see Example 1.2).
References
[1] M. Abert, N. Bergeron, I. Biringer, T. Gelander, N. Nikolov, J. Raimbault, I. Samet, On
the growth of L2 -invariants for sequences of lattices in Lie groups, arXiv:1210.2961v3.
[2] M. Abert, Y. Glasner, B. Virag, Kesten’s theorem for Invariant Random Subgroups,
Duke Math. J. 163, no. 3 (2014), 465-488.
[3] L. Auslander, C. C. Moore, Unitary representations of solvable Lie groups, Memoirs
AMS (1966), 66-77.
[4] B. Bowditch, Tight geodesics in the curve complex, Invent. Math. 171 (2008), no. 2,
281-300.
[5] L. Bowen, R. Grigorchuk, R. Kravchenko, Characteristic random subgroups of geometric groups and free abelian groups of infinite rank, arXiv:1402.3705.
[6] P-E. Caprace, Y. Cornulier, N. Monod, R. Tessera, Amenable hyperbolic groups,
arXiv:1202.3585.
[7] B. Duchesne, Y. Glasner, N. Lazarovich, J. Lécureux, Geometric density for invariant
random subgroups of groups acting on CAT (0) spaces, arXiv:1409.8007.
10
[8] F. Dahmani, V. Guirardel, D. Osin, Hyperbolically embedded subgroups and rotating
families in groups acting on hyperbolic spaces, arXiv:1111.7048; Memoirs AMS, to
appear.
[9] Y. Glasner, Invariant random subgroups of linear groups (with an appendix by Y.
Glasner and T. Gelander), arXiv:1407.2872.
[10] T. Gelander, A lecture on Invariant Random Subgroups, arXiv:1503.08402.
[11] E. Ghys, P. de la Harpe, Sur les groupes hyperboliques d’après Mikhael Gromov.
Progress in Mathematics, 83. Birkhäuser Boston, Inc., Boston, MA, 1990.
[12] M. Gromov, Hyperbolic groups, Essays in Group Theory, MSRI Series, Vol.8, (S.M.
Gersten, ed.), Springer, 1987, 75-263.
[13] E. Ghys, P. de la Harpe, Sur les groupes hyperboliques d’après Mikhael Gromov.
Progress in Mathematics, 83. Birkhäuser Boston, Inc., Boston, MA, 1990.
[14] M. Hamann, Group actions on metric spaces: fixed points and free subgroups,
arXiv:1301.6513.
[15] M. Hull, D. Osin, Transitivity degrees of countable groups and acylindrical hyperbolicity, arXiv:1501.04182.
[16] A. Minasyan, D. Osin, Acylindrical hyperbolicity of groups acting on trees, Math. Ann.
362 (2015), no. 3-4, 1055-1105.
[17] D. Osin, On acylindrical hyperbolicity of groups with positive first ℓ2 -Betti number,
Bull. Lond. Math. Soc. 47 (2015), no. 5, 725-730.
[18] D. Osin, Acylindrically hyperbolic groups, Trans. AMS, to appear; arXiv:1304.1246.
[19] Z. Sela, Acylindrical accessibility for groups, Invent. Math. 129 (1997), no. 3, 527-565.
Denis Osin: Department of Mathematics, Vanderbilt University, Nashville 37240, U.S.A.
E-mail: [email protected]
11
| 4math.GR
|
Noname manuscript No.
(will be inserted by the editor)
A hybrid swarm-based algorithm for single-objective optimization
problems involving high-cost analyses
arXiv:1402.5830v2 [math.OC] 2 May 2016
Enrico Ampellio · Luca Vassio
Received: May 26, 2015 / Accepted: date
Abstract In many technical fields, single-objective optimization procedures in continuous domains involve
expensive numerical simulations. In this context, an improvement of the Artificial Bee Colony (ABC) algorithm, called the Artificial super-Bee enhanced Colony (AsBeC), is presented. AsBeC is designed to provide
fast convergence speed, high solution accuracy and robust performance over a wide range of problems. It implements enhancements of the ABC structure and hybridizations with interpolation strategies. The latter are
inspired by the quadratic trust region approach for local investigation and by an efficient global optimizer for
separable problems. Each modification and their combined effects are studied with appropriate metrics on a
numerical benchmark, which is also used for comparing AsBeC with some effective ABC variants and other
derivative-free algorithms. In addition, the presented algorithm is validated on two recent benchmarks adopted
for competitions in international conferences. Results shows remarkable competitiveness and robustness for
AsBeC.
Keywords modified Artificial Bee Colony · engineering optimization · interpolation strategies · algorithm
comparison
1 Introduction
Numerical optimization in applied sciences usually concerns simulation-based problems in which computational analyses are expensive in terms of time and resources (S. Rao, 2009; K. Deb, 2012). In engineering,
complex non-linear models are commonly used (Lagaros and Papadrakakis, 2015), e.g., computational fluid
dynamics and finite element methods. This field demands fast algorithms for solving optimization problems
using as few function evaluations (FEs) as possible. The present work focuses on single-objective problems
with low, but not minimal dimensionality.
In the literature, there are several architectures and algorithms suitable for global optimization in tough
numerical problems (Floudas and Pardolos, 2009). Among the most studied are surrogates (also called metamodels), heuristic and statistical methods, trust region approaches and nature-based algorithms.
Surrogate models (Koziel et al., 2013) are effectively exploited in a large class of real-world applications.
They are often the first choice and sometimes the only practical one when very expensive analyses are involved. Response surfaces (Box and Draper, 2007), artificial neural networks (Iliadis and Jayne, 2015), radial
E. Ampellio ( )
Dipartimento di Ingegneria Meccanica e Aerospaziale, Politecnico di Torino, 10129 Torino (TO), Italy
E-mail: [email protected]
L. Vassio
Dipartimento di Elettronica e Telecomunicazioni, Politecnico di Torino, 10129 Torino (TO), Italy
E-mail: [email protected]
2
Enrico Ampellio, Luca Vassio
basis functions (Buhmann, 2003) and Kriging (Simpson et al., 2001) are surrogates widely adopted to support
high computing load optimizations. One of the first optimizers fully based on meta-models is the Efficient
Global Optimization (Jones et al., 1998). This algorithm and its successors are expressively targeted to problems where function evaluations are strictly limited by time or cost. Heuristic and statistical techniques such
as Sequential Parameter Optimization (SPO, Bartz-Beielstein et al., 2005) are also related to meta-models and
intensively used in applicative fields. They are focused on selecting and tuning both the optimization methodology and the search methods, but they imply that efficient algorithms and/or surrogates for a specific problem
are a priori available. Nonetheless, ensuring the reliability of surrogates is still a sensitive issue and the training of meta-models could not be cost-effective when considering highly non-linear, noisy, multi-modal and
discontinuous problems. Moreover, this approach becomes computationally impracticable when increasing the
problem dimensionality. As a result, it is attractive to develop direct search algorithms designed for costly
optimizations, especially to cope with higher dimensions (e.g., 10 dimensions).
Among direct search algorithms, derivative-free trust region optimizers (Conn et al., 2000), such as the
quadratic UOBYQA (Powell, 2000) and related approaches (e.g., CONDOR by Berghen and Bersini, 2005;
BOBYQA by Powell, 2009), are effective for heavy computing loads and noisy objective functions. The hybridization with quadratic local meta-models is also used to tackle costly problems, for instance in evolutionary
strategies (lmm-ES by Kern et al., 2006), in which surrogates are periodically called to improve performance
and save expensive function evaluations (FEs).
Meta-heuristic direct-search algorithms play a predominant role when dealing with single-objective optimizations that comprise many continuous variables and noisy multi-modal target functions, and where analytical properties such as differentiability are not known a priori (Talbi, 2009). Several nature-based techniques
have been developed such as genetic algorithms (Goldberg, 1989), particle swarm optimization (Kennedy and
Eberhart, 1995), ant colony optimization (Dorigo and Stützle, 2004), differential evolution (DE by Price et al.,
2005) and evolutionary strategies (Beyer and Schwefel, 2002).
One of the more recent and promising nature-inspired meta-heuristic methods is the Artificial Bee Colony
algorithm (ABC) by Karaboga and Akay (2007). ABC has been compared with other nature-inspired algorithms and it showed good results for a great variety of optimization problems (Karaboga and Akay, 2009)
as well as for real-world applications (Akay and Karaboga, 2010). Besides, a lot of test case validations and
advancements have been done in recent years (Bolaji et al., 2013). The present paper further improves the accuracy and convergence speed of ABC algorithms to make them suitable for continuous optimization problems
with expensive analyses. This means obtaining better solutions using the same maximum number of function evaluations. The main idea behind the proposed algorithm is to hybridize ABC with the basic principles
of derivative-free, quadratic trust region optimizers for high-cost objective functions. Moreover, another hybridization with the fast and robust S.T.E.P method (Swarzberg et al. 1994) is introduced to efficiently cope
with separable functions. In the end, AsBeC aims at showing robust performance over a wide range of problems
by successfully merging the main advantages of swarm-based and interpolative methods.
To introduce the main peculiarities of ABC, a brief review of the original algorithm and its popular modifications is presented in Section 2. The proposed enhancements and hybridizations leading to the novel Artificial
super-Bee enhanced Colony algorithm (AsBeC) are then discussed in Section 3. The analytical benchmarks
and new metrics used for comparisons are respectively introduced in Section 4 and Section 5. The performance
of the new algorithm is accurately evaluated and compared to ABC through extensive simulations in Section 6.
In Section 7, some ABC variants and other state-of-the-art algorithms for direct search and appropriate for
costly optimizations are considered for comparisons. In Section 8, the algorithm is validated using two other
famous benchmarks for expensive optimization. Conclusions and insights about future work on the topic are
presented in Section 9. An electronic appendix gathers all comparative results in expanded form.
2 ABC and its modifications in the literature
The Artificial Bee Colony (ABC) algorithm by Karaboga (2007) is an example of swarm intelligence as it
models the collective behaviour of decentralized and self-organized systems consisting of simple agents. The
algorithm reproduces the behaviour of a honey bee colony searching the best nectar source in a target area.
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
3
Some bees, called the employees, are each assigned to a nectar source and search the space near it. Then, they
come back to the hive and communicate to other bees, called the onlookers, the position of the best nectar
sources found through a dance (Frish, 1967), so that onlookers can help the employed bees in searching the
most promising regions. Nectar sources that reveal themselves non-productive are abandoned in favour of new
positions randomly chosen in the search domain, which are investigated by scout bees.
In the optimization context, nectar sources represent the seed solutions and the nectar amount corresponds
to the objective function value to optimize; ”non-productive” sources are those not improving for some time,
controlled by the limit parameter. An entire optimization sequence composed by the employee group, performing the exploration, the onlooker group, performing the exploitation, and the scout bee, maintaining the
population diversity, is called a cycle. For an ABC pseudo-code refer to Section 2.1 in Karaboga and Akay
(2007).
The basic search movement for a seed solution x j is called mutation equation and it modifies only one
variable at a time, i, chosen uniformly at random. Another seed solution xk 6= x j is also chosen uniformly at
random and the new candidate solution x j,rnd in variable i is:
xij,rnd = xij + rnd[−1, 1] · (xij − xki )
(1)
where rnd[−1, 1] is a uniformly distributed random number in the continuous range [−1, 1].
The ABC algorithm tries to balance exploration and exploitation behaviours, offering worthy global and local search skills at once. As a result, when compared with other methods, ABC demonstrates high-performance
and suitability for a broad spectrum of optimization problems. To enter in details consult Karaboga (2007) and
for extensive simulation comparisons refer to Karaboga and Akay (2009). According to many authors, ABC
algorithm is simple to implement, easy to be effectively parallelized (e.g., Subotic et al., 2011) and hybridized,
driven by few control parameters, flexible and robust (e.g., Karaboga et al. 2014). On the other hand, its semirandom movement is not taking into account the local shape of the function. Moreover, the algorithm does not
exploit the history of the tested solutions.
Since the original paper by Karaboga (2007), many researches were developed to overcome the deficiencies
of the original ABC. In particular, many scientists focused on the common goal to advance the local search
mechanism without worsening the global one. In this regard, specific modifications of various nature have
proven to be convenient (Bolaji et al., 2013). Among them, the popular ABC variants by Gao and Liu (e.g.,
IABC, 2011 and BABC, 2012) and by Kong et al. (e.g., 2013) are focused on initialization enhancements and
DE/PSO inspired mutation equations. Nevertheless, many preliminary tests performed by the authors of the
present paper showed that these modifications are not of major impact when dealing with a limited number of
FEs and small colonies. Their performance in this framework is comparable with the simpler and well-known
GABC (Zhu and Kwong, 2010), which drives the classic bee movement towards the global best solution.
Some other variants give more importance to the exploitation attitude by altering the fitness function or the
assignation methodology for onlookers (e.g., Subotic, 2011). Among them, JA-ABC5 (Sulaiman et al., 2015)
introduces a multi-stage re-organization of the bees which seems particularly beneficial.
Many modifications have been proposed coupling ABC either with classical optimization methods for local
search such as simplex (HSABC by Kang et al., 2009), the Rosenbrock method (RABC by Kang et al., 2011)
and pattern search (HJABC by Kang et al., 2013), or with other robust and efficient well-known techniques
for global search, such as genetic algorithms (GA-ABC by Zhao et al., 2010) or evolutionary strategies (BEABC by Li and Li, 2012). Among these methods, the RABC based on the Rosenbrock directional search
(Rosenbrock, 1960) shows fast convergence to the global optimum.
Despite the great amount of available literature on ABC variants, to the best of our knowledge no paper
underlines a performance gain even with very few function evaluations. As well, no hybridization with methods
similar to trust region techniques has been analysed so far, even though they are recognized to be among the
fastest techniques, especially when dimensionality is small (Rios and Sahinidis, 2013). These reasons justify
the introduction and analysis of ABC modifications specifically targeted to costly optimization and real-world
oriented problems.
4
Enrico Ampellio, Luca Vassio
3 Improving techniques for the AsBeC algorithm
Among all the proposed modifications of the ABC algorithm in the literature, some techniques are on average
more effective or more robust (Liao et al.,2013 and Aydin, 2015) for specific problems. In the present work,
the authors discover that modifying the onlooker assignment stage, exploiting the data history according to
the theory of collective memory in social animals (Couzin et al., 2002) and applying some hybridizations like
local interpolation are especially beneficial when the number of FEs is very limited. The presented techniques
have the purpose of achieving the fastest solution improvement, without clustering the swarm and leading to
premature convergence at the same time.
The sections below describe each technique in details. A minimization problem is considered without loss
of generality. The final AsBeC pseudo-code is outlined and commented in Section 3.6.
3.1 Strictly biased onlooker assignment
In ABC, onlookers are assigned to seed solutions by a stochastic rule. A certain probability is associated to
each of them: it is proportional to the solution quality f , through a fitness function fit. Probability and fitness
definitions are taken from Akay and Karaboga (2012). However, if the value of f for all the seed solutions is
similar, that fitness function returns almost identical values and the relative probabilities are alike. Hence, the
standard fitness formulation is not able to distinguish the solution quality in relative terms. In order to always
strengthen exploitation, it is possible to unbalance onlooker assignation by normalizing the fitness fit to a new
re-scaled function fitres :
fitres (xk ) =
fit(xk ) − mins=1,...,SN f it(xs )
maxs=1,...,SN fit(xs ) − mins=1,...,SN fit(xs )
∀k ∈ {1, . . . , SN}
where xk is a seed solution and SN is the number of seed solutions. The above formulation for fitres (xk )
is generalized and still valid even when changing the definition of fit. Different from the rank-based fitness
transformation of Kang et al. (2011 and 2013), it takes into account also the quality difference among seed
solutions. In fact, fitres scores the seed solutions from 0 to 1 according to their distance from the best.
In the original ABC formulation, there is a possibility to assign few or no bees to the best seed solutions.
The previous scenario is unlikely when dealing with a large-size swarm; however, for small-size colonies a
different approach should be preferred in order to save function evaluations. A deterministic rule for onlookers
assignation is proposed in place of the original stochastic one, in order to associate a prearranged number of
bees nk to seed solution k proportionally to its fitres :
$
ON · fitres (xk )
n =
res s
∑SN
s=1 fit (x )
k
%
∀k ∈ {1, . . . , SN}
where ON is the onlooker number and the operator ⌊·⌋ returns the largest integer smaller than the arguk
ment. In case the number of onlookers assigned, ∑SN
k=1 n , is lower than the total number of onlookers ON, the
k is assigned to the best seed solution. In practice, no bees are ever sent to the worst
difference ON − ∑SN
n
k=1
seed solution and at least one is always sent to the best one. The application of this deterministic rule is similar to the idea of exploiting only few best solutions of the JA-ABC5 algorithm and it always encourages the
repositioning near the global best, as in GABC; however, it is self-adapted during the optimization process.
Notice that this strictly biased onlooker (BO) assignment, through fitness re-scaling and deterministic rule,
does not appreciably affect the global search skills of the swarm. In fact, exploration is essentially up to the employee movement, which remains unaltered. This technique intentionally disregards the worst areas, similarly
to the principle driving other swarm-based algorithms such as the FIREFLY algorithm (Yang, 2009).
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
5
3.2 Postponed onlooker dance
This modification enables the onlooker bees to seek nectar near the best regions for a longer time. The entire
onlooker group performs more than one iteration sequentially, without altering the updating order of seed
solutions inside each iteration. As in the original ABC, a seed solution is immediately updated whenever a
bee finds an improvement for it. In the postponed onlooker dance (PD) iterations, seed solutions are never
reassigned, but their updated versions are the starting configurations for the new movements. In this way, each
onlooker bee has multiple chances to improve the seed solution to which it was assigned and to interact with
the other proposed techniques. Anyway, the number of PD iterations should be maintained small in order not
to affect exploration and decelerate the convergence speed to the global optimum. In this paper, the number of
reiterations on the onlooker group is set to three.
This modification acts like enlarging the onlooker group, but repeating their movements instead of increasing their number. Hence, the best seed solutions have a higher probability to improve and to be selected for
mutation than the others, as onlookers are not distributed on a larger basis.
This postponed onlooker dance among bee groups is inspired by nature. In fact, bees normally come back
to the hive to communicate their seed solutions only after some time, when there is more probability to have
collected new pollen. This technique is inspired by the work of Subotic (MO-ABC, 2011), except that the PD
proposed is a deterministic procedure.
3.3 Local interpolation
In order to quickly improve the best solutions in their local neighbourhood, the concepts of opposition principle
and parabolic interpolation are introduced.
The opposition-based learning for the ABC algorithm has already been suggested in the OABC by El-Abd
(2011). In the present paper, the following version is proposed: whenever a pseudo-random movement for the
seed solution k to a new position xk,rnd does not produce any improvement, it is possible to move in the same
direction with the same step, but in the opposite sense. If this new position xk,opp is better than the previous, the
correspondent seed solution k is updated. As a result, the opposite principle is a linear local estimator when the
step kxk − xk,rnd k is small enough and the objective function is continuously differentiable in the neighborhood
of xk . The operator k·k expresses the Euclidean norm.
On the other hand, the parabolic interpolation estimates the local and directional curvature of the objective function, acting as a second order optimization method with partial Hessian computation. This technique
follows the opposition principle: if the opposite step is not successful, the multi-dimensional parabola passing
through three already known points, i.e., (i) the point xk , (ii) the point generated by the first random movement
xk,rnd and (iii) the opposite point xk,opp , is calculated and its minimum is tested. The opposition principle and
second order interpolation are performed in excluding sequence: if the first random movement is improving,
none of the two following steps is carried out; if the opposite movement is improving, the parabola will not
be estimated; if the two previous steps lead to solutions worse than the seed solution, then the second order
interpolation is computed.
Local interpolation (LI) is used in ABC when at least three onlooker bees are assigned to the same seed solution. When the PD technique is activated, the onlooker group has more chances to exploit local interpolation.
During the onlooker phase, the previous bee movements relative to the same seed solution are recorded and
shared among bees, also considering different PD iterations. When performing local interpolation, it is possible to use candidate solutions generated by different onlooker bees that are assigned to the same seed solution.
Using three iterations for the PD technique, a seed solution with just one onlooker assigned could still perform
a complete sequence of random, opposite and parabolic interpolation movements.
The raw approximation offered by local interpolation is balanced by the low number of function evaluations required, i.e., respectively one and two additional FEs for the opposite movement and the parabolic
interpolation, regardless of the problem dimensionality.
6
Enrico Ampellio, Luca Vassio
3.4 Quadratic prophet
A hybridization with trust region concepts is introduced in order to improve the convergence speed, especially
in the early phases. The quadratic prophet (QP) hybridizes ABC with a fast and robust local search method,
which is the same philosophy of other hybrids such as RABC (Kang et al., 2011). The key idea is to exploit
the function evaluations already performed to create multiple quadratic polynomial interpolators around the
best regions. Hence, these quadratic models are built in the neighborhood of the seed solutions by using their
closest samplings in the so-called hive memory. This memory consists in all the solutions generated by the
AsBeC algorithm so far. The global minima of the quadratic models, if they exist, are then tested. The lmmCMAES algorithm by Kern et al. (2006) is similar, but it uses the quadratic models as response surfaces to
be called in place of direct function evaluations. Instead, QP does not require additional FEs to populate the
samplings for the interpolations; its final cost is at the most equal to the number of seed solutions if all quadratic
models succeeded. Whenever the QP finds a better solution, the related seed solution is immediately updated.
Although this method does not regulate directly the trust region radius, it adapts itself in accordance to
the hive memory. In fact, if the samplings around a given seed solution are dense, the algorithm is exploiting
the area, and thus the trust region radius is small in order to refine the local solution. On the contrary, when
the samplings are sparse, the trust region radius is large to support exploration. This self-adaptation resembles
the mechanism used in the CMA-ES algorithm (Hansen, 2006). Other famous trust region methods, such as
the classic Levenberg-Marquardt (Levenberg, 1944; Marquardt, 1963) or the recent BOBYQA (Powell, 2009)
algorithms follow the same basic principle. The main difference is that the QP technique uses the data that are
already available within the AsBeC algorithm, instead of adopting a costly and complex population strategy to
train and use quadratic approximations. On the other hand, the quality of the QP model depends on the quality
of the samplings, hence on the ability of AsBeC to properly investigate the search space surrounding the
minima. Therefore, the Quadratic Prophet technique can hybridize any optimization algorithm able to provide
well-distributed samplings.
The implementation of QP is based on quadratic polynomial surfaces, whose complete form, QC , and
reduced form, QR , for a generic problem in D dimensions are expressed as:
D
QC (x1 , x2 , ..., xD ) = ∑
D
D
∑ ai j · xi · x j + ∑ bi · xi + R
i=1 j=1
i=1
D
D
i=1
i=1
QR (x1 , x2 , ..., xD ) = ∑ aii · x2i + ∑ bi · xi + R
where ai j , bi and R are the C coefficients of the model, while xi are the independent variables of the problem.
QC needs C = 0.5 · (D + 2) · (D + 1) samplings to be exactly defined, since ai j = a ji . The reduced quadratic
surface QR , without any mixed term, only needs 2 · D + 1 samplings. The reduced model is not able to capture
rotated functions, but it is enough to cope with bowl shape regions. In order to activate this technique as soon as
possible, reduced models QR are built starting when the algorithm provides at least 2 · D + 1 solutions, ending
up with complete models QC when more than C samplings are available.
The model coefficients are obtained in the simplest possible way, that is by solving exactly the linear system
of equations for the C samplings nearest to a given seed solution. If the Hessian matrix of the model is diagonal
positive, then the point in which the Jacobian determinant is zero is tested. This point could represent either
a local minimum, if the Hessian is positive definite, or a saddle point dominated by squared terms. Actually,
the latter can be wrongly recognized as a saddle instead of a minimum due to inaccuracies. Nonetheless,
it represents a key point to better describe the objective function contour in that region. Anyway, eventual
redundant QP predictions are not tested twice, in order to save precious FEs.
Notice that the model could misinterpret the local behaviour of the objective function if the sampling is
ill-conditioned or if f is not quadratic or noisy in that area, as common in real-world problems. Then, if the
exact solution does not lead to a point to test, the system is solved by means of an iterative method with large
tolerance on the same C samplings. This alternative is aimed at obtaining quickly a simple-to-handle coarse
shape interpreter, passing nearby the sample points and not exactly through them. This inaccurate model could
still be able to capture the objective function profile and drive the search towards the most promising regions.
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
7
The tolerance for the iterative resolution of quadratic systems is set to 10−1 by using the QMR method (Roland
and Nachtigal, 1991).
In order to increase the likelihood of a well-conditioned sampling, the points should be heterogeneously
different in their coordinates. There are two main features that affect the sampling diversity: the standard bee
movement, changing only one variable at a time, and the LI technique, producing sequentially aligned points.
Therefore, when the QP is activated, LI is deactivated and the mutation equation is slightly modified. Following
the idea introduced by Akay and Karaboga (2012) for the MABC, more than one variable is chosen when
moving a bee, according to the standard ABC equation. The number of variables to be changed concurrently is
set as half of the total number of dimensions, i.e., D/2. Clearly, these two modifications overweight exploration
and hence have to be restricted to the early phases of the optimization. In this work, they are adopted while the
total number of solutions tested is lower than 2 · C = (D + 1) · (D + 2). In this way, a large pool of samplings
more likely to be well-conditioned is available for complete quadratic models.
This QP technique is a finer exploitation of the seed solutions and, hence, it is implemented as an additional
exploitative stage at every cycle, after the onlooker phase. It is also used once during the initialization. Indeed,
the number of random initial evaluations is set to be (at least) equal to 2 · D + 1, and a fully reduced QR is
already feasible.
It must be clear that the QP is designed to deal with problems in which the objective function evaluation
represents the bottleneck of computational times. In fact, the computational cost of this technique is dominated
by the resolution of the linear system, that in its general form requires O(C 3 ) operations at each cycle, for each
food source. Given that C increases along with D, the technique becomes numerically heavy when addressing
high dimensionality problems.
3.5 Systematic global optimization
The S.T.E.P algorithm by Swarzberg et al. (1994) shows high convergence rate for highly multi-modal and
complex functions of one variable. It is derivative-free, not population-based, without parameters to tune and
it does not assume any property for the function to optimize. It relies on the history of the points tested during
its functioning. The domain boundaries and a random point between them, called the context, are evaluated.
These three points define two partitions. For each partition is computed the curvature a parabola should have
to enclose the best so far solution inside it. This curvature is called the partition difficulty. The S.T.E.P method
iterates by sampling the center of the partition with smallest difficulty, i.e. the one identified with the greatest
chance of improving the best solution found so far, until a given tolerance is reached. The S.T.E.P. authors
showed that around 50 FEs are sufficient to solve difficult univariate problems, such as Michalewicz’s second
function. The simplest multivariate extension to the S.T.E.P. algorithm solves one dimension at a time. The
recent version by Baudiš and Pošı́k (2015) interleaves the steps of the univariate solvers such that all dimensions
are optimized concurrently, in a similar way to Rosenbrock algorithm (1960). Each time, the dimension to
investigate is chosen following a round-robin scheduling.
Given the fast convergence and robust performance of the S.T.E.P. algorithm, it can be hybridized with
ABC, which is more oriented to exploration. However, in a multivariate environment S.T.E.P. based solvers
are very effective only on additively separable or quasi-separable problems, in which the correlations among
variables are weak or only few variables are correlated. As a consequence, this technique is activated if a test
for quasi-separability is successfully passed. The proposed test investigates the separability in the cheapest
possible way, giving just a necessary but not sufficient condition. For an additively separable function, each
first-order partial derivative ∂ f /∂ xi is independent from other variables x j 6= xi , giving ∂ 2 f /∂ xi ∂ x j = 0. For
each design variable xi , another variable x j with i 6= j is randomly chosen to asses the correlation between them.
Then, the partial derivative ∂ f /∂ xi is estimated for two different values of x j , while all other coordinates are
kept constant. If the two estimations vary below a given tolerance, here chosen as 10−3 , the two variables could
be quasi-separable. The same check is repeated for ∂ f /∂ x j for two different values of xi . In order to estimate
the aforementioned derivatives through variations, four FEs are needed. These four solutions are chosen around
the active seed solutions. If the non-separability condition on the derivatives holds for D couples of variables,
the test is passed. The procedure is interrupted whenever a variable is found to be correlated. This test needs at
8
Enrico Ampellio, Luca Vassio
most 2· D +1 function evaluations, since a complete set of partial derivatives in the same point can be estimated
only once and reused for all the couples.
If this test is successfully passed, the interleaved solver by Baudiš and Pošı́k (2015) is run as a one-shot
technique after the initialization phase, using the best so-far solution as context and investigating all the variables, in deterministic sequence. For this reason, the proposed hybridization is called systematic global optimization (SGO). At the end of the SGO phase, the best seed solution is moved to the best solution found. The
maximum number of evaluations per variable is set to 50, according to the results by Swarzberg et al. (1994).
Then, this technique along with its test uses up to 52 · D + 1 function evaluations.
3.6 The AsBeC algorithm
The combination of all the previous techniques, integrated into ABC, gives the AsBeC algorithm. Its pseudocode is presented in Algorithm 1, indicating the sections in which the new techniques added to ABC are
described. BO and PD are enhancements of ABC regarding bees re-organization, while LI, QP and SGO
are hybridizations to estimate the local and global behaviour of the objective function, without changing the
mutation equation itself. Thanks to these hybridizations, the bees assume new extra abilities and are called
super-bees. This explains the name of the new algorithm: the Artificial super-Bee enhanced Colony (AsBeC).
Actually, the authors experimented also other modifications, e.g. in the mutation equation, and other hybridizations with derivative-free directional search techniques, like the discrete gradient method as in Bagirov
et al. (2008). In the end, only the variants recognized to have an effective role for the specific goals of this work
were implemented within AsBeC.
4 Benchmarks
Three different benchmarks for unconstrained problems in the continuous domain are adopted in this paper.
The first, Set A, is used to test AsBeC techniques, to provide standard settings for the new algorithm and to
compare AsBeC with ABC variants and other notable optimizers. The other two, Set B and C, are popular
benchmarks used for competitions in international conferences, herein presented for comparing AsBeC with
the most recent mehods presented at the CEC 2015 and GECCO 2015 conferences.
4.1 Set A
Set A contains unimodal, multimodal, separable and non-separable functions selected among classic analytical
problems. Their definitions, search domains and main characteristics are described in Table 1. They are all
non-negative and have the global minimum value exactly equal to zero. Functions without an exact analytical
characterization of the optimum, e.g., the Styblinski-Tang, are not considered in order to avoid bias on results.
Furthermore, all the functions chosen are scalable to any dimension of the variable space. For a rather complete summary of classical analytical functions used for global optimization refer to Jamil and Yang (2013).
Almost all functions in Table 1 are very popular among ABC modifications and usually adopted for this kind
of assessment (e.g., Gao and Liu, 2011).
Set A is heterogeneous and relatively simple, since it does not contain any shifted rotated, highly illconditioned or extremely complex function. On the other hand, all the functions comprised present a recognizable overall shape and are potentially solvable using limited FEs. For these reasons, Set A is a good environment
to test the techniques integrated into AsBeC.
Each function is investigated with 10 dimensions and a maximum number of FEs equal to 103 . Each experiment is repeated 300 times. This settings fit the main target of the AsBeC algorithm, i.e., finding the solution
of low-dimensional problems using few FEs. Besides, a long term setting is also introduced, investigating 30
dimensions with maximum number of FEs equal to 5 · 104 and each experiment repeated 30 times. This last
setting is in line with other ABC works (e.g., Akay and Karaboga, 2012) and it is important to assess premature
clustering tendency and refining skills on higher dimensional problems when a lot of FEs are allowed.
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
9
Algorithm 1 The AsBeC pseudo-code
{Initialization}
1: Generate max(number of seed solutions,2 · D + 1) starting random solutions in the search area (Sec. 3.4)
2: Evaluate their quality
3: Identify seed solutions as the best starting solutions and assign employees to them
4: Find minimum of the quadratic model near the best seed solution (Sec. 3.4)
5: if a new best solution has been found then update best seed solution end if
{SGO phase (Sec. 3.5)}
6: if quasi-separability test passed then use interleaved S.T.E.P solver end if
7: if a new best solution has been found then update best seed solution end if
8: repeat
{Employees phase}
9:
for all employees do
10:
Generate new pseudo-random solution near its seed solution
11:
Evaluate the quality of solution
12:
if it is better than current employee’s seed solution then update seed solution end if
13:
end for
14:
15:
16:
17:
18:
19:
20:
21:
{Onlookers phase}
Assign onlookers to the seed solutions depending on their quality, according to a biased deterministic rule (Sec. 3.1)
for all onlookers do
for all postponed dance iterations (Sec. 3.2)
Generate new solution near its seed solution, possibly using LI (Sec. 3.3)
Evaluate the quality of solution
if it is better than current onlooker’s seed solution then update seed solution end if
end for do
end for
22:
23:
24:
25:
26:
{Quadratic Prophet phase (Sec. 3.4)}
for all seed solution do
Collect the samplings nearest to the current seed solution
Build quadratic model in its neighbourhood and find its minimum
if it is better than current seed solution then update seed solution end if
end for
27:
{Scout phase}
if a seed solution is not improved for a limited time then replace it with a new random solution end if
{Best-so-far}
28:
if a new best solution has been found then update global best end if
29: until requirements are met
10−16 is considered as the minimum achievable by each function of Set A, i.e., it is the selected tolerance,
in line with the numerical floating point double precision accuracy of MATLAB. The authors of this paper
verified that increasing the value of this tolerance does not affect the qualitative interpretation of results by
using the performance metrics defined in Section 5. Each run of the algorithm over a function stops at the
maximum allowed number of FEs. Repeating many times the experiment helps in reducing the influence of the
random component.
4.2 Set B
The expensive optimization session at CEC conference 20151 is based on a benchmark, here called Set B,
developed by Qu et al. (2016) of 15 functions. The tolerance used is 10−8 and the number of repetitions is set
to 20. The functions are tested with 10 dimensions and 500 FEs and with 30 dimensions and 1500 FEs.
Set B has been explicitly targeted to costly optimization and therefore it is appropriate for testing AsBeC.
Its functions are shifted rotated, thus analytically not separable, and it includes highly ill-conditioned and very
1
http://www3.ntu.edu.sg/home/EPNSugan/index_files/CEC2015/CEC2015.htm
10
Enrico Ampellio, Luca Vassio
N#
Name
Formulation
Domain
Properties
Minimum
f1
Sphere
2
f (x) = ∑D
i=1 xi
[−100, 100]D
U, S, Simple bowl
f (0, ..., 0) = 0
f2
QuarticR
4
f (x) = ∑D
i=1 ixi + rnd[0, 1)
[−1.28, 1.28]D
U, S, Noisy bowl
f (0, ..., 0) = 0
f3
Step
2
f (x) = ∑D
i=1 (⌊xi + 0.5⌋)
[−100, 100]D
U, S, Jagged bowl
f (0.5, ..., 0.5)= 0
f4
DixonPrice
2
2
f (x) = (x1 − 1)2 + ∑D
i=2 i(2xi − xi−1 )
[−10, 10]D
U, N, Valley
f5
Powell
D/4
f (x) = ∑i=1 [(x4i−3 + 10x4i−2 )2 + 5(x4i−1 − x4i )2 + (x4i−2 −
2x4i−1 )2 + 10(x4i−3 − x4i )4 ]
[−4, 5]D
U, N, Valley
f (3, −1, 0, 1, ..., 3, −1, 0, 1)= 0
f6
Rosenbrock
2
2
2
f (x) = ∑D−1
i=1 [100(xi − xi+1 ) + (xi − 1) ]
f7
Schwefel1.2
i
2
f (x) = ∑D
i=1 [∑ j=1 (xi )]
f8
Schwefel2.22
f9
f (1, ..., 2
D −2
−2 D
2
)=0
[−30, 30]D
U, N, Valley
f (1, ..., 1) = 0
[−100, 100]D
U, N, Ridges
f (0, ..., 0) = 0
D
f (x) = ∑D
i=1 |xi | + ∏i=1 |xi |
[−10, 10]D
U, N, Ridges
f (0, ..., 0) = 0
Zakharov
D
4
D
2
2
f (x) = ∑D
i=1 (xi ) + (∑i=1 0.5i(xi )) + (∑i=1 0.5i(xi ))
[−5, 10]D
U, N, Large plate
f (0, ..., 0) = 0
f 10
Alpine
f (x) = ∑D
i=1 |xi sin(xi ) + 0.1xi |
[−10, 10]D
M, S, Many far minima
f (0, ..., 0) = 0
f 11
Rastrigin
2
f (x) = 10D + ∑D
i=1 (xi − 10 cos(2π xi ))
q
1 ∑D x2 ) − exp( 1 ∑D (cos(2π x ))) + 20 + exp(1)
f (x) = −20 exp(−0.2 D
i
D i=1
i=1 i
x21
x
D
D
f (x) = ∑i=1 4000 − ∏i=1 cos ( √i ) + 1
i
f (x) = sin2 (πω1 ) + ∑D−1 (ωi − 1)2 [1 + 10sin2 (πωi + 1)] + (ωD −
i=1
1)2 [1 + sin2 (2πωD )]
x −1
where ωi = 1 + i 4 ∀i
[−5.12, 5.12]D
M, S, Several minima
f (0, ..., 0) = 0
[−32, 32]D
M, N, Narrow hole
f (0, ..., 0) = 0
[−600, 600]D
M, N, Several minima
f (0, ..., 0) = 0
[−10, 10]D
M, S, Many far minima
f (1, ..., 1) = 0
[−50, 50]D
M, N, Many far minima
f (0, ..., 0) = 0
[−50, 50]D
M, N, Many close minima
f (0, ..., 0) = 0
[−100, 100]D
M, N, Few far minima
f (0, ..., 0) = 0
[−10.24, 10.24]D
M, N, Many close minima
f (1, ..., 1) = 0
f 12
Ackley
f 13
Griewank
f 14
Levy
f 15
Penalized
f 16
Penalized2
f 17
Schaffer
f 18
Whitley
π (10 sin2 (π y ) + (S + (y − 1)2 )) + T
f (x) = D
1
D
2
2
where yi = 1 +0.25(xi +1) , S = ∑D−1
i=1 (yi −1) (1 +10 sin (π yi+1 )) and
D
4
T = ∑i=1 [(100(xi − 10) )(xi > 10) + (100(−xi − 10)4 )(xi < −10)]
f (x) = 0.1(sin2 (π xi ) + (S + ((x(D) − 1)2 )(1 + sin2 (2π xd )))) + T
D
2
2
where S = ∑D−1
i=1 (xi − 1) (1 + sin (3π xi+1 )) and T = ∑i=1 [(100(xi −
5)4 )(xi > 5) + (100(−xi − 5)4 )(xi < −5)]
q
(sin( ∑D x2i )−0.5
i=1
f (x) = 0.5 +
2 2
(1+0.001 ∑D
i=1 xi )
D
2
2
2 2
2
f (x) = ∑D
i=1 ∑ j=1 (((100(xi −x j ) +(1 −x j ) ) )/4000 −cos(100(xi −
x j )2 + (1 − x j )2 ) + 1)
Table 1: The analytical Set A. U=unimodal; M=multimodal; S=separable; N=non-separable;
complex problems, also with extremely jagged shapes. Set B is complementary to Set A, since it tests the
adaptability of the optimization methods also to these kinds of functions. Notice that the maximum number
of admitted FEs in 30 dimensions is really small (50 · D), which is likely to favor a simple exploration of the
space. Further details about the settings and functions can be found in the work by Chen et al. (2015).
4.3 Set C
The noiseless Black-Box Optimization Benchmarking (BBOB)2 set used at CEC 2015 and at GECCO 2015
conferences, here called Set C, is developed by Hansen et al. (2009a and 2009b) and contains 24 functions.
The tolerance and the number of repetitions are respectively set to 10−8 and to 20. In the present work, the
functions are tested with 5 dimensions and 500 FEs and with 20 dimensions and 2000 FEs.
The major advantage of using this benchmark is due to its schematic structure and to the huge amount of literature dealing with it; in fact, at least 151 algorithms have been compared so-far on this benchmark from 2009
to 2015. Set C is not explicitly targeted to costly optimization, since its functions are not meant to be solved
within a strictly limited number of FEs. Set C comprises five different groups of functions: (i) separable, (ii)
low-conditioned, (iii) high-conditioned, (iv) multi-modal with adequate global structure and (v) multi-modal
2
http://coco.gforge.inria.fr
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
11
with weak global structure. This well organized heterogeneous architecture allows to interpret the strengths and
weaknesses of each optimization algorithm. For details about this benchmark refer to Hansen et al. (2009b).
5 Mean logarithmic value metrics
Some quantitative metrics are introduced to quickly compare algorithms in terms of the objective value reached
after a given amount of FEs. A performance estimator for an algorithm is defined as the Logarithmic Value LV :
∗ ( f )
(Mrep ( fbest ( f , FEs, rep)) − fbest
LV ( f , FEs) := log
tolerance
where rep is the run repetition, FEs is the number of performed function evaluations, f is the selected
∗
benchmark function, fbest
is the analytical optimal value, fbest is the best objective value obtained and Mrep
represents the median operator over repetitions. The median is chosen in spite of the mean since it is less
sensible to outliers. The numerator represents the residual and the LV represents the median performance of an
algorithm in terms of orders of magnitude missing to the given tolerance. A LV is always nonnegative and it
reaches zero when the median results of an algorithm are optimal, within the given tolerance.
Three averaged forms of the Logarithmic Value, called Mean Logarithmic Values (MLV s), are defined.
MLVFEs is an average over the number of benchmark functions N f , MLV f is an average over the maximum
number of function evaluations FEsmax and MLVA is an average over both functions and FEs:
N
MLVFEs (FEs) =
MLV f ( f ) =
MLVA =
f
1
· ∑ LV ( f , FEs)
N f f =1
FEsmax
1
· ∑ LV ( f , FEs)
FEsmax FEs=1
FEsmax
1
1
· ∑
FEsmax FEs=1 N f
Nf
∑ LV ( f , FEs)
f =1
MLVA helps in comparing different algorithms through just one real number. It is meaningful since it
considers the performance evolution on the whole set of functions, not only after a specific number of FEs.
The main difference of the introduced metric with respect to a simple mean of the median residuals is the
fact that MLV s average the logarithms of the residuals, and not the mean of the actual values. This helps to
assess how much the algorithm is getting closer to the optimum in relative terms.
The Logaritmic Value and its averaged versions are here introduced for the first time by the authors of this
paper. They will be used for comparisons alongside other classical and benchmark-specific metrics.
6 Analysis of the improving techniques on Set A
Having defined the comparison metrics, the main three parameters of the ABC architecture, i.e., the colony
size, the number of cycles and the limit parameter, are chosen. The limit parameter is set on D · SN as advised
for the original ABC (Karaboga and Akay, 2007), while the number of cycles depends on FEs and on the
colony size. The overall number of agents, N, is a key performance-driving parameter for population-based
algorithms. Six configurations for the total number of individuals (4, 8, 16, 32, 64 and 128 individuals) are
directly investigated and MLVA is the metric used to choose the best one. For the original ABC the best colony
size turns out to be 8 (see Table 2). This properly tuned ABC is used for the comparisons.
At this point, the impact of each single technique proposed in Section 3 is evaluated. They are applied
one-by-one to the reference ABC configuration. Fig. 1a illustrates the correspondent MLVA for each technique
and for the basic ABC. All the techniques have positive effects on the algorithm performance. The QP and
SGO clearly take a dominant role with respect to the other techniques, that show a more limited impact.
12
Enrico Ampellio, Luca Vassio
22
18
17.2
17.2
17.2
20
16
18
15
FEs
17
14.5
MLV
MLV A
17.4
14
12
11
11.5
ABC
BO
PD
LI
QP
Improving techniques
14
12
12.8
13
ABC
AsBeC
others
16
SGO
AsBeC
(a) MLVA for each technique, ABC and AsBeC
10
8
200
400
600
800
1000
FEs
(b) MLVFEs for all the combinations of techniques
Fig. 1: Validation of the techniques using the MLV metrics
The differences between the ABC with or without the implemented techniques presented in Fig. 1a are
statistically significant. The MLVA represents the average of the LV over the evaluations and the functions. A
Student significance t-test has been conducted over the two samples (the samples mean will follow a normal
distribution by the Central Limit Theorem). The null hypothesis that the MLVA of the modified algorithm is
greater than the one of ABC has been tested against the alternative hypothesis that the mean has positive value.
In all the cases, the null hypotheses can be rejected with significance level of 0.05.
Then, all possible combinations of techniques are tested in order to capture their mutual interactions. The
MLVFEs comparisons are shown in Fig. 1b, where each line corresponds to a different combination of techniques. The best configuration (AsBeC line) is the one with lowest MLVA , and it corresponds to the AsBeC
(MLVA of 11.5, reported in Fig. 1a). The Quadratic Prophet, based on a trust region approach, is revealed to be
impressively suited for this kind of problems, as expected.
Once the benefit of the techniques is clear and AsBeC is recognized as the best combination of these
techniques, the number of agents is studied for Set A. The best number of bees results 8 for the setting with 10
dimensions and 32 for the setting with 30 dimensions (see Table 2). Indeed, other tests on different dimensions
showed that the number of individuals should be at least equal to 8 and approximatively proportional to the
number of dimensions. The same AsBeC settings discussed in this section for set A will be also used for set B
and C, in order to prove the robustness of this algorithm configuration.
7 Comparison with other algorithms on Set A
The AsBeC effectiveness has to be tested when compared to other state-of-the-art methods of the same class
suitable for the same goal (see Section 1), consistently with what has been done by Akay and Karaboga (2012)
or Gao and Liu (2011) for their algorithms. According to the considerations in Section 2, GABC, JA-ABC5
and RABC are selected as representative of ABC modifications also suited for the specific goal under study.
The authors have selected other three typologies of direct search algorithms for the comparison: FIREFLY
(Yang, 2009), CMA-ES (Hansen, 2006) and BOBYQA (Powell, 2009). FIREFLY is a notable competitor of
the ABC among the swarm-based techniques, while the evolutionary CMA-ES and the trust region BOBYQA
are recognized as state-of-the-art in their fields and have been widely used for engineering applications.
The implementations used for FIREFLY and CMA-ES are the ones made available by the authors themselves. The MATLAB routine used for BOBYQA is instead the one embedded in the NAG optimization toolbox3 .
3
http://www.nag.co.uk/numeric/MB/start.asp
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
13
7.1 Choice of parameter values on Set A
A key point when comparing different methods consists in tuning the parameters of the optimizers to make them
perform at their best. For all the algorithms considered, the parameter selection is addressed as in Section 6
for ABC and AsBeC. Only the number N of individuals for population-based methods and the number N of
interpolation samples for BOBYQA are directly investigated on Set A. The range studied for N is a progression
of the power of two; its boundaries depend on the specific mechanics driving each algorithm. The best size is
selected according to the MLVA metric and is reported in Table 2.
All other parameters assume the values suggested in literature as suitable for many problems (Zhu and
Kwong, 2010 for GABC, Kang et al., 2011 for RABC, Sulaiman et al., 2015 for JA-ABC5, Hansen, 2006
for CMA-ES and Powell, 2009 for BOBYQA). For FIREFLY, the authors use the parameters proposed by
Yang (2010) for the MATLAB implementation of the algorithm, since they reach better average results on the
adopted benchmark with respect to the general suggestions by Yang and He (2013).
Dimensions
10D
30D
ABC
8
16
GABC
8
16
RABC
8
16
JA-ABC5
16
16
FIREFLY
16
32
CMA-ES
8
32
BOBYQA
32
256
AsBeC
8
32
Table 2: Best N for each algorithm according to the MLVA metric
7.2 Results of comparison on Set A
The analysis of MLVFEs for 10 dimensions is reported in Figure 2a. It shows that AsBeC is the most effective on
average for quickly improving the solution as well as refining it. Referring to function-by-function performance
in the electronic appendix, the AsBeC algorithm is able to outperform, or at least to approach, all the others
during the entire FEs envelope, on all the benchmark functions. In short, its performance is very good and
robust, as it is revealed also by the boxplot comparisons in the electronic appendix. None of the algorithms
appears remarkably better than AsBeC in any function. AsBeC performs always better than the original ABC
and it is the only algorithm able to get close to the global optimum of the Shaffer function. Final objective values
analysis, reported in the electronic appendix, recognize AsBeC as the most appropriate for the benchmark.
Moreover, following the same procedure adopted in Section 6, the improvements of the MLVA of AsBeC with
respect to the other algorithms show statistical significance.
Considering the other algorithms, the two variants GABC and JA-ABC5 perform slightly better than the
original ABC. Instead, RABC is in general faster and more capable also at refining the solution. Nonetheless,
20
20
18
MLVFEs
14
15
ABC
GABC
JAABC5
RABC
FIREFLY
CMAES
BOBYQA
AsBeC
MLVFEs
ABC
GABC
JAABC5
RABC
FIREFLY
CMAES
BOBYQA
AsBeC
16
10
12
10
5
100
200
300
400
500
FEs
600
700
(a) 10D
800
900 1000
0.5
1
1.5
2
2.5
FEs
3
3.5
(b) 30D.
Fig. 2: Set A. Comparison of the algorithms using MLVFEs
4
4.5
5
×10 4
14
Enrico Ampellio, Luca Vassio
12.5
12.5
12
12
11.5
11.5
11
MVMO
CMAS-ES_QR
TunedCMAES
iSRPSO
humanCog
AsBeC
10.5
10
9.5
MLVFEs
MLVFEs
11
MVMO
CMAS-ES_QR
TunedCMAES
iSRPSO
humanCog
AsBeC
10.5
10
9.5
9
9
8.5
8.5
8
8
0
100
200
300
400
500
FEs
0
500
1000
1500
FEs
(a) 10D
(b) 30D
Fig. 3: Set B. Comparison of the algorithms using MLVFEs
it is worse than the original ABC in those functions where finding a preferential direction towards the global
optimum is particularly difficult. FIREFLY is among the fastest ones in reaching the best regions, but its refining
ability is poor. The CMA-ES performance is robust and shows a good growth potential, but it needs more total
FEs to improve. BOBYQA presents the most fluctuating results, performing very well on some functions but
getting stuck in local minima in many others.
Looking at the overall results, it is clear that they are in general not convergent and in many functions
they are still far from the global minimum. Nevertheless, the above comparison is still solid. In fact, in reallike problems where the optimal position is unknown, designers are interested in obtaining the best solution
improvement within a limited amount of resources.
Set A with a 30-dimensions setting is not the main focus of the paper, but provides interesting insights on
long-term performance. The analysis of the MLV in Figure 2b and function-by-function performance reported
in the electronic appendix, reveal that AsBeC is the most promising for quickly improving the solution also
in this kind of problems. In Figure 2b, first data are plotted after 103 FEs to improve visualization, showing
that AsBeC reaches a very low value after very few FEs. Moreover, AsBeC is able to reach the given tolerance
in 11/18 of the functions and, among the other functions, only Dixon Price and Rosenbrock can be solved by
other algorithms. In short, AsBeC is the best algorithm up to around 4 · 104 FEs. RABC is the most efficient
at refining if a large number of FEs is available. However, it is clearly much slower than AsBeC. The other
two ABC variants and CMA-ES are quite similar and better than ABC, as revealed by the MLVFEs . BOBYQA
is again among the fastest ones during the very first optimization phases. FIREFLY seems to show a slower
convergence speed with respect to the others when dimensionality rises.
8 Other validations
The primary goal of the two validation sets B and C is to assess the robustness of AsBeC, besides its quality, in
comparison to the most recent and proficient optimization methods. For this reason, the same AsBeC standard
settings discussed in Section 6 for Set A will be also used for Set B and C. Following the guidelines of Section
7.1, the bee colony size is chosen as N = 8 for D = 5 and D = 10, N = 20 for D = 20, and N = 32 for D = 30.
8.1 Results of the comparison on Set B
Public results on this benchmark1 by Qu et al. (2016) include the following algorithms:
– MVMO, a population-based stochastic technique with a mapping function for the offspring;
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
15
– CMAS-ES QR and TunedCMAES, variants of CMAES for expensive scenarios;
– iSRPSO, a PSO implementing a dynamic learning strategy for velocity updating;
– humanCog, a 3-layer architecture that mimics human cognitive behaviour.
The MLVFEs comparisons with the AsBeC are presented in Figure 3. Extended function-by-function plots
of median residual over repetitions as function of FEs and complete tables with mean, median and standard
deviation for final achievements are reported in the electronic appendix. It is evident that AsBeC has a great
overall quality, reaching much lower values that the best method, MVMO. AsBeC is able to solve the quadratic
problems, even if ill-conditioned, to recognize the quasi-separability of the test function 4, and to approach the
best methods in the other functions. Unlike the other algorithms, AsBeC is capable of a stable performance
regardless the dimensionality.
Besides, Qu et al. (2016) defined an official total scoring for this competition as an average of the median
and mean values at the end and at the middle of the computation. The AsBeC total score is one order of
magnitude less than MVMO (full table reported in the electronic appendix). Notice that this scoring favors the
improvements made on the most difficult function, TF 10. Since both AsBeC and MVMO are able to better
minimize this problem, their scores are much lower than those of the other algorithms.
8.2 Results of comparison on Set C
At the CEC and GECCO conferences in 2015, 26 algorithms have competed on this benchmark2 . Indeed, they
are tuned variants of few basic methods. Among them, four top categories can be identified:
–
–
–
–
CMAES derived, including many tuned variants of IPOP-CMAES;
Surrogate based, including the ones that exploit the MATLAB MATSuMoTo Library for metamodels;
DE derived, tuned for cheap, medium, and expensive settings;
Axis-Parallel Brent-S.T.E.P. method, which investigates some variants of the multidimensional BrentS.T.E.P. method.
11.5
12
11
11.5
10.5
11
10
10.5
GP5-CMAES
RAND-2xDefault-MATSuMoTo
R-DE-10e2
Srr
AsBeC
9.5
9
MLVFEs
MLVFEs
For each category, the best algorithm in terms of performance and robustness according to the official
results of the conference is compared to the standard AsBeC through the MLVFEs metric. These four tuned
methods are the GP5-CMAES, RAND- 2xDefaultMATSuMoTo, R-DE-10e2 and Srr.
The MLVFEs metric in Figure 3 recognizes the Srr algorithm as the best one, just slightly better than AsBeC.
However, function-by-function plots and BBOB official ECDF metric, reported in the electronic appendix,
point out that Srr is less robust, especially when addressing non separable multi-modal functions. Srr is the best
method for separable problems (f1-5), but it is the worst on some others, reaching premature convergence (e.g.,
9.5
8.5
9
8
8.5
7.5
GP5-CMAES
RAND-2xDefault-MATSuMoTo
R-DE-10e2
Srr
AsBeC
10
8
7
0
100
200
300
400
500
7.5
0
500
1000
1500
FEs
(a) 5D
(b) 20D
Fig. 4: Set C. Comparison of the algorithms using MLVFEs
2000
16
Enrico Ampellio, Luca Vassio
in f7 and f18). Instead, AsBeC and GP5-CMAES offer overall stable high-quality performance. In particular,
AsBeC reaches very good results when solving separable, low-conditioned and weak multi-modal problems.
In all the functions, AsBeC is at least close to the best methods, even if no specific tuning has been studied
for this set.
9 Concluding remarks
Achieving fast and robust improvements in single-objective optimization problems involving expensive analyses means saving precious time and resources. A new swarm-based algorithm hybridized with interpolation
strategies, called Artificial super-Bee enhanced Colony (AsBeC) algorithm, is proposed. The new algorithm
has been designed for solving expensive problems with low dimensionality, using a limited number of function
evaluations. The ambition of this work is to improve the local search, i.e., the exploitation ability, concurrently
preserving the good global attitude of the original ABC, i.e., exploration ability, especially during the first
search phases.
A meaningful metric for comparison is defined, the Mean Logarithmic Value, which takes into account the
evolution of the optimization process and the relative distance to the analytical optimum. All the implemented
techniques are analysed in order to identify a standard robust setting for them. The standard AsBeC algorithm
is compared with ABC, with some of its relevant modifications and with other state-of-the-art direct search
algorithms. The same standard AsBeC has also been validated on other benchmarks, and it is compared with
some of the latest methods presented at the CEC 2015 and GECCO 2015 conferences, which have specifically been tuned for the benchmarks. The AsBeC algorithm is confirmed to be robust and effective on all the
benchmarks.
This promising outcome paves the way for a useful application of the proposed algorithm, especially in
engineering. Indeed, the basic principles presented in this paper already showed interesting results in the past
when applied by the authors to turbine design (Bertini et al., 2013). Future works will include comparisons on
real-world optimization problems, introduction of parallel strategies and extension to multi-objective problems.
Acknowledgements The authors would like to thank GE Avio S.r.l. and its Engineering Technologies department, especially Ing.
F. Bertini and Ing. E. Spano. Their collaboration was fundamental for shaping the AsBeC algorithm within an industrial application
framework. We are also grateful to Prof. E. Benini (Università di Padova) for his comments and reviews.
References
Akay, B., & Karaboga, D. (2010). Artificial bee colony algorithm for large-scale problems and engineering
design optimization. Journal of Intelligent Manufacturing, 23(4), 1001-1014.
Akay, B., & Karaboga, D. (2012). A modified artificial bee colony algorithm for real-parameter optimization.
Information Sciences, 192(0), 120-142.
Aydin, D. (2015). Composite artificial bee colony algorithms: From component-based analysis to highperforming algorithms. Applied Soft Computing, 32(0), 266-285.
Bagirov, A. M., Karaszen, B. & Sezer, M. (2008). Discrete gradient method: Derivative-free method for nonsmooth optimization. Journal of Optimization Theory and Application, 137(2), 317-334.
Bartz-Beielstein, T., Lasarczyk, C., & Preuß, M. (2005). Sequential parameter optimization. In Proceedings of
the 2005 IEEE Congress on Evolutionary Computation (Vol. 1, pp. 773-780). Piscataway, NJ: IEEE Press.
Baudiš, P., & Pošı́k, P. (2015). Global Line Search Algorithm Hybridized with Quadratic Interpolation and Its
Extension to Separable Functions. In Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation (pp. 257-264). New York, NY: ACM.
Berghen, F. V., & Bersini, H. (2005). CONDOR, a new parallel, constrained extension of Powell’s UOBYQA
algorithm: Experimental results and comparison with the DFO algorithm. Journal of Computational and
Applied Mathematics, 181(1), 157-175.
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
17
Bertini, F., Dal Mas, L., Vassio, L., & Ampellio, E. (2013). Multidisciplinary optimization for gas turbines
design. XXII AIDAA Conference. http://arxiv.org/pdf/1402.0420v1.pdf. Accessed February 5,
2016.
Beyer, H.G. & Schwefel, H.P. (2002). Evolution strategies: A comprehensive introduction. Natural Computing,
1(1), 3-52.
Bolaji, A., Khader, A., Al-Betar, M., & Awadallah, M. (2013). Artificial Bee Colony algorithm, its variants and
applications: a survey. Journal of Theoretical and Applied Information Technology, 47(2), 434-459.
Box, G. E. P., & Draper, N. (2007). Response surfaces, mixtures, and ridge analyses, Second Edition. Hoboken,
NJ: John Wiley & Sons.
Buhmann, M. D. (2003). Radial Basis Functions: Theory and Implementations. Cambridge, UK: Cambridge
University Press.
Chen, Q., Liu, B., Zhang, Q., Liang, J. J., Suganthan, P. N. & Qu B. Y (2015). Problem Definitions and
Evaluation Criteria for CEC 2015 Special Session on Bound Constrained Single-Objective Computationally
Expensive Numerical Optimization. Technical Report, Computational Intelligence Laboratory, Zhengzhou
University, Zhengzhou, China, Nov 2014.
Conn, A. R., Gould, G. & Toint, P. (2000). Trust region methods. Philadelphia, PA: SIAM.
Couzin, I. D., Krause, L., Ruxton, G. D., & Franks, N. R. (2002). Collective memory and spatial sorting in
animal groups. Journal of Theoretical Biology, 128(1), 1-11.
Deb, K. (2012). Optimization for engineering design: Algorithms and examples, Second Edition. New Delhi,
India: PHI Learning Pvt.
Dorigo, M., & Stützle, T. (2004). Ant colony optimization. Cambridge, MA: MIT Press.
El-Abd, M. (2011). Opposition-based artificial bee colony algorithm. In Proceedings of the 13th annual Conference on Genetic and Evolutionary Computation (pp. 109-116). New York, NY: ACM.
Floudas, C., & Pardolos, M. (2009). Encyclopedia of Optimization, Second Edition. New York, NY: Springer.
Frish, K. (1967). The Dance Language and Orientation of Bees. Cambridge, MA: Harvard University Press.
Gao, W., & Liu, S. (2011). Improved artificial bee colony algorithm for global optimization. Information Processing Letters, 111, 871-882.
Gao, W., & Liu, S. (2012). A global best artificial bee colony algorithm for global optimization. Journal of
Computational and Applied Mathematics, 236(11), 2741-2753.
Goldberg, D. (1989). Genetic Algorithms in Search, Optimization and Machine Learning. Boston, MA:
Addison-Wesley Longman Publishing.
Hansen, N. (2006). The CMA Evolution Strategy: A Comparing Review. In Lozano et al., Towards a new
evolutionary computation. Advances in estimation of distribution algorithms (pp. 75-102). Berlin, Germany:
Springer.
Hansen, N., Auger, A., Finck, S., & Ros, R. (2009a). Real-parameter black-box optimization benchmarking
2009: Experimental setup. Technical Report RR-6828, INRIA.
Hansen, N., Finck, S., Ros, R., & Auger, A. (2009b). Real-parameter black-box optimization benchmarking
2009: Noiseless functions definitions. Technical Report RR-6829, INRIA.
Iliadis, J., & Jayne, C. (Eds.). (2015). Engineering Applications of Neural Networks: Proceedings of the 16th
International Conference on Engineering Applications of Neural Networks. Berlin, Germany: Springer.
Jamil. M., & Yang, X. S. (2013). A literature survey of benchmark functions for global optimization problems.
International Journal of Mathematical Modelling and Numerical Optimization, 4(2), 150-194.
Jones, D. R., Schonlau, M., & Welch W. J. (1998). Efficient global optimization of expensive black-box functions. Journal of Global Optimization, 13(4), 455-492.
Kang, F., Li. J., & Xu, Q. (2009). Structural inverse analysis by hybrid simplex artificial bee colony algorithms.
Computers and Structures, 87(13-14), 861-870.
Kang, F., Li. J.,& Ma. Z. (2011). Rosenbrock artificial bee colony algorithm for accurate global optimization
of numerical functions. Information Sciences, 181(16), 3508-3531.
Kang, F., Li, J. & Li, H. (2013). Artificial bee colony algorithm and pattern search hybridized for global
optimization. Applied Soft Computing, 13(4), 1781-1791.
18
Enrico Ampellio, Luca Vassio
Karaboga, D. (2007). A powerful and efficient algorithm for numerical function optimization: artificial bee
colony (ABC) algorithm. Journal of Global Optimization, 39(3), 459-471.
Karaboga, D., & Akay, B. (2007). Artificial Bee Colony (ABC) optimization algorithm for solving constrained
optimization problems. In Melin et al., Foundations of Fuzzy Logic and Soft Computing: Proceedings of
the 12th International Fuzzy Systems Association World Congress 2007 (pp. 789-798). Berlin, Germany:
Springer.
Karaboga, D., & Akay, B. (2009). A comparative study of Artificial Bee Colony Algorithm. Applied Mathematics and Computations, 214(1), 108-132.
Karaboga, D., Gorkemli, B., Ozturk, C., & Karaboga, N. (2014). A comprehensive survey: artificial bee colony
(ABC) algorithm and applications. Artificial Intelligence Review, 42(1), 21-57.
Kennedy, J., & Eberhart, R. (1995). Particle Swarm Optimization. In Proceedings of IEEE International Conference on Neural Networks 1995 (pp. 1942-1948). Piscataway, NJ: IEEE Press.
Kern, S., Hansen, N., & Koumoutsakos, P. (2006). Local meta-models for optimization using evolution strategies. In Runarsson et al., Proceedings of the 9th International Conference of Parallel Problem Solving from
Nature - PPSN IX (pp. 939-948). Berlin, Germany: Springer.
Kong, X., Liu, S., & Wang, Z. (2013). An improved Artificial Bee Colony algorithm and its application. International Journal of Signal Processing, Image Processing and Pattern Recognition, 6(6), 259-274.
Koziel, S., & Leifsson, L. (2013). Surrogate-Based Modeling and Optimization, Applications in Engineering.
New York, NY: Springer.
Lagaros, N., & Papadrakakis, M. (Eds.). (2015). Computational Methods in Applied Sciences: Engineering and
Applied Sciences Optimization. Cham, Switzerland: Springer.
Levenberg, K. (1944). A method for the solution of certain non-linear problems in least square. Quarterly
Journal of Applied Mathematics, 2, 164-168.
Li, B., & Li, Y. (2010). BE-ABC: Hybrid Artificial Bee Colony Algorithm with Balancing Evolution Strategy.
In Proceedings of the 3rd International Conference on Intelligent Control and Information Processing (pp.
217-222). Piscataway, NJ: IEEE Press.
Liao, T., Aydin, D., & Stützle, T. (2013). Artificial bee colonies for continuous optimization: Experimental
analysis and improvements. Swarm Intelligence, 7(4): 327-357.
Marquardt, D. (1963). An Algorithm for Least-Squares Estimation of Nonlinear Parameters. SIAM Journal on
Applied Mathematics, 11(2), 431-441.
Powell, M. (2000). UOBYQA: Unconstrained optimization by quadratic approximation. Technical Report No.
DAMTP2000/14, Department of Applied Mathematics and Theoretical Physics, University of Cambridge,
England.
Powell, M., (2009). The BOBYQA algorithm for bound constrained optimization without derivatives. Technical report NA2009/06, Department of Applied Mathematics and Theoretical Physics, University of Cambridge, England.
Price, K., Storn, R., & Lampinen, J. (2005). Differential Evolution - A Practical Approach to Global Optimization. Berlin, Germany: Springer.
Qu, B. Y., Liang, J. J. , Wang, Z. Y. , Chen, Q., & Suganthan, P. N. (2016). Novel Benchmark Functions for
Continuous Multimodal Optimization with Comparative Results. Swarm and Evolutionary Computation, 26,
23-34.
Rao, S.(2009). Engineering Optimization: Theory and Practice, Fourth Edition. New York, NY: Wiley.
Rios, L. M., & Sahinidis, N. V., (2013). Derivative-free optimization: a review of algorithms and comparison
of software implementations. Journal of Global Optimization, 56(3), 1247-1293.
Roland, F. W., & Nachtigal, N. M. (1991). QMR: A quasi-minimal residual method for non-Hermitian linear
systems. Numerische Mathematik. 60(1), 315-339.
Rosenbrock, H. H., (1960). An automatic method for finding the greatest or least value of a function. The
Computer Journal, 3(3), 175-184.
Simpson, T. W., Mauery, T. M., Korte, J. J., & Mistree, F. (2001). Kriging Models for Global Approximation
in Simulation-Based Multidisciplinary Design Optimization. AIAA Journal, 39(12), 2233-2241.
A hybrid swarm-based algorithm for optimization problems involving high-cost analyses
19
Subotic, M., Tuba, M., & Stanarevic, N. (2011). Different approaches in parallelization of the artificial bee
colony algorithm. International Journal of Mathematical Models and Methods in Applied Sciences, 5(4),
755-762.
Subotic, M. (2011). Artificial bee colony algorithm with multiple onlookers for constrained optimization problems. In Proceedings of the 5th European Computing Conference (pp. 251-256). Stevens Point, WI: World
Scientific and Engineering Academy and Society.
Sulaiman, N., Mohamed-Saleh, J., & Abro, A., G. (2014). New enhanced Artificial Bee Colony (JA-ABC5)
algorithm with application for reactive power optimization. The Scientific World Journal, 2015, 1-11.
Swarzberg, S., Seront, G., & Bersini, H. (1994), S.T.E.P.: the easiest way to optimize a function. In Proceedings
of the 1st IEEE Conference on Evolutionary Computation (pp. 519-524). Piscataway, NJ: IEEE Press.
Talbi, E. (2009). Metaheuristics: from design to implementation. Hoboken, NJ: Wiley.
Yang, X. (2009). Firefly algorithms for multimodal optimization. In Proceedings of the 5th International Conference on Stochastic Algorithms: Foundations and Applications (pp. 169-178). Berlin, Germany: Springer.
Yang, X. (2010). Engineering Optimization: An Introduction with Metaheuristic Applications. Hoboken, NJ:
Wiley.
Yang, X., & He, X. (2013). Firefly algorithm: recent advances and applications. International Journal of Swarm
Intelligence, 1(1), 36-50.
Zhao, H., Pei, Z., Jiang, J., Guan, R., Wang, C., & Shi, X. (2010). A hybrid swarm intelligent method based on
genetic algorithm and artificial bee colony. In Proceedings of the 1st International Conference on Advances
in Swarm Intelligence (pp. 558-565). Berlin, Germany: Springer.
Zhu, G., & Kwong, S. (2010). Gbest-guided artificial bee colony algorithm for numerical function optimization.
Applied Mathematics and Computation, 217(7), 3166-3173.
| 9cs.NE
|
arXiv:1705.01340v1 [stat.ME] 3 May 2017
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
UNDER LEVEL PERMUTATIONS
FABIO RAPALLO AND MARIA PIERA ROGANTIN
Abstract. In this paper we study the behavior of the fractions of a factorial
design under permutations of the factor levels. We focus on the notion of
regular fraction and we introduce methods to check whether a given symmetric
orthogonal array can or can not be transformed into a regular fraction by
means of suitable permutations of the factor levels. The proposed techniques
take advantage of the complex coding of the factor levels and of some tools
from polynomial algebra. Several examples are described, mainly involving
factors with five levels.
1. Introduction
In Design of Experiments the use of fractions of a full factorial design plays
an important role when the observation of the response variable at each of the
possible level combinations of the factors is impracticable. In general, the selected
level combinations must satisfy optimality criteria to conveniently measure the
impact of factors and their interactions on the mean and on the variability of the
response variable. In this framework, orthogonal arrays and regular fractions are
widely used and both are based on properties of orthogonality among factors. The
notion of orthogonality has two main interpretations that coincide in the two level
case: vector orthogonality and combinatorial orthogonality. Vector orthogonality
allows one to construct linear models with non correlated factor effect estimators,
but this concept is relevant with quantitative factors. Combinatorial orthogonality
easily applies both in qualitative and quantitative cases. The complex coding of
factor levels, extensively studied in [20] for asymmetric and multilevel designs, get
together vector and combinatorial interpretations.
In this paper we consider how the orthogonal properties of a fraction change
in presence of permutations of the factor levels, and in particular we give some
methods to check if a given fraction with qualitative factors is or not isomorphic
(or equivalent) to a regular fraction up to permutations of the factor levels. Regular
fractions are special orthogonal arrays where the factors and their interactions are
either orthogonal or totally confounded.
Two fractional factorial designs are said isomorphic if one of them can be obtained from the other by reordering the runs, relabeling the factors and/or permuting the levels of one or more factors. If a factor is quantitative, only the reverse
permutation is allowed. Usually, the isomorphism is referred as geometric when
the factors are quantitative, and as combinatoric when the factors are qualitative.
Mixed situations can occur in practice, see [13].
2010 Mathematics Subject Classification. 62K15; 13P10.
Key words and phrases. Algebraic statistics; complex coding; fractional factorial designs; indicator function; isomorphic fractions; orthogonal arrays; regular fractions.
1
2
FABIO RAPALLO AND MARIA PIERA ROGANTIN
Different methods to check the isomorphism between fractions has been done
in literature. [3], [14], and [13] give necessary and sufficient conditions for isomorphism, based on the Hamming distance between two fractions, while [15] uses the
singular value decomposition of the design matrix for binary designs. [18] and [19]
consider an approach using the centered L2 -discrepancy for geometric isomorphism.
In the latter paper, an algorithm with low complexity is presented. [11] introduce
a different definition of regularity based on the canonical correlation to analyze the
problem. [22] essentially deal with an inverse problem: to find factor level permutations of a regular fraction to reduce contamination of non negligible interactions
on the estimation of linear effects without increasing the run size.
We emphasize that most of the algorithms in literature are specific for binary
factors, and even when defined in general for multilevel factors they are exemplified
in the three levels case. Symmetric qualitative designs with three factor levels can
be partitioned into isomorphic to a regular and not isomorphic to a regular fraction,
while with larger number of levels the situation is more difficult to analyze. In fact,
up to the three levels case, the isomorphism with respect to a regular fraction can
be detected directly using the complex coding of levels and the indicator function,
as argued later in Section 3.
A necessary condition for isomorphism between two fractions is that they have
the same Generalized Word Length Pattern (GWLP) and this condition does not
depend on the level coding, see [23]. [9] derive a formula, computationally easy and
of clear interpretation, for computing the GWLP of a fraction, based on the mean
aberration.
In this paper we use tools from Algebraic statistics, and in particular the polynomial indicator function to specify a fraction and its orthogonality/regularity properties, the polynomial representation of the permutations of the factor levels, and
the representation through Latin squares of orthogonal arrays of strength two. In
this framework, the coding of the s levels of a factor by the s-th roots of the unity
is essential. The methodology introduced here to detect isomorphism between fractions applies to symmetric designs with s prime and the given examples consider
s ≥ 5.
The paper is organized as follows. In Section 2 some relevant results of the
algebraic theory of fractional factorial designs are presented. In Section 3 the polynomial representation of factor level permutations is analyzed. Using polynomial
conditions on the coefficients of such a representation, an algorithm to check if two
fractions are isomorphic is given and some examples, implemented in the symbolic
software CoCoA-5, are shown. In particular fractions of a 53 design are checked to
be isomorphic or not to a regular one. In Section 4 the investigation if a fraction is
isomorphic to a regular one or not is approached using the Latin square representation of orthogonal arrays of strength 2. When the isomorphism exists, the relevant
level permutations and the generating equations of the regular fraction are recovered, by exploiting the properties of the complex coding of the levels. Such a check
is based on the nullity of all the 2 × 2 minors of the multi-layer Latin squares in
the numeric complex field. Several examples are discussed to show how to actually
apply the proposed method. Finally, some further developments are illustrated in
Section 5.
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
3
2. Algebraic characterization of fractional designs
In this section we present some relevant results of the algebraic theory of fractional designs. The interested reader can find further information, including the
proofs of the propositions, in [8], [20], [9].
Let us consider an experiment with m factors observed at s levels each with s a
prime number.
√
Let us code the s levels of the s-th roots of the unity ωk = exp −1 2π
s k ,
k = 0, . . . , s − 1. We denote the level set by Ωs = {ω0 , . . . , ωs−1 }.
As α = β mod s implies ωkα = ωkβ , it is useful to introduce the residue class
ring Zs and the notation [k]s for the residue of k mod s. For integer α, we obtain
(ωk )α = ω[αk]s . We also have ωh ωk = ω[h+k]s . We drop the sub-s notation when
there is no ambiguity.
We denote by D the full factorial design with complex coding: D = Ωm
s ; the
cardinality of the full factorial design is #D = sm .
We denote by L the exponent set of the complex coded design: L = Zm
s . Notice
that L is both the exponent set of the complex coded design and the integer coded
design. The elements of L are denoted in vector notation by α, β, . . .:
L = {α = (α1 , . . . , αm ) : αj = 0, . . . , s − 1, j = 1, . . . , m} ;
[α − β] is the m-tuple ([α1 − β1 ] , . . . , [αm − βm ]).
In order to use polynomials to represent all the functions defined over D, we
define
- Xj , the j-th component function, which maps a point of D to its j-th
component, Xj : D ∋ (ζ1 , . . . , ζm ) 7−→ ζj . The function Xj is a simple
term or, by abuse of terminology, a factor.
α
αm
- X α = X1α1 · . . . · Xm
, α ∈ L = Zm
s i.e., the monomial function X : D ∋
αm
(ζ1 , . . . , ζm ) 7→ ζ1α1 · . . . · ζm
. The function X α is an interaction term. As
s is a prime number, the interaction X α takes values in Ωs .
The set of monomials {X α : α ∈ L} is an orthonormal basis of all the complex
functions defined over D.
Since we will make use occasionally factors with a non-prime number of levels,
the behavior of the factors and of the interactions is summarized below:
- Let Xi be a simple term with level set Ωs . Let us define h = s/gcd(r, s)
and let Ωh be the set of the h-th roots of the unity. The term Xir takes all
the values of Ωh equally often over D.
αj
αj
α
- Let X α = Xj1j1 · · · Xjk k be an interaction term of order k where Xji i takes
values in Ωhji . Let us define h = lcm{hj1 , . . . , hjk }. The interaction X α
takes values in Ωh equally often over D.
Let F be a subset of the full factorial design D. We consider here only fractions
without replications.
Definition 2.1. The indicator function F of a fraction F is a complex polynomial
defined over D such that for each ζ ∈ D, F (ζ) is equal to 1 if ζ ∈ F and equal to
0 if ζ ∈ D \ F . We denote by bα the coefficients of the representation of F on D
using the monomial basis {X α , α ∈ L}:
X
(1)
F (ζ) =
bα X α (ζ), ζ ∈ D, bα ∈ C .
α∈L
4
FABIO RAPALLO AND MARIA PIERA ROGANTIN
Proposition 2.2. Let F be a fraction with indicator function F .
(1) The coefficients bα of F are given by:
(2)
bα =
s−1
1 X
1 X [−α]
X
(ζ) =
nα,s−h ωh
#D
#D
ζ∈F
(2)
(3)
(4)
(5)
h=0
where nα,h is the number of the occurrences of ωh in {X α (ζ) : ζ ∈ F }.
In particular b0 = #F /#D.
The term X α is centered on F if, and only if, bα = b[−α] = 0.
The terms X α and X β are orthogonal on F if, and only if, b[α−β] = 0;
If X α is centered then, for each β and γ such that α = [β −γ] or α = [γ −β],
X β is orthogonal to X γ .
Let s be prime. Then, the term X α is centered on F if, and only if, its s
levels appear equally often: nα,0 = · · · = nα,s−1 .
As stated in the proposition above, the coefficients bα encode interesting properties of the fraction such as orthogonality among factors and interactions, and get
together the combinatorial and vectorial orthogonality.
A common choice to select an experiment is to use orthogonal array. The assumption that interactions above a specified order are not present is translated into
a combinatorial property on the frequency of the levels in the fraction. Let us denote with OA(n, sm , t) a symmetric orthogonal array with n rows and m columns,
in which each column has s symbols, and with strength t, as defined e.g. in [23].
Strength t means that for every selection of t columns of the matrix, all the elements
of Ωts appear equally often in the t columns.
Definition 2.3. Let I be a non-empty subset of {1, . . . , m}. A fraction F factorially
projects on the I-factors if the projection of F on the I-factors is a full factorial
design where each point appears k times. A fraction F is an orthogonal array of
strength t if it factorially projects on any I-factors with #I = t.
The proposition below shows a connection between the combinatorial definition
of orthogonal array introduced above and the coefficients of the indicator function
in Eq. (1).
Proposition 2.4. A fraction is an orthogonal array of strength t if, and only if,
all the coefficients of the indicator function up to the order t are zero:
bα = 0
∀ α of order up to t, α 6= (0, . . . , 0) .
Definition 2.5. Given an interaction X α defined on a fraction F of the full factorial
design D, its aberration, or degree of aliasing, aα is given by the real number
aα =
kbα k22
b20
where kxk22 is the square of the Euclidean norm of the complex number x.
The GWLP A(F ) = (A1 (F ), . . . , Am (F )) of a fraction F is defined as
X
aα j = 1, . . . , m ,
Aj (F ) =
kαk0 =j
where kαk0 is the number of non-null elements of α, i.e., the order of interaction of
X α.
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
5
The following proposition allows us to compute all the aberrations aα without
using complex numbers.
Proposition 2.6. Let X α be a simple or interaction term with values in Ωt . Its
aberration aα is
!
X
t−1
t−1
1 X
2π
aα = 2
nα,i nα,[i−k] .
k
cos
n
t
i=0
k=0
Regular fractions are a subset of the orthogonal arrays. In a regular fraction any
two simple or interaction terms are either orthogonal or totally confounded.
Definition 2.7. A fraction F is regular if there exist a sup-group L of L, a group
homomorphism e from L to Ωs such that the set of equations
(3)
{X α = ωe(α) : α ∈ L}
defines the fraction F . If H is a minimal generator of the group L, the set of
equations {X α = ωe(α) : α ∈ H} is called the set of defining equations of F .
Proposition 2.8. Let F be a fraction. The following statements are equivalent:
(1) The fraction F is regular according to Definition 2.7.
(2) The indicator function of the fraction has the form
1 X
ωe(α) X α (ζ)
ζ∈D .
F (ζ) =
#L
α∈L
where L is a given subset of L and e : L → Ωs is a given mapping.
(3) For each α, β ∈ L the interactions represented on F by the terms X α and
X β are either orthogonal or totally aliased.
Finally, we recall two basic definitions of isomorphic fractions. For details see
e.g. [6].
Definition 2.9. Two fractions are
- combinatorially isomorphic if one can be obtained from the other by reordering the runs, relabeling the factors and/or switching the levels of one
or more factors.
- geometrically isomorphic if one design can be obtained from the other by
reordering the runs, relabeling the factors and/or r eversing the level order
of one or more factors
The combinatorial isomorphism pertains to qualitative factors, while geometric
isomorphism pertains to quantitative factors. In this paper we focus on qualitative
factors mainly.
From the definition of the indicator function, it follows immediately that a reordering of the runs does not affect the polynomial representation of the indicator
function. Moreover, a relabeling of the factors simply permutes the subscripts.
From the expression of the indicator function, it is relatively simple to find the
relevant relabeling. Therefore, the most interesting task in analyzing the isomorphism of fractions is to study the behavior of the fractions under permutations of
the factor levels.
6
FABIO RAPALLO AND MARIA PIERA ROGANTIN
3. Polynomial representation of the factor level permutations
In this section first we give an account of the polynomial representation of the
level permutations for a single factor, then we extend this representation to several factors, and finally we use such characterization to check the combinatorial
isomorphism of two fractions on some examples.
A level permutation is a function from Ωs to Ωs and it always admits a polynomial representation. In special cases, such polynomial reduces to a monomial, and
therefore a permutation of the levels does not affect the regularity of a fraction.
Proposition 3.1. A regular fraction is transformed into a regular fraction by the
group of transformations generated by the following level permutations on the factor
Xj :
(1) Cyclical permutations Xj → ωk Xj with k = 0, . . . , s − 1.
(2) Power permutations, for s prime, Xj → ωk Xjh with h = 1, . . . , s − 1.
Permutations of type 1. and 2., produce s(s − 1) permutations on the factor Xj ,
and produce the following transformed monomial on the term X α :
m
Y
[α h ]
ω[αj kj ] Xj j j .
(4)
j=1
The proof of Prop. 3.1 is in [20]. We observe only that under monomial permutations the absolute values of the indicator function coefficients do not change, so
that a regular fraction is transformed into a regular fraction.
We highlight again that, for factors with two or three levels, all the level permutations have a monomial representation.
Remark 1. From Prop. 3.1 the monomial representation of a geometric isomorphism
follows. In fact, for quantitative factors, the two admissible level permutations are
both in monomial form: Y = X (the identity) and Y = ωs−1 X s−1 (the reversing
of the factor levels).
In the reminder of this section we characterize the polynomial representation of
the permutations for general multilevel factors. Such characterization leads to a
criterion to actually check if a given fraction may be or may be not transformed
into a regular fraction after permutation of the levels of one or more factors.
Let X be a factor with s levels and let π be a permutation of the level set Ωs .
We denote by Y the transformed factor, Y = π(X). The polynomial representation
of Y is
s−1
X
(5)
Y =
uh X h
uh ∈ C
h=0
with π(ωi ) =
system
(6)
Ps−1
h
h=0 uh ωi .
The uh ’s coefficients are the solutions of the linear
ω00
ω10
..
.
ω10
ω11
..
.
0
. . . ωs−1
s−1
. . . ω1
..
.
0
ωs−1
1
ωs−1
s−1
. . . ωs−1
u0
u1
..
.
us−1
=
π(ω0 )
π(ω1 )
..
.
π(ωs−1 )
.
The matrix appearing in Eq. 6 is a Vandermonde matrix V . If we denote by
vh+1,k+1 , with h, k = 0, . . . , s − 1, the generic element of V , it is known that the
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
7
−1
inverse V −1 of V has generic element vh+1,k+1
/s. In our case, from the results in
Section 2:
1
−1
vh+1,k+1 = ωhk = ω[hk] and vh+1,k+1
= ω[s−hk]
s
and the resulting system is
u0
ω0
ω0
. . . ω0
π(ω0 )
u1 1 ω0 ωs−1 . . . ω1 π(ω1 )
(7)
.. = ..
.
..
..
..
. s .
.
.
.
us−1
π(ωs−1 )
ω0
ω1
. . . ωs−1
Full details on the Vandermonde matrices for the roots of the unity, their properties and applications to complex interpolations can be found in, e.g., [4].
Combining the expression of V −1 with the fact that π is a permutation,we obtain
constraints on the coefficients u0 , . . . , us−1 as in the following proposition.
Proposition 3.2. The coefficients ui ’s must satisfy the following equations:
(i) u0 = 0;
(ii) for all q = 2, . . . , s − 1,
s−1
X
(8)
uh1 . . . uhq−1 u[−h1 ...−hq−1 ] = 0 ;
h1 ,...,hs−1 =0
(iii) given a permutation π, we have
Ps−1
(iv) ( h=1 uh )s − 1 = 0 .
Ps−1
h=1
uh = π(ω0 ), and therefore
Proof. Let π be a permutation of Ωs . For item (i), observe that the first row in the
system (7) leads to
s−1
s−1
1X
1X
π(ωi ) =
ωi = 0 .
u0 =
p i=0
p i=0
The constraints in item (ii) are derived in the same way, but using the powers
Ps−1
Y 2 , . . . , Y s−1 . For Y 2 , the term of degree zero is h=0 ui us−i which is the left-hand
side in Eq. (8) for q = 2 and in the same way one writes the corresponding degree
zero terms for q = 3, . . . , s − 1. Now, if s is prime, all the powers Y 2 , . . . , Y s−1
are permutations of Ωs and the result follows from item (i). For general s (not
prime necessarily), define m = gcd(q, s) and note that Y q contains m times all the
elements of the set Ω[s/m] , whose sum is again zero. For item (iii) it is enough to
write
s−1
s−1
s−1
s−1
s−1
X
X
X
1 X −h
1X
uh =
ωk π(ωk ) =
π(ωk )
ω[−kh] =
s
s
h=1
h=0
k=0
k=0
h=0
where the inner sum is always equal to zero except for k = 0, and thus
1
(sπ(ω0 )) = π(ω0 ) .
s
Finally, item (iv) follows from (iii) by noting that the value of π(ω0 ) may take any
values in Ωs .
=
It is interesting to write explicitly the equations in items (i), (ii), and (iv) for
the first values of s.
8
FABIO RAPALLO AND MARIA PIERA ROGANTIN
For s = 2, we have u0 = 0 and u21 − 1 = 0, and such two equations characterize
the only two possible permutations of Ω2 . The same holds for s = 3, where we
obtain
u0 = 0 ;
u1 u2 = 0 ;
(u1 + u2 )3 − 1 = 0 .
From the second equation, we conclude that one among u1 and u2 is zero, providing
an alternative proof to the fact that all the level permutations have a monomial
representation for factors with two or three levels.
We illustrate now an example with s = 4, i.e., a non-prime s. The conditions
are:
u0 = 0 ;
u22 + 2u1 u3 = 0 ;
u2 (u21 + u23 ) = 0 ;
(u1 + u2 + u3 )4 − 1 = 0 .
When s = 4 not all the monomial maps of the form Y = ωh X k , h = 0, . . . , 3,
k = 1, . . . , 3 are the polynomial representation of a permutation. Take for example
the monomial map Y = X 2 . This correspond to the transformation with coefficients
u0 = u1 = u3 = 0 and u2 = 1. This is not a solution of the above equations, since
the second equation is not satisfied.
When s increases the situation becomes computationally less simple, since from
Prop. 3.2 there are s polynomial equations with degrees 1, . . . , s. Therefore, the degree of the polynomial system is s!, which is exactly the number of the permutations
of Ωs .
To check if the system has a finite number of solutions one can apply the finiteness
theorem (see Appendix A). It is enough to compute a Gröbner basis of the ideal
cs−1
generated by the s equations in Prop. 3.2 and check if all the terms uc00 , . . . , us−1
are
all leading terms of polynomials in the Gröbner basis for some exponents c0 , . . . , cs .
Example 3.3. For s = 5 the system in Eq. (7) yields 5 equations, and the Gröbner
basis of the corresponding polynomial ideal is formed by 28 polynomials. Among
them, the five polynomials displayed below have leading term of the form u[i]^c[i]
for appropriate exponents c[i] for all i, and therefore the finiteness theorem in
Appendix A applies. Then the polynomial system has a finite number of solutions.
u[0],
u[1]^5 +u[2]^5 +u[3]^5 +(-20)*u[1]^3*u[3]*u[4] +
+(-20)*u[1]*u[2]*u[4]^3 +u[4]^5 +(-1),
u[2]^6 +(23)*u[1]^4*u[4]^2 +(16)*u[1]^2*u[3]*u[4]^3 +
+(19)*u[3]^2*u[4]^4 +(20)*u[2]*u[4]^5 +(-1)*u[2],
u[3]^6 +(48)*u[1]*u[3]^3*u[4]^2 +(26)*u[2]^3*u[4]^3 +
+(-81)*u[1]^2*u[4]^4 +(27)*u[3]*u[4]^5 +(-1)*u[3],
u[4]^11 +(-12628/625)*u[1]^3*u[3]*u[4]^2 +(-77/625)*u[1]*u[3]^2*u[4]^3 +
+(-12639/625)*u[1]*u[2]*u[4]^4 +(-121/625)*u[4]^6 +(-504/625)*u[4]
Finally, an interesting property of the coefficients ui ’s concern their expression
in terms of the roots of the unity.
Proposition 3.4. Up to the constant 1/s, the coefficients ui are integer nonnegative combinations of the s-th roots of the unity:
uh =
s−1
X
r=0
vr ωr ,
vr ∈ N
When the number of levels s is prime, for all permutations π, such representation
of the coefficients is unique up to an additive integer constant.
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
9
Proof. The first part follows directly from the expression of the inverse of the Vandermonde matrix in Eq. (7). For the uniqueness, see [20].
Now we show how to use the equations above in order to study the isomorphism
between two fractions, by merging polynomial constraints on the support of a full
factorial design and the polynomial constraints in Prop. 3.2. The computations
below are carried out with the free software CoCoA, see [1]. Also in these examples
we will make use of basic tools from Computational Commutative Algebra, such as
polynomial ideal, Gröbner basis, Normal Form. The basic definitions and results
needed here are collected in Appendix A. Useful techniques to handle polynomials
can be found in [2].
We write the polynomial indicator function of the two fractions under investigation, and we do some algebraic manipulations in order to obtain the coefficients
of the (possible) permutation needed to transform the first fraction into the second
one. In particular, we check if a fraction is isomorphic to a regular one. Our examples are given in the 53 case, where there is only one regular orthogonal array with
strength 2 up to monomial transformations. There are several online databases of
orthogonal arrays. The examples analyzed here are taken from [7], generated with
the algorithm introduced in [21].
Let F0 and F1 be the two fractions to compare, with polynomial indicator functions F0 and F1 respectively. Consider a generic transformation π = (π1 , π2 , π3 )
where πj acts on the factor Xj :
πj : Xj −→
4
X
ukj Xjk .
k=0
In particular, if we want to check the regularity of the fraction F1 , then F0 is the
indicator function of the regular fraction with defining equation X1 X2 X3 = ω0 ,
P
(r)
namely F0 = 14 4k=0 (X1 X2 X3 )k
(1) Consider the ring of the indeterminates
- x[1],x[2],x[3], the factors;
- u[0..4, 1..3], the 5 × 3 transformation coefficients;
- w, the 5-th primitive root of the unity, satisfying the equation
1+w+w^2+w^3+w^4=0. The indeterminate w is considered here as a
parameter.
(2) Input F0 and F1, the indicator functions of the two fractions to be compared
(minus one).
(3) Consider I, the ideal generated by the polynomials defining the full factorial
design (x5i − 1, for i = 1, 2, 3) and the 3 × 5 polynomials with the conditions
for the transformation coefficients. The CoCoA code for such polynomials
is:
L:=NewList(5);
L[1] := [u[0,j] |j in 1..3];
L[2] := [(Sum([u[i,j] | i In 1..4]))^5 -1 |j in 1..3];
L[3] := [Sum([u[i,j]*u[Mod(-i,5),j]| i In 1..4]) |j in 1..3];
L[4] := [Sum(Sum([[u[i,j]*u[h,j]*u[Mod(-i-h,5),j]| i In 1..4]
| h In 1..4])) |j in 1..3];
L[5] := [Sum(Sum(Sum([[[u[i,j]*u[h,j]*u[k,j]*u[Mod(-i-h-k,5),j]
|i In 1..4]| h In 1..4]| k In 1..4]))) |j in 1..3];
10
FABIO RAPALLO AND MARIA PIERA ROGANTIN
(4) Compute NF_P_F0, the normal form of the transformation of F0 by π in the
quotient space K/I and compute Coe_F0, the list of the coefficients of the
terms in x[1],x[2],x[3] appearing in NF_P_F0.
(5) Compute Coe_F1, the list of the coefficients of the terms in x[1], x[2],
x[3] appearing in F1.
(6) Compute Coe_Diff, the difference between the coefficients Coe_F1 and the
coefficients Coe_F0, for all the terms in Coe_F0.
(7) Compute the Gröbner Basis of the ideal generated by the polynomials in
Coe_Diff and the polynomials in L[0..4] with the conditions for the transformation coefficients.
If the Gröbner basis is empty, then the two fractions are not isomorphic, otherwise
the Gröbner Basis contains equations on the transformation coefficients that allow
us to find the permutations.
(r)
Notice that, if F0 is the indicator function with generating equation X1 X2 X3 =
1, then the number of terms of NF_P_F0 is 6401, while, obviously, the length of
Coe_F0 is 65, the number of the interactions of order 3 plus the constant.
Example 3.5. Let FA , FB and FC be three fractions of a 53 factorial design,
whose indicator functions are shown in Appendix B. The first two fractions are
listed in [7]. The fraction FB is a regular fraction evidently.
Using the previous algorithm we checked if they are isomorphic to the regular
fraction F (r) above. In the firs case, the Gröbner Basis of the step 7 has only the
element 1; then FA is not isomorphic to any regular fraction of a 53 factorial design.
In the last case the Gröbner Basis has been computed in 29 secs. of CPU time and
contains 91 elements. A solution is:
u.,1 = (0, 1, 0, 0, 0)
u.,3 = (0, 0, 0, 0, 1)
1
u.,2 = (0, 2 − ω2 − ω3 , 2ω1 + ω2 + 2ω3 , ω1 + ω2 − ω3 − 1, 2ω1 − ω2 − 1)
5
that corresponds to no permutation on the first factor, a power permutation on the
third factor and the switch between the levels 0 and 1 on the second one.
4. Regularity check of multi-level orthogonal arrays
In this section we approach the problem of checking the regularity of a multilevel orthogonal array using the complex coding of the factor levels. In particular
we provide a result which leads us to check if a given orthogonal array may be
regarded as a regular fraction under suitable permutations of the factor levels. This
technique exploits the connections between orthogonal arrays and Latin squares,
see for instance [16] and [12], and it is based on the generating equations of the
regular fraction rather than on the whole indicator function.
We focus on orthogonal arrays with strength 2. Let F be an OA(n, sm , 2) with
s prime. A regular fraction with strength 2 has at least one generating equation
involving only 3 factors. Therefore we first look at generating equations involving
three factors. Without loss of generality, let us consider the factors X1 , X2 , X3 . Let
(9)
X1α1 X2α2 X3α3 = ωk
be a generating equation of F , with α1 , α2 , α3 ∈ {1, . . . , s − 1} and ωk ∈ Ωs . For
brevity, we say that X1 , X2 , X3 form a generating equation of the orthogonal array
F.
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
11
If X3 is a function of X1 and X2 , we can consider the s× s table C = X3 (X1 , X2 )
containing the values of X3 as a function of X1 and X2 , i.e., Cj1 ,j2 = x3 given
x1 = ωj1 and x2 = ωj2 . Since the strength of the orthogonal array is 2, the table
C may be regarded as a s × s Latin square with values in Ωs . The main result can
be stated as follows.
Theorem 4.1. Let X1 , X2 , X3 be three factors of an OA(n, sm , 2), s prime. If X3
is a function of X1 and X2 , let X3 (X1 , X2 ) be the corresponding Latin square.
(a) If X1 , X2 , X3 form a generating equation, then X3 (X1 , X2 ) has rank 1 in
C, i.e., all 2 × 2 minors of X3 (X1 , X2 ) vanish in C;
(b) If there is a permutation π3 of Ωs such that (π3 (X3 ))(X1 , X2 ) is a Latin
square with rank 1 in C, then there exist permutations π1 and π2 such that
π1 (X1 ), π2 (X2 ), π3 (X3 ) form a generating equation.
Proof. (a) By hypothesis there exist α1 , α2 , α3 ∈ {1, . . . , s − 1} and ωk ∈ Ωs such
that X1α1 X2α2 X3α3 = ωk . Since s is prime, we can assume α3 = 1. In fact, given a
generating equation X1α1 X2α2 X3α3 = ωk , there exists r such that [rα3 ] = 1 and the
equation
[rα1 ]
X1
[rα2 ]
X2
X3 = ω[rk]
is also a generating equation of the fraction. Consider the Latin square C =
[−rα ] [−rα ]
X3 (X1 , X2 ). The entry Cj1 ,j2 of C is Cj1 ,j2 = ω[rk] ωj1 1 ωj2 2 for j1 , j2 =
0, . . . , s − 1 and the generic 2 × 2 minor of C is
ω[2rk] ω[j11 −rα1 ] ω[j21 −rα2 ] ω[j12 −rα1 ] ω[j22 −rα2 ] −
ω[j11 −rα1 ] ω[j22 −rα2 ] ω[j12 −rα1 ] ω[j21 −rα2 ]
that equals 0 for all pairs of distinct row indices j11 , j12 ∈ {0, . . . , s − 1} and for all
pairs of distinct column indices j21 , j22 ∈ {0, . . . , s − 1}.
(b) Suppose that there is a permutation π3 of X3 such that the table C =
π3 (X3 )(X1 , X2 ) is a Latin square with all 2 × 2 minors equal to 0. Then apply
suitable permutations π1 and π2 to X1 and X2 , respectively, in order to obtain a
Latin square in reduced form, i.e., with the first row and column lexicographically
ordered. Now it is immediate to check that π3 (X3 ) = π1 (X1 )π2 (X2 ) and therefore a
defining equation after the level permutations is π1 (X1 )π2 (X2 )π3 (X3 )[−1] = ω0 .
Some remarks on part (b) of the theorem above are now in order. We can exploit
the monomial representation of the permutations in Prop. 3.1 to reduce considerably the computational cost. First, observe that the permutations to be checked
on X3 are at most (s − 2)!. In fact, we can exclude the powers (to each defining
equation correspond other (s − 2) equivalent ones) and the s cyclic permutations
(they only affect the constant term). For instance, if s = 5, there are 120 level permutations but only six are to be checked. Secondly, the relevant permutations of
X1 and X2 are the permutations needed to put the Latin square in reduced form.
Additionally, if the permutations π1 and π2 can be expressed in monomial form
(powers and/or cyclic permutations), then the defining equation can be written
without actual permutations on X1 and X2 .
Finally, note that the permutations of the factors are not uniquely defined: for
instance, a shift of the form ωh X can be applied to whatever factor and it generates
a unique transformation in the generating equation.
12
FABIO RAPALLO AND MARIA PIERA ROGANTIN
0
1
2
3
4
0
0
1
2
3
4
1 2
1 2
2 3
3 4
4 0
0 1
(a)
3
3
4
0
1
2
4
4
0
1
2
3
0
1
2
3
4
0
2
0
3
4
1
1 2
4 0
2 1
1 4
3 2
0 3
(b)
3
1
3
2
0
4
4
3
4
0
1
2
0
1
2
3
4
0
0
1
2
3
4
1 2
1 2
0 3
3 4
4 1
2 0
(c)
3
3
4
0
2
1
4
4
2
1
0
3
Figure 1. The three orthogonal arrays of Example 4.2.
0
1
2
3
4
0
0
4
2
1
3
1
1
0
3
2
4
2
4
3
1
0
2
3
3
2
0
4
1
4
2
1
4
3
0
Figure 2. The Latin square of the orthogonal array (b) of Example 4.2 after the permutation (ω4 , ω3 , ω0 , ω2 , ω1 ) on X3 .
Before analyzing the general case of symmetric multilevel designs, we present
some applications of the theorem above in the simple case of orthogonal arrays
with 3 factors and strength 2, so that there is only one defining equation.
Example 4.2. In the framework of the 53 full factorial design, consider the three
orthogonal arrays with strength 2 identified by the three Latin squares in Figure 1.
To ease the readability of the tables, we write k in place of ωk . The three designs
are built from the 2 non-isomorphic orthogonal arrays with strength 2 listed in [7].
The array in (a) is the second fraction in [7], the array in (b) is obtained by the
previous one, with some permutations on the factor levels, and the array in (c)
is the first fraction in [7]. The three indicator functions of these fractions are in
Appendix B. In this case of fractions of the 53 design with 25 runs, it is known that
there are no other non-isomorphic orthogonal arrays.
(a) All the 2 × 2 minor vanish in C. The defining equation of this fraction,
without any permutations, is X1 X2 = X3 or equivalently X1 X2 X34 = ω0 .
(b) This fraction has been built from the previous one by applying the permutations (ω3 , ω2 , ω4 , ω1 , ω0 ) to X1 , (ω2 , ω0 , ω1 , ω4 , ω3 ) to X2 and (ω2 , ω4 , ω0 , ω3 , ω1 )
to X3 . Among the five possible (other than the identity) on X3 , we observe that the permutation (ω4 , ω3 , ω0 , ω2 , ω1 ) produces the Latin square in
Figure 2, where all the 2 × 2 minor vanish in C.
Looking at the rows and columns beginning with ω0 , we can read easily
the permutations of the levels of X1 and X2 needed to obtain a Latin square
in reduced form. From the first column we read that the permutation on X1
is (ω0 , ω3 , ω2 , ω4 , ω1 ), while in the first row we read that the permutation
on X2 is (ω0 , ω1 , ω4 , ω3 , ω2 ). Notice that such permutations are those used
in the construction of the orthogonal array up to a shift of four levels for X1
and one level for X2 . In this example, the relevant permutations can not
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
0
1
2
3
4
0
0
1
3
2
4
1
1
0
2
4
3
2
3
2
4
1
0
3
2
4
0
3
1
0
1
2
3
4
0
0
1
4
2
3
4
4
3
1
0
2
0
1
2
3
4
1
1
0
2
3
4
2
4
2
3
1
0
0
0
1
3
4
2
3
2
3
0
4
1
1
1
0
4
2
3
4
3
4
1
0
2
2
3
4
2
1
0
3
4
2
0
3
1
4
2
3
1
0
4
0
0
1
4
3
2
0
1
2
3
4
0
0
1
2
4
3
0
1
2
3
4
1
1
0
3
2
4
2
4
3
2
1
0
3
3
2
0
4
1
1
1
0
4
3
2
2
2
4
3
1
0
3
4
3
0
2
1
13
4
3
2
1
0
4
4
2
4
1
0
3
Figure 3. The five Latin squares obtained from the orthogonal
array (c) of Example 4.2 after the five non-identical relevant permutations.
0
1
2
3
4
5
6
0
0
1
4
6
5
3
2
1
1
0
2
3
4
5
6
2
2
4
3
5
1
6
0
3
3
5
6
0
2
1
4
4
4
2
5
1
6
0
3
5
5
6
0
4
3
2
1
6
6
3
1
2
0
4
5
Figure 4. A Latin square representing a non regular orthogonal
array of strength 2 of the 73 design.
be expressed in monomial form, and thus the permutations of the factor
levels are unavoidable.
(c) The Latin squares after the 5 relevant permutations of X3 are in Figure 3.
The first minor in all tables is ω02 − ω12 = 1 − ω2 6= 0 and this is enough to
conclude that this fraction can not be transformed into a regular one.
Remark 2. In the part (c) of the previous example, note that the 2 × 2 upper-left
matrix has non-zero determinant for all permutations of the levels of X3 . In fact,
if the permutation π3 maps ω0 into ωj0 and ω1 into ωj1 , with j0 6= j1 , one obtains
the minor ωj21 − ωj20 = ω[2j1 −2j0 ] 6= 0. This remark can also be used to build nonregular orthogonal arrays also in case of a large number of levels, as illustrated in
the example below.
Example 4.3. The Latin square in Figure 4 represents an orthogonal array of
strength 2 of the 73 design. It has been defined starting from the upper-left 2 × 2
sub-matrix, and then completed in the remaining entries. By construction, it is a
non-regular design even under permutations of the factor levels.
Theorem 4.1 can be used for constructing an algorithm to check the regularity of
orthogonal arrays with strength 2 and with an arbitrary number of factors, under
permutations of the factor levels.
14
FABIO RAPALLO AND MARIA PIERA ROGANTIN
First, consider an orthogonal array with one defining equation involving m > 3
factors. Theorem 4.1 can be applied recursively layer by layer. For instance fix
m = 4. The factors X1 , X2 , X3 , X4 form a generating equation if and only if
X1α1 X2α2 X3α3 X4 = ωk
for some α1 , α2 , α3 ∈ {1, . . . , s − 1} and ωk ∈ Ωs . Thus, for each x4 = ωj4 ,
j4 = 0, . . . , s − 1, the equation
X1α1 X2α2 X3α3 = ω[k−j4 ]
is satisfied. Conversely, if Theorem 4.1 applies to each layer x4 = ωj4 , j4 = 0, . . . , s−
1, with the same permutation, and there exist permutations π1 , π2 , π3 on X1 , X2 , X3
respectively such that
π1 (X1 )α1 π2 (X2 )α2 π3 (X3 )α3 = ωk(j4 )
with different ωk(j4 ) , for j4 = 0, . . . , s − 1, then k(j4 ) defines a permutation π4 for
X4 and
π1 (X1 )α1 π2 (X2 )α2 π3 (X3 )α3 π4 (X4 )[s−1] = ω0
is a defining equation for the orthogonal array.
For orthogonal arrays with strength 2 and m factors, we check the possible
defining equations in the following order.
(1) First, check all the 3-tuples.
(2) Then, if the defining equations with 3 factors are not sufficient to define
the orthogonal array, check the defining equations with 4 or more factors.
Notice that the number of (independent) defining equations is #D/#F .
The regularity check is based on the following property of regular fractions,
that can be easily proved within the framework of the complex coding of the level
factors. The proof of this result is based on the properties of the elimination ideals,
see Appendix A for some basic definitions.
Proposition 4.4. Let F be a regular fraction of a sm design, and let I ⊂ {1, . . . , m}.
Denote with FI the projection of F onto the I-factors. Apart from the multiplicity,
FI is either a full factorial design or a regular fraction.
Proof. Let I = {1, . . . , m} \ I. Let us define the ideal I(F ) as the ideal in
C[x1 , . . . , xm ] generated by the binomials xsj − 1 = 0, j = 1, . . . , m and by the
generating equations of F . The ideal I(F ) is a binomial ideal. In fact, two factors
or interactions are either orthogonal or totally confused and this yields only binomial equations. The ideal I(FI ) is the elimination ideal of I(F ) with respect to the
variables xj , j ∈ I. From the results in Chapter 3 of [5], I(FI ) is also a binomial
ideal and two cases may arise: (a) the binomials xsj − 1 = 0, j ∈ I generate I(FI ),
and this means that FI is a full factorial design on the I factors; (b) there are other
generators. From the definition of ideal, such polynomials define FI as a regular
fraction.
Remark 3. To ease computations, remember that a defining equation with a given
number of factors cannot include simultaneously all the factors of a defining equation with a lower number of factors.
Example 4.5. Consider the 55−2 regular fraction defined by
(10)
X12 X2 X3 = ω1
X1 X2 X4 X5 = ω 1
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
0
1
2
3
4
0
1
4
3
0
2
1
0
3
2
4
1
2
4
2
1
3
0
3
3
1
0
2
4
15
4
2
0
4
1
3
Figure 5. The Latin square X3 (X1 , X2 ) for Example 4.5.
0
1
2
3
4
0
1
3
0
2
4
1
4
1
3
0
2
2 3
2 0
4 2
1 4
3 1
0 3
X5 = ω 0
0
1
2
3
4
0
0
2
4
1
3
4
3
0
2
4
1
0
1
2
3
4
1
3
0
2
4
1
0
3
0
2
4
1
2 3
1 4
3 1
0 3
2 0
4 2
X5 = ω 3
1
1
3
0
2
4
2 3
4 2
1 4
3 1
0 3
2 0
X5 = ω 1
4
2
4
1
3
0
0
1
2
3
4
4
0
2
4
1
3
0
4
1
3
0
2
1
2
4
1
3
0
0
1
2
3
4
2 3
0 3
2 0
4 2
1 4
3 1
X5 = ω 4
0
2
4
1
3
0
1
0
2
4
1
3
2 3
3 1
0 3
2 0
4 2
1 4
X5 = ω 2
4
4
1
3
0
2
4
1
3
0
2
4
Figure 6. The layers X4 (X2 , X3 ) given X5 for Example 4.5.
and a fractions obtained by permuting the levels of two factors. The indicator
function of this fraction is written in Appendix B and has 289 nonzero monomials.
We apply our technique to the new fraction and we show how to recover the defining
equations (10) and the correct permutations starting from the fraction points.
First we check the defining equations with 3 factors. We obtain a valid Latin
square only with X1 , X2 , X3 as in Figure 5.
Here the minors are all zero and therefore no permutation on X3 is needed. Now,
from the column beginning with ω0 we read the permutation of X1 . It is easy to
see that it is X13 with the switch permutation (ω2 , ω4 ). From the row beginning
with ω0 we obtain the permutation X24 . Thus we have X3 = ω1 π1 (X1 )3 X24 , or
equivalently π1 (X1 )2 X2 X3 = ω1 . Notice that this constant term ω1 can be easily
recovered from the Latin square above, since it is the symbol in the upper-left
position, where π1 (X1 ) = X2 = ω0 and therefore X3 is equal to the constant term
on the right side hand of the defining equation.
As no other defining equations with 3 factors can be obtained, we move to the
interactions of order 4. There are few checks to do at this stage, because there are
only 5 subsets with 4 factors and two of them are impossible: {X1 , X2 , X3 , X4 }
and {X1 , X2 , X3 , X4 } can be excluded as they contain {X1 , X2 , X3 }, see Remark
3. Consider the 4-tuple {X2 , X3 , X4 , X5 }. We look at the layers defined by X5 and
we obtain the five Latin squares in Figure 6.
16
FABIO RAPALLO AND MARIA PIERA ROGANTIN
In all these Latin squares the 2 × 2 minors are all zero, and thus no permutation
is needed on X4 . Looking at the first Latin square, one reads the permutation
(ω2 , ω0 , ω3 , ω1 , ω4 ) for X2 , i.e., ω1 X23 , and the permutation (ω1 , ω3 , ω0 , ω2 , ω4 ) for
X3 , i.e., ω2 X32 . Then the equation is π5 (X5 )X4 = ω3 X22 X33 , and this equation
holds in all the five Latin squares in Figure 6. We only need to find the possible
permutation π5 on X5 . This is accomplished by looking at the constant terms in
the upper-left cell. From the five Latin squares we read
π5 (ω0 )ω1 = ω3
π5 (ω1 )ω3 = ω3
π5 (ω2 )ω2 = ω3
π5 (ω3 )ω0 = ω3
π5 (ω4 )ω4 = ω3
and these equations are satisfied when π5 = (ω2 , ω0 , ω1 , ω3 , ω4 ). Therefore, applying
π5 to X5 we have the defining equation X23 X32 X4 π5 (X5 ) = ω3 . Finally, we check
that this defining equation corresponds to the second equation used in Eq. (10) to
define the array. Indeed, from π1 (X1 )2 X2 X3 = ω1 we have X3 = ω1 π1 (X1 )3 X2 , and
replacing this expression of X3 into the previous equation one obtains immediately
the defining equation π1 (X1 )X2 X4 π5 (X5 ) = ω1 .
5. Conclusions
In this paper we addressed the problem of level permutations for qualitative
factors. In particular we presented two tools to check if a fraction of a sm factorial
design is isomorphic or not to a regular fraction by permutations of factor levels.
Such a problem is very important in the applications because of the special property
of not partial confounding that have the regular fractions. In this framework, the
coding of levels by the s-th roots of the unity and some tools of algebraic statistics
are essential.
Future works will concern the case of designs with non prime number of levels
and mixed designs, where several properties of the roots of the unity do not hold,
and therefore for this class of designs a different approach must be implemented.
Moreover we want to deepen and better define the concept of mean aberration,
already introduced in [9]. In particular, we want to limit the mean only to permutations compatible with the design matrix of the fraction under investigation. In fact,
the aberrations are calculated through the level counts, and they are connected to
each other by a convolution formula presented in the aforementioned article. This
new definition could allow us to clarify which aberrations are compatible with those
of permuted regular fractions.
Finally, in order to generalize the algorithms presented in this paper for specific
examples, we want to provide efficient packages, both in symbolic and statistical
software (i.e., CoCoA and R respectively), to make actual computations regarding
regularity and isomorphism checks under level permutations in a general setting.
Appendix A. Basic facts in Computational Commutative Algebra
In this appendix we collect some basic notions of Computational Commutative
Algebra used in the paper. The structure of this appendix is taken from [10] (Section
1.7 with Roberto Notari). Introductory expositions of the subject can be found in,
e.g, [5] and [17].
Let K be a numeric field. In our applications we consider K = C or K = R or
K = Zp , the finite field with p elements (p prime). Let R = K[x1 , . . . , xm ] be the
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
17
ring of the polynomials in the variables x1 , . . . , xm and with coefficients in K. The
ring operations in R are the usual sum and product of polynomials.
Definition 5.1. A subset I ⊂ R is an ideal if f + g ∈ I for all f, g ∈ I and f g ∈ I
for all f ∈ I and all g ∈ R.
Proposition 5.2. Let f1 , . . . , fr ∈ R. The set hf1 , . . . , fr i = {f1 g1 + · · · + fr gr :
g1 , . . . , gr ∈ R} is the smallest ideal in R with respect to the inclusion that contains
f1 , . . . , fr . The ideal hf1 , . . . , fr i is called the ideal generated by f1 , . . . , fr .
A key theorem in the theory of ideals in a polynomial ring is Hilbert’s basis
theorem, which states that every ideal in R is finitely generated.
Theorem 5.3. Given an ideal I ⊂ R, there exist f1 , . . . , fr ∈ I such that I =
hf1 , . . . , fr i.
The intersection of two ideals is an ideal. The union of two ideals is not an ideal
in general, but the following definition can be stated.
Definition 5.4. Let I, J ⊂ R be ideals. Then,
I + J = {f + g : f ∈ I, g ∈ J}
is the smallest ideal in R with respect to inclusion that contains both I and J, and
it is called the sum of I and J.
Given an ideal I, an equivalence relation is naturally defined on R: f ∼I g if
f − g ∈ I for f, g ∈ R. This relation is compatible with the ring operations: if
f1 ∼I f2 , g1 ∼I g2 then f1 + g1 ∼I f2 + g2 and f1 g1 ∼I f2 g2 . This fact implies that
the quotient space R/I can be defined and it is a ring with the operations inherited
from R.
The definitions below are the starting point of the computational side of Commutative Algebra.
Definition 5.5. A term in R is xa = xa1 1 . . . xamm for a = (a1 , . . . , am ) ∈ Nm . The
set of terms is denoted with Tm .
Definition 5.6. A term-ordering is a well ordering 4 on Tm such that 1 4 xa for
every xa ∈ Tm and xa 4 xb implies xa xc 4 xb xc for every xc ∈ Tm .
in R is a linear combination of a finite set of terms in Tm : f =
P A polynomial
a
m
a∈A ca x where A is a finite subset of N .
Definition 5.7. Let f ∈ R be a polynomial, A the finite set formed by the terms
in f and xb = max{xa : a ∈ A}. Let I ⊂ R be an ideal.
(1) The term LT(f ) = cb xb is called the leading term of f .
(2) The ideal generated by LT(f ) for every f ∈ I is called the order ideal of I
and is indicated as LT(I).
Definition 5.8. Let I ⊂ R be an ideal and let f1 , . . . , ft ∈ I. The set {f1 , . . . , ft }
is a Gröbner basis of I with respect to 4 if
LT(I) = hLT(f1 ), . . . , LT(ft )i.
Gröbner bases are special sets of generators for ideals in R with several applications. We list here only the results used in the paper.
18
FABIO RAPALLO AND MARIA PIERA ROGANTIN
Proposition 5.9. Let I ⊆ R be an ideal. Then, I = R if, and only if, 1 ∈ F ,
where F is a Gröbner basis of I, with respect to any term-ordering 4.
The theorem below is known as the finiteness theorem and is used to determine
whether a system of polynomial equations has a finite number of solutions. Here
V (I) is the variety defined by I, i.e., the set of the affine points (x1 , . . . , xm ) in Km
such that f (x1 , . . . xm ) = 0 for all f ∈ I.
Theorem 5.10. Let I ⊂ R = K[x1 , . . . , xm ] be an ideal and fix a term-ordering 4
on R. If K is algebraically closed, then the following five statements are equivalent:
(i) For each i = 1, . . . m, there is some ci > 0 such that xci i ∈ LT (I).
(ii) Let G be a Gröbner basis of I with respect to 4. Then for each i = 1, . . . , m,
there is some ci > 0 such that xci i = LT (f ) for some f ∈ G.
(iii) The set {xα : xα ∈
/ LT (I)} is finite.
(iv) The set R/I is finite-dimensional as K-vector space.
(v) V (I) is a finite set.
P
a
Definition 5.11. Let I ⊂ R be an ideal. A polynomial f =
a∈A ca x is in
a
normal form with respect to 4 and I if x ∈
/ LT(I) for each a ∈ A.
Proposition 5.12. Let I ⊂ R be an ideal. For every f ∈ R there exists a unique
polynomial, indicated as NF(f ) ∈ R, in normal form with respect to 4 and I such
that f − NF(f ) ∈ I.
The normal form of a polynomial f can be computed from f and a Gröbner basis
of I with respect to 4. The normal form solves the problem of ideal membership:
a polynomial f belongs to the ideal I if and only if NF(f ) = 0.
Definition 5.13. Given an ideal I ⊂ R = K[x1 , . . . , xm ], the l-th elimination ideal
Il is the ideal of K[xl+1 , . . . , xm ] defined by
Il = I ∩ K[xl+1 , . . . , xm ].
The elimination ideal is the generalization of the Gaussian elimination to the
polynomial case. Computing a Gröbner basis of I with respect to a special termordering, we can actually compute the ideal Il . The lex term-ordering on Tm is
defined by: for a 6= b, xa 4 xb if and only if the leftmost nonzero entry of the vector
difference b − a ∈ Zm is positive.
Theorem 5.14. Let I ⊂ R = K[x1 , . . . , xm ] be an ideal and let G be a Gröbner
basis of I with respect to lex term-ordering with x1 > x2 > · · · > x − n. Then, for
every l = 0, . . . , m, the set
Gl = G ∩ K[xl+1 , . . . , xm ]
is a Gröbner basis of the elimination ideal Il .
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
19
Appendix B. Indicator functions of fractions used in the paper
Fraction A in Example 2, first fraction in [7] for the 53 design:
FA =
1
(1 + (5ω3 + 5ω2 + 5ω1 − 2)X14 X24 X34 + (−5ω3 − 5ω2 − 5ω1 − 9)X12 X24 X34 +
5
(−4ω1 + 1)X1 X24 X34 + (5ω3 + 5ω1 − 3)X14 X23 X34 + (−5ω3 − 5ω1 − 8)X13 X23 X34 +
(−4ω1 + 1)X1 X23 X34 + (−5ω3 − 5ω2 − 8ω1 − 3)X14 X22 X34 + (5ω3 − 4)X13 X22 X34 +
(5ω2 + 5ω1 − 3)X12 X22 X34 + (−5ω3 − 7ω1 − 2)X14 X2 X34 + (−3ω2 + 2ω1 + 2)X13 X2 X34 +
(5ω3 − 3ω1 + 2)X12 X2 X34 + (−2ω1 + 13)X1 X2 X34 + (−5ω3 − 5ω1 − 8)X14 X24 X33 +
(5ω3 − 3ω1 + 2)X13 X24 X33 + (5ω1 − 4)X1 X24 X33 + (5ω3 − 4)X14 X23 X33 +
(−5ω3 − 7)X13 X23 X33 + (−4ω1 + 1)X12 X23 X33 + (−3ω2 + 2ω1 + 2)X14 X22 X33 +
(−5ω2 − 7ω1 − 2)X13 X22 X33 + (−2ω1 + 13)X12 X22 X33 + (5ω2 − 3ω1 + 2)X1 X22 X33 +
(5ω2 + 5ω1 − 3)X13 X2 X33 + (−4ω1 + 1)X12 X2 X33 + (−5ω2 − 5ω1 − 8)X1 X2 X33 +
(5ω2 + 5ω1 − 3)X14 X24 X32 + (−4ω1 + 1)X13 X24 X32 + (−5ω2 − 5ω1 − 8)X12 X24 X32 +
(5ω3 − 3ω1 + 2)X14 X23 X32 + (−2ω1 + 13)X13 X23 X32 + (−2ω3 + 3ω2 + 3ω1 + 3)X12 X23 X32 +
(−5ω3 − 5ω2 − 8ω1 − 3)X1 X23 X32 + (−4ω1 + 1)X13 X22 X32 + (−5ω2 − 7)X12 X22 X32 +
(5ω2 − 4)X1 X22 X32 + (−5ω3 − 5ω2 − 5ω1 − 9)X14 X2 X32 + (5ω2 − 3ω1 + 2)X12 X2 X32 +
(5ω3 + 5ω1 − 3)X1 X2 X32 + (−2ω1 + 13)X14 X24 X3 + (5ω2 − 3ω1 + 2)X13 X24 X3 +
(−5ω3 − 5ω2 − 8ω1 − 3)X12 X24 X3 + (5ω3 − 2ω2 + 3ω1 + 3)X1 X24 X3 + (−5ω2 − 5ω1 − 8)X13 X23 X3 +
(5ω2 − 4)X12 X23 X3 + (−3ω2 + 2ω1 + 2)X1 X23 X3 + (−4ω1 + 1)X14 X22 X3 +
(5ω3 + 5ω1 − 3)X12 X22 X3 + (−5ω3 − 5ω1 − 8)X1 X22 X3 + (−4ω1 + 1)X14 X2 X3 +
(5ω1 − 4)X13 X2 X3 + (−5ω1 − 7)X1 X2 X3 )
Fraction B in Example 2, second fraction in [7] for the 53 design:
FB =
4
i
1X
1
1 + X1 X2 X34 + X12 X22 X33 + X13 X23 X32 + X14 X24 X3 =
X1 X2 X34
5
5 i=0
Fraction C in Example 2, other fraction for the 53 design:
1
5 + (ω4 + ω3 + 3ω1 )X1 X24 X34 + (ω4 + 2ω2 + 2ω1 )X1 X23 X34 +
25
(2ω3 + ω2 + 2ω1 )X1 X22 X34 + (ω4 + ω1 + 3)X1 X2 X34 + (ω4 + 2ω2 + 2ω1 )X12 X24 X33 +
FC =
(ω3 + 3ω2 + ω1 )X12 X23 X33 + (ω3 + ω2 + 3)X12 X22 X33 + (2ω4 + ω3 + 2ω2 )X12 X2 X33 +
(2ω3 + ω2 + 2ω1 )X13 X24 X32 + (ω3 + ω2 + 3)X13 X23 X32 + (ω4 + 3ω3 + ω2 )X13 X22 X32 +
(2ω4 + 2ω3 + ω1 )X13 X2 X32 + (ω4 + ω1 + 3)X14 X24 X3 + (2ω4 + ω3 + 2ω2 )X14 X23 X3 +
(2ω4 + 2ω3 + ω1 )X14 X22 X3 + (3ω4 + ω2 + ω1 X14 X2 X3
20
FABIO RAPALLO AND MARIA PIERA ROGANTIN
Fraction D, fraction for the 55 design in Example 5:
1
5 + ω2 X14 X24 X34 + (ω3 + ω2 )X13 X24 X34 + (ω3 + ω1 + 1)X12 X24 X34 +
125
(ω4 + ω3 + ω2 + 1)X1 X24 X34 + (ω2 + 2ω1 + 1)X14 X23 X34 + (ω4 + ω3 + 1)X13 X23 X34 +
FD =
(ω4 + 1)X12 X23 X34 + X1 X23 X34 + (ω3 + ω2 + 2)X14 X22 X34 +
(ω3 + ω1 + 1)X13 X22 X34 + (ω2 + 1)X12 X22 X34 + ω4 X1 X22 X34 +
ω4 X14 X2 X34 + (ω3 + ω1 )X13 X2 X34 + (ω4 + ω1 + 1)X12 X2 X34 +
(ω4 + ω2 + ω1 + 1)X1 X2 X34 + (ω4 + 1)X14 X24 X33 + (ω4 + ω1 + 2)X13 X24 X33 +
ω3 X12 X24 X33 + (ω2 + ω1 + 1)X1 X24 X33 + (ω2 + ω1 + 1)X14 X23 X33 +
ω4 X13 X23 X33 + (ω4 + ω3 + ω1 + 1)X12 X23 X33 + (ω4 + ω1 )X1 X23 X33 +
(ω3 + ω2 + 1)X14 X22 X33 + ω3 X13 X22 X33 + (ω4 + ω3 + ω2 + 1)X12 X22 X33 +
(ω2 + ω1 )X1 X22 X33 + (ω3 + 1)X14 X2 X33 + (ω4 + 2ω2 + 1)X13 X2 X33 +
X12 X2 X33
+ (ω3 + ω1 + 1)X1 X2 X33 + (ω4 + ω2 + 1)X14 X24 X32 + X13 X24 X32 +
(2ω3 + ω1 + 1)X12 X24 X32 + (ω2 + 1)X1 X24 X32 + (ω4 + ω3 )X14 X23 X32 +
(ω3 + ω2 + ω1 + 1)X13 X23 X32 + ω2 X12 X23 X32 + (ω3 + ω2 + 1)X1 X23 X32 +
(ω4 + ω1 )X14 X22 X32 + (ω4 + ω2 + ω1 + 1)X13 X22 X32 + ω1 X12 X22 X32 +
(ω4 + ω3 + 1)X1 X22 X32 + (ω4 + ω3 + 1)X14 X2 X32 + ω2 X13 X2 X32 +
(ω4 + ω1 + 2)X12 X2 X32 + (ω1 + 1)X1 X2 X32 + (ω4 + ω3 + ω1 + 1)X14 X24 X3 +
(ω4 + ω1 + 1)X13 X24 X3 + (ω4 + ω2 )X12 X24 X3 + ω1 X1 X24 X3 +
ω1 X14 X23 X3 + (ω3 + 1)X13 X23 X3 + (ω4 + ω2 + 1)X12 X23 X3 +
(ω3 + ω2 + 2)X1 X23 X3 + X14 X22 X3 + (ω1 + 1)X13 X22 X3 +
(ω2 + ω1 + 1)X12 X22 X3 + (2ω4 + ω3 + 1)X1 X22 X3 + (ω3 + ω2 + ω1 + 1)X14 X2 X3 +
(ω4 + ω2 + 1)X13 X2 X3 + (ω3 + ω2 )X12 X2 X3 + ω3 X1 X2 X3 +
(ω4 + ω3 + 3ω1 )X14 X24 X44 X54 + (ω4 + 2ω2 + 2ω1 )X13 X24 X44 X54 +
(2ω3 + ω2 + 2ω1 )X12 X24 X44 X54 + (ω4 + ω1 + 3)X1 X24 X44 X54 +
(ω3 + ω2 + ω1 )X14 X34 X44 X54 + X13 X34 X44 X54 + (ω4 + ω2 + ω1 + 1)X12 X34 X44 X54 +
(ω2 + 1)X1 X34 X44 X54 + (ω4 + ω3 + ω2 + ω1 )X14 X23 X34 X44 X54 +
(ω3 + ω1 + 1)X13 X23 X34 X44 X54 + (ω4 + ω3 )X12 X23 X34 X44 X54 + ω4 X1 X23 X34 X44 X54 +
(ω2 + 3ω1 + 1)X22 X34 X44 X54 + (ω3 + ω1 + 1)X14 X2 X34 X44 X54 + ω1 X13 X2 X34 X44 X54 +
(2ω4 + ω2 + ω1 )X12 X2 X34 X44 X54 + (ω3 + ω1 )X1 X2 X34 X44 X54 + ω2 X14 X33 X44 X54 +
(ω4 + ω1 )X13 X33 X44 X54 + (ω3 + ω1 + 1)X12 X33 X44 X54 + (ω4 + ω3 + 2ω1 )X1 X33 X44 X54 +
(ω4 + ω1 + 1)X14 X23 X33 X44 X54 + ω3 X13 X23 X33 X44 X54 + (ω2 + 2ω1 + 1)X12 X23 X33 X44 X54 +
(ω2 + ω1 )X1 X23 X33 X44 X54 + (2ω2 + ω1 + 2)X22 X33 X44 X54 + ω3 X14 X2 X33 X44 X54 +
(ω4 + ω3 )X13 X2 X33 X44 X54 + (ω4 + ω2 + ω1 )X12 X2 X33 X44 X54 +
(ω4 + ω3 + ω1 + 1)X1 X2 X33 X44 X54 + (ω3 + 2ω2 + ω1 )X14 X32 X44 X54 +
(ω4 + ω1 + 1)X13 X32 X44 X54 + (ω1 + 1)X12 X32 X44 X54 + ω1 X1 X32 X44 X54 +
(ω4 + ω1 )X14 X23 X32 X44 X54 + (2ω3 + ω1 + 1)X13 X23 X32 X44 X54 + ω1 X12 X23 X32 X44 X54 +
(ω4 + ω2 + ω1 )X1 X23 X32 X44 X54 + (2ω4 + 2ω3 + ω1 )X22 X32 X44 X54 +
(ω4 + ω2 + ω1 + 1)X14 X2 X32 X44 X54 + (ω2 + ω1 + 1)X13 X2 X32 X44 X54 + (ω3 + 1)X12 X2 X32 X44 X54 +
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
21
ω2 X1 X2 X32 X44 X54 + (ω4 + 1)X14 X3 X44 X54 + (ω4 + ω3 + ω2 + ω1 )X13 X3 X44 X54 +
ω3 X12 X3 X44 X54 + (ω4 + ω3 + ω1 )X1 X3 X44 X54 + X14 X23 X3 X44 X54 +
3 3
(ω4 + ω2 )X1 X2 X3 X44 X54 + (ω2 + ω1 + 1)X12 X23 X3 X44 X54 + (ω3 + ω2 + ω1 + 1)X1 X23 X3 X44 X54 +
(ω4 + ω3 + 3ω1 )X22 X3 X44 X54 + (ω1 + 1)X14 X2 X3 X44 X54 + (ω2 + 2ω1 + 1)X13 X2 X3 X44 X54 +
ω4 X12 X2 X3 X44 X54 + (ω3 + ω2 + ω1 )X1 X2 X3 X44 X54 + (ω4 + 2ω2 + 2ω1 )X14 X23 X43 X53 +
(ω3 + 3ω2 + ω1 )X13 X23 X43 X53 + (ω3 + ω2 + 3)X12 X23 X43 X53 +
(2ω4 + ω3 + 2ω2 )X1 X23 X43 X53 + (ω2 + 1)X14 X34 X43 X53 + (2ω4 + ω2 + ω1 )X13 X34 X43 X53 +
ω2 X12 X34 X43 X53 + (ω3 + ω2 + 1)X1 X34 X43 X53 + (2ω3 + ω2 + 2ω1 )X24 X34 X43 X53 +
(ω1 + 1)X14 X22 X34 X43 X53 + (ω4 + ω3 + ω2 + 1)X13 X22 X34 X43 X53 + ω4 X12 X22 X34 X43 X53 +
(ω4 + ω2 + 1)X1 X22 X34 X43 X53 + ω2 X14 X2 X34 X43 X53 + (ω3 + ω2 )X13 X2 X34 X43 X53 +
(ω4 + ω3 + ω2 )X12 X2 X34 X43 X53 + (ω2 + 2ω1 + 1)X1 X2 X34 X43 X53 +
(ω4 + ω3 + ω2 + 1)X14 X33 X43 X53 + (ω4 + ω2 + ω1 )X13 X33 X43 X53 +
(2ω3 + ω2 + 2ω1 )X24 X34 X43 X53 + (ω1 + 1)X14 X22 X34 X43 X53 + (ω4 + ω3 + ω2 + 1)X13 X22 X34 X43 X53 +
ω4 X12 X22 X34 X43 X53 + (ω4 + ω2 + 1)X1 X22 X34 X43 X53 + ω2 X14 X2 X34 X43 X53 +
(ω3 + ω2 )X13 X2 X34 X43 X53 + (ω4 + ω3 + ω2 )X12 X2 X34 X43 X53 +
(ω2 + 2ω1 + 1)X1 X2 X34 X43 X53 + (ω4 + ω3 + ω2 + 1)X14 X33 X43 X53 +
(ω4 + ω2 + ω1 )X13 X33 X43 X53 + (ω4 + 1)X12 X33 X43 X53 + X1 X33 X43 X53 +
(ω4 + 3ω2 + 1)X24 X33 X43 X53 + (ω4 + 2ω3 + ω2 )X14 X22 X33 X43 X53 +
(ω2 + ω1 + 1)X13 X22 X33 X43 X53 + (ω2 + ω1 )X12 X22 X33 X43 X53 + ω2 X1 X22 X33 X43 X53 +
(ω3 + ω1 )X14 X2 X33 X43 X53 + (ω4 + ω3 + ω2 + ω1 )X13 X2 X33 X43 X53 + ω3 X12 X2 X33 X43 X53 +
(ω2 + ω1 + 1)X1 X2 X33 X43 X53 + ω1 X14 X32 X43 X53 + (ω3 + 1)X13 X32 X43 X53 +
(ω3 + ω2 + ω1 )X12 X32 X43 X53 + (ω4 + ω3 + ω2 + ω1 )X1 X32 X43 X53 + (ω3 + 3ω2 + ω1 )X24 X32 X43 X53 +
ω3 X14 X22 X32 X43 X53 + (ω2 + 1)X13 X22 X32 X43 X53 + (ω4 + ω2 + ω1 )X12 X22 X32 X43 X53 +
(ω4 + 2ω2 + 1)X1 X22 X32 X43 X53 + (ω4 + ω2 + 1)X14 X2 X32 X43 X53 + X13 X2 X32 X43 X53 +
(ω4 + ω2 + ω1 + 1)X12 X2 X32 X43 X53 + (ω4 + ω3 )X1 X2 X32 X43 X53 + (ω2 + ω1 + 1)X14 X3 X43 X53 +
ω4 X13 X3 X43 X53 + (ω3 + 2ω2 + ω1 )X12 X3 X43 X53 + (ω3 + ω2 )X1 X3 X43 X53 +
(2ω4 + ω2 + 2)X24 X3 X43 X53 + (ω4 + ω3 + ω2 )X14 X22 X3 X43 X53 + ω1 X13 X22 X3 X43 X53 +
(ω3 + ω2 + ω1 + 1)X12 X22 X3 X43 X53 + (ω3 + ω1 )X1 X22 X3 X43 X53 + (ω4 + 2ω2 + 1)X14 X2 X3 X43 X53 +
(ω3 + ω2 + 1)X13 X2 X3 X43 X53 + (ω4 + ω2 )X12 X2 X3 X43 X53 + ω1 X1 X2 X3 X43 X53 +
(2ω3 + ω2 + 2ω1 )X14 X22 X42 X52 + (ω3 + ω2 + 3)X13 X22 X42 X52 +
(ω4 + 3ω3 + ω2 )X12 X22 X42 X52 + (2ω4 + 2ω3 + ω1 )X1 X22 X42 X52 + (ω3 + ω2 )X14 X34 X42 X52 +
(ω4 + 2ω3 + ω2 )X13 X34 X42 X52 + ω1 X12 X34 X42 X52 + (ω4 + ω3 + 1)X1 X34 X42 X52 +
ω4 X14 X24 X34 X42 X52 + (ω3 + ω1 )X13 X24 X34 X42 X52 + (ω3 + ω2 + 1)X12 X24 X34 X42 X52 +
(2ω3 + ω1 + 1)X1 X24 X34 X42 X52 + (ω4 + ω2 )X14 X23 X34 X42 X52 + (ω4 + ω3 + ω2 + 1)X13 X23 X34 X42 X52 +
ω4 X12 X23 X34 X42 X52 + (ω3 + ω2 + ω1 )X1 X23 X34 X42 X52 + (ω3 + 2ω1 + 2)X2 X34 X42 X52 +
(ω4 + ω3 + ω2 + ω1 )X14 X33 X42 X52 + (ω4 + ω3 + ω2 )X13 X33 X42 X52 + (ω2 + 1)X12 X33 X42 X52 +
ω4 X1 X33 X42 X52 + (ω2 + ω1 )X14 X24 X33 X42 X52 + (ω4 + ω3 + ω1 + 1)X13 X24 X33 X42 X52 +
X12 X24 X33 X42 X52 + (ω3 + ω1 + 1)X1 X24 X33 X42 X52 + (2ω3 + ω1 + 1)X14 X23 X33 X42 X52 +
22
FABIO RAPALLO AND MARIA PIERA ROGANTIN
(ω4 + ω3 + ω1 )X13 X23 X33 X42 X52 + (ω3 + 1)X12 X23 X33 X42 X52 + ω2 X1 X23 X33 X42 X52 +
(ω4 + 3ω3 + ω2 )X2 X33 X42 X52 + X14 X32 X42 X52 + (ω1 + 1)X13 X32 X42 X52 +
(ω4 + ω3 + ω1 )X12 X32 X42 X52 + (ω3 + ω2 + ω1 + 1)X1 X32 X42 X52 + (ω4 + ω3 + 1)X14 X24 X32 X42 X52 +
ω2 X13 X24 X32 X42 X52 + (ω4 + ω3 + ω2 + ω1 )X12 X24 X32 X42 X52 + (ω4 + ω2 )X1 X24 X32 X42 X52 +
ω3 X14 X23 X32 X42 X52 + (ω4 + ω3 )X13 X23 X32 X42 X52 + (ω4 + ω3 + 1)X12 X23 X32 X42 X52 +
(ω3 + 2ω2 + ω1 )X1 X23 X32 X42 X52 + (3ω3 + ω1 + 1)X2 X32 X42 X52 + (ω3 + ω2 + 1)X14 X3 X42 X52 +
ω3 X13 X3 X42 X52 + (ω4 + ω3 + 2ω1 )X12 X3 X42 X52 + (ω3 + 1)X1 X3 X42 X52 +
(2ω4 + ω3 + 1)X14 X24 X3 X42 X52 + (ω3 + ω2 + ω1 )X13 X24 X3 X42 X52 + (ω3 + ω2 )X12 X24 X3 X42 X52 +
ω3 X1 X24 X3 X42 X52 + (ω3 + ω1 + 1)X14 X23 X3 X42 X52 + ω1 X13 X23 X3 X42 X52 +
(ω3 + ω2 + ω1 + 1)X12 X23 X3 X42 X52 + (ω4 + 1)X1 X23 X3 X42 X52 + (2ω4 + ω3 + 2ω2 )X2 X3 X42 X52 +
(ω4 + ω1 + 3)X14 X2 X4 X5 + (2ω4 + ω3 + 2ω2 )X13 X2 X4 X5 + (2ω4 + 2ω3 + ω1 )X12 X2 X4 X5 +
(3ω4 + ω2 + ω1 )X1 X2 X4 X5 + (ω4 + ω2 + ω1 )X14 X34 X4 X5 + ω2 X13 X34 X4 X5 +
(ω4 + ω3 + ω2 + ω1 )X12 X34 X4 X5 + (ω1 + 1)X1 X34 X4 X5 + (ω4 + ω3 + ω2 )X14 X24 X34 X4 X5 +
ω1 X13 X24 X34 X4 X5 + (2ω4 + ω3 + 1)X12 X24 X34 X4 X5 + (ω4 + 1)X1 X24 X34 X4 X5 +
(3ω4 + ω2 + ω1 )X23 X34 X4 X5 + (ω4 + ω3 + ω2 + 1)X14 X22 X34 X4 X5 +
(ω4 + ω3 + 1)X13 X22 X34 X4 X5 + (ω3 + ω1 )X12 X22 X34 X4 X5 + X1 X22 X34 X4 X5 +
ω4 X14 X33 X4 X5 + (ω4 + 1)X13 X33 X4 X5 + (ω4 + ω1 + 1)X12 X33 X4 X5 + (ω4 + 2ω3 + ω2 )X1 X33 X4 X5 +
ω3 X14 X24 X33 X4 X5 + (ω2 + 1)X13 X24 X33 X4 X5 + (ω4 + ω3 + 1)X12 X24 X33 X4 X5 +
(ω4 + ω3 + ω1 + 1)X1 X24 X33 X4 X5 + (ω4 + 2ω2 + 2ω1 )X23 X33 X4 X5 + (ω4 + ω3 + ω1 )X14 X22 X33 X4 X5 +
ω4 X13 X22 X33 X4 X5 + (ω4 + 2ω2 + 1)X12 X22 X33 X4 X5 + (ω4 + ω1 )X1 X22 X33 X4 X5 +
(2ω4 + ω2 + ω1 )X14 X32 X4 X5 + (ω4 + ω2 + 1)X13 X32 X4 X5 + (ω4 + ω1 )X12 X32 X4 X5 +
ω3 X1 X32 X4 X5 + (ω4 + ω2 + ω1 + 1)X14 X24 X32 X4 X5 + (ω4 + ω3 +
ω1 )X13 X24 X32 X4 X5 + (ω2 + ω1 )X12 X24 X32 X4 X5 + ω2 X1 X24 X32 X4 X5 +
(ω4 + 2ω3 + 2)X23 X32 X4 X5 + (ω4 + ω3 )X14 X22 X32 X4 X5 + (2ω4 + ω3 + 1)X13 X22 X32 X4 X5 +
ω2 X12 X22 X32 X4 X5 + (ω4 + ω1 + 1)X1 X22 X32 X4 X5 + (ω3 + 1)X14 X3 X4 X5 +
(ω4 + ω3 + ω1 + 1)X13 X3 X4 X5 + X12 X3 X4 X5 + (ω4 + ω3 + ω2 )X1 X3 X4 X5 +
(ω4 + ω2 )X14 X24 X3 X4 X5 + (ω4 + ω3 + 2ω1 )X13 X24 X3 X4 X5 + ω4 X12 X24 X3 X4 X5 +
(ω4 + ω2 + 1)X1 X24 X3 X4 X5 + (3ω4 + ω3 + 1)X23 X3 X4 X5 + ω1 X14 X22 X3 X4 X5 +
(ω2 + ω1 )X13 X22 X3 X4 X5 + (ω4 + ω2 + 1)X12 X22 X3 X4 X5 + (ω4 + ω3 + ω2 + ω1 )X1 X22 X3 X4 X5
Acknowledgements
We thank Anna Bigatti (Università di Genova) for her valuable help in using CoCoA. This work
is partially funded by INdAM (National Institute for Higher Mathematics) through a GNAMPAINdAM Project 2017. This research is original and has a financial support of the Università del
Piemonte Orientale.
References
[1] J. Abbott, A. M. Bigatti, and G. Lagorio. CoCoA-5: a system for doing Computations in
Commutative Algebra. Available at http://cocoa.dima.unige.it, 2015.
[2] John Abbott. Sparse squares of polynomials. Math. Comp., 71(237):407–413, 2002.
[3] James B. Clark and Angela M. Dean. Equivalence of fractional factorial designs. Statist.
Sinica, 11(2):537–547, 2001.
ALGEBRAIC CHARACTERIZATION OF REGULAR FRACTIONS
23
[4] Robert M. Corless and Nicolas Fillion. A graduate introduction to numerical methods.
Springer, 2013.
[5] David Cox, John Little, and Donal O’Shea. Ideals, Varieties, and Algorithms: An Introduction to Computational Algebraic Geometry and Commutative Algebra. Springer, 2015.
[6] Angela Dean, Max Morris, John Stufken, and Derek Bingham. Handbook of Design and
Analysis of Experiments. CRC Press, 2015.
[7] Pieter
Eendebak.
Complete
series
of
non-isomorphic
orthogonal
arrays.
http://pietereendebak.nl/oapage/ , 2017. Accessed: 2017-03-30.
[8] Roberto Fontana, Giovanni Pistone, and Maria Piera Rogantin. Classification of two-level
factorial fractions. J. Statist. Plann. Inference, 87(1):149–172, 2000.
[9] Roberto Fontana, Fabio Rapallo, and Maria Piera Rogantin. Aberration in qualitative multilevel designs. J. Statist. Plann. Inference, 174:1–10, 2016.
[10] Paolo Gibilisco, Eva Riccomagno, Maria Piera Rogantin, and Henry P. Wynn. Algebraic and
geometric methods in statistics. In Algebraic and geometric methods in statistics, pages 1–24.
Cambridge Univ. Press, Cambridge, 2010.
[11] Ulrike Grömping and Rosemary A. Bailey. Regular fractions of factorial arrays. In Joachim
Kunert, Christine H. Müller, and Anthony C. Atkinson, editors, mODa 11 - Advances in
Model-Oriented Design and Analysis: Proceedings of the 11th International Workshop in
Model-Oriented Design and Analysis held in Hamminkeln, Germany, June 12-17, 2016,
pages 143–151. Springer, Cham, 2016.
[12] A. Samad Hedayat, Neil James Alexander Sloane, and John Stufken. Orthogonal Arrays:
Theory and Applications. Springer, New York, 1999.
[13] Tena I. Katsaounis. Equivalence of factorial designs with qualitative and quantitative factors.
J. Statist. Plann. Inference, 142(1):79–85, 2012.
[14] Tera I. Katsaounis and Angela M. Dean. A survey and evaluation of methods for determination of combinatorial equivalence of factorial designs. J. Statist. Plann. Inference, 138(1),
2008.
[15] Tera I. Katsaounis, Angela M. Dean, and Bradley Jones. On equivalence of fractional factorial
designs based on singular value decomposition. J. Statist. Plann. Inference, 143(11):1950–
1953, 2013.
[16] A. Donald Keedwell and József Dénes. Latin squares and their applications. Elsevier/NorthHolland, second edition, 2015.
[17] Martin Kreuzer and Lorenzo Robbiano. Computational Commutative Algebra 1. Springer,
2008.
[18] Chang-Xing Ma, Kai-Tai Fang, and Dennis K. J. Lin. On the isomorphism of fractional
factorial designs. J. Complexity, 17(1):86–97, 2001.
[19] Fang Pang and Min-Qian Liu. Geometric isomorphism check for symmetric factorial designs.
J. Complexity, 27(5):441–448, 2011.
[20] Giovanni Pistone and Maria-Piera Rogantin. Indicator function and complex coding for mixed
fractional factorial designs. J. Statist. Plann. Inference, 138(3):787–802, 2008.
[21] Eric D. Schoen, Pieter T. Eendebak, and Man V.M. Nguyen. Complete enumeration of purelevel and mixed-level orthogonal arrays. Journal of Combinatorial Designs, 18(2):123–140,
2010.
[22] Yu Tang and Hongquan Xu. Permuting regular fractional factorial designs for screening quantitative factors. Biometrika, 101(2):333–350, 2014.
[23] C.F.Jeff Wu and Michael Hamada. Experiments: Planning, Analysis, and Parameter Design
Optimization. Wiley, 2000.
Department of Science and Technological Innovation, University of Piemonte Orientale, Alessandria, Italy
E-mail address: [email protected]
Department of Mathematics, University of Genova, Genova, Italy
E-mail address: [email protected]
| 10math.ST
|
arXiv:1403.4724v1 [math.AC] 19 Mar 2014
Equivariant class group. II.
Enriched descent theorem
Mitsuyasu Hashimoto
Department of Mathematics, Okayama University
Okayama 700–8530, JAPAN
[email protected]
Abstract
We prove a version of Grothendieck’s descent theorem on an ‘enriched’ principal fiber bundle, a principal fiber bundle with an action
of a larger group scheme. Using this, we prove the isomorphisms of the
equivariant Picard and the class groups arising from such a principal
fiber bundle.
1.
Introduction
This paper is a continuation of [Has3].
Let S be a scheme, and G a flat S-group scheme. In [Has3], we have
defined the equivariant class group Cl(G, X) for a locally Krull G-scheme X.
Utilizing Grothendieck’s descent theorem, we have proved that for a principal
G-bundle ϕ : X → Y , Y is locally Krull, and the inverse image functor ϕ∗
induces an isomorphism Cl(Y ) → Cl(G, X) ([Has3, (5.32)]).
In this paper, we generalize this to an enriched version. Let f : G → H
be an fpqc homomorphism, and N := Ker f . Let ϕ : X → Y be a Gmorphism which is also a principal N-bundle. We call such a morphism a
G-enriched principal N-bundle (the name ‘equivariant principal bundle’ is
2010 Mathematics Subject Classification. Primary 14L30. Key Words and Phrases.
principal fiber bundle, descent theory, class group, Picard group
1
reserved for a different notion, which is deeply related, see Example 7.6).
The purpose of this paper is to prove the enriched descent theorem, which
yields an isomorphism ϕ∗ : Cl(H, Y ) → Cl(G, X). A similar isomorphism
ϕ∗ : Pic(H, Y ) → Pic(G, X) is also proved.
The isomorphism is induced by the corresponding equivalence of the categories of quasi-coherent sheaves, ϕ∗ : Qch(H, Y ) → Qch(G, X). This enriched Grothendieck’s descent is conceptually trivial, and probably checked
relatively easily for simpler cases. In this paper, we do not assume that ϕ is
flat, quasi-compact, quasi-separated, or locally of finite type (these assumptions are automatically satisfied if N → S satisfies the same conditions, see
Lemma 2.12). However, the original Grothendieck’s descent is known under
the same general settings (even more is known, see [Vis, (4.46)]), and we try
to prove the enriched version without these redundant assumptions.
The quasi-inverse of ϕ∗ is (?)N ◦ ϕ∗ , the direct image followed by the Ninvariance. However, this does not mean ϕ∗ preserves the quasi-coherence,
as we do not assume that ϕ is quasi-compact quasi-separated. As we do
not assume that ϕ is flat, there are some technical problems in treating the
quasi-coherent sheaves on the small Zariski site, and it is more comfortable to
treat the big site with flat topology. Finally, as we do not put any finiteness
assumptions on ϕ : X → Y , it is suitable to treat the fpqc topology, not the
fppf topology.
As a biproduct, we have an isomorphism ϕ∗ : Pic(H, Y ) → Pic(G, X) for
non-flat huge groups (Corollary 7.2).
Sections 2 to 5 are devoted to preliminaries.
In section 2, we define quasi-fpqc (qfpqc for short) morphisms of schemes
(Definition 2.2). The topology defined by quasi-fpqc morphisms is the same
as that defined by fpqc morphisms. However, any group scheme is qfpqc, and
it is comfortable to treat qfpqc morphisms when we do not put the flatness
assumption on groups. We point out that various properties of morphisms
of schemes descend with respect to qfpqc morphisms (Lemma 2.4).
In section 3, we give a way to make the big fpqc site skeletally small
in a reasonable way. We measure the ‘size’ of a morphism by a cardinal,
and we only treat the morphisms whose size is bounded by a fixed regular
cardinal. This way, we can avoid set-theoretic problems in treating fpqc
topology. Sizing schemes using a cardinal is treated in [dJ, (3.9)], and our
approach can be viewed as a relative version of that in [dJ]. Our definition
of κ-morphisms for a regular cardinal κ enables us to treat all schemes T (of
arbitrary size) as a base scheme (see the condition 4 in (5.1)).
2
Section 4 is a preliminary on sheaves on ringed sites. As an abstraction
of Kempf’s result [Kem, Theroem 8], we give a sufficient condition for cohomology functors to be compatible with direct limits (Lemma 4.7). Using
this, we compare quasi-coherent sheaves on different ringed sites.
In section 5, we introduce several sites related to diagrams of schemes.
It is convenient to grasp G-equivariant quasi-coherent sheaves on X as a
M
quasi-coherent sheaves over the finite diagram of schemes BG
(X), and a wellbehaved treatment of dualizing complexes is known for this category [Has2].
However, sometimes we want to consider the full simplicial schemes. It seems
that this would be necessary when we pursue the cohomological descent as in
[SD]. Moreover, for our purpose, considering big sites is necessary. Some of
our argument in section 6 does not work for small Zariski site. The point is
the exactness of the inverse image (Lemma 5.13) and the fact that Lipman’s
theta is an isomorphism (Lemma 5.14).
On the other hand, the big sites with non-flat morphisms are too big when
we consider the derived category, since the full subcategory of quasi-coherent
sheaves is not closed under kernels in the whole category of modules. So we
compare these sites, and give a proof for the fact that the categories of quasicoherent sheaves are all equivalent, although it seems that this is well-known
for experts.
In the last section, section 6, we prove the enriched Grothendieck’s descent theorem. Our main theorem is on the equivariant modules on the
fpqc site, Theorem 6.20. The author does not know if the same statement
for the small Zariski site is true. We get the corresponding assertion for
quasi-coherent sheaves (note that the quasi-coherent sheaves are essentially
independent of the choice of the ringed site, as proved in section 5) immediately (Corollary 6.21). We do not know how to prove Corollary 6.21 directly
without using the fpqc site in this generality, although it seems that the additional assumption that G is flat and quasi-compact quasi-separated would
make it possible relatively easily. The choice of the fpqc site is effective in
our proof. See the proof of Lemma 6.18 and Theorem 6.20.
As corollaries, we prove the isomorphisms between the equivariant Picard
and the class groups arising from an enriched principal fiber bundle, see
Corollary 7.2 and Corollary 7.4.
We will see some applications of the enriched descent in the continuation
of this paper [Has4].
Acknowledgment: The author is grateful to Professor Shouhei Ma for
kindly showing him the reference [Vis].
3
2.
Quasi-fpqc morphisms and enriched principal bundles
(2.1) This paper is a continuation of [Has3]. We follow the notation and
terminology there, unless otherwise specified. Throughout this paper, let S
be a scheme.
Definition 2.2. A morphism of schemes ϕ : X → Y is said to be quasi-fpqc
(or qfpqc for short) if there exists some morphism ψ : Z → X such that ϕψ
is fpqc.
Lemma 2.3. The following hold true.
1 A morphism ϕ : X → Y is fpqc if and only if it is qfpqc and flat.
2 A base change of a qfpqc morphism is qfpqc.
3 A composite of qfpqc morphisms is qfpqc.
4 A qfpqc morphism is submersive. In particular, it is surjective.
5 A group scheme G over S is qfpqc over S.
Proof. 1 The ‘only if’ part is trivial. We prove the ‘if’ part. So there is a
morphism ψ : Z → X such that ϕψ is fpqc. Let U be a quasi-compact open
subset of Y . Then there is a quasi-compact open subset V of Z such that
ϕ(ψ(V )) = U. As ψ(V ) is quasi-compact and contained in ϕ−1 (U), there is a
quasi-compact open subset W of ϕ−1 (U) containing ψ(V ). Then ϕ(W ) = U.
So ϕ is fpqc.
2 and 3 are easy. 4 follows from [Gro2, (2.3.12)]. 5 is because S → G → S
is the identity, where the first map S → G is the unit of the group scheme.
Lemma 2.4. Let P be a property of morphisms of schemes. Assume that
S
1 (Zariski local property) If f : X → Y is a morphism and Y = i Ui
is an affine open covering, then f satisfies P if and only if f |f −1 (Ui ) :
f −1 (Ui ) → Ui satisfies P for each i.
2 Let f : X → Y be a morphism, g : Y ′ → Y a morphism, and assume
that Y and Y ′ are affine.
(i) (base change for affine bases) If f satisfies P, then the base change
f ′ : Y ′ ×Y X → Y ′ of f by g satisfies P.
4
(ii) (flat descent for affine bases) If f ′ satisfies P and g is faithfully
flat, then f satisfies P.
Then if f : X → Y is a morphism of schemes, g : Y ′ → Y a qfpqc morphism
of schemes and the base change f ′ satisfies P, then f satisfies P. In particular,
if f ′ satisfies one of separated, quasi-compact, quasi-separated, locally of finite
presentation, proper, affine, finite, flat, faithfully flat, smooth, unramified,
étale, submersive, a closed immersion, an open immersion, an immersion,
an isomorphism, fpqc, and qfpqc, then f satisfies the same property. If Q is
a property of an algebra essentially of finite type over a field such that the
base change and the descent holds for any base field extension, and P is ‘any
fiber satisfies Q,’ then P descends with respect to a qfpqc base change. So if
Q is one of geometrically normal, geometrically reduced, Cohen–Macaulay,
Gorenstein, and local complete intersection, and f ′ satisfies P, then f satisfies
P.
Proof. The first assertion is rather formal, and is left to the reader. Most of
the examples of P are listed in [Vis, (2.36)]. We only prove that the property
P = qfpqc satisfies 2, (ii). As f ′ is qfpqc, there is a morphism h : U → X ′
such that f ′ h is fpqc. As g is also fpqc, the composite gf ′h = f g ′h is fpqc,
where g ′ : X ′ = Y ′ ×Y X → X is the second projection. So f is qfpqc, as
required.
(2.5) From now on, unless otherwise specified, let G be an S-group scheme,
and N ⊂ G a normal subgroup scheme of G. That is, N is a subscheme of G,
N itself is an S-group scheme, and the inclusion N ֒→ G is a homomorphism
such that G × N → G ((g, n) 7→ gng −1) factors through N.
Definition 2.6. We say that ϕ : X → Y is a G-enriched principal N-bundle
if it is a G-morphism, and is a principal N-bundle [Has3, (2.6)].
The following is immediate from the definition.
Lemma 2.7. Let ϕ : X → Y be a G-enriched principal N-bundle, and
h : Y ′ → Y be a G-morphism. If the action of N on Y ′ is trivial, then the
second projection p2 : X ×Y Y ′ → Y ′ is a G-enriched principal N-bundle.
Lemma 2.8. An N-invariant G-morphism ϕ : X → Y is a G-enriched
principal N-bundle if and only if ϕ is qfpqc and the map Φ : N ×X → X ×Y X
given by Φ(n, x) = (nx, x) is an isomorphism.
5
Proof. This is immediate from [Vis, (4.43)].
Lemma 2.9. A qfpqc morphism is an epimorphism. That is, if f : X → Y
is a qfpqc morphism, g and h are morphisms Y → Z, and gf = hf , then
g = h.
Proof. Set S := Spec Z. Then there is a commutative diagram
X0
Y0
/
dX
X
/
Z
dY
f
/
Y
(g,h)
/
∆
Z ×S Z
with cartesian squares. By assumption, dX is an isomorphism. By Lemma 2.4,
dY is an isomorphism, and hence g = h.
Corollary 2.10. Let S be a scheme, F an S-group scheme, and ϕ : X → Y
an qfpqc F -morphism. If the action of F on X is trivial, then the action of
F on Y is trivial.
Proof. Let f = 1F × ϕ : F × X → F × Y . It is qfpqc. Then af = p2 f , as ϕ
is an F -morphism and the action of F on X is trivial, where a : F × Y → Y
is the action, and p2 : F × Y → Y is the second projection. By Lemma 2.9,
a = p2 , and the action of F on Y is trivial.
Lemma 2.11. Let X and Y be G-schemes, and ϕ : X → Y be a Gmorphism. Then the following are equivalent.
1 ϕ is a G-enriched principal N-bundle.
2 There exists some qfpqc G-morphism h : Y ′ → Y such that the second
projection p2 : X ×Y Y ′ → Y ′ is a G-morphism which is a trivial Nbundle.
3 There exists some qfpqc S-morphism h : Y ′ → Y such that the second
projection p2 : X ′ = X ×Y Y ′ → Y ′ is a principal N-bundle.
Proof. 1⇒2 Let h = ϕ, and use Lemma 2.8.
2⇒3 is trivial.
3⇒1 There exists some fpqc h′ : Y ′′ → Y ′ such that the base change
X ′′ → Y ′′ is a trivial N-bundle. As hh′ is qfpqc, there exists some h′′ :
Y ′′′ → Y ′′ such that hh′ h′′ is fpqc. As the base change X ′′′ → Y ′′′ is also a
trivial N-bundle, ϕ : X → Y is a principal N-bundle. As ϕ is assumed to be
a G-morphism, it is a G-enriched principal N-bundle.
6
Lemma 2.12. Let ϕ : X → Y be a principal G-bundle, and P a property of
morphisms of schemes as in Lemma 2.4. If G → S satisfies P, then ϕ also
satisfies P.
Proof. There is an fpqc morphism Y ′ → Y such that the base change ϕ′ :
Y ′ ×Y X → Y ′ is a trivial G-bundle. By the Zariski local property and
the base change for affine bases, it is easy to check the general base change
property, and hence ϕ′ satisfies P. By Lemma 2.4, ϕ satisfies P.
3.
κ-schemes
(3.1) Let κ be a cardinal. A topological space X is said to be κ-compact if
every open cover of X has a subcover of cardinality strictly less than κ. So X
is 1-compact (resp. 2-compact, ℵ0 -compact, ℵ1 -compact) if and only if X is
empty (resp. local [HO, (8.7)], (quasi-)compact, Lindelöf), where ℵ0 = #Q,
and ℵ1 is the successor cardinal of ℵ0 .
If f : X → Y is a surjective continuous map and X is κ-compact, then
so is Y .
(3.2) A cardinal is said to be regular if it is equal to its own cofinality
[End, p. 257]. In our paper, a regular cardinal is required
S to be infinite. So
κ is regular if and only if for any decomposition κ = i∈I Si for a family of
subsets with #I < κ, #Si = κ for at least one i [End, Theorem 9T]. Note
that ℵ0 and infinite successor cardinals are regular. By the definition, an
inaccessible cardinal is regular [End, p. 254].
S
Let κ be regular. Let (Yi )i∈I be a family of subspaces of X. If i Yi = X,
each Yi is κ-compact, and #I < κ, then X is also κ-compact.
(3.3) Let κ be infinite. (The underlying space of) a scheme is κ-compact
if and only if it is covered by an affine open covering of cardinality strictly
less than κ.
(3.4) A morphism of schemes ϕ : X → Y is said to be κ-compact if there
exists some affine open covering (Ui )i∈I of Y such that each ϕ−1 (Ui ) is κcompact. If κ < κ′ and ϕ is κ-compact, then it is κ′ -compact. A quasicompact morphism is nothing but an ℵ0 -compact morphism. So a quasicompact morphism is κ-compact.
(3.5) Note that ϕ : X → Y is κ-compact, and h : Y ′ → Y is another
morphism such that Y ′ is quasi-compact, then Y ′ ×Y X is κ-compact. In
7
particular, a base change of a κ-compact morphism is again κ-compact. If Y
is quasi-compact, ϕ : X → Y is κ-compact, then X is κ-compact.
(3.6) Let κ be regular. If ϕ : X → Y is a morphism of schemes, Y is
κ-compact, and ϕ is κ-compact, then X is κ-compact. It follows that a
composite of κ-compact morphisms is again κ-compact in this case.
(3.7) Let κ be a regular cardinal. A morphism of schemes ϕ : X → Y
is said to be κ-quasi-separated if the diagonal ∆X/Y : X → X ×Y X is κcompact. A quasi-separated morphism is nothing but an ℵ0 -quasi-separated
morphism. As in [Gro, (1.2)], we can prove the following. An immersion
is κ-quasi-separated. A base change of a κ-quasi-separated morphism is κquasi-separated. Let f : X → Y and g : Y → Z be morphisms. If f and g
are κ-quasi-separated, then so is gf . If gf is κ-quasi-separated, then so is f .
If g is κ-quasi-separated and gf is κ-compact, then f is κ-compact.
(3.8) We call a κ-compact κ-quasi-separated morphism a κ-concentrated
morphism. The composition of κ-concentrated morphisms is κ-concentrated.
The base change of a κ-concentrated morphism is κ-concentrated. If gf and
g are κ-concentrated, then so is f .
(3.9) Let λ be an infinite cardinal. Let h : A → B be a map of commutative
rings. We say that h is of λ-type if B is generated by a subset whose cardinal
is strictly less than λ over A. So h is of finite type if and only if h is of
ℵ0 -type. We say that a ring A is λ-type if Z → A is of λ-type. If λ > ℵ0 ,
then A is of λ-type if and only if #A < λ. In particular, if λ > ℵ0 , then a
subring of a ring of λ-type is again of λ-type. If A is a pure subring of B,
and B is a ring of λ-type, then A is also of λ-type. If λ > ℵ0 , then this is
because A ⊂ B. If λ = ℵ0 , then B is of finite type over Z, and hence so is A
by [Has]. In particular, if A → B is faithfully flat and B is of λ-type, then
so is A. Let B be an A-algebra, and C a B-algebra. If B is of λ-type over A
and C is of λ-type over B, then C is of λ-type over A. If C is of λ-type over
A, then it is of λ-type over B.
(3.10) A morphism of schemes ϕ : X → Y is said to be locally of λ-type
if there exists some affine open covering (Ui )i∈I and affine open covering
(Vji )j∈Ji of ϕ−1 (Ui ) for each i such that Γ(Ui , OY ) → Γ(Vji , OX ) is of λ-type.
A morphism locally of finite-type is locally of λ-type. It is easy to see that
a base change of a locally λ-type morphism is again locally of λ-type. Note
that Spec B → Spec A is locally of λ-type if and only if A → B is of λ-type.
8
The composite of two locally λ-type morphisms is again locally of λ-type.
In particular, a morphism of schemes ϕ : X → Y is locally of λ-type if and
only if for any affine open subscheme U = Spec A of Y and any affine open
subscheme V = Spec B of ϕ−1 (U), A → B is of λ-type. If f : X → Y and
g : Y → Z are morphisms and gf is locally of λ-type, then so is f .
(3.11) From now on, until the end of this paper, let κ denote a regular
cardinal. We say that a morphism of schemes ϕ : X → Y is a κ-morphism
if it is locally of κ-type and κ-concentrated. A scheme X is said to be a κscheme if X → Spec Z is a κ-morphism. The composite of two κ-morphisms
is a κ-morphism. In particular, if Y is a κ-scheme and ϕ : X → Y is a
κ-scheme, then X is a κ-scheme. A base change of a κ-morphism is a κmorphism. In particular, a direct product of κ-schemes is a κ-scheme. If
f : X → Y and g : Y → Z are morphisms and gf is a κ-morphism, then so
is f .
(3.12) For a ring A, the number of quasi-compact open subsets of Spec A
is finite or less than or equal to #A, because the compact open subsets of the
form Spec A[1/f ] form an open basis. It follows that if a ring A is of κ-type,
then any subset of Spec A is κ-compact. So any subscheme of a κ-scheme is
again a κ-scheme. If X is a κ-scheme, then a local ring OX,x is a κ-ring. If
ϕ : X → Y is an fpqc morphism and X is a κ-scheme, then Y is a κ-scheme.
For a given set of schemes Ω, there exists some κ such that any element of
Ω is a κ-scheme.
(3.13) We can add the properties of morphisms κ-compact, κ-quasi-separated,
locally of κ-type to the list of properties P in Lemma 2.4. So these properties
descends with respect to qfpqc base change. In particular, by Lemma 5.31,
5 and Lemma 2.12, a principal G-bundle is a κ-morphism, if G → S is so.
(3.14) For a given scheme S, we denote the full subcategory of Sch/S
consisting of objects such that the structure morphisms are κ-morphisms by
(Sch/S)κ . If T ∈ (Sch/S)κ , then (Sch/S)κ /T is the same as (Sch/T )κ .
Lemma 3.15. Let S be a scheme, and κ a regular cardinal. Then the category (Sch/S)κ is skeletally small.
Proof. Replacing κ if necessary, we may assume that S is a κ-scheme. So
(Sch/S)κ is equivalent to (Sch/Z)κ /S. If a category C is skeletally small and
c ∈ C, then C/c is skeletally small. So We may assume that S = Spec Z. Let
9
R be a complete set of representatives of isomorphism classes of Z-algebras of
κ-type. Note that R can be a subset of the set of quotients of the polynomial
ring Z[xα ]α∈κ , and certainly a small set. Thus the category A of affine κschemes is skeletally small.
Let I0 = κ and I1 `
= I0 × I0 × κ. Let I be the small category with the
object set Ob(I) = I0 I1 , and Hom(i, j) is a singleton if j = (j, j ′ , k) ∈ I1
and i ∈ {j, j ′ } ⊂ I0 , and empty otherwise. The category of I op -diagrams
op
of affine κ-schemes B =
S Func(I , A) is skeletally small. Any κ-scheme can
be expressed as X = i∈I0 Ui with Ui ∈ A. Note that each Ui ∩ Uj is κcompact, since
S X is κ-quasi-separated. So for each (i, j), there is a covering
Ui ∩ Uj = k∈κ Uijk with Uijk ∈ A. So the collections determines an object
X = (((Ui ), (Uikj ))) of B, and X is the colimit of X . By the uniqueness of
the colimit, the category of κ-schemes is also skeletally small.
4.
Module sheaves over a ringed site
(4.1) Let C = (C, OC ) be a ringed site. That is, C is a site (a category
equipped with a pretopology) and OC is a sheaf of commutative rings on C.
The category of OC -modules is denoted by Mod(C). A free sheaf on C is
an OC -module which is isomorphic to a direct sum of copies of the OC -module
OC . A sheaf of OC -modules M is said to be quasi-coherent (resp. invertible)
if for any c ∈ C, there exists some covering (cλ → c) such that for each λ,
there is an exact sequence of sheaves of OC |cλ -modules of the form
F1 → F0 → M|cλ → 0
with F1 and F0 free (resp. F1 = 0 and F0 = OC |cλ ). Obviously, a quasicoherent sheaf is invertible. The category of quasi-coherent sheaves (resp.
invertible sheaves) on C is denoted by Qch(C) (resp. Inv(C)).
(4.2) Assume that C has an initial object X. Then for an Γ(X, OC )-module
M, A presheaf M̃ p is defined by Γ(c, M̃ p ) = Γ(c, OC ) ⊗Γ(X,OC ) M. Its sheafification is denoted by M̃ .
An OC -module of the form M̃ is quasi-coherent.
(4.3) Let C and D be sites and f : D → C a functor. In this paper, we say
that f is continuous if for any d ∈ D and any covering (dλ → d), (f dλ → f d)
is a covering again. Let f = (f, η) : (C, OC ) → (D, OD ) be a morphism of
ringed sites. That is, f : D → C is a continuous functor, and η : OD → f∗ OC
10
is a map of sheaves of commutative rings. Then almost by definition, if
M ∈ Mod(D) is quasi-coherent (resp. invertible), then so is f ∗ M.
(4.4) Let C be a site. An EC -morphism is a morphism that appears as a
part of a covering. A subset B of Ob(C) is said to be a basis of the topology
of C if for any U ∈ Ob(C) and any covering (Vλ → U), there is a refinement
(Wµ → U) of (Vλ → U) such that each Wµ belongs to B. An object U of C
is said to be quasi-compact if for any covering (Vλ → U)λ∈Λ of U, there is a
finite subset Λ0 of Λ such that (Vλ → U)Λ0 is also a covering of U. An object
U of C is said to be quasi-separated if for any quasi-compact objects V , W
and any EC -morphisms V → U and W → U, V ×U W is again quasi-compact.
A subset B of Ob(C) is said to be quasi-compact (resp. quasi-separated) if
each element of B is quasi-compact (resp. quasi-separated).
We say that C is locally concentrated if C has a quasi-compact quasiseparated base of topology.
(4.5) Let C be a site with a basis of topology B. Let D be a full subcategory of C such that Ob(D) consists of quasi-compact quasi-separated objects.
Assume that Ob(D) ⊃ B.
We say that a presheaf of abelian groups M on D is a D-sheaf, if for any
U ∈ D, for any finite covering (Vi → U) of U with Vi ∈ B, for any coverings
(Vijk → Vi ×U Vj ) of Vi ×U Vj with Vijk ∈ B, the sequence
Y
YY
Γ(Vijk , M)
0 → Γ(U, M) →
Γ(Vi , M) →
i
i,j
k
is exact. Let Ps(D) and Ps(C) be the category of presheaves of abelian groups,
and let Sh(C) be the category of sheaves of abelian groups. Let Sh(D) be the
category of D-sheaves.
Let ι : D → C be the inclusion. Then there is an obvious commutative
diagram
qC
/ Ps(C) ,
Sh(C)
ι#
Ps
ι#
Sh
Sh(D)
qD
/ Ps(D)
where q are the inclusions.
Lemma 4.6. Let the notation be as above. Then there is a functor γ :
Ps(D) → Sh(C) such that the following are satisfied.
11
1 γι# : Ps(C) → Sh(C) is a sheafification, that is, left adjoint to q.
2 ι# γ : Ps(D) → Sh(D) is a sheafification, that is, left adjoint to q.
Moreover, we have
3 ι#
Sh : Sh(C) → Sh(D) is an equivalence with γq its quasi-inverse.
4 If L ∈ Ps(C) and ι#
Ps (L) is a D-sheaf, then the canonical map Γ(U, L) →
a
Γ(U, L) is an isomorphism for U ∈ D.
Proof. Let M ∈ Ps(D).
0
For each object W of C, define Ȟ (W, M) to be the inductive limit of
Y
Y Y
Ker( Γ(Vi , M) →
Γ(Vijk , M)),
ij k∈Jij
i∈I
where (Vi → W )i∈I runs through the coverings of W with Vi ∈ B (the index
set I may be infinite), and (Vijk → Vi ×W Vj ) runs through the coverings
0
0
of Vi ×W Vj with Vijk ∈ B. It is easy to see that Ȟ (M) = Ȟ (?, M) is a
0
separated presheaf of abelian groups on C, and Ȟ (M)|D is M. Now as can
0
0
be seen easily, N = Ȟ (Ȟ (M)) is a sheaf of abelian groups, and N |D = M.
0
0
This defines a functor γ := Ȟ Ȟ : Ps(D) → Sh(C). By construction, 1 and
2 are satisfied.
The other assertions are proved easily.
Lemma 4.7 (cf. [Kem, Theorem 8]). Let C be a locally concentrated site, and
U a quasi-compact quasi-separated object of C. Let Fα be a filtered inductive
system of abelian sheaves on C. Then the canonical map
lim H i (U, Fα) → H i (U, lim Fα )
−→
−→
is an isomorphism.
Proof. First consider the case that i = 0. Let L be the inductive limit of
(Fα ) in the category of presheaves so that the left hand side is L(U).
Let V = (Vi → U) be any finite covering with Vi quasi-compact quasiseparated, and (Vijk → Vi ×U Vj ) be finite coverings with Vijk quasi-compact
quasi-separated. Now
Y
Y
Fα (Vijk )
0 → Fα (U) →
Fα (Vi ) →
i
i,j,k
12
is exact, since each Fα is a sheaf. Taking the inductive limit,
Y
Y
0 → L(U) →
L(Vi ) →
L(Vijk )
i
i,j,k
is also exact, since the inductive limit is compatible with any finite product. Applying Lemma 4.6, 4 for the full subcategory D of C consisting of
all the quasi-compact quasi-separated objects, we have the case for i = 0
immediately, as the sheafification a L of L is lim Fα .
−→
Now consider the general case. Let Iα• be the injective resolution of Fα .
It suffices to show that lim Iα• is an Γ(U, ?)-acyclic resolution of lim Fα . To
−→
−→
show this, it suffices to show that if (Iα ) is an inductive system of abelian
sheaves which are Γ(U, ?)-acyclic for any quasi-compact quasi-separated U,
then H i (U, lim Iα ) = 0 for any quasi-compact quasi-separated U and i > 0.
−→
This is checked as in [Mil, (2.12)], and is left to the reader.
Corollary 4.8. Let (C, O) be a locally concentrated ringed site. Then for any
quasi-coherent sheaf M over C and c ∈ C, there exists some covering (cλ → c)
such that for each λ, M|cλ is isomorphic to M̃λ for some Γ(cλ , O)-module
Mλ .
Proof. There is a covering (cλ → c) such that for each λ, there exists some
exact sequence of the form
d
λ
Fλ,0 → M|cλ → 0
Fλ,1 −→
with Fλ,1 and Fλ,0 free. We may assume that each cλ is quasi-compact quasiseparated.
Let Mλ be the cokernel of the map
d
λ
Γ(cλ , Fλ,0 ) =: F0 .
F1 := Γ(cλ , Fλ,1 ) −→
Then there is a commutative diagram
F̃1
dλ
/ F̃0
h1
Fλ,1
/
M̃
h0
dλ
/
/
0,
h
Fλ,0
/
M|cλ
/
0
where h1 , h0 and h are canonical maps. Thus it suffices to show that h1 and
h0 are isomorphisms. This is proved easily, using Lemma 4.7.
13
5.
Grothendieck’s descent
(5.1) Let C be a class of morphisms of schemes. We assume that
1 All isomorphisms are in C.
2 If f : X → Y and g : Y → Z are morphisms and g ∈ C, then f ∈ C if
and only if gf ∈ C.
3 If f : X → Y is in C and Y ′ → Y any morphism of schemes, then the
base change f ′ : X ′ = Y ′ ×Y X → Y ′ of f is again in C.
Sometimes we identify C as a subcategory of Sch such that Ob(C) = Ob(Sch)
and Mor(C) = C. When we take the fiber product of two morphisms with
the same codomain in C in the category Sch, it is also a fiber product in C.
For a scheme T , the category C/T is a full subcategory of Sch/T . Further,
we assume that
4 C/T is skeletally small for any scheme T .
(5.2) Let E be a Grothendieck pretopology on Sch/S such as Zariski, étale,
fppf, and fpqc. An E-morphism is a morphism in Sch/S which appear as
a morphism of some covering (such as an open immersion, étale map, flat
map locally of finite presentation, and flat map). We assume that an Emorphism is in C. We also assume that the presheaf O of rings defined by
O(X) = Γ(X, OX ) is a sheaf with respect to E.
(5.3) Let I be a small category, and X• an I op -diagram of S-schemes.
We define (C/X• )E to be the ringed site defined as follows. An object
of (C/X• )E is a pair (i, f : U → Xi ) such that i ∈ Ob(I) and f ∈ C. A
morphism from (j, f : U → Xj ) to (i, g : V → Xi ) is a pair (φ, h) such that
φ : i → j is a morphism in I, and h : U → V is a morphism such that
gh = Xφ f . The composition is given by (φ′ , h′ ) ◦ (φ, h) = (φφ′ , h′ h). Thus
(C/X• )E is a skeletally small category. A family ((φλ , hλ ) : (iλ , Uλ ) → (i, U))
with the same codomain (i, U) is a covering if iλ = i, φλ = idλ , and the
collection (hλ : Uλ → U) is a covering of U in E. The structure sheaf O is
defined in an appropriate way.
14
(5.4) The first example is the case that C = (Zar) is the class of open immersions and E = Zar is the Zariski topology. The resulting site ((Zar)/X• )Zar
is called the small Zariski site and is denoted by Zar(X• ). We denote
Mod(Zar(X• )), Qch(Zar(X• )), and Inv(Zar(X• )) by Mod(X• ), Qch(X• ), and
Inv((X• ), respectively.
(5.5) Let κ be a regular cardinal. Let C = (κ- Mor) be the class of all
κ-morphisms. The resulting site ((κ- Mor)/X• )Zar is called the κ-big Zariski
site, and is denoted by Zarb (X• ) = Zarbκ (X• ).
We denote Mod(Zarbκ (X• )), Qch(Zarbκ (X• )), and Inv(Zarbκ (X• )) by Modbκ X• ),
Qchbκ (X• ), and Invbκ (X• ), respectively. The suffix κ will be omitted if there
is no danger of confusion.
(5.6) There is an obvious inclusion µ : Zar(X• ) → Zar(X• )bκ . It gives
a morphism of ringed sites µ : Zar(X• )bκ → Zar(X• ). It is easy to see
that for each M ∈ Mod(Zar(X• )) and a κ-morphism (i, f : U → Xi ),
Γ((i, f ), µ∗(M)) = Γ(U, f ∗(Mi )).
(5.7) Let E = Fpqc (resp. Qfpqc) be the pretopology whose covering
` (Vλ →
U) is a family of S-morphisms such that the induced morphism λ Vλ → U
is fpqc (resp. qfpqc). Almost by definition, Qfpqc is a saturation of Fpqc. In
what follows, we do not treat Qfpqc, but implicitly use the fact that a sheaf
for Fpqc is also a sheaf for Qfpqc [Vis, (2.48)].
We denote the site ((κ- Mor)/X• )Fpqc by Fpqc(X• ) = Fpqcκ (X• ). By
[Vis, (2.60), (4.22)], the sheaf O of commutative rings is a sheaf on Fpqc(X• ),
and thus Fpqc(X• ) is a ringed site. We denote Mod(Fpqc(X• )), Qch(Fpqc(X• )),
and Inv(Fpqc(X• )) by ModFpqc (X• ), QchFpqc (X• ), and InvFpqc (X• ), respectively.
We have a morphism of ringed sites ν : Fpqc(X• ) → Zarb (X• ).
(5.8) Note that (i, f : U → Xi ) with U affine forms a quasi-compact quasiseparated basis of topology of Zar(X• ). Similarly for Zarb (X• ) and Fpqc(X• ).
Thus these sites are locally concentrated.
(5.9) Considering the case that I = 1 (the discrete category with one
object), we have Zar(X), Zarb (X), and Fpqc(X) for a single scheme X.
(5.10) Until the end of this section, let C and E be as in (5.1) and (5.2).
Let I and X• be as in (5.3).
15
Lemma 5.11. The restriction functor (?)J : Mod((C/X• )E ) → Mod((C/X• )E )
has both a left adjoint and a right adjoint. In particular, (?)J preserves arbitrary limits and colimits. In particular, (?)J is exact.
Proof. The left adjoint exists by Kan’s lemma (see e.g., [Art, Theorem I.2.1].
The existence of the right adjoint is by [Has2, (2.31)].
Lemma 5.12. Let I be a small category, and f• : X• → Y• be a morphism of
I op -diagrams of schemes. Let J be a subcategory of I. Then Lipman’s theta
[Lip, (3.7.2)], [Has2, (1.21)]
θ : (f• )∗J (?)J → (?)J f•∗
is an isomorphism of functors Mod((C/Y• )E ) → Mod((C/X• |J )E ).
Proof. This is proved similarly to [Has2, (6.25)].
Lemma 5.13. Let I be a small category, and f• : X• → Y• be a morphism of I op -diagrams of schemes. If fi ∈ C for any i ∈ I, then f•∗ :
Mod((C/Y• )E ) → Mod((C/X• )E ) is nothing but the restriction. That is,
for M ∈ Mod((C/Y• )E ) and (i, g : U → Xi ) ∈ C/X• , Γ((i, g), f•∗M) =
Γ((i, fi g), M). In particular, f•∗ is an exact functor.
Proof. By Lemma 5.12 applied to J = {i}, we may assume that I = 1,
and the problem is on a morphism f : X → Y of single schemes which lies
in C. Then the inverse image of the presheaf module (f ∗ )p M is given by
Γ(U, (f ∗)p M) = lim Γ(U, OX ) ⊗Γ(V,OY ) Γ(V, M), where the colimit is taken
−→
over pairs (V, h) with V ∈ C/Y and h ∈ Mor(U, V ). But as f ∈ C, the
colimit is taken over a category with the final object (U, 1U ), and hence
Γ(U, (f ∗)p M) = Γ(U, M). So (f ∗ )p M is already a sheaf, and the assertion
follows.
Lemma 5.14. Let I be a small category, and
X•′
g•X
X•
/
f•′
Y•′
f•
g•Y
/
Y•
be a cartesian square of I op -diagrams of S-schemes. Assume that (g•Y )i ∈ C
for any i ∈ I. Then Lipman’s theta
θ : (g•Y )∗ (f• )∗ → (f•′ )∗ (g•X )∗
between the functors Mod((C/X• )E ) → Mod((C/Y•′ )E ) is an isomorphism.
16
Proof. By [Lip, (3.7.2)] and Lemma 5.12, it suffices to prove that
θ : (giY )∗ (fi )∗ → (fi′ )∗ (giX )∗
is an isomorphism for each i ∈ I, and we may assume that I = 1. Then for
M ∈ Mod((C/X)E ) and U ∈ C/Y ′ , θ is the identity map on Γ(U ×Y X, M),
and is an isomorphism.
Lemma 5.15. Let (i, f : U → Xi ) ∈ Zar(X• ) and M a Γ((i, f ), OX• )module. Then for any (φ, h) : (j, g) → (i, f ) ∈ Zar(X• )/(i, f ) with V =
Spec A, the source of g, affine, we have that the canonical map
Γ((j, g), M̃ p ) = A ⊗Γ((i,f ),OX• ) M → Γ((j, g), M̃)
is an isomorphism. Similar results hold for the other two sites.
Proof. This follows immediately from [Vis, (4.22)] and Lemma 4.6.
Lemma 5.16. Let M ∈ Mod(X• ). Then the following are equivalent.
1 M is quasi-coherent.
2 For any (i, f : U → Xi ) ∈ Zar(X• ) (resp. Zarb (X• ), Fpqc(X• )) with
U = Spec A affine, M|(i,f ) is canonically isomorphic to M̃ , where M =
Γ((i, f ), M).
3 For any morphism (φ, h) : (j, g : V → Xj ) → (i, f : U → Xi ) in
Zar(X• ) (resp. Zarb (X• ), Fpqc(X• )) with U = Spec A and V = Spec B
affine, the canonical map B ⊗A Γ((i, f ), M) → Γ((j, g), M) is an isomorphism.
Similar results hold for the other two sites.
Proof. 1⇒2. It suffices to show that M̃ ∼
= M|(i,f ) for some A-module M. By
Corollary 4.8, for some faithfully flat algebra A′ of A, Ñ ∼
= M|(i,f h) , where
′
h : Spec A → Spec A is the canonical map, and N = Γ((i, f h), M). Now we
can find such an M by the descent theory [Vis, (4.21)].
2⇒3 follows from Lemma 5.15.
3⇒1 Let (i, f : U → Xi ) be an object of Zar(X• ) (resp. Zarb (X• ),
Fpqc(X• )) with U = Spec A affine. Then M|(i,f ) ∼
= M̃ for M = Γ((i, f ), M)
by the uniqueness assertion in Lemma 4.6. The assertion follows.
17
Lemma 5.17. Let ϕ : X → Y be an qfpqc morphism of schemes. Let M be
an OY -module. Then M is quasi-coherent (resp. an invertible sheaf) if and
only if ϕ∗ M is so.
Proof. If M is quasi-coherent or invertible, then obviously ϕ∗ M is so. In
order to prove the converse, as the question is local on Y , we may assume that
Y is affine. Then we may assume that X is affine and ϕ is flat. The assertion
follows from [Has2, (10.14)] for quasi-coherence. To prove the assertion for
invertible sheaves, it suffices to show that if A → B is a homomorphims of
commutative rings, M an A-module, and B ⊗A M is rank-one projective,
then M is rank-one projective. As B ⊗A M is finitely presented and flat, M
is so. Thus M is finite projective. It is easy to see that M has rank one.
Lemma 5.18. The functors µ∗ : Qch(X• ) → Qchb (X• ) and µ∗ : Qchb (X• ) →
Qch(X• ) are well-defined and quasi-inverse each other. Similar results hold
for the morphism of ringed sites ν : Fpqc(X• ) → Zarb (X• ).
These equivalences induce equivalences of the categories of invertible sheaves
Inv(X• ), Invb (X• ), and InvFpqc (X• ).
Proof. Well-definedness of µ∗ is (4.3). Well-definedness of µ∗ is checked by
the condition 3 of Lemma 5.16.
If (i, f : Spec A → Xi ) is an object of Zar(X• ) and M an A-module, then
∗
µ (M̃ ) ∼
= M̃ and µ∗ (M̃) ∼
= M̃ . Thus µ∗ and µ∗ are quasi-inverse each other.
Similarly for ν.
It is obvious that the inverse image preserves invertible sheaves. As
µ∗ (M̃ ) ∼
= M̃, it suffices to show that for any faithfully flat ring homomorphism A → B and an A-module M, M is rank-one projective if and only if
B ⊗A M is rank-one projective, in order to prove that µ∗ preserves invertible
sheaves. This is done in the proof of Lemma 5.17.
Similarly for ν.
Lemma 5.19. For a morphism of I op -diagrams of schemes f• : X• → Y• ,
the diagram
µ
ν /
/ Zar(X• )
Fpqc(X• )
Zarb (X• )
f•
Fpqc(Y• )
f•
f•
ν
/
Zarb (Y• )
is commutative.
18
µ
/
Zar(Y• )
Proof. Obvious.
Lemma 5.20. Let ϕ : X → Y be a qfpqc morphism of schemes. Then
β1 −β2
u
0→M−
→ ϕ∗ ϕ∗ M −−−→ ψ∗ ψ ∗ M
is exact for M ∈ Qch(Y ), where u is the unit map for the adjoint pair
(ϕ∗ , ϕ∗ ), pi : X ×Y X → X is the ith projection, ψ = ϕp1 = ϕp2 , and βi is
the composite
u
ϕ∗ ϕ∗ M −
→ ϕ∗ (pi )∗ (pi )∗ ϕ∗ M ∼
= ψ∗ ψ ∗ M.
Proof. This is equivalent to say that µ∗ (M) is already a sheaf in the fpqc
topology, which is equivalent to the qfpqc topology. But this is trivial, as
µ∗ (M) ∼
= ν∗ ν ∗ µ∗ (M).
(5.21) Let M ∈ QchFpqc (X• ), and (i, f : U → Xi ) ∈ Fpqc(X• ) with
U = Spec A affine. Then M|(i,U ) ∼
= M̃ , where M = Γ((i, U), M), by
Lemma 5.16. For any covering of the form U : (i, Spec B) → (i, Spec A)
i
and any A-module M, the Čech cohomology Ȟ (U, M̃ ) = 0 for i > 0. It
i
follows that Ȟ ((i, U), M̃ ) = 0 for i > 0. Quite similarly to the proof of [Mil,
(2.12)], we can prove that H n ((i, U), M) = 0 for U affine, n > 0, and any
quasi-coherent sheaf on Fpqc(X• ).
(5.22) An OX• -module M is said to be equivariant if the canonical map
αφ : Xφ∗ Mi → Mj is an isomorphism for any morphism φ : i → j in I,
see [Has2, (4.14)]. The full subcategory of Mod(X• ) consisting of equivariant OX• -modules is denoted by EM(X• ). Similarly, an equivariant module
on Zarb (X• ) and Fpqc(X• ) are defined, and the full subcategories of these,
denoted by EMb (X• ) and EMFpqc (X• ), respectively, are defined.
(5.23) An object M in Mod(X• ) is said to be locally quasi-coherent if the
restriction Mi is quasi-coherent for each i ∈ I.
Lemma 5.24. Let M ∈ Mod(X• ) (resp. Modb (X• ), ModFpqc (X• )). Then
M is quasi-coherent if and only if it is locally quasi-coherent and equivariant.
Proof. The case M ∈ Mod(X• ) is [Has2, (7.3)]. The other cases are proved
using the discussion of [Has2, (7.3)], Lemma 5.16, and Lemma 5.19.
19
Lemma 5.25. Assume that Xφ ∈ C for φ ∈ Mor(I). If
M1 → M2 → M3 → M4 → M5
is an exact sequence in Mod((C/X• )E ) and M1 , M2 , M4 , and M5 are
equivariant, then M3 is equivariant. In particular, EM((C/X• )E ) is closed
under extensions, kernels, and cokernels in Mod((C/X• )E ), and hence itself
is an abelian category, and the inclusion EM((C/X• )E ) ֒→ Mod((C/X• )E )
is an exact functor.
Proof. For each morphism φ : i → j of I, consider the commutative diagram
Xφ∗ (M1 )i
/ X ∗ (M2 )i
φ
αφ (M1 )
(M1 )j
/
Xφ∗ (M3 )i
αφ (M2 )
/ (M2 )j
/
Xφ∗ (M4 )i
/
αφ (M3 )
(M3 )j
/
Xφ∗ (M5 )i
/
αφ (M4 )
(M4 )j
.
/
αφ (M5 )
(M5 )j
The rows are exact by Lemma 5.11 and Lemma 5.13. By assumption, αφ (Mi )
is an isomorphism for i = 1, 2, 4, 5. By the five lemma, αφ (M3 ) is also an
isomorphism. Hence M3 is equivariant.
(5.26) Let S be a scheme, and G an S-group scheme. Fix κ sufficiently
large so that G → S is a κ-morphism. Let X be a G-scheme.
M
(5.27) We define Zar(G, X) := Zar(BG
(X)) and Zar+ (G, X) := Zar(BG (X)(∆)mon ).
See for the notation, see [Has2]. Strictly speaking, the object set of Zar(X• )
is slightly different in [Has2]. In our definition, an object of Zar(X• ) is a pair
(i, h : U → Xi ) with h an open immersion, while in [Has2], h is required
to be the inclusion map of an open subscheme. But this difference will not
cause any trouble. The category of (G, OX )-modules Mod(Zar(G, X)) is
denoted by Mod(G, X). Mod(Zar+ (G, X)) is denoted by Mod+ (G, X). Similar definitions are done in obvious ways. For example, Fpqc(G, X) means
M
Fpqc(BG
(X)), EMb,+ (G, X) means EM(Zarb,+ (G, X)) = EM(Zarb (BG (X)(∆)mon )),
M
and Invb (G, X) means Inv(Zarb (G, X)) = Inv(Zarb (BG
(X))), and so on.
A (G, OX )-module means an object of Mod(G, X). An fpqc (G, OX )module means an object of ModFpqc (G, X).
(5.28) There are canonical restriction functors
(?)(∆)mon
: Mod+ (G, X) → Mod(G, X),
[0],[1],[2]
20
(?)(∆)mon
: Mod+
Fpqc (G, X) → ModFpqc (G, X),
[0],[1],[2]
and so on. These functors induce equivalences
EM+ (G, X) → EM(G, X),
Qch+ (G, X) → Qch(G, X),
EM+
Fpqc (G, X) → EMFpqc (G, X)
and so on. Thus the six categories
Qch+ (G, X), Qch(G, X), Qchb,+ (G, X),
Qchb (G, X), Qch+
Fpqc (G, X), QchFpqc (G, X)
are equivalent and identified in a natural way.
(5.29) The category EM(G, X) and Qch(G, X) are identified with the category of G-linearized OX -modules and that of G-linearized quasi-coherent
OX -modules defined in [MFK], respectively. Letting Qch /S denote the stack
of quasi-coherent sheaves in (Zariski site) over Sch/S with the fpqc topology
(see [Vis, (4.23)]), Qch(G, X) is also equivalent to the category (Qch /S(X))G
of G-equivariant objects in Qch /S(X) ([Vis, (3.48)]).
(5.30) Let F be an S-group scheme, and X an S-scheme on which F
acts trivially. For an (F, OX )-module M, MF denotes the descended OX module (?)−1 R(∆M ) M, where we consider that M ∈ Mod(F, X) [Has2,
(30.1), (30.2)]. We call MF the F -invariance of M. A similar definition
can be done for the fpqc topology.
Let ρ : p∗ M → p∗ M be the F -linearization, where p = p2 = a : F × X →
X is the (trivial) action, which equals the second projection. Then MF
is the kernel of (1 − ρ)u : M → p∗ p∗ M, where u : M → p∗ p∗ M is the
unit of adjunction of the adjoint pair (p∗ , p∗ ). We say that M is F -trivial if
MF = M. If p is quasi-compact and M is quasi-coherent, then it is easy to
see that MF is also quasi-coherent.
Proposition 5.31. Let ϕ : X → Y be a principal G-bundle. Then the
inverse image ϕ∗ : Qch(Y ) → Qch(G, X) is an equivalence. The functor
(?)G ◦ ϕ∗ : Qch(G, X) → Qch(Y ) is its quasi-inverse. The unit of adjunction
u : Id → (?)G ϕ∗ ϕ∗ is the map induced by the unit of adjunction u′ : Id →
ϕ∗ ϕ∗ of the adjoint pair (ϕ∗ , ϕ∗ ). That is, u is the unique map such that
21
u′ = ιu, where ι : (?)G ֒→ Id is the inclusion. The counit of adjunction
ε : ϕ∗ (?)G ϕ∗ → Id is the composite
ε′
ι
ϕ∗ (?)G ϕ∗ −
→ ϕ∗ ϕ∗ −
→ Id,
where ε′ is the counit of the adjunction of (ϕ∗ , ϕ∗ ).
Proof. (ϕ∗ , (?)G ϕ∗ ) is an adjoint pair. Indeed, the composite
ϕ∗ u
εϕ∗
ϕ∗ −−→ ϕ∗ (?)G ϕ∗ ϕ∗ −−→ ϕ∗
is the identity almost by definition and the equality (ε′ ϕ∗ ) ◦ (ϕ∗ u′ ) = id.
Similarly, the composite
(?)G ϕ∗ ε
u(?)G ϕ∗
(?)G ϕ∗ −−−−→ (?)G ϕ∗ ϕ∗ (?)G ϕ∗ −−−−→ (?)G ϕ∗
is the identity, since ι is a monomorphism, and the diagram
(?)G ϕ∗
u
(?)G ϕ∗ ϕ∗ (?)G ϕ∗
ε
/ (?)G ϕ∗
7
♦
♦
ε♦♦♦♦
ι
ι
ι
♦♦
♦♦♦
u
/ (?)G ϕ∗ ϕ∗ ϕ∗
ϕ∗ PPP
7 ϕ∗
PPP
♥♥♥ O
♥
♥
PPuP
ε♥♥♥
ι
PPP
♥♥♥
PP'
♥
♥
♥
/
ϕ∗ ϕ∗ ϕ∗
@A
id
BC
is commutative. As (ϕ∗ , (?)G ϕ∗ ) is an adjoint pair and ϕ∗ is an equivalence
by [Vis, (4.23), (4.46)], we are done.
Similarly, considering the stack of the all modules ModFpqc,κ /S (it is a
stack, as in [Vis, (4.11)]) over (Sch/S)κ with the fpqc topology, we have a
similar result.
Lemma 5.32. Let ϕ : X → Y be a principal G-bundle. Then the inverse
image ϕ∗ : ModFpqc (Y ) → EMFpqc (G, X) is an equivalence. The functor
(?)G ◦ ϕ∗ is its quasi-inverse. The unit of adjunction u : Id → (?)G ϕ∗ ϕ∗
is the map induced by the usual map Id → ϕ∗ ϕ∗ . The counit of adjunction
ε : ϕ∗ (?)G ϕ∗ → Id is the composite
ϕ∗ (?)G ϕ∗ → ϕ∗ ϕ∗ → Id.
Under this equivalence, QchFpqc (Y ) corresponds to QchFpqc (G, X).
22
Proof. We only need to prove the last assertion. By definition, the quasicoherence is local. So ϕ∗ N is locally quasi-coherent for N ∈ QchFpqc (Y ).
The equivariance of ϕ∗ N is trivial, so ϕ∗ N ∈ QchFpqc (G, X). Conversely, if
ϕ∗ N is quasi-coherent, then N is quasi-coherent simply by the local nature
of quasi-coherence.
6.
Enriched Grothendieck’s descent
(6.1) Let S be a scheme, and G an S-group scheme.
Lemma 6.2. A principal G-bundle ϕ : X → Y is a universal geometric
quotient in the sense of [MFK]. In particular, it is a categorical quotient,
and hence is uniquely determined only by the G-scheme X.
Proof. As a base change of a principal G-bundle is again a principal Gbundle, it is enough to show that a principal G-bundle is a geometric quotient.
Let ϕ : X → Y be a principal G-bundle. By Proposition 5.31, u :
OY → (ϕ∗ OX )G is an isomorphism. By Lemma 2.3, 4, ϕ is submersive. By
assumption, ϕ is G-invariant. The map Φ : G × X → X ×Y X is surjective,
since it is an isomorphism by Lemma 2.8.
(6.3) From now on, until the end of this paper, let f : G → H be a qfpqc
homomorphism of S-group schemes, and N = Ker f . Note that N is a normal
subgroup scheme of G.
(6.4) Note that Φ : N × G → G ×H G given by Φ(n, g) = (ng, g) is an
isomorphism, since Ψ : G ×H G → N × G given by Ψ(g, g1) = (gg1−1, g1 ) is
its inverse. So f is a G-enriched principal N-bundle by Lemma 2.8.
Lemma 6.5. Let X be a G-scheme on which N acts trivially. Then there is
a unique action a′X : H × X → X such that the diagram
(1)
aX
G×X
X
/
f ×1X
H ×X
1X
a′X
/
X
is commutative, where aX : G × X → X is the given action.
Proof. Let X ′ be the S-scheme X with the trivial G-action. Then being a
base change of f , f × 1X : G × X ′ → H × X ′ is a geometric quotient under
23
the action of N by Lemma 6.2. In particular, it is a categorical quotient
under the action of N by [MFK, Proposition 0.1]. As aX : G × X ′ → X is an
N-invariant morphism, there is a unique morphism a′X : H × X → X such
that the diagram (1) is commutative.
We compare the two maps a′X ◦ (µH × 1X ) and a′X ◦ (1H × a′X ) from
H × H × X to X. When we compose f × f × 1X : G × G × X → H × H × X
to the right, they agree by the commutativity of (1) and the facts that f is
a homomorphism, and that aX is an action. By Lemma 2.9, the two maps
agree. It is easy to see that
′
aX
uH ×1X
X∼
= S × X −−−−→ H × X −→ X
is the identity, where uH : S → H is the unit element. Thus a′X is an action
of H on X.
(6.6) Conversely, if a′X : H × X → X is a given H-action on X, then
defining aX = a′X (f × 1X ) (that is, in a unique way so that the diagram (1)
is commutative), aX : G × X → X is an action, and N acts trivially on X.
From now on, we identify an H-scheme and a G-scheme on which N acts
trivially.
Lemma 6.7. Let ϕ : X → Y be a G-morphism which is also a geometric
quotient under the action of N. Let U be a G-stable open subset of X. Then
1 U = ϕ−1 (ϕ(U)).
2 ϕ(U) is a G-stable open subset of Y .
3 ϕ|U : U → ϕ(U) is a base change of ϕ by a G-stable open immersion.
In particular, it is a G-morphism which is a geometric quotient under
the action of N. If, moreover, ϕ is a principal N-bundle (resp. trivial
N-bundle), then so is ϕ|U : U → ϕ(U).
Proof. Set V := ϕ(U). 1 We may assume that G = N. If U 6= ϕ−1 (ϕ(U)),
then there is a geometric point ξ → S such that Φ : G(ξ)×X(ξ) → X(ξ)×Y (ξ)
X(ξ) is surjective, and U(ξ) 6= ϕ−1 (ϕ(U(ξ))). Then there exists some x ∈
U(ξ) and y ∈ ϕ−1 (ϕ(U(ξ))) \ U(ξ) such that ϕ(x) = ϕ(y). There exists
some g ∈ G(ξ) such that Φ(g, x) = (gx, x) = (y, x). This contradicts the
assumption that U is G-stable.
24
2 As ϕ is submersive and ϕ−1 (V ) = U is open, V is open. As ϕ ◦ aX =
aY ◦ (1G × ϕ) as morphisms from G × X to Y ,
G · V = aY (G × V ) = aY (1G × ϕ)(G × U) = ϕaX (G × U) = ϕ(U) = V.
This shows that V is G-stable.
3 follows from 1 and 2.
(6.8) Until the end of this section, we fix a regular cardinal κ sufficiently
large so that the structure morphism G → S is a κ-morphism (such a κ
exists).
Let X be an H-scheme. We want to show that the category Qch(H, X)
of quasi-coherent (H, OX )-modules and the category QchN (G, X) of quasicoherent (G, OX )-modules on which N acts trivially are equivalent. Similarly,
we want to show that EMFpqc (H, X) is equivalent to the category of N-trivial
equivariant (G, OX )-modules EMN,Fpqc (G, X).
For the purpose above, we need the notion of G-linearized OX -modules
by Mumford [MFK] (in the fpqc topology).
Let F be an S-group scheme, and Z be an F -scheme. An F -linearized OZ module is a pair (M, ρ) such that M ∈ ModFpqc (Z), and ρ : a∗ M → p∗2 M
is an isomorphism such that the diagram
(µ × 1)∗ a∗ M
ρ
(µ × 1)∗ p∗2 M
/
∼
=
/
p∗23 pO ∗2 M
∼
=
ρ
(1 × a)∗ a∗ M
ρ
/
(1 × a)∗ p∗2 M
∼
=
/ p∗
∗
23 a M
is commutative, where µ : F × F → F is the product, a : F × Z → Z is the
action, p2 : F ×Z → Z is the second projection, and p23 : F ×F ×Z → F ×Z
is given by p23 (g, g ′, z) = (g ′ , z). When we need not mention ρ, sometimes
we say that M is an F -linearized OZ -module. ρ is called the F -linearization
of M. Let M and N be F -linearized OZ -modules. We say that φ : M → N
is a map of F -linearized OZ -modules if φ is OZ -linear, and the diagram
a∗ M
a∗ φ
a∗ N
/
ρ
p∗2 M
ρ
p∗2 φ
25
/
p∗2 N
is commutative. The category LMFpqc (F, Z) of F -linearized OZ -modules (in
the fpqc topology) is obtained. For M ∈ EMFpqc (F, Z), when we define
ρ : a∗ M0 → p∗2 M0 to be the composite
α−1
δ
αδ0
1
p∗2 M0 ,
a M0 −−→ M1 −−→
∗
then (M0 , ρ) is an F -linearized OZ -module, and we get an equivalence EMFpqc (F, Z) →
LMFpqc (F, Z). We identify an equivariant (F, OZ )-module and an F -linearized
OZ -module.
We say that (M, ρ) ∈ LMFpqc (F, Z) is quasi-coherent if M is so. The
full subcategory of LMFpqc (F, Z) consisting of quasi-coherent F -linearized
OZ -modules is denoted by LQFpqc (F, Z). The equivalence EMFpqc (F, Z) →
LMFpqc (F, Z) induces an equivalence QchFpqc (F, Z) → LQFpqc (F, Z).
(6.9) Let F be as above, and let ϕ : X → Y be an F -morphism. For
M ∈ LMFpqc (F, Y ), M ∈ EMFpqc (F, Y ). So ϕ∗ M ∈ EMFpqc (F, X). So
ϕ∗ M has a structure of F -linearized OX -module. It is easy to check that the
F -linearization is given by the composite
ρ
→ (1G × ϕ)∗ (p2 )∗Y M ∼
a∗X ϕ∗ M ∼
= (p2 )∗X ϕ∗ M.
= (1G × ϕ)∗ a∗Y M −
(6.10) Assume that F → S is a κ-morphism. For N ∈ EMFpqc (F, X), we
have ϕ∗ N ∈ EMFpqc (F, Y ) (this is proved similarly to [Has2, (7.14)]). So
ϕ∗ N has a structure of an F -linearized OY -module. The F -linearization is
given by
θ −1
ρ
θ
a∗Y ϕ∗ N −
→ (1G × ϕ)∗ a∗X N −
→ (1G × ϕ)∗ (p2 )∗X N −−→ (p2 )∗Y ϕ∗ N ,
where θ is Lipman’s theta [Has2, (1.21)], which is an isomorphism (this is
true because we are considering the fpqc site, see Lemma 5.14).
(6.11) Let h : F → F ′ be a homomorphism of S-group schemes. Let Z be
an F ′ -scheme. Then Z is an F -scheme in an obvious way, and we get a map
BhM (Z) : BFM (Z) → BFM′ (Z) defined as
BFM (Z) :
/
F ×F ×Z
/
/
F ×Z
h×h×1Z
BFM′ (Z) :
/
/Z
1Z
h×1Z
F′ × F′ × Z
26
/
/
/
F′ × Z
/
/Z
.
For the definition of BFM (Z), see [Has2, Chapter 29]. This induces the
pull-back BhM (Z)∗ : EM(F ′ , Z) → EM(F, Z) by [Has2, (7.22)]. We denote
′
′
this functor by resFF . Similarly, resFF,Fpqc = BhM (Z)∗ : EMFpqc (F ′, Z) →
EMFpqc (F, Z) is defined.
′
′
(6.12) Corresponding to resFF , we have a functor rFF : LM(F ′ , Z) →
′
LM(F, Z). rFF (M, ρ) = (M, r(ρ)), where r(ρ) is the composite
∗
(h×1X ) ρ
a∗F M ∼
= (p2 )∗F M.
= (h × 1X )∗ a∗F ′ M −−−−−→ (h × 1X )∗ (p2 )∗F ′ M ∼
′
F
Similarly, rF,Fpqc
: LMFpqc (F ′ , Z) → LMFpqc (F, Z) is obtained.
(6.13) Let F be an S-group scheme, and X an F -scheme on which F -acts
trivially. It is easy to see that M ∈ Mod(G, X) (resp. M ∈ ModFpqc (G, X))
{e}
is F -trivial if and only if M ∼
= resF M0 for some M0 ∈ Mod(X) =
Mod({e}, X) (resp. M0 ∈ ModFpqc (X) = ModFpqc ({e}, X)), where {e} is
the trivial group scheme.
(6.14) Let X be an H-scheme. By restriction, M = resH
G M is an equivariant (G, OX )-module for an equivariant (H, OX )-module M. Note that
{e}
H
H
H
∼
∼
Let EMN (G, X)
resG
N resG M = resN M = resN res{e} M is N-trivial.
(resp. QchN (G, X)) denote the category of N-trivial equivariant (resp. quasicoherent) (G, OX )-modules. Then we have a faithful exact functor resH
G :
EM(H, X) → EMN (G, X). The category of N-trivial G-linearized OX modules (resp. quasi-coherent OX -modules) is denoted by LMN (G, X) (resp.
LQN (G, X)). Thus EMN (G, X) is equivalent to LMN (G, X) and QchN (G, X)
is equivalent to LQN (G, X). Similar definitions can be done for the fpqc site.
(6.15) Let X be an H-scheme. Let (M, ρ) ∈ LMN (G, X). Thus M is
an OX -module, and ρ : a∗G M → (p2 )∗G M is a G-linearization, where aG :
G×X → X and (p2 )G : G×X → X are the action and the second projection,
respectively.
Note that N acts on G × X by α := (µ × 1X ) ◦ (ι × 1 × 1) : N × G × X →
G × X, where ι : N → G is the inclusion. That is, n(g, x) = (ng, x). We
define q : N × G × X → G × X by q(n, g, x) = (g, x). Note also that f × 1X :
G × X → H × X is a principal N-bundle. Thus a∗G M ∼
= (f × 1X )∗ a∗H M and
(p2 )∗G M ∼
= (f × 1X )∗ (p2 )∗H M have structures of (N, OG×X )-modules. The
N-linearization η of a∗G M is given by
α∗ a∗G M ∼
= q ∗ a∗G M,
= P ∗M ∼
27
where P (n, g, x) = gx. The N-linearization ζ of (p2 )∗G M is given similarly.
Lemma 6.16. Let the notation be as above. Then ρ : a∗ M → p∗2 M is
a map of N-linearized OG×X -modules, where a = aG and p2 = (p2 )G . A
similar result holds for the fpqc site.
Proof. Note that in the category of OG×G×X -modules Mod(G × G × X),
(µ × 1)∗ a∗ M
ρ
(µ × 1)∗ p∗2 M
/
∼
=
/
p∗23 p∗2 M
O
∼
=
ρ
(1 × a)∗ a∗ M
ρ
/
(1 × a)∗ p∗2 M
∼
=
/ p∗
∗
23 a M
is commutative. Pulling back this diagram by ι × 1G × 1X : N × G × X →
G × G × X, we get a commutative diagram
28
GF
α∗ a∗ M
ρ
/
α∗ p∗2 M
∼
=
(ι × 1 × 1)∗ (µ × 1)∗ a∗ M
ζ
/
q ∗ p∗2 M o
∼
=
ρ
(ι × 1 × 1)∗ (µ × 1)∗ p∗2 M
/
∼
=
∼
=
/
(ι × 1 × 1)∗ p∗23 p∗2 M
O
∼
=
(ι × 1 × 1)∗ (1 × a)∗ a∗ M
ρ
ρ
/
(ι × 1 × 1)∗ (1 × a)∗ p∗2 M
(1 × a)
29
@A
∗
(p2 )∗N M
/
id
/
ζ
(1 × a)
∗
(p2 )∗N M
ρ
(ι × 1 × 1)∗ p∗23 a∗ M
∼
=
∼
=
∼
=
.
ED
∼
=
∗ ∗
q aO M
BC
BC
This shows that ρ : a∗ M → p∗2 M is a map of N-linearized OG×X -modules.
H
Lemma 6.17. resH
G : Qch(H, X) → QchN (G, X) and resG,Fpqc : EMFpqc (H, X) →
EMN,Fpqc (G, X) are equivalences.
Proof. The proof is essentially the same, and we prove the assertion only for
the Zariski site.
H
It suffices to show that rG
: LQ(H, X) → LQN (G, X) is an equivalence.
Let (M, ρ) ∈ LQN (G, X). By Lemma 6.16 and Proposition 5.31 (for the
fpqc topology, use Lemma 5.32), there exists a unique ρ̄ : (aH )∗ M → (p2 )∗H M
such that the composite
ρ̄
a∗ M ∼
→ (f × 1)∗ (p2 )∗H M ∼
= (f × 1)∗ (aH )∗ M −
= p∗2 M
is ρ. By descent, it is easy to verify that ρ̄ is an H-linearization of M. As
(f ×1)∗ is a faithful functor, we have that Q : LQN (G, X) → LQ(H, X) given
by Q(M, ρ) = (M, ρ̄) is a functor. It is easy to see that Q is a quasi-inverse
H
of rG
.
Lemma 6.18. Let F be an S-group scheme, and M its normal subgroup
scheme. Let X be an F -scheme on which M acts trivially. Let F → S be a
κ-morphism. Let M ∈ ModFpqc (F, X). Then MM is the largest M-trivial
(F, OX )-submodule of M. (?)M : LMFpqc (F, X) → LMM,Fpqc (F, X) is a left
exact functor.
Proof. By Lemma 5.13, p∗2 , a∗ : ModFpqc (X) → ModFpqc (F × X) are exact
functors, and hence a∗ MM is a submodule of a∗ M, and p∗2 MM is a submodule of pM
2 (this discussion without the flatness of p2 is possible because we
are working on the fppf site).
We first prove that the F -linearization ρ : a∗ M → p∗2 M maps a∗ MM
into p∗2 MM , where a : F × X → X and p2 : F × X → X are the action and
the second projection, respectively.
It suffices to show that the composite
ρ
u
1−ρ
a∗ MM → a∗ M −
→ p∗2 M −
→ p∗2 p∗ p∗ M −−→ p∗2 p∗ p∗ M
is zero, where p : M × X → X is the second projection, and u is the unit of
30
adjunction. The diagram
u
p∗2 M
/
1−ρ
p∗2 p∗ p∗ M
u
/
p∗2 p∗ p∗ M
θ
θ
d
(µ × 1)∗ (µ × 1)∗ p∗2 M
/
(µ × 1)∗ p∗23 p∗ M
1−ρ
/
(µ × 1)∗ p∗23 p∗ M
is commutative by the naturality of θ and [Has2, (1.26)], where p : M × X →
X is the trivial action, µ : F × M → F is the product, p23 : F × M × X →
M × X is the projection, θ is Lipman’s theta [Has2, (1.21)] with respect to
the cartesian square
F ×M ×X
p23
/
M ×X ,
p
µ×1
p2
F ×X
/
X
and d is the canonical isomorphism [Has2, Chpaters 1, 2]. As θ : p∗2 p∗ p∗ M →
(µ × 1)∗ p∗23 p∗ M is an isomorphism by Lemma 5.14 (this is also because we
treat the fpqc site), through the adjoint isomorphism
HomOF ×X (a∗ MM , (µ × 1)∗ p∗23 p∗ M) ∼
= HomOF ×N×X ((µ × 1)∗ a∗ MN , p∗23 p∗ M),
it suffices to show that the composite
ρ
ρ−1 −1
d
(µ × 1)∗ a∗ MM → (µ × 1)∗ a∗ M −
→ (µ × 1)∗ p∗2 M −
→ p∗23 p∗ M −−−→ p∗23 p∗ M
is zero.
By the definition of the F -linearization, we have a commutative diagram
of OM ×G×X -modules
(2)
(µ′ × 1)∗ a∗ MM
d
/
(1 × a)∗ p∗ MM
ρ=1
/
(1 × a)∗ p∗ MM
ρ
(µ′ × 1)∗ p∗2 M
d
d
/
p∗23 p∗2 M o
ρ
p∗23 a∗ MM
is commutative, where µ′ : M × F → F is the product. Let Θ : F × M ×
X → M × F × X be the isomorphism given by Θ(f, m, x) = (f mf −1 , f, x).
31
Applying Θ∗ to (2), we get the commutative diagram
ρ
(µ × 1)∗ a∗ MM
/
(µ × 1)∗ p∗2 M
d
d
Θ∗ (µ′ × 1)∗ a∗ MM
ρ
Θ∗ (µ′ × 1)∗ p∗2 M
/
d
d
Θ∗ (1 × a)∗ p∗ MM
Θ∗ p∗23 p∗2 M
O
ρ
ρ=1
Θ∗ (1 × a)∗ p∗ MM
d
/
Θ∗ p∗23 a∗ M
because (µ′ × 1)Θ = µ × 1. So the diagram
(3)
(µ × 1)∗ a∗ MM
ρ
/
(µ × 1)∗ p∗2 M
d
d
(1F × p)∗ a∗ MM
ρ
/
(1F × p)∗ p∗2 M
is commutative (note that p23 Θ = 1F × p). On the other hand, by the
definition of F -linearization, the diagram
(4)
(µ × 1)∗ a∗ MM
ρ
/ (µ × 1)∗ p∗ M
d
2
p∗23 p∗ M
/
ρ−1
d
(1 × p)∗ a∗ MM
ρ
/
d
(1 × p)∗ p∗2 M
/
p∗23 p∗ M
is commutative. Combining the commutativity of (3) and (4), we have that
the composite
ρ
ρ−1 −1
d
(µ × 1)∗ a∗ MM −
→ (µ × 1)∗ p∗2 M −
→ p∗23 p∗ M −−−→ p∗23 p∗ M
is zero, as desired.
Next, we want to prove that ρ−1 : p∗2 M → a∗ M maps p∗2 MM into a∗ MM .
This is proved similarly, using the commutativity of
(µ × 1)∗ a∗ M o
ρ−1
(µ × 1)∗ p∗2 MM
d
(1F × p)∗ a∗ M o
d
ρ−1
32
(1F × p)∗ p∗2 MM
and the commutativity of
Θ∗ (µ′ × 1)∗ a∗ M o
ρ−1
O
Θ∗ (µ′ × 1)∗ p∗2 M .
d
d
∗ ∗
Θ∗ (1 × a)
p M
O
Θ∗ p∗23 p∗2 MN
ρ−1
ρ−1
Θ∗ (1 × a)∗ p∗ M o
d
Θ∗ p∗23 a∗ M
Thus ρ : a∗ MM → p∗2 MM is an isomorphism, and thus MM is an F linearized OX -submodule of M. It is M-trivial almost by the definition of
MM . The other assertions are easy.
Lemma 6.19. Let I be a small category, and f• : X• → Y• be a morphism
of I op -diagrams of S-schemes. Let jX• = µν : Fpqc(X• ) → Zar(X• ) be the
canonical morphism of ringed sites. Define jY• in a similar way. Then
∗
∗
is an isomorphism for any sub(?)J → (?)J jX
1 Lipman’s theta θ : jX
•
•
category J of I.
2 The conjugate ξ : (jX• )∗ RJ → RJ (jX• |J )∗ is also an isomorphism.
3 The Lipman’s theta θ : f•∗ (jY• )∗ M → (jX• )∗ f•∗ M is an isomorphism
for locally quasi-coherent M.
Proof. 1 is proved similarly to [Has2, (6.25)]. 2 follows from 1. To prove
3, we may assume that f : X → Y is a map of single schemes. It is easy
to check the assertion for the case that f is an open immersion. So we may
assume that both X = Spec B and Y = Spec A are affine, and M = M̃ .
Then it is easy to see that θ is the identity map (B ⊗A M)˜→ (B ⊗A M)˜.
Theorem 6.20. Let ϕ : X → Y be a G-enriched principal N-bundle.
Then (?)N ϕ∗ : EMFpqc (G, X) → EMFpqc (H, Y ) is an equivalence, and ϕ∗ :
EMFpqc (H, Y ) → EMFpqc (G, X) is its quasi-inverse. Under the equivalence,
QchFpqc (H, Y ) corresponds to QchFpqc (G, X). The unit of adjunction u :
Id → (?)N ϕ∗ ϕ∗ and the counit of adjunction ϕ∗ (?)N ϕ∗ → Id are given as in
Proposition 5.31.
33
Proof. ϕ∗ is a functor from EM(G, X) to EM(G, Y ) (again, we use the commutativity of Lipman’s theta. This is known to be true for quasi-coherent
sheaves over small Zariski site only for the case that G is flat and f is quasicompact quasi-separated, see [Has2, (7.12), (7.14)]). On the other hand, (?)N
is a functor from EM(G, Y ) to EMN (G, Y ) ∼
= EM(H, Y ) by Lemma 6.18 and
N
Lemma 6.17. So (?) ϕ∗ is a functor from EM(G, X) to EM(H, Y ). Conversely,
ϕ∗
res
EM(H, Y ) −→ EMN (G, Y ) −→ EM(G, X)
is a functor.
These two functors are quasi-inverse each other, since u : Id → (?)N ϕ∗ ϕ∗
and ε : ϕ∗ (?)N ϕ∗ → Id given in Lemma 5.32 (the group N here is G there)
are (G, OX )-linear.
The last assertion follows from the last assertion of Lemma 5.32.
Corollary 6.21. Under the same assumption as in the theorem, (?)N ϕ∗ :
Qch(G, X) → Qch(H, Y ) and ϕ∗ : Qch(H, Y ) → Qch(G, X) are quasiinverse each other. The unit map and the counit map are given as in Proposition 5.31.
Proof. It is clear that ϕ∗ : Qch(H, Y ) → Qch(G, X) is a functor. Consider
the composite of equivalences
j∗
(?)N ϕ∗
j∗
Qch(G, X) −
→ Qch(H, Y ),
→ QchFpqc (G, X) −−−−→ QchFpqc (H, Y ) −
where j = µν. As in [Has2, (30.3)], (?)N : Mod(G, Y ) → Mod(Y ) is
(?)[−1] ◦ R∆ . So by Lemma 6.19, the composite is identified with (?)N ϕ∗ j∗ j ∗ :
Qch(G, X) ∼
= Qch(H, Y ). As j∗ j ∗ : Qch(G, X) → Qch(G, X) is an autoequivalence, (?)N ϕ∗ : Qch(G, X) → Qch(H, Y ) is an equivalence. The rest is
easy.
Remark 6.22. The statement of Corollary 6.21 is independent of the choice
of κ, as it is an assertion for the Zariski topology.
7.
Equivariant Picard groups and class groups
(7.1) Let f : G → H be a qfpqc homomorphism of S-group schemes, and
N = Ker f , as above.
34
Corollary 7.2. Let ϕ : X → Y be a G-enriched principal N-bundle. Then
ϕ∗ : Inv(H, Y ) → Inv(G, X) is an equivalence, and (?)N ◦ ϕ∗ : Inv(G, X) →
Inv(H, Y ) is its quasi-inverse. Thus we have an isomorphism
ϕ∗ : [M] 7→ [ϕ∗ M]
from Pic(H, Y ) to Pic(G, X). Its inverse is given by [N ] 7→ [(ϕ∗ N )N ].
Proof. This is immediate from Corollary 6.21 and Lemma 5.17.
(7.3) Now assume that G and N are flat over S. Being a principal Nbundle, f is also flat, and hence is fpqc. So H is also flat over S.
Corollary 7.4. Let ϕ : X → Y be a G-enriched principal N-bundle such that
X is locally Krull. Then ϕ∗ : Ref(H, Y ) → Ref(G, X) is an equivalence, and
(?)N ◦ϕ∗ : Ref(G, X) → Ref(H, Y ) is its quasi-inverse. With this equivalence,
Ref n (H, Y ) corresponds to Ref n (G, X). Thus Ref 1 (H, Y ) and Ref 1 (G, X) are
equivalent, and we have an isomorphism
ϕ∗ : [M] 7→ [ϕ∗ M]
from Cl(H, Y ) to Cl(G, X). Its inverse is given by [N ] 7→ [(ϕ∗ N )N ].
Proof. Note that ϕ is fpqc by Lemma 2.12, as N → S is fpqc. Now the
assertion follows from Corollary 6.21 and [Has3, (5.32)].
(7.5) Temporarily forget our settings on G, H, and N.
Example 7.6. Let N be an S-group scheme, and H another S-group scheme
acting on N as group automorphisms. We say that X is an H-equivariant Nscheme when X is an H-scheme N-scheme such that the action aX : N ×X →
X is an H-morphism. When we set G := N ⋊H, the semidirect product, then
an H-equivariant N-scheme and a G-scheme is the same thing. We define: An
H-equivariant N-morphism is a G-morphism. An H-equivariant N-invariant
morphism is a G-morphism which is N-invariant. An H-equivariant principal
N-bundle is a G-enriched principal N-bundle.
Thus our results also apply to equivariant principal bundles.
Example 7.7. Let k be a field, and N0 a finite étale k-group scheme, and
ϕ : X → Y a principal N0 -bundle. Let k ′ be a finite Galois extension of k
such that k ′ ⊗ N0 is a constant finite group N. That is, N is a finite group
35
and k ′ ⊗ k[N0 ] ∼
= k ′ ⊗ k[N] as k ′ -Hopf algebras. We understand that N also
denotes the constant group scheme over k. So k ′ ⊗ N0 ∼
= k ′ ⊗ N. Note that
the finite group N is identified with the group of k ′ -valued points of N0 ,
N0 (k ′ ) = (Sch/k)(Spec k ′ , N0 ).
Let H be the Galois group of k ′ /k. H acts on N0 trivially, and it also acts
on k ′ ⊗k[N0 ] ∼
= k ′ ⊗k[N]. As an algebra automorphism preserves idempotents
and k[N] is the k-subalgebra generated by the idempotents of k ′ ⊗ k[N], H
acts on k[N], and so H acts on N. Thus the composite
α
→ Spec k ′ × X
N × Spec k ′ × X ∼
= N0 × Spec k ′ × X −
is an action of N on X ′ = Spec k ′ × X, and the action is H-equivariant,
where α(n0 , w, x) = (w, n0x). Now it is easy to see that the base change
ϕ′ : X ′ → Y ′ of ϕ by Spec k ′ → Spec k is an H-equivariant principal Nbundle, as a k-morphism. So it is also a G-enriched principal N-bundle,
where G = N ⋊ H. Note that the diagram of equivalences is commutative
up to natural isomorphisms
Qch(G, X ′) o
Qch(N, X) .
O
O
(ϕ′ )∗
ϕ∗
Qch(H, Y ′ ) o
Qch(Y )
Thus (ϕ′ )∗ does almost the same thing as ϕ∗ , but G and H are constant
groups, and no group scheme appears, while ϕ is a principal N0 -bundle, and
N0 need not be constant.
References
[Art]
M. Artin, Grothendieck Topology, Mimeographed notes, Harvard
University (1962).
[dJ]
J. de Jong et al, Stacks Project, pdf version, available at
http://stacks.math.columbia.edu
[End]
H. B. Enderton, Elements of set theory, Academic Press (1977).
[Gro]
A. Grothendieck, Eléments de Géométrie Algébrique IV, 1e Partie,
IHES Publ. Math. 20 (1964).
36
[Gro2]
A. Grothendieck, Eléments de Géométrie Algébrique IV, 2e Partie,
IHES Publ. Math. 24 (1965).
[Has]
M. Hashimoto, A pure subalgebra of a finitely generated algebra is
finitely generated, Proc. Amer. Math. Soc. 133 (2005), 2233–2235.
[Has2]
M. Hashimoto, Equivariant twisted inverses, Foundations of
Grothendieck Duality for Diagrams of Schemes (J. Lipman, M.
Hashimoto), Lecture Notes in Math. 1960, Springer (2009),
pp. 261–478.
[Has3]
M. Hashimoto, Equivariant class group. I. Finite generation
of the Picard and the class groups of an invariant subring,
arXiv:1309.2367v1
[Has4]
M. Hashimoto, Equivariant class group. III. Almost principal fiber
bundles, in preparation.
[HO]
M. Hashimoto and M. Ohtani, Equivariant Matlis and the local
duality, J. Algebra 324 (2010), 1447–1470.
[Kem]
G. R. Kempf, Some elementary proofs of basic theorems in the
cohomology of quasi-coherent sheaves, Rocky Mountain J. Math.
10 (1980), 637–645.
[Lip]
J. Lipman, Notes on derived functors and Grothendieck duality,
Foundations of Grothendieck Duality for Diagrams of Schemes (J.
Lipman, M. Hashimoto), Lecture Notes in Math. 1960, Springer
(2009), pp. 1–259.
[Mil]
J. S. Milne, Étale Cohomology, Princeton (1980).
[MFK]
D. Mumford, J. Fogarty and F. Kirwan, Geometric Invariant Theory, 3rd ed., Springer (1994).
[SD]
B. Saint-Donat, Techniques de descente cohomologique, Théorie
des topos et cohomologie étale des schémas. Tome 2 (SGA4),
Springer (1972), 83–162.
[Vis]
A. Vistoli, Grothendieck topologies, fibered categories and descent
theory, Fundamental Algebraic Geometry, B. Fantechi et al. (eds.),
AMS (2005), pp. 1–104.
37
| 0math.AC
|
arXiv:1509.06771v2 [math.GT] 19 Aug 2016
CHARACTERIZATION OF FINITE GROUPS GENERATED BY
REFLECTIONS AND ROTATIONS
CHRISTIAN LANGE
Abstract. We characterize finite groups G generated by orthogonal transformations in a
finite-dimensional Euclidean space V whose fixed point subspace has codimension one or
two in terms of the corresponding quotient space V /G.
1. Introduction
We call a finite group generated by reflections and rotations in a finite-dimensional Euclidean space, i.e. by orthogonal transformations with codimension one and two fixed point
subspaces, a reflection-rotation group. We call it a rotation group if it is generated by rotations. For a finite orthogonal group G acting on Rn the quotient space Rn /G inherits different
structures from Rn , e.g. a topology, a metric and a piecewise linear (PL) structure. It is easy
to see that G is a reflection-rotation group, if the quotient space Rn /G is a piecewise linear manifold with boundary and that in this case G contains a reflection if and only if the
boundary of Rn /G is nonempty (cf. Section 4). In this paper we show that the converse also
holds.
Theorem. For a finite subgroup G < On the quotient space Rn /G is a PL manifold with
boundary, if and only if G is a reflection-rotation group. In this case Rn /G is either PL
homeomorphic to the half space Rn−1 × R≥0 and G contains a reflection or Rn /G is PL
homeomorphic to Rn and G does not contain a reflection.
A partial result in this direction is contained in a paper by Mikhaîlova from 1984 [36].
In that paper the if direction of our result is verified for many rotation groups but only in
the topological category. Large classes of reflection-rotation groups are real reflection groups,
their orientation preserving subgroups and unitary reflection groups considered as real groups.
However, there are many more examples (cf. Appendix A), in particular, since a reducible
reflection-rotation group in general does not split as a product of irreducible components.
A complete classification of reflection-rotation groups has only been obtained recently [27]
based on earlier work by Mikhaîlova [33, 34, 35], Huffman and Wales [17, 18, 19, 47] and
others. It comprises both exceptional rotation groups and building blocks of infinite families
of rotation groups that are not considered in [36]. The largest irreducible rotation group
among them occurs in dimension 8. It is an extension of the alternating group on 8 letters
by a nonabelian group of order 27 and in a sense the largest exceptional irreducible rotation
group (cf. [27]). Reducible rotation groups that have not been studied in [36] occur, for
The results of this paper appear in the author’s thesis written at the University of Cologne [26]. The
author was supported by a ‘Kurzzeitstipendium für Doktoranden’ by the German Academic Exchange Service
(DAAD).
1
2
CHRISTIAN LANGE
example, in the product of two copies of a reflection group W of type H3 or H4 due to the
existence of outer automorphisms of W that map reflections onto reflections but cannot be
realized through conjugation by elements in its normalizer.
We adapt some of the methods from [36] as to also work in the piecewise linear category
and describe new methods to prove the if direction of our result by verifying its conclusion
for all reflection-rotation groups. For instance, we apply a result on equivariant smoothing
of piecewise linear manifolds [25] and the equivariant Poincaré conjecture in the case n = 4
(cf. Section 6.3) and the generalized Poincaré conjecture in the case n > 5 (cf. Section 5.3).
In this way we moreover simplify the consideration compared to [36] and make the proofs
therein rigorous (cf. Sections 6.3, 6.7). Finally, in Section 7 we suggest an approach to prove
our main result that would essentially dispense with the classification of reflection-rotation
groups.
Notice that the only-if direction of our result does not hold in the topological category.
The quotient S 3 /P of a 3-sphere by the binary icosahedral group P < SO4 is Poincaré’s
homology sphere and it follows from Cannon’s double suspension theorem that its double
suspension Σ2 (S 3 /P ) is a topological 5-sphere [6]. Therefore, the quotient space R2 × R4 /P
is homeomorphic to R6 , though P is not a rotation group. However, using this paper’s result
it can be shown that the binary icosahedral group is essentially the only counterexample to
the only-if direction in the topological category (cf. [24]).
Acknowledgements. I thank Alexander Lytchak for introducing me to the problem and
for helpful discussions. I thank the referee for carefully reading the paper and for making
valuable comments that helped to improve the exposition.
2. Notation
We denote the cyclic group of order n and the dihedral group of order 2n by Cn and Dn ,
respectively. We denote the symmetric and alternating group on n letters by Sn and An .
Classical Lie groups are denoted like SOn and On . For a finite subgroup G < On we denote its
orientation preserving subgroup as G+ < SOn . In particular, we write W + for the orientation
preserving subgroup of a reflection group W .
3. Preliminaries
3.1. PL spaces and PL structures. We remind of some concepts from piecewise linear
topology. For more details we refer to [16] and [41]. A subset P ⊂ Rn is called a polyhedron
if, for every point x ∈ P , there exists a finite number a simplices contained in P such that
their union is a neighborhood of x in P . Every polyhedron P in Rn is the underlying space of
some (locally finite Euclidean) simplicial complex K in Rn [16, Lem. 3.5]. Such a complex is
called a triangulation of P . Conversely, the underlying space of a simplicial complex K ⊂ Rn
is a polyhedron. A continuous map f : P → Q between polyhedra P ⊂ Rn , Q ⊂ Rm is called
piecewise linear (PL), if its graph {(x, f (x))|x ∈ P } ⊂ Rm+n is a polyhedron.
A PL chart (P, ϕ) for a topological space X is an embedding ϕ : P → X of a compact
polyhedron P (i.e. a Euclidean polyhedron in terms of [16]). Two PL charts (P, ϕ) and (Q, φ)
are said to be compatible if ϕ−1 (φ(Q)) is a compact polyhedron and φ−1 ◦ ϕ : ϕ−1 (φ(Q)) → Q
is piecewise linear. An atlas (base of a PL structure in terms of [16]) on X is a family of
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
3
compatible PL charts for X such that for each point x ∈ X there is a chart (P, ϕ) for which
ϕ(P ) is a topological neighborhood of x. A PL structure on X is a maximal atlas of X. A
second-countable topological Hausdorff space X endowed with a PL structure is called a PL
space (cf. [16, p. 77]). A PL space X is called a PL manifold (with boundary) of dimension
n, if for every point p ∈ X there exists a chart (P = ∆n , ϕ) of X such that p ∈ ϕ(Int(P ))
(p ∈ IntX (ϕ(P ))) (cf. [16, p. 79]). The boundary of a PL n-manifold with boundary, i.e.
the complement of its PL manifold points, is a PL (n − 1)-manifold without boundary. For
instance, the standard simplex ∆n is a PL manifold with boundary, a (standard) PL n-ball.
Its boundary ∂∆n is a PL manifold without boundary, a (standard) PL (n − 1)-sphere.
Every PL space can be triangulated by a simplicial complex K ⊂ RN and can thus be
realized as a polyhedron [16, Lem. 3.5, p. 80][37, Thm. 7.1, p. 53]. Conversely, an (abstract)
locally finite simplicial complex K has a natural PL structure. It is a PL n-manifold (with
boundary), if and only if the link of every vertex is a PL (n − 1)-sphere (or a PL (n − 1)-ball)
(cf. [41, p. 24]). PL maps and PL homeomorphisms between PL spaces are defined in the
usual way via compositions with charts (cf. [16, p. 83]). A map between two PL spaces X
and Y is a PL homeomorphism if and only if there exist triangulations K and L of X and
Y with respect to which f is a simplicial isomorphism between simplical complexes (cf. [16,
p. 84, Thm. 3.6 C]).
The open cone CX of a compact PL space X inherits a natural PL structure from X. If
X is embedded as a polyhedron P in some RN , then CX is PL homeomorphic to the internal
open cone CX = R≥0 · (P + en+1 ) ⊂ Rn+1 . Here en+1 denotes the last canonical basis vector
of Rn+1 . The cone CX is PL homeomorphic to some Rn if and only if X is PL homeomorphic
to the standard PL (n − 1)-sphere ∂∆n . To see the only-if direction of this statement one can
triangulate X by a simplicial complex K and extend this triangulation to a triangulation L of
CX such that the link of the cone point in L is K [41, proof of Prop. 2.9]. Then the statement
follows from the remark above on links in simplical complexes that are PL manifolds.
3.2. PL quotients. Let ∼ be an equivalence relation on a PL space X. We would like to
know whether X/ ∼ is a PL space such that the projection map q : X → X/ ∼ is PL.
More precisely, if there exists a PL space Y and a PL map f : X → Y that induces a
homeomorphism f : X/ ∼→ Y . In general such a pair (Y, f ) does not exist (cf. [46]).
However, if such a pair exists, then the following universal property shows that Y is unique
up to PL homeomorphism and can thus be considered the quotient of X with respect to ∼
in the PL category (cf. [46]). Note that we identify X\ ∼ and Y via f .
Lemma 3.1. If Y 0 is a PL space and f 0 : X → Y 0 a PL map such that x ∼ y for x, y ∈ X
implies f 0 (x) = f 0 (y), then the unique map g : Y → Y 0 is PL.
Proof. Let x ∈ X, y = q(x) and y 0 = f 0 (x). We choose charts (P, ϕ), (Q, φ) and (Q0 , φ0 )
about x, y and y 0 that define topological neighborhoods of the respective points. Since q is
open and since the image of a compact polyhedron under a PL map is a compact polyhedron
[41, Cor. 2.5], we can assume that qϕ(P ) = φ(Q) and P = ϕ−1 f 0−1 φ0 (Q0 ). In particular, we
have φ−1 g −1 φ0 (Q0 ) = Q and graph(φ0−1 gφ) = (φ−1 q × φ0−1 f 0 )(ϕ(P )) ⊂ Q × Q0 is a compact
polyhedron,
i.e. φ0−1 gφ : Q → Q0 is a PL map. Now the general facts that
S
S a finite union
P
of
compact
polyhedra
P
is
a
compact
polyhedron
and
that
a
map
f
:
i
i Pi → Q is PL
i i
if all restrictions f|Pi are PL [41, p. 5, 1.5 (4)], imply the condition for all charts, i.e. the map
g is PL.
4
CHRISTIAN LANGE
A Euclidean vector space Rn carries a natural PL structure with respect to which it is a PL
manifold and with respect to which On acts by PL homeomorphisms on it. In the following
section we realize Rn /G as a simplicial complex and show that the projection from Rn to
Rn /G with the induced PL structure is a PL map, i.e. that the quotient Rn /G is in a natural
way a PL space.
3.3. Admissible triangulations. Let G be a finite group. A simplicial complex K is called
a G-complex, if G acts simplicially on it. It is called a regular G-complex, if for each subgroup
H < G and each tuple of elements g0 , g1 , . . . , gn ∈ H such that both of the sets {v0 , . . . , vn }
and {g0 v0 , . . . , gn vn } describe vertices of a simplex in K, there exists an element g ∈ H such
that gvi = gi vi for all i (cf. [4, Ch. III, Def. 1.2, p. 116]). The second barycentric subdivision
of a G-complex K is always regular and for a regular G-complex one can define in a natural
way a simplicial complex K/G whose underlying space is homeomorphic to the topological
quotient |K|/G [4, p. 117]. The vertices of K/G are the G-orbits of the vertices of K and
a subset of these vertices forms a simplex if and only if there are representatives of these
vertices in K that form a simplex in K.
For a finite subgroup G < On we call a triangulation K of Rn admissible (for the action
of G on Rn ), if K is a regular G-complex that contains the origin as a vertex. The following
lemma shows that admissible triangulations always exist.
Lemma 3.2. For a finite subgroup G < On there exists a triangulation K of Rn that is
admissible for the action of G on Rn .
Proof. We start with any triangulation t : K̃ → Rn that contains the origin as a vertex and
replace it by the common subdivision t : K → Rn of the triangulations g ◦t : |K̃| → Rn , g ∈ G
as in [25, Lem. 2.1] (cf. [26, Lemma 67]). The resulting triangulation defines a G-complex.
Upon passage to the second barycentric subdivision, we can assume that this G-complex is
regular [4, p. 117].
Let K be an admissible triangulation for the action of G < On on Rn . Then Y = K/G
is a simplicial complex and the projection K → K/G maps simplices linearly onto simplices.
In particular, Y is a PL space and the projection K → K/G is PL, i.e. Rn /G is in a natural
way a PL space (cf. Section 3.2). The link of the origin in K is also a regular G-complex
and its quotient by G is simplicially isomorphic to the link of the origin in K/G. Hence, the
PL space Rn /G is a PL manifold (with boundary) if and only if this link in K/G is a PL
(n−1)-sphere (or a PL (n−1)-ball) (cf. Section 3.1). Radially projecting the link of the origin
in K to the unit sphere S n−1 defines a PL structure on S n−1 and induces a PL structure on
S n−1 /G. We call triangulations and PL structures of S n−1 and S n−1 /G that arise in this
way admissible (for the action of G on S n−1 ). Two admissible PL structures on S n−1 /G need
not be identical but are PL homeomorphic (cf. [41, pp. 20-21], “pseudoradial projection”).
Hence, the question if Rn /G is a PL manifold is equivalent to the question if S n−1 /G is a PL
sphere with respect to one and then any admissible PL structure. The following lemma gives
a necessary condition for this to hold. For details we refer to [26].
Lemma 3.3. Let G < On be a finite subgroup and suppose S n−1 is triangulated by a simplicial
complex K in an admissible way for the action of G. Then S n−1 /G is a PL manifold (with
boundary) with respect to the induced PL structure if and only if for every vertex v of K the
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
5
quotient space Tv S n−1 /Gv is a PL manifold (with boundary). In particular, this is the case
if Rn /G is a PL manifold (with boundary).
4. The only-if direction
Let us show by induction that a finite subgroup G < On is a reflection-rotation group, if
the quotient space Rn /G is a PL manifold with boundary and that in this case G contains a
reflection if and only if the boundary of Rn /G is nonempty. For n ≤ 2 the claim is trivially
true. Assume it holds for some n ≥ 2 and let G < On+1 be a finite subgroup such that Rn+1 /G
is a PL manifold with boundary. Then all isotropy groups Gv for v 6= 0 are reflection-rotation
groups by Lemma 3.3 and the induction assumption. Let Grr / G be the reflection-rotation
group generated by all of them and let v ∈ S n . Then the inclusions
Gv = (Gv )rr ⊆ (Grr )v ⊆ Gv
imply that Gv = (Grr )v for all v ∈ S n . This means that the action of G/Grr on S n /Grr is
free. Because of n ≥ 2 the quotient space S n /G is simply connected by assumption and thus
we conclude that G and Grr coincide, i.e. that G is a reflection-rotation group. Again by
the induction assumption, Grr contains a reflection if and only if the boundary of S n /Grr is
nonempty. Hence, G contains a reflection if and only if the boundary of Rn /G is nonempty
and the claim follows by induction.
5. Methods for the if direction
Before proving the if direction of our main result, we provide several methods that allow
us to identify the quotient spaces in question.
5.1. PL linearization principle. The idea of the PL linearization principle is to divide the
determination of the PL quotient Rn /G for a finite subgroup G < On into several steps. Let
H / G be a normal subgroup and assume there exists a PL homeomorphism F : Rn /H → Rn
and a homomorphism r : G → On with kernel H such that the left square in the following
diagram commutes
G × Rn /H
r×F
r(G) × Rn
/ Rn /H
F
/ Rn
/ / Rn /G
F̃
/ / Rn /r(G)
Then we say that the PL linearization principle can be applied to the groups H / G. In this
case F induces a PL homeomorphism F̃ : Rn /G → Rn /r(G) due to Lemma 3.1. This reduces
the determination of Rn /G to the determination of Rn /r(G) and one might look for a suitable
normal subgroup of r(G) in order to apply the PL linearization principle again. If the PL
linearization principle can be applied to H / G and to H̃ / r(G), then it can also directly be
applied to r−1 (H̃) / G.
Lemma 5.1. Suppose that the PL linearization principle can be applied to groups H / G. If
g ∈ G is a reflection (rotation), then so is r(g). In particular, if G is generated by reflections
and rotations, then so is r(G).
6
CHRISTIAN LANGE
The PL linearization principle can be established by describing a homeomorphism f :
S n−1 /H → S n−1 and a homomorphism r : G → On with kernel H such that the following
square commutes and such that the PL structure on S n−1 induced by an admissible PL
structure on S n−1 /H via f is admissible for the linearized action of r(G) on S n−1
G × S n−1 /H
r×f
/ S n−1 /H
f
r(G) × S n−1
/ S n−1
In fact, in this case we can take the open cone at each site and extend the PL homeomorphism
f linearly to a PL homeomorphism F : Rn /H → Rn which makes the first diagram above
commute.
5.2. Uniqueness of compatible PL structures. A map g : P → M from a polyhedron P
to a smooth manifold M is called piecewise differentiable or PD, if P admits a triangulation
such that the restriction of f to each simplex in this triangulation is smooth. It is called
PD homeomorphism, if it is in addition a homeomorphism and its restriction to each simplex
has injective differential at each point. If g is a PD homeomorphism, then, due to a theorem
by Whitehead [48], the polyhedron P is a PL manifold. Moreover, such a polyhedron exists
and is unique up to PL homeomorphisms. According to a result of Illman these statements
also hold equivariantly for a finite group acting smoothly on M [21]. In particular, if M is a
smooth manifold on which a finite group G acts smoothly and gi : Pi → M , i = 1, 2, are two
PD homeomorphisms such that the induced actions of G on the polyhedra P1 and P2 are PL,
then there exists a G-equivariant PL homeomorphism between P1 and P2 .
Assume we have a PL (n − 1)-sphere P1 on which a finite group G acts by PL homeomorphisms, a PD homeomorphism g : P1 → S n−1 and a group homomorphism r : G → On
such that the following diagram commutes
G × P1
/ P1
r×g
g
r(G) × S n−1
/ S n−1
Then, by Section 3.2 there exists an admissible triangulation of S n−1 by a polyhedron P2 with
respect to the action of r(G) and this triangulation defines a PD homeomorphism from the
polyhedron to S n−1 . Therefore P1 and P2 are G-equivariantly PL homeomorphic as explained
above. In other words, we can replace g by another G-equivariant homeomorphism f which
is in addition piecewise linear with respect to an admissible PL structure for the action of
r(G) on S n−1 .
5.3. Generalized Poincaré conjecture. The generalized Poincaré conjecture holds in the
following version. Note that a closed simply connected topological manifold M with H∗ (M ; Z) =
H∗ (S n ; Z) is homotopy equivalent to an n-sphere (cf. [42, Thm. 7.5.9, p. 399] and [14,
Thm. 4.5, p. 346]).
Theorem 5.2. For n 6= 4 a closed simply connected PL manifold M with H∗ (M ; Z) =
H∗ (S n ; Z) is PL homeomorphic to a standard PL n-sphere.
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
7
For n = 1, 2 the statement has long be known by the classification of manifolds in that
dimensions. For n ≥ 6 it follows from the PL h-cobordism theorem (c.f. [41, Thm. A, p. 17]).
For n ≤ 5 every PL manifold admits a compatible smooth structure [15] and thus, according
to the uniqueness part of Whitehead’s theorem [48, 30], the statement can be reduced to the
respective statement in the smooth category. For n = 5 the smooth version of the generalized
Poincaré conjecture follows from the smooth h-cobordism theorem combined with the fact
that every closed, smooth 5-manifold homotopy equivalent to S 5 bounds a smooth, compact,
contractible 6-manifold [22] (cf. [28] for more details). Finally, for n = 3 the smooth Poincaré
conjecture follows from Perelman’s work [38, 39, 40] (cf. [29, 23]).
We would like to apply the statement of Theorem 5.2 in the following situation. For a
rotation group G < SOn the quotient S n−1 /G is simply connected unless n ≤ 2 by the
following lemma (cf. [1], where the statement is proven in greater generality).
Lemma 5.3. Let G < SOn with n ≥ 3 be a finite subgroup generated by elements that fix
some point in S n−1 . Then the quotient space S n−1 /G is simply connected.
Suppose S n−1 and S n−1 /G are equipped with PL structures that are admissible for the
action of G (cf. Section 3.3). According to Lemma 3.3, in order to show that S n−1 /G is a
closed PL manifold it suffices to check that for any point p ∈ S n−1 the isotropy group Gp
is a rotation group acting in Tp S n−1 = Rn−1 such that the quotient space Rn−1 /Gp is a PL
manifold. The condition on the homology groups of S n−1 /G can be verified as follows.
Lemma 5.4. Let a finite subgroup G < SOn and some i ∈ Z≥0 be given. If there are
subgroups H of G with coprime indices and Hi (S n−1 /H; Z) = 0, then Hi (S n−1 /G; Z) = 0. In
particular, the conclusion holds if the subgroups have coprime indices and the quotients Rn /H
are homeomorphic to Rn .
Proof. We choose an admissible triangulation K of S n−1 for the action of G and work with
simplicial homology over Z (in the following the coefficient ring is omitted and understood to
be Z). Let H < G be a subgroup with Hi (S n−1 /H) = 0. For an i-cycle c ∈ Zi (K/G) there
exists an i-cycle c0 ∈ Zi (K/H) such that π(c0 ) = [G : H] · c where πG/H : K/H → K/G is
the natural simplicial projection (e.g. c0 = µG/H (c) in the notation of [4, pp. 118-121]). The
(i + 1)-chain a ∈ Ci+1 (K/H) with ∂a = c0 satisfies ∂π(a) = π(∂a) = [G : H] · c and thus 0 =
(πG/H )∗ (µG/H )∗ ([c]) = [G : H] · [c] in Hi (S n−1 /G) (the induced map (µG/H )∗ : H∗ (K/G) →
H∗ (K/H) is called transfer, cf. [4, Ch. III. 2., pp. 118-121]). Hence, if there are subgroups H
of G with coprime indices and Hi (S n−1 /H) = 0, then it follows that Hi (S n−1 /G) = 0. The
assumption on Hi (S n−1 /H) is in particular fulfilled, if Rn /H is homeomorphic to Rn , since
this implies H∗ (S n−1 /H) = H∗ (S n−1 ) (cf. [14, p. 117]).
5.4. Chevalley’s theorem. For a unitary reflection group G < Un the following theorem
due to Chevalley holds [32, Thm. 3.20, p. 48].
Theorem 5.5. The algebra of invariants of a finite unitary reflection group G < Un is
generated by n algebraically independent homogenous polynomials.
For n such generators f1 , . . . , fn of C[z1 , . . . , zn ]G we will see in Section 6.2 that the map
f = [f1 , . . . , fn ] : Cn −→
Cn
v 7−→ (f1 (v), . . . , fn (v))
8
CHRISTIAN LANGE
descends to a homeomorphism f : Cn /G → Cn and, moreover, that R2n /G is in fact PL
homeomorphic to R2n .
5.5. The fundamental domain of a group. For a finite subgroup G < On there exists a
vector v0 ∈ Rn such that gv0 6= v0 for all g ∈ G\{e}. For such a vector v0 the set
\
Λ=
{v ∈ Rn |(v, v0 ) ≥ (v, gv0 )}
g∈G
n
is
S a fundamental domain for the group G, i.e. the G-translates of Λ cover R and nthe union
g∈G g Λ̊ is disjoint. It inherits a subspace topology and PL structure from R and the
quotient space Rn /G with its quotient topology and PL structure (cf. Sections 3.2 and 3.3)
can be obtained from Λ by identifying certain points on the boundary, namely those which
belong to the same orbit of G.
5.6. Gluing construction. We need the following elementary gluing result for PL balls.
The lemma states that “twisted” PL spheres are standard PL spheres. Its proof is a direct
consequence from the fact that a PL homeomorphism f : ∂∆n → ∂∆n can be linearly
extended in a radial direction to a PL homeomorphism f : ∆n → ∆n [41, Lem. 1.10, p. 8].
Lemma 5.6. Suppose B1n and B2n are PL balls and ϕ : ∂B1n → ∂B2n is a PL homeomorphism.
Then the space B1n ∪ϕ B2n obtained by gluing B1n and B2n together along their boundary via ϕ
is a PL n-sphere.
5.7. Collapsing. Let K be a simplicial complex and let σ, τ ∈ K be simplices such that
(i) τ < σ, i.e. τ is a proper face of σ,
(ii) σ is a maximal simplex in K and τ is not contained in any other maximal simplex
of K,
then τ is called a free face of K. A simplicial collaps of K defined by τ < σ as above is the
removal of all simplices ρ of K with τ ≤ ρ ≤ σ. We say that K collapses onto a subcomplex
L of K if there exists a finite sequence of collapses leading from K to L. A simplicial complex
that collapses onto a point is called collapsible. Being collapsible to a subcomplex is a PL
property, i.e. it does not depend on a specific triangulation (cf. [41, p. 39]). In our proof we
will apply the following characterization (cf. [41, Cor. 3.28, p. 41])
Lemma 5.7. A collapsible PL n-manifold (with or without boundary) is a PL n-ball.
In order to be able to apply this characterization, we will need the following lemma.
Lemma 5.8. Let p : K → K̃ be a simplicial surjection between finite simplicial complexes K
and K̃ that maps simplices of K homeomorphically onto simplices of K̃. Suppose further that
L is a subcomplex of K such that p restricts to a bijection p : K\L → K̃\p(L). If K collapses
onto L, then K̃ collapses onto p(L)
Proof. Let τ < σ ∈ K with τ, σ ∈
/ L and suppose that τ < σ defines a simplicial collaps of K.
By assumption on p we have p(τ ), p(σ) ∈
/ p(L) and thus the claim follows inductively, if we
can show that p(τ ) < p(σ) ∈ K̃ defines a simplicial collapse of K̃. But again, by assumption
on p, and because of τ, σ ∈
/ L it is clear that p(σ) is a maximal simplex of K̃ and that p(τ )
is not contained in any other maximal simplex of K̃.
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
9
6. Proof of the main result
In this section we prove the if direction of our main result by verifying its conclusion for
all reflection-rotation groups. The proof is structured as follows. For each reflection-rotation
group G we either prove the conclusion of our main result directly or we reduce such a proof
to the respective claim on reflection-rotation groups of lower order via the PL linearization
principle (cf. Appendix A for a summary of the classification result). In doing this we
will need to show that for each pair M / Grr of an irreducible reflection-rotation group Grr
that contains a reflection and a nontrivial normal rotation group M / Grr such that Grr is
generated by the reflections it contains and by M , there exists a nontrivial rotation group
H / M normalized by Grr such that the PL linearization principle can be applied to the
groups H / Grr (cf. Section 6.4). All such pairs M / Grr are listed in Theorem A.3. In
each case we will either show this property directly or reduce it to proving our main result
for reflection-rotation groups of order less than Grr . Once we have treated all the cases, the
if direction of our main result follows by induction. References to the sections in which the
respective cases are treated can be found in the appendix.
6.1. Real reflection groups. The fundamental domain Λ of a reflection group W < On
acting on S n−1 is a spherical simplex [7, Thm. 4, p. 595]. Let W + be the orientation
preserving subgroup of W and, if there exists some rotation h ∈ On \W that normalizes W ,
+
set W × = hW, hi and W ×
= hW + , hi. Choose an admissible triangulation for the action of W
+
n−1
(and hence of W ) on S
that refines the triangulation of S n−1 by the fundamental domains
of W . Then the quotient space S n−1 /W is a PL ball, namely the fundamental domain Λ of
W , and the quotient space S n−1 /W + can be obtained by gluing together two copies of Λ along
their boundary, i.e. the resulting space is a PL sphere by Lemma 5.6. Moreover, a coset s
of a reflection s ∈ W interchanges the two copies. Therefore the PL linearization principle
can be applied to the groups W + / W . If h exists, then its action on S n−1 /W + commutes
with the action of a reflection s ∈ W on S n−1 /W + , since h normalizes W by assumption.
+
In particular, the quotient space S n−1 /W ×
can be realized as the suspension of ∂Λ/h whose
+
cone points are interchanged by s. Therefore, it is clear that S n−1 /W ×
is a PL sphere, that
n−1
×
S
/W is a PL ball and that the PL linearization principle can be applied to the groups
+
W + / W × and W ×
/ W × . Hence, we have proven
Lemma 6.1. In the notation used above our main theorem holds for groups of type W , W + ,
+
+
W×
, W × and the PL linearization principle can be applied to W + /W , W + /W × and W ×
/W × .
+
In particular, it can be applied to W (Dn ) / W (BCn ).
6.2. Reflection groups induced by unitary reflection groups. For a unitary reflection
group G < Un we choose n algebraically independent homogenous generators f1 , . . . , fn ∈
C[z1 , . . . , zn ]G given by Chevalley’s theorem (cf. Section 5.4). The continuous map
f = [f1 , . . . , fn ] : Cn −→
Cn
v 7−→ (f1 (v), . . . , fn (v))
descends to a continuous map f : Cn /G → Cn . The map f is injective, since the algebra of
invariants of G separates its orbits [32, Thm. 3.5, p. 41], and also onto [32, Thm. 3.15, p. 45].
Moreover, since C[z1 , . . . , zn ] is a finitely generated C[z1 , . . . , zn ]G -module [43, Thm. 1.3,
p. 478], the map f is a finite and hence proper morphism of complex affine varieties [13,
10
CHRISTIAN LANGE
6.1.11, 5.5.3]. Therefore, the map f is also proper with respect to the usual topology in the
sense of [2, Ch. 1, §10, no. 1, Def. 1] by [12, Ch. XII., Prop. 3.2]. In particular, the map f
is closed. Consequently f is a homeomorphism and thus R2n /G and R2n are homeomorphic
where G is regarded as a real rotation group.
−1
can alternatively be shown by induction as follows. According to a
Continuity of f
theorem by Steinberg isotropy groups of unitary reflection groups are again unitary reflection
groups [44, Thm. 1.5, p. 394] (cf. [32, Thm. 9.44, p. 186] and [31]). Hence, it follows by
induction that Cn /G − {0} is a topological manifold, where 0 is the coset of 0 ∈ Cn in Cn /G.
−1
Using the domain invariance theorem it is then not difficult to conclude that f is continuous
(cf. [26] for more details).
Finally, since isotropy groups of unitary reflection groups are again unitary reflection groups
by Steinbergs theorem, it follows by induction as explained in Section 5.3 that the PL quotient
R2n /G for a unitary reflection group G regarded as a rotation group is PL homeomorphic to
R2n , i.e. we have
Lemma 6.2. Our main theorem holds for unitary reflection groups considered as real groups.
6.3. Reflection-rotation groups in dimension up to four. Groups up to dimension
three can be easily treated by hand. For instance, all finite subgroups of SO2 and SO3 are
orientation preserving subgroups of reflection groups which have been treated in Section 6.1.
In dimension four a classification based proof becomes rather long-winded and cumbersome
(cf. [36, §3]).
The following approach dispenses with the classification in dimensions up to four. The
proof is based on induction on the dimension. For simplicity let us suppose that we have
already treated the cases n < 4, i.e. we formulate the proof for n = 4. The same arguments
apply in dimenions n < 4. All isotropy groups of a rotation group G < SO4 are again rotation
groups. Therefore, the quotient S 3 /G is a closed simply connected PL manifold with respect
to a PL structure induced by K/G where K is an admissible triangulation for the action of
G on S 3 (cf. Lemma 3.3). Hence, by the PL version of the Poincaré conjecture (cf. Section
5.3), the PL quotient S 3 /G is PL homeomorphic to the standard PL 3-sphere and R4 /G is
PL homeomorphic to R4 , i.e. we have
Lemma 6.3. Our main theorem holds for rotation groups in dimension up to four.
Now let G < O4 be a finite group and suppose that H / G is a rotation group (again, the
case of lower dimensions works analogous). We endow S 3 with an admissible PL structure for
the action of G (and hence H). According to [25] (cf. [26] and [45, p. 208]) the quotient S 3 /H
admits a smooth structure such that the identity map of S 3 /H is a PD homeomorphism and
such that G/H acts smoothly on it. Since smoothings of PL manifolds in dimension three are
unique up to diffeomorphism [45, Thm. 3.10.9], this action is smoothly conjugate to a smooth
action of G/H on the standard sphere S 3 and by [8, Thm. E] it is thus smoothly conjugate
to a linear action on S 3 . Therefore, we have a PD homeomorphism g : S 3 /H → S 3 and a
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
11
homomorphism r : G → SO4 such that the following diagram commutes
G × S 3 /H
r×g
r(G) × S 3
/ S 3 /H
g
/ S3
According to Section 5.2 we can replace g by a PD homeomorphism f such that the induced
PL structure on S 3 is admissible with respect to the action of r(G). Therefore, the PL
linearization principle can be applied to the groups H / G (cf. Section 5.1). In particular,
taking G as a reflection-rotation group and H as its orientation preserving subgroup proves
our main theorem for all reflection-rotation groups up to dimension four. Summarizing we
have
Lemma 6.4. Our main theorem holds for reflection-rotation groups in dimension up to four.
If H < On , n ≤ 4, is a rotation group normalized by a finite group G < On , then the PL
linearization principle can be applied to the groups H / G.
Note that in principle the usage of the Poincaré conjecture can be avoided by applying other
means such as the algebra of polynomial invariants or explicit constructions of fundamental
domains (cf. [36] and Section 6.7 for illustrations of these methods). However, proofs along
such lines are cumbersome and so we do not refrain from using the Poincaré conjecture as a
convenient tool.
6.4. Reducible reflection-rotation groups. Now let G < On be a reducible reflectionrotation group and let Rn = V1 + . . . + Vk be a decomposition into irreducible components.
Let Hi /G be the normal subgroup generated by rotations that only act in Vi (i.e. by rotations
of the first kind in Vi in terms of [27]) and let Gi be the projection of G to the i-th factor.
We can assume that Hi 6= Gi because otherwise Hi splits of as a direct factor. The pairs
Hi / Gi that occur in this way are classified in [27, Thm. 3]. It is shown that this classification
amounts to a classification of pairs M / Grr of an irreducible reflection-rotation group Grr
that contains a reflection and a normal subgroup M / Grr generated by rotations such that
Grr is generated by its reflections and by M (cf. the remark following [27, Thm. 3]). All such
pairs with nontrivial M are listed in Theorem A.3 in the appendix.
Suppose there is some i ∈ {1, . . . , k} and some nontrivial rotation group H < Hi normalized
by Gi such that the PL linearization principle can be applied to H / Gi . Then the PL
linearization principle can also be applied to H / G. Hence, the following lemma holds (cf.
Section 5.1).
Lemma 6.5. Let G be a reducible reflection-rotation group and suppose that some Hi is
nontrivial. Suppose further that our main theorem holds for all reflection-rotation groups of
smaller order than G. If for each pair M / Grr occurring in Theorem A.3 there exists some
nontrivial rotation group H / M normalized by Grr such that the PL linearization principle
can be applied to the groups H / Grr , then our main theorem holds for the group G.
Note that it is necessary to verify the assumption on the pair M / Grr in each case of
Theorem A.3. In fact, given such a pair M / Grr a reducible rotation group G can be
constructed with two irreducible components and with H1 = H2 = M , G1 = G2 = Grr (cf.
12
CHRISTIAN LANGE
[27, Thm. 4]). In the course of the proof, in each case of Theorem A.3 we either verify the
condition on M / Grr directly or reduce such a proof to showing our main result for rotation
groups of order less than Grr . A reference to the lemma in which we do this for a specific
pair M / Grr can also be found in Theorem A.3.
The only case in which we cannot apply Lemma 6.5 is, in the notation above, when all the
Hi are trivial, i.e. when there are no rotations in G that act in a single Vi factor. Suppose
this is the case. If the group G does not split as a product of nontrivial factors of lower order,
it is either a reflection group of type A1 , a rotation group of type W + (A1 × · · · × A1 ) or a
rotation group of the form
∆ϕ (W × W ) := {(g, ϕ(g)) ∈ W × W |g ∈ W } < Om × Om
for some reflection group W < Om and some isomorphism ϕ : W → W that maps reflections
onto reflections [27, Thm. 4, Cor. 64]. The first two cases are treated in Lemma 6.1. For m < 3
the third case is treated in Lemma 6.4. If all labels of the Coxeter graph of W lie in {3, 4, 6},
then every automorphism of W that maps reflections onto reflections can be realized through
conjugation by an orthogonal transformation in the normalizer of W in Om [11, Cor. 19,
p. 7]. In this case the quotient R2m /∆ϕ (W × W ) is PL homeomorphic to R2m /∆id (W × W ).
Hence this case is subject of Lemma 6.2, since ∆ = ∆id (W × W ) < W × W preserves the
complex structure
0
1m
J=
.
−1m 0
and can thus be regarded as a unitary reflection group acting on Cm . The only remaining
cases are W = W (H3 ) and W = W (H4 ) (cf. Theorem A.3 or [27, Thm. 3]) and indeed, in
these cases there exist outer automorphisms of W that map reflections onto reflections but
cannot be realized through conjugation in Om (cf. [10, pp. 31-32]). Note that the argument in
[36] breaks down for groups ∆ϕ (W × W ) for which ϕ cannot be realized through conjugation
since the proof of [36, Thm. 1.2] does not work in this case. Summarizing, we have
Lemma 6.6. In the notation above, let G be a reducible reflection-rotation group such that
all the Hi are trivial. Suppose that our main theorem holds for all reflection-rotation groups
of smaller order than G. If G is different from ∆ϕ (W × W ) for W of type H3 and H4 , then
our main theorem holds for the group G.
The two exceptional cases excluded in the lemma are treated in Section 6.7.
6.5. Monomial reflection-rotation groups. Let D(n) be the diagonal subgroup of On and
let D+ (n) be its orientation preserving subgroup. For a permutation group H < Sn consider
the monomial groups M = D+ (n)oH < SOn and M × = D(n)oH < On . The subgroup H <
M × leaves the spherical simplex Λ = {x ∈ S n−1 |xi ≥ 0, i = 1, . . . , n} invariant and acts on its
boundary ∂Λ. Choose an admissible triangulation for the action of M × on S n−1 that refines
the triangulation of S n−1 by the D(n) translates of Λ. The quotient S n−1 /D+ (n) of S n−1 by
the diagonal subgroup D+ (n) can be obtained by gluing together two copies of Λ along their
boundaries. It can be realized as the suspension of ∂Λ whose cone points are interchanged by
the reflections in D(n). Therefore the PL linearization principle can be applied to the groups
D+ (n) / M × . In particular, we have (note that D(W + (BCn )) = D(W + (Dn )) = D+ (n))
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
13
Lemma 6.7. The PL linearization principle can be applied to the groups D(W + (BCn )) /
W (BCn ) and thus also to the groups D(W + (Dn )) / W (Dn ).
Note that with respect to the constructed linearization the reflection in M × /D+ (n) acts in
a 1-dimensional subspace orthogonal to a subspace Rn−1 in which H ∼
= M/D+ (n) acts. Now
suppose that M is a rotation group. Then the linearization of H acts as a rotation group (cf.
Section 5.1). With these observations we obtain
Lemma 6.8. In the notation above, suppose that M is a rotation group and that our main
theorem holds for all reflection-rotation groups of smaller order than M . Then our main
theorem holds for the reflection-rotation groups M and M × and the PL linearization principle can be applied to the pairs M / M × . In particular, this applies in the cases of M =
M5 , M6 , M7 , M8 , M (Dn ) in Theorem A.1, (v), (a).
The exceptional monomial rotation groups M (Q7 ) = M7p < SO7 and M (Q8 ) = M8p < SO8
(cf. Theorem A.1, (v), (a)) are treated in Section 6.7.
6.6. Imprimitive reflection-rotation groups. Let µm < C∗ be the cyclic subgroup of
m-th roots of unity. For a factor p of m set
n
o
A(m, p, n) := (θ1 , . . . , θn ) ∈ µnm |(θ1 . . . θn )m/p = 1 .
Let G(m, p, n) be the semidirect product of A(m, p, n) with the symmetric group Sn . Then the
natural realization of G(m, p, n) in the unitary group Un is an imprimitive unitary reflection
group (cf. [32, Ch. 2, p. 25]). In the following we consider G(m, p, n) as a real rotation group
in SO2n .
In this section we treat the irreducible imprimitive rotation groups
+
G×
(kq, k, n) = hG(kq, k, n), τ i < SO2n
where n > 2, q ∈ N, k = 1, 2, kq ≥ 3 and where τ is a rotation that conjugates the first two
coordinates, i.e.
τ (z1 , z2 , z3 . . . , zn ) = (z 1 , z 2 , z3 . . . , zn ),
and the corresponding irreducible imprimitive reflection-rotation groups
G× (kq, k, n) = hG(kq, k, n), si < SO2n ,
where s is a reflection that conjugates the first coordinate, i.e.
s(z1 , z2 , z3 . . . , zn ) = (z 1 , z2 , . . . , zn )
(cf. [27] for more details on the constructions of these groups). Let R2n = V1 + · · · + Vn
be a decomposition into components of a system of imprimitivity of G× (kq, k, n) (and hence
+
of G×
(kq, k, n)), i.e. a decomposition into subspaces that are permuted by the group. Let
H / G× (kq, k, n) be the normal subgroup generated by rotations in G× (kq, k, n) that only act
in one of the factors Vi , i = 1, . . . , n. The projection Hi of H to O(Vi ) is a cyclic group of
order q and the group H splits as a product of these projections, i.e. H = H1 × · · · × Hn .
Because of k ∈ {1, 2} and kq ≥ 3, the group H is nontrivial. Due to Lemma 6.4 the PL
linearization principle can be applied to the groups H / G× (kq, k, n). Hence, we have
14
CHRISTIAN LANGE
+
Lemma 6.9. Let G be a reflection-rotation group of type G× (kq, k, n) or G×
(kq, k, n) with
n > 2, k = 1, 2, kq ≥ 3 and suppose that our main theorem holds for reflection-rotation groups
of smaller order than G. Then our main result holds for the group G.
Moreover, we see
+
Lemma 6.10. Let M / Grr be a pair occurring in Theorem A.3 of type G×
(kq, k, n) /
×
×
+
×
G (kq, k, n) or G (2q, 2, n) / G (2q, 1, n), n > 2, k = 1, 2, kq ≥ 3. Then there exists a
nontrivial rotation group H < M normalized by Grr such that the PL linearization principle
can be applied to the groups H / Grr .
Observe that by now we have verified the conclusion of the preceding lemma for all pairs
of groups M / Grr occurring in Theorem A.3, and hence established the respective condition
in Lemma 6.5 on reducible reflection-rotation groups.
6.7. Exceptional rotation groups. The only indecomposable rotation groups which we
have not treated in view of our main result yet are the exceptional irreducible rotation groups
M (R5 ), M (S6 ), M (Q7 ), M (Q8 ) and M (T8 ) and the exceptional reducible rotation groups
∆ϕ (W ×W ) for W of type H3 and H4 (cf. Section 6.4 and [27, Sect. 4.6]). The proofs in [36] in
the cases of M (R5 ), M (S6 ), M (Q7 ), M (Q8 ) in principle work [36, II)-IV) in Thm. 1.4, p. 105]
but lack some arguments. This manifests in the fact that isotropy groups, which determine
the local structure of the respective quotient, are not examined. The cases of M (T8 ) and
∆ϕ (W × W ) for W of type H3 and H4 are not considered in [36].
In the cases n > 5 we will make use of the PL version of the generalized Poincaré conjecture
(cf. Theorem 5.2). For n = 5 this tool is not available, which is why we have to perform a
computation by hand in the case of M (R5 ). Our arguments turn the approach in [36, p. 102]
to this case into a rigorous proof.
Lemma 6.11. Our main result holds for the rotation group M (R5 ), i.e. for G = M (R5 ) <
SO5 the PL quotient R5 /G is PL homeomorphic to R5 .
Proof. The outline of the proof is as follows. First we construct a fundamental domain Λ on
S 4 and choose an admissible triangulation that refines the tesselation of S 4 by the translates
of Λ. With respect to the induced PL structure Λ is a PL 4-ball. We can choose a PL collar
A of ∂Λ in Λ that collapses onto ∂Λ [41, Cor. 3.17, Cor. 3.30] (cf. Section 5.7). The closure
of the complement of A in Λ, say B, is a PL 4-ball. We set Q = ∂Λ/ ∼ and N = A/ ∼
where ∼ denotes the equivalence relation induced by G. Then we can recover S 4 /G by gluing
together N and B along the PL 3-sphere ∂B which is not affected by ∼. In particular, we
see that S 4 /G is a PL 4-sphere, if we can show that N is a PL 4-ball. Due to Section 6.3 and
the fact that all isotropy groups of G are rotation groups [27, Lem. 27], we already know that
N is a PL 4-manifold with boundary (cf. Lemma 3.3). Moreover, since our triangulation is
admissible, the projection A → N is simplicial and maps simplices homeomorphically onto
simplices (cf. Section 3.3). Hence, N collapses onto Q by Lemma 5.8. Therefore, according
to Lemma 5.7 it is sufficient to show that Q is collapsible in order to prove the lemma.
The group G is isomorphic to the alternating group A5 and a specific set of generators in A6
is given by (12)(34), (15)(23), (16)(24) (cf. [36, p. 102]) where G is regarded as the restriction
of the permutation action of S6 on R6 to the subspace R5 = {(x1 , . . . , x6 ) ∈ R6 |x1 +. . .+x6 =
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
15
0} of R6 . A fundamental domain Λ for the action of G on S 4 is constructed in [36, p. 103] as
follows: For v0 = {−1, −1, −1, 0, 1, 2} we have gv0 6= v0 for all g ∈ G and thus
\
Λ=
{v ∈ S 4 |(v, v0 ) ≥ (v, gv0 )}
g∈G
is a fundamental domain for the action of G on S 4 ⊆ R5 ⊆ R6 . It can be described by the 8
inequalities
−x3 + x4
−x4 + x5
−x5 + x6
−x1 + x5
≥ 0,
≥ 0,
≥ 0,
≥ 0,
−x2 + x6
−x1 − 2x2 + x4 + 2x5
−2x1 − x2 + x4 + 2x5
−2x2 − x3 + x4 + 2x5
≥ 0,
≥ 0,
≥ 0,
≥0
and has vertices
1
v1 = √ (−5, 1, 1, 1, 1, 1),
30
1
v3 = √ (1, 1, −5, 1, 1, 1),
30
1
v5 = √ (1, −1, −1, −1, 1, 1),
6
1
v7 = √ (−1, 1, −1, −1, 1, 1),
6
1
v2 = √ (1, −5, 1, 1, 1, 1),
30
1
v4 = √ (−1, −1, −1, −1, −1, 5),
30
1
v6 = √ (1, 1, −5, −5, 4, 4),
84
1
v8 = √ (−5, 4, −5, 1, 1, 4).
84
Let Pi be the boundary of the half-space in R6 determined by the ith inequality above. The
faces of the fundamental domain are the following three-dimensional polytopes: In the plane
P1 the pentagonal pyramid v1 v2 v4 v5 v6 v7 with vertex v4 ; in P2 the double pyramid v1 v2 v3 v4 v8
with vertices v2 and v8 ; in P3 the pentagonal pyramid v1 v2 v3 v5 v6 v7 with vertex v3 ; in P4
the simplex v2 v3 v4 v5 ; in P5 the simplex v1 v3 v7 v8 ; in P6 the double pyramid v3 v4 v6 v7 v8 with
vertices v6 and v8 ; in P7 the simplex v3 v4 v5 v6 and in P8 the simplex v1 v4 v7 v8 . The boundary
of the fundamental domain is illustrated in Figure 1.
It can be checked that the identifications induced by G on ∂Λ are generated by the following
identifications
(12)(34) in P1 :v1
v2 , v4
v4 , v6
v6 , v5
v7 ,
(13)(45) in P2 :v1
v3 , v2
v2 , v4
v4 , v8
v8 ,
(12)(56) in P3 :v1
v2 , v3
v3 , v5
v7 , v6
v6 ,
(15)(23) in P4 :v2
v3 , v4
v4 , v5
v5 ,
(13)(26) in P5 :v1
v3 , v7
v7 , v8
v8 ,
(14)(25) in P6 :v3
v3 , v4
v4 , v6
v8 , v7
v7 ,
(13425) :v1 * v3 , v4
v4 , v7 * v5 , v8 * v6 ,
(15243) :v1 ) v3 , v4
v4 , v7 ) v5 , v8 ) v6 ,
16
CHRISTIAN LANGE
Figure 1. Boundary of a fundamental domain Λ for the action of M (S5 ) =
R5 (A5 ) on S 4 cut into two pieces.
The first six of them correspond to “pasting in half” the faces of Λ lying in the planes
P1 , . . . , P6 . Since points in the interior of such a face are not identified with points outside
the interior of this face, we see that the images of these faces in the quotient Q can be collapsed
to the images of their boundary in Q. In particular, we see that Q collapses onto the images
of the faces v3 v4 v5 v6 and v1 v4 v7 v8 of Λ. Since these faces are identified with each other by
(13425), we see that Q collapses onto the image of v3 v4 v5 v6 in Q. Examining the list of
generating identifications shows that this image is a 3-simplex itself. Hence, Q is collapsible
and thus the claim follows by the remarks above.
In principle, the proofs in [36] in the cases M (S6 ), M (Q7 ) and M (Q8 ) can be made rigorous
in the same way. However, in order to avoid long computations with fundamental domains
and identifications, we provide the following alternative argument.
In each of the remaining cases choose an admissiable triangulation of S n−1 for the action of
G. Because of n > 5 in these cases, it is sufficient to show that S n−1 /G is a simply connected
PL manifold with H∗ (S n−1 /G) = H∗ (S n−1 ) in order to prove that Rn /G is piecewise linear
homeomorphic to Rn (cf. Section 5.3, Theorem 5.2). According to [27, Lem. 27] all isotropy
groups of the remaining irreducible exceptional rotation groups M (S6 ), M (Q7 ), M (Q8 ) and
M (T8 ) are rotation groups. The same statement is true for reducible rotation groups of type
∆ϕ (W × W ) (cf. Section 6.4), since isotropy groups of real reflection groups are generated
by the reflections they contain [20, Thm. 1.12 (c), p. 22] (apply this result twice). Hence, we
have (recall that the quotient S n−1 /G is simply connected by Lemma 5.3)
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
17
Lemma 6.12. Let G be one of the remaining rotation groups and suppose that our main
result holds for all rotation groups of smaller order. Then S n−1 /G is a simply connected PL
manifold.
In particular, we have H1 (S n−1 /G) = π1 (S n−1 /G)ab = 0 [14, Thm. 2A.1]. By Poincaré
duality and the universal coefficient theorem, it is sufficient to show that H2 (S n−1 /G) = 0
or H2 (S n−1 /G) = H3 (S n−1 /G) = 0, respectively, depending on whether n ≤ 6 or 6 < n ≤ 8
[14, Thm. 3.2, Thm. 3.30], in order to prove H∗ (S n−1 /G) = H∗ (S n−1 ). In Lemma 5.4 we
have seen that the existence of subgroups H < G with coprime indices and Hi (S n−1 /H) = 0
implies Hi (S n−1 /G) = 0. The latter condition in particular holds for rotation groups H for
which we have already shown that Rn /H is homeomorphic to Rn (cf. the long exact sequence
in [14, p. 117]). Hence, in order to treat the remaining cases, it is sufficient to find suitable
subgroups.
Lemma 6.13. If our main result holds for all rotation groups of smaller order than M (S6 ),
then it also holds for M (S6 ).
Proof. The group M (S6 ) ∼
= PSL2 (7) has order 23 ·3·7 and thus contains a subgroup H of order
7. Since M (S6 ) can be realized as the isomorphic image of a permutation group in S7 < SO7
to SO6 (cf. [27, Lem. 16]), this subgroup of order 7 is generated by a 7-cycle in S7 and acts
thus freely on the unit sphere S 5 ⊂ R6 . The facts that H2 (Z7 ) = 0 [5, (3.1), p. 35] and that
H2 (S 5 /H) = H2 (H) for groups acting freely on S 5 [5, p. 20] imply that H2 (S 5 /H) = 0. A
rotation group of order 23 · 3 contained in M (S6 ) is described in [27, Lem. 28]. Hence, the
claim follows by the remarks above.
The reflection groups W (H3 ) and W (H4 ) properly contain reflection groups with coprime
indices (cf. [9, Table 8, Table 9]) and thus the rotation groups ∆ϕ (W × W ) for W of type H3
and H4 properly contain rotation groups with coprime indices. The rotation groups M (Q7 )
and M (T8 ) also properly contain rotation groups with coprime indices [27, Lem. 29, Lem. 31].
By the remarks above, we obtain
Lemma 6.14. Let G be a rotation group of type M (Q7 ), M (T8 ) or ∆ϕ (W × W ) for W of
type H3 or H4 . If our main result holds for all rotation groups of smaller orden than G, then
it also holds for G.
The group M (Q7 ) does not contain rotation groups with coprime indices. However, the
following lemma is proven in [27, Lem. 30].
Lemma 6.15. The rotation group M (Q8 ) of order 21504 = 210 · 3 · 7 contains a reducible
rotation group G of order 1536 = 29 · 3 with k = 2 and
(Gi , Hi , Fi , Gi /Hi ) = (W (D4 ), D(W + (D4 )), W (D4 ), W (A3 )),
i = 1, 2 (see below for an explanation of this notation), which is normalized by an element
τ of order two that interchanges the two components of G. Moreover, it contains a rotation
group M (S6 ) of order 168 = 23 · 3 · 7.
More precisely, with respect to the action of the group G < SO8 in the preceding lemma
splits into two four-dimensional irreducible subspaces R8 = V1 ⊕ V2 . The subgroup Hi of
G generated by rotations that only act in Vi is of type D(W + (D4 )). The quotient of G by
R8
18
CHRISTIAN LANGE
H = H1 × H2 is abstractly isomorphic to W (A3 ) ∼
= S4 and its linearization on R8 acts as a
diagonal group ∆ϕ (W × W ) for W of type A3 (cf. Section 6.4).
∼ C(S 3 ∗ RP3 ) and
Lemma 6.16. For N = hG, τ i as in the preceding lemma we have R8 /N =
hence H∗ (S 7 /N ) = H∗ (Σ4 (RP3 )). In particular, H2 (S 7 /N ) = H3 (S 7 /N ) = 0.
Proof. Let R8 = V1 + V2 be a decomposition into irreducible components with respect to G.
We can assume that V1 and V2 are two identical copies of R4 such that the projections of G
to V1 and V2 coincide. Since τ normalizes G it has the form
0 h
τ=
h−1 0
for some h ∈ NO4 (W (D4 )). After conjugation we can assume that h = id and thus we can
apply the linearization principle to the groups H1 × H2 / N . It then remains to show that
R8 /N ∼
= C(S 3 ∗ RP3 ) for N = G, σ with σ = τ and G = Gϕ = {(g, ϕ(g))|g ∈ W (A3 )} ⊂
W (A3 ) × W (A3 ) < O4 × O4 for some ϕ ∈ Aut(W (A3 )). Since the symmetric group S4 has
no outer automorphisms we can assume that ϕ = id after conjugation. Then σ has the form
0 g0
< SO8
σ=
g0−1 0
for some g0 ∈ W (A3 ) < O4 . Now σ ∈ N (G) implies g02 ∈ CO4 (W (A3 )) = {±id3 } × {±id1 } <
O3 × O1 . Because of g0 ∈ W (A3 ) and since W (A3 ) has a trivial center, we have g02 = 1, i.e.
g0 = g0−1 . If we identify R8 with C4 then G can be regarded as a unitary reflection group
and the action of σ is given by σ ((z1 , z2 , z3 , z4 )) = i · g0 (z 1 , z 2 , z 3 , z 4 ) where g0 permutes
the coordinates. Let si be the elementary symmetric polynomial of degree i in the zj , i, j =
1, 2, 3, 4. Then the map
f:
C4 /G
→ C4
(z1 , z2 , z3 , z4 ) 7→ (s1 , s2 , s3 , s4 )
defines a homeomorphism as explained in Section 5.4. Since the si are invariant under coordinate permutations, the induced action of N /G on C4 is given by
σ(s1 , s2 , s3 , s4 ) = (i · s1 , −s2 , −i · s3 , −s4 ).
∼
It follows that
= C(S 3 ∗ RP3 ). In particular, we have H2 (S 7 /N ) = H3 (S 7 /N ) = 0 by
the long exact sequence in [14, p. 117].
R8 /N
With the preceding two lemmas and the remarks above we obtain
Lemma 6.17. If our main result holds for all rotation groups of smaller order than M (Q8 ),
then it also holds for M (Q8 ).
Since we have by now treated all cases, our main result follows by induction.
7. Towards a classification free proof
As a corollary of our main result reflection-rotation groups share the following properties
(i) − (iii) (cf. Lemma 3.3). Property (i) generalizes an isotropy theorem by Steinberg on
unitary reflection groups [44, Thm. 1.5, p. 394] (cf. [3, Ch. V, Exercise 8] and [31] for
alternative proofs by Bourbaki and Lehrer).
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
19
(i) Isotropy groups of reflection-rotation groups are generated by the reflections and
rotations they contain.
(ii) For a rotation group G < SOn we have
H∗ (S n−1 /G; Z) = H∗ (S n−1 ; Z).
(iii) For a reflection-rotation group G < On that contains a reflection we have
H∗ (S n−1 /G; Z) = H∗ ({∗}; Z).
Conversely, these properties together with the PL h-cobordism theorem and some extra work
in low dimensions imply the if direction of our result by induction as explained in Section 5.3.
Hence, in order to essentially dispense with the classification of reflection-rotation groups in
the proof of our result one might first try to find conceptual proofs for properties (i), (ii) and
(iii) that do not depend on a classification of reflection-rotation groups.
A. Appendix
Classification of rotation-reflection groups. We provide the classification result from
[27] omitting details that are not needed in the present paper. For more detailed descriptions
we refer to the respective sections above and to [27].
Theorem A.1. Every irreducible rotation group occurs, up to conjugation, in precisely one
of the following cases.
(i) Orientation preserving subgroups W + of irreducible real reflection groups W (Lem.
6.1).
(ii) Irreducible unitary reflection groups that are not the complexification of a real reflection group considered as real groups (Lem. 6.2).
+
(iii) The imprimitive rotation groups G×
(km, k, l) < SOn for n = 2l > 4, k ∈ {1, 2} and
km ≥ 3 (Lem. 6.9).
+
(iv) The unique extensions W ×
of W + by a normalizing rotation if such exists (Lem. 6.1).
(v) The following rotation groups M of type P5 , P6 , P7 , P8 , Q7 , Q8 , R5 , S6 and T8 .
More precisely (in the same order),
(a) monomial rotation groups M5 , M6 , M7 , M8 (Lem. 6.8) and M7p (Lem. 6.14)
and M8p (Lem. 6.17).
(b) primitive rotation groups R5 (A5 ) (Lem. 6.11) and R6 (PSL2 (7)) (Lem. 6.13)
given as the image of the unique irreducible representations of A5 in SO5 and of
PSL2 (7) in SO6 ,
(c) a primitive rotation group in SO8 isomorphic to an extension of the alternating
group A8 by a nonabelian group of order 28 (Lem. 6.14).
(vi) Other examples in dimension 4 (Lem. 6.4).
The references in the parentheses indicate where the respective group is treated in view of our
main result.
For a real reflection group W we denote by W × its unique extension by a normalizing
+
rotation, provided such exists, i.e. W × = hW ×
, W i. For a monomial rotation group M
×
we denote by M its extension by a coordinate reflection (cf. Section 6.5). Finally, for an
imprimitive rotation group of type G(km, k, l) let G× (km, k, l) be its extension by a reflection
20
CHRISTIAN LANGE
s of the form s(z1 , . . . , zl ) = (z 1 , z2 , . . . , zl ) (cf. Section 6.6). The following result holds (cf.
[27, Thm. 2]).
Theorem A.2. Every irreducible reflection-rotation group either appears in Theorem A.1 or
it contains a reflection and occurs, up to conjugation, in one of the following cases.
(i) Irreducible real reflection groups W (Lem. 6.1).
(ii) The groups W × generated by a reflection group W of type A4 , D4 , F4 , A5 or E6 and
a normalizing rotation (Lem. 6.1).
(iii) The monomial groups M × of type Dn , P5 , P6 , P7 , P8 , i.e. M × (Dn ) := D(W (BCn ))o
An , M5× , M6× , M7× and M8× (Lem. 6.8).
(iv) The imprimitive groups G× (km, k, l) < SOn with n = 2l, k = 1, 2 and km ≥ 3 (Lem.
6.9).
The references in the parentheses indicate where the respective group is treated in view of our
main result.
For a reducible reflection-rotation group G < On let Rn = V1 ⊕ . . . ⊕ Vk be a decomposition
into irreducible components, let Gi be the projection of G to O(Vi ) and let Hi / Gi be the
normal subgroup of G generated by rotations in G that only act in Vi . Then the group Gi
is generated by its reflections and by Hi . If Gi does not split off as a direct factor, then
Gi contains a reflection and is thus distinct from Hi . Conversely, every pair M / Grr of an
irreducible reflection-rotation group Grr that contains a reflection and a normal subgroup
M / Grr generated by rotations such that Grr is generated by its reflections and by M occurs
in this way. The following theorem is proven in [27, Thm. 3] (cf. the remark following [27,
Thm. 3]).
Theorem A.3. Let Grr < On be an irreducible reflection-rotation group that contains a
reflection and let M / Grr be a nontrivial normal subgroup generated by rotations. If Grr is
generated by the reflections it contains and by M , then the pair M / Grr occurs in precisely
one of the following cases.
(i) M / M × for M = M5 , M6 , M7 , M8 , M (Dn ) = W + (Dn )
(Lem. 6.8)
+
(ii) G×
(km, k, l) / G× (km, k, l) for k = 1, 2, km ≥ 3 and n = 2l
(Lem. 6.10)
+
(iii) G×
(2m, 2, l) / G× (2m, 1, l) for m ≥ 2 and n = 2l
(Lem. 6.10)
+
(iv) W / W for any irreducible reflection group W .
(Lem. 6.1)
(v) D(W + (BCn )) / W (BCn )
(Lem. 6.7)
(vi) W + (Dn ) / W (BCn )
(Lem. 6.1)
(vii) D(W (Dn )) / W (Dn )
(Lem. 6.7)
+
(viii) W ×
/ W × for a reflection group W of type A5 or E6 .
(Lem. 6.1)
(ix) Other examples in dimensions n ≤ 4
(Lem. 6.4)
In the lemmas referred to, we either show the existence of a nontrivial rotation group H / M
normalized by Grr for which the linearization principle can be applied to the groups H / Grr
directly or reduce such a claim to the validity of our main result for reflection-rotation groups
of order less than Grr .
References
[1]
M. A. Armstrong, The fundamental group of the orbit space of a discontinuous group, Proc. Cambridge
Philos. Soc. 64 (1968), 299–301.
CHARACTERIZATION OF REFLECTION-ROTATION GROUPS
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
21
N. Bourbaki, Éléments de mathématique. Topologie générale. Chapitres 1 à 4, Hermann, Paris, 1971.
N. Bourbaki, Groupes et algèbres de Lie, Hermann, Paris, 1968 (French).
Glen E. Bredon, Introduction to Compact Transformation Groups, Academic Press, New York-London,
1972. Pure and Applied Mathematics, Vol. 46.
Kenneth S. Brown, Cohomology of groups, Graduate Texts in Mathematics, vol. 87, Springer-Verlag,
1994.
J. W. Cannon, Shrinking cell-like decompositions of manifolds. Codimension three, Ann. of Math. (2)
110 (1979), no. 1, 83–112.
H. S. M. Coxeter, Discrete groups generated by reflections, Ann. of Math. (2) 35 (1934), no. 3, 588–621.
Jonathan Dinkelbach and Bernhard Leeb, Equivariant Ricci flow with surgery and applications to finite
group actions on geometric 3-manifolds, Geom. Topol. 13 (2009), no. 2, 1129–1173.
J. Matthew Douglass, Götz Pfeiffer, and Gerhard Röhrle, On reflection subgroups of finite Coxeter groups,
Comm. Algebra 41 (2013), no. 7, 2574–2592.
William N. Franzsen, Automorphisms of Coxeter Groups, PhD thesis, University of Sydney, 2001.
W. N. Franzsen and R. B. Howlett, Automorphisms of nearly finite Coxeter groups, Adv. Geom. 3 (2003),
no. 3
A. Grothendieck, Revêtements étales et groupe fondamental (SGA 1), Séminaire de géométrie algébrique
du Bois Marie 1960–61 (French).
A. Grothendieck, Éléments de géométrie algébrique. II. Étude globale élémentaire de quelques classes de
morphismes, Inst. Hautes Études Sci. Publ. Math. No. 8 (1961), 222 pp.
Allen Hatcher, Algebraic topology, Cambridge University Press, Cambridge, 2002.
Morris W. Hirsch and Barry Mazur, Smoothings of piecewise linear manifolds, Princeton University Press,
Princeton, N. J.; University of Tokyo Press, Tokyo, 1974. Annals of Mathematics Studies, No. 80.
J. F. P. Hudson, Piecewise linear topology, University of Chicago Lecture Notes prepared with the assistance of J. L. Shaneson and J. Lees, W. A. Benjamin, Inc., New York-Amsterdam, 1969.
W. Cary Huffman, Imprimitive linear groups generated by elements containing an eigenspace of codimension two, J. Algebra 63 (1980), no. 2, 499–513.
W. Cary Huffman, Linear groups containing an element with an eigenspace of codimension two, J. Algebra
34 (1975), no. 2, 260–287.
W. C. Huffman and D. B. Wales, Linear groups of degree n containing an element with exactly n–2 equal
eigenvalues, Linear and Multilinear Algebra 3 (1975/76), no. 1/2, 53–59.
James E. Humphreys, Reflection groups and Coxeter groups, Cambridge Studies in Advanced Mathematics, vol. 29, Cambridge University Press, Cambridge, 1990.
Sören Illman, Smooth equivariant triangulations of G-manifolds for G a finite group, Math. Ann. 233
(1978), no. 3, 199–220.
M. A. Kervaire and J. W. Milnor, Groups of homotopy spheres, I, Ann. of Math. (2) 77 (1963), 504–537.
B. Kleiner and J. Lott, Notes on Perelman’s papers, arXiv:0605667, (2006).
Christian Lange, When is the underlying space of an orbifold a topological manifold?, preprint, arXiv
(2013).
Christian Lange, Equivariant smoothing of piecewise linear manifolds, submitted (2016).
Christian Lange, Some results on orbifold quotients and related objects, thesis, Universität zu Köln (2016).
Christian Lange and Marina A. Mikhaîlova, Classification of finite groups generated by reflections and
rotations, Transformation Groups, (2016), DOI 10.1007/s00031-016-9385-6.
John Milnor, Lectures on the h-cobordism theorem, Notes by L. Siebenmann and J. Sondow, Princeton
University Press, Princeton, N.J., 1965
John Morgan and Gang Tian, Ricci Flow and the Poincaré Conjecture, Clay Mathematics Monographs,
Volume 3, arXiv:math/0607607v2 (2007)
James R. Munkres, Elementary differential topology, Lectures given at Massachusetts Institute of Technology, Fall, vol. 1961, Princeton University Press, Princeton, N.J., 1966.
G. I. Lehrer, A new proof of Steinberg’s fixed-point theorem, Int. Math. Res. Not. 28 (2004), 1407–1411.
Gustav I. Lehrer and Donald E. Taylor, Unitary reflection groups, Australian Mathematical Society
Lecture Series, vol. 20, Cambridge University Press, Cambridge, 2009.
22
CHRISTIAN LANGE
[33] М. А. Меарчик (девичья фамилия М. А. Михайловой), Конечные гриппы, порождённые псевдоотражениями в четырёхмерном евклидовом пространцтве, Труды киргиз. гос. унив., сер. мат. маук
11 (1976), 66-72. [M. A. Maerchik (maiden name of Mikhaîlova), Finite groups generated by pseudoreflections in four-dimensional Euclidean space, Trudy Kirgiz Gos. Univ. Ser. Mat. Nauk 11 (1976), 66-72.
(Russian)].
[34] М. А. Михайловa, Конечные импримитивные группы, порожленные псевдоотражениями, в сб.:
Исследования по леометрии и алгебре, киргиз. гос. униб., Фрунзе, 1978, 82-93. [M. A. Mikhaîlova,
Finite imprimitive groups generated by pseudoreflections, Studies in geometry and algebra, Kirgiz. Gos.
Univ., Frunze, 1978, 82–93 (Russian)].
[35] М. А. Михайловa, Конечные группы, порождённые псеvбоотражениями, деп. в ВИНИТИ, no. 124882, 1982. [M. A. Mikhaîlova, Finite reducible groups generated by pseudoreflections, deposited at VINITI,
manuscript no. 1248-82, 1982 (Russian).]
[36] М. А. Михайловa, О фаторпространстве по действию конечной группы, порождённой псевдоотражениями, Изв. АН СССР, Сер. матем. 48 (1984), вып. 1, 104–126. Engl. trans.: M. A. Mikhaîlova,
On the quotient space modulo the action of a finite group generated by pseudoreflections, Math. USSRIzvestiya 24 (1985), no. 1, 99–119.
[37] Edwin E. Moise, Geometric topology in dimensions 2 and 3, Springer-Verlag, New York-Heidelberg, 1977.
Graduate Texts in Mathematics, Vol. 47.
[38] G. Perelman, The entropy formula for the Ricci flow and its geometric applications, preprint,
arXiv:0211159, (2002).
[39] G. Perelman, Finite extinction time for the solutions to the Ricci flow, preprint, arXiv:0307245, (2003).
[40] G. Perelman, Ricci flow with surgery on three-manifolds. preprint, arXiv:0303109, (2003).
[41] C. P. Rourke and B. J. Sanderson, Introduction to piecewise-linear topology, Springer-Verlag, New YorkHeidelberg, 1972. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 69.
[42] E. H. Spanier, Algebraic topology, McGraw-Hill, New York, 1966.
[43] Richard P. Stanley Invariants of finite groups and their applications to combinatorics, Bull. Amer. Math.
Soc. (N.S.) 1 (1979), no. 3, 475–511
[44] Robert Steinberg, Differential equations invariant under finite reflection groups, Trans. Amer. Math. Soc.
112 (1964), no. 3, 392–400.
[45] William P. Thurston, Three-dimensional geometry and topology, Vol. 1, Princeton Mathematical Series,
vol. 35, Princeton University Press, Princeton, NJ, 1997.
[46] Jože Vrabec, Polyhedral quotient spaces, Colloq. Math. 62 (1991), no. 1, 145–151.
[47] D. B. Wales, Linear groups containing an involution with two eigenvalues −1, II, J. Algebra 53 (1978),
no. 2, 58–67.
[48] J. Whitehead, On C 1 complexes, Annals of Math. 41 (1940), 809–832.
Christian Lange, Mathematisches Institut der Universität zu Köln, Weyertal 86-90, 50931
Köln, Germany
E-mail address: [email protected]
| 4math.GR
|
arXiv:1803.07105v1 [math.AC] 19 Mar 2018
A New Bound on Hrushovski’s Algorithm for Computing
the Galois Group of a Linear Differential Equation
Mengxiao Sun
Abstract
The complexity of computing the Galois group of a linear differential equation is of general
interest. In a recent work, Feng gave the first degree bound on Hrushovski’s algorithm for
computing the Galois group of a linear differential equation. This bound is the degree bound
of the polynomials used in the first step of the algorithm for finding a proto-Galois group (see
Definition 2.8) and is sextuply exponential in the order of the differential equation. In this
paper, we use Szántó’s algorithm of triangular representation for algebraic sets to analyze the
complexity of computing the Galois group of a linear differential equation and we give a new
bound which is triple exponential in the order of the given differential equation.
keywords: differential Galois groups, linear differential equations, algorithms, triangular sets
1
Introduction
The differential Galois group (see Definition 2.6) is an analogue for a linear differential equation of
the classical Galois group for a polynomial equation. An important application of the differential
Galois group is that a linear differential equation can be solved by integrals, exponentials and
algebraic functions if and only if the connected component of its differential Galois group is solvable
[8, 19]. For example [9, Appendix], the differential Galois group of Bessel’s equation t2 y ′′ + ty ′ +
(t2 − ν 2 )y = 0 over C(t) is isomorphic to SL2 (C) (not solvable) when ν 6∈ 21 + Z. In other
words, Bessel’s equation cannot be solved by integrals, exponentials and algebraic functions unless
ν ∈ 12 + Z. Computing the differential Galois groups would help us determine the existence of
the solutions expressed in terms of elementary functions (integrals, exponentials and algebraic
functions) and understand the algebraic relations among the solutions.
Hrushovski in [5] first proposed an algorithm for computing the differential Galois group of a
general linear differential equation over k(t) where k is a computable algebraically closed field of
characteristic zero. Recently, Feng approached finding a complexity bound of the algorithm in [4],
which is the degree bound of the polynomials used in the first step of the algorithm for finding
a proto-Galois group (see Definition 2.8), but not for the whole algorithm. The bound given by
Feng is sextuply exponential in the order n of the differential equation.
In this paper, we present a triple exponential degree bound using triangular sets instead of
Gröbner bases for representing the algebraic sets. In general, the degrees of defining equations
of a differential Galois group cannot be bounded by a function of n only. For example [13, Ex1
y over C(t) is isomorphic to Z/mZ
ample 1.3.7, page 12], the differential Galois group of y ′ = mt
where m is a positive integer, which implies that the degree of the defining equation xm − 1 is m.
A crucial point of Hrushovski’s algorithm is that one can find a proto-Galois group which is
an algebraic subgroup of GLn (k), provided that the degree bound of the defining equations of the
proto-Galois group is computed. The differential Galois group can then be recovered from the
proto-Galois group (more details in [4, 5]). Therefore, a bound for the proto-Galois group plays
an important role in determining the complexity of Hrushovski’s algorithm. Following Feng’s
approach, we prove that such a proto-Galois group exists by constructing a family F of algebraic
subgroups such that the identity component of any algebraic subgroup H ′ ⊆ GLn (k) is contained
1
in some H of F and [H ′ H : H] is uniformly bounded. We also prove that the degrees of the
defining equations of any element of F are bounded and provide a numerical bound d¯ depending
on the order n of the given differential equation. This is stated as Theorem 3.1. Then by collecting
¯ we obtain a
the algebraic subgroups H̄ such that there is some H of F such that [H̄ : H] ≤ d,
family F̄ of algebraic subgroups in which one can always find a proto-Galois group for any linear
differential equation. Moreover, we give a numerical degree bound of the defining equations of any
algebraic subgroup of F̄. This is stated as Corollary 3.1.
Using degrees of defining equations of algebraic subgroups to bound F̄ , one needs an upper
degree bound of the defining equations of the algebraic subgroups of F and an upper bound
of [H̄ : H]. Hence, a double-exponential degree bound for computing Gröbner bases would be
involved if one chooses to represent an algebraic subgroup by the generating set of its defining
ideal (generated by the defining equations). In order to give a better bound, we represent an
algebraic subgroup by the triangular sets (see Definition 3.2) instead of the generating set in the
process of constructing F . In such a process, we need to take the differences between Gröbner
bases and triangular sets into account. We apply Szántó’s modified Wu-Ritt type decomposition
algorithm [15, 16] which has been proved to be more efficient than computing a Gröbner basis and
make use of the numerical bound for Szántó’s algorithm [1] to adapt to the complexity analysis
of Hrushovski’s algorithm. In doing this, we are able to avoid working with Gröbner bases to get
a better bound of the degrees of the defining equations of the algebraic subgroups of F which is
triple exponential in the order n of the given differential equation. Additionally, we are able to
not increase the degree bound of the defining equations of the algebraic subgroups of F̄ . Each
element H̄ of F̄ is a union of at most d¯ cosets of some element H of F . The degree bound for the
generating set of the ideal generated by the defining equations of H̄ would be raised to an exponent
¯ which results in a big increase on the degrees of the defining equations of the algebraic
at most d,
subgroups of F̄ . However, this issue has been resolved when expressing the algebraic subgroups
by triangular sets. More details can be found in sections 2and 3.
Besides Hrushovski’s general algorithm, there are other algorithmic results in the Galois theory
of linear differential equations. Kovacic in [10] presented an algorithm for computing the Galois
group of a second order linear differential equation. The Galois groups of second and third order
linear differential equations were studied by Singer and Ulmer in [14]. Compoint and Singer in [2]
proposed an algorithm for computing the Galois group if the differential equation is completely
reducible. The numeric-symbolic computation of differential Galois groups was presented by van
der Hoeven in [17]. For more details on the differential Galois theory from an algorithmic point of
view, readers are referred to [18] and [13].
This paper is organized as follows. In sections 2.1 and 2.2, we introduce the notations, definitions and facts from triangular sets and differential Galois groups. In section 3.1, we state and
prove the preparation lemmas which we use in analyzing the complexity of the algorithm. In section 3.2, we present and prove the new complexity bound of Hrushovski’s algorithm. In section 4,
we compare our bounds when n = 2 with the ones in [4, Proposition B.11, Proposition B.14].
2
2.1
Preliminaries
Triangular sets and Szántó’s algorithm
Definition 2.1. Fix a monomial ordering with x1 < x2 < · · · < xn in a polynomial ring
k[x1 , . . . , xn ]. Let f ∈ k[x1 , . . . , xn ]. Then class (f ) denotes the highest indeterminate in f .
Definition 2.2. Fix a monomial ordering with x1 < x2 < · · · < xn in a polynomial ring
k[x1 , . . . , xn ]. A sequence of polynomials {g1 , . . . , gm } is called a triangular set if class (gi ) <
class (gj ) for all i < j.
The pseudo division is a method of division for multivariate polynomials which is a generalization of the method of division for univariate polynomials.
2
Let f ∈ k[x1 , . . . , xn ] and G = {g1 , . . . , gm } be a triangular set in k[x1 , . . . , xn ]. We denote the
leading coefficient of gi by lc (gi ). There exist polynomials q1 , . . . , qm and α1 , . . . , αm ∈ N such
that lc (g1 )α1 . . . lc (gm )αm f = Σm
i=1 qi gi + f0 where degxji (f0 ) < degxji (gi ) if class (gi ) = xji for
1 ≤ i ≤ m. We call f0 the pseudo remainder of f by G, denoted by prem (f, G). If f = prem (f, G),
then we say that f is reduced modulo G.
Definition 2.3. We say that an ideal I ⊆ k[x1 , . . . , xn ] is represented by a triangular set G if
I = {f ∈ k[x1 , . . . , xn ] : prem (f, G) = 0}, denoted by I = rep (G).
√
Definition 2.4. We say that an ideal I ⊆ k[x1 , . . . , xn ] has a triangular representation if I is
expressed by an intersection of radical ideals Ii such that for each i, Ii = rep (Gi ) where Gi is a
triangular set in Ii .
Szántó in [15] and [16] presented a Wu-Ritt type decomposition algorithm expressing the radical
of an ideal as an intersection of radical ideals represented by unmixed triangular sets. An unmixed
triangular set is a triangular set with some certain conditions. For our purposes, we state the
following theorem without emphasizing the "unmixed" property of triangular sets.
Theorem 2.1. [16, T heorem 4.1.7] Let I ⊆ k[x1 , . . . , xn ] be an ideal. Then there is an algorithm
which computes
a triangular representation of I, that is computing a family of triangular sets Gi
√
T
m
such that I = i=1 Ii where Ii = rep (Gi ) for each i.
Definition 2.5. An ideal I ⊆ k[x1 , . . . , xn ] is bounded by d if any polynomial pj in the generating
set {p1 , . . . , pm } of I has degree not greater than d.
In order to get a numerical complexity bound for Hrushovski’s algorithm, we need a numerical
complexity bound for Szántó’s algorithm, which is stated as the following theorem.
Theorem 2.2. [1, T heorem 3.5] Suppose that n > 1 and I ⊆ k[x1 , . . . , xn ] is an ideal bounded
by d. Then all the polynomials appearing during the computations and in the output of Szántó’s
3
algorithm have degrees not greater than nd5.5n .
2.2
Differential Galois groups and Hrushovski’s algorithm
We consider a linear differential equation in the matrix form:
δ(Y ) = AY
(1)
where Y is a vector containing n unknowns and A is an n × n matrix with entries in k(t). Denote
d
the Picard-Vessiot extention field of the differential field k(t) by K with the derivation δ = dt
and
the solution space of (1) by V in K. Let F ∈ GLn (K) be a fundamental matrix of (1). Let GL(V )
be the group of automorphisms of the solution space V . Then there is a group isomorphism ΦF :
GL(V ) → GLn (k) sending σ ∈ GL(V ) to Mσ ∈ GLn (k) where F Mσ = σ(F ).
Definition 2.6. The Galois group G of (1) is the group of k(t)-automorphisms of K which commutes with the derivation and fixes k(t) pointwise.
Definition 2.7. An algebraic subgroup H of GLn (k) is bounded by d if there exist finitely many
polynomials p1 , . . . , pm ∈ k[xi,j ]1<i,j<n of degrees not greater than d such that H = zero (p1 , . . . , pm )∩
GLn (k).
Let H ⊆ GLn (k) be an algebraic subgroup. Let H 0 be the identity component of H and
ΦF (G)0 be identity component of ΦF (G). Let (H 0 )t be the intersection of kernels of all characters
of H 0 .
The definition of a proto-Galois group of (1) was introduced by Feng in [4], which is as follows:
3
Definition 2.8 ([4, Definition 1.1]). If there is an algebraic subgroup H of GLn (k) bounded by an
integer d3 such that
(H 0 )t E ΦF (G)0 ⊆ ΦF (G) ⊆ H,
then H is called a proto-Galois group of (1).
1
y over C(t). The differExample 2.1. Consider the first order linear differential equation y ′ = 3t
√
√
1
3
3
1
ential Galois group is the subgroup {1, − 2 + 2 i, − 2 − 2 i} ⊆ C∗ , where C∗ is the multiplicative
group of complex numbers [13, Example 1.3.7, page 12]. The identity component of C∗ is itself.
The intersection of kernels of all characters of C∗ is trivial because the identity map of C∗ is a
character. So in this case C∗ is a proto-Galois group.
Example 2.2. Consider the Airy equation y ′′ = ty over C(t). The differential Galois group is
the subgroup SL2 (C) ⊆ GL2 (C) [19, Example 8.15, page 250]. The identity component of SL2 (C)
is itself because SL2 (C) is connected. The identity component of GL2 (C) is itself because GL2 (C)
is connected. A character φ of GL2 (C) is of the form ∀g ∈ GL2 (C) φ(g) = (det (g))n where det
is the determinant map and n ∈ N. So the intersection of kernels of all characters of GL2 (C) is
SL2 (C). So GL2 (C) is a proto-Galois group. The intersection of kernels of all characters of the
identity component of the differential Galois group is a normal subgroup of the identity component
of the differential Galois group, so the differential Galois group itself is also a proto-Galois group.
Example 2.3. Consider the second order linear differential equation y ′′ + 1t y ′ = 0 over C(t). The
differential Galois group is
1 c
:c∈C
0 1
[3, Example 4.1, page 90]. The identity component of GL2 (C) is itself because GL2 (C) is connected. A character φ of GL2 (C) is of the form ∀g ∈ GL2 (C) φ(g) = (det (g))n where det is the
determinant map and n ∈ N. So the intersection of kernels of all characters of GL2 (C) is SL2 (C)
which is not contained in the identity component of the differential Galois group. So in this case
GL2 (C) cannot be a proto-Galois group. The identity component of SL2 (C) is itself because SL2 (C)
is connected. The intersection of kernels of all characters of SL2 (C) is itself which is not contained in the identity component of the differential Galois group. So in this case SL2 (C) cannot be
a proto-Galois group. The intersection of kernels of all characters of the identity component of the
differential Galois group is a normal subgroup of the identity component of the differential Galois
group, so the differential Galois group itself is a proto-Galois group.
Example 2.4. Consider the second order linear differential equation y ′′ − 2ty ′ − 2y = 0 over C(t).
The differential Galois group is
a c
: a, b, c ∈ C, ab 6= 0
0 b
[11, Example 6.10, pages 81, 82, 83]. The identity component of GL2 (C) is itself because GL2 (C)
is connected. A character φ of GL2 (C) is of the form ∀g ∈ GL2 (C) φ(g) = (det (g))n where det
is the determinant map and n ∈ N. So the intersection of kernels of all characters of GL2 (C) is
SL2 (C) which is not contained in the identity component of the differential Galois group. So in
this case GL2 (C) cannot be a proto-Galois group. The differential Galois group is not a subgroup
of SL2 (C), so in this case SL2 (C) cannot be a proto-Galois group. The differential Galois group
itself is a proto-Galois group because the intersection of kernels of all characters of the identity
component of the differential Galois group is a normal subgroup of the identity component of the
differential Galois group.
Remark. The proto-Galois group of a linear differential equation is not unique. In addition, as
shown in the above examples, the proto-Galois group can be far from the differential Galois group.
4
Hrushovski in [5, Corollary 3.7] proved that such an integer d3 is computable. Feng in [4,
Propostion B.14] gave the first explicit bound for d3 which is sextuply exponential in the order n
of the given linear differential equation.
To understand the key role of the integer d3 in analyzing the complexity of Hrushovski’s algorithm, we separate the algorithm in three main steps following the way in which Feng in [4]
described it.
Definition 2.9. Let Ṽ = {F h : h ∈ GLn (k)}. A subset V0 of Ṽ is defined by finitely many
polynomials p1 , . . . , pm if V0 = zero (p1 , . . . , pm ) ∩ Ṽ where zero (p1 , . . . , pm ) denotes the zero set of
{p1 , . . . , pm } in k n×n . If p1 , . . . , pm have coefficients in k, we say that V0 is k-definable subset of
Ṽ .
• In the first step, we compute a proto-Galois group H of (1) bounded by d3 . The existence of
H is guaranteed by [5, Corollary 3.7]. Let Nd3 (Ṽ ) be the set of all subsets of Ṽ defined by
finitely many polynomials of degrees not greater than d3 . Then one can compute H by the
intersection of the stabilizers of k-definable elements in Nd3 (Ṽ ).
• In the second step, we compute the identity component (ΦF (G))0 of ΦF (G). Let χ1 , . . . , χl
0
be the generators of the character group of ΦF (G) . Let k̂ be an algebraic extension of
k(t). Then χ ΦF (G)0 is the Galois group of some exponential extension K̂ of k̂ where
χ = (χ1 , . . . , χl ). K̂ can be obtained by computing hyperexponential solutions of some
0
symmetric power system of (1). ΦF (G) can be found by the pre-image of χ ΦF (G)0 in
0
ΦF (H) .
• In the last step, we compute the differential Galois group G of (1). Let G 0 be the pre-image
0
of ΦF (G) . Find a Galois extension kG of k(t) and a kG -definable subset VkG of Ṽ such
that G 0 = stab (VkG ) where stab (VkG ) is the stabilizer of VkG . Then
G=
m
[
i=0
{σ ∈ GL(V )|σ(VkG ) = Vi }
where Vi is the orbit of VkG under the action of Gal kG /k(t) .
From the first step of the algorithm, we can see that the integer d3 determines the complexity
of computing a proto-Galois group. The differential Galois group is obtained by recovering the
proto-Galois group in the next two steps. Therefore d3 plays an important role in determining the
complexity of the whole algorithm.
3
3.1
Complexity bound on Hrushovski’s algorithm
Preparation
√
Definition
3.1. An ideal I ⊆ k[x1 , . . . , xn ] has a triangular representation bounded by d if I =
Tm
i=1 Ii where, for each i, Ii = rep (Gi ) and Gi is a triangular set in Ii such that every polynomial
in Gi has degree not greater than d.
Definition 3.2. An algebraic subgroup H ⊆ GLn (k) is said to have a triangular representation if
the ideal generated by the defining equations of H has a triangular representation.
Let I ⊆ k[x1 , . . . , xn ] be an ideal. In [4, Proposition B.2], Feng gave a degree bound for
I ∩ k[x1 , . . . , xr ] which is double-exponential in n using the computation of Gröbner bases. In the
following lemma, we give a degree bound for the triangular representation of I ∩k[x1 , . . . , xr ] which
is polynomial exponential in n.
5
Lemma 3.1. Assume that n > 1. Let I ⊆ k[x1 , . . . , xn ] be an ideal bounded by d and 1 ≤ r ≤ n.
3
Then I ∩ k[x1 , . . . , xr ] has a triangular representation bounded by nd5.5n .
√
T
Proof. Assume that I = i Ii is a triangular representation of I where Ii = rep (Gi ) and Gi is a
triangular set of Ii . Since
p
√
I ∩ k[x1 , . . . , xr ] = I ∩ k[x1 , . . . , xr ],
it suffices to show that for each i
(2)
rep Gi ∩ k[x1 , . . . , xn ] ∩ k[x1 , . . . , xr ] = rep (Gi ) ∩ k[x1 , . . . , xr ].
Let g ∈ LHS of (2). Then prem g, Gi ∩ k[x1 , . . . , xr ] = 0. If Gi ⊆ k[x1 , . . . , xr ], then
prem (g, Gi ) = prem g, Gi ∩ k[x1 , . . . , xr ] .
So prem (g, Gi ) = 0. If Gi 6⊆ k[x1 , . . . , xr ], suppose that Gi = {gi,1 , . . . , gi,s } and gi,j is the smallest
element in Gi containing xr+1 term. Then
Gi ∩ [x1 , . . . , xr ] = {gi,1 , . . . , gi,j−1 }.
Since g ∈ k[x1 , . . . , xr ], g is reduced modulo gi,j , . . . , gi,s . Then prem (g, Gi ) = 0. So g ∈ RHS of
(2). Let f ∈ RHS of (2). Then f ∈ k[x1 , . . . , xr ] and prem (f, Gi ) = 0. Since f ∈ k[x1 , . . . , xr ], f
is reduced modulo polynomials containing terms larger than xr in Gi . So
prem f, Gi ∩ k[x1 , . . . , xr ] = 0.
So f ∈ LHS of (2). Therefore, I ∩ k[x1 , . . . , xr ] has a triangular representation which is
\
p
Ii′
I ∩ k[x1 , . . . , xr ] =
i
where
Ii′ = rep Gi ∩ k[x1 , . . . , xr ] .
3
By Theorem 2.1, the triangular representation of I ∩ k[x1 , . . . , xr ] is bounded by nd5.5n .
Lemma 3.2. Let I, J ⊆ k[x1 , . . . , xn ] be ideals. Assume that I and J have triangular representations bounded by d. Then IJ has a triangular representation bounded by d.
√
√
T
T
Proof. Suppose that I = i Ii and J = j Jj are triangular representations of I and J. Then
√
√
√
T
T
IJ = I ∩ J =
Ii ∩
Jj . So if Ii = rep (GIi ) and Jj = rep (GJj ) for some triangular
√
T
T
rep (GJj ) . Therefore, IJ has a triangular
rep GIi ) ∩
sets GIi and GJj , then IJ =
representation bounded by d.
Definition 3.3. An algebraic subgroup H in GLn (k) is said to have a triangular representation
bounded by d if the ideal generated by the defining equations of H has a triangular representation
bounded by d.
Definition 3.4. We say that a family F of algebraic subgroups in GLn (k) is represented by a
family of triangular sets if any H ∈ F has a triangular representation, and F is bounded by d if
the triangular representations of any H ∈ F are bounded by d.
6
Let H ⊆ GLn (k) be an algebraic subgroup. Let τ : H −→ GLl (k) be a homomorphism where
P
l is a positive integer. Assume that τ = ( Qi,j ), where P and Q are polynomials with coefficients in
k and 1 ≤ i, j ≤ l. The homomorphism τ is said to be bounded by d if the polynomials Pi,j and
Q have degrees not greater than d.
for the generating set of the ideal generated
In [4, Lemma B.5], Feng gave a degree bound
by the defining equations of τ −1 H ′ ∩ τ (H) where H ⊆ GLn (k) and H ′ ⊆ GLl (k). We use
a similar argument
to give in the following lemma a bound for the triangular representation of
τ −1 H ′ ∩ τ (H) .
Lemma 3.3. Assume that n > 1. Let H ⊆ GLn (k) be an algebraic subgroup whose triangular
representation bounded by d and H ′ ⊆ GLl (k) be an algebraic subgroup whose triangular representation bounded by d′ . Assume that the homomorphism τ : H −→ GLl (k) is bounded by m. Then
5.5n3
τ −1 H ′ ∩ τ (H) has a triangular representation bounded by n max (d, md′ )
.
Proof. Let I(H) be the ideal generated by the defining equations of H and I(H ′ ) be the ideal
generated by the defining equations of H ′ . Let X be the set of indeterminates xα,β , 1 ≤ α, β ≤ n
P
and Y be the set of indeterminates yi,j , 1 ≤ i, j ≤ l. Assume that τ = ( Qi,j )1≤i,j≤l where Pi,j and
Q are polynomials in k[X]. Assume that H has a triangular representation
\
p
I(H) =
rep (Gr )
r
and H ′ has a triangular representation
\
p
I(H ′ ) =
rep (Fw ),
w
S
where Gr are triangular sets in k[X] and Fw are triangular sets in k[Y ]. Since zero ( w Fw ) =
H ′ , composing every polynomial in each Fw with τ S
and clearing the denominators, we
S can get
−1
′
E
)
=
τ
(H
).
Since
zero
(
the sets ES
of
polynomials
in
k[X]
such
that
zero
(
w
w w
Sr Gr ) =
S
.
Let
J
be
the
ideal
generated
by
(
H, zero ( r Gr ) ∪ ( w Ew ) = τ −1 H ′ ∩ τ (H)
r Gr ) ∪
S
−1
′
H ∩ τ (H) . Since the degrees of polynomials in Gr are not
( w Ew ) . Thus, zero (J) = τ
greater than d and the degrees of polynomials in Ew are not greater than md′ , by Theorem 2.2, J
5.5n3
has a triangular representation bounded by n max (d, md′ )
. That is, τ −1 H ′ ∩ τ (H) has a
5.5n3
triangular representation bounded by n max (d, md′ )
.
In [4, Proposition B.6], Feng uniformly bounded the homomorphisms defined in the following
lemma by considering the bound for the generating set of the ideal generated by the defining
equations of an algebraic subgroup. Instead, we present a bound for such homomorphisms by
making use of the bound for the triangular representation of an algebraic subgroup. The proof is
similar to the one in [4, Proposition B.6].
Lemma 3.4. Assume that n > 1. Let H and H ′ be algebraic subgroups of GLn (k) such that HEH ′ .
Assume that H has a triangular representation bounded by d. Then there exists a homomorphism
τH ′ ,H : H ′ −→ GLd∗ (k)
bounded by n∗ with ker(τH ′ ,H ) = H, d∗ = max
i
(n2 +d)2
d
and n∗ = d∗ d
i
n2 +d
d
.
Proof. The existence of such a homomorphism is guaranteed by [6, Theorem, page 82]. Let G(H)
be the family of triangular sets in a triangular representation of H. Then G(H) is a k-vector space
7
with a finite dimension. Let k[xi,j ]≤d be the set of polynomials of degrees not greater than d where
1 ≤ i, j ≤ n. Let
l
^
k[xi,j ]≤d , 1 ≤ i, j ≤ n
E=
which is the lth exterior power of k[xi,j ]≤d . Since k[xi,j ]≤d is a k-vector space with dimension
2
n2 +d
(n d+d) and Vl C(H) = kv for some v ∈ E where Vl C(H) is the lth exterior
,
dim
(E)
=
k
d
l
power of C(H). By a similar argument in the proof of [4, Lemma B.6], we can construct a desired
homomorphism bounded by n∗ .
Let U be a subgroup generated by unipotent elements of GLn (k). In [4, Lemma B.8], Feng
gave a degree bound for U which is double exponential in n. In the following lemma, we bound
the triangular representation of U . The bound we give is polynomial exponential in n.
Lemma 3.5. Assume that n > 1. Let U be a subgroup generated by unipotent elements of GLn (k).
Then U has a triangular representation bounded by
148.5n6
.
3n2 2n2 (n − 1)
Proof. By [6, Lemma C, page 96], any one-dimensional subgroup H generated by unipotent elements of GLn (k) has the form
H = {In + M x +
M n−1 n−1
M2 2
x + ··· +
x
: x ∈ C}
2!
(n − 1)!
where M ∈ Matn (k) with M n = 0. By [6, Proposition, page 55], U is a product of at most
2 dim (U ) one-dimensional subgroups generated by unipotent elements. Hence,
2 dim (U)
U=
Y
Hi
i=1
M n−1
M2
i
where Hi = {In + Mi xi + 2!i x2i + · · · + (n−1)!
xin−1 : xi ∈ C} is a one-dimensional subgroup
generated by unipotent elements of GLn (k) and Mi ∈ Matn (k) with Min = 0. Since dim (U ) ≤ n2 ,
the defining equations of U contain at most 3n2 variables and have degrees not greater than
2n2 (n − 1). By Theorem 2.2, the ideal generated by the defining equations of U has a triangular
representation bounded by
148.5n6
5.5(3n2 )3
.
= 3n2 2n2 (n − 1)
3n2 2n2 (n − 1)
Jordan in [7] proved that there exists a positive integer J(n) depending on n such that every
finite subgroup of GLn (k) contains a normal abelian subgroup of index at most J(n). Schur in
[12] provided an explicit bound which is
J(n) ≤
√
√
2n2
2n2
8n + 1
−
8n − 1
.
We use Schur’s bound in our computations. Assume that n > 1. Let
148.5n6
D = 3n2 2n2 (n − 1)
,
d1 = max
i
8
n2 +D
D
i
2
,
(3)
(4)
d2 = d1 D
d2 (d21
d3 = n
2
n +D
,
D
(5)
2
2 !5.5n3
d1 + 1
,
+ 1) max
i
i
and
(6)
2
2 !
d1 + 1
.
max
i
i
d¯ = J
(7)
Next we give numerical bounds for D, d1 , d2 , d3 and d¯ which will be used in the following theorems.
Since
148.5n6
6
≤ 3n2 (2n3 )148.5n
D = 3n2 2n2 (n − 1)
and
2
n2
6
2
2
8
n +D
e(n2 + D)
≤
≤ (e + 3e(2n3 )148.5n )n ≤ 18n (2n3 )148.5n ,
n2
n2
n2 +D
D
d1 ≤ 2(
3 149n8
d2 ≤ 4(2n
)
) 2 ≤ 218n
6
2
(2n3 )148.5n
8
2
2
3 149n8
≤ 2(2n
8
)
2
3 149n8
≤ 4(2n
3 149n8
2
3n2 (2n3 )148.5n 18n (2n3 )148.5n = 3n2 18n 4(2n
8
6
3 149n8
3 149n8
≤n (2n3 )149n +149n 4(2n )
8(2n )
+1 4
3 149n8
8
6
3 149n8
8(2n )
4
≤n 2(2n3 )149n +149n 4(2n )
8(2n
8(2n
3 )149n8
3 )149n8
,
8
)
2
8
6
3 149n8
3 149n8
d3 ≤n 3n2 18n (2n3 )148.5n +148.5n 4(2n )
8(2n )
+1 4
)
(2n3 )(148.5n
8(2n
3 )149n8
5.5n3
+148.5n6 )
,
5.5n3
+1
+1
5.5n3
+1
5.5n3
8
8
3 149n8
3 149n8 +149n6 (2n3 )149n (2n3 )149n 8(2n )
≤n 8(2n )
4
8
4
8
(2n3 )149n
≤n 88
and
5.5n3
8
3 (2n3 )149n
= n85.5n
8
,
p
p
2n2
2n2
2
2
8 · 4d1 +1 + 1
8 · 4d1 +1 − 1
−
8
8
√
√
2n2
2n2
(2n3 )149n
(2n3 )149n
+1
−
−1
32 · 28
≤ 32 · 28
8
8
√
(2n3 )149n 2n2
2
2 (2n3 )149n
= 210n 4n 8
.
≤ 2 32 · 28
d¯ ≤
3.2
Complexity of Hrushovski’s Algorithm
In this section, when we say that a family F of algebraic subgroups of GLn (k) is bounded by
an integer we mean that the triangular representations of all elements in F are bounded by that
integer. We prove the following theorem and corollaries following the way in which Feng proved
[4, Proposition B.11, Lemma B.12, Lemma B.13]. But in order to improve the bounds, we replace
the use of Gröbner bases with the triangular representations. Our main result is stated as the
following theorem.
9
Theorem 3.1. Assume that n > 1. There is an integer
2
2 (2n
d¯ ≤ 210n 4n 8
3 )149n8
and a family F of algebraic subgroups of GLn (k) whose triangular representations are bounded by
8
3 (2n3 )149n
d3 ≤ n85.5n
8
with the following properties: for every algebraic subgroup H ′ ⊆ GLn (k), there exists an algebraic
subgroup H of F such that
(a) (H ′ )◦ ⊆ H
(b) H E H ′ H ⊆ GLn (k)
(c) [H ′ : H ∩ H ′ ] = [H ′ H : H] ≤ d¯
(d) Every unipotent element of H is in (H ′ )◦
where (H ′ )◦ is the identity component of H ′ .
Proof. In the first case we assume that H ′ is a finite subgroup in GLn (k). Since every finite
subgroup of GLn (k) contains a normal abelian subgroup of index at most J(n), we choose such
a normal abelian subgroup H̃ ′ ⊆ H ′ . H̃ ′ is diagonalizable, so H̃ ′ is in some maximal torus of
GLn (k). Let H be the intersection of maximal tori containing H̃ ′ in GLn (k). We prove that H
satisfies (a) − (d) for H ′ . (a) is true because of the construction of H. (b) is true because H
normalizes H ′ . (c) is true because
[H ′ H : H] = [H ′ : H ∩ H ′ ] ≤ [H ′ : H̃ ′ ] ≤ J(n).
So we can choose d¯ = J(n). (d) is true because there is only one unipotent element of H which is
the identity. Let F be the family of all the intersections of maximal tori in GLn (k). Then F is
the desired family of algebraic subgroups of GLn (k) with d3 = 1.
In the second case we assume that H ′ is a subgroup whose identity component is a torus. Let
T be the intersection of all maximal tori containing (H ′ )◦ in GLn (k). Then T has a triangular
representation bounded by 1. Let S be the normalizer of T in GLn (k). By Lemma 3.4, there is a
homomorphism
τS,T : S −→ GLn′ (k)
bounded by
2
2
n +1
(n + 1) max
i
i
2
2
such that ker (τS,T ) = T and n′ = maxi n i+1 . Since the identity component of H ′ is contained
in T , τS,T (H ′ ) is a finite subgroup of GLn′ (k). Let F1 be the family of all the intersections of
maximal tori of GLn′ (k). By the first case, there exists HF1 ∈ F1 such that (a) − (c) are true for
τS,T (H ′ ) with d¯ = J(n′ ). Let
−1
H = τS,T
τS,T (S) ∩ HF1 .
2
We prove that (a) − (d) are true for H and H ′ . Since the identity component (H ′ )◦ of H ′ is a
torus and T ⊆ H, (H ′ )◦ ⊆ H. This proves (a). Let h′ ∈ H ′ . Then
τS,T (h′ Hh′−1 ) = τS,T (h′ ) τS,T (S) ∩ HF2 τS,T (h′−1 ).
Since HF1 E τS,T (H ′ )HF1 ,
τS,T (h′ ) τS,T (S) ∩ HF1 τS,T (h′−1 ) = τS,T (S) ∩ HF1 = τS,T (H).
10
So h′ Hh′−1 ⊆ H and H E H ′ . Hence, H E H ′ H ⊆ GLn (k). This proves (b). Since H ′ ⊆ S and
[τS,T (H ′ ) : HF1 ∩ τS,T (H ′ )] ≤ J(n′ ),
[H ′ : H ∩ H ′ ] = [H ′ H : H] = [τS,T (H ′ H) : τS,T (H)] = [τS,T (H ′ )τS,T (H) : τS,T (H)]
= [τS,T (H ′ ) : τS,T (H) ∩ τS,T (H ′ )] = [τS,T (H ′ ) : HF1 ∩ τS,T (H ′ )] ≤ J(n′ ).
2 2
We can choose d¯ = J(n′ ) = J maxi n i+1
. This proves (c). Let h ∈ H be a unipotent
element. Then τS,T (h) is a unipotent element in HF1 . Since every element in HF1 is semi-simple,
τS,T (h) = 1. So h must be in ker (τS,T ) = T . By the definition of T , T is in some torus of GLn (k).
So h = 1. Hence, every unipotent element of H is in (H ′ )◦ . This proves (d). By Lemma 3.3, H
has a triangular representation bounded by
5.5n3
2
n(n + 1)
2
11n3
n +1
.
max
i
i
Let F be the family of such subgroups H. Then F is the desired family with
3
d3 ≤ n(n2 + 1)5.5n max
i
2
11n3
n +1
.
i
The general case is proved as follows. Let Hu′ be the intersection of kernels of all characters of
(H ′ )◦ . (H ′ )◦ is a connected subgroup of GLn (k), so Hu′ is generated by all unipotent elements by
[4, Lemma B.10]. By Lemma 3.5, Hu′ has a triangular representation bounded by D. Let N be
the normalizer of Hu′ in GLn (k). By Lemma 3.4, there exists a homomorphism
τN,Hu′ : N −→ GLd1 (k)
bounded by
n2 + D
d2 = d1 D
D
such that ker(τN,Hu′ ) = Hu′ and
d1 = max
i
n2 +D
D
i
2
.
The identity component of τN,Hu′ (H ′ ) is a torus in GLd1 (k), by the second case, there exists
H ′′ ⊆ GLd1 (k) whose triangular representation bounded by
(d21
2
2
d1 + 1
+ 1) max
i
i
such that (a) − (d) are true for τN,Hu′ (H ′ ) with
d¯ = J
2
2 !
8
2
2 (2n3 )149n
d1 + 1
≤ 210n 4n 8
max
.
i
i
Let
H = τN,Hu′ −1 H ′′ ∩ τN,Hu′ (N ) .
By Lemma 3.3, H has a triangular representation bounded by
2
2 !5.5n3
8
3 (2n3 )149n
d
+
1
1
≤ n85.5n 8
d3 = n d2 (d21 + 1) max
.
i
i
11
By a similar argument in the proof of [4, Proposition B.11], (a) − (d) are true for H and H ′ . Let
F be the family of such algebraic subgroups H. Then F is the desired family with
3 (2n3 )149n
d3 ≤ n85.5n
8
8
.
Corollary 3.1. Assume that n > 1. There exists a family F̄ of algebraic subgroups of GLn (k)
whose triangular representations are bounded by
8
3 (2n3 )149n
d3 ≤ n85.5n
8
such that for any algebraic subgroup H ′ ⊆ GLn (k) there exists H̄ of F̄ such that H ′ ⊆ H̄ and every
unipotent element of H̄ is in (H ′ )◦ .
¯ Let H ′ ⊆ GLn (k)
Proof. Let F̄ = {H̄ : there exists H ∈ F such that H E H̄ and [H̄ : H] ≤ d}.
be an algebraic subgroup. By Theorem 3.1, there is an H ∈ F such that (a) − (d) are true for H
and H ′ . Let H̄ = H ′ H. By (b) and (c) in Theorem 3.1, H̄ ∈ F̄ . By (d) in Theorem 3.1, every
unipotent element of H is in (H ′ )◦ . Since every unipotent element of H̄ is in H̄ ◦ ⊆ H ◦ . Hence,
every unipotent element of H̄ is in (H ′ )◦ . Since H̄ is the union of the cosets of some element in
F and every element of F has a triangular representation bounded by d3 , by Lemma 3.2, we have
that H̄ has a triangular representation bounded by
8
3 (2n3 )149n
d3 ≤ n85.5n
8
8
3 (2n3 )149n
Therefore, F̄ is bounded by d3 ≤ n85.5n
8
.
.
Corollary 3.2. Let F̄ be the family in Corollary 3.1. Then for any algebraic subgroup H ′ ⊆
GLn (k), there exists H̄ of F̄ such that
(H̄ ◦ )t E (H ′ )◦ ⊆ H ′ ⊆ H̄.
Proof. By Corollary 3.1, there exists H̄ of F̄ such that H ′ ⊆ H̄ and every unipotent element of
H̄ is in (H ′ )◦ . Since H̄ ◦ is a connected subgroup of GLn (k), (H̄ ◦ )t is generated by all unipotent
elements in H̄ ◦ by [4, Lemma B.10]. Since (H̄ ◦ )t E H̄ ◦ and every unipotent element in H̄ ◦ is in
(H ′ )◦ , (H ◦ )t E (H ′ )◦ . Therefore, (H̄ ◦ )t E (H ′ )◦ ⊆ H ′ ⊆ H̄.
Remark. Since the differential Galois group of (1) is an algebraic subgroup in GLn (k), by Corollary 3.2, there exists an algebraic subgroup H̄ bounded by d3 such that (H̄ ◦ )t E (H ′ )◦ ⊆ H ′ ⊆ H̄.
By the definition of the proto-Galois group, this algebraic subgroup H̄ is a proto-Galois group of
(1).
4
Comparison
We compute d¯ and d3 explicitly for n = 2. We plug in n = 2 to the equations (3), (4), (5), (6),
and (7) instead of the formulas in Theorem 3.1 and Corollary 3.1 to do calculations, which would
give us more refined bounds. Feng in [4] roughly estimated that d¯ is quintuply exponential in n
and d3 is sextuply exponential in n, but he did not give numerical bounds for them. In order to
compare our bounds with the ones in [4], we also give numerical bounds in [4, Proposition 11,
Proposition 14]. In [4, Proposition 11, Proposition 14], d3 is denoted as d˜ and d¯ is denoted as I(n)
respectively. The numerical bounds of d˜ and I(n) are as follows:
d˜ ≤ 32
2
2(24n )
2(2n)
22
(12n2 )
(2n)2
, I(n) ≤ 4
12
22
.
When n = 2,
218
and
2
d¯ ≤ 22
17
22
d3 ≤ 22
296
22
, I(n) ≤ 22
194
22
22
, d˜ ≤ 22
,
.
Acknowledgement. This work has been partially supported by the NSF grants CCF-0952591,
CCF-1563942, and DMS-1413859 and by PSC-CUNY grant #60098-00 48. I am grateful to Eli
Amzallag and Alexey Ovchinnikov for discussions on this work.
References
[1] E. Amzallag, G. Pogudin, M. Sun, and N. T. Vo. Complexity of triangular representations of
algebraic sets. Preprint, 2016. URL https://arxiv.org/abs/1609.09824.
[2] E. Compoint and M. F. Singer. Computing Galois groups of completely reducible differential equations. Journal of Symbolic Computation, 28(4-5):473–494, 1999. URL
https://dx.doi.org/10.1006/jsco.1999.0311.
[3] T. Crespo, Z. Hajto, and J. J. M. Ruiz. Introduction to differential Galois theory.
Wydawnictwo PK, Cracow, Poland, 2007.
[4] R. Feng.
Hrushovski’s algorithm for computing the Galois group of a linear
differential equation.
Advances in Applied Mathematics, 65:1–37, 2015.
URL
http://dx.doi.org/10.1016/j.aam.2015.01.001.
[5] E. Hrushovski. Computing the Galois group of a linear differential equation. Banach Center
Publications, 58(1):97–138, 2002. URL https://dx.doi.org/10.4064/bc58-0-9.
[6] J. E. Humphreys. Linear Algebraic Groups. Springer-Verlag, New York, 1975. URL
http://dx.doi.org/10.1007/978-1-4684-9443-3.
[7] C. Jordan.
Mémoire sur les équations différentielles linéaires à intégrale algèbriques. Journal für die reine und angewandte Mathematik, 84:89–215, 1877. URL
http://eudml.org/doc/148348.
[8] E. R. Kolchin. Algebraic matric groups and the Picard-Vessiot theory of homogeneous
linear ordinary differential equations. Annals of Mathematics, 49(1):1–42, 1948. URL
https://dx.doi.org/10.2307/1969111.
[9] E. R. Kolchin. Algebraic groups and algebraic dependence. American Journal of Mathematics,
90(4):1151–1164, 1968. URL https://dx.doi.org/10.2307/2373294.
[10] J. J. Kovacic.
An algorithm for solving second order linear homogeneous differential equations.
Journal of Symbolic Computation, 2(1):3–43, 1986.
URL
https://dx.doi.org/10.1016/s0747-7171(86)80010-4.
[11] A. R. Magid. Lectures on Differential Galois Theory. American Mathematical Society, 1994.
URL https://doi.org/10.1090/ulect/007.
[12] I. Schur. Über gruppen periodischer substitutionen. Sitzber. Preuss. Akad. Wiss, pages 619–
627, 1911.
[13] M. F. Singer. Introduction to the Galois theory of linear differential equations. Preprint, 2007.
URL https://arxiv.org/abs/0712.4124.
[14] M. F. Singer and F. Ulmer.
Galois groups of second and third order linear differential equations.
Journal of Symbolic Computation, 16(1):9–36, 1993.
URL
https://dx.doi.org/10.1006/jsco.1993.1032.
[15] Á. Szántó. Complexity of the Wu-Ritt decomposition. Proceedings of the Second International Symposium on Parallel Symbolic Computation, pages 139–149, 1997. URL
https://dx.doi.org/10.1145/266670.266716.
13
[16] Á. Szántó. Computation with polynomial systems. PhD thesis, Cornell University, 1999.
[17] J. van der Hoeven.
Around the numeric–symbolic computation of differential Galois groups.
Journal of Symbolic Computation, 42(1-2):236–264, 2007.
URL
https://dx.doi.org/10.1016/j.jsc.2006.03.007.
[18] M. van der Put.
Galois theory and algorithms for linear differential equations.
Journal of Symbolic Computation, 39(3-4):451–463, 2005.
URL
https://dx.doi.org/10.1016/j.jsc.2004.11.013.
[19] M. van der Put and M. F. Singer. Galois theory of linear differential equations. SpringerVerlag, Berlin, 2003. URL https://dx.doi.org/10.1007/978-3-642-55750-7.
14
| 0math.AC
|
Bacterial foraging optimization based brain magnetic resonance
image segmentation
Abdul Kayom Md Khairuzzaman1, *
1
*
Department of Electrical Engineering, National Institute of Technology, Silchar, India
[email protected]
Abstract: Segmentation partitions an image into its constituent parts. It is essentially the pre-processing
stage of image analysis and computer vision. In this work, T1 and T2 weighted brain magnetic resonance
images are segmented using multilevel thresholding and bacterial foraging optimization (BFO) algorithm.
The thresholds are obtained by maximizing the between class variance (multilevel Otsu method) of the
image. The BFO algorithm is used to optimize the threshold searching process. The edges are then
obtained from the thresholded image by comparing the intensity of each pixel with its eight connected
neighbourhood. Post processing is performed to remove spurious responses in the segmented image. The
proposed segmentation technique is evaluated using edge detector evaluation parameters such as figure of
merit, Rand Index and variation of information. The proposed brain MR image segmentation technique
outperforms the traditional edge detectors such as canny and sobel.
1. Introduction
Segmentation partitions an image into certain regions of interest or clusters. It has numerous
applications ranging from computer vision to target matching for military applications. Another interesting
application is in medical science. For example, image segmentation helps in diagnosing abnormality in
brain or any part of the body from the MRI or PET scan. MRI is a powerful non-invasive technique for
diagnosis and treatment planning of various diseases such as multiple sclerosis (MS), Alzheimer’s disease,
Parkinson’s disease, epilepsy, cerebral atrophy, presence of any lesion like glioma etc [1]. Effective image
segmentation helps in classifying and analysing these disorders.
Histogram based thresholding is the most popular and simple technique for image segmentation. Otsu
[2] proposed a method for automatic threshold selection by maximizing the between class variance in a
gray level image. Kapur’s [3] method utilizes maximization of posterior entropy that indicates
homogeneity of the segmented classes. In general, the Kapur and Otsu methods are known for their better
shape and uniformity measures. These methods are originally developed for bi-level thresholding and later
on extended to multilevel thresholding. All these methods have one common problem that the
computational complexity increases exponentially when extended to multilevel thresholding. This is due
1
to the exhaustive search for the optimal thresholds, which limits their usage for multilevel thresholding
applications [4].
Nature has always been an inspiration in solving computationally complex problems. Particle swarm
optimization (PSO) [5, 6], Ant colony optimization (ACO) [7] and BFO [8] algorithms are inspired from
the foraging behaviour of natural animals. Computational complexity of multilevel thresholding is greatly
reduced by using these nature inspired optimization algorithms.
Genetic Algorithms (GA) [9], Particle Swarm Optimization (PSO) have been successfully applied in
multilevel thresholding [10, 11]. Sathya et al. [4] showed that BFO algorithm performs better than PSO
and GA for multilevel thresholding in terms of accuracy, speed and stability of the solution.
Bacteria Foraging Algorithm proposed by Passino [8], mimicked the foraging behaviour of E. Coli
bacteria present in human intestine. In foraging theory, it is assumed that the objective of the animal is to
search for and obtain nutrients such that the energy intake per unit time is maximized [8]. Maitra et al. [1]
applied BFO algorithm with Kapur method for multilevel thresholding of brain MR Image and showed its
superiority over PSO based multilevel thresholding.
Bi-level thresholding assumes that an image contains an object and the background. It is too harsh on
an image and eliminates lot of information and hence fails in most of the cases. In case of medical image
segmentation one has to be very careful while eliminating any information as it may affect diagnosis.
Therefore, it is necessary to perform multilevel thresholding which retains most of the important
information to help better diagnosis. However, multilevel thresholding without optimization is time
consuming. MRI is state of the art technique for diagnosis of various diseases. An effective segmentation
of MR image is therefore necessary for better diagnosis.
Our main objective is to determine optimal thresholds, so that the image can be subdivided into several
classes with different gray levels for their easier analysis and interpretation. For this, we propose a
multilevel thresholding technique based on multilevel Otsu method optimized with BFO algorithm for
2
brain MR Image segmentation. First Multilevel thresholding is performed on the brain MR image. The
edges in the thresholded image are then detected by comparing the intensity of each pixel with its eight
connected neighbourhood. The proposed algorithm is then evaluated objectively using Rand index [12],
variation of information and Pratt Figure of Merit [13]. For objective evaluation, a reference image is
created by manually segmenting the original image. The proposed brain MR image segmentation
technique out performs the traditional edge detection techniques such as Canny and Sobel.
2. Multilevel thresholding
Thresholding is a popular image segmentation technique because of its simplicity and effectiveness.
Normally, image histogram is used to determine the thresholds. There exist a number of methods for
threshold selection. Kapur et al. [3] and Otsu [2] are the most widely acknowledged histogram based
automatic threshold selection methods. Kapur’s method is based on maximizing posterior entropy of the
thresholded image whereas Otsu method maximizes the variance between segmented classes. Here we
explain briefly the Otsu method extended for multilevel thresholding [4, 2].
Let there be L gray levels in a given image, i.e. 0,1,2,..., L 1 and let us define Pi
h(i)
,
N
0 i (L 1) where h i number of pixels with gray level i and N = total number of pixels in the
L 1
image = h(i) .
i 0
Now, the multilevel thresholding problem can be configured as an m-dimensional optimization
problem for determination of m optimal thresholds [t1 , t2 ,.., tm ] which divide the original image into m 1
classes: C0 for [0,..., t1 1] , C1 for [t1 ,..., t2 1] ,... and Cm for [tm ,..., L 1] . The thresholds are obtained by
maximizing the following objective function:
J (t1, t2 ,..., tm ) 02 12 22 ... m2 .......... (1)
Where
02 0 (0 T )2 ,
12 1 (1 T )2 ,
22 2 (2 T )2 ,...,
m2 m (m T )2 are the variances of the segmented classes. Where the class probabilities are,
3
t1 1
t2 1
t3 1
L 1
i 0
i t1
i t2
i tm
0 Pi , 1 Pi , 2 Pi ,..., m Pi
And the mean levels 0 , 1 ,.., m for classes C0 , C1 ,..., Cm are as follows:
t1 1
0
i 0
ipi
0
t2 1
, 1
i t1
ipi
1
L 1
,..., m
i tm
ipi
m
Let T be the mean intensity for the whole image, then we have,
0 0 11 2 2 ... m m T And 0 1 2 ... m 1 .
3. Bacterial Foraging Optimization
Foraging strategies are methods of locating, handling and ingesting food. Natural selection eliminates
animals with poor foraging strategies. This facilitates the propagation of genes of most successful foraging
strategies. After many generations, the poor foraging strategies are either eliminated or redesigned into
better ones. A foraging animal looks to maximize energy intake per unit time spent on foraging within its
environmental and physiological constraints.
The E. Coli bacteria present in human intestine follows foraging behaviour, which consists of
chemotaxis, swarming, reproduction and elimination or dispersal. Passino [8] has modelled this
evolutionary technique as an effective optimization tool.
3.1 Chemotaxis
The bacterial movement of swimming (in a predefined direction) and tumbling (altogether in different
directions) in presence of attractant and repellent chemicals from other bacteria is called chemotaxis. A
chemotactic step is a tumble followed by a tumble or run. To represent a tumble, a unit length random
direction, ( j) is generated which is then used to model chemotaxis as follows,
X i j 1, k , l X i ( j, k , l ) C(i)( j)
.......... (2)
Where X i ( j, k , l ) represents the i th bacterium at j th chemotactic, k th reproductive and l th elimination or
dispersal event. C (i ) is the step size in the direction of movement specified by tumble (run length unit).
4
3.2 Swarming
Bacterium which reaches a good food source produces chemical attractant to invite other bacteria to
swarm together. While swarming, they maintain a minimum distance between any two bacteria by
secreting chemical repellent. Swarming is represented mathematically as,
Jcc( X , P( j , k , l ))
S
J cc( X , X ( j, k , l ))
i
i
i 1
S
m
i 1
n 1
[dattract exp(attract ( X n X i )2 )]
S
m
i 1
n 1
[hrepellant exp(repellant ( X n X i )2 )] .......... (3)
Where Jcc( X , P( j, k, l)) is the value of the cost function to be added to the actual cost function to be
optimized to simulate swarming behaviour. S is the total number of bacteria, m is the number of
parameters (dimension of the optimization problem) to be optimized. d attract , attract , repellant , hrepellant are
the coefficients to be chosen properly.
3.3 Reproduction
After completion of N c chemotactic steps, a reproductive step follows. Health of i th bacterium is
determined as,
Nc
J i health J sw (i, j, k , l ) .......... (4)
j 1
Then, the bacteria are sorted in descending order of their health. The least healthy bacteria die and the
other healthier bacteria take part in reproduction. In reproduction, each healthy bacterium splits into two
bacteria each containing identical parameters that of the parent keeping population of bacteria constant.
3.4 Elimination and dispersal
5
The bacterial population in a habitat may change gradually due to constraint of food or suddenly, due
to environmental or any other factor. All the bacteria in a region may be killed or a group may be
dispersed into a new location. It may have the possibility of destroying chemotactic progress, but it also
has the possibility of assisting chemotaxis, since dispersal event may place the bacteria to near good food
sources [1].
3.5 Bacterial foraging optimization algorithm
The original BFO algorithm given by Passino [8] is modified by Sathya et al. [4] which increases
the convergence speed and global searching ability of the algorithm. In the modified BFO algorithm,
they take the global best bacterium for the calculation of swarm attraction function. Instead of
averaging all the objective function values, they consider the best value for each bacterium. Let us
briefly explain the BFO algorithm in step by step manner.
Step 1
Initialize the number of variables to be optimized p , number of E. Coli bacteria S , number of
chemotactic steps N c , maximum swimming length N s , number of reproduction steps N re , number of
elimination or dispersal events N ed , the probability of elimination or dispersal Ped , and the step size
C i , i 1, 2,...S .
Step 2
Elimination-dispersal loop: ell ell 1
Step 3
Reproduction loop: k k 1
Step 4
Chemo-taxis loop: j j 1
6
Step 4.1
For i 1,2,..., S take a chemotactic step for i th bacterium as:
Step 4.2
Calculate the value of objective function J i, j, k , ell
Step 4.3
Find the global best bacterium X gn , from all the objective functions evaluated till this point.
Step 4.4
Calculate J sw , i.e. the cost function value J which is to be added with the swarm attractant
cost J cc . So J sw can be expressed by the following equation:
J sw i, j, k , ell J i, j, k , ell J cc X gn ( j , k , ell ), X ( j , k , ell )
Step 4.5
Let Jlast J sw i, j, k , ell save this value for finding a better cost via run.
Step 4.6
Tumble: Generate a random vector i with each element n i , n 1,2,..., m , a random
number in 1,1 .
Step 4.7
Move:
X i j 1, k , ell X i j , k , ell C (i)
(i)
(i)(i)
T
which results in a step of size C (i ) in the direction of the tumble for i th bacterium.
Step 4.8
Calculate J i, j 1, k , ell and let
J sw i, j 1, k , ell J i, j , k , ell J cc X ( j 1, k , ell ), P( j 1, k , ell )
i
7
Step 4.9
Swim:
set m=0 (counter for swim length).
While m<Ns (if have not climbed down too long ).
m=m+1.
If J sw i, j 1, k , ell Jlast (if doing better),
Let J last J sw (i, j 1, k , ell ) and
X i ( j 1, k , ell ) X i ( j, k , ell ) C(i)
(i)
(i)(i)
T
And use this X i ( j 1, k , ell ) to calculate the new J sw (i, j 1, k , ell ) .
Else, set m=Ns.
Go to next bacterium (i 1) if i S.
That is go to step 4.2; process the next bacterium.
Step 5
if j Nc , go to step 4. In this case, continue chemotaxis since the life of the bacteria is not over.
Step 6
Reproduction:
Step 6.1
For the given k and ell , and for each i 1,2,..., S , compute J health value.
Step 6.2
Sort the bacteria in order of descending J health value.
Step 6.3
8
The Sr bacteria with the lowest J health values die and the remaining Sr bacteria with best values
split (it is performed by the copies that are already made, now placed at the same location as their
parents).
Step 7
If k Nre , go to step 3. The next generation of the bacteria starts.
Step 8
Elimination-dispersal:
For i 1,2,..., S , eliminate and disperse each bacterium with probability Ped . To do this, if a bacterium
is eliminated, simply disperse another one to a random location on the optimization domain. If
ell Ned , go to step 2; otherwise, stop.
4. Proposed technique
The flowchart of fig-1 shows the proposed technique of segmenting the MR images. An MR image is
taken and optimal multilevel thresholding is performed based on its histogram. Then the edges in the
thresholded image are detected by comparing the intensity values. Eight connected neighbourhood pixels
are considered and intensity value of the centre pixel is compared with each of the neighbouring pixels.
Whenever we get a difference of nonzero value, one of the pixels is marked as edge pixel. The procedure
is performed throughout the whole image to get the complete edge image. Then post processing is
performed to get the final segmented image.
9
Figure 1: The proposed technique
4.1 Determine the number of thresholds
There are four contrast levels in brain MR Images (T1 and T2) corresponding to fat, white matter, grey
matter and cerebrospinal fluid [14]. Histogram of the image should carefully be examined before
determining the number of thresholds. We have experimented with various numbers of thresholds on a
number of T1 and T2 weighted brain MR images. Usually, 3-4 level of thresholding in brain MR (T1 and
T2) Images works well.
4.2 Perform multilevel thresholding
After selecting the suitable number of thresholds, the MR image is thresholded using bacterial foraging
optimization algorithm by maximizing the objective function which in this case is between class variance
as given by the equation -1 as discussed in section-2.
4.3 Detect the edges
Edges are detected by simply comparing the intensities in the thresholded image. To search for the
edge pixels and mark them, we compute the intensity difference of the pixel I (x, y) with the eight
10
connected neighborhood in the thresholded image. Wherever we find that the intensity difference is
nonzero, the greater intensity pixel is marked as edge (or the lower intensity pixel can also be marked as
edge). It must be remembered that only one of them should be marked to map it with an edge. The
intensity comparison is performed throughout the whole image to obtain the complete edge image.
4.4 Post processing
Post processing is done on the segmented image. The function ‘bwareaopen (bw,p)’ in MATLAB is
used to remove small objects (such as spurious and noise contents) in the segmented image. This function
removes all connected components (objects) having less than p number of pixels.
4.5 Objective evaluation
The accuracy of the segmentation is measured using objective evaluation techniques such as “figure of
merit”, “Rand Index”, and “Variation of information”.
5. Results and discussion
Brain MR Image segmented results are presented in fig-2 to fig-9. The evaluation criteria Pratt figure
of merit, rand index and variation of information are used to assess the performance of the proposed
segmentation technique. A greater value of figure of merit specifies better performance of the edge
detector. The Sobel and Canny edge is obtained using matlab toolbox. The figure of merit value obtained
for BFO edge image is 0.8135 which is greater than the corresponding value for Canny and Sobel edge
images as given in table1. The table2 gives the Rand Index, and variation of information values for BFO,
Canny and Sobel edge images. Highest Rand Index and the lowest Variation of Information are achieved
in the proposed BFO segmentation technique. A greater value of Rand Index and lesser value of variation
of information indicate better performance of the segmentation method. Thus the parameter values
obtained indicate better performance of the proposed technique than the traditional Canny and Sobel
11
methods. Figure-9 shows the convergence of the BFO algorithm as the graph becomes progressively flat
towards the right hand side.
Figure 2: T2w brain MR image
Figure 4: thresholded image
Figure 3: Histogram of the image
Figure 5: BFO segmented image
Figure 6: Canny edge image
12
Figure 7: Sobel edge image
Figure 8: Manually segmented image
Figure 9: Convergence of the BFO algorithm
Table 1: Objective evaluation of the edge detectors
Edge detector
Figure of merit
Canny
0.4696
Sobel
0.3831
Proposed
0.8135
Table 2: Objective evaluation of the edge detectors
Parameters
Rand Index (RI)
Sobel
0.8913
Canny
0.8988
Proposed
0.9639
Variation of
information(VI)
0.4647
0.4092
0.2102
6. Conclusions
This paper presents a multilevel thresholding based brain MR image segmentation technique for T1
and T2 weighted brain MR images. The visual quality of the segmented results and the values of the
objective evaluation parameters demonstrate the superiority of the proposed technique over traditional
methods such as canny and sobel. We have achieved a significantly higher figure of merit by the proposed
13
technique than that of canny and sobel edge detection techniques. Moreover, the RI and VI parameters as
obtained by the proposed method are also better than the canny and sobel edge detection techniques for T1
and T2 weighted brain MR images.
References
[1] Maitra, Madhubanti, and Amitava Chatterjee. "A novel technique for multilevel optimal magnetic.
[2] Otsu, Nobuyuki. "A threshold selection method from gray-level histograms." Automatica 11.285-296
(1975): 23-27.
[3] Kapur, Jagat Narain, Prasanna K. Sahoo, and Andrew KC Wong. "A new method for gray-level picture
thresholding using the entropy of the histogram." Computer vision, graphics, and image processing 29.3
(1985): 273-285.
[4] Sathya, P. D., and R. Kayalvizhi. "Modified bacterial foraging algorithm based multilevel thresholding
for image segmentation." Engineering Applications of Artificial Intelligence 24.4 (2011): 595-615.
[5] Kennedy, James. "Particle swarm optimization." Encyclopedia of Machine Learning. Springer US,
2010. 760-766.
[6] Poli, Riccardo, James Kennedy, and Tim Blackwell. "Particle swarm optimization." Swarm
intelligence 1.1 (2007): 33-57.
[7] Dorigo et al., “Distributed optimization by ant colonies”, European conference on artificial life, paris,
france, elsevier publishing, 134–142.
[8] Passino, Kevin M. "Biomimicry of bacterial foraging for distributed optimization and control." Control
Systems, IEEE 22.3 (2002): 52-67.
[9] John Henry Holland. Adaptation in natural and artificial systems: an introductory analysis with
applications to biology, control, and artificial intelligence. MIT press, 1992.
14
[10] Maitra and Chatterjee, “A hybrid cooperative–comprehensive learning based PSO algorithm for
image segmentation using multilevel thresholding”, Elsevier, Expert Systems with Applications 34 (2008)
1341–1350.
[11] Kamal et al., “A multilevel automatic thresholding method based on a genetic algorithm for a fast
image segmentation”, Elsevier, Computer Vision and Image Understanding 109 (2008) 163–175.
[12] Unnikrishnan, Ranjith, Caroline Pantofaru, and Martial Hebert. "A measure for objective evaluation
of image segmentation algorithms." Computer Vision and Pattern Recognition-Workshops, 2005. CVPR
Workshops. IEEE Computer Society Conference on. IEEE, 2005.
[13] William K. Pratt, “Digital image processing”, A Wiley-Interscience Publication, 490-494.
[14] Jerrold et al., “The essential physics of medical imaging”, Lippincot Williams and Wilkins, 2002.
395-400.
15
| 9cs.NE
|
arXiv:1711.04584v1 [math.GR] 13 Nov 2017
PROPERTY (T), FINITE-DIMENSIONAL
REPRESENTATIONS, AND GENERIC REPRESENTATIONS
MICHAL DOUCHA, MACIEJ MALICKI, AND ALAIN VALETTE
Abstract. Let G be a discrete group with property (T). It is a standard
fact that, in a unitary representation of G on a Hilbert space H, almost
invariant vectors are close to invariant vectors, in a quantitative way.
We begin by showing that, if a unitary representation has some vector
whose coefficient function is close to a coefficient function of some finitedimensional unitary representation σ, then the vector is close to a subrepresentation isomorphic to σ: this makes quantitative a result of P.S.
Wang [12]. We use that to give a new proof of a result by D. Kerr, H.
Li and M. Pichot [9], that a group G with property (T) and such that
C ∗ (G) is residually finite-dimensional, admits a unitary representation
which is generic (i.e. the orbit of this representation in Rep(G, H) under
the unitary group U (H) is comeager). We also show that, under the
same assumptions, the set of representations equivalent to a Koopman
representation is comeager in Rep(G, H).
1. Introduction
Let G be a discrete group and π be a unitary representation of G on
some Hilbert space H. For a finite set F ⊂ G and ε > 0, a vector ξ ∈ H
is (F, ε)-invariant if maxg∈F kπ(g)ξ − ξk < ε. Recall that π almost has
invariant vectors if, for every pair (F, ε), the group G has (F, ε)-vectors;
and that the group G has Kazhdan’s property (T) or is a Kazhdan group
if every unitary representation of G almost having invariant vectors, has
non-zero invariant vectors; see e.g. [2] for Property (T). The definition can
be reformulated in terms of weak containment of representations: G has
Property (T) if every unitary representation weakly containing the trivial
representation of G, contains it strongly (see Remark 1.1.2 in [2]). Crucial
for us is an equivalent characterization due to P.S. Wang (Corollary 1.9
and Theorem 2.1 in [12]): the group G has property (T) if and only if for
some (hence every) irreducible finite-dimensional unitary representation σ
of G, every unitary representation π of G that contains σ weakly, contains
it strongly.
2010 Mathematics Subject Classification. 22D10.
Key words and phrases. property (T), Wang’s theorem, unitary representations, generic
representations, Koopman representations.
1
2
MICHAL DOUCHA, MACIEJ MALICKI, AND ALAIN VALETTE
It is a simple but useful fact that, if G has property (T) and π is a unitary
representation almost having invariant vectors, “almost invariant vectors are
close to invariant vectors”. More precisely:
Proposition 1.1 (Proposition 1.1.9 in [2]). Let G be a Kazhdan group.
If S is a finite generating set of G and ε0 is the corresponding Kazhdan
constant, then for every δ ∈]0, 1[ and every unitary representation π of G,
any (S, ε0 δ)-invariant vector ξ satisfies kξ − P ξk ≤ δkξk, where P is the
orthogonal projection on the subspace of π(G)-invariant vectors.
For a Kazhdan group G and a unitary representation π of G, fix a unit
vector ξ and look at the coefficient function
φπ,ξ (g) = hπ(g)ξ, ξi (g ∈ G).
The question we first address in this paper is: if φπ,ξ is close to some coefficient of an irreducible finite-dimensional unitary representation σ of G,
must ξ be close to a finite-dimensional invariant subspace of π carrying a
sub-representation isomorphic to σ? We will see that, in analogy to Proposition 1.1, the answer is positive - with some care.
Definition 1.2. Let G be a finitely generated group with a symmetric finite
generating set S ⊆ G and let φ be some normalized positive definite function on G associated with a unitary irreducible representation σ, of finite
dimension d. Let π be some unitary representation of G on H. Let ε > 0.
2
Say that a unit vector ξ ∈ H is (π, φ, ε)-good if for every s ∈ S 2d +1 we have
|φπ,ξ (s) − φ(s)| < ε.
Note that S k is just the ball of radius k centered at the identity in G.
So there is a certain lack of uniformity in Definition 1.2: we require an
approximation of φπ,ξ by φ on a ball whose size depends on the dimension
of the representation d. Our main result, proved in section 2, can be viewed
as a quantitative version of Wang’s result.
Theorem 1.3. Let G be a discrete Kazhdan group, S a finite symmetric
generating set with e ∈ S, and let φ be a normalized positive definite function
on G associated with a finite-dimensional unitary irreducible representation
σ of G. For every 0 < δ < 1 there exists εφ,δ > 0 such that for every unitary
representation π of G on a Hilbert space H, and every unit vector x ∈ H
that is (π, φ, εφ,δ )-good, there exists a unit vector x′ ∈ H with kx − x′ k ≤ δ
such that the restriction of π to the span of π(G)x′ is isomorphic to σ.
In section 3, we apply Theorem 1.3 to the study of the global structure of
the space of unitary representations of Kazhdan groups. Let us start with
the notation. Let G be an arbitary countable group and let H be a separable
infinite-dimensional Hilbert space. The set Rep(G, H) of all homomorphisms
from G into the unitary group U (H) can be viewed as a closed subset of
the product space U (H)G , when we equip U (H) with the strong operator
topology. With this identification, Rep(G, H) is a Polish (i.e. separable
PROPERTY (T) AND UNITARY REPRESENTATIONS
3
and completely metrizable) space. We refer the reader to the monograph
[8], especially to the section on the spaces of unitary representations, for
more information about this point of view on unitary representations. Recall that two unitary representations π1 , π2 ∈ Rep(G, H) are isomorphic,
or unitarily equivalent if there is a unitary operator φ ∈ U (H) such that
π1 (g) = φπ2 (g)φ∗ , for every g ∈ G. Notice that this is an orbit equivalence relation given by the action of the unitary group U (H) on the space
Rep(G, H) by conjugation. Kechris raised a question (see again the section
on the space of unitary representations in [8]) if there are countable groups
with a generic unitary representation, where “generic” here means its conjugacy class is large in the sense of Baire category, i.e. a representation
whose class under the unitary equivalence contains a dense Gδ subset. As
a matter of fact, we mention that it follows from the topological zero-one
law that for every countable group G either there is a generic representation
in Rep(G, H), or all conjugacy classes are meager (see e.g. Theorem 8.46
in [7]; to apply it, note that there is a dense conjugacy class in Rep(G, H)
— indeed, take some countable dense set of representations from Rep(G, H)
and consider their direct sum).
Here as an application of Theorem 1.3 we prove the following result.
Theorem 1.4. Let G be a discrete Kazhdan group such that finite-dimensional
representations are dense in the unitary dual Ĝ. Then there is a generic unitary representation of G.
We note that, although not explicitly stated there, this result already
follows from a more general result of Kerr, Li and Pichot from [9], where
they prove (see Theorem 2.5 there) that if A is a separable C*-algebra where
finite-dimensional representations are dense in Â, then there is a dense Gδ
class in Rep(A, H). Theorem 1.4 is then a special case for A = C ∗ (G). Our
proof is nevertheless done by more elementary means, in particular it does
not invoke Voiculescu’s theorem (see the proof of Theorem 2.5 in [9] for details).
Another open question posed by Kechris as Problem H.16 in [8] is whether
the subset of those representations π ∈ Rep(G, H), where G is still a countable group, that are equivalent to Koopman representations is meager in
Rep(G, H). Such representations are called realizable by an action in [8].
Let us recall the terminology first. Let (X, µ) be a standard probability
space (i.e. a space isomorphic to the unit interval [0, 1] equipped with the
Lebesgue measure). Let α : G y (X, µ) be an action of a countable group G
on X by measure preserving measurable transformations. Consider the unitary representation πα : G → L2 (X, µ) defined by πα (g)f (x) = f (α(g −1 , x)),
for every f ∈ L2 (X, µ). The Koopman representation of α is the restriction
of πα to the invariant subspace L20 (X, µ), which is the orthogonal complement of the invariant subspace of constant functions.
4
MICHAL DOUCHA, MACIEJ MALICKI, AND ALAIN VALETTE
In section 4 we prove the following result addressing the question of
Kechris.
Theorem 1.5. Let G be a discrete Kazhdan group such that finite-dimensional
representations are dense in the unitary dual Ĝ. Then the set of representations realizable by an action is comeager in Rep(G, H).
Let us mention that the condition that finite-dimensional representations
are dense in the unitary dual Ĝ is, by the result of Archbold from [1], equivalent with the statement that the full group C*-algebra C ∗ (G) is residually
finite-dimensional. That is in turn, by the result of Exel and Loring from [6]
(see also [11]), equivalent with the statement that finite-dimensional representations are dense in Rep(G, H), which we shall use in the proof. Note that
we call a representation π ∈ Rep(G, H) finite-dimensional if the subalgebra
π(G) generates in B(H) is finite-dimensional.
The existence of infinite discrete Kazhdan groups with residually finitedimensional C*-algebras seems to be open — see Question 7.10 in [2] and
also Question 6.5 of Lubotzky and Shalom in [10] where they ask if there are
infinite discrete Kazhdan groups with property FD, which is strictly stronger
than having a residually finite-dimensional C*-algebra (a group has property
FD if representations factoring through finite groups are dense in the unitary
dual).
Question 1.6. It is known that being residually finite is not a sufficient
condition to have a residually finite-dimensional C*-algebra by a result of
Bekka [3]. However how about being LERF? (Recall that a finitely generated
group is LERF if any finitely generated subgroup is the intersection of the finite index subgroups containing it). Ershov and Jaikin-Zapirain constructed
in [5] a Kazhdan group which is LERF. Is its group C*-algebra residually
finite-dimensional?
Remark 1.7. We note that on the other hand we cannot exclude that it is
possible to prove by a different argument that for every infinite group G,
all classes in Rep(G, H) are meager. That would together with Theorem
1.4 give that there are no infinite Kazhdan groups with a residually finitedimensional C*-algebra.
Acknowledgements: The first named author was supported by the GAČR
project 16-34860L and RVO: 67985840.
2. A quantitative version of Wang’s theorem
Let G be an infinite, finitely generated group. Let S be a finite, symmetric, generating set of G, with e ∈ S. Let CG be the complex group ring of
G.
2.1. Quantifying the Burnside theorem. Let σ be an irreducible unitary representation of dimension d, i.e. a homomorphism σ : G → Ud (C)
PROPERTY (T) AND UNITARY REPRESENTATIONS
5
such that σ(G) has no proper invariant subspace. The classical Burnside
theorem says that σ(CG) = Md (C), i.e. σ(G) contains a basis of Md (C).
Definition 2.1. Set k(σ) = min{k > 0 : dimC span σ(S k ) = d2 }.
Lemma 2.2. There is a constant C > 0 (only depending on S) such that
C log d ≤ k(σ) ≤ d2 .
Proof. We have
d2 = dimC span σ(S k(σ) ) ≤ |σ(S k(σ) )| ≤ |S k(σ) | ≤ |S|k(σ) .
Taking logarithms:
2
log |S|
log d ≤ k(σ). To prove the upper bound, observe
that the sequence span σ(S k ) of subspaces of Md (C), is strictly increasing for
k < k(σ). Indeed, assume that k is such that span σ(S k ) = span σ(S k+1 ):
this means that span σ(S k ) is stable by left multiplication by σ(S), hence
by σ(G) as S is generating. Since the identity matrix is in σ(S k ), we have
σ(G) ⊂ span σ(S k ), hence k ≥ k(σ). From this it is clear that k(σ) ≤ d2 .
Let v be a unit vector in Cd . Since v is cyclic for σ(G), the map:
Tv : CS k(σ) → Cd : f 7→ σ(f )v
is onto. Let (ker Tv )⊥ denote the orthogonal of ker Tv in CS k(σ) , let Uv be
the inverse of the map Tv |(ker Tv )⊥ . Endow CS k(σ) with the ℓ1 -norm, and
let kUv k2→1 be the corresponding operator norm of Uv . So for every w a
unit vector in Cd , there exists f ∈ CS k(σ) with kf k1 ≤ kUv k2→1 , such that
σ(f )v = w.
Lemma 2.3. There exists M > 0 such that for every two unit vectors
v, w ∈ Cd , there exists f ∈ CS k(σ) with kf k1 ≤ M , such that σ(f )v = w.
Proof. This is the preceding observation plus compactness of the unit sphere
in Cd : the constant M = maxkvk=1 kUv k2→1 does the job.
2.2. From weak containment to weak containment à la Zimmer.
Recall that, if π, ρ are unitary representations of a discrete group G, the
representation π is weakly contained in the representation ρ (i.e. π ρ) if
every function of positive type associated with π can be pointwise approximated by finite sums of positive definite type associated with ρ. If π is
irreducible, this is equivalent to require that every normalized function of
positive type associated with π can be pointwise approximated by normalized functions of positive type associated with ρ (see Proposition F.1.4 in
[2]).
Zimmer introduced in Definition 7.3.5 of [13] a different, inequivalent
notion of weak containment. A n × n-submatrix of π is a function
G → Mn (C) : g 7→ (hπ(g)ei , ej i)1≤i,j≤n
where {e1 , ..., en } is an orthonormal family in Hπ . Say that π is weakly
contained in ρ in Zimmer’s sense (i.e. π Z ρ) if, for every n > 0, every
6
MICHAL DOUCHA, MACIEJ MALICKI, AND ALAIN VALETTE
n×n-submatrix of π can be pointwise approximated by n×n-submatrices of
ρ. The exactly relation with the classical notion recalled above, is worked out
in Remark F.1.2(ix) in [2]; in particular, when π is irreducible, π ρ implies
π Z ρ. Our first goal will be to make the latter statement quantitative.
For this we need a definition.
Let φ be associated with σ, as in Definition 1.2. Let v be a unit vector
in Hσ such that φ = φσ,v . Let e1 , ..., ed be an orthonormal basis of Cd ; by
lemma 2.3, we find functions f1 , ..., fd ∈ CS k(σ) , with maxi kfi k1 ≤ M , such
that σ(fi )v = ei (i = 1, ..., d).
Lemma 2.4. Let π ∈ Rep(G, H) be a unitary representation. Assume there
is ε > 0 and a unit vector η ∈ H such that for s ∈ S 2k(σ)+1 we have
|hπ(s)η, ηi − hσ(s)v, vi| < ε. Set ηi = π(fi )η. Then for i, j = 1, ..., d and
g ∈ S:
|hσ(g)ei , ej i − hπ(g)ηi , ηj i| ≤ εM 2 .
Proof. For g ∈ S:
|hσ(g)ei , ej i − hπ(g)ηi , ηj i| = |hσ(g)σ(fi )v, σ(fj )vi − hπ(g)π(fi )η, π(fj )ηi|
X
=|
fi (s)fj (t)(hσ(t−1 gs)v, vi − hπ(t−1 gs)η, ηi)|
s,t∈G
≤
X
|fi (s)||fj (s)||hσ(t−1 gs)v, vi − hπ(t−1 gs)η, ηi|.
s,t∈G
Since the supports of the fi ’s are contained in S k(σ) , and t−1 gs ∈ S 2k(σ)+1
for s, t ∈ S k(σ) , we get using the assumption:
X
|hσ(g)ei , ej i − hπ(g)ηi , ηj i| ≤ ε
|fi (s)||fj (t)| = εkfi k1 kfj k1 ≤ εM 2 .
s,t∈G
In the previous proof, by applying the Gram-Schmidt orthonormalization process to the ηi ’s, it is possible to show that the d × d-submatrix
(hσ(·)ei , ej i)1≤i,j≤d of σ, is close on S to some d × d-submatrix of α, with an
explicit bound; but we don’t need it at this point.
2.3. Quantifying Wang’s theorem. Let Hσ be the (d-dimensional) Hilbert
space of σ, and let Hσ be the conjugate Hilbert space (with complex conjugate scalar multiplication and complex conjugate inner product), equipped
with the conjugate representation σ. Form the tensor product Hσ ⊗Hπ , carP
rying the representation σ ⊗ π. Set ξi = ei ⊗ ηi and ξ = di=1 ξi ∈ Hσ ⊗ Hπ ,
where the ei ’s and ηi ’s are as in the section above; observe that the ξi ’s are
pairwise orthogonal. We need an estimate on how ξ is moved by σ ⊗ π.
kξ − (σ ⊗ π)(g)ξk2 = 2kξk2 − 2Reh(σ ⊗ π)(g)ξ, ξi
d
d
X
X
Reh(σ ⊗ π)(g)ξi , ξj i
kξi k2 − 2
=2
i=1
i,j=1
PROPERTY (T) AND UNITARY REPRESENTATIONS
=2
d
X
2
kηi k − 2
d
X
7
Rehej , σ(g)ei ihπ(g)ηi , ηj i.
i,j=1
i=1
P
Observe that for every g ∈ G we have: d = di,j=1 hσ(g)ei , ej ihej , σ(g)ei i as
the ei ’s are an orthonormal basis. Subtracting and adding 2d to the previous
formula we get:
d
d
X
X
2
Rehej , σ(g)ei i(hπ(g)ηi , ηj i−hσ(g)ei , ej i)
kξ−(σ⊗π)(g)ξk = 2[ (kηi k −1)]−2
2
i,j=1
i=1
hence, using Cauchy-Schwarz:
2
kξ −(σ⊗π)(g)ξk ≤ 2
d
X
i=1
2
|kηi k −1|+2
d
X
|hπ(g)ηi , ηj i−hσ(g)ei , ej i| (2.1)
i,j=1
Theorem 1.3 will follow immediately form the next Proposition, together
with lemma 2.2
Proposition 2.5. Let G be a discrete Kazhdan group, S a finite symmetric
generating set with e ∈ S, and let φ be a normalized positive definite function
on G associated with a finite-dimensional unitary irreducible representation
σ of G. For every 0 < δ < 1 there exists εφ,δ > 0 such that for every
π ∈ Rep(G, H), and every unit vector x ∈ H such that |φ(s)− φπ,x (s)| < εφ,δ
for s ∈ S 2k(σ)+1 , there exists a unit vector x′ ∈ H with kx − x′ k ≤ δ such
that the restriction of π to the span of π(G)x′ is isomorphic to σ.
Proof. Set d = dim σ, let v be a unit vector in Hσ such that φ(g) = hσ(g)v, vi
for every g ∈ G. As in section 2.2, for an orthonormal basis e1 , ..., ed of
Cd , we find functions f1 , ..., fd ∈ CS k(σ) , with maxi kfi k1 ≤ M , such that
σ(fi )v = ei (i = 1, ..., d).
Let 0 < ε0 < 2 be such that (S, ε0 ) is a Kazhdan pair for G. Fix δ with
0 < δ < 1, and set
δ2 ε20
.
εφ,δ = ε =
24d(d + 1)M 2
Let π ∈ Rep(G, H) and x ∈ H be a unit vector with |φπ,x (s) − φ(s)| < ε
for s ∈ S 2k(σ)+1 . Set ηi = π(fi )x. We may assume that e1 = v and
the function f1 is δe , so that η1 = x. We want to prove that the vector
P
ξ = di=1 (ei ⊗ ηi ) ∈ Hσ ⊗ H is (S, tε0 )-invariant for some 0 < t < 1, in order
to apply Proposition 1.1.
For g ∈ S we have, by lemma 2.4 and the inequality 2.1:
δ2 ε20
12
2
Again by lemma 2.4, evaluated at g = e, we have: |kηi k − 1| ≤ εM 2 < 12 ,
P
hence 21 ≤ kηi k2 ≤ 32 and d2 ≤ kξk2 = di=1 kηi k2 ≤ 3d
2 . So that, for g ∈ S:
kξ − (σ ⊗ π)(g)ξk2 ≤ 2dεM 2 + 2d2 εM 2 = 2d(d + 1)εM 2 =
kξ − (σ ⊗ π)(g)ξk2 ≤
δ2 ε20
kξk2 .
6d
8
MICHAL DOUCHA, MACIEJ MALICKI, AND ALAIN VALETTE
By Proposition 1.1, there exists a G-fixed ξ ′ ∈ Hσ ⊗ H such that kξ −
Pd
Pd
δ2
′ 2
ξ ′ k2 ≤ 6d
kξk2 . Write ξ ′ =
i=1 ei ⊗ ζi , so that kξ − ξ k =
i=1 kηi −
ζi k2 . Identify Hσ ⊗ H with the space of linear operators from Hσ to H
(endowed with the Hilbert-Schmidt norm), via u ⊗ y 7→ (w 7→ hw, uiy).
P
Then ξ ′ identifies with the operator w 7→ di=1 hw, ei iζi , which is therefore
an intertwining operator between σ and π. The image of this operator, which
is span{ζ1 , ..., ζd }, carries a sub-representation of π unitarily equivalent to
σ (by Schur’s lemma). Set x′′ = ζ1 , then:
′′ 2
2
kx − x k = kη1 − ζ1 k ≤
d
X
kηi − ζi k2 = kξ − ξ ′ k2 ≤
i=1
δ
2.
Finally, set x′ =
i.e. kx − x′′ k ≤
triangle inequality:
x′′
kx′′ k ,
δ2
δ2 3d
δ2
kξk2 ≤
= ,
6d
6d 2
4
a unit vector in H. Then by the
kx − x′ k ≤ kx − x′′ k + kx′′ − x′ k = kx − x′′ k + kx′′ k|1 −
1
|
kx′′ k
= kx − x′′ k + |kx′′ k − kxk| ≤ 2kx − x′′ k ≤ δ.
This concludes the proof.
Question 2.6. In the previous proof, the constant εφ,δ depends on σ through
the dimension d and the constant M from lemma 2.3. By Theorem 2.6 in
[12], a discrete Kazhdan group has finitely many unitary irreducible representations of a given finite dimension (up to unitary equivalence), so Theorem
1.3 can be made uniform over all unitary irreducible representations σ with
dimension less than a given dimension. Can it be made uniform over all
finite-dimensional unitary representations?
3. Proof of Theorem 1.4
Let {Un } be a countable basis of open sets in the unit sphere K of H, and
let Φ be the set of all positive definite functions on G defining irreducible
finite dimensional representations. Notice that the set X ′ ⊆ Rep(G, H) of
all representations π such that for every n ∈ N and every δ > 0 there exist
m > 0, x ∈ Un , xi ∈ K, ci ∈ C \P{0}, and φi ∈ Φ, i ≤ m, such that the
xi ’s are pairwise orthogonal, x =
ci xi , and each xi is (π, φi , εφi ,δi′ )-good,
δ
′
where δi = |ci |.m , and εφi ,δi′ is given by Theorem 1.3, is a Gδ set. Indeed,
for fixed n, δ, m, x, x̄ = (x1 , . . . , xm ), c̄ = (c1 , . . . , cm ), φ̄ = (φ1 , . . . , φm ) as
above, the set
n,δ,m
Vx,x̄,c̄,
= {π ∈ Rep(G, H) : each xi is (π, φi , εφi ,δi′ )-good}
φ̄
n,δ,m
is clearly open. We also put Vx,x̄,c̄,
to be the empty set if the xi ’s are not
φ̄
P
pairwise orthogonal or x =
6
ci xi . Now we can define X ′ by
\ \ [ [ [ [ [
n,ǫ,m
Vx,x̄,c̄,
,
X′ =
φ̄
n∈N δ∈Q+ m∈N x∈Un x̄∈K m c̄∈Cm φ̄∈Φm
PROPERTY (T) AND UNITARY REPRESENTATIONS
9
which is a Gδ condition.
Moreover, X ′ is dense in Rep(G, H) as it contains all direct sums of
finite-dimensional representations, which, by our assumption, are dense in
Rep(G, H). This is because it is easy to see that
P for every such sum π there
are densely many elements x ∈ K of the form
ci xi , where xi are pairwise
orthogonal unit vectors, and each xi is (π, φi , δ)-good for some φi and every
δ > 0.
Now we show that every representation in X ′ is a direct sum of finitedimensional representations. Fix π ∈ X ′ . Using Zorn’s lemma, we can
decompose H into H0 and H1 such that H0 is the direct sum of all finitedimensional representations contained in π. For i = 0, 1, let PHi be the
orthogonal projection of H on Hi . Suppose that H1 is not trivial, and fix
0 < δ < 1, x ∈ P
K, pairwise orthogonal xi ∈ K and ci ∈ C \ {0}, i ≤ m,
such that x =
ci xi , each xi is (π, φi , εφi , δ )-good for some φi ∈ Φ,
|ci |.m
and kx − PH0 xk > δ (the last condition can be satisfied by choosing x in
an appropriate Un .) By Theorem 1.3, there exist x′i ∈ K, i ≤ m, inducing
irreducible finite-dimensional representations, and such that kxi − x′i k <
P ′
δ
ci xi k < δ. But then, clearly, x′i0 6∈ H0 for some i0 ≤ m,
|ci |.m , that is, kx −
P
as if it was not the case, we would get that kx − ci x′i k ≥ kx − PH0 xk > δ.
Since PH1 is a G-intertwiner, the image under PH1 of the linear span of
π(G)x′i0 , is an invariant subspace of H1 , which is a contradiction.
Now let X ′′ be the set of all those representations that contain every finite
dimensional representation with infinite multiplicity. As G is a Kazhdan
group, we can see that X ′′ is given by a Gδ condition. Indeed, for [σ] the
isomorphism class of a finite-dimensional unitary irreducible representation
of G, and n > 0, let V[σ],n be the set of representations π ∈ Rep(G, H) such
that [σ] appears in π with multiplicity at least n. Clearly V[σ],n is open and
\\
X ′′ =
V[σ],n ,
[σ] n
where the intersection is countable because there are countably many [σ]’s.
By our assumption on C ∗ (G), the set X ′′ is dense. Thus, X = X ′ ∩
′′
X is a dense Gδ set, all the representations of which are direct sums of
finite dimensional representations, each appearing with infinite multiplicity.
Clearly, all elements in X are conjugate.
Remark 3.1. The converse of Theorem 1.4 also follows from Theorem 2.5
in [9]. That is, if either G does not have property (T), or C ∗ (G) is not
residually finite-dimensional, then all classes in Rep(G, H) are meager. Indeed, Theorem 2.5 from [9] says: if for a separable C*-algebra A the set
of isolated points in  is not dense, then the restriction of the action of
U (H) by conjugation on a dense Gδ invariant subset of Rep(A, H) is turbulent. That, by the definition of turbulence, in particular implies that every
class in Rep(G, H) is meager. Now take A = C ∗ (G): as isolated points in
10
MICHAL DOUCHA, MACIEJ MALICKI, AND ALAIN VALETTE
Ĝ correspond to finite-dimensional representations, it follows that when G
does not have property (T), Ĝ does not have isolated points, by Theorem
2.1 in Wang [12]; when C ∗ (G) is not residually finite-dimensional, then the
isolated points in Ĝ are not dense by Archbold’s main result in [1].
4. Proof of Theorem 1.5
For a unitary representation π, we denote by ∞ · π the ℓ2 -direct sum of
countably many copies of π.
Lemma 4.1. Let H be a locally compact group. Assume that H has (up to
unitary equivalence) countably many finite-dimensional irreducible unitary
representations σ1 , σ2 , .... Then the representation ⊕∞
n=1 ∞ · σn is unitarily
equivalent to a Koopman representation.
Proof. View σn as a continuous homomorphism H → U (Nn ). Let Kn denote
the closure of σn (H) in U (Nn ), so that Kn is a compact group (on which H
acts by left translations by elements of σn (H)). Let mn denote normalized
Haar measure on Kn , and let λn denote the regular representation of Kn
on L2 (Kn , mm ). For p ≥ 1, let K
`n,p denote a copy of Kn endowed with
the measure 2−n−p mn . Set X = n,p Kn,p , endowed with the H-invariant
probability measure µ = ⊕n,p 2−n−p mn . Note that the H-representations on
L2 (X, µ) and on L20 (X, µ) are equivalent, as L2 (X, µ) contains the trivial
representation with infinite multiplicity.
So it is enough to prove that the H-representation π on L2 (X, µ) is equivalent to ⊕∞
n=1 ∞·σn . To see this, first observe that π is equivalent to ⊕n ∞·πn ,
where πn = λn ◦ σn . By Peter-Weyl, πn decomposes as a direct sum of
finite-dimensional irreducible representations of H, hence of certain σk ’s,
and moreover σn is a sub-representation of πn (because the natural representation of Kn on CNn is irreducible, hence appears as a sub-representation
of λn ). This shows that ⊕n ∞ · πn is equivalent to ⊕n ∞ · σn .
To prove Theorem 1.5, observe that a discrete Kazhdan group G satisfies
the assumption of lemma 4.1 (by Theorem 2.6 in [12]). Let (σn )n∈N be an
enumeration of all finite-dimensional irreducible unitaryLrepresentations of
∞
G. By Theorem 1.4 and its proof, the representation
n=1 ∞ · σn has a
comeager conjugacy class.
In particular, we get the following statement which was proved in [4] only
for finite abelian groups.
Corollary 4.2. Let G be a finite group. Then the set of unitary representations realizable by an action is comeager in Rep(G, H).
Remark 4.3. Kechris proves (see section (F) in Appendix H of [8]) that, if G
is torsion-free abelian, then the set of representations realizable by an action
is meager in Rep(G, H).
PROPERTY (T) AND UNITARY REPRESENTATIONS
11
References
[1] R. J. Archbold, On residually finite-dimensional C*-algebras, Proc. Amer. Math. Soc.
123 (1995), no. 9, 2935–2937.
[2] B. Bekka, P. de la Harpe, A. Valette, Kazhdan’s property (T), New Mathematical
Monographs, 11. Cambridge University Press, Cambridge, 2008.
[3] B. Bekka, On the full C*-algebras of arithmetic groups and the congruence subgroup
problem, Forum Math. 11 (1999), no. 6, 705–715.
[4] M. Doležal, Unitary representations of finite abelian groups realizable by an action,
Topology Appl. 164 (2014), 87–94.
[5] M. Ershov, A. Jaikin-Zapirain, Groups of positive weighted deficiency and their applications, J. reine angew. Math. 677 (2013), 71–134.
[6] R. Exel, T. Loring, Finite-dimensional representations of free product C*-algebras,
Internat. J. Math. 3 (1992), no. 4, 469–476.
[7] A. Kechris, Classical descriptive set theory, Graduate Texts in Mathematics, 156.
Springer-Verlag, New York, 1995.
[8] A. Kechris, Global aspects of ergodic group actions, Mathematical Surveys and Monographs, 160. American Mathematical Society, Providence, RI, 2010.
[9] D. Kerr, H. Li, M. Pichot, Turbulence, representations, and trace-preserving actions,
Proc. Lond. Math. Soc. (3) 100 (2010), no. 2, 459–484.
[10] A. Lubotzky, Y. Shalom, Finite representations in the unitary dual and Ramanujan
groups, Discrete geometric analysis, 173–189, Contemp. Math., 347, Amer. Math.
Soc., Providence, RI, 2004.
[11] V. Pestov, V. Uspenskij, Representations of residually finite groups by isometries of
the Urysohn space, J. Ramanujan Math. Soc. 21 (2006), no. 2, 189–203.
[12] P. S. Wang, On isolated points in the dual spaces of locally compact groups, Math.
Ann. 218 (1975), no. 1, 19–34.
[13] R.J. Zimmer, Ergodic theory and semisimple groups, Mongraphs in Math., Springer,
1984.
Institute of Mathematics, Czech Academy of Sciences, Žitná 25, 115 67
Praha 1, Czech republic
E-mail address: [email protected]
Department of Mathematics and Mathematical Economics, Warsaw School
of Economics, al. Niepodleglości 162, 02-554 Warsaw, Poland
E-mail address: [email protected]
Institut de Mathématiques, Université de Neuchâtel, Unimail, 11 Rue Emile
Argand, 2000 Neuchâtel, Switzerland
E-mail address: [email protected]
| 4math.GR
|
On the Complexity of Realizing Facial Cycles?
Giordano Da Lozzo1 and Ignaz Rutter2
arXiv:1607.02347v1 [cs.CG] 8 Jul 2016
1
Department of Engineering, Roma Tre University, Italy
2
Karlsruhe Institute of Technology, Germany
Abstract. We study the following combinatorial problem. Given a planar graph G = (V, E) and a set of simple cycles C in G, find a planar
embedding E of G such that the number of cycles in C that bound a
face in E is maximized. We establish a tight border of tractability for
this problem in biconnected planar graphs by giving conditions under
which the problem is NP-hard and showing that relaxing any of these
conditions makes the problem polynomial-time solvable. Moreover, we
give a 2-approximation algorithm for series-parallel graphs and a (4 + ε)approximation for biconnected planar graphs.
1
Introduction
A planar graph is a graph that can be embedded into the plane without crossings.
While there exist infinitely many such embeddings, the embeddings for connected
graphs can be grouped into finitely many equivalence-classes of combinatorial
embeddings, where two embeddings are equivalent if the clockwise cyclic order of
the edges around each vertex is the same. Since a graph may admit exponentially
many different such embeddings, many drawing algorithms for planar graphs
simply assume that one embedding has been fixed beforehand and draw the
graph with this fixed embedding. Often, however, the quality of the resulting
drawing depends strongly on this embedding; examples are the number of bends
in orthogonal drawings, or the area requirement of planar straight-line drawings.
Consequently, there is a long line of research that seeks to optimize quality
measures over all combinatorial embeddings. Not surprisingly, except for a few
notable cases such as minimizing the radius of the dual graph [1,3,13], many
of these problems have turned out to be NP-complete. For example it is NPcomplete to decide whether there exists a planar embedding that allows for a
planar orthogonal drawing without bends or for an upward planar drawing [11].
While there has been quite a bit of work on solving these problems for special
cases, e.g., for the orthogonal bend minimization problem [4,5], to the best of
our knowledge, approximation algorithms have rarely been considered.
Another way of describing a combinatorial embedding of a connected graph
G is by describing its facial walks, i.e., by listing the walks of G that bound
a face. In the case of biconnected planar graphs, the facial walks are simple,
?
This work was partially supported by MIUR Project “AMANDA” under PRIN
2012C4E3KT and by DFG grant WA 654/21-1.
and we refer to them as facial cycles. In this paper we consider the problem of
optimizing the set of facial cycles, i.e., given a list C of cycles in a biconnected
graph G, we seek an embedding E of G such that as many cycles of C as possible
are facial cycles of E. The research on this problem was initiated by Mutzel
and Weiskircher [14], who gave an integer linear program (ILP) for a weighted
version of the problem. Woeginger [16] showed that the problem is NP-complete
by showing that it is NP-complete to maximize the number of facial cycles
that have size at most 4. Da Lozzo et al. [6] consider the problem of deciding
whether there exists an embedding such that the maximum face size is k. They
give polynomial-time algorithms for k ≤ 4 and show NP-hardness for k ≥ 5
and give a factor-6 approximation for minimizing the size of the largest face.
Finally, Dornheim [9] studies a decision problem subject to so-called topological
constraints, which specify for certain cycles of a planar graph two subsets of
edges of the graph that have to be embedded inside and outside the respective
cycle; note that a cycle is a facial cycle if its interior is empty. He proved NPcompleteness and reduced the connected case to the biconnected case.
We note that, given a biconnected planar graph G and a set C of cycles of
G, it can be efficiently decided whether there exists a planar embedding of G
in which all cycles of C are facial cycles; for each cycle C ∈ C, we subdivide
each edge of C once and connect the subdivision vertex to a new vertex vC . If
the resulting graph is planar, the desired embedding of G can be obtained by
removing all vertices vC and their incident edges.
Contribution and Outline. In this paper, we thoroughly study the problem
Max Facial C-Cycles of maximizing the number of cycles from a given set C
that bound a face of a biconnected planar graph. We start with preliminaries
concerning graphs and their combinatorial embeddings in Section 2. In Section 3
we show that Max Facial C-Cycles is NP-complete even if each cycle in C
intersects any other cycle in C in at most two vertices and intersects at most
three other cycles of C. In Section 4 we complement these results with efficient
algorithms for series-parallel and general planar graphs when the cycles intersect
only few other cycles in more than one vertex. Finally, in Section 5, we develop
an efficient approximation algorithm for the problem. For series-parallel graphs
we give a 2-approximation, and for biconnected planar graphs we achieve a
(4 + ε)-approximation for ε > 0.
2
Preliminaries
A planar drawing Γ of a graph maps vertices to points in the plane and edges to
internally disjoint curves. Drawing Γ partitions the plane into topologically connected regions, called faces. The bounded faces are internal and the unbounded
face is the outer face. A planar drawing determines a circular ordering of the
edges incident to each vertex. Two planar drawings of a connected planar graph
are equivalent if they determine the same orderings and have the same outer
face. A combinatorial embedding is an equivalence class of planar drawings.
2
v4
C4
C2
v1
v2
C3
C1 e
ve
v3
Fig. 1: Illustrations for the proof of Theorem 1. (a) Graph H (black) and its
planar dual H ∗ (red). Vertex v1 is the only vertex in the MIS of H. (b) An
embedding of graph G in which cycle C1 (corresponding to the face of H ∗ that
is dual to the vertex v1 of H) bounds a face.
For the definition of the SPQR-tree of a biconnected graph and the concepts
of skeleton skel(µ) and pertinent graph pert(µ) of a node µ of an SPQR-tree, and
that of virtual edge of a skeleton, and expansion graph of a virtual edge we refer
the reader to [6]; for convenience we also provide a definition in Appendix A.
3
Complexity
In this section we study the computational complexity of the underlying decision problem Facial C-Cycles of Max Facial C-Cycles, which given a
biconnected planar graph G, a set C of simple cycles of G, and a positive integer
k ≤ |C| asks whether there exists a planar embedding E of G such that at least
k cycles in C are facial cycles of E. Facial C-Cycles is in NP, as we can guess
a set C 0 ⊆ C of k cycles and then check whether an embedding of G exists in
which all cycles in C 0 are facial cycles in polynomial time. We show NP-hardness
for general graphs and for series-parallel graphs.
Theorem 1. Facial C-Cycles is NP-complete, even if each cycle C ∈ C
(i) intersects any other cycle in C in at most two vertices, and
(ii) intersects at most three other cycles of C in more than one vertex.
Proof (sketch). We give a reduction from Maximum Independent Set in 3connected cubic planar graphs, which we recently showed to be NP-complete [7].
Let H be a 3-connected cubic planar graph. Observe that H has a unique combinatorial embedding up to a flip. We construct an instance hG, C, ki of Facial
C-Cycles as follows; see Fig. 1. Take the planar dual H ? of H and take C as
the set of facial cycles of H ? . Observe that H ? is a planar triangulation, since
H is cubic and 3-connected. The graph G is obtained from H ? by adding for
each edge e = uv ∈ E(H ? ) an edge vertex ve with neighbors u and v. It is not
hard to see that H admits an independent set of size k if and only if G admits
a combinatorial embedding where k cycles in C are facial (see appendix). By
construction C satisfies the restrictions in the statement of the theorem.
t
u
3
ub
2
1
2
ud
ua1
ua
3
3
2
3
1
1
a
s
sa
3
1
2
uc
ua2 ua3
uc1 uc3 uc2 ua3 ua1 ua2 ub1 ub3 ub2 ud2 ud3 ud1
vc
va
ta
va
vb
vd
t
b
c
Fig. 2: Illustrations for the proof of Theorem 2. (a) Cubic graph H with a Hamiltonian circuit Q (thick, colored edges). (b) Gadget Ga for a vertex a ∈ V (H).
(c) Combinatorial embedding of G corresponding to the Hamiltonian circuit Q
(facial cycles have the same color as the corresponding edge in Q).
Theorem 2. Facial C-Cycles is NP-complete for series-parallel graphs, even
if any two cycles in C share at most three vertices.
Proof (sketch). We reduce from Hamiltonian Circuit, which is known to be
NP-complete even for cubic graphs [10]. Let H be any such a graph.
Each vertex a ∈ V (H) is represented by the following gadget Ga . It consists of
the graph K2,3 , where the vertices in the partition of size 2 are denoted sa and v a
and the other vertices are denoted ua1 , ua2 , ua3 , and of an additional vertex ta
adjacent to v a ; see Fig. 2b. The graph G is obtained by merging the vertices sa
into a single vertex s and the vertices ta into a single vertex t.
To define C, we number the incident edges of each vertex of H from 1 to 3.
If ab is the i-th edge for a and the j-th edge for b, we define Cab ∈ C as the
cycle (s, uai , v a , t, v b , ubj , s); see Fig. 2a and 2c. We claim that G admits a combinatorial embedding with |V (H)| facial cycles in C if and only if H is Hamiltonian.
If Q is a Hamiltonian circuit of H, we embed G such that the order of
the gadgets Ga is the same as the order of the vertices in Q. We then choose
embedding of the gadgets such that for each edge ab of Q the cycle Cab bounds
the face between Ga and Gb ; this yields the claimed number of facial cycles
in C. Conversely, observe that if Cab is a facial cycle of an embedding of G,
then Ga and Gb , where ab is an edge of H, must be consecutive in the circular
order around s. If G has |V (H)| facial cycles in C it follows that the vertices
corresponding to the gadgets form a Hamiltonian circuit in this order.
t
u
4
Polynomial-time Solvable Cases
In this section we discuss special cases of Max Facial C-Cycles that admit a
polynomial-time solution. In particular, we show that strengthening any of the
conditions in Theorem 1 or Theorem 2 makes the problem tractable.
4
4.1
General Planar Graphs
In this section we study Max Facial C-Cycles when each cycle in C intersects
at most two other cycles in C in more than one vertex. In this setting we give in
Theorem 5 a linear-time algorithm for biconnected planar graphs. Further, for
the class of series-parallel graphs we present in Theorem 4 an FPT-algorithm
with respect to the maximum number of cycles in C sharing two or more vertices
with any cycle in C. We remark that our algorithms imply that strengthening
any of the two conditions of Theorem 1 results in a polynomial-time solvable
problem. In particular, Max Facial C-Cycles is polynomial-time if any two
cycles in C share at most one vertex.
We compute the optimal solution in these cases by a dynamic program that
works bottom-up in the SPQR-tree T of G. Let µ be a node of T . We call a
cycle C ∈ C relevant for µ (or for skel(µ)) if it projects to a cycle in skel(µ), that
is, the vertices of C in skel(µ) and the edges of skel(µ) that contain vertices or
edges in C form a cycle C 0 in skel(µ) with at least two edges. The cycle C 0 is
the projection of the cycle C in skel(µ). Similarly, we also define the projection
of a cycle C ∈ C to pert(µ). We denote the set of relevant cycles and of interface
cycles of a node µ by R(µ) and by I(µ), respectively. Clearly, I(µ) ⊆ R(µ). We
denote I(µ) = {X ⊆ I(µ) | |X| ≤ 2}.
Let µ be a node of T . We have the following two important observations.
Observation 1 If each cycle in C intersects at most two other cycles in C in
more than one vertex, then |I(µ)| ≤ 3.
Observation 2 In any combinatorial embedding E of G at most two interface
cycles of µ can simultaneously bound a face in E.
Observation 1 holds since all interface cycles of a node µ share at least the
poles of µ. Observation 2 holds since each interface cycle can only bound one of
the two faces incident to the virtual edge representing the parent of µ in skel(µ).
Thus to the rest of G the only
u
relevant information about a combinatorial embedding of pert(µ) is
(a) the number of facial cycles in C
and (b) the set of cycles in C projectv
ing to the facial cycles incident to the
parent edge.
Fig. 3: Graph and a P-node skeleton
If E is a combinatorial embedding (shaded) with three virtual edges corof pert(µ) and the elements of I ∈ responding to children from left to right
I(µ) has project to distinct faces inci- realizing none, the green and the red and
dent to the parent edge in pert(µ), we only the red cycle, respectively. The red
say that E realizes I; see Fig. 3.
cycle bounds a face since, in addition,
For any node µ and any set I ∈ the second and third child are adjacent
I(µ), we denote by T [µ, I] the maxi- in the embedding of the skeleton.
mum number k such that there exists
a combinatorial embedding E of pert(µ) that realizes I and such that k cycles
5
in C bound a face of E that is not incident to the parent edge of pert(µ). If no
such embedding exists, we set T [µ, I] = −∞. Due to Observation 2, for convenience we extended the definition of T to the case in which the size of I is larger
than 2; in this case, we define T [µ, I] = −∞.
We show how to compute the entries of T in a bottom-up fashion in the
SPQR-tree T of G. It is not hard to modify the dynamic program to additionally
output a corresponding combinatorial embedding of G. We root T at an arbitrary
Q-node ρ. Let φ be the unique child of ρ. Note that the maximum number of
facial cycles in C for any combinatorial embedding of G is maxI∈I(φ) |I| + T [φ, I].
For any leaf Q-node µ, we have that T [µ, I] = 0 for each I ∈ I(µ). The following
lemmata deal with the different types of inner nodes in an SPQR-tree.
Lemma 1. Let µ be an S-node with children µi , i = 1, . . . , k. Then T [µ, I] =
Pk
i=1 T [µi , I], for I ∈ I(µ). Each entry T [µ, I] can be computed in O(k) time.
Proof. The lemma follows easily from the observation that a combinatorial embedding of pert(µ) realizes I if and only if each of its children realizes I.
t
u
Lemma 2. Let µ be a P-node with children µ1 , . . . , µk . Then
T [µ, I] =
max
k
X
(
T [µi , Cµi ] + f (C)) ,
I⊆C⊆R(µ)
i=1
where (i) Cµi = C ∩ I(µi ) and (ii) f (C) = |C \ I| if skel(µ) admits a planar
embedding E where (a) each two virtual edges ei and ej corresponding to children
µi and µj of µ, respectively, such that |Cµi ∩ Cµj | = 1 are adjacent in E, and
where (b) the virtual edges e0 and e00 corresponding to the children µ0 and µ00 of
µ such that Cµ0 ∩ I 6= ∅ and Cµ00 ∩ I 6= ∅, respectively, are incident to the outer
face of E, and f (C) = −∞ otherwise.
Proof. Consider an embedding of pert(µ) that embeds T [µ, I] cycles of C as facial
cycles and the corresponding embedding E of skel(µ). Let C ⊆ R(µ) denote the
set of cycles in C that are facial cycles in E or that are in I. Obviously, to make
a cycle c ∈ C \ I a facial cycle, each of the two children of µ that contain c
in their interface (i) must be adjacent in E and (ii) must both realize cycle c.
Also, in order for the cycles in I to bound the outer-face of the embedding
of pert(µ), the two children of µ containing such interface cycles (i) must be
incident to the outer-face of E and (ii) must each realize one of these cycles in
their interface. Hence T [µ, C] is a lower bound on the number of facial cycles
in C in the embedding of pert(µ). On the other hand, it is not hard to see that
by picking the maximum over all subsets C ⊆ R(µ) this bound is attained for
the correct set of cycles C.
t
u
We note that the existence of a corresponding embedding for a P-node µ with
k children can be tested in O(k) time for any set C ⊆ R(µ), thus allowing us
to evaluate f (C) efficiently as follows. Consider the auxiliary multigraph O that
contains a vertex for each virtual edge of skel(µ), except for the edge representing
6
the parent of µ, and two such edges are adjacent if and only if there is a cycle
in C \ I that contains edges from both expansion graphs. Also, if there exist two
virtual edges in skel(µ) containing edges from cycles in I, multigraph O contains
an edge connecting them. A corresponding embedding exists if and only if O
is either a simple cycle or it is a collection of paths. In latter case, O can be
augmented to a simple cycle and the order of the virtual edges along this cycle
defines a suitable embedding of skel(µ).
Generally, the number of cycles in R(µ) can be large. However, if every cycle
C ∈ C shares two or more vertices with at most r other cycles in C, the running
time can be bounded as follows.
Lemma 3. Let µ be a P-node with children µ1 , . . . , µk such that any cycle
of R(µ) shares two or more vertices with at most r other cycles in R(µ). For
each set I ∈ I(µ), table T [µ, I] can be computed in O(r2 2r · k) time from T [µi , ·]
with i = 1, . . . , k.
Proof. We employ Lemma 2. It is |R(µ)| ≤ r + 1, and |I(µ)| = O(r2 ). For each
I ∈ I(µ) we need to consider all the sets C ⊆ R(µ) such that I ⊆ C. There are
O(2r ) such sets C and for each of them we evaluate f (C) in O(k) time.
t
u
We now deal with R-nodes. Let µ be an R-node with k children µ1 , . . . , µk ,
let I ∈ I(µ) and let C ⊆ R(µ) with C ⊇ I be a set of cycles that project to
distinct facial cycles of skel(µ). Note that relevant cycles of µ that do not project
to a facial cycle of skel(µ) can never bound a face, and we can hence assume
that such cycles have been removed from C in a preprocessing step, i.e., every
relevant cycle of µ projects to a facial cycle of skel(µ). We define
gain(C, I) =
k
X
i=1
(T [µi , C ∩ I(µi )] − T [µi , I ∩ I(µi )]) + |C \ I(µ)|.
It is not hard to see that, for two such sets of cycles C1 , C2 ⊆ R(µ) with I ⊆
(C1 ∪ C2 ) and such that no two cycles C 0 ∈ C1 and C 00 ∈ C2 share a virtual edge
of skel(µ), we have gain(C1 ∪ C2 , I) = gain(C1 , I) + gain(C2 , I).
Let H be the subgraph of the dual of skel(µ) induced by the faces that
are projections of cycles in R(µ). Since we assume that any two cycles in C
share two or more vertices with at most two other cycles in C, the maximum
degree of H is at most 2. Once we have chosen an interface I ∈ I(µ) and for
each of the connected components H1 , . . . , Hc of H a set Ci of cycles that we
want
PcHi , the overall number of faces can be expressed
Pkto realize as faces for
as i=1 T [µi , I ∩ I(µi )] + i=1 gain(Ci ). In particular,
T [µ, I] =
k
X
i=1
T [µi , I ∩ I(µi )] +
I⊆
Sc max
i=1 Ci ⊆R(µ)
c
X
gain(Ci , I),
i=1
where the maximization considers only those sets Ci whose cycles project to
distinct faces of skel(µ) and whose dual vertices are in Hi .
7
It thus remains to choose for each connected component Hi of H a set of
cycles in C that project to faces that are vertices of Hi and that maximize the
gain. We exploit the fact that these graphs have maximum degree 2 to give an
efficient algorithm via dynamic programming.
Let H 0 be such a connected component, which is either a path or a cycle. We
observe that, if H contains vertices corresponding to the faces incident to the
parent edge, then they are contained in the same connected component of H.
In the following, we assume that H 0 does not contain these vertices. The other
case can be treated similarly, but requires also to take into account that a set
I ∈ I(µ) of cycles has to be realized.
Assume that H 0 is a path v1 , . . . , vh . Each vertex vi is associated with a
set Ci ⊆ C of potential cycles that can realize the face that corresponds to vi .
Observe that, since each such vertex corresponds to a facial cycle that is the
projection of some cycle in C, it follows that |Ci | = 1 for i = 2, . . . , h − 1 and
that |C1 |, |Ch | ≤ 2 for h ≥ 2 and |C1 | ≤ 3 if h = 1. Otherwise one such cycle
would intersect too many other cycles in two or more vertices.
We now compute the optimal solution by dynamic programming along the
path. More precisely, we define P [i, C 0 ] with i ∈ {1, . . . , h} and C 0 ⊆ Ci with |C 0 | ≤
1 as the maximum gain obtainable by any set of cycles in C1 ∪ · · · ∪ Ci−1 ∪ C 0 .
Clearly P [1, ∅] = 0 and P [1, C] = gain({C}). For i > 1, observe that P [i, ∅] =
maxC 0 ⊆Ci−1 ,|C 0 |≤1 P [i−1, C 0 ] and for C ∈ Ci , it is P [i, {C}] = maxC 0 ⊆Ci−1 ,|C 0 |≤1 P [i−
1, C 0 ] + gain(C 0 ∪ {C}, C 0 ). Note that gain(C 0 ∪ {C}, C 0 ) describes the gain of realizing C in addition to C 0 .
This recurrence allows us to compute the optimal gain value in O(k) time
if H 0 is a path of length k. Now assume that H 0 is a cycle of length k. Observe
that, in this case, each facial cycle has at most one candidate cycle in C. We
exploit that either all these cycles are chosen, or at least one of them is not
chosen. It is not hard to compute the gain of the solution that chooses all facial
cycles. Further, we try each facial cycle as the one that is not chosen, leaving us
with an instance that forms a path, to which we apply the previous algorithm.
Altogether, in this way we can compute the optimal gain value when H 0 is a
cycle of length k in O(k 2 ) time. It is not hard to adapt the dynamic program to
realize a given set of cycles in I(µ). We thus have the following lemma.
Lemma 4. Let µ be an R-node with children µ1 , . . . , µk . There is an O(k 2 )-time
algorithm for computing T [µ, ·] from T [µi , ·] for i = 1, . . . , k, provided that cycles
in C shares two or more vertices with at most two other cycles from C.
Altogether, Lemmas 1, 3, and 4 imply the following theorem.
Theorem 3. Max Facial C-Cycles can be solved in O(n2 ) time if every cycle
in C intersects at most two other cycles in more than one vertex.
4.2
Series-Parallel Graphs
In this section, we consider Max Facial C-Cycles on series-parallel graphs.
Note Combining the results from Lemma 1 and Lemma 3 yields the following.
8
Theorem 4. Max Facial C-Cycles is solvable in O(r2 2r · n) time for seriesparallel graphs if any cycle in C intersects at most r other cycles.
Corollary 1. Max Facial C-Cycles is solvable in O(n) time for series-parallel
graphs if any cycle in C intersects at most two other cycles.
In the following we show that Max Facial C-cycles can be solved in polynomial time for series-parallel graphs if any two cycles in C share at most two
vertices. The next lemma shows the special structure of relevant cycles in Pnodes of the SPQR-tree in this case.
Lemma 5. Let G be a series-parallel graph and C be a set of cycles in G such
that any two cycles share at most two vertices. For each P-node µ any two
relevant are either edge-disjoint in skel(µ) or they share the unique virtual edge
of skel(µ) that corresponds to a Q-node child of µ, if any.
We again use a bottom-up traversal traversal of the SPQR-tree of a seriesparallel graph to obtain the following theorem. The S-nodes are handled using
Lemma 1 and the structural properties guaranteed by Lemma 5 allow for a
simple handling of the P-nodes.
Theorem 5. Max Facial C-Cycles is solvable in O(n) for series-parallel
graphs if any two cycles in C share at most two vertices.
5
Approximation Algorithms
In this section we derive constant-factor approximations for Max Facial CCycles in series-parallel graphs and in biconnected planar graphs. Again, we use
dynamic programming on the SPQR-tree. This time, however, instead of computing T [µ, I], we compute an approximate version T̃ [µ, I] of it. A table T̃ [µ, ·]
is a c-approximation of T [µ, ·] if 1/c · T [µ, I] ≤ T̃ [µ, I] ≤ T [µ, I], for all I ∈ I(µ).
For P-nodes, we give an algorithm that approximates each entry within a factor
of 2, for R-nodes, we achieve an approximation ratio of (4 + ε) for any ε > 0. In
the following lemmas we deal separately with S-, P-, and R-nodes.
Lemma 6. Let µ be an S-node with children µ1 , . . . , µk . Assume that T̃ [µi , I] is
Pk
a c-approximation of T [µi , I] for i = 1, . . . , k. Then setting T̃ [µ, I] = i=1 T̃ [µi , I]
yields a c-approximation of T [µ, I].
Pk
Proof. To see this, observe that by Lemma 1, it is 1/c·T [µ, I] = 1/c· i=1 T [µi , I] ≤
Pk
Pk
Pk
t
u
i=1 T [µi , I] = T [µ, I].
i=1 T̃ [µi , I] and
i=1 T̃ [µi , I] ≤
Next we deal with a P-node µ with children µ1 , . . . , µk . The algorithm works
as follows. Fix an set I ∈ I(µ). We construct an auxiliary graph H as follows.
The vertices of H are the children µ1 , . . . , µk of µ. Two vertices µi and µj
are adjacent in H if and only if there exists a cycle C ∈ C that intersects µi
and µj such that T̃ [µx , I ∩ I(µx ) ∪ {C}] = T̃ [µx , I ∩ I(µx )] for x ∈ {i, j}, i.e.,
9
according to the approximate table T̃ additionally realizing C in the interface of
the children µi and µj does not cause additional costs. If |I| = 2, assume that µ1
and µ2 are the two children intersected by the cycles in I. Unless µ1 and µ2
are the only children of µ, we remove the edge µ1 µ2 from H if it is there. This
reflects the fact that, due to the restrictions imposed by I, it is not possible to
realize a corresponding cycle. Now compute a maximum matching M in H. The
matching M corresponds to a set CM ⊆ R(µ) of relevant cycles of µ that are
Pk
pairwise edge-disjoint. We set T̃ [µ, I] = i=1 T̃ [µi , (I ∪ CM ) ∩ I(µi )] + |M |.
We claim that this gives a max{2, c}-approximation of T [µ, ·] if the input is
a c-approximation for T [µi , ·] for i = 1, . . . , k.
Lemma 7. Let µ be a P-node an let T̃ [µ, ·] denote the table computed in the
above fashion. Then T̃ [µ, ·] is a max{2, c}-approximation of T [µ, ·] if T̃ [µi , ·] is
a c-approximation of T [µi , ·].
Proof. We first show that T̃ [µ, ·] ≤ T [µ, ·]. To this end, it suffices to show that, for
any I ∈ I(µ), there exists an embedding of pert(µ) that realizes I and has T̃ [µ, I]
realized cycles from C. Consider the multigraph with vertex set {µ0 , µ1 , . . . , µk }
and edge set CM ∪ I. This graph has maximum degree 2 and, due to our special
treatment, unless k = 2, none of its connected components is a cycle. We can
thus always complete this graph into a cycle, which defines a circular order
of µ0 , . . . , µk , and hence an embedding of skel(µ). In this embedding, all the
cycles in CM ∪I project to facial cycles. Realizing all these cycles yields T̃ [µ, I] =
Pk
Pk
i=1 T̃ [µi , (I ∪ CM ) ∩ I(µi )] + |M | ≤
i=1 T [µi , (I ∪ CM ) ∩ I(µi )] + |M | realized
cycles. By the definition of the T [µi , ·] we get embeddings for the pert(µi ) with
a corresponding number of cycles in C and by combining them according to
the embedding of skel(µ) chosen above we obtain an embedding of pert(µ) that
realizes I and has at least T̃ [µ, I] facial cycles in C. Hence T̃ [µ, I] ≤ T [µ, I].
Conversely, consider T [µ, I] and a corresponding embedding of skel(µ). Denote by Copt the set of cycles realized by an optimal solution that project to
facial cycles of skel(µ). We consider two cycles in Copt as adjacent if they intersect the same child of µ. Clearly, each child µi is intersected by at most two
cycles in Copt and, moreover, the two faces of skel(µ) incident to the parent edge
are not realized. Hence the corresponding graph is a collection of paths. It is
0
hence possible to edge-color it with two colors. Let Copt
be the cycles in the
0
0
larger color class. We have |Copt | ≥ |Copt |/2 and no two distinct cycles in Copt
0
intersect the same child µi of µ, i.e., interpreting the cycles in Copt as edges on
the vertex set {µ1 , . . . , µk } yields a matching M 0 . We would like to argue that
our matching M in the auxiliary graph H is larger than M 0 , and hence we realize
at least half of the cycles of the optimum. However, this argument is not valid,
since M 0 may contain edges that are not present in H due to approximation errors in the T̃ [µi , ·]. We will show that the contribution of these edges is irrelevant
and hence the intuition about comparing the matching sizes indeed applies.
Let M10 = M 0 \ E(H) and M20 = M 0 ∩ E(H). Let J = {1, . . . , k} and
let J1 = {i ∈ J | ∃C ∈ M10 that intersects µi } be the indices of children that are
intersected by a cycle in M10 . The set J2 = J \ J1 contains the remaining indices.
10
Pk
Clearly, we have T [µ, I] = i=1 T [µi , (I ∪ Copt ) ∩ I(µi )] + |Copt | according
0
to Lemma 2. Realizing instead of Copt just the set of cycles CM 0 = Copt
corre0
sponding to M drops at most |Copt |/2 facial cycles in C, while imposing weaker
interface constraints on the children. We therefore have
k
X
T [µ, I] =
i=1
k
X
T [µi , (I ∪Copt )∩I(µi )]+|Copt | ≤
i=1
T [µi , (I ∪CM 0 )∩I(µi )]+2|M 0 |
We now use the fact that the T̃ [µi , ·] are a c-approximation of the T [µi , ·], and
hence also a c0 -approximation for c0 = max{c, 2}, and we also separate the sum
by the index set J1 and J2 and consider the two matchings M10 and M20 separately.
k
X
i=1
T [µi , (I ∪ CM 0 ) ∩ I(µi )] + 2|M 0 | ≤ c0 ·
0
+c ·
X
i∈J2
X
i∈J1
T̃ [µi , (CM10 ∪ I) ∩ I(µi )] + 2|M10 |
T̃ [µi , (CM20 ∪ I) ∩ I(µi )] + 2|M20 |.
(1)
Observe that the indices of the children intersected by cycles that form a matching M2 in H are all contained in J2 . By the definition of H, we thus have T̃ [µi , (CM20 ∪
I) ∩ I(µi )] = T̃ [µi , I ∩ I(µi )], for i ∈ J2 .
For the first term, observe that, for each edge µi µj ∈ M10 , we have T̃ [µx , (M10 ∪
I) ∩ I(µx )] ≤ T̃ [µx , I ∩ I(µx )] − 1 for at least one x ∈ {i, j}. Otherwise the edge
would be in H, and hence in M20 . Let J10 ⊆ J1 denote the set of indices where
this happens and let J100 = J1 \ J10 . Observe that |J10 | ≥ |M10 |. We thus have
c0 ·
X
i∈J1
= c0 ·
≤ c0 ·
T̃ [µi , (CM10 ∪ I) ∩ I(µi )] + 2|M10 |
X
T̃ [µi , (CM10 ∪ I) ∩ I(µi )] + c0 ·
i∈J10
X
i∈J10
(T̃ [µi , I ∩ I(µi )] − 1) + c0 ·
X
i∈J100
X
i∈J100
T̃ [µi , (CM10 ∪ I) ∩ I(µi )] + 2|M10 |
T̃ [µi , I ∩ I(µi )] + 2|M10 |
!
≤ c0 ·
X
i∈J1
T̃ [µi , I ∩ I(µi )] − |J10 |
+ 2|M10 | ≤ c0 ·
X
i∈J1
T̃ [µi , I ∩ I(µi )],
where the last step uses the fact that c0 ≥ 2. Plugging this information into
Eq. 1, yields the following.
X
X
c0 ·
T̃ [µi , (CM10 ∪ I) ∩ I(µi )] + 2|M10 | + c0 ·
T̃ [µi , (CM20 ∪ I) ∩ I(µi )] + 2|M20 |
i∈J1
≤ c0 ·
0
≤c ·
i∈J2
k
X
i=1
T̃ [µi , I ∩ I(µi )] + 2|M20 | ≤ c0 ·
k
X
i=1
k
X
i=1
!
T̃ [µi , I ∩ I(µi )] + |M |
11
T̃ [µi , I ∩ I(µi )] + 2|M |
Here the last two steps use the fact that M ⊆ E(H) is a maximum matching,
and hence larger than M20 , and that c0 ≥ 2, respectively.
t
u
We note that the bottleneck for computing T [µ, I] is finding a maximum
matching in a graph with O(| skel(µ)|) vertices
and O(|C|) edges. Hence the
p
≤ |C|2 ,
running time for one step is O(| skel(µ)| + | skel(µ)| · |C|). Since |I(µ)|
p
the running time for processing a single P-node µ is O(| skel(µ)||C|2 √
+ | skel(µ)|·
|C|3 ). The total time for processing all P-nodes then is O(n|C|2 + n|C|3 ).
Theorem
6. There is a 2-approximation algorithm with running time O(n|C|2 +
√
3
n|C| ) for Max Facial C-Cycles in series-parallel graphs.
Next we deal with R-nodes. Let µ be an R-node with children µ1 , . . . , µk and
let J = {1, . . . , k}. For each face f of skel(µ) let Jf denote the indices of the
children µi whose corresponding virtual edge in skel(µ) is incident to f .
Fix I ∈ I(µ). We propose the following algorithm for computing T̃ [µ, I].
Consider the subgraph H of the dual of skel(µ) induced by those vertices v
corresponding to a face f not incident to the parent edge of skel(µ) and such
that there exists a cycle Cv ∈ C that projects to the boundary of f and such
that T̃ [µi , ({Cv } ∪ I) ∩ I(µ)] = T̃ [µi , I ∩ I(µ)], i.e., requiring that Cv is realized
in µi does not change the approximate number of faces realized by pert(µi ).
Now we compute a (1 + ε/4)-approximation of a maximum independent
set of H, which can be done in time polynomial in | skel(µ)| (and exponential
in (1/ε)) [2]. Let X denote this independent set, and let CX = {Cv | v ∈ X} be a
Pk
set of corresponding cycles in C. We set T̃ [µ, I] = i=1 T̃ [µi , (I∪X)∩I(µi )]+|X|,
and claim that in this fashion T̃ [µ, ·] is a max{c, (4 + ε)}-approximation provided that T̃ [µi , ·] is a c-approximation of T [µi , ·]. The proof 4-colors the facial
cycles Copt that are realized by an optimal solution and considers the largest
color class, which is an independent set of size at least |Copt |/4. The proof is
similar to that of Lemma 7.
Lemma 8. Let T̃ [µ, ·] denote the table computed in the above fashion. Then T̃ [µ, ·]
is a max{c, (4 + ε)}-approximation of T [µ, ·] provided that T̃ [µi , . . . ] is a capproximation of T [µi , ·].
Overall, we obtain the following theorem.
Theorem 7. Max Facial C-Cycles for biconnected planar graphs admits an
efficient (4 + ε)-approximation algorithm for any ε > 0.
6
Conclusion
In this paper we showed NP-hardness of Max Facial C-Cycles under restrictive conditions, showed that even stronger conditions make the problem tractable
and gave constant-factor approximations for series-parallel and biconnected planar graphs with approximation guarantees of 2 and 4 + ε, respectively.
12
We remark that it is possible to adapt all our algorithmic results to the
weighted case where each facial cycle has a positive weight and one seeks a
planar embedding that maximizes the total weight of the facial cycles in C. We
leave open the question whether similar algorithmic results can be obtained for
arbitrary, not necessarily biconnected, planar graphs.
13
References
1. Angelini, P., Di Battista, G., Patrignani, M.: Finding a minimum-depth embedding
of a planar graph in O(n4 ) time. Algorithmica 60, 890–937 (2011)
2. Baker, B.S.: Approximation algorithms for np-complete problems on planar graphs.
J. ACM 41(1), 153–180 (Jan 1994), http://doi.acm.org/10.1145/174644.174650
3. Bienstock, D., Monma, C.L.: On the complexity of covering vertices by faces in a
planar graph. SIAM J. Comput. 17(1), 53–76 (1988)
4. Bläsius, T., Lehmann, S., Rutter, I.: Orthogonal graph drawing with inflexible
edges. Comput. Geom. 55, 26–40 (2016), http://dx.doi.org/10.1016/j.comgeo.
2016.03.001
5. Bläsius, T., Rutter, I., Wagner, D.: Optimal orthogonal graph drawing with convex
bend costs. In: Fomin, F.V., Freivalds, R., Kwiatkowska, M.Z., Peleg, D. (eds.)
Automata, Languages, and Programming - 40th International Colloquium, ICALP
2013, Riga, Latvia, July 8-12, 2013, Proceedings, Part I. LNCS, vol. 7965, pp. 184–
195. Springer (2013), http://dx.doi.org/10.1007/978-3-642-39206-1
6. Da Lozzo, G., Jelı́nek, V., Kratochvı́l, J., Rutter, I.: Planar embeddings with small
and uniform faces. In: Ahn, H., Shin, C. (eds.) Algorithms and Computation 25th International Symposium, ISAAC 2014, Jeonju, Korea, December 15-17, 2014,
Proceedings. LNCS, vol. 8889, pp. 633–645. Springer (2014)
7. Da Lozzo, G., Rutter, I.: Strengthening hardness results to 3-connected planar graphs. http://www.dia.uniroma3.it/~dalozzo/papers/augmentation.pdf
(2016)
8. Di Battista, G., Tamassia, R.: On-line graph algorithms with SPQR-trees. In:
Paterson, M.S. (ed.) Automata, Languages and Programming (ICALP’90). LNCS,
vol. 443, pp. 598–611. Springer (1990)
9. Dornheim, C.: Planar graphs with topological constraints. J. Graph Algorithms Appl. 6(1), 27–66 (2002), http://www.cs.brown.edu/publications/jgaa/
accepted/2002/Dornheim2002.6.1.pdf
10. Garey, M.R., Johnson, D.S., Tarjan, R.E.: The planar hamiltonian circuit problem
is NP-complete. SIAM J. Comput. 5(4), 704–714 (1976)
11. Garg, A., Tamassia, R.: On the computational complexity of upward and rectilinear
planarity testing. SIAM J. on Comput. 31(2), 601–625 (2001)
12. Gutwenger, C., Mutzel, P.: A linear time implementation of SPQR-trees. In: Marks,
J. (ed.) Graph Drawing (GD’00). LNCS, vol. 1984, pp. 77–90. Springer (2001)
13. Kammer, F.: Determining the smallest k such that G is k -outerplanar. In: Arge,
L., Hoffmann, M., Welzl, E. (eds.) Algorithms - ESA 2007, 15th Annual European
Symposium, Eilat, Israel, October 8-10, 2007, Proceedings. LNCS, vol. 4698, pp.
359–370. Springer (2007)
14. Mutzel, P., Weiskircher, R.: Optimizing over all combinatorial embeddings of a
planar graph. In: Cornuéjols, G., Burkard, R.E., Woeginger, G.J. (eds.) Integer
Programming and Combinatorial Optimization, 7th International IPCO Conference, Graz, Austria, June 9-11, 1999, Proceedings. LNCS, vol. 1610, pp. 361–376.
Springer (1999), http://dx.doi.org/10.1007/3-540-48777-8_27
15. Robertson, N., Sanders, D.P., Seymour, P., Thomas, R.: Efficiently four-coloring
planar graphs. In: STOC ’96. pp. 571–575. ACM (1996)
16. Woeginger, G.J.: Embeddings of planar graphs that minimize the number of longface cycles. Oper. Res. Lett. 30(3), 167–168 (2002), http://dx.doi.org/10.1016/
S0167-6377(02)00119-0
14
A
Connectivity and SPQR-trees
A graph G is connected if there is a path between any two vertices. A cutvertex
is a vertex whose removal disconnects the graph. A separating pair is a pair of
vertices {u, v} whose removal disconnects the graph. A connected graph is 2connected if it does not have a cutvertex and a 2-connected graph is 3-connected
if it does not have a separating pair. A 2-connected plane graph G is internally
3-connected if G can be extended to a 3-connected planar graph by adding a
vertex in the outer face and joining it to all the vertices incident to the outer
face.
We consider uv-graphs with two special pole vertices u and v, which can be
constructed in a fashion very similar to series-parallel graphs. Namely, an edge
(u, v) is an uv-graph with poles u and v. Now let Gi be an uv-graph with poles
ui , vi for i = 1, . . . , k and let H be a planar graph with two designated vertices u
and v and k + 1 edges uv, e1 , . . . , ek . We call H the skeleton of the composition
and its edges are called virtual edges; the edge uv is the parent edge and u and v
are the poles of the skeleton H. To compose the Gi into an uv-graph with poles
u and v, we remove the edge uv and replace each ei by Gi for i = 1, . . . , k by
removing ei and identifying the poles of Gi with the endpoints of ei . In fact, we
only allow three types of compositions: in a series composition the skeleton H
is a cycle of length k + 1, in a parallel composition H consists of two vertices
connected by k + 1 parallel edge, and in a rigid composition H is 3-connected.
It is known that for every 2-connected graph G with an edge uv the graph
G−st is an uv-graph with poles u and v. Much in the same way as series-parallel
graphs, the uv-graph G \ uv gives rise to a (de-)composition tree T describing
how it can be obtained from single edges. The nodes of T corresponding to edges,
series, parallel, and rigid compositions of the graph are Q-, S-, P-, and R-nodes,
respectively. To obtain a composition tree for G, we add an additional root Qnode representing the edge uv. To fully describe the composition, we associate
with each node µ its skeleton denoted by skel(µ). For a node µ of T , the pertinent
graph pert(µ) is the subgraph represented by the subtree with root µ. Similarly,
for a virtual edge ε of a skeleton skel(µ), the expansion graph of ε, denoted by
exp(ε) is the pertinent graph pert(µ0 ) of the neighbour µ0 of µ corresponding to
ε when considering T rooted at µ.
The SPQR-tree of G with respect to the edge uv, originally introduced by
Di Battista and Tamassia [8], is the (unique) smallest decomposition tree T for
G. Using a different edge u0 v 0 of G and a composition of G − u0 v 0 corresponds
to rerooting T at the node representing u0 v 0 . It thus makes sense to say that
T is the SPQR-tree of G. The SPQR-tree of G has size linear in G and can
be computed in linear time [12]. Planar embeddings of G correspond bijectively
to planar embeddings of all skeletons of T ; the choices are the orderings of the
parallel edges in P-nodes and the embeddings of the R-node skeletons, which are
unique up to a flip. When considering rooted SPQR-trees, we assume that the
embedding of G is such that the root edge is incident to the outer face, which is
equivalent to the parent edge being incident to the outer face in each skeleton.
We remark that in a planar embedding of G, the poles of any node µ of T are
15
incident to the outer face of pert(µ). Hence, in the following we only consider
embeddings of the pertinent graphs with their poles lying on the same face and
refer to such embeddings as regular.
Let µ be a node of T , we denote the poles of µ by u(µ) and v(µ), respectively.
In the remainder of the paper, we will assume edge (u(µ), v(µ)) to be part of
skel(µ) and pert(µ). The outer face of a (regular) embedding of pert(µ) is the
one obtained from such an embedding after removing the (u(µ), v(µ)) connecting
its poles. Also, the two paths incident to the outer face of pert(µ) between u(µ)
and v(µ) are called boundary paths of pert(µ).
B
Omitted Proofs from Section 3
Lemma 9. Facial C-Cycles is in NP.
Proof. Let hG, C, ki be an instance of Facial C-Cycles. A non-deterministic
Turing machine can guess in polynomial-time a combinatorial embedding E of
G and test whether at least k cycles in C are facial cycles in E.
t
u
Theorem 1. Facial C-Cycles is NP-complete, even if
(i) each cycle in C intersects any other cycle in C in at most two vertices, and
(ii) each cycle in C intersects at most three other cycles of C in more than one
vertex.
Proof. We give a reduction from Maximum Independent Set in 3-connected
cubic planar graphs, which is NP-complete [7]. Let H be a 3-connected cubic
planar graph. Observe that H has a unique combinatorial embedding up to a
flip. We construct an instance hG, C, ki of Facial C-Cycles as follows. We take
the planar dual H ? of H and take C as the set of facial cycles of H ? . Observe that
H ? is a planar triangulation, since H is cubic and 3-connected. The graph G is
obtained from H ? by adding for each edge e = uv ∈ E(H ? ) an edge vertex ve that
is adjacent to both u and v; see Fig. 1. We claim that H admits an independent
set of size k if and only if G admits a combinatorial embedding where k cycles
in C are facial.
Note that the embedding of H ? ⊆ G is unique up to a flip. The only embedding choices for G are to decide, for each edge e ∈ E(H ? ), in which of the two
faces incident to e in H ? the vertex ve is embedded. A cycle in C bounding a face
of H ? forms a facial cycle in the embedding of G if and only if no edge vertex
is embedded inside it. Note that no two cycles in C sharing an edge e ∈ E(H ? )
can both bound a face of G since the shared edge vertex ve must be embedded
in the interior of one of the two faces of H ? incident to e. It follows that an
embedding with k facial cycles in C induces a set of independent faces in H ? ,
and thus an independent set of size k in H. Conversely, by embedding the edge
vertices outside the faces of H ? corresponding to an independent set of size k
in H, we obtain an embedding of G with k facial cycles in C.
Observe that since H ? is a 3-connected, no two faces of H ? (resp. no two
cycles in C sharing an edge) share more than two vertices and, moreover, since
16
H ? is a planar triangulation, no cycle of C shares two vertices with more than
three other cycles. By construction C satisfies the restrictions in the statement
of the theorem.
t
u
Theorem 2. Facial C-Cycles is NP-complete for series-parallel graphs, even
if any two cycles in C share at most three vertices.
Proof. We give a reduction from Hamiltonian Circuit, which is known to be
NP-complete even for cubic graphs [10]. Let H be any such a graph.
Each vertex a ∈ V (H) is represented by the following gadget Ga . It consists of
the graph K2,3 , where the vertices in the partition of size 2 are denoted sa and v a
and the other vertices are denoted ua1 , ua2 , ua3 , and of an additional vertex ta
adjacent to v a ; see Fig. 2b. The graph G is obtained by merging the vertices sa
into a single vertex s and the vertices ta into a single vertex t.
To define C, we number the incident edges of each vertex of H from 1 to 3.
If ab is the i-th edge for a and the j-th edge for b, we define Cab ∈ C as the
cycle (s, uai , v a , t, v b , ubj , s); see Fig. 2a and 2c. We claim that G admits a combinatorial embedding with |V (H)| facial cycles in C if and only if H is Hamiltonian.
Assume that Q is a Hamiltonian circuit of H. We embed the graph G such
that the order of the gadgets Ga around s is the same as the order of the vertices
along Q. Now, for each edge ab ∈ Q the gadgets Ga and Gb are adjacent in this
order, say with Ga before Gb . Assume that ab is the i-th edge for a and the j-th
edge for b. We choose the order of the vertices uai in Ga and the vertices ubj in Gb
such that uai and ubj are incident to the face shared by Ga and Gb . Thus Cab
bounds a face. The resulting embedding clearly has |V (H)| facial cycles in C.
Conversely, assume that G has a combinatorial embedding with at least k =
|V (H)| facial cycles in C. Consider any two adjacent gadgets Ga and Gb . Since
there are |V (H)| facial cycles in C, it follows that the face between Ga and Gb
must be bounded by the cycle Cab in C. But this implies that a and b are adjacent
in H. Hence the circular order of the gadgets around s determines a Hamiltonian
circuit of H. Observe that any two cycles of C share at most three vertices. t
u
C
Omitted proofs from Section 4
Lemma 5. Let G be a series-parallel graph and C be a set of cycles in G
such that any two cycles share at most two vertices. For each P-node µ any two
relevant are either edge-disjoint in skel(µ) or they share the unique virtual edge
of skel(µ) that corresponds to a Q-node child of µ, if any.
Proof. Let C and C 0 be two relevant cycles for some P-node µ with poles u
and v. Clearly C and C 0 share the two poles u and v. Now assume that C
and C 0 additionally share a virtual edge e of skel(µ). Consider the expansion
graph Ge of e and observe that {u, v} cannot be a separation pair of Ge , since µ
is a P-node. Thus the corresponding child ν of µ must be either a Q- or an
S-node. If it is an S-node, however, then Ge contains a cutvertex c, which is
contained in both C and C 0 , a contradiction. Further observe that a P-node
may have at most one child that is a Q-node. This concludes the proof.
t
u
17
Theorem 5. Max Facial C-Cycles is solvable in O(n) for series-parallel
graphs if any two cycles in C share at most two vertices.
Proof. We use again a bottom-up approach as in the previous section. Q-nodes
can be handled trivially as before, and S-nodes can be handled by Lemma 1. It
remains to deal with the P-nodes. Here we use the special structure guaranteed
by Lemma 5.
Let µ be a P-node with k children and let I ∈ I(µ). First consider all cycles
in R(µ) whose projections do not traverse a Q-node child of µ. By Lemma 5 they
are pairwise disjoint, and we realize each such cycle C if it has positive gain, i.e.,
if and only if gain({C}, I) > 0. For the remaining cycles, which all share the same
virtual edge e that corresponds to a Q-node child ν of µ, we observe that at most
two of them can be realized, and their gains are again independent since they
share only e and they are disjoint from the other realized cycles. Altogether,
this allows to fill the table T [µ, I] in O(k) time for each I ∈ I(µ). Note that
|I(µ)| ≤ 1 unless the parent µ0 of µ is a Q-node, in which case the algorithm
has reached the root of the SPQR-tree, and we simply choose I greedily so that
it contains up to two cycles whose projections contain the parent edge and that
have positive gain.
t
u
D
Omitted Proofs from Section 5
Lemma 8. Let T̃ [µ, ·] denote the table computed in the above fashion. Then T̃ [µ, ·]
is a max{c, (4 + ε)}-approximation of T [µ, ·] provided that T̃ [µi , . . . ] is a capproximation of T [µi , ·].
Proof. We first show T̃ [µ, ·] ≤ T [µ, ·] by constructing for each I ∈ I(µ) an
embedding of pert(µ) that realizes I and that has T̃ [µ, I] realized cycles from C.
Fix I ∈ I(µ) and let CX denote the set of cycles in C determined as above. Then,
Pk
Pk
it is T̃ [µ, I] = i=1 T̃ [µi , (I ∪CX )∩I(µ)]+|X| ≤ i=1 T [µi , (I ∪CX )∩I(µ)]+|X|
since the T̃ [µi , ·] are c-approximations of T [µi , ·]. Thus, there exist embeddings of
the pert(µi ) with the corresponding number of facial cycles in C. By combining
them according to the embedding of skel(µ), we obtain a planar embedding of
pert(µ), which in addition to the facial cycles of the pert(µi ) in C has |X| facial
cycles in C that project to faces of skel(µ). This proves the claim.
Conversely, consider T [µ, I] and a corresponding embedding of pert(µ). Let
Copt denote the faces of C \ I that bound faces of pert(µ). By the 4-color theorem [15], it is possible to 4-color the faces in Copt such that two faces have the
same color only if they are disjoint. Let C 0 denote the largest color class and
observe that |C 0 | ≥ |Copt |/4.
Again, as in Lemma 7, we would like to compare the sizes of C 0 and X and
argue that X cannot be much smaller than C 0 since it is an approximation of
a maximum independent set in H and C 0 corresponds to an independent set of
faces in skel(µ). Again, the problem is that C 0 may contain cycles that project
to faces of skel(µ) for which no vertex is contained in H due to approximation
18
errors in the tables T̃ [µi , ·]. As before, we argue that the contribution of these
cycles is irrelevant in the approximation, and hence this consideration applies.
Let C10 = C 0 \ V (H) and let C20 = C 0 \ C10 . Recall that J = {1, . . . , k} and
let J1 = {i ∈ J | ∃C ∈ C10 that intersects µi } be the indices of children that are
intersected by a cycle in C10 . The set J2 = J \ J1 contains the remaining indices.
Pk
Clearly, we have T [µ, I] =
i=1 T [µi , (I ∪ Copt ) ∩ I(µi )] + |Copt | according to Lemma 4. Realizing instead of Copt just the set of cycles C 0 drops at
most |Copt |/4 facial cycles in C, while imposing weaker interface constraints on
the children. We therefore have
T [µ, I] =
k
X
i=1
T [µi , (I ∪ Copt ) ∩ I(µi )] + |Copt | ≤
k
X
i=1
T [µi , (I ∪ C 0 ) ∩ I(µi )] + 4|C 0 |
(2)
We now use the fact that the T̃ [µi , ·] are a c-approximation of the T [µi , ·], and
hence also a c0 -approximation for c0 = max{c, 4}, and we also separate the
sum by the index set J1 and J2 and consider the two sets of cycles C10 and C20
separately.
k
X
i=1
X
T [µi , (I ∪ C 0 ) ∩ I(µi )] + 4|C 0 | ≤ c0 ·
i∈J1
X
0
+c ·
i∈J2
T̃ [µi , (C10 ∪ I) ∩ I(µi )] + 4|C10 |
T̃ [µi , (C20 ∪ I) ∩ I(µi )] + 4|C20 |. (3)
Observe that the indices of the children intersected by cycles that correspond
to an vertex in H are all contained in J2 . By the definition of H, we thus
have T̃ [µi , (C20 ∪ I) ∩ I(µi )] = T̃ [µi , I ∩ I(µi )], for i ∈ J2 .
For the first term, observe that, for each cyclde C ∈ C10 , we have T̃ [µx , ({C10 }∪
I) ∩ I(µx )] ≤ T̃ [µx , I ∩ I(µx )] − 1 for at least one x ∈ Jf , where f is the facial
cycle of skel(µ) to which C projects. Otherwise the vertex would be in H, and
hence in C20 . Let J10 ⊆ J1 denote the set of indices where this happens and
let J100 = J1 \ J10 . Observe that |J10 | ≥ |C10 |. We thus have the following.
c0 ·
X
i∈J1
= c0 ·
≤ c0 ·
T̃ [µi , (C10 ∪ I) ∩ I(µi )] + 4|C10 |
X
i∈J10
X
i∈J10
(4)
T̃ [µi , (C10 ∪ I) ∩ I(µi )] + c0 ·
(T̃ [µi , I ∩ I(µi )] − 1) + c0 ·
X
i∈J100
X
i∈J100
T̃ [µi , (C10 ∪ I) ∩ I(µi )] + 4|C10 |
T̃ [µi , I ∩ I(µi )] + 4|C10 |
!
≤ c0 ·
X
i∈J1
T̃ [µi , I ∩ I(µi )] − |J10 |
+ 2|M10 | ≤ c0 ·
X
i∈J1
T̃ [µi , I ∩ I(µi )]
Where the last step uses the fact that c0 ≥ 4. Plugging this information into
Eq. 3, yields the following.
19
c0 ·
X
i∈J1
T̃ [µi , (C10 ∪ I) ∩ I(µi )] + 4|C10 | + c0 ·
X
i∈J2
T̃ [µi , (C20 ∪ I) ∩ I(µi )] + 4|C20 |
(5)
≤ c0 ·
k
X
i=1
0
T̃ [µi , I ∩ I(µi )] + 4|C20 | ≤ c0 ·
≤ max{c , 4 + ε} ·
k
X
i=1
k
X
i=1
T̃ [µi , I ∩ I(µi )] + 4(1 + ε/4)|X|
!
T̃ [µi , I ∩ I(µi )] + |X|
= max{c, 4 + ε}T̃ [µi , I]
Where the last third and second to last steps us the fact that X ⊆ V (H)
is a (1 + ε/4) approximation of a maximum independent set in H maximum
matching, and hence (1 + ε/4)|X| is at least as large as |C20 |, and that c0 ≥ 4,
respectively.
t
u
20
| 8cs.DS
|
arXiv:cs/0401016v3 [cs.LO] 14 Mar 2006
Generalized Strong Preservation
by Abstract Interpretation
F RANCESCO R ANZATO F RANCESCO TAPPARO
Dipartimento di Matematica Pura ed Applicata, Università di Padova
Via Belzoni 7, 35131 Padova, Italy
[email protected] [email protected]
Abstract
Standard abstract model checking relies on abstract Kripke structures which approximate concrete
models by gluing together indistinguishable states, namely by a partition of the concrete state space.
Strong preservation for a specification language L encodes the equivalence of concrete and abstract
model checking of formulas in L . We show how abstract interpretation can be used to design abstract
models that are more general than abstract Kripke structures. Accordingly, strong preservation is generalized to abstract interpretation-based models and precisely related to the concept of completeness in
abstract interpretation. The problem of minimally refining an abstract model in order to make it strongly
preserving for some language L can be formulated as a minimal domain refinement in abstract interpretation in order to get completeness w.r.t. the logical/temporal operators of L . It turns out that this refined
strongly preserving abstract model always exists and can be characterized as a greatest fixed point. As
a consequence, some well-known behavioural equivalences, like bisimulation, simulation and stuttering,
and their corresponding partition refinement algorithms can be elegantly characterized in abstract interpretation as completeness properties and refinements.
Keywords: Abstract interpretation, abstract model checking, strong preservation, completeness, refinement, behavioural equivalence.
1 Introduction
The design of an abstract model checking framework always includes a preservation result, roughly stating
that for any formula ϕ specified in some temporal language L , if ϕ holds on an abstract model then ϕ
also holds on the concrete model. On the other hand, strong preservation means that a formula of L holds
on an abstract model if and only if it holds on the concrete model. Strong preservation is highly desirable
since it allows to draw consequences from negative answers on the abstract side [10].
Generalized Strong Preservation. The relationship between abstract interpretation and abstract model
checking has been the subject of a number of works (see e.g. [9, 11, 16, 17, 19, 20, 28, 39, 40, 41, 43, 47]).
This paper follows the standard abstract interpretation approach [13, 14] where abstract domains are specified by Galois connections, namely pairs of abstraction and concretization maps α/γ. We deal with generic
(temporal) languages L of state formulae that are inductively generated by some given sets of atomic
propositions and operators. The interpretation p of atomic propositions p ∈ AP as subsets of States and
of operators f ∈ Op as mappings f on ℘(States) is determined by a suitable semantic structure S, e.g. a
Kripke structure, so that the concrete semantics JϕKS ∈ ℘(States) of a formula ϕ ∈ L is the set of states
making ϕ true w.r.t. S. Abstract semantics can be systematically defined by standard abstract interpretation. The powerset ℘(States) plays the role of concrete semantic domain so that abstract domains range
in AbsDom(℘(States)). Any abstract domain A ∈ AbsDom(℘(States)) induces an abstract semantic
structure S A where atoms p are abstracted to α(p) and operators f are interpreted as best correct approximations on A, that is α◦ f ◦ γ. Thus, A determines an abstract semantics JϕKA
S ∈ A that evaluates formulae
ϕ ∈ L in the abstract domain A.
It turns out that this approach generalizes standard abstract model checking [9, 10]. Given a Kripke
1
structure K = (States, →) (for simplicity we omit here a labeling function for atomic propositions), a
standard abstract model is specified as an abstract Kripke structure A = (AStates, →♯ ) where the set
AStates of abstract states is defined by a surjective map h : States → AStates. Thus, AStates determines a partition of States and vice versa. It turns out that state partitions are particular abstract domains.
In fact, the lattice of partitions of States is an abstract interpretation of the lattice of abstract domains
AbsDom(℘(States)) so that the abstract state space AStates corresponds to a particular abstract domain
ad(AStates) ∈ AbsDom(℘(States)). Abstract domains that can be derived from a state partition are
called partitioning. The interpretation of the language L w.r.t. the abstract Kripke structure A determines
an abstract semantic function JϕKA ∈ AStates. The abstract Kripke structure A strongly preserves L
when for any ϕ ∈ L and s ∈ States, it turns out that h(s) ∈ JϕKA ⇔ s ∈ JϕKK .
Strong preservation can then be generalized from standard abstract models to abstract interpretationbased models. Given a generalized abstract model A ∈ AbsDom(℘(States)), the induced abstract semanA
tics J·KA
S is strongly preserving for L when for any ϕ ∈ L and S ∈ ℘(States), α(S) ≤A JϕKS ⇔ S ⊆
♯
JϕKS . It turns out that this is an abstract domain property, because any abstract semantics J·K : L → A that
evaluates formulae in the abstract domain A is strongly preserving for L if and only if J·KA
S is. Standard
strong preservation becomes a particular instance, namely an abstract Kripke structure strongly preserves
L if and only if the corresponding partitioning abstract model strongly preserves L . On the other hand,
generalized strong preservation may work where standard strong preservation may fail, namely it may happen that although a strongly preserving abstract semantics on a partitioning abstract model ad(AStates)
exists this cannot be derived from a strongly preserving abstract Kripke structure on AStates.
Generalized Strong Preservation and Complete Abstract Interpretations. Given a language L and
a Kripke structure K = (States, →), a well-known key problem is to compute the smallest abstract
state space AStates L , when this exists, such that one can define an abstract Kripke structure AL =
(AStates L , →♯ ) that strongly preserves L . This problem admits solution for a number of well-known
temporal languages like CTL (or, equivalently, the µ-calculus), ACTL and CTL-X (i.e. CTL without the
next-time operator X). A number of algorithms for solving this problem exist, like those by Paige and
Tarjan [42] for CTL, by Henzinger et al. [35], Bustan and Grumberg [5] and Tan and Cleaveland [48]
for ACTL, and Groote and Vaandrager [32] for CTL-X. These are coarsest partition refinement algorithms: given a language L and a partition P of States, which is determined by a state labeling, these
algorithms can be viewed as computing the coarsest partition PL that refines P and strongly preserves
L . It is worth remarking that most of these algorithms have been designed for computing well-known
behavioural equivalences used in process algebra like bisimulation (for CTL), simulation (for ACTL) and
divergence-blind stuttering (for CTL-X) equivalence. Our abstract interpretation-based framework allows
to give a generalized view of the above partition refinement algorithms. We show that the most abstract domain ADL ∈ AbsDom(℘(States)) that strongly preserves a given language L always exists. It turns out
that ADL is a partitioning abstract domain if and only if L includes full propositional logic, that is when
L is closed under logical conjunction and negation. Otherwise, a proper loss of information occurs when
abstracting ADL to the corresponding partition PL . Moreover, for some languages L , it may happen that
one cannot define an abstract Kripke structure on the abstract state space PL that strongly preserves L
whereas the most abstract strongly preserving semantics in AbsDom(℘(States)) instead exists.
The concept of complete abstract interpretation is well known [14, 31]. This encodes an ideal situation where the abstract semantics coincides with the abstraction of the concrete semantics. We establish a precise correspondence between generalized strong preservation of abstract models and completeness in abstract interpretation. Our results are based on the notion of forward complete abstract domain.
An abstract domain A is forward complete for a concrete semantic function f when for any a ∈ A,
f (γ(a)) = γ(α(f (γ(a)))), namely when no loss of precision occurs by approximating in A a computation f (γ(a)). This notion of forward completeness is dual and orthogonal to the standard definition of
completeness in abstract interpretation. Giacobazzi et al. [31] showed how complete abstract domains can
be systematically and constructively derived from noncomplete abstract domains by minimal refinements.
This can be done for forward completeness as well. Given any domain A, the most abstract domain that
refines A and is forward complete for f does exist and can be characterized as a greatest fixpoint. Such a
domain is called the forward complete shell of A for f . It turns out that strong preservation is related to forward completeness as follows. As described above, the most abstract domain ADL that strongly preserves
2
L always exists. It turns out that ADL coincides with the forward complete shell for the operators of L
of a basic abstract domain determined by the state labeling. This characterization provides an elegant generalization of partition refinement algorithms used in standard abstract model checking. As a consequence
of these results, we derive a novel characterization of the corresponding behavioural equivalences in terms
of forward completeness of abstract domains. For example, it turns out that a partition P is a bisimulation
on some Kripke structure K if and only if the corresponding partitioning abstract domain ad(P ) is forward
complete for the standard predecessor transformer pre→ in K.
2 Basic Notions
2.1 Notation and Preliminaries
Let X be any set. Fun(X) denotes the set of functions f : X n → X, for some n ≥ 0, called arity of
f . Following a standard convention, when n = 0, f is meant to be a specific object of X. The arity
of f is also denoted by ♯(f ) ≥ 0. id denotes the identity map. If F ⊆ Fun(X) and Y ⊆ X then
def
F (Y ) = {f (~y) | f ∈ F, ~
y ∈ Y ♯(f ) }, namely F (Y ) is the set of images of Y for each function in F . If
f : X → Y then the image of f is also denoted by img(f ) = {f (x) ∈ Y | x ∈ X}. If f : X → Y
and g : Y → Z then g ◦ f : X → Z denotes the composition of f and g, i.e. g ◦ f = λx.g(f (x)). The
complement operator for the universe set X is ∁ : ℘(X) → ℘(X), where ∁(S) = X r S. When writing
a set S of subsets of a given set, like a partition, we often write S in a compact form like {1, 12, 13} or
{[1], [12], [13]} that stand for {{1}, {1, 2}, {1, 3}}. Ord denotes the proper class of ordinals and ω ∈ Ord
denotes the first infinite ordinal.
Let hP, ≤i be a poset. Posets are often denoted also by P≤ . We use the symbol ⊑ to denote pointwise
ordering between functions: If X is any set and f, g : X → P then f ⊑ g if for all x ∈ X, f (x) ≤ g(x).
A mapping f : P → Q on posets is continuous when f preserves least upper bounds (lub’s) of countable
chains in P , while, dually, it is co-continuous when f preserves greatest lower bounds (glb’s) of countable
chains in P . A complete lattice C≤ is also denoted by hC, ≤, ∨, ∧, ⊤, ⊥i where ∨, ∧, ⊤ and ⊥ denote,
respectively, lub, glb, greatest element and least element in C. A mapping f : C → D between complete
lattices is additive (co-additive) when for any Y ⊆ C, f (∨C Y ) = ∨D f (Y ) (f (∧C Y ) = ∧D f (Y )). We
denote by lfp(f ) and gfp(f ), respectively, the least and greatest fixpoint, when they exist, of an operator f
on a poset. The well-known Knaster-Tarski’s theorem states that any monotone operator f : C → C on a
complete lattice C admits a least fixpoint and the following characterization holds:
lfp(f ) = ∧{x ∈ C | f (x) ≤ x} = ∨α∈Ord f α,↑ (⊥)
where the upper iteration sequence {f α,↑ (x)}α∈Ord of f in x ∈ C is defined by transfinite induction on α
as usual:
– α = 0: f 0,↑ (x) = x;
– successor ordinal α = β + 1: f β+1,↑ (x) = f (f β,↑ (x));
– limit ordinal α: f α,↑ (x) = ∨β<α f β,↑ (x).
It is well known that if f is continuous then lfp(f ) = ∨α∈ω f α,↑ (⊥). Dually, f also admits a greatest
fixpoint and the following characterization holds:
gfp(f ) = ∨{x ∈ C | x ≤ f (x)} = ∧α∈Ord f α,↓ (⊤),
where the lower iteration sequence {f α,↓ (x)}α∈Ord of f in x ∈ C is defined as the upper iteration sequence
but for the case of limit ordinals: f α,↓ (x) = ∧β<α f β,↓ (x).
Let Σ be any set. PreOrd(Σ) denotes the set of preorder relations on Σ, that is R ⊆ Σ × Σ is a
preorder on Σ if R is reflexive and transitive. Part(Σ) denotes the set of partitions of Σ. Sets in a partition
P are called blocks of P . If ≡ ⊆ Σ × Σ is an equivalence relation then we denote by P≡ ∈ Part(Σ) the
corresponding partition of Σ. Vice versa, if P ∈ Part(Σ) then ≡P ⊆ Σ × Σ denotes the corresponding
3
equivalence relation on Σ. Part(Σ) is endowed with the following standard partial order 4: P1 4 P2 , i.e.
P2 is coarser than P1 (or P1 refines P2 ) iff ∀B ∈ P1 .∃B ′ ∈ P2 .B ⊆ B ′ . It is well known that hPart(Σ), 4i
is a complete lattice.
A transition system T = (Σ, ) consists of a (possibly infinite) set Σ of states and a transition relation
⊆ Σ × Σ. As usual [10], we assume that the relation is total, i.e., for any s ∈ Σ there exists some
t ∈ Σ such that st, so that any maximal path in T is necessarily infinite. T is finitely branching when for
any s ∈ Σ, {t ∈ Σ | st} is a finite set. The pre/post transformers on ℘(Σ) are defined as usual:
def
– pre = λY.{a ∈ Σ | ∃b ∈ Y. ab};
def
– pre
f = ∁ ◦ pre ◦∁ = λY.{a ∈ Σ | ∀b ∈ Σ.(ab ⇒ b ∈ Y )};
def
– post = λY.{b ∈ Σ | ∃a ∈ Y. ab};
g def
– post
= ∁ ◦ post ◦∁ = λY.{b ∈ Σ | ∀a ∈ Σ.(ab ⇒ a ∈ Y )}.
g are co-additive.
Let us observe that pre and post are additive operators on ℘(Σ)⊆ while pre
f and post
∃∃
∀∃
If R ⊆ Σ1 × Σ2 is any relation then the relations R , R ⊆ ℘(Σ1 ) × ℘(Σ2 ) are defined as follows:
– (S1 , S2 ) ∈ R∃∃ iff ∃s1 ∈ S1 .∃s2 ∈ S2 . (s1 , s2 ) ∈ R;
– (S1 , S2 ) ∈ R∀∃ iff ∀s1 ∈ S1 .∃s2 ∈ S2 . (s1 , s2 ) ∈ R.
2.2 Abstract Interpretation and Completeness
2.2.1 Abstract Domains
In standard Cousot and Cousot’s abstract interpretation, abstract domains can be equivalently specified
either by Galois connections, i.e. adjunctions, or by upper closure operators (uco’s) [13, 14]. Let us recall
these standard notions.
Galois Connections and Insertions. If A and C are posets and α : C → A and γ : A → C are
monotone functions such that ∀c ∈ C. c ≤C γ(α(c)) and α(γ(a)) ≤A a then the quadruple (α, C, A, γ) is
called a Galois connection (GC for short) between C and A. If in addition α ◦ γ = λx.x then (α, C, A, γ)
is a Galois insertion (GI for short) of A in C. In a GI, γ is 1-1 and α is onto. Let us also recall that the
notion of GC is equivalent to that of adjunction: if α : C → A and γ : A → C then (α, C, A, γ) is a GC
iff ∀c ∈ C.∀a ∈ A. α(c) ≤A a ⇔ c ≤C γ(a). The map α (γ) is called the left- (right-) adjoint to γ (α).
It turns out that one adjoint map α/γ uniquely determines the other adjoint map γ/α as follows. On the
one hand, a map α : C → A admits a necessarily unique right-adjoint map γ : A → C iff α preserves
def
arbitrary lub’s; in this case, we have that γ = λa. ∨C {c ∈ C | α(c) ≤A a}. On the other hand, a map
γ : A → C admits a necessarily unique left-adjoint map α : C → A iff γ preserves arbitrary glb’s; in
def
this case, α = λc. ∧A {a ∈ A | c ≤C γ(a)}. In particular, we have that in any GC (α, C, A, γ) between
complete lattices it turns out that α is additive and γ is co-additive. Also, if (α, C, A, γ) is a GI and C is a
complete lattice then A is a complete lattice as well and hA, ≤A i is order-isomorphic to himg(γ), ≤C i.
We assume the standard abstract interpretation framework, where concrete and abstract domains, C and
A, are complete lattices related by abstraction and concretization maps α and γ forming a GC (α, C, A, γ).
A is called an abstraction of C and C a concretization of A. The ordering relations on concrete and abstract
domains describe the relative precision of domain values: x ≤ y means that y is an approximation of x or,
equivalently, x is more precise than y. Galois connections allow to relate the concrete and abstract notions
of relative precision: an abstract value a ∈ A approximates a concrete value c ∈ C when α(c) ≤A a, or,
equivalently (by adjunction), c ≤C γ(a). As a key consequence of requiring a Galois connection, it turns
out that α(c) is the best possible approximation in A of c, that is α(c) = ∧{a ∈ A | c ≤C γ(a)} holds. If
(α, C, A, γ) is a GI then each value of the abstract domain A is useful in representing C, because all the
values in A represent distinct members of C, being γ 1-1. Any GC can be lifted to a GI by identifying in
an equivalence class those values of the abstract domain with the same concretization. Abs(C) denotes the
set of abstract domains of C and we write A ∈ Abs(C) to mean that the abstract domain A is related to
4
C through a GI (α, C, A, γ). An abstract domain A is disjunctive when the corresponding concretization
map γ is additive.
Closure Operators. An (upper) closure operator, or simply a closure, on a poset P≤ is an operator
µ : P → P that is monotone, idempotent and extensive, i.e., ∀x ∈ P. x ≤ µ(x). Dually, lower closure
operators are monotone, idempotent, and restrictive, i.e., ∀x ∈ P. µ(x) ≤ x. uco(P ) denotes the set of
closure operators on P . Let hC, ≤, ∨, ∧, ⊤, ⊥i be a complete lattice. A closure µ ∈ uco(C) is uniquely
determined by its image img(µ), which coincides with its set of fixpoints, as follows: µ = λy. ∧ {x ∈
img(µ) | y ≤ x}. Also, X ⊆ C is the image of some closure operator µX on C iff X is a Moore-family of
def
C, i.e., X = M(X) = {∧S | S ⊆ X} — where ∧∅ = ⊤ ∈ M(X). In other terms, X is a Moore-family
of C when X is meet-closed. In this case, µX = λy. ∧ {x ∈ X | y ≤ x} is the corresponding closure
operator on C. For any X ⊆ C, M(X) is called the Moore-closure of X in C, i.e., M(X) is the least
(w.r.t. set inclusion) subset of C which contains X and is a Moore-family of C. Moreover, it turns out
that for any µ ∈ uco(C) and any Moore-family X ⊆ C, µimg(µ) = µ and img(µX ) = X. Thus, closure
operators on C are in bijection with Moore-families of C. This allows us to consider a closure operator
µ ∈ uco(C) both as a function µ : C → C and as a Moore-family img(µ) ⊆ C. This is particularly
useful and does not give rise to ambiguity since one can distinguish the use of a closure µ as function or
set according to the context.
It turns out that hµ, ≤i is a complete meet subsemilattice of C, i.e. ∧ is its glb, but, in general, it is not
a complete sublattice of C, since the lub in µ — defined by λY ⊆ µ. µ(∨Y ) — might be different from
that in C. In fact, it turns out that µ is a complete sublattice of C (namely, img(µ) is also join-closed) iff µ
is additive.
If C is a complete lattice then uco(C) endowed with the pointwise ordering ⊑ is a complete lattice
denoted by huco(C), ⊑, ⊔, ⊓, λx.⊤, λx.xi, where for every µ, η ∈ uco(C), {µi }i∈I ⊆ uco(C) and x ∈ C:
– µ ⊑ η iff ∀y ∈ C. µ(y) ≤ η(y) iff img(η) ⊆ img(µ);
– (⊓i∈I µi )(x) = ∧i∈I µi (x);
– x ∈ ⊔i∈I µi ⇔ ∀i ∈ I. x ∈ µi ;
– λx.⊤ is the greatest element, whereas λx.x is the least element.
Thus, the glb in uco(C) is defined pointwise, while the lub of a set of closures {µi }i∈I ⊆ uco(C) is the
closure whose image is given by the set-intersection ∩i∈I µi .
The Lattice of Abstract Domains. It is well known since [14] that abstract domains can be equivalently
specified either as Galois insertions or as closures. These two approaches are completely equivalent. On the
one hand, if µ ∈ uco(C) and A is a complete lattice which is isomorphic to img(µ), where ι : img(µ) → A
and ι−1 : A → img(µ) provide the isomorphism, then (ι ◦ µ, C, A, ι−1 ) is a GI. On the other hand, if
def
(α, C, A, γ) is a GI then µA = γ ◦ α ∈ uco(C) is the closure associated with A such that himg(µA ), ≤C i
is a complete lattice which is isomorphic to hA, ≤A i. Furthermore, these two constructions are inverse of
each other. Let us also remark that an abstract domain A is disjunctive iff µA is additive. Given an abstract
domain A specified by a GI (α, C, A, γ), its associated closure γ ◦ α on C can be thought of as the “logical
meaning” of A in C, since this is shared by any other abstract representation for the objects of A. Thus, the
closure operator approach is particularly convenient when reasoning about properties of abstract domains
independently from the representation of their objects.
Abstract domains specified by GIs can be pre-ordered w.r.t. precision as follows: if A1 , A2 ∈ Abs(C)
then A1 is more precise (or concrete) than A2 (or A2 is an abstraction of A1 ), denoted by A1 A2 , when
µA1 ⊑ µA2 . The pointwise ordering ⊑ between uco’s corresponds therefore to the standard ordering used
to compare abstract domains with respect to their precision. Also, A1 and A2 are equivalent, denoted by
A1 ≃ A2 , when their associated closures coincide, i.e. µA1 = µA2 . Hence, the quotient Abs(C)/≃ gives
rise to a poset that, by a slight abuse of notation, is simply denoted by hAbs(C), ⊑i. Thus, when we write
A ∈ Abs(C) we mean that A is any representative of an equivalence class in Abs(C)/≃ and is specified by
5
a Galois insertition (α, C, A, γ). It turns out that hAbs(C), ⊑i is a complete lattice, called the lattice of abstract interpretations of C [13, 14], because it is isomorphic to the complete lattice huco(C), ⊑i. Lub’s and
glb’s in Abs(C) have therefore the following reading as operators on domains. Let {Ai }i∈I ⊆ Abs(C):
(i) ⊔i∈I Ai is the most concrete among the domains which are abstractions of all the Ai ’s; (ii) ⊓i∈I Ai is
the most abstract among the domains which are more concrete than every Ai — this latter domain is also
known as reduced product of all the Ai ’s.
2.2.2 Completeness
Let C be a concrete domain, f : C → C be a concrete semantic function1 and let f ♯ : A → A be
a corresponding abstract function on an abstract domain A ∈ Abs(C) specified by a GI (α, C, A, γ).
Then, hA, f ♯ i is a sound abstract interpretation when α ◦ f ⊑ f ♯ ◦ α holds. The abstract function f ♯ is
called a correct approximation on A of f . This means that a concrete computation f (c) can be correctly
approximated in A by f ♯ (α(c)), namely α(f (c)) ≤A f ♯ (α(c)). An abstract function f1♯ : A → A is more
precise than f2♯ : A → A when f1♯ ⊑ f2♯ . Since α ◦ f ⊑ f ♯ ◦ α holds iff α ◦ f ◦ γ ⊑ f ♯ holds, the abstract
def
function f A = α ◦ f ◦ γ : A → A is called the best correct approximation of f in A.
Completeness in abstract interpretation corresponds to requiring that, in addition to soundness, no loss
of precision occurs when f (c) is approximated in A by f ♯ (α(c)). Thus, completeness of f ♯ for f is encoded
by the equation α ◦ f = f ♯ ◦ α. This is also called backward completeness because a dual form of forward
completeness may be considered. As a very simple example, let us consider the abstract domain Sign
representing the sign of an integer variable, namely Sign = {⊥, Z<0 , 0, Z>0 , ⊤} ∈ Abs(℘(Z)⊆ ). Let us
def
consider the binary concrete operation of integer addition on sets of integers, that is X + Y = {x + y | x ∈
2
2 def
X, y ∈ Y }, and the square operator on sets of integers, that is X = {x | x ∈ X}. It turns out that
the best correct approximation +Sign of integer addition in Sign is sound but not complete — because
α({−1} + {1}) = 0 <Sign ⊤ = α({−1})+Sign α({1}) — while it is easy to check that the best correct
approximation of the square operation in Sign is instead complete.
A dual form of completeness may be considered. The soundness condition α ◦ f ⊑ f ♯ ◦ α can be
equivalently formulated as f ◦ γ ⊑ γ ◦ f ♯ . Forward completeness for f ♯ corresponds to requiring that
the equation f ◦ γ = γ ◦ f ♯ holds, and therefore means that no loss of precision occurs when a concrete
computation f (γ(a)), for some abstract value a ∈ A, is approximated in A by f ♯ (a). Let us notice
that backward and forward completeness are orthogonal concepts. In fact: (1) as observed above, we
have that +Sign is not backward complete while it is forward complete because for any a1 , a2 ∈ Sign,
γ(a1 ) + γ(a2 ) = γ(a1 +Sign a2 ); (2) the best correct approximation (·)2Sign of the square operator on Sign
is not forward complete because γ(Z>0 )2 ( γ(Z>0 ) = γ((Z>0 )2Sign ) while, as observed above, it is
instead backward complete.
Giacobazzi et al. [31] observed that completeness uniquely depends upon the abstraction map, i.e. upon
the abstract domain: this means that if f ♯ is backward complete for f then the best correct approximation
f A of f in A is backward complete as well, and, in this case, f ♯ indeed coincides with f A . Hence, for any
abstract domain A, one can define a backward complete abstract operation f ♯ on A if and only if f A is
backward complete. Thus, an abstract domain A ∈ Abs(C) is defined to be backward complete for f iff the
equation α ◦ f = f A ◦ α holds. This simple observation makes backward completeness an abstract domain
property, namely an intrinsic characteristic of the abstract domain. Let us observe that α ◦ f = f A ◦ α
holds iff γ ◦ α ◦ f = γ ◦ f A ◦ α = γ ◦ α ◦ f ◦ γ ◦ α holds, so that A is backward complete for f when
µA ◦ f = µA ◦ f ◦ µA . Thus, a closure µ ∈ uco(C), that defines some abstract domain, is backward
complete for f when µ ◦ f = µ ◦ f ◦ µ holds. Analogous observations apply to forward completeness,
which is also an abstract domain property: A ∈ Abs(C) is forward complete for f (or forward f -complete)
when f ◦ µA = µA ◦ f ◦ µA , while a closure µ ∈ uco(C) is forward complete for f when f ◦ µ = µ ◦ f ◦ µ
holds.
Let us also recall that, by a well-known result (see, e.g., [14, Theorem 7.1.0.4], [1, Fact 2.3] and [21,
Lemma 4.3]), backward complete abstract domains are “fixpoint complete” as well. This means that if
A ∈ Abs(C) is backward complete for a concrete monotone function f : C → C then α(lfp(f )) =
lfp(f A ). Moreover, if α and f are both co-continuous then this also holds for greatest fixpoints, namely
1 For
simplicity of notation we consider here unary functions since the extension to generic n-ary functions is straightforward.
6
α(gfp(f )) = gfp(f A ). As far as forward completeness is concerned, the following result holds.
Lemma 2.1. If A ∈ Abs(C) is forward complete for a monotone f then α(gfp(f )) = gfp(f A ). Moreover,
if γ and f are both continuous and γ(⊥A ) = ⊥C then α(lfp(f )) = lfp(f A ).
Proof. Let us show that α(gfp(f )) = gfp(f A ). On the one hand, since gfp(f ) ≤ γ(α(gfp(f ))), we
have that gfp(f ) = f (gfp(f )) ≤ f (γ(α(gfp(f )))), therefore, by using forward completeness, gfp(f ) ≤
γ(f A (α(gfp(f )))). Thus, α(gfp(f )) ≤ f A (α(gfp(f ))), from which follows that α(gfp(f )) ≤ gfp(f A ).
On the other hand, by using forward completeness, f (γ(gfp(f A ))) = γ(f A (gfp(f A ))) = γ(gfp(f A )),
so that γ(gfp(f A )) ≤ gfp(f ), and therefore, by applying α, we obtain that gfp(f A ) = α(γ(gfp(f A ))) ≤
α(gfp(f )).
Assume now that γ and f are both continuous and γ(⊥A ) = ⊥C . Let us show by induction on k that for
any k ∈ N, γ((f A )k,↑ (⊥A )) = f k,↑ (⊥C ).
(k = 0): By hypothesis, γ((f A )0,↑ (⊥A )) = γ(⊥A ) = ⊥C = f 0,↑ (⊥C ).
(k + 1):
γ((f A )k+1,↑ (⊥A )) =
γ(f A ((f A )k,↑ (⊥A ))) =
A k,↑
f (γ((f )
f (f
(⊥A ))) =
k,↑
[by forward completeness]
[by inductive hypothesis]
(⊥C )) =
f k+1,↑ (⊥C )).
Thus, by applying α, we obtain that for any k ∈ N,
(f A )k,↑ (⊥A ) = α(f k,↑ (⊥C )).
(†)
Since γ and f are continuous and α is always additive, we have that f A = α ◦ f ◦ γ is continuous because
it is a composition of continuous functions. Hence:
lfp(f A ) =
∨k∈N (f A )k,↑ (⊥A ) =
[by Knaster-Tarski’s theorem]
[by (†)]
∨k∈N α(f
k,↑
(⊥C )) =
[as α is additive]
α(∨k∈N f
k,↑
(⊥C )) =
[by Knaster-Tarski’s theorem]
α(lfp(f ))
and this concludes the proof.
It is worth noting that concretization maps of abstract domains which satisfies the ascending chain
conditions (i.e., every ascending chain is eventually stationary) are always trivially continuous.
2.2.3 Shells
Refinements of abstract domains have been studied from the beginning of abstract interpretation [13, 14]
and led to the notion of shell of an abstract domain [26, 29, 31]. Given a generic poset P≤ of semantic
objects — where x ≤ y intuitively means that x is a “refinement” of y — and a property P ⊆ P of these
objects, the generic notion of shell goes as follows: the P-shell of an object x ∈ P is defined to be an
object sx ∈ P such that:
(i) sx satisties the property P,
(ii) sx is a refinement of x, and
(iii) sx is the greatest among the objects satisfying (i) and (ii).
7
Note that if a P-shell exists then it is unique. Moreover, if the P-shell exists for any object in P then it
turns out that the operator mapping x ∈ P to its P-shell is a lower closure operator on P, being monotone,
idempotent and reductive: this operator will be called the P-shell refinement. We will be particularly
interested in shells of abstract domains and partitions, namely shells in the complete lattices of abstract
domains and partitions. Given a state space Σ and a partition property P ⊆ Part(Σ), the P-shell of
P ∈ Part(Σ) is the coarsest refinement of P satisfying P, when this exists. Also, given a concrete
domain C and a domain property P ⊆ Abs(C), the P-shell of A ∈ Abs(C), when this exists, is the most
abstract domain that satisfies P and refines A. Giacobazzi et al. [31] gave a constructive characterization of
backward complete abstract domains, under the assumption of dealing with continuous concrete functions.
As a consequence, they showed that backward complete shells always exist when the concrete functions
are continuous. In Section 6 we will follow this same idea for forward completeness and this will provide
the link between strongly preserving abstract models and complete abstract interpretations.
2.3 Abstract Model Checking and Strong Preservation
Standard temporal languages like CTL, CTL∗ , ACTL, the µ-calculus, LTL, etc., are interpreted on models specified as Kripke structures. Given a set AP of atomic propositions (of some language), a Kripke
structure K = (Σ, , ℓ) over AP consists of a transition system (Σ, ) together with a state labeling funcdef
tion ℓ : Σ → ℘(AP ). We use the following notation: for any s ∈ Σ, [s]ℓ = {s′ ∈ Σ | ℓ(s) = ℓ(s′ )},
def
while Pℓ = {[s]ℓ | s ∈ Σ} ∈ Part(Σ) denotes the state partition that is induced by ℓ. The notation s|=K ϕ
means that a state s ∈ Σ satisfies in K a state formula ϕ of some language L , where the specific definition
of the satisfaction relation |=K depends on the language L (interpretations of standard logical/temporal
operators can be found in [10]).
Standard abstract model checking [9, 10] relies on abstract Kripke structures that are defined over
partitions of the concrete state space Σ. A set A of abstract states is related to Σ by a surjective abstraction h : Σ → A that maps concrete states into abstract states and thus gives rise to a state partition
def
Ph = {h−1 (a) | a ∈ A} ∈ Part(Σ). Thus, in standard abstract model checking, formulae are interpreted
on an abstract Kripke structure A = (A, ♯ , ℓ♯ ) whose states are an abstract representation in A of some
block of the partition Ph . Given a specification language L of state formulae, a weak preservation result
for L guarantees that if a formula in L holds on an abstract Ktipke structure A then it also holds on the
corresponding concrete structure K: for any ϕ ∈ L , a ∈ A and s ∈ Σ such that h(s) = a, if a|=A ϕ
then s|=K ϕ. Moreover, strong preservation (s.p. for short) for L encodes the equivalence of abstract and
concrete validity for formulae in L : for any ϕ ∈ L , a ∈ A and s ∈ Σ such that h(s) = a, a|=A ϕ if and
only if s|=K ϕ.
The definition of weakly/strongly preserving abstract Kripke structures depends on the language L .
Let us recall some well-known examples [9, 10, 33]. Let K = (Σ, , ℓ) be a concrete Kripke structure
h : Σ → A be a surjection.
(i) Consider the language ACTL∗ . If Ph Pℓ then the abstract Kripke structure A = (A, ∃∃
h , ℓh )
weakly preserves ACTL∗ , where ℓh (a) = ∪{ℓ(s) | s ∈ Σ, h(s) = a} and ∃∃
⊆
A
×
A
is
defined
h
′
′
′
′
′
′
as: h(s1 ) ∃∃
h h(s2 ) ⇔ ∃s1 , s2 . h(s1 ) = h(s1 ) & h(s2 ) = h(s2 ) & s1 s2 .
(ii) Let Psim ∈ Part(Σ) be the partition induced by simulation equivalence on K. If Ph = Psim (this
also holds when Ph Psim ) then the abstract Kripke structure A = (A, ∀∃
h , ℓh ) strongly preserves
′
′
′
′
′
′
ACTL∗ , where h(s1 ) ∀∃
h(s
)
⇔
∀s
.
h(s
)
=
h(s
).
∃s
.
h(s
)
=
h(s
2
1
2 ) & s1 s2 .
1
1
2
2
h
(iii) Let Pbis ∈ Part(Σ) be the partition induced by bisimulation equivalence on K. If Ph = Pbis (this
also holds when Ph Pbis ) then the abstract Kripke structure A = (A, ∃∃
h , ℓh ) strongly preserves
∗
CTL .
Following Dams [19, Section 6.1] and Henzinger et al. [36, Section 2.2], the notion of strong preservation can be also given w.r.t. a mere state partition rather than w.r.t. an abstract Kripke structure. Let
J·KK : L → ℘(Σ) be the semantic function of state formulae in L w.r.t. a Kripke structure K = (Σ, , ℓ),
def
i.e., JϕKK = {s ∈ Σ | s|=K ϕ}. Then, the semantic interpretation of L on K induces the following logical
8
w
GFED
@ABC
R
stop
stop
ONML
/ HIJK
RY
GFED
/ @ABC
G
go
go
GFED
/ @ABC
Y
Figure 1: A U.K. traffic light.
equivalence ≡K
L ⊆ Σ × Σ:
′
′
s ≡K
L s iff ∀ϕ ∈ L . s ∈ JϕKK ⇔ s ∈ JϕKK .
Let PL ∈ Part(Σ) be the partition induced by ≡K
L (the index K denoting the Kripke structure is omitted). Then, a partition P ∈ Part(Σ) is strongly preserving2 for L (when interpreted on K) if P 4 PL .
Thus, PL is the coarsest partition that is strongly preserving for L . For a number of well known temporal
languages, like ACTL∗ , CTL∗ (see, respectively, the above points (ii) and (iii)), CTL∗ -X and the fragments of the µ-calculus described by Henzinger et al. [36], it turns out that if P is strongly preserving for
L then the abstract Kripke structure (P, ∃∃ , ℓL ) is strongly preserving for L , where, for any B ∈ P ,
ℓL (B) = ∪s∈B ℓ(s). In particular, (PL , ∃∃ , ℓL ) is strongly preserving for L and, additionally, PL is
the smallest possible abstract state space, namely if A = (A, ♯ , ℓ♯ ) is an abstract Kripke structure that
strongly preserves L then |PL | ≤ |A|.
However, given a language L and a Kripke structure K where formulae of L are interpreted, the
following example shows that it is not always possible to define an abstract Kripke structure A on the
partition PL such that A strongly preserves L .
Example 2.2. Consider the following simple language L :
L ∋ ϕ ::= stop | go | AXXϕ
and the Kripke structure K depicted in Figure 1, where superscripts determine the labeling function. K
models a four-state traffic light controller (like in the U.K. and in Germany): Red → RedYellow →
Green → Yellow. According to the standard semantics of AXX, we have that s|=K AXXϕ iff for any
path s0 s1 s2 . . . starting from s0 = s, it happens that s2 |=K ϕ. It turns out that JAXXstopKK = {G, Y }
and JAXXgoKK = {R, RY }. Thus, we have that PL = {{R, RY }, {G, Y }}. However, let us show
that there exists no abstract transition relation ♯ ⊆ PL × PL such that the abstract Kripke structure
A = (PL , ♯ , ℓL ) strongly preserves L . Assume by contradiction that such an abstract Kripke structure
A exists. Let B1 = {R, RY } ∈ PL and B2 = {G, Y } ∈ PL . Since R|=K AXXgo and G|=K AXXstop,
by strong preservation, it must be that B1 |=A AXXgo and B2 |=A AXXstop. Hence, necessarily, B1 ♯ B2
A
and B2 ♯ B1 . This leads to the contradiction B1 6|= AXXgo. In fact, if ♯ = {(B1 , B2 ), (B2 , B1 )} then
A
we would have that B1 6|= AXXgo. On the other hand, if, instead, B1 ♯ B1 (the case B2 ♯ B2 is analogous), then we would still have that B1 6|=A AXXgo. Even more, along the same lines it is not hard to show
that no proper abstract Kripke structure that strongly preserves L can be defined, because even if either
B1 or B2 is split we still cannot define an abstract transition relation that is strongly preserving for L .
3 Partitions as Abstract Domains
Let Σ be any (possibly infinite) set of states. Following [15, Section 5], a partition P ∈ Part(Σ) can be
viewed as an abstraction of ℘(Σ)⊆ as follows: any S ⊆ Σ is over approximated by the unique minimal
cover of S in P , namely by the union of all the blocks B ∈ P such that B ∩ S 6= ∅. A graphical example is
depicted on the left-hand side of Figure 2. This abstraction is formalized by a GI (αP , ℘(Σ)⊆ , ℘(P )⊆ , γP )
where:
def
def
αP (S) = {B ∈ P | B ∩ S 6= ∅}
γP (B) = ∪B∈B B.
Hence, any partition P ∈ Part(Σ) induces an abstract domain adp (P ) ∈ Abs(℘(Σ)), and an abstract
domain A ∈ Abs(℘(Σ)) is called partitioning when A is equivalent to adp (P ) for some partition P .
Observe that the closure adp (P ) = γP ◦ αP associated to a partitioning abstract domain is defined as
2 Dams
[19] uses the term “fine” instead of “strongly preserving”.
9
Figure 2: Partitions as abstract domains: over-approximation on the left and under-approximation on the
right.
adp (P ) = λS. ∪ {B ∈ P | B ∩ S 6= ∅}. Accordingly, a closure µ ∈ uco(℘(Σ)) that coincides with
γP ◦ αP , for some partition P , is called partitioning. We denote by Abspar (℘(Σ)) and ucopar (℘(Σ))
the sets of, respectively, partitioning abstract domains and closures on ℘(Σ). As noted in [16], a surjective abstraction h : Σ → A used in standard abstract model checking that maps concrete states into
abstract states (cf. Section 2.3) gives rise to a partitioning Galois insertion (αh , ℘(Σ)⊆ , ℘(A)⊆ , γh ) where
def
def
αh = λS ⊆ Σ.{h(s) ∈ A | s ∈ S} and γh = λX ⊆ A.{s ∈ Σ | h(s) ∈ X}.
Partitions can be also viewed as dual abstractions when a set S is under approximated by the union of
all the blocks B ∈ P such that B ⊆ S. A graphical example of this under approximation is depicted on the
right-hand side of Figure 2. This dual abstraction is formalized by the GI (e
αP , ℘(Σ)⊇ , ℘(P )⊇ , e
γP ) where
the ordering on the concrete domain ℘(Σ) is given by the subset relation and
def
α
eP (S) = {B ∈ P | B ⊆ S}
def
γ
eP (B) = ∪B∈B B.
In the following, we will be interested in viewing partitions as over approximations, that is partitions as
abstract domains of ℘(Σ)⊆ .
Thus, partitions can be viewed as representations of abstract domains. On the other hand, it turns out
that abstract domains can be abstracted to partitions. An abstract domain A ∈ Abs(℘(Σ)⊆ ) induces a state
equivalence ≡A on Σ by identifying those states that cannot be distinguished by A:
s ≡A s′ iff α({s}) = α({s′ }).
def
For any s ∈ Σ, [s]A = {s′ ∈ Σ | α({s}) = α({s′ })} is a block of the state partition par(A) induced by A:
def
par(A) = {[s]A | s ∈ Σ}.
Thus, par : Abs(℘(Σ)) → Part(Σ) is a mapping from abstract domains to partitions.
Example 3.1. Let Σ = {1, 2, 3, 4} and let us specify abstract domains as uco’s on ℘(Σ). The uco’s
µ1 = {∅, 12, 3, 4, 1234}, µ2 = {∅, 12, 3, 4, 34, 1234}, µ3 = {∅, 12, 3, 4, 34, 123, 124, 1234}, µ4 =
{12, 123, 124, 1234} and µ5 = {∅, 12, 123, 124, 1234} all induce the same partition P = par(µi ) =
{12, 3, 4} ∈ Part(Σ). For example, µ5 ({1}) = µ5 ({2}) = {1, 2}, µ5 ({3}) = {1, 2, 3} and µ5 ({4}) =
{1, 2, 3, 4} so that par(µ5 ) = P . Observe that µ3 is the only partitioning abstract domain because
adp (P ) = µ3 .
Abstract domains of ℘(Σ) carry additional information other than the underlying state partition and
this additional information allows us to distinguish them. It turns out that this can be precisely stated by
abstract interpretation since the above mappings par and adp allows us to show that the whole lattice of
partitions of Σ can be viewed as a (“higher-order”) abstraction of the lattice of abstract domains of ℘(Σ).
Theorem 3.2. (par, Abs(℘(Σ))⊒ , Part(Σ) , adp ) is a Galois insertion.
Proof. Let A ∈ Abs(℘(Σ)) and P ∈ Part(Σ) and let µA ∈ uco(℘(Σ)) be the closure associated with the
abstract domain A. Let us prove that P par(A) ⇔ adp (P ) ⊑ µA .
10
(⇒) For S ∈ ℘(Σ) we have to prove that adp (P )(S) ⊆ µA (S). Consider s ∈ adp (P )(S). Hence, there
exists some B ∈ P such that s ∈ B and B ∩ S 6= ∅. Let q ∈ B ∩ S. Since P par(A), there exists
some block [r]A ∈ par(A) such that B ⊆ [r]A . Thus, for any x, y ∈ B, α({x}) = α({r}) = α({y}), in
particular, α({s}) = α({q}). Consequently, since q ∈ S and therefore µA ({q}) ⊆ µA (S), we have that
µA ({s}) = µA ({q}) ⊆ µA (S), so that s ∈ µA (S).
(⇐) Consider a block B ∈ P and some s ∈ B. We show that B ⊆ [s]A , namely if s′ , s′′ ∈ B
then α({s′ }) = α({s′′ }). Since adp (P ) ⊑ µA , if s′ , s′′ ∈ B then adp (P )({s′ }) = B ⊆ µA ({s′ })
so that s′′ ∈ µA ({s′ }) and therefore µA ({s′′ }) ⊆ µA ({s′ }). Likewise, µA ({s′ }) ⊆ µA ({s′′ }) so that
µA ({s′ }) = µA ({s′′ }) and in turn α({s′ }) = α({s′′ }).
Finally, observe that adp is 1-1 so that the above adjunction is indeed a Galois insertion.
Let us observe that, as recalled in Section 2.2, the adjoint maps par and adp give rise to an order
isomorphism between the lattices hPart(Σ), i and hAbspar (℘(Σ)), ⊑i.
Corollary 3.3. Let A ∈ Abs(℘(Σ)). The following statements are equivalent:
(1) A is partitioning.
(2) γ is additive and {γ(α({s}))}s∈Σ is a partition of Σ. In this case, par(A) = {γ(α({s}))}s∈Σ .
(3) A is forward complete for the complement operator ∁.
Proof. Let A ∈ Abs(℘(Σ)) and let µA = γ ◦ α ∈ uco(℘(Σ)) be the corresponding uco.
(1) ⇒ (2) By Theorem 3.2, A ∈ Abspar (℘(Σ)) iff adp (par(A)) = A. Thus, if adp (par(A)) = A then
µA = γ ◦ α is obviously additive. Moreover, s ≡A s′ iff α({s}) = α({s′ }) iff γ(α({s})) = γ(α({s′ })),
so that, for any s ∈ Σ, [s]A = γ(α({s})) and therefore par(A) = {γ(α({s}))}s∈Σ .
(2) ⇒ (1) Since {γ(α({s}))}s∈Σ = P ∈ Part(Σ) we have that for any s ∈ Σ, [s]A = γ(α({s})): in fact,
if s′ ∈ γ(α({s})) then α({s′ }) ≤ α({s}), hence γ(α({s′ })) ⊆ γ(α({s})) and therefore γ(α({s′ })) =
γ(α({s})). Thus, par(A) = P . Moreover, since γ is additive, for any S ⊆ Σ, ∪s∈S γ(α({s})) =
γ(∨s∈S α({s})) = γ(α(S)) ∈ µA . Hence, since adp (P ) = {∪s∈S γ(α({s})) | S ⊆ Σ} we have that
adp (par(A)) = A.
(1) ⇒ (3) Assume that A ∈ Abspar (℘(Σ)). It is enough to prove that for any s ∈ Σ, ∁(µA ({s})) ∈ µA : in
fact, by (1) ⇒ (2), γ is additive and therefore µA is additive (because it is a composition of additive maps)
and therefore if S ∈ µA then S = ∪s∈S µA ({s}) so that ∁(S) = ∩s∈S ∁(µA ({s})). Let us observe the
following fact (∗): for any s, s′ ∈ Σ, s 6∈ µA ({s′ }) ⇔ µA ({s}) ∩ µA ({s′ }) = ∅; this is a consequence of
the fact that, by (1) ⇒ (2), {µA ({s})}s∈Σ is a partition. For any s ∈ Σ, we have that ∁(µA ({s})) ∈ µA
because:
µA (∁(µA ({s}))) = µA ({s′ ∈ Σ | s′ 6∈ µA ({s})}
= ∪{µA ({s′ }) | s′ 6∈ µA ({s})}
[by additivity of µA ]
[by the above fact (∗)]
= ∪{µA ({s′ }) | µA ({s′ }) ∩ µA ({s}) = ∅}
= ∪{µA ({s′ }) | µA ({s′ }) ⊆ ∁(µA ({s}))}
⊆ ∁(µA ({s}))
(3) ⇒ (1) Assume that µA is forward complete for ∁, i.e. µA is closed under complements. By (2) ⇒ (1),
it is enough to prove that γ is additive and that {µA ({s})}s∈Σ ∈ Part(Σ).
(i) γ is additive. Observe that γ is additive iff µA is additive iff µA is closed under arbitrary unions. If
{Si }i∈I ⊆ µA then ∪i Si = ∁(∩i ∁(Si )) ∈ µA , because, µA is closed under complements (and arbitrary
intersections).
(ii) {µA ({s})}s∈Σ ∈ Part(Σ). Clearly, we have that ∪s∈Σ µA ({s}) = Σ. Consider now s, r ∈ Σ such that
µA ({s}) ∩ µA ({r}) 6= ∅. Let us show that µA ({s}) = µA ({r}). In order to show this, let us prove that
s ∈ µA ({r}). Notice that µA ({s})rµA ({r}) = µA ({s})∩∁(µA ({r})) ∈ µA , because µA is closed under
complements. If s 6∈ µA ({r}) then we would have that s ∈ µA ({s}) r µA ({r}) ∈ µA , and this would
imply µA ({s}) ⊆ µA ({s}) r µA ({r}) ⊆ µA ({s}), namely µA ({s}) = µA ({s}) r µA ({r}). Thus, we
would obtain the contradiction µA ({s}) ∩ µA ({r}) = ∅. Hence, we have that s ∈ µA ({r}) and therefore
µA ({s}) ⊆ µA ({r}). By swapping the roles of s and r, we also obtain that µA ({r}) ⊆ µA ({s}), so that
µA ({s}) = µA ({r}).
11
def
Let us remark that P = adp ◦ par is a lower closure operator on hAbs(℘(Σ)), ⊑i and that for any
A ∈ Abs(℘(Σ)), A is partitioning iff P(A) = A. Hence, P is exactly the partitioning-shell refinement,
namely P(A) is the most abstract refinement of A that is partitioning.
4 Abstract Semantics of Languages
4.1 Concrete Semantics
We consider temporal specification languages L whose state formulae ϕ are inductively defined by:
L ∋ ϕ ::= p | f (ϕ1 , ..., ϕn )
where p ranges over a (typically finite) set of atomic propositions AP, while f ranges over a finite set Op
of operators. AP and Op are also denoted, respectively, by AP L and Op L . Each operator f ∈ Op has
an arity3 ♯(f ) > 0.
Formulae in L are interpreted on a semantic structure S = (Σ, I) where Σ is any (possibly infinite) set
of states and I is an interpretation function I : AP ∪ Op → Fun(℘(Σ)) that maps p ∈ AP to I(p) ∈ ℘(Σ)
and f ∈ Op to I(f ) : ℘(Σ)♯(f ) → ℘(Σ). I(p) and I(f ) are also denoted by, respectively, p and f .
def
def
Moreover, AP = {p ∈ ℘(Σ) | p ∈ AP } and Op = {f : ℘(Σ)♯(f ) → ℘(Σ) | f ∈ Op}. Note that the
def
interpretation I induces a state labeling ℓI : Σ → ℘(AP) by ℓI (s) = {p ∈ AP | s ∈ I(p)}. The concrete
state semantic function J·KS : L → ℘(Σ) evaluates a formula ϕ ∈ L to the set of states making ϕ true
w.r.t. the semantic structure S:
JpKS = p and Jf (ϕ1 , ..., ϕn )KS = f (Jϕ1 KS , ..., Jϕn KS ).
Semantic structures generalize the role of Kripke structures. In fact, in standard model checking a semantic
structure is usually defined through a Kripke structure K so that the interpretation of logical/temporal
operators is defined in terms of standard logical operators and paths in K. In the following, we freely
use standard logical and temporal operators together with their corresponding usual interpretations: for
example, I(∧) = ∩, I(∨) = ∪, I(¬) = ∁, I(EX) = preR , I(AX) = pre
f R , etc. As an example, consider
the standard semantics of CTL:
CTL ∋ ϕ ::= p | ϕ1 ∧ ϕ2 | ¬ϕ | AXϕ | EXϕ | AU(ϕ1 , ϕ2 ) | EU(ϕ1 , ϕ2 ) | AR(ϕ1 , ϕ2 ) | ER(ϕ1 , ϕ2 )
with respect to a Kripke structure K = (Σ, R, ℓ). Hence, K determines a corresponding interpretation I
for atoms in AP and operators of Op CTL , namely I(AX) = pre
f R , I(EX) = preR , etc., and this defines
the concrete semantic function J·KK : CTL → ℘(Σ).
If g is any operator with arity ♯(g) = n > 0 whose interpretation is given by g : ℘(Σ)n → ℘(Σ) and
S = (Σ, I) is a semantic structure then we say that a language L is closed under g for S when for any
ϕ1 , ..., ϕn ∈ L there exists some ψ ∈ L such that g(Jϕ1 KS , ..., Jϕn KS ) = JψKS . For instance, if Op L
includes EX and negation with their standard interpretations, i.e. I(EX) = preR and I(¬) = ∁, then L
is closed under AX with its standard interpretation pre
f R because pre
f R = ∁ ◦ preR ◦∁. This notion can be
extended in a straightforward way to infinitary operators: for instance, T
L is closed under infinite logical
conjunction for S iff for any Φ ⊆ L , there exists some ψ ∈ L such that ϕ∈Φ JϕKS = JψKS . In particular,
let us remark that if L is closed under infinite logical conjunction then it must exist some ψ ∈ L such
that ∩∅ = Σ = JψKS , namely L is able to express the tautology true. Let us remark that if the state space
Σ is finite and L is closed under logical conjunction then we always mean that there exists some ψ ∈ L
such that ∩∅ = Σ = JψKS . Finally, note that L is closed under negation and infinite logical conjunction
if and only if L includes propositional logic.
3 It would be possible to consider generic operators whose arity is any possibly infinite ordinal, thus allowing, for example, infinite
conjunctions or disjunctions.
12
p
?>=<
89:;
1
p
>=<
/ ?89:;
2
p,q
?>=<
89:;
3
>=<
/ ?89:;
4 j
r
⊤
@@
~~~ @
a5 A
a4 A
A
A
}}} A }}} A
a3
a2
a1 OO
OOO
oo
O oooo
* ?>=<
89:;
5
q
⊥
Figure 3: A Kripke structre on the left and an abstract domain on the right.
4.2 Abstract Semantics
In the following, we apply the standard abstract interpretation approach for defining abstract semantics
[13, 14]. Let L be a language and S = (Σ, I) be a semantic structure for L . An abstract semantic
structure S ♯ = (A, I ♯ ) is given by an abstract domain A ∈ Abs(℘(Σ)⊆ ) and by an abstract interpretation
function I ♯ : AP ∪ Op → Fun(A). An abstract semantic structure S ♯ therefore induces an abstract
semantic function J·KS ♯ : L → A that evaluates formulae in L to abstract values in A. The abstract
interpretation I ♯ is a correct over-approximation (respectively, under-approximation) of I on A when for
any p ∈ AP , γ(I ♯ (p)) ⊇ I(p) (respectively, γ(I ♯ (p)) ⊆ I(p)) and for any f ∈ Op, γ ◦ I ♯ (f ) ⊒
I(f ) ◦ γ (respectively, γ ◦ I ♯ (f ) ⊑ I(f ) ◦ γ). If I ♯ is a correct over-approximation (respectively, underapproximation) of I and the semantic operations in Op are monotone then the abstract semantics is an
over-approximation (respectively, under-approximation) of the concrete semantics, namely for any ϕ ∈ L ,
γ(JϕKS ♯ ) ⊇ JϕKS (respectively, γ(JϕKS ♯ ) ⊆ JϕKS ).
In particular, the abstract domain A always induces an abstract semantic structure S A = (A, I A ) where
I A is the best correct approximation of I on A, i.e. I A interprets atoms p and operators f as best correct
approximations on A of, respectively, p and f : for any p ∈ AP and f ∈ Op,
I A (p) = α(p) and I A (f ) = f A .
def
def
Thus, the abstract domain A systematically induces an abstract semantic function J·KS A : L → A, also
denoted by J·KA
S , which is therefore defined by:
A
A
A
A
JpKA
S = α(p) and Jf (ϕ1 , ..., ϕn )KS = f (Jϕ1 KS , ..., Jϕn KS ).
As usual in abstract interpretation, observe that the concrete semantics is a particular abstract semantics,
namely it is the abstract semantics induced by the “identical abstraction” (id, ℘(Σ), ℘(Σ), id).
Example 4.1. Let L ∋ ϕ ::= p | q | r | ϕ1 ∧ ϕ2 | EXϕ. Let us consider the Kripke structure K = (Σ, , ℓ)
and the lattice A both depicted in Figure 3. Let S be the semantic structure induced by the Kripke structure
K so that EX = pre . Let us consider the formulae EXr and EX(p ∧ q), whose concrete semantics are as
follows: JEXrKS = {3, 5} and JEX(p ∧ q)KS = {1, 2}. A is an abstract domain of ℘(Σ) where the Galois
insertion (α, ℘(Σ), A, γ) is determined by the following concretization map:
γ(⊥) = ∅; γ(a1 ) = {1, 2}; γ(a2 ) = {3}; γ(a3 ) = {3, 4};
γ(a4 ) = {1, 2, 3}; γ(a5 ) = {3, 4, 5}; γ(⊤) = {1, 2, 3, 4, 5}.
Note that, by Corollary 3.3, A is not partitioning because γ is not additive: γ(a2 ) ∪ γ(a3 ) = {3, 4} (
{3, 4, 5} = γ(a2 ∨ a3 ). It turns out that:
JEXrKA
S
= α(pre (γ(JrKA
S )) = α(pre (γ(α(r)))) = α(pre (γ(a3 )))
= α(pre ({3, 4})) = α({1, 2, 3, 5}) = ⊤;
JEX(p ∧ q)KA
S
A
= α(pre (γ(JpKA
S ∧ JqKS ))) = α(pre (γ(α(p) ∧ α(q))))
= α(pre (γ(a4 ∧ a5 ))) = α(pre (γ(a2 ))) = α(pre (3)) = α({1, 2}) = a1 .
Observe that the abstract semantics JEXrKA
S is a proper over-approximation of JEXrKS because JEXrKS (
).
On
the
other
hand,
the
concrete
semantics JEX(p ∧ q)KS is precisely represented in A because
γ(JEXrKA
S
)
=
JEX(p
∧
q)K
.
γ(JEX(p ∧ q)KA
S
S
13
5 Generalized Strong Preservation
We showed in Section 3 how a state partition P can be viewed as a partitioning abstract domain adp (P )
specified by the GI (αP , ℘(Σ)⊆ , ℘(P )⊆ , γP ). Thus, given a language L and a corresponding semantic
structure S = (Σ, I), it turns out that
any partition P ∈ Part(Σ) systematically induces a corresponddef
adp (P )
: L → adp (P ) that evaluates a formula in L to a (possibly
ing abstract semantics J·KP
S = J·KS
empty) union of blocks of P . Strong preservation for a partition P can be characterized in terms of the
corresponding abstract domain adp (P ) as follows.
Lemma 5.1. P ∈ Part(Σ) is s.p. for L iff ∀ϕ ∈ L and S ⊆ Σ, αP (S) ⊆ JϕKP
S ⇔ S ⊆ JϕKS .
Proof. (⇒): Let us first observe that for any ϕ ∈ L , γP (αP (JϕKS )) = JϕKS : in fact, for any s ∈ JϕKS ,
αP ({s}) is the block of P containing s; since P PL , we have that αP ({s}) ⊆ JϕKS , and from this
αP (JϕKS ) ⊆ JϕKS and in turn γP (αP (JϕKS )) = JϕKS .
Let us now prove by structural induction on ϕ ∈ L that JϕKS = γP (JϕKP
S ):
– ϕ ≡ p ∈ AP L : by using the above observation, JpKS = γP (αP (JpKS )) = γP (JpKP
S ).
– ϕ ≡ f (ϕ1 , . . . , ϕn ):
Jf (ϕ1 , . . . , ϕn )KS = [by the above observation]
γP (αP (Jf (ϕ1 , . . . , ϕn )KS )) = [by definition]
γP (αP (f (Jϕ1 KS , . . . , Jϕn KS ))) = [by inductive hypothesis]
P
γP (αP (f (γP (Jϕ1 KP
S ), . . . , γP (Jϕn KS )))) = [by definition]
γP (Jf (ϕ1 , . . . , ϕn )KP
S ).
P
Now, consider any ϕ ∈ L . If S ⊆ JϕKS then αP (S) ⊆ αP (JϕKS ) = αP (γP (JϕKP
S )) = JϕKS . Conversely,
P
if αP (S) ⊆ JϕKP
then
S
⊆
γ
(JϕK
)
=
JϕK
.
P
S
S
S
(⇐): Consider a block B ∈ P and s, s′ ∈ B so that αP ({s}) = B = αP ({s′ }). By hypothesis, for
′
P
′
any ϕ ∈ L , we have that s ∈ JϕKS iff αP ({s}) ⊆ JϕKP
S iff αP ({s }) ⊆ JϕKS iff s ∈ JϕKS . Thus,
′
s ≡L s .
This states that a partition P ∈ Part(Σ) is s.p. for L if and only if to check whether some set S of
states satisfies some formula ϕ ∈ L , i.e. S ⊆ JϕKS , is equivalent to check whether the abstract state
P
αP (S) is more precise than the abstract semantics JϕKP
S , that is S is over-approximated by JϕKS . The
key observation here is that in our abstract interpretation-based framework partitions are particular abstract
domains. This allows us to generalize the notion of strong preservation from partitions to generic abstract
semantic functions as follows.
Definition 5.2. Let L be a language, S = (Σ, I) be a semantic structure for L and S ♯ = (A, I ♯ ) be a
corresponding abstract semantic structure. The abstract semantics J·KS ♯ is strongly preserving for L (w.r.t.
S) if for any ϕ ∈ L and S ⊆ Σ,
α(S) ≤A JϕKS ♯ ⇔ S ⊆ JϕKS .
Definition 5.2 generalizes standard strong preservation from partitions, as characterized by Lemma 5.1,
both to an arbitrary abstract domain A ∈ Abs(℘(Σ)) and to a corresponding abstract interpretation function
I ♯ . Likewise, standard weak preservation can be generalized as follows. Let K = (Σ, R, ℓ) be a concrete
Kripke structure that induces the concrete semantics JϕKK = {s ∈ Σ | s|=K ϕ}. Let h : Σ → A be a
surjective abstraction and let (αh , ℘(Σ), ℘(A), γh ) be the corresponding partitioning abstract domain. Let
A = (A, R♯ , ℓ♯ ) be an abstract Kripke structure on A that gives rise to the abstract semantics JϕKA = {a ∈
A | a|=A ϕ}. Then, A weakly preserves L when
∀φ ∈ L .∀S ⊆ Σ. αh (S) ⊆ JϕKA ⇒ S ⊆ JϕKK .
Hence, weak preservation can be generalized to generic abstract domains and abstract semantics accordingly to Definition 5.2.
14
p
?>=<
89:;
1
p
>=<
/ ?89:;
2
p
>=<
/ ?89:;
3
p
ONML
HIJK
[12]
h
p
HIJK
/ ONML
[3]
r
Figure 4: A Kripke structure K on the left and an abstract Kripke structure A on the right.
5.1 Strong Preservation is an Abstract Domain Property
Definition 5.2 is a direct and natural generalization of the standard notion of strong preservation in abstract
model checking. It can be equivalently stated as follows.
Lemma 5.3. J·KS ♯ is s.p. for L iff for any ϕ ∈ L , JϕKS = γ(JϕKS ♯ ).
Proof. (⇒) On the one hand, γ(JϕKS ♯ ) ⊆ JϕKS iff α(γ(JϕKS ♯ )) ≤ JϕKS ♯ iff JϕKS ♯ ≤ JϕKS ♯ , which is
trivially true. On the other hand, JϕKS ⊆ γ(JϕKS ♯ ) iff α(JϕKS ) ≤ JϕKS ♯ iff JϕKS ⊆ JϕKS , that is trivially
true.
(⇐) We have that S ⊆ JϕKS iff S ⊆ γ(JϕKS ♯ ) iff α(S) ≤ JϕKS ♯ .
In particular, it is worth noting that if J·KS ♯ is s.p. for L then J·KS ♯ = α ◦ J·KS holds.
Lemma 5.4. Let A ∈ Abs(℘(Σ)).
(1) Let S1♯ = (A, I1♯ ) and S2♯ = (A, I2♯ ) be abstract semantic structures on A. If J·KS ♯ and J·KS ♯ are both
1
2
s.p. for L then J·KS ♯ = J·KS ♯ .
1
2
(2) Let S ♯ = (A, I ♯ ) be an abstract semantic structure on A. If J·KS ♯ is s.p. for L then J·KA
S is s.p. for L .
Proof. (1) By Lemma 5.3, for any ϕ ∈ L , γ(JϕKS ♯ ) = JϕKS = γ(JϕKS ♯ ), so that, by applying α,
1
2
JϕKS ♯ = α(γ(JϕKS ♯ )) = α(JϕKS ) = α(γ(JϕKS ♯ )) = JϕKS ♯ .
1
1
2
2
(2) Let us first observe that for any ϕ ∈ L , γ(α(JϕKS )) = JϕKS . In fact, γ(α(JϕKS )) ⊆ JϕKS ⇔
α(γ(α(JϕKS ))) ≤ JϕKS ♯ ⇔ α(JϕKS ) ≤ JϕKS ♯ ⇔ JϕKS ⊆ JϕKS . As a consequence of this fact, by
structural induction on ϕ ∈ L , analogously to the proof of Lemma 5.1, it is easy to prove that γ(JϕKA
S) =
JϕKS . Thus, by Lemma 5.3, J·KA
is
s.p.
for
L
.
S
Thus, it turns out that strong preservation is an abstract domain property. This means that given any
abstract domain A ∈ Abs(℘(Σ)), it is possible to define an abstract semantic structure S ♯ = (A, I ♯ ) on
A such that the corresponding abstract semantics J·KS ♯ is s.p. for L if and only if the induced abstract
semantics J·KA
S : L → A is s.p. for L . In particular, this also holds for the standard approach: if A =
(A, R♯ , ℓ♯ ) is an abstract Kripke structure for L , where h : Σ → A is the corresponding surjection, then
the standard abstract semantics J·KA strongly preserves L if and only if the abstract semantics induced by
the partitioning abstract domain (αh , ℘(Σ), ℘(A), γh ) strongly preserves L , and in this case this abstract
semantics coincides with the standard abstract semantics J·KA . Strong preservation is an abstract domain
property and therefore can be defined without loss of generality as follows.
Definition 5.5. An abstract domain A ∈ Abs(℘(Σ)) is strongly preserving for L (w.r.t. a semantic structure S) when J·KA
S is s.p. for L (w.r.t. S). We denote by SPL ⊆ Abs(℘(Σ)) the set of abstract domains
that are s.p. for L .
Example 5.6. Let us consider Example 4.1. It turns out that the abstract domain A is not s.p. for L
because, by Lemma 5.3, JEXrKS = {3, 5} ( {1, 2, 3, 4, 5} = γ(⊤) = γ(JEXrKA
S ).
Example 5.7. Let us consider the simple language L ∋ ϕ ::= p | EXϕ and the Kripke structure K
depicted in Figure 4. The Kripke structure K induces the semantic structure S = ({1, 2, 3}, I) such that
I(p) = {1, 2, 3} and I(EX) = pre . Hence, we have that JpKS = {1, 2, 3}, JEXpKS = {1, 2, 3} and, for
k > 1, JEXk pKS = {1, 2, 3}. Let us consider the partitioning abstract domain A induced by the partition
P = {[12], [3]} and related to ℘(Σ) by α and γ. Let us consider two different abstract semantic structures
on A.
15
– The abstract semantic structure S A = (A, I A ) is induced as best correct approximation of I by A.
– The abstract semantic structure S A = (A, I A ) is instead induced by the abstract Kripke structure
A = (A, ♯ , ℓ♯ ) in Figure 4. Hence, I A (p) = {[12], [3]} and I A (EX) = pre♯ .
S A is different from S A because I A (EX) 6= I A (EX). In fact, I A (EX)({[12]}) = α(pre (γ({[12]}))) =
α(pre ({1, 2})) = α({1}) = {[12]}, while I A (EX)({[12]}) = pre♯ ({[12]}) = ∅.
Let us show that both the abstract semantics J·KA
S and J·KS ♯ are s.p. for L .
A
– We have that JpKA
S = {[12], [3]}, JEXpKS = α(pre ({1, 2, 3})) = α({1, 2, 3}) = {[12], [3]} and,
k A
for k > 1, JEX pKS = {[12], [3]}. Thus, for any ϕ ∈ L, JϕKS = γ(JϕKA
S ).
– We have that JpKS A = {[12], [3]}, JEXpKS A = pre♯ ({[12], [3]}) = {[12], [3]} and, for k > 1,
JEXk pKS A = {[12], [3]}. Thus, for any ϕ ∈ L, JϕKS = γ(JϕKS A ).
Consequently, by Lemma 5.3, both abstract semantics are s.p. for L .
5.2 The Most Abstract Strongly Preserving Domain
As recalled in Section 2.3, a language L and a semantic structure S for L induce a corresponding logical
partition PL ∈ Part(Σ). By Lemma 5.1, it turns out that PL is the coarsest strongly preserving partitioning abstract domain for L . This can be generalized to arbitrary abstract domains as follows. Let us define
ADL by:
def
ADL = M({JϕKS | ϕ ∈ L }).
Hence, ADL is the closure under arbitrary intersections of the set of concrete semantics of formulae in L .
Observe that ADL ∈ Abs(℘(Σ)) because it is a Moore-family of ℘(Σ).
Theorem 5.8. For any A ∈ Abs(℘(Σ)), A ∈ SPL iff A ⊑ ADL .
Proof. Let µ = γ ◦ α ∈ uco(℘(Σ)) and let µL ∈ uco(℘(Σ)) be the uco associated to ADL , that is
µL (S) = ∩{JϕKS | ϕ ∈ L , S ⊆ JϕKS }. Recall that A ⊑ ADL iff for any ϕ ∈ L , JϕKS ∈ µ.
(⇒) For any ϕ ∈ L , we have that γ(α(JϕKS )) = JϕKS because, by Lemma 5.3, γ(α(JϕKS )) =
A
γ(α(γ(JϕKA
S ))) = γ(JϕKS ) = JϕKS .
(⇐) By hypothesis, γ(α(JϕKS )) = JϕKS for any ϕ. Let us show by structural induction on ϕ ∈ L that
JϕKS = γ(JϕKA
S ).
– ϕ ≡ p ∈ AP L : by using the hypothesis, JpKS = γP (αP (JpKS )) = γP (JpKA
S ).
– ϕ ≡ f (ϕ1 , . . . , ϕn ):
Jf (ϕ1 , . . . , ϕn )KS = [by hypothesis]
γ(α(Jf (ϕ1 , . . . , ϕn )KS )) = [by definition]
γ(α(f (Jϕ1 KS , . . . , Jϕn KS ))) = [by inductive hypothesis]
A
γ(α(f (γ(Jϕ1 KA
S ), . . . , γ(Jϕn KS )))) = [by definition]
γ(Jf (ϕ1 , . . . , ϕn )KA
S ).
Thus, by Lemma 5.3, A ∈ SPL .
Thus, ADL is the most abstract domain that is s.p. for L w.r.t. S. As a consequence, it turns out that
A is s.p. for L if and only if A represents with no loss of precision the concrete semantics of any formula
in L , that is ∀ϕ ∈ L . γ(α(JϕKS )) = JϕKS . Lemma 5.4 states that if a s.p. abstract semantics on a given
abstract domain exists then this is unique. Nevertheless, Example 5.7 shows that this unique s.p. abstract
semantics may be induced from different abstract semantic structures, i.e. different abstract interpretation
functions. However, when L is closed under conjunction, it turns out that on the most abstract s.p. domain
ADL , the abstract interpretation function is unique and is given by the best correct approximation I ADL .
16
Theorem 5.9. Let L be closed under infinite logical conjunction and let S ♯ = (ADL , I ♯ ) be an abstract
semantic structure on ADL . If J·KS ♯ is s.p. for L then I ♯ = I ADL .
Proof. Since L is closed under arbritrary logical conjunctions we have that ADL = {JϕKS | ϕ ∈ L }.
L
Thus, for any a ∈ ADL , there exists some ϕ ∈ L such that a = JϕKS ♯ = JϕKAD
. In fact, if a ∈ ADL
S
then a = JϕKS , for some ϕ ∈ L , so that, by Lemmata 5.3 and 5.4, a = JϕKS = γ(JϕKS ♯ ) = JϕKS ♯ =
L
JϕKAD
.
S
L
Let p ∈ AP. Then, by Lemma 5.4, JpKS ♯ = JpKAD
so that I ♯ (p) = I ADL (p).
S
Let f ∈ Op. Then,
I ♯ (f )(a1 , ..., an ) =
♯
I (f )(Jϕ1 KS ♯ , ..., Jϕn KS ♯ ) =
Jf (ϕ1 , ..., ϕn )KS ♯ =
L
Jf (ϕ1 , ..., ϕn )KAD
=
S
L
L
I ADL (f )(Jϕ1 KAD
, ..., Jϕn KAD
)=
S
S
ADL
I
(f )(a1 , ..., an ).
[by the observation above]
[by definition]
[by Lemma 5.4]
[by definition]
[by the observation above]
Thus, I ♯ = I ADL .
Hence, in the most abstract s.p. domain ADL there is a unique choice for interpreting atoms and operations
of L .
In our generalized framework, strong preservation for partitions becomes a particular instance through
the Galois insertion par/adp . Moreover, when L is closed under infinite conjunction, it turns out that the
most abstract s.p. domain ADL is partitioning if and only if L is also closed under negation.
Proposition 5.10.
(1) PL = par(ADL ) and adp (PL ) = P(ADL ).
(2) P is strongly preserving for L iff P 4 par(ADL ) iff adp (P ) ⊑ ADL .
(3) Let L be closed under conjunction. Then, ADL is partitioning iff L is closed under logical negation.
Proof. (1) Let µL ∈ uco(℘(Σ)) be the uco associated to ADL . We have that par(ADL ) = {[s]ADL | s ∈
Σ}, where [s]ADL = {s′ ∈ Σ | µL ({s′ }) = µL ({s})}. We also have that s ≡L s′ iff ∀ϕ ∈ L .s ∈
JϕKS ⇔ s′ ∈ JϕKS iff µL ({s}) = µL ({s′ }), so that PL = par(ADL ). Moreover, adp (PL ) =
adp (par(ADL )) = P(ADL ).
(2) P is s.p. for L iff P 4 PL iff, by Point (1), P 4 par(AL ) iff, by Theorem 3.2, adp (P ) ⊑ ADL .
(3) Since L is closed under infinite logical conjunction, ADL = {JϕKS | ϕ ∈ L }. Thus, L is closed
under logical negation iff ADL is closed under complementation ∁ and this exactly means that ADL is
forward complete for the complement ∁. By Corollary 3.3, this latter fact happens iff ADL is partitioning.
In particular, when L is closed under conjunction but not under negation, it turns out that adp (PL ) ⊏
ADL , i.e. a proper loss of information occurs when the domain ADL is abstracted to the partition
par(ADL ) = PL . On the other hand, when L is closed under conjunction and negation, we have that
adp (PL ) = ADL and therefore, by Theorem 5.9, the abstract interpretation function on the partitioning
abstract domain adp (PL ) is uniquely determined.
Example 5.11. Let us consider the traffic light controller K in Example 2.2. As already observed, formulae
of L have the following semantics in K:
JstopKK = {R, RY }; JgoKK = {G, Y }; JAXXstopKK = {G, Y }; JAXXgoKK = {R, RY }
so that
ADL = M({JϕKK | ϕ ∈ L }) = {∅, {R, RY }, {G, Y }, {R, RY, G, Y }}
and PL = par(ADL ) = {{R, RY }, {G, Y }}. We denote by µL the uco associated to ADL . As
shown in Example 2.2, it turns out that no abstract Kripke structure that properly abstracts K and strongly
17
p
?>=<
89:;
1
p
>=<
/ ?89:;
2
p
?>=<
89:;
3
>=<
/ ?89:;
4 j
p
GFED
@ABC
[12]
p
p
GFED
@ABC
[3]
* ?>=<
89:;
5
q
p
GFED
/ @ABC
[4] k
+ @ABC
GFED
[5]
q
Figure 5: Concrete (on the left) and abstract (on the right) Kripke structures.
preserves L can be defined. In our approach, the abstract domain ADL induces a corresponding strongly
L
preserving abstract semantics J·KAD
: L → ADL , where the best correct approximation of the operator
K
AXX : ℘(Σ) → ℘(Σ) on ADL is:
µL ◦ AXX = {∅ 7→ ∅, {R, RY } 7→ {G, Y }, {G, Y } 7→ {R, RY },
{R, RY, G, Y } 7→ {R, RY, G, Y }}.
Example 5.12. Consider the language CTL and the Kripke structure K = (Σ, R, ℓ) depicted in Figure 5,
where the interpretation of temporal operators of CTL on K is standard. It is well known that the coarsest
s.p. partition PCTL can be obtained by refining the initial partition P = {1234, 5} induced by the labeling
ℓ through the Paige-Tarjan [42] algorithm, since PCTL coincides with bisimulation equivalence on K. It
is easy to check that PCTL = {12, 3, 4, 5}. This partition determines (see point (2) in Section 2.3) the
s.p. abstract Kripke structure depicted in Figure 5. Since CTL is closed under conjunction and negation,
by Proposition 5.10 (1) and (3), it turns out that the most abstract s.p. domain ACTL is partitioning and
coincides with the following partitioning closure:
adp (PCTL ) = {∅, 12, 3, 4, 5, 34, 35, 45, 122, 124, 125, 345, 1234, 1235, 1245, 12345}.
Let us now consider the following language L ∋ ϕ ::= p | q | ϕ1 ∧ ϕ2 | EF[0,2] ϕ, where EF[0,2] is
a time bounded reachability operator that is useful for quantitative temporal analysis [24], e.g., of discrete
real-time systems [10, Chapter 16]. The standard interpretation of EF[0,2] is as follows: s|=K EF[0,2] ϕ iff
there exists a path s0 s1 s2 s3 . . . in K starting from s = s0 and some n ∈ [0, 2] such that sn |=K ϕ. Let us
characterize the semantics of formulae in L :
JpKK = {1, 2, 3, 4}; JqKK = {5}; JEF[0,2] pKK = {1, 2, 3, 4, 5};
JEF[0,2] qKK = {3, 4, 5}; JEF[0,2] (EF[0,2] q)KK = {1, 2, 3, 4, 5};
Jp ∧ EF[0,2] qKK = {3, 4}; JEF[0,2] (p ∧ EF[0,2] q)KK = {1, 2, 3, 4, 5}.
Thus, ADL = M({JϕKK | ϕ ∈ L }) = {∅, 5, 34, 345, 1234, 12345}. On the other hand, by Proposition 5.10 (1), PL = par(ADL ) = {12, 34, 5}. In this case, it turns out that adp (PL ) ⊏ ADL . Moreover,
analogously to Example 2.2, let us show that there exists no abstract transition relation ♯ ⊆ PL ×PL that
determines an abstract Kripke structure A = (PL , ♯ , ℓL ) which strongly preserves L . Let B = {1, 2},
B ′ = {3, 4} and B ′′ = {5} be the blocks in PL . Assume by contradiction that such an abstract Kripke
structure A exists.
(i) On the concrete model K we have that 3|=K EF[0,2] q. Thus, by strong preservation, it must be that
B ′ |=A EF[0,2] q. On the other hand, if B ′ ♯ B and B♯ B ′′ then B|=A EF[0,2] q and therefore, by
weak preservation, we would have that 1|=K EF[0,2] q, which is a contradiction. Thus, necessarily,
B ′ ♯ B ′′ .
(ii) Let us observe that 1|=K EF[0,2] EF[0,2] q. Hence, by strong preservation, B|=A EF[0,2] EF[0,2] q. If
B♯ B ′′ then, as in point (i), we would still have that 1|=K EF[0,2] q, i.e. a contradiction. Hence,
necessarily, B♯ B ′ .
18
(iii) From B♯ B ′ and B ′ ♯ B ′′ , we would obtain that B|=A EF[0,2] q that, as observed in point (ii), is a
contradiction.
Thus, this shows that it is not possible to define an abstract Kripke structure on the abstract state space PL
L
that strongly preserves L . The abstract domain ADL induces a corresponding abstract semantics J·KAD
K
that instead strongly preserves L . In this case, the best correct approximation of the operator EF[0,2] on
ADL is:
µL ◦ EF[0,2] = {∅ 7→ ∅, 5 7→ 345, 34 7→ 12345, 345 7→ 12345,
1234 7→ 12345, 12345 7→ 12345}.
6 Strong Preservation and Completeness
In this section we establish a precise correspondence between generalized strong preservation of abstract
models and completeness of abstract interpretations, so that the problem of minimally refining an abstract
model in order to get strong preservation can be formulated as a complete domain refinement in abstract
interpretation.
6.1 Forward Complete Shells
Let us consider forward completeness of abstract domains A ∈ Abs(C) for generic n-ary concrete operations f : C n → C, with n ≥ 0. Hence, A is forward complete for f , or simply f -complete,
when f ◦ hµA , ..., µA i = µA ◦ f ◦ hµA , ..., µA i, that is, for any ~x ∈ C n , f (µA (x1 ), ..., µA (xn )) =
µA (f (µA (x1 ), ..., µA (xn ))). Equivalently, A is f -complete when for any ~a ∈ An , f (γ(a1 ), ..., γ(an )) =
γ(α(f (γ(a1 ), ..., γ(an )))). For a set of operations F ⊆ Fun(C), A is F -complete when A is f -complete
for each f ∈ F . Observe that F -completeness for an abstract domain A means that the associated closure
µA is closed under the image of functions in F , namely F (µA ) ⊆ µA . Also note that when k : C 0 → C,
i.e. k ∈ C is a constant, A is k-complete iff k is precisely represented in A, i.e. γ(α(k)) = k. Let
us also note that an abstract domain A ∈ Abs(C) is always forward meet-complete because any uco is
Moore-closed.
Let us first note that forward F -complete shells always exist. Let SF : Abs(C) → Abs(C) be defined
def
as SF (A) = ⊔ {X ∈ Abs(C) | X ⊑ A, X is F -complete}.
Lemma 6.1. SF (A) is the F -complete shell of A.
Proof. Let η = ⊔{ρ ∈ uco(C) | ρ ⊑ µA , ρ is F -complete} = ∩{ρ ∈ uco(C) | ρ ⊑ µA , ρ is F -complete}.
Let f ∈ F , with ♯(f ) = n > 0 (if ♯(f ) = 0 then, trivially, f ∈ η) and ~c ∈ η n . Consider any ρ ∈ uco(C)
that is F -complete and such that ρ ⊑ µ. Since η ⊆ ρ, we have that ~c ∈ ρn and therefore f (~c) ∈ ρ because
ρ is F -complete. Thus, f (~c) ∈ η, i.e., η is F -complete.
A forward complete shell SF (A) is a more concrete abstraction than A. How to characterize SF (A)?
It is here useful to view abstract domains as closure operators on the concrete domain, i.e. as subsets
of C. Hence, A is viewed as the subset img(µA ) = γ(A) of the concrete domain C so that SF (A)
can be characterized as the least Moore-closed subset of C that contains img(µA ) and is forward F complete. We need to characterize the least amount of concrete information that must be added to γ(A) in
order to get forward completeness. It turns out that forward complete shells admit a constructive fixpoint
def
characterization. Let F uco : uco(C) → uco(C) be defined as follows: F uco (ρ) = M(F (ρ)), namely
uco
F (ρ) is the most abstract domain that contains the image of F on ρ. Observe that the operator λρ.µA ⊓
F uco (ρ) : uco(C) → uco(C) is monotone.
Lemma 6.2. SF (A) = gfp(λρ.µA ⊓ F uco (ρ)).
Proof. Observe that a uco ρ is F -complete iff F (ρ) ⊆ ρ iff M(F (ρ)) = F uco (ρ) ⊆ ρ iff ρ ⊑ F uco (ρ).
Thus, we have that SF (A) = ⊔{ρ ∈ uco(C) | ρ ⊑ µA , ρ is F -complete} = ⊔{ρ ∈ uco(C) | ρ ⊑ µA , ρ ⊑
F uco (ρ)} = ⊔{ρ ∈ uco(C) | ρ ⊑ µA ⊓ F uco (ρ)} = gfp(λρ.µA ⊓ F uco (ρ)).
19
Thus, it turns out that the lower iteration sequence of λρ.µA ⊓F uco (ρ) in uco(C) converges to the complete
shell SF (µA ).
Example 6.3. Let us consider the square operator on sets of integers sq : ℘(Z) → ℘(Z), i.e. sq(X) =
X 2 = {x2 | x ∈ X}, and the abstract domain Sign = {∅, Z<0 , {0}, Z>0 , Z}. As observed in Section 2.2.2, Sign is not forward complete for the square operator. Let us apply Lemma 6.2 in order to
compute the forward complete shell Ssq (Sign). Observe that
∅2 = ∅ ∈ Sign; {0}2 = {0} ∈ Sign; Z2<0 = Z2>0 = Z2 6∈ Sign.
Thus, the first step of iteration refines Sign to Sign ∪ {Z2 } (notice that this is an abstract domain because
2
it is Moore-closed). Then, (Z2 )2 = Z2 6∈ Sign ∪ {Z2 }, so that on the second step of iteration we obtain
2
k
Sign ∪ {Z2 , Z2 }. In general, for n ≥ 1, the n-th step of iteration provides Sign ∪ {Z2 | k ∈ [1, n]}, so
n
that the complete shell Ssq (Sign) coincides with the least fixpoint Sign ∪ {Z2 | n ≥ 1}.
Finally, the following easy observation will be useful later on.
Lemma 6.4. Let F, G ⊆ Fun(C). Then, SF = SG if and only if for any A ∈ Abs(C), A is F -complete
⇔ A is G-complete.
Proof. (⇒) If A is F -complete then A = SF (A) = SG (A) and therefore A is G-complete as well.
(⇐) This follows from SF (A) = ⊔{X ∈ Abs(C) | X ⊑ A, X is F -complete} = ⊔{X ∈ Abs(C) | X ⊑
A, X is G-complete} = SG (A).
6.2 Strong Preservation and Complete Shells
Let L be a language with atoms in AP L and operators in Op L and let S = (Σ, I) be a semantic structure
for L so that AP L and OpL denote, respectively, the corresponding sets of semantic interpretations of
atoms and operators. It turns out that forward completeness for AP L and OpL implies strong preservation for L .
Lemma 6.5. If A ∈ Abs(℘(Σ)) is forward complete for AP L and OpL then A is s.p. for L .
Proof. By Theorem 5.8, we show that A ⊑ ADL . Let us show by induction that for any ϕ ∈ L ,
JϕKS = γ(α(JϕKS )).
– ϕ ≡ p ∈ AP L : since A is forward complete for p, JpKS = p = γ(α(p)) = γ(α(JpKS )).
– ϕ ≡ f (ϕ1 , . . . , ϕn ) with f ∈ Op L :
Jf (ϕ1 , ..., ϕn )KS = [by definition]
f (Jϕ1 KS , ..., Jϕn KS ) = [by inductive hypothesis]
f (γ(α(Jϕ1 KS )), ..., γ(α(Jϕn KS ))) = [since A is forward complete for f ]
γ(α(f (γ(α(Jϕ1 KS )), ..., γ(α(Jϕn KS ))))) = [by inductive hypothesis and by definition]
γ(α(Jf (ϕ1 , ..., ϕn )KS )).
On the other hand, the converse is not true, that is strong preservation does not imply forward completeness, as shown by the following example.
Example 6.6. Let us consider again Example 5.7 where we showed that the partitioning abstract domain A = ℘(P )⊆ is s.p. for L . However, A is not forward complete for OpL = {pre }. In fact:
γ(α(pre (γ(α({3}))))) = γ(α(pre ({3}))) = γ(α({2, 3})) = {1, 2, 3} while pre (γ(α({3}))) =
pre ({3}) = {2, 3}.
Instead, it turns out that most abstract s.p. domains can be characterized as forward complete shells.
20
6.2.1 Complete Shells as Strongly Preserving Abstract Domains
Partition refinement algorithms for computing behavioural equivalences like bisimulation [42], simulation
equivalence [5, 35, 48] and (divergence blind) stuttering equivalence [32] are used in standard abstract
model checking to compute the coarsest strongly preserving partition of temporal languages like CTL∗ or
the µ-calculus for the case of bisimulation equivalence, ACTL∗ for simulation equivalence and CTL∗ -X
for stuttering equivalence.
Given a language L and a concrete state space Σ, these partition refinement algorithms work by iteratively refining an initial partition P within the lattice of partitions Part(Σ) until the fixpoint PL is
reached. The input partition P determines the set AP P of atoms and their interpretation IP as foldef
def
lows: AP P = {pB | B ∈ P } and IP (pB ) = B. More in general, any X ⊆ ℘(Σ) determines a set
{pX }X∈X of atoms with interpretation IX (pX ) = X. In particular, this can be done for an abstract domain A ∈ Abs(℘(Σ)) by considering its concretization γ(A) ⊆ Σ, namely A is viewed as a set of atoms
with interpretation IA (a) = γ(a). Thus, an abstract domain A ∈ Abs(℘(Σ)) together with a set of functions F ⊆ Fun(℘(Σ)) determine a language LA,F , with atoms in A, operations in F and endowed with
a semantic structure SA,F = (Σ, IA ∪ IF ) such that for any a ∈ A, IA (a) = γ(a) and for any f ∈ F ,
IF (f ) = f . Therefore, the most abstract s.p. domain ADLA,F generalizes in our framework the output
of a partition refinement algorithm for some language. Accordingly, we aim at characterizing ADLA,F
as the output of a refinement process of the initial domain A within the lattice Abs(℘(Σ)) of abstract domains. The following result shows that forward completeness for the operations in F is the right notion of
refinement to be used for the case of abstract domains.
Theorem 6.7. Let A ∈ Abs(℘(Σ)), F ⊆ Fun(℘(Σ)) and assume that LA,F is closed under infinite
logical conjunction. Then, ADLA,F = SF (A).
Proof. Since LA,F is closed under conjunction we have that ADLA,F = {JϕKSA,F | ϕ ∈ LA,F }. Let us
first prove that {JϕKSA,F | ϕ ∈ LA,F } ⊆ SF (A) by structural induction on ϕ ∈ LA,F :
– ϕ ≡ a ∈ A: JaKSA,F = IA (a) = γ(a) ∈ γ(A) ⊆ SF (A).
– ϕ ≡ f (ϕ1 , ..., ϕn ) with f ∈ F : Jf (ϕ1 , ..., ϕn )KSA,F = f (Jϕ1 KSA,F , ..., Jϕn KSA,F ), where, by
inductive hypothesis, Jϕi KSA,F ∈ SF (A). Therefore, since SF (A) is forward f -complete, we have
that f (Jϕ1 KSA,F , ..., Jϕn KSA,F ) ∈ SF (A).
Let us now prove the opposite inclusion. Let us first observe that ADLA,F is forward F -complete. For
simplicity of notation, consider f ∈ F with ♯(f ) = 1. If JϕKSA,F ∈ ADLA,F , where ϕ ∈ LA,F , then,
f (ϕ) ∈ LA,F and f (JϕKSA,F ) = Jf (ϕ)KSA,F ∈ ADLA,F . By Lemma 6.2, we know that SA (A) =
⊓α∈Ord (λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ), so that it is sufficient to prove by transfinite induction on
α ∈ Ord that
(λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F .
– α = 0: (λρ.µA ⊓ M(F (ρ)))0,↓ (⊤uco(℘(Σ)) ) = ⊤uco(℘(Σ)) = {Σ} ∈ γ(A) ⊆ ADLA,F .
– α + 1: By inductive hypothesis, (λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F . Moreover,
ADLA,F is Moore-closed and forward F -complete (hence closed under F ). Thus, M(F ((λρ.µA ⊓
M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ))) ⊆ ADLA,F , namely (λρ.µA ⊓M(F (ρ)))α+1,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F .
– limit ordinal α: This follows from
(λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ) = ⊓β<α (λρ.µA ⊓ M(F (ρ)))β,↓ (⊤uco(℘(Σ)) )
because, by inductive hypothesis, (λρ.µA ⊓ M(F (ρ)))β,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F , for any β < α.
21
6.2.2 Strongly Preserving Abstract Domains as Complete Shells
Let us consider a language L , with atoms in AP L and operators in Op L , and a semantic structure
S = (Σ, I). As an immediate consequence of Theorem 6.7, the most abstract s.p. domain ADL for L
w.r.t. S can be characterized as the forward AP L ∪ OpL -complete shell of the most abstract domain
{Σ}.
Corollary 6.8. Let L be closed under infinite logical conjunction. Then, ADL = SAP L ∪OpL ({Σ}).
Let us also observe that ADL can be equivalently characterized as the forward OpL -complete shell
of an initial abstract domain M(AP L ) induced by atoms: ADL = SOpL (M(AP L )).
6.2.3 Strongly Preserving Partitions
Theorem 6.7 and Corollary 6.8 provide an elegant generalization of partition refinement algorithms for
strong preservation from an abstract interpretation perspective.
Given a language L with operators in Op L and a corresponding semantic structure S = (Σ, I), as
recalled in Section 6.2.1, an input partition P ∈ Part(Σ) for a partition refinement algorithm determines
the set AP L = {pB | B ∈ P } of atoms of L and their interpretation I(pB ) = B. Thus, M(AP L ) =
M(P ) = P ∪ {∅, Σ}. It turns out that the coarsest s.p. partition PL for L can be characterized in our
abstract domain-based approach as follows.
Corollary 6.9. Let L be closed under infinite logical conjunction.
(1) PL = par(SOpL (M(P ))).
(2) Let L be closed under logical negation. Then, adp (PL ) = SOpL (M(P )).
Proof. (1) By Corollary 6.8, ADL = SOpL (M(P )) and by Proposition 5.10 (1), PL = par(ADL ) =
par(SOpL (M(P ))).
(2) By Proposition 5.10 (1) and (3), Corollary 6.8 and point (1), adp (PL ) = adp (par(ADL )) = ADL =
SOpL (M(P )).
It is worth remarking that when L is not closed under negation, by Proposition 5.10 (3) and Corollary 6.9 (2), it turns out that adp (PL ) ⊏ SOpL (M(P )). This means that when L is not closed under
negation the output partition PL of any partition refinement algorithm for achieving strong preservation
for L is not optimal within the lattice of abstract domains.
Example 6.10. Let us consider the language L and the concrete Kripke structure K in Example 5.12.
The labeling determines the initial partition P = {p = 1234, q = 5} ∈ Part(Σ), so that M(P ) =
{∅, 1234, 5, 12345} ∈ Abs(℘(Σ)). Here, Op L = {∧, EF[0,2] }. Abstract domains are Moore-closed so
that SOpL = SEF[0,2] . Let us compute SEF[0,2] (M(P )).
A0 = M(P ) = {∅, 1234, 5, 12345}
A1 = A0 ⊓ M(EF[0,2] (A0 )) = M(A0 ∪ EF[0,2] (A0 ))
= M({∅, 1234, 5, 12345} ∪ {EF[0,2] ({5}) = 345}) = {∅, 5, 34, 1234, 12345}
A2 = A1
(fixpoint)
As already observed in Example 5.12, PL = {12, 34, 5} is such that adp (PL ) ⊏ µL and it is not possible
to define a strongly preserving abstract Kripke structure on the abstract space PL .
7 An Application to some Behavioural Equivalences
It is well known that some temporal languages like CTL, ACTL and CTL-X induce state logical equivalences that coincide with standard behavioural equivalences like bisimulation equivalence for CTL, (divergence blind) stuttering equivalence for CTL-X and simulation equivalence for ACTL. We derive here
a novel characterization of these behavioural equivalences in terms of forward completeness of abstract
interpretations.
22
7.1 Bisimulation Equivalence
Let K = (Σ, , ℓ) be a Kripke structure over some set AP of atomic propositions. A relation R ⊆ Σ × Σ
is a bisimulation on K if for any s, s′ ∈ Σ such that sRs′ :
(1) ℓ(s) = ℓ(s′ );
(2) For any t ∈ Σ such that st, there exists t′ ∈ Σ such that s′ t′ and tRt′ ;
(3) s′ Rs, i.e. R is symmetric.
Since the empty relation is a bisimulation and bisimulations are closed under union, it turns out that the
largest (as a set) bisimulation relation exists. This largest bisimulation is an equivalence relation called
bisimulation equivalence and is denoted by ∼bis while Pbis ∈ Part(Σ) denotes the corresponding partition.
Thus, a partition P ∈ Part(Σ) is a bisimulation on K when P Pbis .
It is well known [4] that when K is finitely branching, bisimulation equivalence coincides with the
state equivalence induced by CTL, i.e., Pbis = PCTL (the same holds for CTL∗ and the µ-calculus, see
e.g. [19, Lemma 6.2.0.5]). Moreover, it is known (see e.g. [49, Section 12]) that it is enough to consider
finitary Hennessy-Milner logic [34], i.e. a language L1 including propositional logic and the existential
next operator in order to have that PL1 = Pbis :
L1 ∋ ϕ ::= p | ϕ1 ∧ ϕ2 | ¬ϕ | EXϕ
where, as usual, the interpretation EX of EX in K is pre . A number of algorithms for computing
bisimulation equivalence exists [3, 23, 38, 42]. The Paige-Tarjan algorithm [42] runs in O(|| log(|Σ|))time and is the most time-efficient algorithm that computes bisimulation equivalence.
We recalled above that PL1 = PCTL . In our framework, this can be obtained as a consequence of the
fact that the most abstract s.p. domains for CTL and L1 coincide.
Lemma 7.1. Let K be finitely branching. Then, ADCTL = ADL1 = adp (Pbis ).
Proof. Let OpCTL = {∩, ∁, AX, EX, AU, EU, AR, ER} be the set of standard interpretations of the
operators of CTL on K, so that AX = pre
f and EX = pre . We show that µ ∈ uco(℘(Σ)) is forward
complete for OpCTL iff µ is forward complete for {∁, pre }. Assume that µ is forward complete for
{∁, pre }. Let us first prove that µ is forward complete for pre
f = AX:
µ ◦ pre
f ◦µ =
µ ◦ ∁ ◦ pre ◦∁ ◦ µ =
[by definition of pre
f ]
[as µ is complete for ∁]
µ ◦ ∁ ◦ pre ◦µ ◦ ∁ ◦ µ =
µ ◦ ∁ ◦ µ ◦ pre ◦µ ◦ ∁ ◦ µ =
[as µ is complete for pre ]
[as µ is complete for ∁]
∁ ◦ µ ◦ pre ◦µ ◦ ∁ ◦ µ =
∁ ◦ pre ◦µ ◦ ∁ ◦ µ =
[as µ is complete for pre ]
[as µ is complete for ∁]
∁ ◦ pre ◦∁ ◦ µ =
pre
f ◦µ
[by definition of pre
f ]
The following fixpoint characterizations are well known [10]:
– AU(S1 , S2 ) = lfp(λZ.S2 ∪ (S1 ∩ pre
f (Z)));
– EU(S1 , S2 ) = lfp(λZ.S2 ∪ (S1 ∩ pre (Z)));
– AR(S1 , S2 ) = gfp(λZ.S2 ∩ (S1 ∪ pre
f (Z)));
– ER(S1 , S2 ) = gfp(λZ.S2 ∩ (S1 ∪ pre (Z))).
23
Let us show that µ is forward complete for AU. The proofs for the remaining operators in OpCTL are
analogous. We need to show that µ(lfp(λZ.µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (Z)))) = lfp(λZ.µ(S2 ) ∪ (µ(S1 ) ∩
pre
f (Z))). Let us show that µ is forward complete for the function λZ.µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (Z)):
µ(µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (µ(Z)))) =
µ(µ(S2 ) ∪ (µ(S1 ) ∩ µ(pre
f (µ(Z))))) =
µ(µ(S2 ) ∪ µ(µ(S1 ) ∩ µ(pre
f (µ(Z))))) =
µ(S2 ) ∪ µ(µ(S1 ) ∩ µ(pre
f (µ(Z)))) =
µ(S2 ) ∪ (µ(S1 ) ∩ µ(pre
f (µ(Z)))) =
µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (µ(Z))).
[as µ is complete for pre
f ]
[as µ is complete for ∩]
[as µ is complete for ∪]
[as µ is complete for ∩]
[as µ is complete for pre
f ]
Observe that since µ is additive (and therefore continuous) we have that µ(∅) = ∅. Moreover, let us
show that from the hypothesis that K is finitely branching it follows that pre
f is continuous. First, notice
that pre
f is continuous iff pre is co-continuous. Hence, let us check that pre is co-continuous. Let
{Xi }i∈N be a decreasing chain of subsets of Σ and let x ∈ ∩i∈N pre (Xi ). Since K is finitely branching,
post ({x}) is finite so that there exists some k ∈ N such that for any j > 0, post ({x}) ∩ Xk =
post ({x}) ∩ Xk+j . Hence, there exists some z ∈ ∩i∈N Xi ∩ post ({x}), so that x ∈ pre (∩i∈N Xi ).
Therefore, since pre
f is continuous we also have that λZ.µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (Z)) is continuous. We
can therefore apply Lemma 2.1 so that µ(lfp(λZ.µ(S2 )∪(µ(S1 )∩ pre
f (Z)))) = lfp(λZ.µ(S2 )∪(µ(S1 )∩
pre
f (Z))).
Thus, by Lemma 6.4, S{∁,pre } = SOpCTL , so that, by Corollary 6.8, ADL1 = ADCTL . Finally, since
K is finitely branching and L1 is closed under conjunction and negation, adp (PL1 ) = adp (Pbis ) =
adp (PL1 ) = ADL1 .
As a consequence of this and of the results in Section 6 (in particular of Corollary 6.9), any partition refinement algorithm Algbis for computing bisimulation equivalence on a finitely branching Kripke structure,
like those in [3, 23, 38, 42], can be characterized as a complete shell refinement as follows:
Algbis (P ) = par(S{∁,pre } (M(P ))).
Thus, Algbis is viewed as an algorithm for computing a particular abstraction, that is par, of a particular complete shell, that is S{∁,pre } . In particular, this holds for the Paige-Tarjan algorithm [42] and
leads to design a generalized Paige-Tarjan-like procedure for computing most abstract strongly preserving
domains [45].
Finally, our abstract intepretation-based approach allows us to give the following nice characterization of bisimulation for a partition P in terms of forward completeness for the corresponding partitioning
abstract domain adp (P ).
Theorem 7.2. Let P ∈ Part(Σ). Then, P is a bisimulation on K iff adp (P ) is forward complete for
{p | p ∈ AP } ∪ {pre }.
Proof. We view adp (P ) as a uco so that adp (P ) = {∪i Bi ∈ ℘(Σ) | {Bi } ⊆ P }. Let us first observe that P Pℓ iff adp (P ) is forward complete for {p ⊆ Σ | p ∈ AP }. On the one hand, since
p = {s ∈ Σ | p ∈ ℓ(s)}, if s ∈ p and s ∈ B, for some B ∈ P , then B ⊆ [s]ℓ ⊆ p. Hence, p is a union of
some blocks of P and therefore p ∈ adp (P ). On the other hand, if adp (P ) contains {p ⊆ Σ | p ∈ AP }
then, for any p ∈ AP , p is a union of some blocks in P . Thus, for any B ∈ P , either B ⊆ p or B ∩ p = ∅.
Consequently, if s ∈ B then B ⊆ [s]ℓ ∈ Pℓ .
Let us now note that adp (P ) is forward complete for pre iff for any block B ∈ P , pre (B) is a (possibly empty) union of blocks of P : this holds because pre is additive, and therefore if {Bi } ⊆ P
then pre (∪i Bi ) = ∪i pre (Bi ). The fact that, for some B ∈ P , pre (B) = ∪i Bi , for some blocks
{Bi } ⊆ P , implies that if s ∈ pre (B), i.e., st for some t ∈ B, then s ∈ Bj , for some j, and if s′ ∈ Bj
then s′ ∈ pre (B), i.e., s′ t′ for some t′ ∈ B, namely condition (2) of bisimulation for P holds. On
the other hand, if condition (2) of bisimulation for P holds then if s, s′ ∈ B ′ and s ∈ pre (B), for some
B, B ′ ∈ P , then s′ t′ for some t ∈ B, i.e., s′ ∈ pre (B), and therefore pre (B) is a union of blocks of
P . This closes the proof.
24
7.1.1 On the Smallest Abstract Transition Relation
As recalled in Section 2.3, the abstract Kripke structure A = (Pbis , ∃∃ , ℓ∃ ) strongly preserves CTL,
where B1 ∃∃ B2 iff there exist s1 ∈ B1 and s2 ∈ B2 such that s1 s2 , and ℓ∃ (B) = ∪s∈B ℓ(s). As a
simple and elegant consequence of our approach, it is easy to show that ∃∃ is the unique (and therefore
the smallest) abstract transition relation on Pbis that induces strong preservation for CTL.
Let K = (Σ, , ℓ) be finitely branching so that, by Lemma 7.1, ADL1 = adp (Pbis ) = ℘(Pbis ). Recall
that the concrete interpretation I induced by K is such that I(EX) = pre . By Theorem 5.9, the unique
interpretation of atoms and operations in L1 on the abstract domain ℘(Pbis ) that gives rise to a s.p. abstract
semantics is the best correct approximation I ℘(Pbis ) . Hence, if A = (Pbis , ♯ , ℓ♯ ) is strongly preserving for
CTL then the interpretation pre♯ of EX induced by A must coincide with I ℘(Pbis ) (EX). Consequently,
pre♯ = α ◦ pre ◦γ so that for any B1 , B2 ∈ Pbis , we have that B1 ♯ B2 iff B1 ∈ α(pre (γ({B2 }))).
Therefore, we conclude by observing that B1 ∈ α(pre (γ({B2 }))) iff B1 ∃∃ B2 .
We believe that a similar reasoning could be also useful for other languages L in order to prove that
the smallest abstract transition relation on PL that induces strong preservation exists. For example, this
has been proved for the case of ACTL by Bustan and Grumberg [5].
7.2 Stuttering Equivalence
Lamport’s criticism [37] of the next-time operator X in CTL/CTL∗ is well known. This motivated the
study of temporal logics CTL-X/CTL∗ -X obtained from CTL/CTL∗ by removing the next-time operator
and this led to study notions of behavioural stuttering-based equivalences [4, 22, 32]. We are interested
here in divergence blind stuttering (dbs for short) equivalence. Let K = (Σ, , ℓ) be a Kripke structure
over a set AP of atoms. A relation R ⊆ Σ × Σ is a divergence blind stuttering relation on K if for any
s, s′ ∈ Σ such that sRs′ :
(1) ℓ(s) = ℓ(s′ );
(2) If st then there exist t0 , ..., tk ∈ Σ, with k ≥ 0, such that: (i) t0 = s′ ; (ii) for all i ∈ [0, k − 1],
ti ti+i and sRti ; (iii) tRtk ;
(3) s′ Rs, i.e. R is symmetric.
Observe that condition (2) allows the case k = 0 and this simply boils down to requiring that tRs′ . Since
the empty relation is a dbs relation and dbs relations are closed under union, it turns out that the largest
dbs relation relation exists. It turns out that this largest dbs relation is an equivalence relation called
dbs equivalence and is denoted by ∼dbs while Pdbs ∈ Part(Σ) denotes the corresponding partition. In
particular, a partition P ∈ Part(Σ) is a dbs relation on K when when P Pdbs .
De Nicola and Vaandrager [22, Theorem 3.2.5] showed that for finite Kripke structures and for an
interpretation of universal/existential path quantifiers over all the, possibly finite, prefixes, dbs equivalence
coincides with the state equivalence induced from the language CTL-X (this also holds for CTL∗ -X),
that is Pdbs = PCTL-X . This is not true with the standard interpretation of path quantifiers over infinite
paths, since this requires a divergence sensitive notion of stuttering (see the details in [22]). Groote and
Vaandrager [32] presented a partition refinement algorithm that computes the partition Pdbs in O(|Σ|||)time.
We provide a characterization of divergence blind stuttering equivalence as the state equivalence induced by the following language L2 that includes propositional logic and the existential until operator EU,
where the interpretation of the existential path quantifier is standard, i.e. over infinite paths:
L2 ∋ ϕ ::= p | ϕ1 ∧ ϕ2 | ¬ϕ | EU(ϕ1 , ϕ2 )
Since the transition relation is assumed to be total, let us recall that the standard semantics EU :
℘(Σ)2 → ℘(Σ) of the existential until operator is as follows:
EU (S1 , S2 ) = S2 ∪ {s ∈ S1 | ∃s0 , ..., sn ∈ Σ, with n ≥ 0, such that (i) s0 = s,
(ii) ∀i ∈ [0, n − 1]. si ∈ S1 and si si+1 , (iii) sn ∈ S2 }.
25
The following result characterizes a dbs partition P in terms of forward completeness for the corresponding
partitioning abstract domain adp (P ).
Theorem 7.3. Let P ∈ Part(Σ). Then, P ∈ Part(Σ) is a dbs partition on K iff adp (P ) is forward
complete for {p | p ∈ AP } ∪ {EU }.
Proof. As already shown in the proof of Theorem 7.2, it turns out that P Pℓ iff adp (P ) is forward
complete for {p ⊆ Σ | p ∈ AP }. Thus, it remains to show P ∈ Part(Σ) satisfies condition (2) of the
definition of dbs relation iff adp (P ) is forward complete for EU . Let us first observe that P ∈ Part(Σ)
satisfies this condition (2) iff for any B1 , B2 ∈ P , EU (B1 , B2 ) = B1 ∪ B2 .
(⇒) If B1 = B2 then EU (B1 , B1 ) = B1 . Otherwise, assume that B1 6= B2 . If B2 ( EU (B1 , B2 ) ⊆
B1 ∪ B2 then there exists s ∈ EU (B1 , B2 ) such that s ∈ B1 . Thus, if s′ ∈ B1 then, by condition (2),
s′ ∈ EU (B1 , B2 ). This implies that EU (B1 , B2 ) = B1 ∪ B2 .
(⇐) Let B ∈ P , s, s′ ∈ B and st. If t ∈ B then condition (2) is satisfied. Otherwise, t ∈ B ′ , for some
B ′ ∈ P , with B 6= B ′ . Thus, s ∈ EU (B, B ′ ) and therefore EU (B, B ′ ) = B ∪ B ′ . This means that
condition (2) is satisfied for P .
To complete the proof it is now sufficient to show that if, for any B1 , B2 ∈ P , EU (B1 , B2 ) = B1 ∪ B2
then adp (P ) is forward complete for EU , i.e., for any {Bi }i∈I , {Bj }j∈J ⊆ P , EU (∪i Bi , ∪j Bj ) =
∪k Bk , for some {Bk }k∈K ⊆ P . The function EU is additive in its second argument, thus we only need
to show that, for any B ∈ P , EU (∪i Bi , B) = ∪k Bk , namely if s ∈ EU (∪i Bi , B) and s ∈ B ′ , for
some B ′ ∈ P , then B ′ ⊆ EU (∪i Bi , B). If s ∈ EU (∪i Bi , B) and s ∈ B ′ , for some B ′ ∈ {Bi }i ,
then there exist n ≥ 0 and s0 , ..., sn ∈ Σ such that s0 = s, ∀j ∈ [0, n − 1].sj ∈ ∪i Bi and sj sj+1 , and
sn ∈ B. Let us prove by induction on n ∈ N that if s′ ∈ B ′ then s′ ∈ EU (∪i Bi , B).
(n = 0): In this case s ∈ ∪i Bi and s ∈ B = B ′ . Hence, for some k, s ∈ Bk = B = B ′ and therefore
s ∈ EU (B, B). By hypothesis, EU (B, B) = B. Moreover, EU is monotone on its first component
and therefore B ′ = B = EU (B, B) ⊆ EU (∪i Bi , B).
(n+1): Suppose that there exist s0 , ..., sn+1 ∈ Σ such that s0 = s, ∀j ∈ [0, n].sj ∈ ∪i Bi and sj sj+1 , and
sn+1 ∈ B. Let sn ∈ Bk , for some Bk ∈ {Bi }i∈I . Then, s ∈ EU (∪i Bi , Bk ) and s = s0 s1 ...sn .
Since this finite path has length n, by inductive hypothesis, s′ ∈ EU (∪i Bi , Bk ). Hence, there exist
r0 , ..., rm ∈ Σ, with m ≥ 0, such that s′ = r0 , ∀j ∈ [0, m − 1].rj ∈ ∪i Bi and rj rj+1 , and rm ∈ Bk .
Moreover, since sn sn+1 , we have that sn ∈ EU (Bk , B). By hypothesis, EU (Bk , B) = Bk ∪ B,
and therefore rm ∈ EU (Bk , B). Thus, there exist q0 , ..., ql ∈ Σ, with l ≥ 0, such that rm = q0 ,
∀j ∈ [0, l − 1].qj ∈ Bk and qj qj+1 , and ql ∈ B. We have thus found the following finite path:
s′ = r0 r1 ...rm = q0 q1 ...ql , where all the states in the sequence but the last one ql belong to
∪i Bi , while ql ∈ B. This means that s′ ∈ EU (∪i Bi , B).
As a consequence, we obtain a characterization of dbs equivalence as the state equivalence induced by
the standard interpretation of the language L2 .
Corollary 7.4. Let Σ be finite. Then, Pdbs = PL2 .
Proof. By definition, Pdbs = gPart(Σ) {P ∈ Part(Σ) | P is a dbs relation on K}. By Theorem 7.3,
Pdbs = gPart(Σ) {P ∈ Part(Σ) | adp (P ) is complete for {p | p ∈ AP } ∪ {EU }}. By Theorem 3.2, adp is co-additive on Part(Σ) , that is adp preserves lub’s in Part(Σ) . Hence, adp (Pdbs ) =
⊔Abs(℘(Σ)) {adp (P ) ∈ Abs(℘(Σ)) | P ∈ Part(Σ), adp (P ) is complete for {p | p ∈ AP} ∪ {EU }}.
By Theorem 3.2, Abspar (℘(Σ)) = {adp (P ) | P ∈ Part(Σ)} so that adp (Pdbs ) = ⊔Abs(℘(Σ)) {A ∈
Abspar (℘(Σ)) | A is complete for {p | p ∈ AP} ∪ {EU }}. By Corollary 3.3, A ∈ Abspar (℘(Σ))
iff A is forward complete for ∁, so that adp (Pdbs ) = ⊔Abs(℘(Σ)) {A ∈ Abs(℘(Σ)) | A is complete
for {p | p ∈ AP} ∪ {∁, EU }}. Then, we note that A is forward complete for {p | p ∈ AP } iff
A ⊑ M({p | p ∈ AP }). Hence, adp (Pdbs ) = ⊔Abs(℘(Σ)) {A ∈ Abs(℘(Σ)) | A ⊑ M({p | p ∈ AP}), A
is complete for {∁, EU }} = S{∁,EU } (M({p | p ∈ AP})). Finally, since Σ is finite and therefore closure under infinite conjunction boils down to closure under finite conjunction, by Corollary 6.8,
S{∁,EU } (M({p | p ∈ AP })) = ADL2 . Thus, by Proposition 5.10 (1), adp (Pdbs ) = ADL2 , so that
Pdbs = par(adp (Pdbs )) = par(ADL2 ) = PL2 .
26
As a consequence of Corollary 6.9, the Groote-Vaandrager algorithm [32] GV for computing dsb equivalence on a finite Kripke structure can be characterized as a complete shell refinement as follows:
GV(P ) = par(S{∁,EU } (M(P ))).
7.3 Simulation Preorder and Equivalence
Simulations are possibly nonsymmetric bisimulations, that is R ⊆ Σ × Σ is a simulation on a Kripke
structure K = (Σ, , ℓ) if for any s, s′ ∈ Σ such that sRs′ :
(1) ℓ(s′ ) ⊆ ℓ(s);
(2) For any t ∈ Σ such that st, there exists t′ ∈ Σ such that s′ t′ and tRt′ .
The empty relation is a simulation and simulation relations are closed under union, so that the largest simulation relation exists. It turns out that the largest simulation is a preorder relation called similarity preorder
and denoted by Rsim ∈ PreOrd(Σ). Therefore, a preorder relation R ∈ PreOrd(Σ) is a simulation on K
when R ⊆ Rsim . Simulation equivalence ∼simeq ⊆ Σ × Σ is the symmetric closure of Rsim : s ∼simeq s′
iff there exist two simulation relations R1 and R2 such that sR1 s′ and s′ R2 s. Psimeq ∈ Part(Σ) denotes
the partition corresponding to ∼simeq .
A number of algorithms for computing simulation equivalence have been proposed [2, 5, 12, 27, 35]
and some of them like [2, 35] first compute the similarity preorder and then from it they obtain simulation
equivalence. The problem of computing simulation equivalence is important in model checking because,
as recalled in Section 2.3, simulation equivalence strongly preserves ACTL so that Psimeq = PACTL (see
[33, Section 4]). Recall that ACTL is obtained by restricting CTL, as defined in Section 4.1, to universal
quantifiers and by allowing negation on atomic propositions only:
ACTL ∋ ϕ ::= p | ¬p | ϕ1 ∧ ϕ2 | ϕ1 ∨ ϕ2 | AXϕ | AU(ϕ1 , ϕ2 ) | AR(ϕ1 , ϕ2 )
It turns out that the most abstract s.p. domain for ACTL can be obtained as the most abstract s.p.
domain for the following sublanguage L3 :
L3 ∋ ϕ ::= p | ¬p | ϕ1 ∧ ϕ2 | ϕ1 ∨ ϕ2 | AXϕ
Lemma 7.5. Let K be finitely branching. Then, ADACTL = ADL3 .
Proof. Let OpACTL = {∩, ∪, AX, AU, AR} be the set of standard interpretations of the operators of
ACTL on K, so that AX = pre
f . Analogously to the proof of Lemma 7.1, as a consequence of
the least/greatest fixpoint characterizations of AU and AR, it turns out that for any A ∈ Abs(℘(Σ)),
A is forward complete for OpACTL iff A is forward complete for {∪, pre
f }. Thus, by Lemma 6.4,
=
AD
.
,
so
that,
by
Corollary
6.8,
AD
S{∪,pre
=
S
ACTL
L
Op
f }
3
ACTL
Thus, by Proposition 5.10 (1), PACTL = par(ADACTL ) = par(ADL3 ) = PL3 , so that Psimeq = PL3 .
As a further consequence, by Corollary 6.9, any algorithm Algsimeq that computes simulation equivalence
can be viewed as a partitioning abstraction of the {∪, pre
f }-complete shell refinement:
Algsimeq (P ) = par(S{∪,pre } (M(P ))).
An instantiation of the generalized Paige-Tarjan-like procedure in [45] for the complete shell S{∪,pre }
allows to design a new efficient abstract intepretation-based algorithm for computing simulation equivalence [46] whose space and time complexity is comparable with that of state-of-the-art algorithms like
[5, 27].
27
7.3.1 Preorders as Abstract Domains
Simulations give rise to preorders rather than equivalences like in the case of bisimulations and dbs relations. Thus, in order to characterize simulation for preorders as forward completeness for abstract domains
we need to view preorders as abstract domains. This can be obtained by generalizing the abstraction in
Section 3 from partitions to preorders.
def
Let R ∈ PreOrd(Σ) and for any x ∈ Σ let us define Rpre = {preR ({x}) ⊆ Σ | x ∈ Σ}. The preorder
R gives rise to an abstract domain ℘(Rpre )⊆ which is related to ℘(Σ)⊆ through the following abstraction
and concretization maps:
def
αR (S) = {preR ({x}) ⊆ Σ | x ∈ S}
def
γR (X ) = ∪X∈X X.
It is easy to check that from the hypothesis that R is a preorder it follows that (αR , ℘(Σ)⊆ , ℘(Rpre )⊆ , γR )
is indeed a GI. Hence, any R ∈ PreOrd(Σ) induces an abstract domain denoted by add (R) ∈ Abs(℘(Σ)).
Also, note that γR ◦ αR = preR , namely preR is the closure associated to add (R). The notation add comes
from the fact that an abstract domain A is equivalent to some add (R) if and only if A is disjunctive.
Lemma 7.6. {add (R) ∈ Abs(℘(Σ)) | R ∈ PreOrd(Σ)} = {A ∈ Abs(℘(Σ)) | A is disjunctive}.
Proof. Observe that γR is trivially additive, so that any add (R) is disjunctive. On the other hand, let
A ∈ Abs(℘(Σ)) be disjunctive and consider the relation RA = {(x, y) | α({x}) ≤A α({y})} which is
trivially a preorder. Thus, add (RA ) is disjunctive so that in order to conclude that add (RA ) is equivalent to
A it is enough to observe that for any y ∈ Σ, preRA ({y}) = γ(α({y})): this is true because γ(α({y})) =
{x ∈ Σ | α({x}) ≤A α({y})} = preRA ({y}).
Let us observe that add indeed generalizes adp from partitions to preorders because for any P ∈
Part(Σ), adp (P ) = add (R): this is a simple consequence of the fact that for a partition P viewed as an
equivalence relation and for x ∈ Σ, Px is exactly a block of P so that αP (S) = {preP ({x}) | x ∈ S}. On
the other hand, an abstract domain A ∈ Abs(℘(Σ)) induces a preorder relation preord(A) ∈ PreOrd(Σ)
as follows:
(x, y) ∈ preord(A) iff α({x}) ≤A α({y}).
It turns out that the maps add and preord allows to view the lattice of preorder relations as an abstraction
of the lattice of abstract domains.
Theorem 7.7. (preord, Abs(℘(Σ))⊒ , PreOrd(Σ)⊇ , add ).
Proof. Let A ∈ Abs(℘(Σ)) and R ∈ PreOrd(Σ). Let us prove that R ⊆ preord(A) ⇔ add (R) ⊑ γ ◦ α.
(⇒) Let S ⊆ Σ and let us show that add (R)(S) = preR (S) ⊆ γ(α(S)). If x ∈ preR (S) then xRy for
some y ∈ S, so that (x, y) ∈ preord(A), i.e. α({x}) ≤A α({y}). Thus, by applying γ, x ∈ γ(α({x})) ⊆
γ(α({y})) ⊆ γ(α(S)).
(⇐) Let (x, y) ∈ R and let us show that α({x}) ≤ α({y}). Note that x ∈ preR ({y}) = add (R)({y}) ⊆
γ(α({y})), so that α({x}) ≤A α({y}), namely (x, y) ∈ preord(A).
def
Let us remark that D = add ◦ preord is a lower closure operator on hAbs(℘(Σ)), ⊑i and that, by
Lemma 7.6, for any A ∈ Abs(℘(Σ)), A is disjunctive iff D(A) = A. Hence, D coincides with the
disjunctive-shell refinement, also known as disjunctive completion [14], namely D(A) is the most abstract
disjunctive refinement of A.
We can now provide a characterization of simulation preorders in terms of forward completeness.
Theorem 7.8. Let R ∈ PreOrd(Σ). Then, R is a simulation on K iff add (R) is forward complete for
{p | p ∈ AP } ∪ {pre
f }.
Proof. Recall that preR is the closure associated to add (R). We first observe that (sRs′ ⇒ ℓ(s′ ) ⊆ ℓ(s))
iff preR is forward complete for AP. On the one hand, if p ∈ AP and s ∈ preR (p) then sRs′ for some
s′ ∈ p, so that, from ℓ(s′ ) ⊆ ℓ(s), we obtain s ∈ p, and therefore preR (p) = p. On the other hand, if
sRs′ and s′ ∈ p, for some p ∈ AP , then s′ ∈ p = preR (p) so that preR ({s′ }) ⊆ preR (preR (p)) =
preR (p) = p and therefore from s ∈ preR ({s′ }) we obtain s ∈ p.
28
Thus, it remains to show that R satisfies condition (2) of the definition of simulation iff preR is forward
complete for pre .
(⇒) We prove that for any S, preR (pre
f (preR (S))) ⊆ pre
f (preR (S)). Let x ∈ preR (pre
f (preR (S)))
so that there exists some y ∈ pre
f (preR (S)) such that xRy. If xx′ , for some x′ , then, by simulation,
there exists some y ′ such that yy ′ and x′ Ry ′ . Hence, y ′ ∈ preR (S) and this together with x′ Ry ′ , as R is
transitive, gives x′ ∈ preR (S). Therefore, x ∈ pre
f (preR (S)).
(⇐) Observe that in order to show that R is a simulation it is enough to show that if xRy then x ∈
pre
f (preR (post ({y}))). The following implications hold, where post ({y}) ⊆ preR (post ({y}))
holds because preR is a uco:
post ({y}) ⊆ preR (post ({y})) ⇒
pre
f (post ({y})) ⊆ pre
f (preR (post ({y}))) ⇒
{y} ⊆ pre
f (preR (post ({y}))) ⇒
preR ({y}) ⊆ preR (pre
f (preR (post ({y})))) ⇒
preR ({y}) ⊆ pre
f (preR (post ({y}))) ⇒
x ∈ pre
f (preR (post ({y})))
[as pre
f is monotone]
[as y ∈ pre
f (post ({y}))]
[as preR is monotone]
[as preR is forward complete for pre
f ]
[as x ∈ preR ({y})]
and this closes the proof.
8 Related work
Loiseaux et al. [39] generalized the standard approach to abstract model checking to more general abstract models where an abstraction relation σ ⊆ States × A is used instead of a surjective function
h : States → A. However, the results of strong preservation given there (cf. [39, Theorems 3 and 4])
require the hypothesis that the relation σ is difunctional, i.e. σ = σσ −1 σ. In this case the abstraction
relation σ can indeed be derived from a function, so that the class of strongly preserving abstract models
in Loiseaux et al.’s framework is not really larger than the class of standard partition-based abstract models
(see the detailed discussion by Dams et al. [20, Section 8.1]).
Giacobazzi and Quintarelli [28] first noted that strong preservation is related to completeness in abstract interpretation by studying the relationship between complete abstract interpretations and Clarke et
al.’s [6, 7, 8] spurious counterexamples. Given a formula ϕ of ACTL, a model checker running on a standard abstract Kripke structure defined over a state partition P may provide a spurious counterexample π ♯
for ϕ, namely a path of abstract states, namely blocks of P , which does not correspond to a real concrete
counterexample. In this case, by exploiting the spurious counterexample π ♯ , the partition P is refined to P ′
by splitting a single block of P . As a result, this refined partition P ′ does not admit the spurious counterexample π ♯ for ϕ so that P ′ is given as a new refined abstract model for ϕ to the model checker. Giacobazzi
and Quintarelli [28] cast spurious counterexamples for a partition P as a lack of (standard) completeness
in the abstract interpretation sense for the corresponding partitioning abstract domain adp (P ). Then, by
applying the results in [31] they put forward a method for systematically refining abstract domains in order
to eliminate spurious counterexamples. The relationship between completeness and spurious counterexamples was further studied in [18], where it is also shown that a block splitting operation in Paige and
Tarjan [42] partition refinement algorithm can be characterized in terms of complete abstract interpretations. More in general, the idea of systematically enhancing the precision of abstract interpretations by
refining the underlying abstract domains dates back to the early works by Cousot and Cousot [14], and
evolved to the systematic design of abstract interpretations by abstract domain refinements [26, 29, 31].
9 Conclusion
This work shows how the abstract interpretation technique allows to generalize the notion of strong preservation from standard abstract models specified as abstract Kripke structures to generic domains in abstract
interpretation. For any inductively defined language L , it turns out that strong preservation of L in a
standard abstract model checking framework based on partitions of the space state Σ becomes a particular
29
instance of the property of forward completeness of abstract domains w.r.t. the semantic operators of the
language L . In particular, a generalized abstract model can always be refined through a fixpoint iteration
to the most abstract domain that strongly preserves L . This generalizes in our framework the idea of
partition refinement algorithms that reduce the state space Σ in order to obtain a minimal abstract Kripke
structure that is strongly preserving for some temporal language.
This work deals with generic temporal languages consisting of state formulae only. As future work,
it would be interesting to study whether the ideas of our abstract interpretation-based approach can be
applied to linear languages like LTL consisting of formulae that are interpreted as sets of paths of a Kripke
structure. The idea here is to investigate whether standard strong preservation of LTL can be generalized to
abstract interpretations of the powerset of traces and to the corresponding completeness properties. Fairness
can be also an interesting topic of investigation, namely to study whether our abstract interpretation-based
framework allows to handle fair semantics and fairness constraints [10].
Finally, let us mention that the results presented in this paper led to design a generalized Paige-Tarjan
refinement algorithm based on abstract interpretation for computing most abstract strongly preserving domains [45]. As shown in Section 6, a most abstract strongly preserving domain can be characterized as
a greatest fixpoint computation in Abs(℘(Σ)). It is shown in [45] that the Paige-Tarjan algorithm [42]
can be viewed exactly as a corresponding abstract greatest fixpoint computation in Part(Σ). This leads to
an abstract interpretation-based Paige-Tarjan-like refinement algorithm that is parameteric on any abstract
interpretation of the lattice Abs(℘(Σ)) of abstract domains of ℘(Σ) and on any generic inductive language
L.
Acknowledgements. We wish to thank Mila Dalla Preda and Roberto Giacobazzi who contributed to the
early stage of this work. This paper is an extended and revised version of [44]. This work was partially
supported by the FIRB Project “Abstract interpretation and model checking for the verification of embedded
systems” and by the COFIN2004 Project “AIDA: Abstract Interpretation Design and Applications”.
References
[1] K.R. Apt and G.D. Plotkin. Countable nondeterminism and random assignment. J. ACM, 33(4):724–767, 1986.
[2] B. Bloom and R. Paige. Transformational design and implementation of a new efficient solution to the ready
simulation problem. Sci. Comp. Program., 24(3):189–220, 1995.
[3] A. Bouajjani, J.-C. Fernandez and N. Halbwachs. Minimal model generation. In Proc. of the 2nd Internat. Conf.
on Computer Aided Verification (CAV’90), LNCS 531, pp. 197–203, Springer, 1990.
[4] M.C. Browne, E.M. Clarke and O. Grumberg. Characterizing finite Kripke structures in propositional temporal
logic. Theoret. Comp. Sci., 59:115–131, 1988.
[5] D. Bustan and O. Grumberg. Simulation-based minimization. ACM Trans. Comput. Log., 4(2):181–204, 2003.
[6] E.M. Clarke, O. Grumberg, S. Jha, Y. Lu and H. Veith. Counterexample-guided abstraction refinement. In Proc.
of the 12th Internat. Conf. on Computer Aided Verification (CAV’00), LNCS 1855, pp. 154–169, Springer, 2000.
[7] E.M. Clarke, O. Grumberg, S. Jha, Y. Lu and H. Veith. Counterexample-guided abstraction refinement for
symbolic model checking. J. ACM, 50(5):752–794, 2003.
[8] E.M. Clarke, S. Jha, Y. Lu and H. Veith. Tree-like counterexamples in model checking. In Proc. of the 17th IEEE
Symp. on Logic in Computer Science (LICS’02), pp. 19–29, IEEE Press, 2002.
[9] E.M. Clarke, O. Grumberg and D. Long. Model checking and abstraction. ACM Trans. Program. Lang. Syst.,
16(5):1512–1542, 1994.
[10] E.M. Clarke, O. Grumberg and D.A. Peled. Model checking. The MIT Press, 1999.
[11] R. Cleaveland, S.P. Iyer, D. Yankelevich. Optimality in abstractions of model checking. In Proc. 2nd Intern.
Static Analysis Symposium (SAS’95), LNCS 983, pp. 51–63, Springer, 1995.
[12] R. Cleaveland, J. Parrow and B. Steffen. The Concurrency Workbench: a semantics based tool for the verification
of concurrent systems. ACM Trans. Program. Lang. Syst., 15(1):36–72, 1993.
[13] P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by
construction or approximation of fixpoints. In Proc. 4th ACM POPL, pp. 238–252, 1977.
30
[14] P. Cousot and R. Cousot. Systematic design of program analysis frameworks. In Proc. 6th ACM POPL, pp. 269–
282, 1979.
[15] P. Cousot and R. Cousot. Higher-order abstract interpretation (and application to comportment analysis generalizing strictness, termination, projection and PER analysis of functional languages). In Proc. IEEE Int. Conf. on
Computer Languages (ICCL’94), pp. 95–112, 1994.
[16] P. Cousot and R. Cousot. Refining model checking by abstract interpretation. Automated Software Engineering
Journal, 6(1):69–95, 1999.
[17] P. Cousot and R. Cousot. Temporal abstract interpretation. In Proc. 27th ACM POPL, pp. 12–25, 2000.
[18] M. Dalla Preda. Completeness and stability in abstract model checking. Laurea Thesis (in Italian), Univ. of
Verona, Italy, 2003.
[19] D. Dams. Abstract interpretation and partition refinement for model checking. Ph.D. Thesis, Eindhoven University of Technology, The Netherlands, 1996.
[20] D. Dams, O. Grumberg and R. Gerth. Abstract interpretation of reactive systems. ACM Trans. Program. Lang.
Syst., 16(5):1512–1542, 1997.
[21] J.W. De Bakker, J.-J.C. Meyer and J.I. Zucker. On infinite computations in denotational semantics. Theoret.
Comp. Sci., 26(1-2):53–82, 1983.
[22] R. De Nicola and F. Vaandrager. Three logics for branching bisimulation. J. ACM, 42(2):458–487, 1995
[23] A. Dovier, C. Piazza and A. Policriti. An efficient algorithm for computing bisimulation equivalence. Theoret.
Comp. Sci., 311(1-3):221–256, 2004.
[24] E.A. Emerson, A.K. Mok, A.P. Sistla and J. Srinivasen. Quantitative temporal reasoning. In Proc. of the 2nd
Internat. Conf. on Computer Aided Verification (CAV’90), LNCS 531, pp. 136–145, Springer, 1990.
[25] E.A. Emerson and E.M. Clarke. Characterizing correctness properties of parallel programs using fixpoints. In
Proc. ICALP’80, LNCS 85, pp. 169–181, Springer, 1980.
[26] G. Filé, R. Giacobazzi and F. Ranzato. A unifying view of abstract domain design. ACM Comput. Surv.,
28(2):333–336, 1996.
[27] R. Gentilini, C. Piazza and A. Policriti. From bisimulation to simulation: coarsest partition problems. J. Automated Reasoning, 31(1):73-103, 2003.
[28] R. Giacobazzi and E. Quintarelli. Incompleteness, counterexamples and refinements in abstract model checking.
In Proc. 8th Intern. Static Analysis Symposium (SAS’01), LNCS 2126, pp. 356–373, Springer, 2001.
[29] R. Giacobazzi and F. Ranzato. Refining and compressing abstract domains. In Proc. 24th ICALP, LNCS 1256,
pp. 771–781, Springer, 1997.
[30] R. Giacobazzi and F. Ranzato. Optimal domains for disjunctive abstract interpretation. Sci. Comp. Program.,
32:177–210, 1998.
[31] R. Giacobazzi, F. Ranzato and F. Scozzari. Making abstract interpretations complete. J. ACM, 47(2):361–416,
2000.
[32] J.F. Groote and F. Vaandrager. An efficient algorithm for branching bisimulation and stuttering equivalence. In
Proc. ICALP’90, LNCS 443, pp. 626-638, Springer, 1990.
[33] O. Grumberg and D.E. Long. Model checking and modular verification. ACM Trans. Program. Lang. Syst.,
16(3):843–871, 1994.
[34] M. Hennessy and R. Milner. Algebraic laws for nondeterminism and concurrency. J. ACM, 32(1):137–161, 1985.
[35] M.R. Henzinger, T.A. Henzinger and P.W. Kopke. Computing simulations on finite and infinite graphs. In Proc.
36th FOCS, pp. 453–462, IEEE Press, 1995.
[36] T.A. Henzinger, R. Maujumdar and J.-F. Raskin. A classification of symbolic transition systems. ACM Trans.
Comput. Log., 6(1), 2005.
[37] L. Lamport. What good is temporal logic? In Information Processing ’83, pp. 657–668, IFIP North-Holland,
1983.
[38] D. Lee and M. Yannakakis. Online minimization of transition systems. In Proc. 24th ACM STOC, pp. 264–274,
1992.
[39] C. Loiseaux, S. Graf, J. Sifakis, A. Bouajjani and S. Bensalem. Property preserving abstractions for the verification of concurrent systems. Formal Methods in System Design, 6:1–36, 1995.
31
[40] D. Massé. Semantics for abstract interpretation-based static analyzes of temporal properties. In Proc. 9th Intern.
Static Analysis Symposium (SAS’02), LNCS 2477, pp. 428–443, Springer, 2002.
[41] D. Massé. Abstract domains for property checking driven analysis of temporal properties. In Proc. 10th Intern.
Conf. on Algebraic Methodology and Software Technology (AMAST’04), LNCS 3116, pp. 349–363, Springer,
2004.
[42] R. Paige and R.E. Tarjan. Three partition refinement algorithms. SIAM J. Comput., 16(6):973–989, 1987
[43] F. Ranzato and F. Tapparo. Making abstract model checking strongly preserving. In Proc. 9th Intern. Static
Analysis Symposium (SAS’02), LNCS 2477, pp. 411–427, Springer, 2002.
[44] F. Ranzato and F. Tapparo. Strong preservation as completeness in abstract interpretation. In Proc. 13th European
Symposium on Programming (ESOP’04), LNCS. 2986, pp. 18–32, Springer, 2004.
[45] F. Ranzato and F. Tapparo. An abstract interpretation-based refinement algorithm for strong preservation. In
Proc. 11th Intern. Conf. on Tools and Algorithms for the Construction and Analysis of Systems (TACAS’05),
LNCS 3440, pp. 140–156, Springer, 2005.
[46] F. Ranzato and F. Tapparo. An efficient algorithm for computing simulation equivalence based on abstract interpretation. In preparation, 2006.
[47] D.A. Schmidt. Closed and logical relations for over- and under-approximation of powersets. In Proc. 11th Intern.
Static Analysis Symposium (SAS’04), LNCS 3148, pp. 22–37, Springer, 2004.
[48] L. Tan and R. Cleaveland. Simulation revisited. In In Proc. 7th Intern. Conf. on Tools and Algorithms for the
Construction and Analysis of Systems (TACAS’01), LNCS 2031, pp. 480-495, Springer, 2001.
[49] R.J. van Glabbeek. The linear time - branching time spectrum. In Handbook of Process Algebra, pp. 3–99,
Elsevier, 2001.
32
| 2cs.AI
|
People, Penguins and Petri Dishes: Adapting Object Counting Models To New
Visual Domains And Object Types Without Forgetting
arXiv:1711.05586v1 [cs.CV] 15 Nov 2017
Mark Marsden1 , Kevin McGuinness1 , Suzanne Little1 , Ciara E. Keogh2 , Noel E. O’Connor1
1
Insight Centre for Data Analytics
Dublin City University, Ireland
[email protected] {kevin.mcguinness,suzanne.little,noel.oconnor}@dcu.ie
2
Conway Institute, School of Medicine
University College Dublin, Ireland
[email protected]
Abstract
In this paper we propose a technique to adapt a convolutional neural network (CNN) based object counter to additional visual domains and object types while still preserving
the original counting function. Domain-specific normalisation and scaling operators are trained to allow the model
to adjust to the statistical distributions of the various visual
domains. The developed adaptation technique is used to
produce a singular patch-based counting regressor capable
of counting various object types including people, vehicles,
cell nuclei and wildlife. As part of this study a challenging
new cell counting dataset in the context of tissue culture
and patient diagnosis is constructed. This new collection,
referred to as the Dublin Cell Counting (DCC) dataset, is the
first of its kind to be made available to the wider computer
vision community. State-of-the-art object counting performance is achieved in both the Shanghaitech (parts A and
B) and Penguins datasets while competitive performance
is observed on the TRANCOS and Modified Bone Marrow
(MBM) datasets, all using a shared counting model.
Figure 1: Examples of the challenging images encountered
in different object counting domains.
1. Introduction
and perspective issues, visual occlusion and poor illumination. To date, these highly related counting tasks have been
tackled separately with algorithms engineered specifically
to perform object counting in a given domain. Examples of
these challenges are shown across several visual domains in
Figure 1.
Vision based object counting has been tackled using both
object detection [1, 2, 3] and count regression techniques
[4, 5, 6, 7, 8] with both types of approach having their own
specific weakness (visual occlusion and model overfitting
respectively). The utilisation of convolutional neural net-
Vision-based object counting is an important analysis step
in many observational scenarios including wildlife studies,
microscopic imaging and CCTV surveillance. An accurate
object counting system can provide valuable insights such
as the congestion level of a public square (crowd counting),
the level of traffic on a motorway (vehicle counting), the
inferred migration patterns of a penguin colony (wildlife
counting) or the proliferation of cancerous cells in a patient
(cell counting). There is a common set of challenges for
vision-based object counting which limit counting accuracy
in all of these domains. These challenges include object scale
1
works (CNN) and hardware accelerated optimisation has lead
to significant improvements in the accuracy of regressionbased counting approaches across several visual domains
[9, 10, 11]. This paper looks to build upon this work and
investigate if a patch-based object count regressor can be
adapted to additional object types and visual domains while
maintaining accuracy in the original counting task. The potential benefits of a modular counting architecture include
transfer learning and the removal of redundant model parameters. The significant variation in statistical distribution between visual domains is one of the main obstacles
to domain adaptation in computer vision [12]. This challenge is addressed in our approach by including a set of
domain-specific scaling and normalization layers distributed
throughout the network which make up only a small fraction
of the overall parameter count [12]. A sequential training
procedure inspired by the work of Rebuffi et al. [12] allows
the network to be extended to new counting tasks over time
while preserving all previously learned counting functions.
The set of count estimates produced using the proposed
patch-based counting model are then refined using an efficient fully convolutional neural network which utilises the
wider scene context to mitigate possible count errors. The
proposed framework is also extended to perform a visual
domain classification in the event of the observed domain
being unknown, highlighting the versatility and modular nature of the approach. The core contributions of this paper
can be summarised as follows:
• An extendable CNN architecture for patch-based, multidomain object counting is developed for the first time;
• A fully convolutional neural network is utilised to refine
a set of count estimates for an entire image;
• A challenging, representative dataset for cell counting
in a tissue culture/patient diagnosis setting is proposed;
• Domain adaptation in object counting is shown to produce more efficient, higher accuracy counting models;
• State-of-the-art counting accuracy is observed on leading benchmarks for several object types including the
Shanghaitech [13] and Penguins [14] datasets.
The remainder of this paper is organised as follows: Section 2 presents the related work found in the literature. Section 3 describes the proposed multi-domain object counting
approach while Section 4 details the construction of a challenging new cell counting dataset. Finally, Section 5 presents
a comprehensive set of experiments highlighting the development of our technique and the benefits associated with
multi-domain object counting.
2. Related work
Crowd Counting. Crowd counting has been approached
using a wide variety of techniques, from HOG-based head
detectors [15] to CNN-based regressors [9]. Heatmap-based
crowd counting using a fully convolutional neural network
was firstly investigated by Zhang et al. [13] and subsequently
by Marsden et al. [16], with notable performance gains observed. A novel model switching technique for crowd counting was proposed by Sam et al. [17] which firstly classifies
the crowd density level of an image region before performing
heatmap-based counting using a network which has been
optimised for the detected crowd density level. The standard
datasets for evaluating crowd counting techniques include
the UCF CC 50 [15] and Shanghaitech [13] datasets.
Cell counting. Cell nuclei counting techniques have
evolved from SIFT-based heatmap estimation [6] to fully
convolutional neural networks [18, 19], with significant improvements in accuracy observed. Significant variation in
cell morphology and visual occlusion limit the accuracy of
these techniques. The VGG Cell dataset [20] (made up entirely of synthetic images) is the main public benchmark used
to compare cell counting techniques. Recently the Modified
Bone Marrow (MBM) dataset [19] was introduced, containing images of real cells observed in histological slides (a
2D cross section of a 3D tissue structure). However, a cell
counting dataset in the context of tissue culture and patient
diagnosis is still needed. Cell counting in tissue culture is often done by hand on a daily basis, especially in cases where
highly expensive automated systems are not available. A
vision-based cell counting system for tissue culture can lead
to a more affordable solution and faster diagnoses overall.
Vehicle counting. Vehicle counting has been tackled using a variety of techniques including SIFT-based regression
[21], CNN regression [10], density heatmap generation [22]
as well as an LSTM (Long Short-Term Memory) based approach from Zhang et al. [23] which adds a temporal dimension to vehicle counting. Vehicle counting techniques are
trained and evaluated using the WebCamT [23] and TRANCOS [21] datasets.
Wildlife counting. Counting of wildlife in ecological
studies has not received significant attention from the computer vision community. The core work in this area is that
of Arteta et al. [14] who produced a large-scale dataset for
counting penguin colonies.
Domain adaptation and shared learning models.
While this paper focuses primarily on domain adaptation it
also touches on concepts including transfer learning, feature
extraction and learning without forgetting (LwF). The ability
to train a machine learning model to perform additional tasks
over time while maintaining accuracy in previous tasks has
generated significant interest in the computer vision community. This notion is inspired largely by the human visual
system, which learns a universal representation for vision in
the early life and uses this representation for a variety of problems [12]. Fine-tuning is a common process used to adapt a
given neural network to a new task, the main downside being
that the original function is often lost during optimisation.
Multi-Task learning (MTL) approaches attempt to train a
model to simultaneously perform several tasks, often within
a specific visual domain. MTL approaches typically involve
including sets of task specific layers at end of a given neural
network and jointly training to perform all tasks. This type
of approach is cumbersome to extend to new tasks and visual
domains as all tasks must be retrained.
While intra-domain, multi-objective learning has been utilized heavily to extend the functionality of neural networks,
training a model to perform tasks in various distinct visual
domains (CCTV scenes, medical imaging) has proven to be
more challenging due to the observed variation in statistical distributions between domains. These issues have been
addressed by Bilen et al. [24] who trained a CNN model
to learn a universal vision representation which can jointly
perform non-related tasks from distinct visual domains by
including domain-specific scaling and normalisation layers
throughout the network. This work was extended by Rebuffi et al. [12] who added domain-specific convolutions and
proposed a sequential training procedure for learning new
tasks over time without discarding the previously learned
functions.
connected network consists of 5 layers with the following
configuration of neurons 256-128-64-64-1. Rectified linear
unit (ReLU) activations are applied after each fully connected layer. The trainable portion of the overall network
consists of just 330,000 parameters. A variety of pre-trained
object classification networks and image patch sizes are investigated in Section 5.
3. Our approach
A sequential training procedure inspired by the work of
Rebuffi et al. [12] is used to allow the proposed counting
regressor to be extended to new domains over time while
still performing the original set of counting tasks. First the
network is primed by training to convergence on a given visual domain (e.g. crowd counting), after which the domainagnostic parameters (i.e. the fully connected layers) are
frozen and only the 5% subset of domain-specific modules
are trained to convergence for each of the remaining visual
domains, one at a time. Freezing the domain-agnostic parameters enables the network to retain the original counting
regression function it has learned. Euclidean distance, given
in equation 1 is optimised during training. Θ corresponds
to the set of network parameters to optimise, N is the batch
size, Xi is the ith batch image while Fi is the corresponding
ground truth count value. F (Xi ; Θ) is the estimated count
value for a given batch image Xi .
The proposed object counting technique consists of a
patch-based CNN regressor with significant inter-domain
parameter sharing that can be quickly switched between a
learned set of visual domains by interchanging a subset of
domain specific parameters. The estimated object count for a
given image is calculated by summing the object count for all
patches. A patch based approach to counting enables a more
robust regressor to be learned as objects within a smaller
image region are observed to be approximately uniform
in size. A lightweight fully convolutional neural network
is then used to refine a set of patch estimates for a given
image by including the context of adjacent patch estimates
to mitigate possible errors.
3.1. Base object counting regressor
The base object counting regressor (shown in Figure 2)
consists of two distinct steps. First a set of high-level features are extracted from each image patch using a pre-trained
image classification network, the parameters of which are
frozen during model training. The N feature maps generated by the given image classification network’s final convolutional layer are average pooled globally to produce an
N -dimensional feature representation. This N -dimensional
feature representation is then mapped to an object count
value using a fully connected neural network. This fully
3.2. Domain-specific layers
To enable the proposed counting regressor to adapt to
various visual domains (e.g. people, vehicles, wildlife, cell
nuclei) a set of domain-specific modules are included before each fully connected layer and then after the final fully
connected layer. These domain-specific modules are interchanged during training and inference depending on the chosen visual domain (this switching concept is highlighted in
Figure 3). Each domain-specific module contains the residual adapter of Rebuffi et al. [12] (shown in Figure 4). These
residual adapter modules allows for the network to adapt
to the distinct statistical distributions of the various visual
domains through domain-specific normalisation and scaling
[25]. Including these domain-specific modules throughout
the network increases the trainable parameter count by just
5% for each additional visual domain learned.
3.3. Sequential training
Ll2 (Θ) =
N
1 X
2
kF (Xi ; Θ) − Fi k2 .
2N i=1
(1)
The AdaGrad optimiser [26] is used to avoid learning
rate selection issues with the initial learning rate set to 1 ×
10−1 . L2 weight regularisation (i.e. weight decay) is also
included during model training with λ set to 1×10−3 . Model
weights are initalised using the uniform initaliser of Glorot
and Bengio [27] while the bias terms are initalised to zero.
Figure 2: Network architecture for the proposed patch-based object counting regressor
A refinement model can be trained for the count regressor of a given domain by firstly producing a grid of patch
estimates for each training image used as well as a corresponding ground truth grid. Training is then carried out using
these grid pairs for 10,000 iterations with Euclidean distance
again minimised. Once trained, a refinement model can be
applied to a set of count estimates of any width and height
due to the fully convolutional nature of the model. During
inference this step results in a near negligible increase in
processing time.
Figure 3: Domain specific modules are interchanged during
training and inference depending on the chosen counting
task (red path).
Figure 5: Fully convolutional network used for count estimation refinement.
Figure 4: The residual adapter module of Rebuffi et al. [12].
Training is carried out for 10,000 iterations for each domain.
The choice of visual domain used to prime the network is
investigated in section 5.
3.4. Fully convolutional refinement network
A patch-based counting regressor does not include the
wider scene context when processing a given image as
patches are analysed on an individual basis. To address
this, a fully convolutional neural network is trained to refine
a grid of patch estimates for a given image. The proposed
fully convolutional network, shown in Figure 5, consists of
4 convolutional layers with the following number of kernels
per layer: 16,16,16,1. Rectified linear unit (ReLU) activations are applied after each layer. All convolutional kernels
are 3 × 3 resulting in a total parameter count of 4950.
3.5. Domain classification
If the visual domain observed during inference is unknown this can be predicted by extending the core network
to also perform domain classification. To accomplish this
the final fully connected layer is interchanged with a Kneuron fully connected layer, where K is the number of
visual domains to distinguish between. Following this final
layer a softmax activation is applied and the subsequent domain adapter module is not included. Training can then be
performed by freezing the common model parameters (the
initial 4 fully connected layers) and creating a fresh set of
adapter modules. Categorical cross-entropy loss, defined as:
LCCE (Θ) = −
N K
1 XX
Sij log(Ŝij ),
N i=1 j=1
(2)
is then minimized. Θ corresponds to the set of trainable
network parameters, N and K are the batch size and number
of visual domains to be distinguished between while Ŝij is
the predicted probability score for concept j in the ith batch
image and Sij is the corresponding ground truth value.
4. Dublin cell counting dataset
When considering the application of computer vision
to tissue culture and patient diagnosis there is a clear lack
of publicly available and fully annotated cell microscopy
datasets. The main dataset used to evaluate techniques for
this task consists entirely of synthetic images [18]. To address this, the Dublin Cell Counting (DCC) dataset was
constructed.
This dataset consists of 177 images containing a wide
array of tissues and species. Amongst these are examples
of stem cells derived from embryonic mice, isolated human lung adenocarcinoma and examples of primary human
monocytes isolated from a healthy human volunteer. Several factors were varied during image capture to provide a
more representative set of images. First, the density of cells
loaded onto the slide naturally varies as cell lines proliferate
at different rates. Second, the morphology and size of the
cells for each cell line can vary significantly. Furthermore,
the objective lens used during imaging was varied as was
the diameter of the diaphragm which controls the amount
of light hitting the sample. Finally, the haemocytometer
grid size was varied to produce a representative set of noncellular image artifacts. Cell images were obtained via a
camera mounted on an Olympus CKX41 microscope using
both 4× and 10× objectives. The high levels of variation in
this collection allow for a more robust cell counting function
to be learned. After the full set of image were acquired, a
dot annotation process was performed by a domain expert
with a background in molecular biology.
The mean cell count across these images is 34.1 with a
standard deviation of 21.8, showing the significant variation
in cell density. 100 images are used for training and validation while the remaining 77 form an unseen test set. Sample
images from this collection are shown in Figure 6.
5. Experiments
The proposed multi-domain object counting technique
is evaluated using a challenging and representative dataset
for each visual domain including the proposed DCC dataset.
These collections are detailed in full in Table 1. Horizontal
flips are used for training set augmentation in all cases. A
30% subset of the provided training data is set aside as
a validation set for all model selection experiments, apart
from where an explicit validation set is provided. No count
estimate refinement is applied until subsection 5.5. Counting
performance is evaluated for all visual domains using Mean
Absolute Error (MAE) and Root Mean Squared Error (MSE),
which are defined as follows:
Figure 6: DCC dataset examples showing the significant
variation within this collection.
N
1 X
|zi − ži | ,
N i=1
v
u
N
u1 X
(zi − ži )2 ,
MSE = t
N i=1
MAE =
(3)
(4)
All network optimisation and testing is performed using an
NVIDIA GeForce GTX 970 GPU with a batch size of 256
and implemented using the Keras API [28] with a Tensorflow
backend [29].
5.1. Patch size selection
Several image patch sizes (50 × 50, 100 × 100, 200 ×
200) are compared by evaluating crowd counting validation
performance on the Shanghaitech dataset (part A). Feature
extraction for the base counting regressor is performed using
the VGG16 network of Simonyan and Zisserman [30] for
all runs. A given patch size is used during both dataset
construction and inference. Both the domain-agnostic and
domain-specific network parameters are optimised for each
run in this experiment as the network has not been primed
on any other dataset. Table 2 highlights the performance
of the various patch sizes. 100 × 100 patches result in the
best overall performance and will be used in all subsequent
experiments. This size likely strikes the optimal balance
between uniform object size and the inclusion of wider scene
context needed to learn a robust function. On the other hand,
the use of 200 × 200 patches results in significantly inferior
Visual Domain
Dataset
No. of Images
Crowd
Vehicles
Wildlife
Cells
Shanghaitech (Part A) [13]
TRANCOS [21]
Penguins [14]
Dublin Cell Counting
482
1641
80095
177
Count
Mean
501.4
36.54
7.18
34.1
Count
STD
456.4
14.9
5.71
21.8
Count
Range
33-3139
9-95
0-67
0-101
Table 1: Details of the datasets used to evaluate the proposed multi-domain object counting technique.
performance due to the non-uniform object sizes observed
in larger image regions.
Patch Size
50 × 50
100 × 100
200 × 200
MAE
100.8
97.5
158.5
MSE
152.5
145.4
247.3
Table 2: Crowd counting validation performance of various image patch sizes on a validation set taken from the
Shanghaitech (Part A) dataset.
5.2. Feature extractor selection
Several pre-trained image classification networks are evaluated as feature extractors for object counting. These networks include the ResNet50 network of He et al. [31] the
VGG16 network of Simonyan and Zisserman [30] as well
as the MobileNet architecture of Howard et al. [32]. Crowd
counting validation performance is again evaluated on the
Shanghaitech (Part A) dataset with a patch size of 100 ×
100 used. Table 3 details the various pre-trained networks
and presents the performance achieved by each. The number of parameters associated with each network does not
include the original set of fully connected layers (which
have been replaced entirely). The MobileNet feature extractor achieves the best overall performance despite having
significantly fewer parameters and is used for all subsequent
experiments. The use of this lightweight architecture also
potentially makes the developed method more suitable for
edge processing on low power devices.
Net
VGG16 [30]
Resnet50 [31]
MobileNet (α = 1.0) [32]
Parameters
14M
24M
3.5M
MAE
101.2
97.5
94.1
MSE
151.4
145.4
138.9
Table 3: Validation performance on the Shanghaitech dataset
(part A) for various image classification networks being used
for feature extraction.
5.3. Priming the network
In the work of Rebuffi et al. [12] a multi-domain image
classification network is primed on the ImageNet dataset
[33] due to it’s size and variety before being adapted to other
domains. However, for the proposed multi-domain object
counting model the choice of which domain with which to
prime the network is not clear. Therefore all 4 domains
(crowds, vehicles, wildlife, cells) are evaluated for this purpose by priming the network from scratch and then adapting
to the remaining 3 domains one by one using the sequential
training process discussed previously. Table 4 presents the
MAE score observed for each pairing. Each row in this table
corresponds to the domain used to prime the network while
each column corresponds to the MAE performance achieved
when adapting the model. The diagonal entries correspond
to the performance achieved when training the network from
scratch for each domain.
It can be seen that priming the network on the cell domain using the DCC dataset results in the best overall performance, achieving optimal MAE on 3 of the 4 domains.
Adapting from the cell domain achieves performance superior to training from scratch on both the crowd and wildlife
domains, which is noteworthy given the small number of
domain-specific parameters trained (5%). The high performance observed across domains when adapting from a cell
counting model is likely due to the significant morphological
variation observed between cell objects in the DCC dataset,
resulting in a broader set of learned features. In all subsequent experiments the counting network is primed on the
cell counting task.
5.4. Comparison with feature extraction based network adaptation
Feature extraction based network adaptation involves
freezing the majority of a given neural network and retraining the final few layers for a new task (with the task-specific
parameters then interchanged as required). This approach
retains the original function but often requires a large number of new parameters to be trained. In this experiment we
compare the Rebuffi domain adapter [12] with feature extraction as a domain adaptation strategy. The approaches
are evaluated in terms of MAE performance in the target domain and the quantity of new model parameters introduced.
Visual Domain
Crowd (Primed)
Vehicles (Primed)
Wildlife (Primed)
Cells (Primed)
Crowd (Tested)
94.1
96.2
95.3
90.2
Vehicles (Tested)
10.3
9.9
10.3
10.2
Wildlife (Tested)
6.8
6.1
6.05
5.7
Cells (Tested)
12.9
11.5
10.2
9.5
Best Overall
0/4
1/4
0/4
3/4
Table 4: The MAE validation performance achieved when priming the network using each of the 4 visual domains.
The number of retrained fully connected layers is varied to
investigate the difference in performance. All adapter modules placed before trainable layers are re-trained while those
before frozen layers are also frozen. A given cell counting
model is adapted to perform crowd counting on the Shanghaitech dataset (part A) in each case. Figure 5 highlights the
superior performance of the Rebuffi domain adapter when
adapting to crowd counting, despite using significantly fewer
parameters.
Approach
From Scratch
Final 4 layers re-trained
Final 3 layers re-trained
Final 2 layers re-trained
Rebuffi et al. [12]
Extra Parameters
330K
50K
15K
5K
16K
MAE
94.1
128.3
232.2
245.5
90.2
Table 5: MAE validation performance on the Shanghaitech
dataset (part A) for various domain adaptation strategies.
5.5. Prediction refinement
A fully convolutional refinement model is trained for
each of the 4 visual domains and compared to the validation
performance of the base regressor model in each case. Results of this experiment are shown in table 6. This efficient
post-processing step has a near negligible impact on inference time but results in significant performance boosts in the
crowd and cell counting benchmarks. Examples of the final
count regressor in action across all 4 domains are shown in
figure 7.
Visual Domain
Crowds
Vehicles
Wildlife
Cells
Base MAE
90.2
10.2
5.7
9.5
Refined-MAE
86.5
10.1
5.6
8.4
Table 6: The MAE observed on the validation sets of all 4
visual domains for the base regressor and after the proposed
refinement step has been applied.
5.6. Domain classification performance
A domain classifier is trained by firstly producing a representative dataset taken from all 4 visual domains used in this
study. 300 full scene images are taken from each collection
and a 6:4 training/test split is then applied, ensuring an even
number of images are taken from each domain. Horizontal
image flips are applied to produce a 300 image set from the
smaller DCC dataset. Table 7 compares classifier accuracy
when training from scratch, when adapting from a cell counting network using a fresh set of adapter modules and when
adapting from a cell counting network and just training the
final fully connected layer.
Near perfect domain classification accuracy is observed
when training the network from scratch and also when adapting from the a cell counting network using a new set of
adapter modules. However, when we only train the final
fully connected layer to perform classification we observe
significant performance degradation, showing the importance of the included domain adapter modules despite their
low parameter count.
Primer
None
Cells
Cells
Training Approach
Entire Network Trained
Adapter Modules + Final Layer
Final Layer Only
Accuracy
99.7%
98.2%
56.3%
Table 7: Domain classification accuracy as the training approach is varied and domain adapter modules are included.
5.7. Comparison with the state-of-the-art
The developed object counting technique is compared
to the leading techniques across the 4 visual domains used
in this study. Evaluation is performed in all cases on a
previously unseen test set. The network is primed on cell
counting using the DCC dataset. The proposed refinement
step is applied for each domain. Table 8 compare crowd
counting performance on the Shanghaitech dataset (parts
A and B), with state-of-the-art performance achieved on
both. The commonly used UCF CC 50 dataset [15] has
been deemed inappropriate for benchmarking as it contains
crowds too dense for the human eye to count [34] and is
therefore not used to evaluate the proposed technique. Table
9 compares the proposed technique with the leading vehicle
a) Domain: Crowd
Predicted: 801.3 GT: 819
b) Domain: Cells
Predicted: 25.6 GT: 25
c) Domain: Vehicles
Predicted: 29.2 GT: 29
d) Domain: Wildlife
Predicted: 25.2 GT: 26
Figure 7: The robust performance of the proposed counting model is highlighted on images taken from each of the 4 visual
domains used in this study.
Method
Zhang et al. [13]
Marsden et al. [16]
Switch-CNN [17]
Our-Approach
Part A
MAE MSE
110.2 173.2
126.5 173.5
90.4
135.0
85.7
131.1
Part B
MAE MSE
26.4
41.3
23.76 33.12
21.6
33.12
17.7
28.6
Table 8: Comparing the performance of various crowd counting approaches on the Shanghaitech dataset.
Method
Hydra-CNN [10]
Our Approach
Zhang et al. [23]
MAE
10.99
9.7
4.2
Table 9: Comparing performance of various vehicle counting
approaches on the TRANCOS dataset.
Method
[18]
Ours
[11]
N=5
28.9 ± 22.6
23.6± 4.6
12.6± 3.0
N=10
22.2± 11.6
21.5± 4.2
10.7± 2.5
N=15
21.3± 9.4
20.5± 3.5
8.8±2.3
Table 10: Cell counting MAE performance on the MBM
dataset. Out of the 44 images in this collection, N are used
for training, N for validation and an unseen 14 images for
testing. At least 10 runs using random dataset splits are
performed for the each N value.
counting techniques on the TRANCOS dataset while table 10
highlights cell counting performance on the Modified Bone
Marrow dataset [11]. Finally, Table 11 shows the superior
performance of our technique on the Penguins dataset [14].
6. Conclusion
In this paper we propose a new multi-domain object counting technique that employs significant parameter sharing and
achieves state-of-the-art benchmarking performance for sev-
Method
Arteta et al. [14]
Our Approach
MAE
8.11
5.8
Table 11: Comparing performance of various counting techniques on the Penguins dataset test set. MAE is computed
w.r.t the max count on each image (as there are multiple
annotators). The separate site dataset split is used and no
depth information is utilised.
eral visual domains. This model can be extended to new
counting tasks over time while still maintaining identical
performance in all prior tasks. The benefits of this singular
approach to object counting include the removal of redundant model parameters as well as noticeable increases in
counting accuracy over single-domain baseline runs. Each
new counting task requires just 20,000 additional model parameters (including the proposed estimate refinement step).
The Dublin Cell Counting (DCC) dataset was introduced
as part of this study. This new collection is the first of it’s
kind and contains a challenging and highly-varied set of
cellular images in the context of tissue culture and patient
diagnosis. Using the developed single-model approach stateof-the-art object counting performance was observed in the
Shanghaitech dataset (parts A and B) as well as the Penguins
dataset. Future work in this area will look to extend the
proposed counting model to perform additional regression
and classification tasks in fields including medical imaging,
crowd behaviour analysis and scene understanding.
Acknowledgements
This publication has emanated from research conducted
with the financial support of the Irish Research Council
and Science Foundation Ireland (SFI) under grant numbers
SFI/12/RC/2289 and 15/SIRG/3283. A special thanks is also
given to the Cummin’s group in the School of Medicine, University College Dublin for making the DCC dataset images
available to the wider research community.
References
[1] Sheng-Fuu Lin, Jaw-Yeh Chen, and Hung-Xin Chao. Estimation of number of people in crowded scenes using perspective
transformation. IEEE Transactions on Systems, Man, and
Cybernetics-Part A: Systems and Humans, 31(6):645–654,
2001. 1
[2] Bo Wu and Ramakant Nevatia. Detection of multiple, partially
occluded humans in a single image by bayesian combination
of edgelet part detectors. In Tenth IEEE International Conference on Computer Vision (ICCV’05) Volume 1, volume 1,
pages 90–97. IEEE, 2005. 1
[3] Weina Ge and Robert T Collins. Marked point processes for
crowd counting. In Computer Vision and Pattern Recognition,
2009. CVPR 2009. IEEE Conference on, pages 2913–2920.
IEEE, 2009. 1
[4] Chen Change Loy, Shaogang Gong, and Tao Xiang. From
semi-supervised to transfer counting of crowds. In Proceedings of the IEEE International Conference on Computer Vision, pages 2256–2263, 2013. 1
[5] Ke Chen, Chen Change Loy, Shaogang Gong, and Tony Xiang. Feature mining for localised crowd counting. In BMVC,
volume 1, page 3, 2012. 1
[6] Victor Lempitsky and Andrew Zisserman. Learning to count
objects in images. In Advances in Neural Information Processing Systems, pages 1324–1332, 2010. 1, 2
[7] Tongliang Liu and Dacheng Tao. On the robustness and
generalization of cauchy regression. In 2014 4th IEEE International Conference on Information Science and Technology,
pages 100–105. IEEE, 2014. 1
[8] Antoni B Chan and Nuno Vasconcelos. Counting people with
low-level features and bayesian regression. IEEE Transactions on Image Processing, 21(4):2160–2177, 2012. 1
[9] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang
Yang. Cross-scene crowd counting via deep convolutional
neural networks. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
volume 07-12-June, pages 833–841, 2015. 2
[10] Daniel Onoro-Rubio and Roberto J López-Sastre. Towards
perspective-free object counting with deep learning. In European Conference on Computer Vision, pages 615–629.
Springer, 2016. 2, 8
[11] Joseph Paul Cohen, Henry Z Lo, and Yoshua Bengio. Countception: Counting by fully convolutional redundant counting.
arXiv preprint arXiv:1703.08710, 2017. 2, 7, 8
[12] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi.
Learning multiple visual domains with residual adapters.
2017 Conference on Neural Information Processing Systems
(NIPS), 2017. 2, 3, 4, 6, 7
[13] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao,
and Yi Ma. Single-image crowd counting via multi-column
convolutional neural network. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 589–597, 2016. 2, 6, 8
[14] Carlos Arteta, Victor Lempitsky, and Andrew Zisserman.
Counting in the wild. In European Conference on Computer
Vision, pages 483–498. Springer, 2016. 2, 6, 7, 8
[15] Haroon Idrees, Imran Saleemi, Cody Seibert, and Mubarak
Shah. Multi-source multi-scale counting in extremely dense
crowd images. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 2547–2554,
2013. 2, 7
[16] Mark Marsden, Kevin McGuiness, Suzanne Little, and Noel E
O’Connor. Fully Convolutional Crowd Counting On Highly
Congested Scenes. In The International Conference on Computer Vision Theory and Applications, 2017. 2, 8
[17] Deepak Babu Sam, Shiv Surya, and R Venkatesh Babu.
Switching convolutional neural network for crowd counting.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, volume 1, page 6, 2017. 2, 8
[18] Weidi Xie, J Alison Noble, and Andrew Zisserman. Microscopy cell counting and detection with fully convolutional
regression networks. Computer Methods in Biomechanics
and Biomedical Engineering: Imaging & Visualization, pages
1–10, 2016. 2, 5, 8
[19] Joseph Paul Cohen, Genevieve Boucher, Craig A Glastonbury,
Henry Z Lo, and Yoshua Bengio. Count-ception: Counting
by fully convolutional redundant counting. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 18–26, 2017. 2
[20] Weidi Xie, J. Alison Noble, and Andrew Zisserman. Microscopy cell counting and detection with fully convolutional
regression networks. Computer Methods in Biomechanics
and Biomedical Engineering: Imaging & Visualization, pages
1–10, 2016. 2
[21] Ricardo Guerrero-Gómez-Olmedo, Beatriz Torre-Jiménez,
Roberto López-Sastre, Saturnino Maldonado-Bascón, and
Daniel Oñoro-Rubio. Extremely overlapping vehicle counting.
In Iberian Conference on Pattern Recognition and Image
Analysis, pages 423–431. Springer, 2015. 2, 6
[22] Di Kang, Zheng Ma, and Antoni B Chan. Beyond counting: Comparisons of density maps for crowd analysis
tasks-counting, detection, and tracking. arXiv preprint
arXiv:1705.10118, 2017. 2
[23] Shanghang Zhang, Guanhang Wu, João P Costeira, and
José MF Moura. Fcn-rlstm: Deep spatio-temporal neural
networks for vehicle counting in city cameras. arXiv preprint
arXiv:1707.09476, 2017. 2, 8
[24] Hakan Bilen and Andrea Vedaldi. Universal representations:
The missing link between faces, text, planktons, and cat
breeds. arXiv preprint arXiv:1701.07275, 2017. 3
[25] Sergey Ioffe and Christian Szegedy. Batch normalization:
Accelerating deep network training by reducing internal covariate shift. 2015. 3
[26] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization.
Journal of Machine Learning Research, 12(Jul):2121–2159,
2011. 3
[27] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Aistats,
volume 9, pages 249–256, 2010. 3
[28] François Chollet. Keras. https://github.com/fchollet/keras,
2015. 5
[29] Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo,
Zhifeng Chen, Craig Citro, Greg Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale
machine learning on heterogeneous distributed systems. 2015.
5
[30] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv
preprint arXiv:1409.1556, 2014. 5, 6
[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Deep residual learning for image recognition. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 770–778, 2016. 6
[32] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry
Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv
preprint arXiv:1704.04861, 2017. 6
[33] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei.
ImageNet: A Large-Scale Hierarchical Image Database. In
CVPR09, 2009. 6
[34] Yaocong Hu, Huan Chang, Fudong Nian, Yan Wang, and
Teng Li. Dense crowd counting from still images with convolutional neural networks. Journal of Visual Communication
and Image Representation, 38:530–539, 2016. 7
| 1cs.CV
|
A Near-Linear Pseudopolynomial Time Algorithm for Subset Sum
Karl Bringmann∗
arXiv:1610.04712v2 [cs.DS] 8 Jan 2017
January 10, 2017
Abstract
Given a set Z of n positive integers and a target value t, the SubsetSum problem asks
whether any subset of Z sums to t. A textbook pseudopolynomial time algorithm by Bellman
from 1957 solves SubsetSum in time O(n
√ t). This has been improved to O(n max Z) by Pisinger
[J. Algorithms’99] and recently to Õ( n t) by Koiliaris and Xu [SODA’17].
Here we present a simple randomized algorithm running in time Õ(n + t). This improves
upon a classic algorithm and is likely to be near-optimal, since it matches conditional lower
bounds from SetCover and k-Clique.
We then use our new algorithm and additional tricks to improve the best known polynomial
space solution from time Õ(n3 t) and space Õ(n2 ) to time Õ(n t) and space Õ(n log t), assuming
the Extended Riemann Hypothesis. Unconditionally, we obtain time Õ(n t1+ε ) and space Õ(n tε )
for any constant ε > 0.
∗
Max Planck Institute for Informatics, Saarland Informatics Campus, Germany, [email protected].
1
1
Introduction
Given a (multi-)set Z of n positive integers and a target t, the SubsetSum problem asks whether
there is a subset Y of Z summing to exactly t. This classic NP-hard problem draws its significance
from the fact that it lies at the core of many other NP-hard (optimization) problems, e.g. SubsetSum can easily be reduced to the Knapsack problem, Constrained Shortest Path (also
known as Restricted Shortest Path), and many other problems.
Bellman showed in 1957 that SubsetSum can be solved in pseudopolynomial time O(nt) by
dynamic programming [5]. This algorithm is being taught for decades in undergraduate algorithms
courses and thus had a great influence on computer science. Since this pseudopolynomial time
algorithm is a fundamental part of our curriculum, and SubsetSum is one of the core NP-hard
problems, it is an important question whether the running time of O(nt) can be improved.
Pisinger [22] used basic RAM parallelism (allowing to operate on Θ(log t) bits in constant time)
to obtain the first improvement: an O(nt/ log t) algorithm. Considering s := max Z, Pisinger [21]
presented an O(ns) algorithm, which is faster than O(nt) in some situations. The first algorithm
√
breaking through the O(nt) barrier by a polynomial factor in the worst case was the recent Õ( n t)
algorithm1 by Koiliaris and Xu [17]. Roughly speaking, they use hashing to solve SubsetSum
quickly if Z is contained in a small interval. Then they reduce the general case to the small interval
case by appropriately splitting Z into smaller sets. They also presented an Õ(t4/3 ) algorithm.
There is reason to believe that SubsetSum has no t1−ε nO(1) algorithm for any ε > 0, since this
′
would yield an 2(1−ε )n (nm)O(1) algorithm for SetCover (on m sets over a universe of size n) via
the reductions in [7], and thus break the SetCover hypothesis. A second reason is that any com′
binatorial t1−ε nO(1) algorithm for SubsetSum would yield a combinatorial O(n(1−ε )k ) algorithm
for k-Clique, for some large constant k, via the reduction in [1]2 . This means that polynomial
improvements over running time t, even in terms of n, are unlikely. However, these arguments give
no evidence that the additional factor n of the O(nt) dynamic programming algorithm is necessary
√
(or the factor n of Koiliaris and Xu). Specifically, they leave the open problem whether there is
an Õ(n + t) algorithm.
Note that an Õ(n + t) algorithm would also up to polylogarithmic factors dominate the O(ns)
algorithm by Pisinger [21], where s = max Z, since any non-trivial instance satisfies s ≥ t/n.
Near-linear Time We present an algorithm in time Õ(n + t). This improves the classic dynamic
√
programming solution by a factor Ω̃(n) and the best known algorithm by a factor Ω̃( n). Moreover,
we match the conditional lower bounds, so any further polynomial improvement would beat the
O∗ (2n ) SetCover algorithm and possibly the O(nk ) combinatorial k-Clique algorithm.
Theorem 1.1 (Section 3). SubsetSum can be solved in time Õ(n + t) by a randomized, one-sided
error algorithm with error probability (n + t)−Ω(1) .
More precisely, if S(Z; t) is the set of all sums generated by subsets of Z and bounded by t, then
we can compute in time O(n + t log t log3 (n/δ) log n) a set S ⊆ S(Z; t) containing any s ∈ S(Z; t)
1
For a running time T depending on the input size n and possibly more parameters, such as t, we write Õ(T ) as
shorthand for O(T polylog(T )). Similarly, Ω̃(T ) denotes Ω(T /polylog(T )).
2
Abboud et al. [1] presented a reduction from k-Clique to O(k2 )-SUM on O(f (k)n2 ) integers in the range
{1, . . . , O(f (k)(n1+o(1) )k )} for some (explicit) function f . Note that k′ -SUM can be reduced to SubsetSum by
introducing O(log k′ ) leading bits that ensure choosing exactly k′ integers. This yields a reduction from k-Clique to
SubsetSum on O(f ′ (k)n2 ) integers in {1, . . . , O(f ′ (k)(n1+o(1) )k )} for some function f ′ . Let ε > 0, c ∈ R be constants
and let k be sufficiently large (k ≥ 8c/ε). Then for all sufficiently large n ≥ n0 , such that the no(1) -factor is less
than nε/4 , a combinatorial O(t1−ε nc ) algorithm for SubsetSum would yield a combinatorial O(nk−εk/2 ) algorithm
for k-Clique, refuting the combinatorial k-Clique conjecture.
2
with probability at least 1 − δ.
We briefly also consider the UnboundedSubsetSum problem, where each input number may
be used multiple times instead of just once, and present a simple deterministic Õ(n + t) algorithm.
In Section 5 we prove this result and discuss why it is much simpler than our main result.
Theorem 1.2. UnboundedSubsetSum can be solved deterministically in time O(n + t log t).
Polynomial Space A surprising, relatively recent result is that SubsetSum can be solved in
pseudopolynomial time and polynomial space: Lokshtanov and Nederlof [19] presented an algorithm
in time Õ(n3 t) and space Õ(n2 ). In contrast, the O(nt) dynamic programming algorithm as well
as all improvements need space Θ̃(t), which can be much larger than poly(n). We use our new
SubsetSum algorithm and some additional tricks to improve upon their bounds. This almost
answers an open problem by J. Nederlof [14] for a time Õ(nt) and space Õ(n) algorithm.
Theorem 1.3 (Section 4). SubsetSum has a randomized, one-sided error algorithm with error
probability (n + t)−Ω(1) in
• time Õ(n t) and space Õ(n log t) assuming the Extended Riemann Hypothesis (ERH), and
• time Õ(n t1+ε ) and space Õ(n tε ) for any constant ε > 0 unconditionally.
We leave it as an open problem to find an algorithm with time Õ(n + t) and space Õ(n log t).
1.1
Techniques
Our techniques are very different from previous improvements for SubsetSum. In particular, all
previous algorithms are deterministic. Our near-linear time algorithm is simple and elegant and
makes extensive use of randomization and the fast Fourier transform (FFT). In the following we
discuss the main ingredients of our algorithms.
Sumset Computation For sets of integers A, B we define the sumset A⊕ B as the set of all sums
a + b with a ∈ A ∪ {0}, b ∈ B ∪ {0}. Note that here we allow to not choose any value in one or both
sets by adding {0}. Often we consider the t-capped sumset A ⊕t B, which is simply the restriction
of A ⊕ B to {0, . . . , t}. Sumset computation is the most important primitive of our algorithm. Easy
reductions transform this problem to Boolean convolution and further to integer multiplication,
which has well-known algorithms using FFT. This yields an algorithm for computing A ⊕t B in
time O(t log t) (see Section 2 for details).
Color-Coding Color-coding is an algorithmic technique that was first used for the k-Path problem: Given a graph G, decide whether it contains a path of length k [3]. The idea is to randomly
color the vertices of G with k colors, so that for a fixed path of length k in G with probability 1/k!
it is colored (1, 2, . . . , k), in which case we can find it by a simple dynamic programming algorithm
on the layered graph obtained from keeping only the edges in G from color class i to i + 1 (for
each i). Over O(k! log n) repetitions we find a k-path with high probability, if one exists. Research
on color-coding has led to various improvements and derandomizations of this technique [3, 20, 23].
The derandomizations also apply to our use of color-coding in this paper, however, other parts of
our algorithm seem impossible to derandomize and we leave this as an open problem.
We make use of color-coding for finding all sums generated by small subsets. More precisely,
given a SubsetSum instance (Z, t) and a threshold k, we compute a set S ⊆ S(Z; t) containing any
sum generated by a subset Y ⊆ Z of size |Y | ≤ k with constant probability. This can be boosted
3
to any high probability by repeating and taking the union. The main trick is to randomly partition
Z = Z1 ∪ . . . ∪ Zk2 by assigning to any element z ∈ Z a color in {1, . . . , k2 } independently and
uniformly at random. Consider the sumset Z1 ⊕t . . . ⊕t Zk2 . Note that this set consists of sums
generated by subsets of Z, in particular we never add up the same number twice since the Zi are
3
disjoint. We say that the random
P partition splits Y if each Zi contains at most one element of Y .
In this case, the sum Σ(Y ) := y∈Y y is contained in the sumset Z1 ⊕t . . . ⊕t Zk2 . Indeed, by the
definition of ⊕t , in each position i we can choose a number in Zi ∪ {0}, so if Zi contains exactly
one element of Y then we can choose this element, while if Zi contains no element of Y we can
choose 0, to write Σ(Y ) as a sum in the sumset Z1 ⊕t . . .⊕t Zk2 . It remains to argue that the random
partition splits Y with constant probability. The color-coding assigns each element in Y a random
color in {1, . . . , k 2 }, and thus we can view the partitioning, restricted to Y , as placing k balls (the
elements of Y ) into k2 bins (the subsets Zi ). This process is well understood, in particular the
birthday paradox implies for our choice of k 2 bins that with constant probability some Zi contains
more than one element of Y – which is a bad event in our situation. However, since the bound of
the birthday paradox is tight, also with constant probability we have |Zi ∩ Y | ≤ 1 for all i, and thus
the partitioning splits Y . Note that the running time for randomly partitioning and computing the
sumset Z1 ⊕t . . . ⊕t Zk2 is Õ(n + k2 t), which is near-linear in n + t if k is polylogarithmic in n and t.
Layer Splitting Any SubsetSum instance (Z, t) can be partitioned into log n layers Zi ⊆
[t/2i , t/2i−1 ], plus a set Z0 ⊆ [0, t/n] that can be treated very similarly to layers and that we
ignore here for simplicity. We obtain the desired sumset S(Z; t) by combining the sumsets S(Zi ; t)
of the layers in a straight-forward way using sumset computations.
It remains to compute S(Zi ; t). By the definition of layers, any set Y ⊆ Zi summing to at most t
has size |Y | ≤ 2i . Thus, the color-coding algorithm with k = 2i computes S(Zi ; t), and it runs in
time Õ(n + tk 2 ). To obtain an Õ(n + t) algorithm we need to eliminate the factor k2 . We remove
one factor k by observing that all items in Zi are bounded by O(t/k), which allows us to implement
the sumset computations in the color-coding algorithm more efficiently. The remaining factor k
stems from color-coding partitioning Zi into k2 subsets. We remove this factor by a two-stage
approach: In the first stage, we partition Zi into roughly k = 2i sets Zi,j . This k is too small to
split Y entirely, i.e., to have |Zi,j ∩ Y | ≤ 1 for all j with high probability. Indeed, for this property
we would need to partition Zi into k2 sets. However, we still have |Zi,j ∩ Y | ≤ O(log n) with high
probability. Hence, in the second stage we can run the color-coding algorithm with size bound
k′ = O(log n) on each Zi,j , and then combine their computed sumsets in a straight-forward way.
This removes the factor-k overhead from partitioning into k2 sets. Carefully implementing these
ideas yields time Õ(n + t).
Polynomial Space The polynomial space algorithm by Lokshtanov and Nederlof interprets a
SubsetSum algorithm as a circuit, where each gate performs the convolution or pointwise addition
of two vectors of some fixed length f (n, t). This circuit C is transfered to the Fourier domain by
replacing every convolution gate by pointwise multiplication. The new circuit C ′ computes the
Fourier transform of the output vector of C. Using that the inverse Fourier transform can be
written as a simple sum, we can evaluate an entry of C by evaluating all entries of C ′ one-by-one.
Writing g(n, t) for the number of gates in C, since all operations in C ′ are pointwise, computing an
entry of C ′ can be done in Õ(g(n, t)) arithmetic operations, storing Õ(g(n, t)) numbers. Evaluating
all entries of C ′ one-by-one thus can be done in Õ(g(n, t) · f (n, t)) arithmetic operations, storing
Õ(g(n, t)) numbers.
3
In contrast to typical uses of color-coding, we want at most one element instead of exactly one element.
4
One problem in the algorithm by Lokshtanov and Nederlof is that entries can become as large
as 2Ω(n) , so they need to work with Õ(n) bits of precision. This yields total time Õ(n ·g(n, t)f (n, t))
and space Õ(n·g(n, t)). We work with a variant of their algorithm using modular arithmetic instead
of complex numbers. This allows us to work modulo a random prime p, thus reducing the precision
from Õ(n) to O(log p) bits, and improving time and space by a factor Õ(n/ log p). A technical
difficulty is that for using the Fourier transform the field Zp has to contain a primitive t-th root
of unity. This is guaranteed by choosing p in the arithmetic progression 1 + t · N. However, to be
able to choose p at random from a sufficiently large ground set, we need to choose a threshold x
such that there are many primes p ≤ x in the arithmetic progression 1 + t · N. This requires a
quantitative version of Dirichlet’s theorem. The best such result assumes ERH and allows us to
choose p ≤ x = nO(1) , yielding an improvement factor Õ(n/ log p) = Õ(n).
The specific SubsetSum circuit of Lokshtanov and Nederlof needs g(n, t) = O(n) gates and
vector length f (n, t) = Õ(nt). A priori it is not easy to improve f (n, t) to Õ(t). We show that the
circuit induced by our new SubsetSum algorithm allows us to set f (n, t) = Õ(t) and g(n, t) = Õ(n),
thus improving the running time by another factor Õ(n).
1.2
Further Related Work
SubsetSum has been studied extensively, see e.g. [16]. The best-known running time in terms of n
is O∗ (2n/2 ) [13]. There is a large literature improving this running time for inputs that are in some
sense “structured”, see, e.g., [4] and the references therein. Using number-theoretic arguments,
certain dense cases of SubsetSum are solvable in near-linear time, e.g., if t ≪ n2 then there is an
Õ(n) algorithm [11]. A (1 − ε)-approximation for SubsetSum yields a set with sum in [(1 − ε)t, t] if
one exists. The best known approximation algorithm runs in time Õ(min{n/ε, n+1/ε2 }) [18, 12, 15].
2
Preliminaries
All logarithms (log) in this paper are base 2.
P
Sums and Sumsets For a set S of integers we denote the sum of its elements by Σ(S) := s∈S s.
Given a set Z of n positive integers and target t, the SubsetSum problem asks whether there exists
Y ⊆ Z with Σ(Y ) = t. We often solve the more general problem of computing the set of all subset
sums of Z bounded by t, i.e.,
S(Z; t) := {Σ(Y ) | Y ⊆ Z} ∩ {0, . . . , t}.
We represent a set S of integers in {0, . . . , m} by its characteristic vector of length m + 1. For
sets A, B of non-negative integers, we define their sumset, in a slightly non-standard way, as the
set of all sums of at most one element of A and at most one element of B, i.e.,
A ⊕ B = {a + b | a ∈ A ∪ {0}, b ∈ B ∪ {0}}.
For any t > 0, we define the t-capped sumset as A ⊕t B := (A ⊕ B) ∩ {0, . . . , t}.
Sumset Computation For Boolean W
vectors x, y of length t we define their convolution as the
Boolean vector z of length 2t with zi = ij=1 xj ∧ yi−j , where out-of-bounds values are interpreted
as false. Observe that the convolution of the characteristic vectors of integer sets A, B equals the
characteristic vector of the sumset A ⊕ B (if 0 ∈ A, B). Thus, sumset computation can be reduced
to Boolean convolution.
5
A simple algorithm for Boolean convolution is to reduce to integer multiplication: Convert the
Boolean vector x to a number by interpreting true as 1 and false as 0 and padding with log t zeroes
between any two bits of x. Do the same with y. Then from the product of the constructed numbers
we can infer z, since a block of 1 + log t bits is identically 0 if and only if the corresponding bit
of z is 0. Hence, if we can multiply two t-bit numbers in time M (t), then Boolean convolution and
thus sumset computation can be performed in time O(M (t log t)). The best known bound for M (t)
depends on the machine model. For simplicity, in this paper we work on the RAM model with cell
size w = Θ(log t), where typical operations on log(t)-bit numbers can be performed in constant
time. In this model it is known that M (t) = O(t), see, e.g., [10]4 . This yields the following:
Proposition 2.1. Given sets A, B of non-negative integers, the t-capped sumset A ⊕t B can be
computed in time O(t log t).
Preprocessing Multisets So far we assumed that the input Z is a set. Note that it also makes
sense to allow Z to be a multi-set, and define a subset Y ⊆ Z to be a multi-set where each z ∈ Z
has multiplicity in Y at most its multiplicity in Z. Thus, any z ∈ Z with multiplicity m may be
used at most m times in any subset sum. E.L. Lawler [18] showed that the general case of multi-sets
can be reduced to multi-sets with multiplicities bounded by 2. To this end, for any z ∈ Z with
multiplicity 2k + 1 we reduce its multiplicity to 1 and add k times the integer 2z to Z (i.e. increase
the multiplicity of 2z by k). Observe that the resulting set generates the same subset sums, since
we have {i · z | 0 ≤ i ≤ 2k + 1} = {i · z + j · 2z | 0 ≤ i ≤ 1 and 0 ≤ j ≤ k}. Similarly, if z has
multiplicity 2k + 2 we reduce its multiplicity to 2 and add k times 2z. Repeating this step for all
z ∈ Z (in increasing order) yields an equivalent SubsetSum instance where all multiplicities are
bounded by 2, so we obtain the following. For a multi-set Z, by the size |Z| we denote the sum of
all multiplicities of elements in Z.
Proposition 2.2. Given a SubsetSum instance (Z, t), where Z is a multi-set of size n, we can
in time O(n + t) compute an equivalent instance (Z ′ , t) where Z ′ is a multi-set with multiplicities
bounded by 2 and |Z ′ | ≤ min{n, 2t}.
Koiliaris and Xu [17] extended this preprocessing as follows. Write the multi-set Z ′ given by
Proposition 2.2 as a union of two sets Z1 , Z2 ⊆ {0, . . . , t}. Compute S(Z1 ; t) and S(Z2 ; t). Then we
obtain S(Z ′ ; t) as S(Z1 ; t) ⊕t S(Z1 ; t). This yields a reduction to sets, running in time O(n + t log t).
A disadvantage is that we have to compute the whole sets S(Z1 ; t), S(Z2 ; t) instead of just solving
a decision problem, but this is irrelevant for the algorithms presented in this paper.
By running this reduction as a preprocessing of all our algorithms, throughout the paper we
can assume that the input Z ⊆ {0, . . . , t} is a set.
3
Near-Linear Time Algorithm
Our goal is to compute, given a set Z of n positive integers and target t, the set S(Z; t) := {Σ(Y ) |
Y ⊆ Z} ∩ {0, . . . , t}, since checking whether t ∈ S(Z; t) decides the given SubsetSum instance.
In this section, we design a simple algorithm that solves SubsetSum in time Õ(n + t), proving
Theorem 1.1. As discussed in Section 1.1, our algorithm consists of two parts: We first show how
to find sums generated by small subsets using color-coding in Section 3.1, and then use a two-stage
approach on layers for finding all subset sums in Section 3.2.
4
On other machine models M (t) can be larger, e.g., on 2-tape Turing machines or as circuits we have M (t) ≤
∗
t log t · 2O(log t) [9, 8].
6
3.1
Color-Coding: An algorithm for small solution size
We describe an algorithm ColorCoding (see below) for solving SubsetSum if the solution size is
small, i.e., an algorithm that finds all sums Σ(Y ) ≤ t generated by sets Y ⊆ Z of size |Y | ≤ k,
for some given (small) k. We randomly partition Z = Z1 ∪ . . . ∪ Zk2 , i.e., we assign any z ∈ Z
to a set Zi where i is chosen independently and uniformly at random in {1, . . . , k2 }. We say that
this random partition splits Y if |Y ∩ Zi | ≤ 1 holds for all 1 ≤ i ≤ k2 . If this happens, then the
sumset Z1 ⊕t . . . ⊕t Zk2 contains Σ(Y ). Indeed, by definition of ⊕ in each position i we can choose
a number in Zi ∪ {0}, so for |Y ∩ Zi | = 1 we can choose the unique number in the set Y ∩ Zi , while
for |Y ∩ Zi | = 0 we can choose 0, to generate Σ(Y ) as a sum in Z1 ⊕t . . . ⊕t Zk2 . Also note that the
sumset Z1 ⊕t . . . ⊕t Zk2 only contains valid sums in S(Z; t), since no z ∈ Z may be used twice.
Repeating this procedure sufficiently often with fresh randomness, and taking the union over all
computed sumsets Z1 ⊕t . . . ⊕t Zk2 , yields a set S ⊆ S(Z; t) containing any Σ(Y ) ≤ t with Y ⊆ Z
and |Y | ≤ k with probability at least 1 − δ. In fact, O(log 1/δ) repetitions are sufficient, since a
random partition splits Y with constant probability, which follows from the birthday paradox, or
more precisely the tightness of the birthday paradox bound. We briefly prove this standard claim
for completeness. Since Z ⊆ Z1 ⊕t . . . ⊕t Zk2 , we can assume k ≥ 2. For any Y ⊆ Z with |Y | ≤ k,
the probability of the random partition splitting Y is the same as the probability of |Y | balls falling
into |Y | different bins, when throwing |Y | balls into k2 bins. This is equivalent to the second ball
falling into a different bin than the first one, the third ball falling into a different bin than the first
two, and so on, which has probability
2
k − (|Y | − 1) |Y |
1 k 1 2 1
k2 − (|Y | − 1)
k2 − 1 k2 − 2
≥
= .
≥
1
−
·
.
.
.
≥
k2
k2
k2
k2
k
2
4
Hence, r := ⌈log4/3 (1/δ)⌉ repetitions yield the desired success probability of 1 − (1 − 1/4)r ≥ 1 − δ.
This finishes the analysis of ColorCoding and proves the following lemma. For the running time,
note that we perform O(log 1/δ) repetitions of computing k2 sumsets, each taking time O(t log t).
Lemma 3.1. ColorCoding(Z, t, k, δ) computes in time O(tk2 log t log(1/δ)) a set S ⊆ S(Z; t) such
that for any Y ⊆ Z with |Y | ≤ k and Σ(Y ) ≤ t we have Σ(Y ) ∈ S with probability ≥ 1 − δ.
Algorithm 1 ColorCoding(Z, t, k, δ): Given a set Z of positive integers, target t, size bound k ≥ 1
and error probability δ > 0, we solve SubsetSum with solution size at most k
for j = 1, . . . , ⌈log4/3 (1/δ)⌉ do
randomly partition Z = Z1 ∪ . . . ∪ Zk2
3:
Sj := Z1 ⊕t . . . ⊕t Zk2
S
4: return j Sj
1:
2:
Standard techniques allow to derandomize this algorithm by iterating over a (deterministic)
family of partitions of Z that is guaranteed to contain a partition splitting Y . This comes at
the cost of an increased polynomial factor in k, e.g., [20, Lemma 2] gives a factor O(k6 log k).
Alternatively, Koiliaris and Xu provide a very different algorithm running in deterministic time
O(tk 2 log(tk) log n) [17, Lemma 2.12].
3.2
Layer Splitting
Let (Z, t) be a SubsetSum instance and |Z| = n. For ℓ ≥ 1, we call (Z, t) an ℓ-layer instance if
Z ⊆ [t/ℓ, 2t/ℓ]
Z ⊆ [0, 2t/ℓ] and ℓ ≥ n.
or
7
In both cases we have Z ⊆ [0, 2t/ℓ]. Moreover, observe that any Y ⊆ Z summing to at most t has
size |Y | ≤ ℓ. In the second case, this holds since |Y | ≤ |Z| = n. Thus, the ColorCoding algorithm
from the last section with size bound k = ℓ solves ℓ-layer instances. In this section, we will show
that the running time of ColorCoding can be improved for layers, thereby essentially removing the
quadratic dependence on ℓ entirely.
However, before discussing how to solve SubsetSum on layers, we show that any given instance
(Z, t) can be split into O(log n) layers, see Algorithm 2 below. We simply split set Z at t/2i for i =
1, . . . , ⌈log n⌉−1; this yields O(log n) layers Z1 , . . . , Z⌈log n⌉ . On each layer we then run the algorithm
ColorCodingLayer presented below, and we combine the resulting sumsets Si in a straight-forward
way. The error probabilities of the calls to ColorCodingLayer are chosen sufficiently small so
that they sum up to at most δ. Correctness and the running time bound O(t log t log3 (n/δ) log n)
immediate follow from Lemma 3.2 below. This yields algorithm FasterSubsetSum below that
proves Theorem 1.1.
Algorithm 2 FasterSubsetSum(Z, t, δ): Returns a set S ⊆ S(Z; t) containing any s ∈ S(Z; t)
with probability at least 1 − δ, and runs in time O(t log t log3 (n/δ) log n)
1:
2:
3:
4:
5:
6:
split Z into Zi := Z ∩ (t/2i , t/2i−1 ] for i = 1, . . . , ⌈log n⌉ − 1, and Z⌈log n⌉ := Z ∩ [0, t/2⌈log n⌉−1 ]
S=∅
for i = 1, . . . , ⌈log n⌉ do
Si := ColorCodingLayer(Zi , t, 2i , δ/⌈log n⌉)
S := S ⊕t Si
return S
It remains to design a fast algorithm given an ℓ-layer instance (Z, t) and error probability δ.
Let m := ℓ/ log(ℓ/δ) rounded up to the next power of 2. We randomly partition Z into subsets
Z1 , . . . , Zm . For each Zj we run ColorCoding with size bound k = 6 log(ℓ/δ), target 12 log(ℓ/δ)t/ℓ,
and error probability δ/ℓ, yielding a set Sj . We combine the sets S1 , . . . , Sm in a natural, binarytree-like way by computing S1 ⊕ S2 , S3 ⊕ S4 , . . . , Sm−1 ⊕ Sm in the first round, S1 ⊕ S2 ⊕ S3 ⊕
S4 , . . . , Sm−3 ⊕Sm−2 ⊕Sm−1 ⊕Sm in the second round, and so on, until we reach the set S1 ⊕. . .⊕Sm .
Note that in the h-th round of this procedure we combine 2h sets Sj , initially containing integers
bounded by 12 log(ℓ/δ)t/ℓ. Thus, we may use ⊕2h ·12 log(ℓ/δ)t/ℓ in the h-th round. This explains the
following algorithm.
Algorithm 3 ColorCodingLayer(Z, t, ℓ, δ): See Lemma 3.2 for guarantees
1:
2:
3:
4:
5:
6:
if ℓ < log(ℓ/δ) then return ColorCoding(Z, t, ℓ, δ)
m := ℓ/ log(ℓ/δ) rounded up to the next power of 2
randomly partition Z = Z1 ∪ . . . ∪ Zm
γ := 6 log(ℓ/δ)
for j = 1, . . . , m do
Sj := ColorCoding(Zj , 2γt/ℓ, γ, δ/ℓ)
for h = 1, . . . , log m do
for j = 1, . . . , m/2h do
9:
Sj := S2j−1 ⊕2h ·2γt/ℓ S2j
7:
⊲ combine the sumsets Sj in a binary-tree-like way
8:
10:
return S1 ∩ {0, . . . , t}
8
Lemma 3.2. For an ℓ-layer instance (Z, t) and δ ∈ (0, 1/4], the method ColorCodingLayer(Z, t, ℓ, δ)
computes in time O(t log t log3 (ℓ/δ)) a set S ⊆ S(Z; t) containing any s ∈ S(Z; t) with probability
at least 1 − δ.
Proof. The case ℓ < log(ℓ/δ) follows from Lemma 3.1. The inclusion S ⊆ S(Z; t) also follows from
Lemma 3.1
and since we only compute sumsets over partitionings, using the fact S(Z1 ; t1 ) ⊕t′
S(Z2 ; t2 ) ∩ {0, . . . , t} ⊆ S(Z; t) for a partitioning Z = Z1 ∪ Z2 and any t, t′ , t1 , t2 ≥ 1.
For the error probability, fix a subset Y ⊆ Z with Σ(Y ) ≤ t, and let Yj := Y ∩ Zj for 1 ≤ j ≤ m.
A crucial property is that with sufficiently high probability the size |Yj | is smaller than 6 log(ℓ/δ),
thus allowing us to run ColorCoding with size bound k = 6 log(ℓ/δ).
Claim 3.3. We have Pr[|Yj | ≥ 6 log(ℓ/δ)] ≤ δ/ℓ.
Proof. Note that |Yj | is distributed as the sum of |Y | independent Bernoulli random variables with
success probability 1/m. In particular, µ := E[|Yj |] = |Y |/m. A standard Chernoff bound yields
that Pr[|Yj | ≥ λ] ≤ 2−λ for any λ ≥ 2eµ. Recall that (Z, t) is an ℓ-layer instance, and thus Y has
size at most ℓ. This allows us to bound µ = |Y |/m ≤ ℓ/m ≤ log(ℓ/δ), by definition of m. The
concentration inequality thus holds for λ = 6 log(ℓ/δ), and we obtain Pr[|Yj | ≥ 6 log(ℓ/δ)] ≤ δ/ℓ.
By the above claim, we may assume that |Yj | ≤ 6 log(ℓ/δ) holds for each 1 ≤ j ≤ m; this happens
with probability at least 1−m·δ/ℓ. Since Z ⊆ [0, 2t/ℓ], any subset of Zj of size at most 6 log(ℓ/δ) has
sum at most 12 log(ℓ/δ)t/ℓ. It follows that the call ColorCoding(Zj , 12 log(ℓ/δ)·t/ℓ, 6 log(ℓ/δ), δ/ℓ)
finds Σ(Yj ) with probability at least 1 − δ/ℓ. Assume that this event holds for each 1 ≤ j ≤ m;
this happens with probability at least 1 − m · δ/ℓ. Then Sj contains Σ(Yj ), and the tree-like sumset
computation indeed yields a set containing Σ(Y1 )+. . .+Σ(Ym ) = Σ(Y ). The total error probability
is 2mδ/ℓ. Since ℓ ≥ 1 and δ ≤ 1/4 we have log(ℓ/δ) ≥ 2 and obtain m ≤ ℓ/2. Hence, the total
error probability is bounded by δ.
Regarding the running time, observe that by Lemma 3.1 each call to ColorCoding takes time
O(t/ℓ · log4 (ℓ/δ) log t). Since there are m = Θ(ℓ/ log(ℓ/δ)) calls to ColorCoding, we obtain the
claimed time O(t log3 (ℓ/δ) log t). The remaining time for combining the sets S1 , . . . , Sm is
O
log
Xm
h=1
m h
· 2 log(ℓ/δ)t/ℓ · log t
2h
= O(t log t log m),
which is dominated by the total time for calling ColorCoding.
4
4.1
Polynomial Space Algorithm
Setup
We first give the setup of Lokshtanov and Nederlof [19] and then present a version of their main
result using modular arithmetic instead of complex numbers.
Circuits For a set F and binary operators O1 , O2 on F , a circuit C over (F ; O1 , O2 ) is a directed
acyclic graph D = (N, A) with parallel arcs, such that every node of D is either a constant gate
(indegree 0), O1 gate (indegree 2), or O2 gate (indegree 2). For an indegree 2 node, its two inneighbours are its children. The output of a constant gate is the element it is labeled with. The
output of an Oi gate is the result of performing Oi on the output of its two children. The output
9
of C, denoted by out(C), is the output of a specific gate c marked as the output gate. The size
of C is the size of the underlying graph (number of vertices plus number of edges). With some
abuse of notation we will denote a gate of C and the output of that gate by the same symbol. A
probabilistic circuit is a distribution C over circuits over (F ; O1 , O2 ).
Vector Operations Fix a ring R with operations +, ·. For a vector a ∈ Rt we write a[i],
0 ≤ i < t, for its i-th entry. We call a a singleton if it has at most one non-zero entry. We
write ⊞ for pointwise addition and ⊡ for pointwise multiplication, i.e., for vectors a, b ∈ Rt we
have (a ⊞ b)[i]
P = a[i] + b[i] and (a ⊡ b)[i] = a[i] · b[i]. We write ⊠ for convolution, i.e., we have
(a ⊠ b)[i] = ij=0 a[j] · b[i − j]. The length of a vector a is the maximal index of any non-zero
entry of a, i.e., len(a) := max{0 ≤ i < t | a[i] 6= 0} or 0, if a is the all-zeroes-vector. We say that
the convolution a ⊠ b overflows if len(a) + len(b) ≥ t. If a ⊠ b does not overflow and a, b are the
coefficients of polynomials P (x), Q(x), respectively, then a⊠b are the coefficients of the polynomial
P (x) · Q(x). Throughout the paper we will always work withP
non-overflowing convolutions. Finally,
we also consider the sum of all entries of a, i.e., sum(a) := t−1
i=0 a[i].
Lemma 4.1. Let a1 , . . . , ak ∈ Rt . We have
(1) len(a1 ⊞ . . . ⊞ ak ) = maxi len(ai ),
P
(2) sum(a1 ⊞ . . . ⊞ ak ) = i sum(ai ),
(3) if there is no overflow then len(a1 ⊠ . . . ⊠ ak ) =
P
(4) if there is no overflow then sum(a1 ⊠ . . . ⊠ ak ) =
i
Q
len(ai ), and
i sum(ai ).
Proof. (1), (2), and (3) are immediate. For (4), note that any product a1 [j1 ] · . . . · ak [jk ], with
0 ≤ ji ≤ len(ai ), contributes to exactly one entry of b := a1 ⊠ . . . ⊠ ak , and each entry of b can be
written as a sum of such products.
Discrete Fourer Transform Fix a prime p and t ≥ 1. Let ω ∈ Zp be a primitive t-th root of
unity, i.e., ω t = 1 and ω k 6= 1 for 1 ≤ k < t. The t-point Discrete Fourier Transform (DFT) over
modular arithmetic in Zp is the linear function F mapping a vector a ∈ Ztp to F(a) ∈ Ztp with
F(a)[i] =
t−1
X
ω ij a[j].
j=0
Pt−1
i
In other words, if a is the vector of coefficients of the polynomial P (x) =
i=0 a[i]x then F
evaluates P (x) at all powers of ω, i.e., F(a) = (P (1), P (ω), P (ω 2 ), . . . , P (ω t−1 )). The inverse
function F −1 is given by
t−1
1 X −ij
−1
ω a[i].
F (a)[j] =
t
i=0
P
−ij ω ik =
That F −1 is indeed the inverse operation of F is a consequence of the identity t−1
i=0 ω
t · [j = k], which follows from ω being a primitive t-th root of unity. Here and in the remainder
we use Iverson’s bracket notation, i.e., [B] is 1 if B is true, and 0 otherwise. The most important
property of DFT for our purposes is the Convolution Theorem (see, e.g., [6]), stating that for any
a, b ∈ Ztp such that a ⊠ b does not overflow we have
F(a ⊠ b) = F(a) ⊡ F(b).
10
Modular Variant of Lokshtanov-Nederlof Lokshtanov and Nederlof [19] mention without
proof that a version of their main result holds for modular arithmetic. For completeness, we prove
such a result here.
Theorem 4.2. Let p be prime, t ≥ 1, and suppose that Ztp contains a t-th root of unity ω. Let
C be a circuit over (Ztp , ⊞, ⊠) with only singleton constants. Suppose that no convolution gate
overflows. Then given p, t, ω, and 0 ≤ x < t we can compute out(C)[x] in time Õ(|C|t log p) and
space O(|C| log p).
We discuss how to find an appropriate root of unity ω later in Lemma 4.5. We remark that
Lokshtanov and Nederlof have the depth of C as an additional factor in their time and space bounds,
which seems to be due to their choice of the complex Fourier transform and does not appear in the
modular version (this is a lower-order improvement).
Proof. The proof by Lokshtanov and Nederlof works almost verbatim. From the circuit C we
construct another circuit C ′ over (Ztp , ⊞, ⊡) with the same directed graph as C, but with different
gates. Each constant gate a ∈ C is replaced with the constant gate a′ = F(a). Each convolution
gate ⊠ is replaced with a pointwise multiplication gate ⊡. Then an easy inductive argument shows
that for any gate a ∈ C and its corresponding gate a′ ∈ C ′ we have a′ = F(a). Indeed, if a is a
constant gate then the claim holds by definition. For the inductive step, consider let b, c be the
children of a. By the inductive hypothesis we have b′ = F(b) and c′ = F(c). If a is an addition
gate, then we conclude a′ = b′ ⊞ c′ = F(b) ⊞ F(c) = F(b ⊞ c) = F(a). If a is a convolution gate,
then the Convolution Theorem implies a′ = b′ ⊡ c′ = F(b) ⊡ F(c) = F(b ⊠ c) = F(a).
For ease of notation we write f := out(C). We now use C ′ to compute f [x]. Plugging the
definition of F −1 into the identity f = F −1 (F(f )) yields
t−1
f [x] =
1 X −ix
ω (F(f ))[i].
t
i=0
Note that ω s , for s ≤ poly(t), can be computed with O(log t) arithmetic operations in Zp by repeated
squaring. Thus, in order to compute f [x] it suffices to compute (F(f ))[0], . . . , (F(f ))[t − 1].
For computing (F(f ))[j], we convert circuit C ′ to a circuit Cj′ over (Zp , +, ·) with the same
directed graph as C ′ . Each constant gate a′ ∈ C ′ is replaced by a′ [j]. Each pointwise addition
gate ⊞ is replaced with + and each pointwise multiplication gate ⊡ is replaced with ·. It follows
immediately that Cj′ computes (F(f ))[j].
Consider a (singleton) constant gate a ∈ C with a[k] = v and a[ℓ] = 0 for all ℓ 6= k. By
definition of F we have (F(a))[j] = ω jk v. Thus, any constant gate in Cj′ can be computed with
O(log t) arithmetic operations in Zp . Computing the remaining gates of Cj′ takes |C ′ | = |C|
arithmetic operations in Zp . As arithmetic operations in Zp can be performed in time Õ(log p), we
can compute any value (F(a))[j] in time Õ(|C| log p log t) and space O(|C| log p). Summing over
all j = 0, . . . , t − 1 yields time Õ(|C|t log p log t) = Õ(|C|t log p) and space O(|C| log p).
4.2
FasterSubsetSum as a Circuit
In the remainder of the paper, set the error probability δ := 1/n. We now describe how to convert
the algorithm FasterSubsetSum into a probabilistic circuit over (NÕ(t) ; ⊞, ⊠). This yields:
Lemma 4.3. Given a SubsetSum instance (Z, t) with |Z| = n, we can construct a probabilistic
′
circuit C over (Nt ; ⊞, ⊠), where t′ = O(t log3 n) is a power of two, satisfying
11
(1) for any 0 ≤ s ≤ t, s 6∈ S(Z; t) we have PrC∈C [out(C)[s] = 0] = 1, and
(2) for any 0 ≤ s ≤ t, s ∈ S(Z; t) we have PrC∈C [out(C)[s] = 0] ≤ 1/n.
Any circuit C ∈ C has size |C| = Õ(n), uses only singleton constants, has no overflowing convolution gates, and any gate a ∈ C satisfies sum(a) ≤ 2Õ(n) . Sampling a circuit C from C can be
performed in time Õ(n).
Proof. We carefully inspect the algorithm FasterSubsetSum, transforming any ∪-operation into a
⊞-gate and any ⊕t -operation into a ⊠-gate. The details are as follows.
′
Set t′ := 300 log 3 (2n2 )t, rounded up to a power of two. For any vector a ∈ Nt we say that a
represents the set {0 ≤ i < t′ | a[i] > 0}. Note that if a, b represent sets A, B then a ⊞ b represents
A ∪ B. Moreover, if a ⊠ b does not overflow then a ⊠ b represents A ⊕t′ B, which in this case equals
A ⊕ B. For any 0 ≤ i < t′ we denote by ei the singleton vector containing a 1 at position i and a
0 at all other entries. Note that ei represents {i}.
We now use this representation to convert algorithm FasterSubsetSum to a probabilistic circuit
′
over (Nt ; ⊞, ⊠). Fix the randomness of FasterSubsetSum, i.e., fix all random partitions performed
in ColorCoding and ColorCodingLayer. All remaining basic operations in FasterSubsetSum,
ColorCoding, and ColorCodingLayer are (1) forming a subset Z ′ ⊆ Z, (2) taking the union of two
sets, and (3) computing the sumset of two sets. Hence, we can view FasterSubsetSum as a circuit
where each node computes a set. Each leaf is directly assigned a subset Z ′ ⊆ Z. Each inner node
either computes the union ∪ or the sumset ⊕t′′ , for some t′′ , of its children.
Now we replace each leaf, corresponding to Z ′ ⊆ Z, by the circuit computing ⊞z∈Z ′ ez , which
is a vector representing Z ′ . Note that this circuit has size |Z ′ | and uses only singleton constants.
Moreover, we replace each ∪-gate by ⊞ and each ⊕t′′ -gate by ⊠. This yields a circuit C over
′
(Nt ; ⊞, ⊠). Over the randomness of FasterSubsetSum used for the random partitionings, we
obtain a probabilistic circuit C.
We show that t′ is chosen sufficiently large so that circuit C has no overflowing convolution gates.
To this end, we use Lemma 4.1 to bound the length len(a) = max{0 ≤ i < t′ | a[i] 6= 0} of any gate a.
For the output of ColorCoding(Z, t, k, δ), as well as any of its inner gates, we can bound the length
by k2 · max Z, since we compute a union (which does not increase the length) over a k2 -wise sumset
computation (multiplying the length by k2 ) over subsets of Z (having length at most max Z). For
ColorCodingLayer(Z, t, ℓ, δ) we can bound the length by ℓ · 36 log 2 (ℓ/δ) max Z, since we compute
an m-wise sumset, for some m ≤ ℓ, over the output of ColorCoding with k = 6 log(ℓ/δ). Finally,
for FasterSubsetSum(Z, t, δ) we perform a sumset computation over layers i = 1, . . . , ⌈log n⌉. For
the i-th layer we run ColorCodingLayer on ℓ = 2i and Zi with max Zi ≤ t/2i−1 , which by the
above bound yields length at most 2i · 36 log 2 (2i /δ)t/2i−1 ≤ 72 log 2 (2n/δ)t. Summing over all i =
1, . . . , ⌈log n⌉ yields the desired bound on the length of ⌈log n⌉ · 72 log2 (2n/δ)t ≤ 144 log 3 (2n/δ)t <
t′ /2, since we set δ = 1/n. This bound holds at all gates of any circuit C ∈ C. In particular, this
shows that no convolution gate overflows, since t′ is chosen sufficiently large.
With this property it is immediate that C computes a vector out(C) representing a set S ′
such that for S := S ′ ∩ {0, . . . , t} we have FasterSubsetSum(Z, t, 1/n) ⊆ S ⊆ S(Z; t). Indeed,
by the same reasoning as for FasterSubsetSum we argue that the resulting circuit only performs
set partitionings and sumset computations and thus only computes valid subset sums, proving
S ′ ⊆ S(Z; t′ ) and thus S ⊆ S(Z; t). For the other direction, note that we drop the index from a
⊕t′′ -gate when converting it to a ⊠-gate. Thus, the circuit C does not compute the same set S as
FasterSubsetSum(Z, t, 1/n), but S may be a proper superset. However, this change cannot remove
elements, so we obtain the inclusion FasterSubsetSum(Z, t, 1/n) ⊆ S.
12
From this, correctness follows immediately. Indeed, for s ∈ {0, . . . , t}, s 6∈ S(Z; t) we obtain
s 6∈ S and thus out(C)[s] = 0. Also, for s ∈ S(Z; t) we have s ∈ FasterSubsetSum(Z, t, 1/n) with
probability at least 1 − 1/n and thus s ∈ S with probability at least 1 − 1/n.
Since we almost always simply partition the current set Z, except for ColorCoding where we
take a union over O(log(n)) partitionings, each item in Z appears in O(log(n)) leafs of the circuit.
It is also easy to see that the depth of the circuit is O(log(n)). This allows us to bound |C| ≤ Õ(n),
and it is easy to see that C ∈ C can be sampled in thePsame time bound.
t′ −1
Finally, we bound the sum of entries sum(a) = i=0
a[i] for any gate a ∈ C with C ∈ C,
using Lemma 4.1.(4). For a subset Z ′ ⊂ Z we implement a circuit representing Z ′ with sum of
2
entries equal to |Z ′ | ≤ n. For ColorCoding(Z, t, k, δ) we have sum of entries O(log(1/δ)nk ), since
we take the union over O(log(1/δ)) rounds (multiplying the sum of entries by O(log(1/δ))) over
k2 -wise sumset computations (raising the sum of entries to the k2 -th power) over subsets of Z,
which have sum of entries at most n. For ColorCodingLayer(Z, t, ℓ, δ) we can bound the sum of
2
entries by nO(log (ℓ/δ)·ℓ) , since we compute an m-wise sumset, for some m ≤ ℓ, over the output of
ColorCoding with k = O(log(ℓ/δ)). Finally, for FasterSubsetSum(Z, t, 1/n) the sum of entries
2
4
is bounded by nO(log n·log (n)·n) = 2O(log (n)·n) = 2Õ(n) , since we perform a sumset computation
over O(log n) layers (which raises the sum of entries to the O(log n)-th power) over the result of
ColorCodingLayer on ℓ ≤ 2n.
Note that the bound sum(a) ≤ 2Õ(n) also bounds the maximal entry of any gate a in any
circuit C ∈ C. Thus, we may replace N by Zp for some p ≤ 2Õ(n) (with sufficiently many hidden
logarithmic factors) and still obtain the same result. This yields a probabilistic circuit C ′ over
′
(Ztp , ⊞, ⊠). Running Theorem 4.2 on C ′ would yield a randomized SubsetSum algorithm with
time Õ(nt log p) = Õ(n2 t) and space Õ(n log p) = Õ(n2 ), if we were able to efficiently compute the
required root of unity (which is a non-trivial problem).
In the next section, we further reduce time and space and we provide an algorithm for computing
an appropriate root of unity.
4.3
Reducing the Domain Size
′
In the last section, we designed a probabilistic circuit C over (Nt ; ⊞, ⊠) and then replaced N by
Zp for sufficiently large p = 2Õ(n) . Now we will pick p as a random prime, chosen uniformly from
a set P of size Ω̃(n2 ). This randomness is independent from the randomness of picking a circuit
C ∈ C. Since any number k ≤ 2Õ(n) has at most Õ(n) prime factors, for any C ∈ C the probability
′
that p divides out(C) is at most Õ(1/n). Interpreting a circuit C ∈ C as a circuit over (Ztp ; ⊞, ⊠),
′
i.e., performing all arithmetic operations of C modulo p, yields a circuit Cp over (Ztp ; ⊞, ⊠) that
computes the value out(C) mod p. Taking into account the randomness of C ∈ C, we obtain a
′
probabilistic circuit Cp over (Ztp ; ⊞, ⊠) with the following properties (see Lemma 4.3):
(1) for any 0 ≤ s ≤ t, s 6∈ S(Z; t) we have Prp∈P [PrCp ∈Cp [out(Cp )[s] = 0]] = 1, and
(2) for any 0 ≤ s ≤ t, s ∈ S(Z; t) we have Prp∈P [PrCp ∈Cp [out(Cp )[s] = 0]] ≤ Õ(1/n).
Note that Theorem 4.2 and Lemma 4.3 now yield total time Õ(|C|t log(max P )) = Õ(nt log(max P ))
and space Õ(|C| log(max P )) = Õ(n log(max P )) for solving SubsetSum.
It remains to choose an appropriate set of primes P . From the above discussion we have the
following requirements, where we need (1) for running DFT, see Theorem 4.2, (2) to avoid Õ(n)
prime factors, and (3) since max P appears in the running time.
13
Lemma 4.4. For any power of two τ , and n ≤ τ , there is a set P of primes such that
(1) Zp contains a τ -th root of unity ωp for any p ∈ P ,
(2) |P | = Ω̃(n2 ), and
(3) assuming ERH we have max P = τ O(1) , while unconditionally we have max P ≤ O(exp(τ ε ))
for any ε > 0 (with the polylog(n) in |P | = Ω̃(n2 ) depending on ε).
We can pick a random p ∈ P and compute ωp in randomized time (log(max P ))O(1) , with error
probability 1/poly(τ ).
Recall from Theorem 4.2 that the time and space bounds incur a factor O(log p). Assuming ERH
we have O(log p) = O(log τ ) = O(log t), which yields the desired time Õ(nt) and space Õ(n log t),
proving the first statement of Theorem 1.3. The unconditional statement follows similarly.
Proof. We first simplify requirement (1). For any prime p, by Fermat’s little theorem any a ∈ Zp ,
a 6= 0, is a (p−1)-th root of unity. The additional property ak 6= 1 for all 1 ≤ k < p−1 of a primitive
(p − 1)-th root of unity is satisfied for any generator g of the multiplicative group Z∗p = Zp \ {0}.
Finally, if τ divides p−1 and g is a primitive (p−1)-th root of unity then g(p−1)/τ is a primitive τ -th
root of unity. Note that if τ divides p − 1 then p is in the arithmetic progression {1 + k · τ | k ∈ N},
which we write as 1 + τ · N. Hence, we many replace requirement (1) by
(1’) P ⊆ 1 + τ · N.
The above argument shows existence of a primitive τ -th root of unity, but it does not yield
an efficient algorithm for determining one, since there is no efficient algorithm known for finding
a generator of Z∗p (the known algorithms have to compute a prime factorization of p − 1). We
circumvent this problem by using that τ is a power of two. In this situation, we can compute a
τ -th root of unity as follows.
Lemma 4.5. Let p be prime and τ be a power of two, where τ divides p − 1. Pick a ∈ Z∗p uniformly
at random and set ω := a(p−1)/τ . If ω τ 6= 1 or ω τ /2 = 1 then restart this procedure. This method
finds a primitive τ -th root of unity ω ∈ Zp in expected time O(polylog(p)).
Note that by halting after O(polylog(p)) iterations we may obtain an algorithm with worst-case
running time O(polylog(p)) and error probability 1/poly(p) ≤ 1/poly(t).
Proof. First observe that if the method finishes then ω is a primitive τ -th root of unity. Indeed, ω
is a τ -th root of unity since ω τ = 1. For showing that ω is primitive, consider the minimal ℓ ≥ 1
with ω ℓ = 1. From ω τ = 1 it follows that ℓ divides τ , and thus ℓ is a power of two. In particular, if
ℓ < τ then ω τ /2 = ω ℓ·τ /(2ℓ) = 1τ /(2ℓ) = 1, contradicting the break condition ω τ /2 6= 1. Hence, ℓ = τ ,
and for any 1 ≤ k < τ we have ω k 6= 1, showing that ω is primitive.
Since arithmetic operations in Zp can be performed in time O(polylog(p)), it remains to bound
the success probability of one round of the method by Ω(1/polylog(p)). To this end, it suffices to
show that a random a ∈ Z∗p is likely to be a generator of Z∗p , since then a is a primitive (p − 1)-th
root of unity and ω = a(p−1)/τ is a primtive τ -th root of unity. For bounding the probability of a
being a generator, we combine the well-known fact that the number of generators is φ(p − 1), where
φ is Euler’s totient function, and the asymptotic lower bound φ(n) = Ω(n/ log log n).
14
It remains to find a set P ⊆ 1 + τ · N of Ω̃(n2 ) primes. Dirichlet’s theorem shows that the
arithmetic progression 1 + τ · N contains infinitely many primes. However, in order to bound max P
we need a quantitative version of Dirichlet’s theorem. In particular, let π(x, τ ) be the number of
primes in (1 + τ · N) ∩ [0, x]. We need an upper bound on the minimal x such that π(x, τ ) = Ω̃(n2 ).
The best unconditional bound is given by (a well-known corollary of) the Siegel-Walfisz theorem [24]: For any constant C > 0 there is a constant C ′ > 0 such that if τ ≤ (log x)C then
p
π(x, τ ) = Li(x)/φ(τ ) ± O(x exp(−C ′ log x)),
where Li denotes the offset logarithmic integral. Using Li(x) = Θ(x/ log x) and φ(τ
√) ≤ τ ≤
C
′
(log x) , we obtain Li(x)/φ(τ ) = Ω̃(x), which dominates the error term O(x exp(−C log x)) for
any sufficiently large x. Hence, we obtain π(x, τ ) = Ω̃(x) for τ ≤ (log x)C . Setting C = 1/ε and
rearranging, we have π(x, τ ) = Ω̃(x) for x ≥ exp(τ ε ).
Hence, for the threshold x := n2 + exp(τ ε ) = O(exp(τ ε )) we have π(x, τ ) = Ω̃(n2 ). We now
set P as the set of primes in R := (1 + τ · N) ∩ [0, x]. Clearly, we have |P | = Ω̃(n2 ). Moreover,
note that the set R contains (x − 1)/τ ≤ x numbers, Ω̃(x) of which are prime. Thus, a random
number in R is prime with probability Ω(1/polylog(x)) = τ −O(ε) . Now to pick a random p ∈ P ,
we repeatedly pick a random number r ∈ R until it is prime. Recall that checking primality can be
done in polynomial time O(polylog(x)) = τ O(ε) [2]. The expected number of repetitions until we
find a prime is also τ O(ε) . By halting after essentially the same number of repetitions we can also
obtain a worst-case time bound with sufficiently small error probability, say 1/τ . The total time
for finding a random prime p ∈ P is thus τ O(ε) . This procedure has to be performed only once, so
this time is negligible compared to the remaining running time Õ(nt). Note that the number of
log-factors in the bound |P | = Ω̃(n2 ) depends on ε, and since the constants in the Siegel-Walfisz
theorem are ineffective, we cannot give an explicit bound on the number of log-factors in terms
of ε.
Assuming the Extended Riemann Hypothesis, very good asymptotic bounds on π(x, τ ) are
known. Using a bound by Titchmarsh, in [8, Proposition 4.3] it is shown, along the lines of the
last two paragraphs, that already for threshold x := Õ(τ (τ + n2 )) (with sufficiently many hidden
logarithmic factors) the set R contains Ω̃(n2 ) primes, and sampling random elements of R yields a
prime in expected time O(polylog(x)) = O(polylog(τ )). This finishes the proof.
5
Unbounded Subset Sum
We briefly discuss the unbounded variant of SubsetSum, where each input number can be chosen
arbitrarily often (whereas in the bounded variant that we studied so far each input number can be
chosen at most once). Specifically, given a set Z of n positive integers and target t, the task is to
determine whether any sequence over Z sums to exactly t. Note that here it makes no sense to
define Z as a multi-set and thus n ≤ t. We present a simple Õ(t) algorithm.
First consider the problem whether any sequence over Z of length at most k sums to t. Clearly,
this problem can be solved in time Õ(kt), by computing the k-fold sumset Z ⊕t Z ⊕t . . . ⊕t Z,
and checking whether it contains t. Note that for the usual bounded variant of SubsetSum this
property breaks down, since already the sumset Z ⊕t Z is not necessarily contained in S(Z; t), as
it contains sums of the form z + z for z ∈ Z, so we have to resort to sumsets Z1 ⊕t . . . ⊕t Zk over
partitionings Z = Z1 ∪ . . . ∪ Zk . This is a reason why (bounded) SubsetSum is much harder than
the unbounded version UnboundedSubsetSum.
Algorithm 4 solves the UnboundedSubsetSum problem for each t′ ≤ t at once, i.e., it computes
unb
S (Z; t) := {0, . . . , t} ∩ {a1 + . . . + ak | k ∈ N≥0 , a1 , . . . , ak ∈ Z}. We use the fact that the classic
dynamic programming algorithm computes S unb (Z; t) in time O(nt).
15
Algorithm 4 UnboundedSubsetSum in time Õ(t). Given n positive integers Z and target t,
the algorithm computes for each t′ ≤ t whether a sequence over Z sums to t′ .
compute S0 := S unb (Z; t/n) using the classic dynamic program in time O(n · t/n) = O(t)
2: for i = 1, . . . , ⌈log n⌉ do
3:
ti := 2i t/n
4:
Si := Si−1 ⊕ti Si−1 ⊕ti Z
1:
5:
return S⌈log n⌉ ∩ {0, . . . , t}
Since the i-th iteration takes time Õ(ti ) = Õ(2i t/n), the total running time of Algorithm 4 is
Õ(t), more precisely O(t log t). For correctness, we argue inductively that Si = S unb (Z; 2i t/n), and
thus S⌈log n⌉ ∩ {0, . . . , t} = S unb (Z; t). Consider any sequence A over Z summing to an integer in
[0, 2i t/n]. If Σ(A) ≤ 2i−1 t/n, then
Σ(A) ∈S unb (Z; 2i−1 t/n) = Si−1 ⊆ Si−1 ⊕ti Si−1 ⊕ti Z = Si .
Thus, assume Σ(A) > 2i−1 t/n. We split A = (a1 , . . . , ak ) at the smallest index j with a1 + . . . +
aj > 2i−1 t/n into A1 = (a1 , . . . , aj−1 ), aj , and A2 = (aj+1 , . . . , ak ). Observing that aj ∈ Z and
Σ(A1 ), Σ(A2 ) ≤ 2i−1 t/n and hence Σ(A1 ), Σ(A2 ) ∈ Si−1 , we obtain Σ(A) ∈ Si−1 ⊕ti Si−1 ⊕ti Z = Si .
Hence, S unb (Z; 2i t/n) ⊆ Si . Correctness follows, since clearly the converse Si ⊆ S unb (Z; 2i t/n)
holds as well.
6
Conclusion
A textbook algorithm solves SubsetSum in pseudopolynomial time O(nt) on inputs consisting of
n positive numbers with target t. As our main result we present an improved algorithm running
in time Õ(n + t). This improves upon a classic algorithm and is likely to be near-optimal, since
it matches conditional lower bounds of t1−ε for any ε > 0 from SetCover and combinatorial
k-Clique. Our algorithm heavily uses randomization and has one-sided error. The main tricks are
to find small solutions using color-coding, and to split large solutions by a two-stage color-codinglike process.
We also improve the best known polynomial space algorithm for SubsetSum to time Õ(nt) and
space Õ(n) assuming the Extended Riemann Hypothesis, and to time Õ(n t1+ε ) and space Õ(n tε )
for any ε > 0 unconditionally. This improvement is achieved by combining our Õ(n + t) algorithm
with the previously best polynomial space algorithm by Lokshtanov and Nederlof [19], and by
working modulo a random prime in an arithmetic progression, which explains the connection to
ERH. We leave it as an open problem to obtain an algorithm with time Õ(n + t) and space Õ(n).
√
Our techniques are very different from the recent Õ( n t) algorithm by Koiliaris and Xu [17],
which is the fastest known deterministic algorithm.
Acknowledgements The author wants to thank Marvin Künnemann and Jesper Nederlof for
providing useful comments on a draft of this paper.
References
[1] A. Abboud, K. Lewi, and R. Williams. Losing weight by gaining edges. In ESA’14, pages
1–12, 2014.
16
[2] M. Agrawal, N. Kayal, and N. Saxena. PRIMES is in P. Annals of Mathematics, pages
781–793, 2004.
[3] N. Alon, R. Yuster, and U. Zwick. Color-coding. Journal of the ACM, 42(4):844–856, 1995.
[4] P. Austrin, P. Kaski, M. Koivisto, and J. Nederlof. Subset sum in the absence of concentration.
In STACS’15, pages 48–61, 2015.
[5] R.E. Bellman. Dynamic programming. Princeton University Press, 1957.
[6] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction
to Algorithms, Third Edition. The MIT Press, 3rd edition, 2009.
[7] M. Cygan, H. Dell, D. Lokshtanov, D. Marx, J. Nederlof, Y. Okamoto, R. Paturi, S. Saurabh,
and M. Wahlström. On problems as hard as CNF-SAT. ACM Transactions on Algorithms,
12(3):41, 2016.
[8] A. De, P. P. Kurur, C. Saha, and R. Saptharishi. Fast integer multiplication using modular
arithmetic. SIAM Journal on Computing, 42(2):685–699, 2013.
[9] M. Fürer. Faster integer multiplication. SIAM Journal on Computing, 39(3):979–1005, 2009.
[10] M. Fürer. How fast can we multiply large integers on an actual computer? In LATIN’14,
pages 660–670, 2014.
[11] Z. Galil and O. Margalit. An almost linear-time algorithm for the dense subset-sum problem.
SIAM Journal on Computing, 20(6):1157–1189, 1991.
[12] G. V. Gens and E. V. Levner. Fast approximation algorithms for knapsack type problems. In
Optimization Techniques, pages 185–194. Springer, 1980.
[13] E. Horowitz and S. Sahni. Computing partitions with applications to the knapsack problem.
Journal of the ACM, 21(2):277–292, 1974.
[14] T. Husfeldt, R. Paturi, G.B. Sorkin, and R. Williams. Exponential Algorithms: Algorithms and
Complexity Beyond Polynomial Time (Dagstuhl Seminar 13331). Dagstuhl Reports, 3(8):40–
72, 2013.
[15] H. Kellerer, R. Mansini, U. Pferschy, and M. G. Speranza. An efficient fully polynomial
approximation scheme for the subset-sum problem. Journal of Computer and System Sciences,
66(2):349–370, 2003.
[16] H. Kellerer, U. Pferschy, and D. Pisinger. Knapsack problems. Springer, 2004.
[17] K. Koiliaris and C. Xu. A faster pseudopolynomial time algorithm for subset sum. In SODA’17,
2017. To appear. Preprint available at arxiv.org/abs/1507.02318.
[18] E. L. Lawler. Fast approximation algorithms for knapsack problems. Mathematics of Operations Research, 4(4):339–356, 1979.
[19] D. Lokshtanov and J. Nederlof. Saving space by algebraization. In STOC’10, pages 321–330,
2010.
[20] M. Naor, L. J. Schulman, and A. Srinivasan. Splitters and near-optimal derandomization. In
FOCS’95, pages 182–191, 1995.
17
[21] D. Pisinger. Linear time algorithms for knapsack problems with bounded weights. Journal of
Algorithms, 33(1):1–14, 1999.
[22] D. Pisinger. Dynamic programming on the word RAM. Algorithmica, 35(2):128–145, 2003.
[23] J. P. Schmidt and A. Siegel. The spatial complexity of oblivious k-probe hash functions. SIAM
Journal on Computing, 19(5):775–786, 1990.
[24] A. Walfisz. Zur additiven Zahlentheorie. II. Mathematische Zeitschrift, 40(1):592–607, 1936.
18
| 8cs.DS
|
Clique-Based Lower Bounds for
Parsing Tree-Adjoining Grammars
arXiv:1803.00804v1 [cs.CC] 2 Mar 2018
Karl Bringmann∗
Philip Wellnitz†
Abstract
Tree-adjoining grammars are a generalization of context-free grammars that are well
suited to model human languages and are thus popular in computational linguistics. In
the tree-adjoining grammar recognition problem, given a grammar Γ and a string s of
length n, the task is to decide whether s can be obtained from Γ. Rajasekaran and
Yooseph’s parser (JCSS’98) solves this problem in time O(n2ω ), where ω < 2.373 is the
matrix multiplication exponent. The best algorithms avoiding fast matrix multiplication
take time O(n6 ).
The first evidence for hardness was given by Satta (J. Comp. Linguist.’94): For
a more general parsing problem, any algorithm that avoids fast matrix multiplication
and is significantly faster than O(|Γ| n6 ) in the case of |Γ| = Θ(n12 ) would imply a
breakthrough for Boolean matrix multiplication.
Following an approach by Abboud et al. (FOCS’15) for context-free grammar recognition, in this paper we resolve many of the disadvantages of the previous lower bound.
We show that, even on constant-size grammars, any improvement on Rajasekaran and
Yooseph’s parser would imply a breakthrough for the k-Clique problem. This establishes tree-adjoining grammar parsing as a practically relevant problem with the unusual
running time of n2ω , up to lower order factors.
1
Introduction
Introduced in [14, 15], tree-adjoining grammars (TAGs) are a system to manipulate certain
trees to arrive at strings, see Section 2 for a definition. TAGs are more powerful than contextfree grammars, capturing various phenomena of human languages which require more formal
power; in particular TAGs have an “extended domain of locality” as they allow “long-distance
dependencies” [16]. These properties, and the fact that TAGs are efficiently parsable [29],
make them highly desirable in the field of computer linguistics. This is illustrated by the large
literature on variants of TAGs (see, e.g., [9, 21, 24, 30]), their formal language properties
(see, e.g., [16, 29]), as well as practical applications (see, e.g., [2, 13, 25, 26]), including
the XTAG project which developed a tree-adjoining grammar for (a large fraction of) the
English language [10]. In fact, TAGs are so fundamental to computer linguistics that there
is a biannual meeting called “International Workshop on Tree-Adjoining Grammars and
Related Formalisms” [7], and they are part of their undergraduate curriculum (at least at
Saarland University).
∗
Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany,
[email protected]
†
Saarland
University,
Saarland
Informatics
Campus,
Saarbrücken,
Germany,
[email protected]
1
The prime algorithmic problems on TAGs are parsing and recognition. In the recognition
problem, given a TAG Γ and a string s of length n, the task is to decide whether Γ can
generate s. The parsing problem is an extended variant where in case Γ can generate s
we should also output a sequence of derivations generating s. The first TAG parsers ran
in time1 O(n6 ) [23, 29], which was improved by Rajasekaran and Yooseph [20] to O(n2ω ),
where ω < 2.373 is the exponent of (Boolean) matrix multiplication.
A limited explanation for the complexity of TAG parsing was given by Satta [22], who
designed a reduction from Boolean matrix multiplication to TAG parsing, showing that
any TAG parser running faster than O(|Γ| n6 ) on grammars of size |Γ| = Θ(n12 ) yields a
Boolean matrix multiplication algorithm running faster than O(n3 ). This result has several
shortcomings: (1) It holds only for a more general parsing problem, where we need to
determine for each substring of the given string s whether it can be generated from Γ. (2)
It gives a matching lower bound only in the unusual case of |Γ| = Θ(n12 ), so that it cannot
exclude time, e.g., O(|Γ|2 n4 ). (3) It gives matching bounds only restricted to combinatorial
algorithms, i.e., algorithms that avoid fast matrix multiplication2 . Thus, so far there is no
satisfying explanation of the complexity of TAG parsing.
Context-free grammars The classic problem of parsing context-free grammars, which
has important applications in programming languages, was in a very similar situation as TAG
parsing until very recently. Parsers in time O(n3 ) were known since the 60s [8, 11, 17, 31].
In a breakthrough, Valiant [27] improved this to O(nω ). Finally, a reduction from Boolean
matrix multiplication due to Lee [18] showed a matching lower bound for combinatorial
algorithms for a more general parsing problem in the case that the grammar size is Θ(n6 ).
Abboud et al. [1] gave the first satisfying explanation for the complexity of context-free
parsing, by designing a reduction from the classic k-Clique problem, which asks whether
there are k pairwise adjacent vertices in a given graph G. For this problem, for any fixed k
the trivial running time of O(nk ) can be improved to O(nωk/3 ) for any k divisible by 3 [19]
(see [12] for the case of k not divisible by 3). The fastest combinatorial algorithm runs in
time O(nk / logk n) [28]. The k-Clique hypothesis states that both running times are essentially optimal, specifically that k-Clique has no O(n(ω/3−ε)k ) algorithm and no combinatorial
O(n(1−ε)k ) algorithm for any k ≥ 3, ε > 0. The main result of Abboud et al. [1] is a reduction
from the k-Clique problem to context-free grammar recognition on a specific, constant-size
grammar Γ, showing that any O(nω−ε ) algorithm or any combinatorial O(n3−ε ) algorithm
for context-free grammar recognition would break the k-Clique hypothesis. This matching conditional lower bound removes all disadvantages of Lee’s lower bound at the cost of
introducing the k-Clique hypothesis, see [1] for further discussions.
Our contribution We extend the approach by Abboud et al. to the more complex setting
of TAGs. Specifically, we design a reduction from the 6k-Clique problem to TAG recognition:
Theorem 1. There is a tree-adjoining grammar Γ of constant size such that if we can decide
in time T (n) whether a given string
of length n can be generated from Γ, then 6k-Clique can
be solved in time O T (nk+1 log n) , for any fixed k ≥ 1. This reduction is combinatorial.
1
In most running time bounds we ignore the dependence on the grammar size, as we are mostly interested
in constant-size grammars in this paper.
2
The notion of “combinatorial algorithms” is informal, intuitively meaning that we forbid unpractical
algorithms such as fast matrix multiplication. It is an open research problem to find a reasonable formal
definition.
2
Via this reduction, any O(n2ω−ε ) algorithm for TAG recognition would prove that
6k-Clique is in time Õ(n(2ω−ε)(k+1) ) = O(n(ω/3−ε/9)6k ), for sufficiently large3 k. Furthermore, any combinatorial O(n6−ε ) algorithm for TAG recognition would yield a combinatorial
algorithm for 6k-Clique in time Õ(n(6−ε)(k+1) ) = O(n(1−ε/9)6k ), for sufficiently large k. As
both implications would violate the 6k-Clique conjecture, we obtain tight conditional lower
bounds for TAG recognition. As our result (1) works directly for TAG recognition instead of
a more general parsing problem, (2) holds for constant size grammars, and (3) does not need
the restriction to combinatorial algorithms, it overcomes all shortcomings of the previous
lower bound based on Boolean matrix multiplication, at the cost of using the well-established
k-Clique hypothesis, which has also been used in [1, 3, 4, 5, 6].
We thus establish TAG parsing as a practically relevant problem with the quite unusual
running time of n2ω , up to lower order factors. This is surprising, as the authors are aware of
only one other problem with a (conjectured or conditional) optimal running time of n2ω±o(1) ,
namely 6-Clique.
Techniques The essential difference of tree-adjoining and context-free grammars is that
the former can grow strings at four positions, see Figure 3a. Writing a vertex v1 in one
position of the string, and writing the neighborhoods of vertices v2 , v3 , v4 at other positions
in the string, a simple tree-adjoining grammar can test whether v1 is adjacent to v2 , v3 ,
and v4 . Extending this construction, for k-cliques C1 , C2 , C3 , C4 we can test whether C1 ∪
C2 , C1 ∪ C3 , and C1 ∪ C4 form 2k-cliques. Using two permutations of this test, we ensure
that C1 ∪ C2 ∪ C3 ∪ C4 forms an almost-4k-clique, i.e., only the edges C3 × C4 might be
missing (in Figure 2b below this situation is depicted for cliques C2 , C5 , C1 , C6 instead of
C1 , C2 , C3 , C4 ). Finally, we use that a 6k-clique can be decomposed into 3 almost-4k-cliques,
see Figure 2a.
In the constructed string we essentially just enumerate 6 times all k-cliques of the given
graph G, as well as their neighborhoods, with appropriate padding symbols (see Section 3).
We try to make the constructed tree-adjoining grammar as easily accessible as possible
by defining a certain programming language realized by these grammars, and phrasing our
grammar in this language, which yields subroutines with an intuitive meaning (see Section 4).
2
Preliminaries on tree-adjoining grammars
In this section we define tree-adjoining grammars and give examples. Fix a set T of terminals
and a set N of non-terminals. In the following, conceptually we partition the nodes of any
tree into its leaves, the root, and the remaining inner nodes. An initial tree is a rooted tree
where
• the root and each inner node is labeled with a non-terminal,
• each leaf is labeled with a terminal, and
• each inner node can be marked for adjunction.
See Figure 1a for an example; nodes marked for adjunction are annotated by a rectangle.
An auxiliary tree is a rooted tree where
• the root and each inner node is labeled with a non-terminal,
3
For this and the next statement it suffices to set k > 18/ε.
3
• exactly one leaf, called the foot node, is labeled with the same non-terminal as the root,
• each remaining leaf is labeled with a terminal, and
• each inner node can be marked for adjunction.
Initial trees are the starting points for derivations of the tree-adjoining grammar. These
trees are then extended by repeatedly replacing nodes marked for adjunction by auxiliary
trees. Formally, given an initial or auxiliary tree t that contains at least one inner node v
marked for adjunction and given an auxiliary tree a whose root r has the same label as v,
we can combine these trees with the following operation called adjunction, see Figure 1 for
an example.
1. Replace a’s foot node by the subtree rooted at v.
2. Replace the node v with the tree obtained from the last step.
Note that these steps make sense, since r and v have the same label. Note that adjunction
does not change the number leaves labeled with a non-terminal symbol, i.e., an initial tree
will stay an initial tree and an auxiliary tree will stay an auxiliary tree.
A
B
A
a B c
B
A
B
a
a
b
a
A
B
A
B b c
c
A
B b c
b
(a) An initial tree (left) and an auxiliary tree
(right); the internal nodes labeled A and B are
marked for adjunction.
(b) Resulting tree after adjoining the auxiliary
tree into the initial tree.
Figure 1: The basic building blocks and operation of tree-adjoining grammars.
A tree-adjoining grammar is now defined as a tuple Γ = (I, A, T, N ) where
• I is a finite set of initial trees and
• A is a finite set of auxiliary trees,
using the same terminals T and non-terminals N as labels. The set D of derived trees of Γ
consists of all trees that can be generated by starting with an initial tree in I and repeatedly
adjoining auxiliary trees in A. (Note that each derived tree is also an initial tree, but not
necessarily in I.) Finally, a string s over alphabet T can be generated by Γ, if there is a
derived tree t in D such that
• t contains no nodes marked for adjunction and
• s is obtained by concatenating the labels of the leaves of t from left to right.
The language L(Γ) is then the set of all strings that can be generated by Γ.
4
3
Encoding a graph in a string
Given a graph G = (V, E), in this section we construct a string GGk (G) (the graph gadget )
that encodes its k-cliques, over the terminal alphabet T = {0, 1, $, #, |, §, e, l1 , . . . , l6 , r1 , . . . , r6 }
of size 19. In the next section we then design a tree-adjoining grammar Γ that generates
GGk (G) if and only if G contains a 6k-clique. We assume that V = {1, . . . , |V |}, and we
denote the binary representation of any v ∈ V by v and the neighborhood of v by N (v). For
two strings a and b, we use a ◦ b to denote their concatenation and aR to denote the reverse
of a.
We start with node and list gadgets, encoding a vertex and its neighborhood, respectively:
NG(v) := $ v $ and LG(v) :=
NG(u) =
u∈N (v)
$u$
u∈N (v)
Note that u and v are adjacent iff NG(u) is a substring of LG(v).
Next, we build clique versions of these gadgets, that encode a k-clique C and its neighborhood, respectively:
k
k
CNG(C) :=
(# NG(v) #) and CLG(C) :=
# LG(v) #
v∈C
v∈C
Note that two k-cliques C and C ′ form a 2k-clique if and only if the substring of CNG(C)
between the i-th and (i+ 1)-th symbol # is a substring of the substring of CLG(C ′ ) between
the i-th and (i + 1)-th symbol #, for all i. Indeed, every pair of a vertex in C and a vertex
in C ′ is tested for adjacency.
Conceptually, we split any 6k-clique into six k-cliques. Thus, let Ck be the set of all
k-cliques in G. Our final encoding of the graph is:
GGk (G) :=
C∈Ck
◦
C∈Ck
◦
C∈Ck
◦e
◦
C∈Ck
◦
C∈Ck
◦
C∈Ck
| CNG(C) § CLG(C)R l1 r1 CLG(C) § CLG(C)R |
| CNG(C) § CLG(C)R l2 r2 CLG(C) § CLG(C)R |
| CNG(C) § CLG(C)R l3 r3 CLG(C) § CLG(C)R |
| CLG(C) § CLG(C)R l4 r4 CNG(C) § CLG(C)R |
| CLG(C) § CLG(C)R l5 r5 CNG(C) § CLG(C)R |
| CLG(C) § CLG(C)R l6 r6 CNG(C) § CLG(C)R |
As we will show, there is a tree-adjoining grammar of constant size that generates the string
GGk (G) iff G contains a 6k-clique. The structure of this test is depicted in Figure 2. The
clique-gadgets of the same highlighting style together allow us to test for an almost-4k-clique,
as it is depicted in Figure 2a. The two gadgets of the same highlighting style then test for
two claws of cliques, as depicted in Figure 2b.
As the graph has n nodes, for any node u the node and list gadgets NG(u), LG(u) have a
length of O(n log n), and for a k-clique C the clique neighborhood gadgets CNG(C), CLG(C)
thus have a length of O(k2 n log n). As our encoding of the graph consists of O(nk ) clique
neighborhood gadgets, the resulting string length is O(k2 nk+1 log n) = O(nk+1 log n). It is
easy to see that it is also possible to construct all gadgets and in particular the encoding of
a graph in linear time with respect to their length.
5
C2
C2
C1
C1
C3
C6
C4
C6
C5
C5
(b) We will generate an almost-4k-clique as in
(a) by generating two claws. (This tests the
edges (C1 , C6 ), (C2 , C5 ), and (C3 , C4 ) in (a)
twice.)
(a) Each Ci is a k-clique and there is an edge between two k-cliques of some highlighting style
if the clique gadgets of that style ensure that
these two cliques together form a 2k-clique.
Figure 2: Structure of our test for 6k-cliques.
4
Programming with trees
It remains to design a clique-detecting tree-adjoining grammar. To make our reduction more
accessible, we will think of tree-adjoining grammars as a certain programming language. In
the end, we will then present a “program” that generates (a suitable superset of) the set all
strings that represent a graph containing a 6k-clique. We start by defining programs.
A normal tree N with input NIn and output NOut is an auxiliary tree where:
• the root is labeled with NIn ,
• exactly one node is marked for adjunction, and
• this node lies on the path from the root to the foot node and is labeled NOut .
See Figure 3a for an illustration. The special structure of a normal tree N allows us to split
its nodes into four categories (excluding the path from N ’s root to its foot node): subtrees
of left children of the path from N ’s root to NOut , subtrees of left children of the path from
NOut to N ’s foot node, subtrees of right children of the path from NOut to N ’s foot node,
and the remaining nodes (i.e., subtrees of right children of the path from N ’s root to NOut ).
The concatenation of all terminal symbols in N ’s leaves from left to right can then be split
into four parts n1 , n2 , n3 , n4 where each part contains symbols from exactly one category.
We say that the normal tree N generates the tuple (n1 , n2 , n3 , n4 ).
Lemma 1. Given normal trees N with input NIn , output NOut and M with input MIn =
NOut , output MOut , the derived tree N · M obtained by adjoining M into N is a normal tree
with input NIn and output MOut . Further, if N and M generate the tuples (n1 , n2 , n3 , n4 )
and (m1 , m2 , m3 , m4 ), then N · M generates the tuple (n1 ◦ m1 , m2 ◦ n2 , n3 ◦ m3 , m4 ◦ n4 ).
Proof. See Figure 3.
We now define a program P with input PIn and output POut as a set of normal trees that
contains a tree with input PIn and a tree with output POut . Note that all trees derived by
starting with a tree in P and repeatedly adjoining trees from P are normal, by Lemma 1. An
6
NIn
NIn
NOut
NOut
MOut
NIn
NOut
n1 n2
n3 n4
NIn
n 1 m1 m2 n 2
(a) A normal tree N .
n 3 m3 m4 n 4
(b) The tree resulting after adjoining M into N .
Figure 3: Adjoining normal trees.
execution of the program P is a derived tree of P with input PIn and output POut . Further,
the set computed by P, denoted by L(P), is the set of all tuples generated by P’s executions.
We will later use programs as subroutines of tree-adjoining grammars. Let N (P) be the
set of non-terminals of P. Formally, we say that P is a subroutine of a grammar Γ if
• the set of trees P is a subset of the auxiliary trees of Γ, and
• no remaining auxiliary tree of Γ has a root label in N (P) \ {POut }.
These restrictions ensure that any “call” to the program P terminates at POut . Indeed, consider any sequence of adjunctions in Γ ending in a tree without nodes marked for adjunction.
If this sequence contains an adjunction of a node labeled PIn , meaning that program P is
called, then this adjunction must be followed by an execution of P, i.e., it must generate a
derived tree of P with output POut . Indeed, any derived tree of P is normal and thus contains
exactly one node marked for adjunction. To get rid of this node, we have to adjoin some
auxiliary tree, but the remaining auxiliary trees can only adjoin to POut . We will frequently
make use of this observation that ensures coherence of programs.
We now show how to perform two programs sequentially one after another. To avoid
interference, we ensure that the two programs have disjoint non-terminals, except for their
input and output. In particular, we will model two sequential calls to the same program by
creating two copies of the program.
Lemma 2 (Combining programs). For programs P and Q, let Q′ denote the program obtained
from Q by replacing each non-terminal by a fresh copy, ensuring that P and Q′ have disjoint
non-terminals. Further, let Q′′ denote the program obtained from Q′ by replacing Q′ In by
POut . Then P · Q := P ∪ Q′′ is a program computing the set
L(P · Q) := {(a ◦ a′ , b′ ◦ b, c ◦ c′ , d′ ◦ d) | (a, b, c, d) ∈ L(P), (a′ , b′ , c′ , d′ ) ∈ L(Q)}.
Proof. As every execution of P and Q” is a normal tree, the claim follows from Lemma 1.
We can think of · as an operator on programs; the above lemma shows that it is associative.
4.1
Basic programs
We now present some easy programs that will later be used as subroutines.
7
Writing characters We start by demonstrating a program that writes exactly one character to each of the four positions. Formally, given a 4-tuple of characters (a, b, c, d), let the
program W(a, b, c, d) be defined by the following auxiliary tree:
W(a, b, c, d)In
a
W(a, b, c, d) Out
d
b W(a, b, c, d) In c
Clearly, this tree is normal with input W(a, b, c, d)In and output W(a, b, c, d) Out , so that
W(a, b, c, d) is a program. The tree itself is an execution of the program, and it is the only
execution. Thus, this program computes the set L(W(a, b, c, d)) = {(a, b, c, d)}. We write
W(a) to denote the program W(a, a, a, a).
Testing equality We give a program that tests equality of four strings, by writing the
same arbitrary string to all four positions. Formally, for any terminal alphabet Σ, let the
program Eq(Σ) be defined by the following set of |Σ| + 1 auxiliary trees:
Eq(Σ)In
Eq(Σ)In
σ
Eq(Σ)Out
Eq(Σ)In
σ
∀σ ∈ Σ
σ Eq(Σ)In σ
Eq(Σ)In
A simple induction shows that L(Eq(Σ)) = {(v, v R , v, v R ) | v ∈ Σ∗ }.
Writing anything We will need to write appropriate strings surrounding some carefully
constructed substrings. As it turns out, being able to write anything will be sufficient; this
is achieved by the following program. Given an alphabet Σ, let the program A(Σ) be defined
by the following set of 4|Σ| + 1 trees:
A(Σ)In
A(Σ)In
A(Σ)Out
A(Σ)In σ
A(Σ)In
A(Σ)In
A(Σ)In
A(Σ)In
A(Σ)In
σ A(Σ)In
A(Σ)In
A(Σ)In
A(Σ)In
σ A(Σ)In
∀σ ∈ Σ
A(Σ)In σ
As this program allows writing anything, it is easy to see that A(Σ) computes the set
L(A(Σ)) = {(v1 , v2 , v3 , v4 ) | v1 , v2 , v3 , v4 ∈ Σ∗ } = (Σ∗ )4 .
4.2
Detecting Cliques
With the help of the above programs, we now design programs that detect a 6k-clique.
8
Detecting claws
Our next program can detect whether four nodes form a claw graph.
NC := W(#) · A({0, 1, $}) · W($) · Eq({0, 1}) · W($) · A({0, 1, $}) · W(#)
Lemma 3. For any nodes v1 , v2 , v3 , v4 , the program NC generates the tuple
(a, b, c, d) := (# NG(v1 ) #, # LG(v2 )R #, # LG(v3 ) #, # LG(v4 )R #)
and any of its cyclic rotations (i.e., (b, c, d, a), (c, d, a, b), and (d, a, b, c)) if and only if v1 is
adjacent to each one of v2 , v3 , and v4 .
Proof. By Lemma 2 and the properties of basic programs, we see that NC computes all
tuples of the form
(# α1 $ v $ α2 #, # α3 $ v R $ α4 #, # α5 $ v $ α6 #, # α7 $ v R $ α8 #)
where v ∈ {0, 1}∗ and α1 , . . . , α8 ∈ {0, 1, $}∗ . From the construction of node and list gadgets
we see that all tuples (#NG(v1 )#, #LG(v2 )R #, #LG(v3 )#, #LG(v4 )R #) are of this form.
For the other direction, for any generated tuple (a, b, c, d), where a is # $ v $ #, it holds
that $ v $ is a substring of c and its reverse is a substring of b and d. Hence, NG(v1 ) is
a substring of LG(v2 ), LG(v3 ), and LG(v4 ). This implies that v1 is adjacent to v2 , v3 , and
v4 .
Detecting claws of cliques We now extend NC to a program that can detect claws of
k-cliques, see Figure 2b. We define the program CC by the following set of 3 trees (additional
to the trees of NC):
CCIn
NCOut
NCOut
NCOut
NCIn
CCOut
CCIn
NCOut
NCOut
Each execution of CC starts with the first tree, then repeatedly adjoins the second tree
followed by some execution of NC, and finally adjoins the last tree. As the number of
repetitions is arbitrary, the program CC can perform any number of sequential calls to NC.4
Lemma 4. For any k-cliques C1 , C2 , C3 , C4 in G, the program CC generates the tuple
(a, b, c, d) := (CNG(C1 ), CLG(C2 )R , CLG(C3 ), CLG(C4 )R ) and all of its cyclic rotations
(i.e., (b, c, d, a), (c, d, a, b), and (d, a, b, c)) if and only if C1 ∪ C2 , C1 ∪ C3 , and C1 ∪ C4 each
form a 2k-clique in G.
Proof. For any nodes vij , with i ∈ [4], j ∈ [m], m ≥ 1, set
ni :=
# N G(vij ) #
and
ℓi :=
# LG(vij ) #.
j∈[m]
j∈[m]
As program CC can perform any number of calls to NC, and by Lemma 3, program CC
generates the tuple (n1 , ℓ2 , ℓ3 , ℓ4 ) if and only if v1j is adjacent to v2j , v3j , and v4j for all j.
4
Actually, we already know how many calls to NC we want to perform, namely k2 . However, encoding
this number into the grammar would result in a grammar size depending on k, which we want to avoid.
9
Observe that for any k-cliques C1 = {v1 , . . . , vk }, C2 = {u1 , . . . , uk }, both CNG(C) and
CLG(C) can be split into k2 blocks by splitting between two consecutive #-characters:
CNG(C1 ) = # NG(v1 ) ## NG(v1 ) # · · · # NG(v1 ) ## NG(v2 ) # · · ·
CLG(C2 ) = # LG(u1 ) ## LG(u2 ) # · · · # LG(uk ) ## LG(u1 ) # · · ·
This layout is chosen so that each node vi in C1 is paired up with each node uj in C2 exactly
once. The claim follows from these two insights.
Detecting almost-4k-cliques We now use CC twice to test for two claws, thus detecting
“almost-4k-cliques”, as depicted in Figure 2b:
C := CC · W(§) · CC.
Lemmas 4 and 2 directly imply the following, see Figure 2b.
Lemma 5. For any k-cliques Ca , Cb , Cc , Cd the program C generates the tuple
(CNG(Ca ) § CLG(Ca )R , CLG(Cb ) § CLG(Cb )R , CLG(Cc ) § CLG(Cc )R , CNG(Cd ) § CLG(Cd )R )
if and only if Ca ∪ Cb ∪ Cd and Ca ∪ Cc ∪ Cd both form a 3k-clique. A similar statement
holds if we pick any two other positions in the tuple for the CNG(·) gadgets.
Detecting 6k-cliques As in Figure 2a, we now want to test for three almost-4k-cliques
to detect a 6k-clique. Recall that T = {0, 1, $, #, |, §, e, l1 , . . . , l6 , r1 , . . . , r6 } is the terminal
alphabet that we constructed our strings over. The following programs will generate the
highlighted groups in Figure 2a:
P(1, 3, 4, 6) := A(T ) · W(|) · C · W(l1 , r3 , l4 , r6 )
P(1, 2, 5, 6) := W(r1 , l2 , r5 , l6 ) · C · W(|) · A(T ),
P(2, 3, 4, 5) := W(r2 , l3 , r4 , l5 ) · C · W(|) · A(T ),
We now deviate from our notion of normal trees by explicitly not marking P(1, 2, 5, 6) Out and
P(2, 3, 4, 5) Out for adjunction. Our final tree-adjoining grammar Γ consists of the following
initial and auxiliary trees (as well as all auxiliary trees used by its subroutines):
P(1, 3, 4, 6) Out
S
P(1, 2, 5, 6) In
P(1, 3, 4, 6)In
P(2, 3, 4, 5) In
P(1, 3, 4, 6) Out
e
Note that the latter tree is the only one in Γ that has more than one node marked for
adjunction, so it needs special treatment in the analysis.
Lemma 6. For any graph G, the grammar Γ generates the encoding GGk (G) if and only if
G contains a 6k-clique. Moreover, Γ has constant size (independent of k).
10
S
P(1, 3, 4, 6) In
P(1, 3, 4, 6) Out
P(1, 2, 5, 6) In
P(1, 2, 5, 6) Out
P(1, 2, 5, 6) In
P(2, 3, 4, 5) In
P(2, 3, 4, 5) Out
P(2, 3, 4, 5) In
P(1, 3, 4, 6) Out
P(1, 3, 4, 6) In
. . . | (C1 ) l1 r1 (C1 ) | . . . . . . | (C2 ) l2 r2 (C2 ) | . . . . . . | (C3 ) l3 r3 (C3 ) | . . .
e
. . . | (C4 ) l4 r4 (C4 ) | . . .. . . | (C5 ) l5 r5 (C5 ) | . . . . . . | (C6 ) l5 r6 (C6 ) | . . .
Figure 4: Global structure of a parsing of GGk (G) by Γ. (Clique gadgets are abbreviated.)
Proof. First, assume that Γ can generate GGk (G). Then there is a derived tree whose leaves,
if read from left to right, yield GGk (G). All derivations of Γ start with the single initial
tree, and then adjoin an execution of the program P(1,3,4,6) into it. (As P(1,3,4,6) is a
subroutine, only a full execution can be adjoined.) This execution generates some tuple of
strings ( x1 , x2 , x3 , x4 ) and leaves exactly the node labeled P(1, 3, 4, 6) Out as the sole node
marked for adjunction. Therefore, in the next step the auxiliary tree rooted with that node
will be adjoined, which in turn leaves exactly the nodes P(1, 2, 5, 6)In and P(2, 3, 4, 5) In
as nodes marked for adjunction. Again, these are input nodes of subroutines, therefore
at both nodes one (complete) execution of the corresponding programs must be adjoined.
The program execution of program P(1, 2, 5, 6) generates a tuple of strings ( y1 , y2 , y3 , y4 ),
and the execution of P(2, 3, 4, 5) generates ( z1 , z2 , z3 , z4 ). The grammar Γ ensures that
these tuples will be placed in the order ( x1 , y1 , y2 , z1 , z2 , x2 , x3 , z3 , z4 , y3 , y4 , x4 ), see
Figure 4 for a visualization. At this point, no more adjunctions are possible, since we
explicitly forced P(1, 2, 5, 6) Out and P(2, 3, 4, 5)Out not to be marked for adjunction. (Also
note that this structure is the only possibility to obtain a tree containing no more nodes
marked for adjunction.) Hence, GGk (G) can be partitioned as:
GGk (G) = x1 ◦ y1 ◦ y2 ◦ z1 ◦ z2 ◦ x2 ◦ x3 ◦ z3 ◦ z4 ◦ y3 ◦ y4 ◦ x4 .
Consider the strings x1 and y1 . By the definitions of P(1,3,4,6) and P(1,2,5,6), and
Lemma 2, we know that x1 must end with the terminal symbol l1 and that y1 must start
with the symbol r1 . Whenever l1 r1 occurs in GGk (G), it does so in the string
| CNG(C1 ) § CLG(C1 )R l1 r1 CLG(C1 ) § CLG(C1 )R |,
for some k-clique C1 . Since x1 ◦ y1 is a substring of GGk (G), and the program C cannot produce a |-terminal, but the W(|) part of P(·, ·, ·, ·) will always write such a |-character, x1 must
have | CNG(C1 ) § CLG(C1 )R l1 as a suffix and y1 must have r1 CLG(C1 ) § CLG(C1 )R |
as a prefix. This also means that the program C must generate the string between | and l1
in x1 and between r1 and | in y1 .
11
Similar statements hold for the other ten strings. In total we obtain that the program
C generates the following tuples for some k-cliques C1 , . . . , C6 :
• t1 := ( CNG(C1 ) § CLG(C1 )R , CLG(C3 ) § CLG(C3 )R ,
CLG(C4 ) § CLG(C4 )R , CNG(C6 ) § CLG(C6 )R ) in P(1,3,4,6),
• t2 := ( CLG(C1 ) § CLG(C1 )R , CNG(C2 ) § CLG(C2 )R ,
CNG(C5 ) § CLG(C5 )R , CLG(C6 ) § CLG(C6 )R ) in P(1,2,5,6), and
• t3 := ( CLG(C2 ) § CLG(C2 )R , CNG(C3 ) § CLG(C3 )R ,
CNG(C4 ) § CLG(C4 )R , CLG(C5 ) § CLG(C5 )R ) in P(2,3,4,5).
By Lemma 5, this implies that all Ci ∪ Cj form a 2k-clique and thus C1 ∪ . . . ∪ C6 forms
a 6k-clique (see Figure 2a to check that all pairs are covered).
For the other direction, consider a graph G that contains a 6k-clique C ∗ . Then we can
split C ∗ into 6 vertex-disjoint k-cliques C1 , . . . , C6 . Further we know that every three of
these six k-cliques together form a 3k-clique. Thus, the program C generates the tuples
t1 , t2 , t3 as above. We can then use the three programs P(·, ·, ·, ·) to generate such tuples
surrounded with symbols |, li , and ri at appropriate positions. Adding the surrounding
strings by A(T ) and following the global structure of Γ generates the encoding GGk (G).
To see that Γ is of constant size, note that we only use constantly many programs. Thus
using a new set of terminal symbols for every instance of a program will still yield a constant
total number of non-terminal symbols. Further, we only use 19 terminal symbols.
The above lemma and the bound |GGk (G)| = O(nk+1 log n) imply the main theorem.
References
[1] A. Abboud, A. Backurs, and V. V. Williams. If the current clique algorithms are
optimal, so is Valiant’s parser. In 56th Annual Symposium on Foundations of Computer
Science, FOCS’15, pages 98–117, 2015.
[2] A. Abeillé. Parsing french with tree adjoining grammar: some linguistic accounts. In
12th Conference on Computational Linguistics, COLING’88, pages 7–12, 1988.
[3] A. Backurs, N. Dikkala, and C. Tzamos. Tight hardness results for maximum weight
rectangles. In 43rd International Colloquium on Automata, Languages, and Programming, ICALP’16, volume 55, pages 81:1–81:13, 2016.
[4] A. Backurs and C. Tzamos. Improving Viterbi is hard: Better runtimes imply faster
clique algorithms. In 34th International Conference on Machine Learning, ICML’17,
pages 311–321, 2017.
[5] K. Bringmann, A. Grønlund, and K. G. Larsen. A dichotomy for regular expression
membership testing. In 58th Annual IEEE Symposium on Foundations of Computer
Science, FOCS’17, pages 307–318, 2017.
[6] Y. Chang. Hardness of RNA folding problem with four symbols. In 27th Annual
Symposium on Combinatorial Pattern Matching, CPM’16, pages 13:1–13:12, 2016.
12
[7] D. Chiang and A. Koller, editors. Proc. 12th International Workshop on Tree Adjoining
Grammars and Related Formalisms (TAG+12), June 29 - July 1, 2016, Heinrich Heine
University, Düsseldorf, Germany. The Association for Computer Linguistics, 2016.
[8] J. Cocke and J. T. Schwartz. Programming languages and their compilers: Preliminary
notes. Technical report, CIMS, NYU, 1970.
[9] V. Demberg, F. Keller, and A. Koller. Incremental, predictive parsing with psycholinguistically motivated tree-adjoining grammar. Computational Linguistics, 39(4):1025–
1066, 2013.
[10] C. Doran, D. Egedi, B. A. Hockey, B. Srinivas, and M. Zaidel. XTAG system: a
wide coverage grammar for English. In 15th Conference on Computational Linguistics,
COLING’94, pages 922–928, 1994.
[11] J. Earley. An efficient context-free parsing algorithm. C. ACM, 13(2):94–102, 1970.
[12] F. Eisenbrand and F. Grandoni. On the complexity of fixed parameter clique and
dominating set. Theoretical Computer Science, 326(1-3):57–67, 2004.
[13] K. Forbes, E. Miltsakaki, R. Prasad, A. Sarkar, A. Joshi, and B. Webber. D-LTAG
system: Discourse parsing with a lexicalized tree-adjoining grammar. Journal of Logic,
Language and Information, 12(3):261–279, 2003.
[14] A. K. Joshi. Tree adjoining grammars: How much context-sensitivity is required to provide reasonable structural descriptions? In Natural Language Processing, Theoretical,
Computational and Psychological Perspectives. Cambridge University Press, 1985.
[15] A. K. Joshi, L. S. Levy, and M. Takahashi. Tree adjunct grammars. JCSS, 10(1):136–
163, 1975.
[16] A. K. Joshi and Y. Schabes. Tree-adjoining grammars. In Handbook of Formal Languages, pages 69–123. Springer, 1997.
[17] T. Kasami. An efficient recognition and syntax algorithm for context-free languages.
Technical report, AFCRL-65-758, Air Force Cambridge Research Lab, Bedford, MA.,
1965.
[18] L. Lee. Fast context-free grammar parsing requires fast boolean matrix multiplication.
JACM, 49(1):1–15, 2002.
[19] J. Nešetřil and S. Poljak. On the complexity of the subgraph problem. Commentationes
Mathematicae Universitatis Carolinae, 26(2):415–419, 1985.
[20] S. Rajasekaran and S. Yooseph. TAL recognition in O(M (N 2 )) time. JCSS, 56(1):83–
89, 1998.
[21] P. Resnik. Probabilistic tree-adjoining grammar as a framework for statistical natural
language processing. In 14th Conference on Computational Linguistics, COLING’92,
pages 418–424, 1992.
[22] G. Satta. Tree-adjoining Grammar Parsing and Boolean Matrix Multiplication. Comput. Linguist., 20(2):173–191, June 1994.
13
[23] Y. Schabes and A. K. Joshi. An Earley-type parsing algorithm for tree adjoining
grammars. In 26th Annual Meeting of the Association for Computational Linguistics,
ACL’88, pages 258–269, 1988.
[24] S. M. Shieber and Y. Schabes. Synchronous tree-adjoining grammars. In 13th Conference on Computational Linguistics, COLING’90, pages 253–258, 1990.
[25] M. Stone and C. Doran. Sentence planning as description using tree adjoining grammar.
In 35th Annual Meeting of the Association for Computational Linguistics, ACL’97,
pages 198–205, 1997.
[26] Y. Uemura, A. Hasegawa, S. Kobayashi, and T. Yokomori. Tree adjoining grammars
for RNA structure prediction. Theoretical Computer Science, 210(2):277–303, 1999.
[27] L. G. Valiant. General context-free recognition in less than cubic time. JCSS, 10(2):308–
315, 1975.
[28] V. Vassilevska. Efficient algorithms for clique problems. Information Processing Letters,
109(4):254–257, 2009.
[29] K. Vijay-Shankar and A. K. Joshi. Some computational properties of tree adjoining
grammars. In 23rd Annual Meeting of the Association for Computational Linguistics,
ACL’85, pages 82–93, 1985.
[30] K. Vijay-Shanker and A. K. Joshi. Feature structures based tree adjoining grammars.
In 12th Conference on Computational Linguistics, COLING’88, pages 714–719, 1988.
[31] D. H. Younger. Recognition and parsing of context-free languages in time n3 . Information and Control, 10(2):189–208, 1967.
14
S
P(1, 3, 4, 6) In
P(1, 3, 4, 6) Out
P(1, 2, 5, 6) In
P(1, 2, 5, 6) Out
P(1, 2, 5, 6) In
15
P(2, 3, 4, 5) In
P(2, 3, 4, 5) Out
P(2, 3, 4, 5) In
P(1, 3, 4, 6) Out
P(1, 3, 4, 6) In
. . . | (C1 ) l1 r1 (C1 ) | . . . . . . | (C2 ) l2 r2 (C2 ) | . . . . . . | (C3 ) l3 r3 (C3 ) | . . .
e
. . . | (C4 ) l4 r4 (C4 ) | . . .. . . | (C5 ) l5 r5 (C5 ) | . . . . . . | (C6 ) l5 r6 (C6 ) | . . .
Figure 5: Enlarged version of Figure 4. Global structure of a parsing of GGk (G) by Γ. (Clique gadgets are abbreviated.)
| 8cs.DS
|
Tree Edit Distance Cannot be Computed in Strongly Subcubic
Time (unless APSP can)
arXiv:1703.08940v1 [cs.DS] 27 Mar 2017
Karl Bringmann∗
Paweł Gawrychowski†
Shay Mozes‡
Oren Weimann†
Abstract
The edit distance between two rooted ordered trees with n nodes labeled from an alphabet Σ
is the minimum cost of transforming one tree into the other by a sequence of elementary operations consisting of deleting and relabeling existing nodes, as well as inserting new nodes. Tree
edit distance is a well known generalization of string edit distance. The fastest known algorithm
for tree edit distance runs in cubic O(n3 ) time and is based on a similar dynamic programming
solution as string edit distance. In this paper we show that a truly subcubic O(n3−ε ) time
algorithm for tree edit distance is unlikely: For |Σ| = Ω(n), a truly subcubic algorithm for tree
edit distance implies a truly subcubic algorithm for the all pairs shortest paths problem. For
|Σ| = O(1), a truly subcubic algorithm for tree edit distance implies an O(nk−ε ) algorithm for
finding a maximum weight k-clique.
Thus, while in terms of upper bounds string edit distance and tree edit distance are highly
related, in terms of lower bounds string edit distance exhibits the hardness of the strong exponential time hypothesis [Backurs, Indyk STOC’15] whereas tree edit distance exhibits the
hardness of all pairs shortest paths. Our result provides a matching conditional lower bound for
one of the last remaining classic dynamic programming problems.
1
Introduction
Tree edit distance is the most common similarity measure between labelled trees. Algorithms
for computing the tree edit distance are being used in a multitude of applications in various domains including computational biology [24, 37, 52, 58], structured text and data processing (e.g.,
XML) [30, 31, 36], programming languages and compilation [38], computer vision [22, 43], character
recognition [49], automatic grading [14], answer extraction [65], and the list goes on and on.
Let F and G be two rooted trees with a left-to-right order among siblings and where each vertex
is assigned a label from an alphabet Σ. The edit distance between F and G is the minimum cost of
transforming F into G by a sequence of elementary operations (at most one operation per node):
changing the label of a node v, deleting a node v and setting the children of v as the children of
v’s parent (in the place of v in the left-to-right order), and inserting a node v (the complement of
delete1 ). See Figure 1. The cost of these elementary operations is given by two functions: cdel (a)
is the cost of deleting or inserting a vertex with label a, and cmatch (a, b) is the cost of changing the
label of a vertex from a to b.
∗
Max Planck Institute for Informatics, Saarland Informatics Campus
University of Haifa. Partially supported by the Israel Science Foundation grant 794/13.
‡
IDC Herzliya. Partially supported by the Israel Science Foundation grant 794/13.
1
Since a deletion in F is equivalent to an insertion in G and vice versa, we can focus on finding the minimum cost
of a sequence of just deletions and relabelings in both trees that transform F and G into isomorphic trees.
†
1
relabel node x to y
delete node y
y
x
D
A
B
C
relabel node y to x
D
A
C
B
D
A
insert node y
B
C
Figure 1: The three editing operations on a tree with vertex labels.
The Tree Edit Distance (TED) problem was introduced by Tai in the late 70’s [55] as a generalization of the well known string edit distance problem [57]. Since then it was extensively studied.
Tai gave an O(n6 )-time algorithm for TED which was subsequently improved to O(n4 ) in the late
80’s [53], then to O(n3 log n) in the late 90’s [41], and finally to O(n3 ) in 2007 [34]. Many other
algorithms have been developed for TED, see the popular survey of Bille [24] (this survey alone has
more than 600 citations) and the books of Apostolico and Galil [16] and Valiente [56]. For example,
Pawlik and Augsten [48] recently defined a class of dynamic programming algorithms that includes
all the above algorithms for TED, and developed an algorithm whose performance on any input
is not worse (and possibly better) than that of any of the existing algorithms. Other attempts
achieved better running time by restricting the edit operations or the scoring schemes [31,51,54,66],
or by resorting to approximation [12, 17]. However, in the worst case no algorithm currently beats
Ω(n3 ) (not even by a logarithmic factor).
Due to their importance in practice, many of the algorithms described above, as well as additional
heuristics and optimizations were studied experimentally [39,48]. Solving tree edit distance in truly
subcubic O(n3−ε ) time is arguably one of the main open problems in pattern matching, and the
most important one in tree pattern matching.
The fact that, despite the significant body of work on this problem, no truly subcubic time
algorithm has been found, leads to the following natural conjecture that no such algorithm exists.
Conjecture 1. For any > 0 Tree Edit Distance on two n-node trees cannot be solved in O(n3− )
time.
In the same paper proving the O(n3 ) upper bound for TED [34], Demaine et al. prove that their
algorithm is optimal within a certain class of dynamic programming algorithms for TED. However,
proving Conjecture 1 seems to be beyond our current lower bound techniques.
A recent development in theoretical computer science suggests a more fine-grained classification
of problems in P. This is done by showing lower bounds conditioned on the conjectured hardness
of certain archetypal problems such as All Pairs Shortest Paths (APSP), 3-SUM, k-Clique, and
Satisfiability, i.e., the Strong Exponential Time Hypothesis (SETH).
The APSP Conjecture. Given a directed or undirected graph with n vertices and integer edge
weights, many classical algorithms for APSP (such as Dijkstra or Floyd-Warshall) run in O(n3 ) time.
The fastest to date is the recent algorithm of Williams [60] that runs faster than O(n3 / logC n) time
for all constants C. Nevertheless, no truly subcubic O(n3−ε ) time algorithm for APSP is known.
This led to the following conjecture assumed in many papers, e.g. [5, 6, 8, 9, 11, 18, 21, 50, 61, 62].
2
Conjecture 2 (APSP). For any ε > 0 there exists c > 0, such that All Pairs Shortest Paths on n
node graphs with edge weights in {1, . . . , nc } cannot be solved in O(n3−ε ) time.
The (Weighted) k-Clique Conjecture. The fundamental k-Clique problem asks whether a
given undirected unweighted graph on n nodes and O(n2 ) edges contains a clique on k nodes.
This is the parameterized version of the famously NP-hard Max-Clique [40]. k-Clique is amongst
the most well-studied problems in theoretical computer science, and it is the canonical intractable
(W[1]-complete) problem in parameterized complexity. A naive algorithm solves k-Clique in O(nk )
time. A faster O(nωk/3 )-time algorithm (where ω < 2.373 is the exponent of matrix multiplication)
can be achieved via a reduction to Boolean matrix multiplication on matrices of size nk/3 × nk/3
if k is divisible by 3 [47]2 . Any improvement to this bound immediately implies a faster algorithm
for MAX-CUT [59, 64]. It is a longstanding open question whether improvements to this bound are
possible [46, 63]. The k-Clique conjecture asserts that for no k ≥ 3 and ε > 0 the problem has an
O(nωk/3−ε ) time algorithm, or an O(nk−ε ) algorithm avoiding fast matrix multiplication, and has
been used e.g. in [3, 28].
We work with a conjecture on a weighted version of k-Clique. In the Max-Weight k-Clique
problem, the edges have integral weights and we seek the k-clique of maximum total weight. When
the edge weights are small, one can obtain an Õ(nk−ε ) time algorithm [13, 47]. However, when the
weights are larger than nk , the trivial O(nk ) algorithm is the best known (ignoring no(k) improvements). This gives rise to the following conjecture, which has been used e.g. in [10, 18, 21].
Conjecture 3 (Max-Weight k-Clique). For any ε > 0 there exists a constant c > 0, such that for
any k ≥ 3 Max-Weight k-Clique on n-node graphs with edge weights in {1, . . . , nck } cannot be solved
in O(nk(1−ε) ) time.
In 2014, with the burst of the conditional lower bound paradigm, Abboud [1] highlighted seven
main open problems in the field: The first two were to prove quadratic n2−o(1) lower bounds for String
Edit Distance and Longest Common Subsequence, which were soon resolved in STOC’15 [19] and
FOCS’15 [4,29] conditional on SETH. The third problem was to show a cubic n3−o(1) lower bound for
RNA-Folding. Surprisingly, in FOCS’16 [27] it was shown that RNA-Folding can actually be solved
in truly subcubic time, thus ruling out the possibility of such a lower bound. The remaining four
problems remain open. In fact, two of them, showing a cubic lower bound for Graph Diameter and
an ndk/2e−o(1) lower bound for k-SUM, have actually been used as hardness conjectures themselves,
e.g., in SODA’15 [6] and ICALP’13 [8]. Until the present work, no progress has been made on
the last problem posed by Abboud: A cubic lower bound for Tree Edit Distance. In the absence
of progress on either upper bounds or conditional lower bounds for TED, one might think that
Conjecture 1 is yet another fragment in the current landscape of fine grained complexity, and is
unrelated to other common conjectures.
1.1
Our Results
In this paper we resolve the complexity of tree edit distance by showing a tight connection between
edit distance of trees and all pairs shortest paths of graphs. We prove that Conjecture 2 implies
Conjecture 1, and that Conjecture 3 implies Conjecture 1, even for constant alphabet.
2
For the case that k is not divisible by 3 see [35].
3
Theorem 1. A truly subcubic algorithm for tree edit distance on alphabet size |Σ| = Ω(n) implies a
truly subcubic algorithm for APSP. A truly subcubic algorithm for tree edit distance on sufficiently
large alphabet size |Σ| = O(1) implies an O(nk(1−ε) ) algorithm for Max-Weight k-Clique.
Note that the known upper bounds for string edit distance and tree edit distance are highly
related. The O(n2 ) algorithm for strings and the O(n3 ) algorithm for trees (and forests) are both
based on a similar recursive solution: The recursive subproblems in strings (forests) are obtained
by either deleting, inserting, or matching the rightmost or leftmost character (root). In strings, it
is best to always consider the rightmost character. The recursive subproblems are then prefixes and
the overall running time is O(n2 ). In trees however, sticking with the rightmost (or leftmost) root
may result in an O(n4 ) running time. The specific way in which the recursion switches between
leftmost and rightmost roots is exactly what enables the O(n3 ) solution. It is interesting that while
the upper bounds for both problems are so similar, the lower bounds string edit distance exhibits
the hardness of the SETH while tree edit distance exhibits the hardness of APSP.
While a considerable share of the recent conditional lower bounds is on string pattern matching
problems [3, 4, 7, 10, 15, 19, 20, 28, 29, 32, 45], the only conditional lower bound for a tree pattern
matching problem is the recent SODA’16 quadratic lower bound for exact pattern matching [2]
(the problem of deciding whether one tree is a subtree of another). We solve the main remaining
open problem in tree pattern matching, and one of the last remaining classic dynamic programming problems. Indeed, apart from the problems discussed above, for most of the classic dynamic
programming problems a conditional lower bound or an improved algorithm have been found recently. This includes the Fréchet distance [25], bitonic TSP [33], context-free grammar parsing [3],
maximum weight rectangle [18], and pseudopolynomial time algorithms for subset sum [26]. Tree
edit distance was one of the few classic dynamic programming problems that so far resisted this
approach. Two notable remaining dynamic programming problems without matching bounds are
the optimal binary search tree problem (O(n2 )) [44] and knapsack (pseudopolynomial O(nW )) [23].
1.2
Our Reductions
APSP to TED. In order to prove APSP-hardness, by [61] it suffices to show a reduction from
the negative triangle detection problem, where we are given an n-node graph G with edge weights
w(., .) and want to decide whether there are i, j, k with w(i, j) + w(j, k) + w(i, k) < 0. Our first
result is a reduction from negative triangle detection to tree edit distance, which produces trees
of size O(n) over an alphabet of size O(n). This yields the matching conditional lower bound of
O(n3−ε ).
Our reduction constructs trees that are of a very special form: Both trees consist of a single
path (called spine) of length O(n) with a single leaf pending from every node (see Figure 2). Such
instances already have been identified as difficult for a restricted class of algorithms based on a
specific dynamic programming approach [34]. In our setting we cannot assume anything about the
algorithm, and hence need a deeper insight on the structure of any valid sequence of edit operations
(see Figure 2 and Lemma 1). Using this structural understanding, we then show that it is possible
to carefully construct a cost function so that any optimal solution must obey a certain structure
(Figure 3). Namely, for some i, j, k we match the two leaves in depth k, we match the right spine
node in depth k to the left leaf in depth i (which encodes w(i, k)), we match the left spine node in
depth k to the right leaf in depth j (which encodes w(j, k)), and we match as many spine nodes
above depth i and j as possible (which together encode w(i, j) by a telescoping sum).
4
Constant alphabet size. The drawback of the above reduction is the large alphabet size |Σ|, as
essentially each node needs its own alphabet symbol. There are two major difficulties to improving
this to constant alphabet size.
First, the instances identified as hard by the above reduction (and by Demaine et al. [34] for a
restricted class of algorithms) are no longer hard for small alphabet! Indeed, in Section 4 we give
an O(n2 |Σ|2 log n) algorithm for these instances, which is truly subcubic for constant alphabet size.
This algorithm is the first to break the barrier by Demaine et al. for such trees, and we believe it
is of independent interest. Regarding lower bounds, this algorithm shows that for a reduction with
constant alphabet size our trees necessarily need to be more complicated, making it much harder to
reason about the structure of a valid edit sequence. We will construct new hard instances by taking
the previous ones and attaching small subtrees to all nodes.
The second difficulty is that, since the input size of TED is Õ(n + |Σ|2 ), a reduction from
negative triangle detection to TED with constant alphabet size would need to considerably compress
the Ω(n2 ) input size of negative triangle detection. It is a well-known open problem whether such
compressing reductions exist. To circumvent this barrier, we assume the stronger Max-Weight
k-Clique Conjecture, where the input size Õ(n2 ) is very small compared to the running time O(nk ).
Max-Weight k-Clique to TED. Given an instance of Max-Weight k-Clique on an n-node graph
G and weights bounded by nO(k) we construct a TED instance on trees of size O(nk/3+2 ) over
an alphabet of size O(k). One can verify that an O(n3−ε ) algorithm for TED now implies an
O(nk(1−ε/6) ) algorithm for Max-Weight k-Clique, for any sufficiently large k = k(ε).
We roughly follow the reduction from negative triangle detection; now each spine node corresponds to a k/3-clique in G. To cope with the small alphabet, we simulate the previous matching
costs with small gadgets. In particular, to each spine node, corresponding to some k/3-clique U , we
add a small subtree T (U ) of size O(n2 ) such that the edit distance between T (U ) and T (U 0 ) encodes
the total weight of edges between U and U 0 . This raises two issues. First, we need to represent a
weight w ∈ {0, . . . , nO(k) } by trees over an alphabet of size O(k) (that is, constant). This is solved
PO(k)
by writing w in base n as i=0 αi ni and constructing αi nodes of type i, such that the cost of
matching two type i nodes is ni . A second issue is that we need the small subtree T (U ) to interact
with every other small subtree T (U 0 ). So, in a sense, T (U ) needs to “prepare” for any possible U 0 ,
and yet its size needs to be small. We achieve this by creating in T (U 0 ), for every node u in G, a
separate component responsible for counting the total weight of all edges between u and all nodes
in U 0 . Then, in T (U ) we have a separate component for every node u ∈ U , and make sure that it
is necessarily matched to the appropriate component in T (U 0 ).
The final and most intricate component of our reduction is to enforce that in any optimal solution
we have some control on which small subtrees can be matched to which. A similar issue was present
in the negative triangle reduction, when we require control over which spine nodes above depth i
are matched to which spine nodes above depth j. This is handled in the negative triangle reduction
by assigning a different matching cost depending on the node’s depth. Now however, we cannot
afford so many different costs. We overcome this with yet another gadget, called an I-gadget, that
achieves roughly the same goal, but in a more “distributed” manner.
Both of our reductions are highly non-trivial and introduce a number of new tricks that could
be useful for other problems on trees.
5
2
Reducing APSP to TED
We re-define the cost of matching two nodes to be the original cost minus the cost of deleting both
nodes. Then, the goal of TED amounts to choosing a subset of red nodes in both trees, so that
the subtrees defined by the red nodes are isomorphic (i.e., their left-right and ancestor-descendant
relation is the same in both trees) and the total cost of matching the corresponding red nodes is
minimized. See Figure 2. We work with this formulation from now on.
f i1
g j1
f ip
g jp
gjp+1
fip+1
fip+2
gjp+2
fip+q+1
gjp+q+1
fip+q+2
gjp+q+2
Figure 2: Macro structure of the hard instance for TED: A tree F composed of a single spine with
leaves hanging to the right and a tree G composed of a single spine with leaves hanging to the left.
It turns out that a hard instance for TED is given by two seemingly simple caterpillar trees.
These two trees F and G, also called left and right, are shown in Figure 2. Each tree consists of
spine nodes and leaf nodes. If u is a spine node then we denote by u0 the (unique) leaf node attached
to u. For any such hard instance of TED, the red nodes in any matching have the structure given
by Lemma 1 below. Informally, it states that starting from the top of the left tree and ordering the
nodes by depth, the matching consists of (1) a prefix of a matching subsequence of spine nodes in
both trees, (2) a suffix of a matching subsequence of leaf nodes that are in reverse order in the other
tree, and (3) at most one final spine node in each of the trees matching a leaf node in the other tree
that is located between the prefix part (1) and the suffix part (2).
Lemma 1. Let f1 , f2 , . . . and g1 , g2 , . . . denote the spine nodes of F and G, respectively, ordered by
the depth. Then, for some p, q ≥ 0 and some i1 < i2 · · · < ip < ip+1 < · · · < ip+q+1 < ip+q+2 and
j1 < j2 · · · < jp < jp+1 < · · · < jp+q+1 < jp+q+2 the set of red nodes consists of:
(1) Spine nodes fi1 , fi2 , . . . , fip matched respectively to spine nodes gj1 , gj2 , . . . , gjp ,
(2) Leaf nodes fi0p+2 , fi0p+3 , . . . , fi0p+q+1 matched respectively to leaf nodes gj0 p+q+1 , gj0 p+q , . . . , gj0 p+2 (note
the reversed order),
(3) Optionally, a spine node fip+q+2 matched to leaf node gj0 p+1 . Also optionally, a spine node gjp+q+2
matched to a leaf node fi0p+1 .
6
Proof. Consider the subtree defined by the red nodes. It has two isomorphic copies, one in F and
one in G. Its nodes are all the red nodes. The children of node u are all red nodes v1 , v2 , . . . , vk
whose lowest red ancestor is u. The order is such that vi precedes vi+1 in a left-to-right preorder
traversal of F (or equivalently of G). Let u be a red node with two or more children v1 , v2 , . . . , vk ,
k ≥ 2. Observe that u must correspond to spine nodes in both F and G. Further observe that
at most one vi can correspond to a spine node (otherwise, for two spine nodes one must be an
ancestor of the other). Consider any ` ∈ {1, 2, . . . , k − 1}. It is not hard to see that node v`+1 must
correspond to a leaf node in F and node v` must correspond to a leaf node in G. This implies that
both v` and v`+1 are leaves in the red subtree. Moreover, v1 is the only node that may correspond to
a spine node in F and vk is the only node that may correspond to a spine node in G. Consequently,
the red subtree has a particularly simple structure: it consists of nodes u1 , u2 , . . . , up such that for
every ` = 1, 2, . . . , p − 1 the only child of u` is u`+1 , and nodes v1 , v2 , . . . , vk (for some k ≥ 1) that
are all children of up .
For every ` = 1, 2, . . . , p, the node u` must correspond to a spine node fi` ∈ F and gj` ∈ G.
We immediately obtain (1) that i1 < i2 < . . . < ip and that j1 < j2 < . . . < jp . The nodes
v1 , v2 , . . . , vk are all children of up in the subtree. It is possible that all vi are mapped to leaf nodes
fi0p+2 , fi0p+3 , . . . , fi0p+q+1 and gj0 p+2 , gj0 p+3 , . . . , gj0 p+q+1 . In this case, they must be mapped in reverse
order since a left-to-right preorder traversal visits the leaves of G in order of their depth and in
reverse-depth order in F . This implies (2) that ip ≤ ip+2 < . . . < ip+q+1 and jp ≤ jp+2 < . . . <
jp+q+1 . Recall however that v1 may be mapped to a spine node fip+q+2 in F and a leaf node gj0 p+1
in G. The requirement that ip+q+2 > ip+q+1 and that jp+2 > jp+1 ≥ jp follows from the fact that
these nodes correspond to a leftmost leaf in the subtree. For symmetric reasons, vk may be matched
to a spine node gjp+q+2 ∈ G for some jp+q+2 > jp+q+1 and ip+2 > ip+1 ≥ ip . This implies (3) and
concludes the proof.
The above lemma characterizes the structure of a solution to what we call the hard instance of
TED. We next show how to reduce the negative triangle detection problem to TED on the hard
instance. Negative triangle detection is known to be subcubic equivalent to APSP [61]. Given a
complete weighted n-node undirected graph, where w(i, j) denotes the weight of the edge (i, j), the
problem asks whether there are i, j, k such that w(i, j) + w(j, k) + w(i, k) < 0. To solve negative
triangle detection, we clearly only need to find i, j, k that minimize w(i, j) + w(j, k) + w(i, k). We
will show how to construct, given such a graph, a hard instance of TED of size O(n), such that
mini,j,k w(i, j) + w(j, k) + w(i, k) can be extracted from the edit distance.
Lemma 2. Given a complete undirected n-node graph G with weights w(., .) in {1, . . . , nc }, we
construct, in linear time in the output size, an instance of TED of size O(n) with alphabet size
|Σ| = O(n) such that the minimum weight of a triangle in G can be extracted from the edit distance.
Consequently, an O(n3− ) time algorithm for TED implies an O(n3− ) algorithm for negative
triangle detection, and thus an O(n3−/3 ) algorithm for APSP by a reduction in [61].
We create a hard instance of TED consisting of two trees F and G as in Figure 3. Every tree is
divided into a top and a bottom part. The spine nodes of these parts are denoted by a1 , a2 , . . . , an
for the top left part, b1 , b2 , . . . , bn+1 for the bottom left part, c1 , c2 , . . . , cn for the top right part, and
d1 , d2 , . . . , dn+1 for the bottom right part. The labels of all nodes are distinct (hence the alphabet
size |Σ| is Θ(n)). We set the cost cmatch (u, v) of matching two nodes u and v as described below,
where M denotes a sufficiently large number to be specified later. Intuitively, our assignment of
costs ensures that any valid solution to TED must match b0k to d0k , bk+1 to c0j , and dk+1 to a0i for
7
a1
c1
a2
c2
cj
ai
an
cn
b1
d1
b2
d2
bk
dk
bn
dn
bn+1
dn+1
Figure 3: A hard instance of TED constructed for a given instance of negative triangle detection.
Appropriately chosen costs ensure that any optimal solution has a specific structure.
some i, j, k (as shown in Figure 3). Furthermore, the optimal solution (i.e., of minimum cost) must
choose i, j, k that minimize w(i, k) + w(k, j) + w(i, j). The costs are assigned as follows:
(1) cmatch (b0k , d0k ) = −M 2 − 2M · k for every k = 1, 2, . . . , n.
(2) cmatch (bk+1 , c0j ) = −M 2 + M · k + M · j + w(k, j) for every k = 1, 2, . . . , n and j = 1, 2, . . . , n.
(3) cmatch (a0i , dk+1 ) = −M 2 + M · k + M · i + w(i, k) for every i = 1, 2, . . . , n and k = 1, 2, . . . , n.
(4) cmatch (ai , cj ) = −2M + w(i, j) − w(i − 1, j − 1) for every i = 2, 3, . . . , n and j = 2, 3, . . . , n.
(5) cmatch (ai , c1 ) = −M (i + 1) + w(i, 1) for every i = 1, 2, . . . , n.
(6) cmatch (a1 , cj ) = −M (j + 1) + w(1, j) for every j = 1, 2, . . . , n.
All the remaining costs cmatch (u, v) are set to ∞. The following theorem proves that these costs
imply the required structure on the optimal solution as described above. Intuitively, by choosing
sufficiently large M , because of the −M 2 addend in (1), (2) and (3) we can ensure that any optimal
solution matches b0k to d0k , bk0 +1 to c0j , and dk00 +1 to a0i , for some i, j, k and k 0 , k 00 ≤ k. Then, because
of the M · k in (2) and (3), in any optimal solution actually k = k 0 = k 00 and the total cost of all
these matchings is w(k, j) + w(i, k). Finally, because of the −2M in (4), the −M (i + 1) in (5), and
the −M (j + 1) in (6), in any optimal solution ai is matched to cj , ai−1 to cj−1 , ai−2 to cj−2 and so
on. The total cost of these matching is w(i, j) since the w(i, j) − w(i − 1, j − 1) terms in (4) form
a telescoping sum.
8
Theorem 2. For sufficiently large M , the total cost of an optimal matching in a hard instance with
costs (1)-(6) is −3M 2 + mini,j,k w(i, k) + w(k, j) + w(i, j).
Proof. Consider i, j, k minimizing w(i, k) + w(k, j) + w(i, j). We assume without loss of generality
that i ≥ j. It is easy to see that it is possible to choose the following matching (see Figure 3):
1. b0k to d0k with cost −M 2 − 2M · k.
2. bk+1 to c0j with cost −M 2 + M · k + M · j + w(k, j).
3. dk+1 to a0i with cost −M 2 + M · k + M · i + w(i, k).
4. ai to cj , ai−1 to cj−1 , ai−2 to cj−2 , . . . , ai−j+2 to c2 with costs −2M + w(i, j) − w(i − 1, j −
1), −2M + w(i − 1, j − 1) − w(i − 2, j − 2), . . . , −2M + w(i − j + 2, 2) − w(i − j + 1, 1).
5. ai−j+1 to c1 with cost −M · (i − j + 2) + w(i − j + 1, 1).
Summing up and telescoping, the total cost is −M 2 −2M ·k−M 2 +M ·k+M ·j+w(k, j)−M 2 +M ·k+
M ·i+w(i, k)−2M ·(j −1)+w(i, j)−M ·(i−j +2) which is equal to −3M 2 +w(i, k)+w(k, j)+w(i, j).
For the other direction, we need to prove that every solution has cost at least −3M 2 +
mini,j,k w(i, k) + w(k, j) + w(i, j). We first observe that, by Lemma 1, a solution can match b0k
to d0k at most once for some k. Similarly, it can match bk0 +1 to c0j at most once for some j and k 0 ,
and dk00 +1 to a0i at most once for some i and k 00 . Furthermore, for M large enough, either the cost
is larger than −3M 2 or all three such pairs of nodes are matched for some k, i, j, and k 0 , k 00 ≥ k.
Furthermore, if k 0 > k and M is large enough then we can decrease k 0 by one thus decreasing the
total cost, and similarly if k 00 > k. It is enough to consider an optimal solution and hence we can
assume that k = k 0 = k 00 .
Again by Lemma 1, the only possible additional matched pairs of nodes are a subsequence of
spine nodes a1 , . . . , ai and c1 , . . . , cj . We show that an optimal solution matches ai with cj , ai−1 with
cj−1 , ..., ai−j+1 with c1 . To this end, suppose that axz is matched to cyz , for every z = 1, 2, . . . , L,
where 1 ≤ x1 < . . . < xL ≤ i and 1 ≤ y1 < . . . < yL ≤ j. For every z this contributes, up to lower
order terms less than M , −2M if xz , yz > 1, or −M (yz + 1) if xz = 1, or −M (xz + 1) if yz = 1. In
an optimal solution we have x1 = 1 or y1 = 1, as otherwise we can match a1 to c1 to decrease the
total cost. First, assume that y1 = 1. Then, if xz 0 + 1 < xz 0 +1 for some z 0 ∈ {1, . . . , L} (where we
define xL+1 = i + 1), we can increase all x1 , x2 , . . . , xz 0 by 1 to decrease the total cost by M , up to
lower order terms. So xL = i, xL−1 = i − 1, . . . , x1 = i − L + 1. Now if L < j then x1 > 1 (recall that
we assumed i ≥ j) and also yz 0 + 1 < yz 0 +1 for some z 0 ∈ {1, . . . , L} (again, we define yL+1 = j + 1).
This means that we can increase all y1 , y2 , . . . , yz 0 by 1 and then additionally match ax1 −1 with c1
to decrease the total cost by M , up to lower order terms. Second, if x1 = 1 a symmetric argument
applies. We obtain that indeed L = min(i, j) = j, and ai−j+1 is matched to c1 , ai−j+2 is matched
to c2 , ..., ai is matched to cj . Now, by the same calculations as in the previous paragraph, the total
cost is −3M 2 + w(i, k) + w(k, j) + w(i, j).
3
Reducing Max-Weight k-Clique to TED
The drawback of the reduction described in Section 2 is the large size of the alphabet. That is, given
a complete weighted n-node undirected graph it creates two trees of size O(n) where labels of nodes
are distinct, and therefore |Σ| = Θ(n). We would like to refine the reduction so that |Σ| = O(1).
9
However, as the input size of TED on n-node trees and alphabet Σ with O(log n)-bit integer weights
is Õ(n + |Σ|2 ), such a reduction would need to compress the Õ(n2 ) input size of negative triangle
detection considerably. To circumvent this barrier, we assume the stronger Max-Weight k-Clique
Conjecture, where the input size Õ(n2 ) is very small compared to the running time bound O(nk ).
Lemma 3. Given a complete undirected n-node graph G with weights in {1, . . . , nck }, we construct,
in linear time in the output size, an instance of TED of size O(nk/3+2 ) with alphabet size |Σ| = O(ck)
such that the maximum weight of an k-clique in G can be extracted from the edit distance.
0
Thus, an O(n3− ) time algorithm for TED for sufficiently large |Σ| = O(1) implies an
0
O(n(k/3+2)(3− ) ) time algorithm for max-weight k-Clique. Setting = 0 /6, we obtain that, for
every c > 0, there exists k = d6/e such that max-weight k-Clique can be solved in time
0
0
0
O(n(k/3+2)(3− ) ) = O(nk− k/3+6−2 ) = O(nk(1−)−k+6 ) = O(nk(1−) ),
so Conjecture 3 is violated.
The reduction starts with enumerating all k3 -cliques in the graph and identifying them with
numbers 1, 2, . . . , N , where N ≤ nk/3 . Let Q(i) denote the set of nodes in the i-th clique. Then,
for i, j such that Q(i) ∩ Q(j) = ∅, W (i, j) is the total weight of all edges connecting two nodes in
the i-th clique or a node in the i-th clique with a node in the j-th clique. Our goal is to calculate
the maximum value of W (i, z) + W (z, j) + W (j, i) over i, j, z such that Q(i), Q(j) and Q(z) are
pairwise disjoint. If we define w(u, u) = 0 and increase every other weight w(u, v) by Λ := k 2 nck ,
this is equivalent to maximising
over all i, j, z. Indeed, if Q(i),
disjoint, the
Q(j),
Q(z)ckare pairwise
k
k
k
total weight is at least 2 Λ, and otherwise it is at most 2 − 1 (Λ + n ) < 2 Λ. Note that the
new weights are still bounded by nO(ck) .
Our construction of a hard instance of size O(N · poly(n)) is similar to Section 2, however, the
costs are set up differently and we attach small additional gadgets to some of the nodes (which is
necessary, cf. Section 4). The original two trees (with some extra spine nodes without any leaves)
are called the macro structure and all small gadgets are called the micro structures. With notation
as in Section 2, the following micro structures are created for every i = 1, 2, . . . , N (see Figure 4):
1. A0i attached to the leaf a0i ,
2. a copy of I attached as the left child of the leaf c0i ,
3. Ci0 attached as the right child of the leaf c0i ,
4. Ai attached to the spine node ai−1 between the previously existing children ai and a0i−1 ,
5. Bi attached to the spine node bi between the previously existing children bi+1 and b0i ,
6. Ci attached to the spine node ci−1 as the rightmost child,
7. Di attached to the spine node di between the previously existing children d0i and di+1 .
Notice that Ai is attached above ai (and similarly Cj is attached above cj ). Therefore, we need to
create dummy spine nodes a0 and c0 . We also insert an additional spine node b00i between bi and
bi+1 and similarly d00i between di and di+1 , for every i = 1, 2, . . . , N − 1. See Figure 4.
The costs in the macro structure are chosen as follows, where again M is a sufficiently large
value (picking M = nO(ck) will suffice):
10
a0
a1
c0
A1
c1
C1
cj
Cj
A01
ai
Ai
Cj0
A0i
z
AN
CN
cN
}|
I
d1
B1
D1
z
b001
0
CN
{
A0N I
b1
n−1
aN
d001
}|
n−1
I
{
z
d00z0 −1
dz0
}|
Dz0
{
Bz
n−1
b00z−1
bz
b00N −1
d00N −1
bN
dN
BN
DN
Figure 4: A hard instance of TED constructed for a given instance of max-weight k-clique.
1. cmatch (bz , c0j ) = −M 8 for every z = 1, 2, . . . , N and i = 1, 2, . . . , N ,
2. cmatch (a0i , dz 0 ) = −M 8 for every i = 1, 2, . . . , N and z 0 = 1, 2, . . . , N ,
3. cmatch (b0z , d0z 0 ) = −M 7 · 2 for every z = 1, 2, . . . , N and z 0 = 1, 2, . . . , N ,
4. cmatch (ai , cj ) = −M 3 · 2 + M 2 for every i = 1, 2, . . . , N and j = 1, 2, . . . , N .
Additionally, the extra spine nodes b00i and d00i can be matched to some of the nodes of I. Each
copy of I is a path consisting of k/3 segments I0 , I1 , . . . , Ik/3−1 of length n − 1, where the root
of the whole I belongs to I0 . The label of every u ∈ Ii is the same and the costs are set so that
cmatch (u, u) = −M 7 · ni . The label of every b00z (and also d00z 0 ) is chosen as the label of every u ∈ Im ,
where nm is the largest power of n dividing N − z. The cost of matching any other two labels used
in the macro structure is set to infinity. For the nodes belonging to the other micro structures, the
cost of matching is at least −M 6 and will be specified precisely later. This is enough to enforce the
following structural property.
11
Lemma 4. For sufficiently large M , any optimal matching has the following structure: there exist
i, j, z such that a0i is matched to dz , c0j is matched to bz , b01 is matched to d0z−1 , b02 is matched to
d0z−2 , . . . , b0z−1 is matched to d01 . Furthermore, if z < N then b00z is matched to a descendant of c0j
and d00z is matched to a descendant of a0i . Ignoring the spine nodes a1 , . . . , ai , c1 , . . . , ci and all micro
structures that are not copies of I the cost of any such solution is −M 8 · 2 − M 7 · 2(N − 1).
Proof. For sufficiently large M , any optimal solution must match a0i to dz and c0j to bz 0 , for some
i, j, z, z 0 , as otherwise its cost is larger than −M 8 · 2. By the reasoning described in Lemma 1, these
i, j, z, z 0 are uniquely defined for any optimal solution.
Nodes from the copy of I attached as the left child of the leaf c0j can be matched to some spine
nodes below bz , nodes from the copy of I attached as the right child of the leaf a0i can be matched to
some spine nodes below dz 0 , and no other nodes from the copies of I can be matched. We claim that
the total contribution of these nodes is −M 7 (N − z) and −M 7 (N − z 0 ), respectively. By symmetry,
it is enough to justify the former. Observe that the cost of matching a single u ∈ Ii is smaller
than the total cost of matching all nodes from I0 ∪ . . . Ii−1 , therefore an optimal solution must
match as many nodes to nodes from Ik/3−1 as possible. Looking at the expansions of all numbers
Pk/3−1
N − z, N − (z + 1), . . . , N − (N − 1) in base n, where N − z = i=0 αi ni , we see that there are
αk/3−1 such nodes, namely the nodes b0N −z 0 with z ≤ z 0 < N and N − z 0 divisible by nk/3−1 . Then,
an optimal solution must match as many nodes to nodes from Ik/3−2 as possible to nodes above the
topmost node matched to a node from Ik/3−1 . Looking again at the same expression, we see that
there are αk/3−2 such nodes, namely the nodes b0N −z 0 with z ≤ z 0 < N − αk/3−1 nk/3−1 and N − z 0
divisible by nk/3−2 . Continuing in the same fashion, we obtain that there are αi nodes matched to
nodes from Ii , making the total cost −M 7 (N − z) as claimed.
We assume without loss of generality that z ≥ z 0 . Then, an optimal solution must match d0z 0 −1
to b0xz0 −1 , d0z 0 −2 to b0xz0 −2 , . . . , and d01 to b0x1 , for some z ≥ x1 > . . . > xz 0 −1 ≥ 1, as otherwise its
cost is larger than −M 8 · 2 − M 7 (2N − z − z 0 ) − M 7 · 2(z 0 − 1). Rewriting the cost we obtain
−M 8 · 2 − M 7 (2N − 2 − z + z 0 ), so recalling our assumption z ≥ z 0 we see that in fact z = z 0 as
otherwise its cost is larger than −M 8 · 2 − M 7 · 2(N − 1).
We are now ready to state properties of the remaining micro structures. Let cmatch (T1 , T2 ) denote
the cost of matching two trees T1 and T2 . Then, we require that:
1. cmatch (A0i , Dz 0 ) = −M 6 − M 3 (N − i) − W (i, z 0 ) for every i = 1, 2, . . . , N and z 0 = 1, 2, . . . , N ,
2. cmatch (Bz , Cj0 ) = −M 6 − M 3 (N − j) − W (z, j) for every z = 1, 2, . . . , N and j = 1, 2, . . . , N .
3. cmatch (Ai , Cj ) = −M 2 − W (j, i) + W (j − 1, i − 1) for every i = 2, 3, . . . , N and j = 2, 3, . . . , N .
4. cmatch (Ai , C1 ) = −M 5 − M 3 (i − 1) − W (1, i) for every i = 1, 2, . . . , N ,
5. cmatch (A1 , Cj ) = −M 5 − M 3 (j − 1) − W (j, 1) for every j = 1, 2, . . . , N .
The labels of the nodes in the micro structures should be partitioned into disjoint subsets corresponding to the following micro structures:
1. {A01 , A02 , . . . , A0N , D1 , D2 , . . . , DN },
0 },
2. {B1 , B2 , . . . , BN , C10 , C20 , . . . , CN
12
3. {A1 , A2 , . . . , AN , C1 , C2 , . . . , CN },
so that two nodes can be matched only if their labels belong to the same subset. The cost of
matching any node of A0i , Dz 0 , Bz , Cj0 should be at least −M 6 . The cost of matching any node of
Ai , Cj should be at least −M 2 , except that the root of Ai (Cj ) can be matched to the root of C1
(A1 ) with cost larger than −M 5 − M but at most −M 5 , and, for any non-root node of Ai (Cj ) and
for any non-root node of C1 (A1 ), the cost of matching is larger than −M 4 . Finally, every Ai and
Cj should consist of O(n2 ) nodes. Now we can show that, assuming these properties, any optimal
solution has a specific structure.
Lemma 5. For sufficiently large M , the total cost of an optimal matching is
−M 8 · 2 − M 7 · 2(N − 1) − M 6 · 2 − M 5 − M 3 · 2N + M 2 − max{W (i, z) + W (z, j) + W (j, i)}.
i,j,z
Proof. Consider i, j, z maximizing W (i, z) + W (z, j) + W (j, i). We may assume that i ≥ j. Then,
it is possible to choose the following matching:
1. bk to c0j with cost −M 8 ,
2. some nodes from the copy of I being the left child of c0j to some spine nodes below bz with
total cost −M 7 (N − z),
3. a0i to dk with cost −M 8 ,
4. some nodes from the copy of I being the right child of a0i to some spine nodes below dz with
total cost −M 7 (N − z),
5. b01 to d0z−1 , b02 to d0z−2 , . . . , b0z−1 to d01 with cost −M 7 · 2 each,
6. ai to cj , ai−1 to cj−1 , . . . , ai−j+1 to c1 with cost −M 3 · 2 + M 2 each,
7. A0i to Dz with cost −M 6 − M 3 (N − i) − W (i, z),
8. Bz to Cj0 with cost −M 6 − M 3 (N − j) − W (z, j),
9. Ai to Cj , Ai−1 to Cj−1 , . . . , Ai−j+2 to C2 with costs −M 2 − W (j, i) + W (j − 1, i − 1),
−M 2 − W (j − 1, i − 1) + W (j − 2, i − 2), . . . , −M 2 − W (2, i − j + 2) + W (1, i − j + 1).
10. Ai−j+1 to C1 with cost −M 5 − M 3 (i − j) − W (1, i − j + 1).
13
Summing up and telescoping, the total cost is
−M 8
−M 7 (N − z)
−M 8
−M 7 (N − z)
−M 7 · 2(z − 1)
−M 3 · 2j + M 2 · j
−M 6 − M 3 (N − i) − W (i − z)
−M 6 − M 3 (N − j) − W (z, j)
−M 2 (j − 1) − W (j, i) − M 5 − M 3 (i − j)
= − M 8 · 2 − M 7 · 2(N − 1) − M 6 · 2 − M 5 − M 3 · 2N + M 2 − W (i, z) − W (z, j) − W (j, i).
For the other direction, we need to argue that every solution has cost at least −M 8 · 2 − M 7 ·
2(N −1)−M 6 ·2−M 5 −M 3 ·2N +M 2 −maxi,j,z {W (i, z)+W (z, j)+W (j, i)}. We start with invoking
Lemma 4 and analyse the remaining small micro structures. Due to leaves b01 , . . . , b0z−1 , d01 , . . . , d0z−1
being already matched, no node from B1 , . . . , Bz−1 , D1 , . . . , Dz−1 can be matched (as they can in
general only be matched to A0∗ ’s and C∗0 ’s). Then, due to b00z and d00z being already matched (or
z = N ) no node from Bz+1 , . . . , BN , Dz+1 , . . . , DN can be matched, and nodes from Bz or Dz
can be only matched to nodes from Cj0 or A0i , respectively. The cost incurred by all such nodes is
cmatch (A0i , Dz ) + cmatch (Bz , Cj0 ), making the total cost −M 8 · 2 − M 7 · 2(N − 1) − M 6 · 2 − M 3 (2N − i −
j)−W (i, z)−W (z, j). It remains to analyse the contribution of all spine nodes a1 , . . . , aN , b1 , . . . , bN
and nodes from micro structures A1 , . . . , AN , C1 , . . . , CN .
Consider the micro structures C1 and A1 . Matching their roots to roots of some Ai0 and Cj 0 ,
respectively, decreases the total cost by at least −M 5 , which is much smaller than the cost of
matching the remaining nodes. Furthermore, it is not possible to match both the root of C1 to the
root of some Ai0 and the root of A1 to the root of some Cj 0 at the same time, unless the root of
A1 is matched to the root of C1 . Therefore, an optimal solution matches exactly one of them or
both to each other, say we match the root of C1 to the root of some Ai0 , thus adding cmatch (Ai0 , C1 )
to the total cost. Due to a0i being matched to dz , i0 ≤ i holds. Now, unless i0 = 1, no node from
A1 can be matched to a node from Cj 0 , so the cost of matching any ai0 to cj 0 is now much smaller
than the cost of matching nodes in the remaining micro structures (for each such node, the cost is
at least −M 2 , and there are at most O(n2 ) of them in a single micro structure, so the total cost
contributed by a single micro structure is larger than −M 3 for M large enough) and, by Lemma 1,
only nodes a1 , . . . , ai , c1 , . . . , cj can be matched, so an optimal solution matches as many such pairs
as possible. Due to the root of C1 being matched to the root of Ai0 , only nodes ai0 , ai0 +1 , . . . , ai and
c1 , . . . , cj can be matched, so there are min(i − i0 + 1, j) such matched pairs. If i − i0 + 1 < j and
i0 > 1 then C1 can be matched with Ai0 −1 instead of Ai0 which allows for an additional pair and
decreases the total cost (because matching a pair (a∗ , c∗ ) adds −M 3 · 2 to the cost while decreasing
i0 by one adds M 3 to the cost cmatch (Ai0 , C1 ), up to lower order terms). If i − i0 + 1 < j and
i0 = 1 then A1 can be matched with C2 instead of C1 while keeping the number of matched pairs
intact to decrease the total cost. So i − i0 + 1 ≥ j (implying i ≥ j, which is due to our initial
assumption that the root of C1 is matched to the root of some Ai0 ). Then, if i0 < i − j + 1, C1 can
14
be matched with Ai0 +1 instead of Ai0 without changing the number of matched pairs to decrease
the total cost. Thus, i0 = i − j + 1 and ai is matched to cj , ai−1 to cj−1 , . . . , and ai−j+1 to c1 ,
Then nodes from Ai can be only matched to nodes from Cj , nodes from Ai−1 only to nodes from
Cj−1 , and so on. By the same calculations as in the previous paragraph, the total cost is therefore
−M 8 · 2 − M 7 · 2(N − 1) − M 6 · 2 − M 5 − M 3 · 2N + M 2 − maxi,j,z {W (i, z) + W (z, j) + W (j, i)}.
To complete the proof we need to design the remaining micro structures. We start with describing
some preliminary gadgets that will be later appropriately composed to obtain the micro structures
Ai , A0i , Bz , Cj , Cj0 , Dz 0 with the required properties. Each such gadget consists of two trees, called
left and right, and we are able to exactly calculate the cost of matching them. The main difficulty
here is that we need to keep the size of the alphabet small, so for instance we are not able to create
a distinct label for every node of the original graph. At this point it is also important to note that
we can assume M = nO(ck) , i.e., there is a constant d = O(ck) such that all weights constructed
above have absolute value less than nd .
Decrease gadget D(x). For any x ∈ {0, . . . , nd − 1}, the edit distance of the left and right tree
of D(x) is −x, and furthermore the right tree does notP
depend on the value of x.
i
This is obtained by representing x in base n as x = d−1
i=0 αi n . The left tree is a path composed
of d segments, the i-th segment consisting of αi nodes. The right tree is a path composed of d
segments, each consisting of n − 1 nodes. Nodes from the i-th segment of the left tree can be
matched with nodes from the i-th segment of the right tree with cost −ni , so the total cost is −x,
see Figure 5 (left). We reuse the same set of distinct labels in every decrease gadget of the same
type, hence we need only O(d) distinct labels in total. Furthermore, note that the cost of matching
the left tree of D(x) with any tree is at least −x and the cost of matching any node of D(x) is −ni
for some i ∈ {0, 1, . . . , d − 1}.
Equality gadget E(u, v, c= ). For any u, v ∈ {1, . . . , n} and c= ∈ {0, . . . , nd −1}, the edit distance
of the left and right tree of E(u, v, c= ) is −c= · n if u = v and at least −c= · n + c= otherwise. Also,
the left tree does not depend on v and the right tree does not depend on u.
The left tree is a path composed of a segment of length u and a segment of length n − u. The
right tree is a path composed of a segment of length v and a segment of length n − v. Nodes from
the first segment of the left tree can be matched with nodes from the first segment of the right tree
with cost −c= , and similarly for the second segments. Then, if u = v we can match all nodes in
both trees, so the total cost is −c= · n. Otherwise, we can match at most n − 1 nodes, making
the total cost at least −c= · n + c= , see Figure 5 (right). Furthermore, note that the total cost of
matching the left tree of E(u, v, c= ) with any tree is at least −c= · n and the cost of matching any
node of E(u, v, c= ) is −c= .
Connection gadget C(i, j, M ). For any i, j ∈ {1, . . . , N } and sufficiently large M ∈ {0, . . . , nd −
1}, the edit distance of the left and right tree of C(i, j, M ) is −M − W (i, j). The left tree does not
depend on j and the right tree does not depend on i.
Let {u1 , . . . , uk/3 } and {v1 , . . . , vk/3 } be the k/3-cliques corresponding to i and j, respectively,
where u1 < u2 < . . . , uk/3 and v1 < v2 < . . . < vk/3 . Recall that W (i, j) denotes the total weight of
all edges connecting two nodes in the i-th clique or a node in the i-th clique with a node in the j-th
clique, where we assume that w(u, u) = 0. We construct the gadget C(i, j, M ) as follows. The root
15
−1
−1
−1
−n
z
−n
−n
}|
n−1
−n
{
−n2
{
v
}|
z
u
}|
z
{
−n2
{
n−u
}|
z
{
z
n−v
}|
Figure 5: Left: Decrease gadget built for d = 3, n = 6 and x = n2 · 2 + n · 4 + 3. Right: Equality
gadget for u = 3, v = 6.
of the left tree has degree 1 + k/3 and the root of the right tree has P
degree 1 + n. Their rightmost
children correspond to the root of the left and the right trees of D( x<y w(ux , uy )), respectively.
Every other child of the left root can be matched with every other child of the right root with cost
−M2 (we fix M1 and M2 later). Intuitively, we would like the x-th child
P of the the left root to
be matched with the ux -th child of the right root, and then contribute − y w(ux , vy ) to the total
cost, so that summing up over x = 1, 2, . . . , k/3 we obtain the desired sum. To this end, we attach
the left tree of E(ux , ·, M1 ) and the right tree of D(·) to the
Px-th child of the left root. Similarly, we
attach the right tree of E(·, t, M1 ) and the left tree of D( y w(t, vy )) to the t-th child of the right
root. Here we use · to emphasise that a particular tree does not depend on the particular value of
the parameter. All decrease gadgets are of the same type. See Figure 6.
We can clearly construct a solution with total cost −M2 · k/3 − M1 · n · k/3 − W (i, j) (because
16
P
D( x,y w(ux , uy ))
−M2
rig
ht
t
lef
right
right
t
lef
ft
le
P
D( y w(7, vy ))
E(u2 , 7, M1 )
Figure 6: Schematic illustration of a connection gadget for k/3 = 2 and n = 8.
we have enumerated the clique corresponding to i so that u1 < u2 < . . . < uk/3 ).
P We claim that,
for appropriately chosen M1 and M2 , no better solution is possible. Let W = u,v w(u, v). We
fix M1 = W · (k/3 + 1). This is enough to guarantee that the total cost contributed by nodes in
all decrease gadgets is at least −M1 . The total cost contributed by nodes in all equality gadgets is
at least −M1 · n · k/3. Consequently, setting M2 = M1 · n · k/3 + M1 guarantees that any optimal
solution must match all children of the left root, so in fact, for every x = 1, 2, . . . , k/3 we must
match the x-th child of the left root to some child of the right root. Because matching the left tree
of any decrease gadget contributes at least −W to the total cost, by the choice of M1 an optimal
solution in fact must match the x-th child of the left root with the ux -th child of the right root, as
otherwise we lose at least M1 due to the corresponding equality gadget that cannot be compensated
byP
matching its accompanying decrease gadget. Finally, the corresponding decrease gadget adds
− y w(ux , vy ) to the total cost. Therefore, as long as M ≥ M2 · k/3 + M1 · n · k/3 the total cost is
indeed −M −W (i, j) after choosing the cost of matching the roots to be −M +M2 ·k/3+M1 ·n·k/3.
For any node in a decrease gadget, the cost of matching is at least −W , for any node in an equality
gadget, the cost of matching is −M1 , and finally the cost of matching the children of the roots is
−M2 , so the cost of matching any node of C(i, j, W ) is at least −M . For the correctness of the
construction it is enough that M is at least
M2 · k/3 + M1 · n · k/3 =M1 ((nk/3 + 1)k/3 + nk/3)
=W (k/3 + 1)k/3(nk/3 + 1 + n)
=W (k/3 + 1)k/3(n(k/3 + 1) + 1)
≤W · n(k/3 + 1)3 = nO(ck) .
Micro structures A0i , Dz 0 , Bz , Cj0 . We only explain how to construct A0i and Dz 0 , for any i =
1, 2, . . . , N and z 0 = 1, 2, . . . , N , as the construction of Bz and Cj0 is symmetric. Recall that we
require cmatch (A0i , Dz 0 ) = −M 6 − M 4 (N − i) − W (i, z 0 ) and for every node in A0i and Dz 0 the cost of
matching should be at least −M 6 .
A0i consists of a root to which we attach the left tree of D(M 6 + M 4 (N − i) − M ) and the left
tree of C(i, ·, M ), while Dz 0 consists of a root to which we attach the right tree of D(·) and the right
17
tree of C(·, z 0 , M ). All decrease gadgets attached as the left children of A0i and Dz 0 are of the same
type, and all decrease gadgets used inside the connection gadgets attached as the right children
are also of the same but other type. This guarantees that the total cost of matching A0i to Dz 0 is
simply −M 6 − M 4 (N − i) + M − M − W (i, j) = −M 6 − M 4 (N − i) − W (i, j). For sufficiently large
M ≥ W · n(k/3 + 1)3 , the cost of matching any node in D(M 6 + M 4 (N − i) − M ) is at least −M 6
and the cost of matching any node in C(i, j, M ) is at least M .
Micro structures Ai , Cj . Here the situation is a bit more complex, as we simultaneously require
that cmatch (Ai , Cj ) = −M 2 − W (j, i) + W (j − 1, i − 1) for every i = 2, 3, . . . , N and j = 2, 3, . . . , N
and cmatch (Ai , C1 ) = −M 5 − M 3 (i − 1) − W (1, i), and cmatch (A1 , Cj ) = −M 5 − M 3 (j − 1) − W (j, 1)
for every i = 1, 2, . . . , N and j = 1, 2, . . . , N . We must also make sure that the cost of matching a
node of Ai to a node of Cj should be at least −M 2 , except that the root of Ai (Cj ) can be matched
to the root of C1 (A1 ) with cost larger than −M 5 − M but at most −M 5 and, for any non-root
node of Ai (Cj ) and for any non-root node of C1 (A1 ), the cost of matching is larger than −M 4 .
For every i > 1 (j > 1), Ai (Cj ) consists of two subtrees, called left and right, attached to
the common root, while A1 (C1 ) consists of a single subtree connected to a root. For every i > 1
(j > 1), the left subtree of Ai (the right subtree of Cj ) consists of a root with two subtrees, called
left-right and left-right (right-left and right-right). For every i > 1, nodes of the right subtree of
Ai can only be matched to nodes of C1 and nodes of the left subtree of Ai can only be matched to
nodes of the right subtree of Cj for any j > 1. For every j > 1, nodes of the left subtree of Cj can
only be matched to nodes of A1 and nodes of the right subtree of Cj can only be matched to nodes
of the left subtree of Ai for any i > 1. Nodes of A1 can be matched to nodes of the left subtree of
Cj , for any j > 1. Nodes of C1 can be matched to nodes of the right subtree of Ai , for any i > 1.
Additionally, the root of A1 can be matched to the root of C1 with cost −M 5 −W (1, 1) > −M 5 −M ,
and for any i > 1 (j > 1), the root of Ai (Cj ) can be matched to the root of C1 (A1 ) with cost
−M 5 . The subtrees are constructed as follows:
1. the right subtree of Ai is the left tree of D(M 3 (i − 1) + W (1, i)),
2. the only subtree of A1 is the right tree of D(·),
3. the left subtree of Cj is the left tree of D(M 3 (j − 1) + W (j, 1)),
4. the only subtree of C1 is the right tree of D(·).
It remains to fully define the left subtree of every Ai and the right subtree of every Cj , for i, j > 1.
Recall that the goal is to ensure cmatch (Ai , Cj ) = −M 2 − W (j, i) + W (j − 1, i − 1). We define a
new n-node graph with weight function w0 (u, v) = M − w(u, v) for any u 6= v (for sufficiently large
M , the new weights are positive). Then, let C 0 (i, j, M ) denote the connection gadget C(i, j, M )
constructed for the new graph. Nodes of the left-left (left-right) subtree of Ai can be only matched
to nodes of the right-left (right-right) subtree of Cj . The subtrees are constructed as follows:
1. the left-left subtree of Ai is the left tree of C(i, ·, M · (k/3)2 ),
2. the right-left subtree of Cj is the right tree of C(·, j, M · (k/3)2 ),
3. the left-right subtree of Ai is the left tree of C 0 (i − 1, ·, M 2 ),
4. the right-right subtree of Cj is the right tree of C 0 (·, j − 1, M 2 ).
18
See Figure 7. For the construction of C(i, j, M · (k/3)2 ) and C(i − 1, j − 1, M 2 ) to be correct we
need that M · (k/3)2 ≥ W · n(k/3 + 1)3 and M 2 ≥ M · n2 · n(k/3 + 1)3 , respectively, which holds
for sufficiently large M .
Now we calculate cmatch (Ai , Cj ). C(i−1, j−1, M 2 ) contributes −M 2 minus the total cost of edges
connecting two subsets of k/3 nodes in the new graph. As the weights in the new graph are defined
as w0 (u, v) = M − w(u, v), this is exactly −M 2 − (M · (k/3)2 − W (i − 1, j − 1)). C(i, j, M · (k/3)2 )
contributes −M · (k/3)2 − W (i, j), so cmatch (Ai , Cj ) = −M · (k/3)2 − W (i, j) − M 2 − (M · (k/3)2 −
W (i − 1, j − 1)) = −M 2 − W (i, j) + W (i − 1, j − 1) as required.
It remains to bound the cost of matching nodes. Nodes in the left subtree of Ai (Cj ) can be
matched only to nodes of C1 (A1 ) with cost at least −M 3 · n > −M 4 , except that the roots can
be matched with cost −M 5 . The cost of matching a node of Ai to a node of Cj , for i, j > 1, is
either at least −M · (k/3)2 (for the nodes of C(i, j, M · (k/3)2 ) or at least −M 2 (for the nodes of
C(i, j, M 2 )), so for sufficiently large M at least −M 2 .
−M 5 − W (1, 1)
A1
−M
5
5
right
right
−M
C1
D(M 3 (j − 1) + W (j, 1))
lef
D(M 3 (i − 1) + W (1, i))
left
t
Ai
le
ft
C(i, j, M 3 · (k/3)2 )
Cj
t
righ
left
rig
ht
C 0 (i − 1, j − 1, M 4 )
Figure 7: Micro structures A1 , C1 and Ai , Cj for i, j > 1.
Wrapping up. We have shown how to construct, given a complete undirected n-node graph G,
two trees such that the weight of the max-weight k-clique in G can be extracted from the cost of
an optimal matching (and, as mentioned in the beginning of Section 2, by a simple transformation
this is equal to the edit distance). To complete the proof of Lemma 3, we need to bound the size of
both trees and also the size of the alphabet used to label their nodes.
19
Initially, each tree consists of 2N original spine nodes, where N ≤ nk/3 , 2N leaf nodes, and N
additional spine nodes. Then, we attach appropriate microstructures to the original spine nodes
and leaf nodes. The microstructures are A0i , I, Cj0 , Ai , Bz , Cj , Dz 0 . Every copy of I consists of k/3 · n
nodes. To analyse the size of the remaining microstructures, first note that if x ∈ {0, . . . , nd } then
the decrease gadget D(x) consists of O(d · n) nodes. The equality gadget always consists of O(n)
nodes. Finally, the connection gadget E(·, ·, M ) with M ∈ {0, . . . , nd } consists of O(n(n + d · n) + d ·
n) = O(d · n2 ) nodes. Let M = nd with d to be specified later. Now, the size of the microstructures
can be bounded as follows: A0i and Dz 0 (and also Bz and Cj0 ) consist of O(6d · n + d · n2 ) = O(d · n2 )
nodes. The right subtree of Ai (and the left subtree of Cj ) consists of O(3d · n) nodes, while the
left subtree of Ai (and the right subtree of Cj ) consist of O(k 2 d · n2 + 2d · n2 ) = O(k 2 d · n2 ) nodes.
Thus, the total size of all microstructures is O(N · k 2 d · n2 ). It remains to bound M . Recall that
we require M ≥ W · n(k/3 + 1)3 , M · (k/3)2 ≥ W · n(k/3 + 1)3 and M ≥ n3 (k/3 + 1)3 , where
W ≤ n2 · nO(ck) = nO(ck) . Hence, it is sufficient that M ≥ 8W · n3 k 3 . Setting d = Θ(ck) is therefore
enough. The size of the whole instance thus is O(nk/3+2 · ck) = O(nk/3+2 ).
We also have to bound the size of the alphabet. We need k/3 distinct labels for the nodes of I.
We need O(d) distinct labels for the nodes of all decrease gadgets of the same type. There is a
constant number of types, and all other nodes require only a constant number of distinct labels
(irrespectively on c and k), so the total size of the alphabet is O(ck) = O(1).
4
Algorithm for Caterpillars on Small Alphabet
In this section, we show that the hard instances of TED from Section 2 can be solved in time
O(n2 |Σ|2 log n), where n is the size of the trees and Σ is the alphabet. Recall that in such an
instance we are given two trees F and G both consisting of a single path (called spine) of length
O(n) with a single leaf pending from every node, and all these leafs are to the right of the path in
F and to the left of the path in G (see Figure 2). In the following we use the same notation as in
Lemma 1. At a high level, we want to guess the rootmost non-spine node in the left tree fi0p+1 and
the rootmost non-spine node in the right tree gj0 p+1 . The optimal matching of spine nodes above
these non-spine nodes can be precomputed in O(n2 ) total time with a simple dynamic programming
algorithm. It might be tempting to say the same about the situation below, but this is much more
complicated due to the fact that leaf nodes in this part are matched in reversed order. To overcome
this difficulty, we need the following tool.
Lemma 6. For strings s[1..n] and t[1..m] over alphabet Σ and matching cost cmatch (c, d) for any
two letters c, d ∈ Σ, we define the optimal matching of s and the reverse of t as
min
k
nX
o
cmatch (s[i` ], t[j` ]) k ≥ 0, 1 ≤ i1 < . . . < ik ≤ n, 1 ≤ jk < . . . < j1 ≤ m .
`=1
Given two strings s[1..n], t[1..n], in O(n2 log n) total time we can calculate, for every i and j, the
optimal matching of s[1..i] and the reverse of t[1..j].
Proof. We construct an (n + 1) × (n + 1) grid graph on nodes vi,j , where i, j = 0, 1, . . . , n as follows.
For every i, j = 0, 1, . . . , n, we create a directed edge from vi,j to vi+1,j and vi,j+1 with length zero.
Also, we create a directed edge from vi,j to vi+1,j+1 with length cmatch (s[i], t[n + 1 − j]). Then, paths
20
from v1,n+1−j to vi,n+1 are in one-to-one correspondence with matchings of s[1..i] to the reverse of
t[1..j]. Therefore, the cheapest such path corresponds to the optimal matching.
The grid is a planar directed graph, and all starting nodes v1,n+1−j lie on the external face, so
we can use the multiple-source shortest paths algorithm of Klein [42] to compute, in O(n2 log n)
time, a representation of shortest paths from all starting nodes v1,n+1−j to all nodes of the grid.3
This representation can be then queried in O(log n) time to extract the length of any path from
v1,n+1−j to vi,n+1 . Thus, the total time is O(n2 log n).
To see how Lemma 6 can be helpful, consider the (simpler) case when there are no additional
spine nodes fip+q+2 and gjp+q+2 . We construct two strings s and t by writing down the labels of leaf
0
0
nodes fn0 , fn−1
, . . . , f10 and gn0 , gn−1
, . . . , g10 , respectively, and preprocess them using Lemma 6. Then,
to find the optimal matching we guess ip+2 and jp+2 . As explained above, optimal matching of spine
nodes above fi0p+2 and gj0 p+2 can be precomputed in O(n2 ) time in advance. Then, we need to match
some of the leaf nodes fi0p+2 , fi0p+2 +1 , fi0p+2 +2 , . . . to some of the leaf nodes gj0 p+2 , gj0 p+2 +1 , gj0 p+2 +2 , . . .
in the reversed order. This exactly corresponds to matching s[1, n + 1 − ip+2 ] to the reverse of
t[1, n + 1 − jp+2 ] and thus is also precomputed. Iterating over all possible ip+2 and jp+2 gives us
the optimal matching in O(n2 log n) total time.
Now consider the general case. We assume that both optional spine nodes fip+q+2 and gjp+q+2
exist; if only one of them is present the algorithm is very similar. As in the simpler case, we iterate
over all possible ip+1 and jp+1 . The natural next step would be to iterate over all possible ip+q+2
and jp+q+2 , but this is too expensive. However, because no spine nor leaf nodes below fip+q+2 (or
gjp+q+2 ) are matched, we can as well replace fip+q+2 with the lowest spine node with the same label
(and similarly for gjp+q+2 ). Thus, instead of guessing ip+q+2 we can guess the label of fip+q+2 and
choose the lowest spine node with such label (and similarly for jp+q+2 ). Now we retrieve the precomputed optimal matching of spine nodes above fi0p+1 and gj0 p+1 . Then we need to find the optimal
matching of leaf nodes fi0p+1 +1 , fi0p+1 +2 , . . . , fi0p+q+2 −1 and gi0p+1 +1 , gi0p+1 +2 , . . . , gj0 p+q+2 −1 . This can
be precomputed in O(n2 |Σ|2 log n) time with Lemma 6. Indeed, there are only |Σ| possibilities
for ip+q+2 − 1 and also |Σ| possibilities for jp+q+2 − 1, as both of them are defined by the lowest
occurrence of a label among the spine nodes of the left and the right tree, respectively. For each
such combination, we construct two strings s and t by writing down the labels of leaf nodes above
0
0
fp+q+2
and gp+q+2
in the bottom-up order and preprocess them in O(n2 log n) time. This allows us
to retrieve the optimal matching of leaf nodes and then we only have to add cmatch (fi0p+1 , gjp+q+2 ) and
cmatch (fip+q+2 , gj0 p+1 ) to obtain the total cost. Thus, after O(n2 |Σ|2 log n) preprocessing, we can find
the optimal matching by iterating over n2 |Σ|2 possibilities and checking each of them in constant
time.
References
[1] A. Abboud. Hardness for easy problems. In YR-ICALP Satellite Workshop of ICALP, 2014.
[2] A. Abboud, A. Backurs, T. D. Hansen, V. V. Williams, and O. Zamir. Subtree isomorphism
revisited. In SODA, pages 1256–1271, 2016.
3
In the presence of negative-length edges, Klein’s algorithm requires an initial shortest paths tree from some node
on the external face to all other nodes. In our case, computing this initial shortest path tree can easily be done in
O(n2 ) time as our graph is a directed acyclic graph.
21
[3] A. Abboud, A. Backurs, and V. V. Williams. If the current clique algorithms are optimal, so
is Valiant’s parser. In FOCS, pages 98–117, 2015.
[4] A. Abboud, A. Backurs, and V. V. Williams. Tight hardness results for LCS and other sequence
similarity measures. In FOCS, pages 59–78, 2015.
[5] A. Abboud and S. Dahlgaard. Popular conjectures as a barrier for dynamic planar graph
algorithms. In FOCS, pages 476–486, 2016.
[6] A. Abboud, F. Grandoni, and V. V. Williams. Subcubic equivalences between graph centrality
problems, APSP and diameter. In SODA, pages 1681–1697, 2015.
[7] A. Abboud, T. D. Hansen, V. V. Williams, and R. Williams. Simulating branching programs
with edit distance and friends or: a polylog shaved is a lower bound made. In STOC, pages
375–388, 2016.
[8] A. Abboud and K. Lewi. Exact weight subgraphs and the k-SUM conjecture. In ICALP, pages
1–12, 2013.
[9] A. Abboud and V. V. Williams. Popular conjectures imply strong lower bounds for dynamic
problems. In FOCS, pages 434–443, 2014.
[10] A. Abboud, V. V. Williams, and O. Weimann. Consequences of faster alignment of sequences.
In ICALP, pages 39–51, 2014.
[11] A. Abboud, V. V. Williams, and H. Yu. Matching triangles and basing hardness on an extremely
popular conjecture. In STOC, pages 41–50, 2015.
[12] T. Akutsu, D. Fukagawa, and A. Takasu. Approximating tree edit distance through string edit
distance. In ISAAC, volume 4288, pages 90–99, 2006.
[13] N. Alon, Z. Galil, and O. Margalit. On the exponent of the all pairs shortest path problem.
JCSS, 54(2):255–262, 1997.
[14] R. Alur, L. D’Antoni, S. Gulwani, D. Kini, and M. Viswanathan. Automated grading of dfa
constructions. In Proceedings of the Twenty-Third International Joint Conference on Artificial
Intelligence, IJCAI ’13, pages 1976–1982, 2013.
[15] A. Amir, T. M. Chan, M. Lewenstein, and N. Lewenstein. On hardness of jumbled indexing.
In ICALP, pages 114–125, 2014.
[16] A. Apostolico and Z. Galil, editors. Pattern matching algorithms. Oxford University Press,
Oxford, UK, 1997.
[17] T. Aratsu, K. Hirata, and T. Kuboyama. Approximating tree edit distance through string edit
distance for binary tree codes. Fundam. Inform., 101(3):157–171, 2010.
[18] A. Backurs, N. Dikkala, and C. Tzamos. Tight hardness results for maximum weight rectangles.
In ICALP, pages 81:1–81:13, 2016.
[19] A. Backurs and P. Indyk. Edit distance cannot be computed in strongly subquadratic time
(unless SETH is false). In STOC, pages 51–58, 2015.
22
[20] A. Backurs and P. Indyk. Which regular expression patterns are hard to match? In FOCS,
pages 457–466, 2016.
[21] A. Backurs and C. Tzamos. Improving Viterbi is hard: Better runtimes imply faster clique
algorithms. Arxiv 1607.04229, 2016.
[22] J. Bellando and R. Kothari. Region-based modeling and tree edit distance as a basis for gesture
recognition. In Proceedings 10th International Conference on Image Analysis and Processing,
pages 698–703, 1999.
[23] R. Bellman. Dynamic programming. Princeton University Press, 1957.
[24] P. Bille. A survey on tree edit distance and related problems. Theoretical Computer Science,
337(1-3):217–239, 2005.
[25] K. Bringmann. Why walking the dog takes time: Frechet distance has no strongly subquadratic
algorithms unless SETH fails. In FOCS, pages 661–670, 2014.
[26] K. Bringmann. A near-linear pseudopolynomial time algorithm for subset sum. In SODA,
pages 1073–1084, 2017.
[27] K. Bringmann, F. Grandoni, B. Saha, and V. V. Williams. Truly sub-cubic algorithms for
language edit distance and rna-folding via fast bounded-difference min-plus product. In 57th
FOCS, pages 375–384, 2016.
[28] K. Bringmann, A. Grønlund, and K. G. Larsen. A dichotomy for regular expression membership
testing. CoRR, abs/1611.00918, 2016.
[29] K. Bringmann and M. Künnemann. Quadratic conditional lower bounds for string problems
and dynamic time warping. In FOCS, pages 79–97, 2015.
[30] P. Buneman, M. Grohe, and C. Koch. Path queries on compressed XML. In VLDB, pages
141–152, 2003.
[31] S. Chawathe. Comparing hierarchical data in external memory. In VLDB, pages 90–101, 1999.
[32] R. Clifford.
Matrix multiplication and pattern matching under Hamming norm.
http://www.cs.bris.ac.uk/Research/Algorithms/events/BAD09/BAD09/Talks/BAD09Hammingnotes.pdf.
[33] M. de Berg, K. Buchin, B. M. P. Jansen, and G. Woeginger. Fine-grained complexity analysis
of two classic TSP variants. In ICALP, volume 55, pages 5:1–5:14, 2016.
[34] E. Demaine, S. Mozes, B. Rossman, and O. Weimann. An optimal decomposition algorithm
for tree edit distance. ACM Transactions on Algorithms, 6(1):1–19, 2009. Preliminary version
in ICALP 2007.
[35] F. Eisenbrand and F. Grandoni. On the complexity of fixed parameter clique and dominating
set. Theoretical Computer Science, 326(1-3):57–67, 2004.
[36] P. Ferragina, F. Luccio, G. Manzini, and S. Muthukrishnan. Compressing and indexing labeled
trees, with applications. J. ACM, 57:1–33, 2009.
23
[37] D. Gusfield. Algorithms on strings, trees and sequences: computer science and computational
biology. Cambridge University Press, 1997.
[38] C. M. Hoffmann and M. J. O’Donnell. Pattern matching in trees. J. ACM, 29(1):68–95, 1982.
[39] E. Ivkin. Comparison of tree edit distance algorithms. B.Sc. thesis, Charles University in
Prague, 2012.
[40] R. M. Karp. Reducibility among combinatorial problems. In Complexity of Computer Computations, pages 85–103, 1972.
[41] P. N. Klein. Computing the edit-distance between unrooted ordered trees. In ESA, pages
91–102, 1998.
[42] P. N. Klein. Multiple-source shortest paths in planar graphs. In SODA, pages 146–155, 2005.
[43] P. N. Klein, S. Tirthapura, D. Sharvit, and B. B. Kimia. A tree-edit-distance algorithm for
comparing simple, closed shapes. In SODA, pages 696–704, 2000.
[44] D. E. Knuth. Optimum binary search trees. Acta Informatica, 1(1):14–25, 1971.
[45] K. G. Larsen, J. I. Munro, J. S. Nielsen, and S. V. Thankachan. On hardness of several string
indexing problems. Theoretical Computer Science, 582:74–82, 2015.
[46] Miscellaneous Authors. Queries and problems. SIGACT News, 16(3):38–47, 1984.
[47] J. Nešetřil and S. Poljak. On the complexity of the subgraph problem. Commentationes Math.
Universitatis Carolinae, 026(2):415–419, 1985.
[48] M. Pawlik and N. Augsten. Efficient computation of the tree edit distance. ACM Trans.
Database Syst., 40(1):3:1–3:40, Mar. 2015.
[49] J. R. Rico-Juan and L. Micó. Comparison of AESA and LAESA search algorithms using string
and tree-edit-distances. Pattern Recognition Letters, 24(9-10):1417–1426, 2003.
[50] L. Roditty and U. Zwick. On dynamic shortest paths problems. Algorithmica, 61(2):389–401,
2011.
[51] S. Selkow. The tree-to-tree editing problem. Information Processing Letters, 6(6):184–186,
1977.
[52] B. A. Shapiro and K. Zhang. Comparing multiple RNA secondary structures using tree comparisons. Computer Applications in the Biosciences, 6(4):309–318, 1990.
[53] D. Shasha and K. Zhang. Simple fast algorithms for the editing distance between trees and
related problems. SIAM Journal on Computing, 18(6):1245–1262, 1989.
[54] D. Shasha and K. Zhang. Fast algorithms for the unit cost editing distance between trees.
Journal of Algorithms, 11(4):581–621, 1990.
[55] K. Tai. The tree-to-tree correction problem. J. ACM, 26(3):422–433, 1979.
24
[56] G. Valiente. Algorithms on Trees and Graphs. Springer-Verlag, 2002.
[57] R. A. Wagner and M. J. Fischer. The string-to-string correction problem. J. ACM, 21(1):168–
173, 1974.
[58] M. Waterman. Introduction to computational biology: maps, sequences and genomes, Chapters
13, 14. Chapman and Hall, 1995.
[59] R. Williams. A new algorithm for optimal 2-constraint satisfaction and its implications. Theor.
Comput. Sci., 348(2-3):357–365, 2005.
[60] R. Williams. Faster all-pairs shortest paths via circuit complexity. In STOC, pages 664–673,
2014.
[61] V. V. Williams and R. Williams. Subcubic equivalences between path, matrix and triangle
problems. In FOCS, pages 645–654, 2010.
[62] V. V. Williams and R. Williams. Finding, minimizing, and counting weighted subgraphs. SIAM
J. Comput., 42(3):831–854, 2013.
[63] G. J. Woeginger. Space and time complexity of exact algorithms: Some open problems. In
IWPEC, pages 281–290, 2004.
[64] G. J. Woeginger. Open problems around exact algorithms. Discr. Appl. Math., 156(3):397–405,
2008.
[65] X. Yao, B. V. Durme, C. Callison-Burch, and P. Clark. Answer extraction as sequence tagging
with tree edit distance. In HLT-NAACL 2013, pages 858–867, 2013.
[66] K. Zhang. Algorithms for the constrained editing distance between ordered labeled trees and
related problems. Pattern Recognition, 28(3):463–474, 1995.
25
| 8cs.DS
|
1
Sprout: A functional caching approach to minimize
service latency in erasure-coded storage
arXiv:1609.09827v3 [cs.DC] 21 Jul 2017
Vaneet Aggarwal, Yih-Farn R. Chen, Tian Lan, and Yu Xiang
Abstract—Modern distributed storage systems often use erasure codes to protect against disk and node failures to increase
reliability, while trying to meet the latency requirements of the
applications and clients. Storage systems may have caches at
the proxy or client ends in order to reduce the latency. In
this paper, we consider a novel caching framework with erasure
code called functional caching. Functional Caching involves using
erasure-coded chunks in the cache such that the code formed by
the chunks in storage nodes and cache combined are maximaldistance-separable (MDS) erasure codes. Based on the arrival
rates of different files, placement of file chunks on the servers, and
service time distribution of storage servers, an optimal functional
caching placement and the access probabilities of the file request
from different disks are considered. The proposed algorithm
gives significant latency improvement in both simulations and a
prototyped solution in an open-source, cloud storage deployment.
I. I NTRODUCTION
Erasure coding has seen itself quickly emerged as a promising technique to reduce the storage cost for a given reliability
as compared to fully-replicated systems [2, 3]. It has been
widely adopted in modern storage systems by companies like
Facebook [4], Microsoft [5] and Google [6]. In these systems,
the rapid growth of data traffic such as those generated by
online video streaming, Big Data analytics, social networking
and E-commerce activities has put a significant burden on
the underlying networks of datacenter storage systems. Many
researchers have begun to focus on latency analysis in erasure
coded storage systems [7–14] and to investigate algorithms
for joint latency optimization and resource management [12,
14–18].
Historically, a key solution to relieve this traffic burden is
caching [19]. By storing chunks of popular data at different
locations closer to end-users, caching can greatly reduce
congestion in the network and improve service delay for
processing file requests. For example, Figure 1 shows a typical
video storage architecture with video proxies and multiple
video clients. It is very common for 20% of the video content
to be accessed 80% of the time, so caching popular content at
The author names are written in an alphabetical order. V. Aggarwal is with
the School of IE, Purdue University, West Lafayette, IN 47907, USA, email:
[email protected]. Y. R. Chen and Y. Xiang are with AT&T Labs-Research,
Bedminster, NJ 07921, USA, email: {chen,yxiang}@research.att.com. T. Lan
is with the Department of ECE, George Washington University, DC 20052,
USA, email: [email protected]
This work was supported in part by the National. Science Foundation under
grant CNS-1618335. This work was presented in part in Proc. IEEE ICDCS
2016 [1].
proxies significantly reduces the overall latency on the client
side.
Fig. 1. Caching proxies for video storage to reduce the latency of client
accesses for popular Video-on-Demand content
However, caching for datacenters where the files are encoded with erasure codes gives rise to new challenges. The
current results fall short of addressing the impact of erasure
coding on latency and thus fail to providing insights on the optimal caching policy. First, using an (n, k) maximum-distanceseparable (MDS) erasure code, a file is encoded into n chunks
and can be recovered from any subset of k distinct chunks.
Thus, file access latency in such a system is determined by
the delay to access file chunks on hot storage nodes with
the slowest performance. Significant latency reduction can
be achieved by caching only a few hot chunks of each file
(and therefore alleviating system performance bottlenecks),
whereas caching additional chunks or even complete files (e.g.,
[20–23]) only has diminishing benefits. It is an open problem
to design a caching policy that optimally apportions limited
cache capacity among all files in an erasure coded storage to
minimize overall access latency.
More importantly, caching the most popular data chunks is
often optimal because the cache-miss rate and the resulting
network load are proportional to each other. However, this
may not be true for an erasure-coded storage, where cached
chunks need not be identical to the chunks transferred from
storage nodes. More precisely, leveraging the existing erasure
coding, a function of the data chunks can be computed and
cached, so that the constructed new chunks (i.e., d), along
with the existing chunks, satisfy the property of being a new
MDS code. It effectively expands the existing (n, k) code to
an (n + d, k), which leads to much lower access latency [12].
In this paper, we propose a new functional caching approach
called Sprout that can efficiently capitalize on existing file
coding in erasure-coded storage systems. In contrast to exact
caching that stores d chunks identical to original copies, our
functional caching approach forms d new data chunks, which
together with the existing n chunks satisfy the property of
2
being an (n + d, k) MDS code. Thus, the file can now be
recovered from any k out of n + d chunks (rather than k
out of n under exact caching), effectively extending coding
redundancy, as well as system diversity for scheduling file
access requests. The proposed functional caching approach
saves latency due to more flexibility to obtain k − d chunks
from the storage system at a very minimal additional computational cost of creating the coded cached chunks. To the best
of our knowledge, this is the first work studying functional
caching for erasure-coded storage and proposing an analytical
framework for cache optimization.
Most of the caching strategies in today’s data centers cache
complete files [24–26]. This flexibility without erasure coding
has been recently explored in [27]. Partial number of chunks in
cache give more flexibility as compared to caching the entire
file. This is because the connection from some servers may
be better than others and caching a smaller number of chunks
can aid avoiding the bottleneck. The proposed optimization
includes as a special case complete file caching and thus the
performance can be no worse than caching the entire file in the
cache. This paper gives a novel approach, functional caching,
which is an efficient approach to cache partial chunks since
any k−d of the n servers can be used thus increasing flexibility
as compared to using any k − d of the remaining n − d servers
when d exact chunks are copied in the cache. This additional
flexibility shows that the latency with functional caching is
no higher than the strategy where part of the chunks on the
servers are cached as such.
While quantifying service latency in erasure-coded storage
systems is an open problem, we generalize previous results on
probabilistic scheduling policy [11, 12] that distributes the file
requests to cache and storage nodes with optimized probabilities, and derive a closed-form upper bound on mean service
latency for the proposed functional caching approach. The
latency bound is obtained using order-statistic [28] analysis
and it works on erasure-coded storage systems with arbitrary
cache size and data chunk placement.
This analytical latency model for functional caching enables
us to formulate a cache-content optimization problem. This
problem is an integer optimization problem, which is very
difficult to solve. Towards this end, for given data chunk
placement and file request arrival rates, we propose a heuristic
algorithm that iteratively identifies files whose service latency
benefits most from caching and constructs new functional data
chunks until the cache is filled up. The algorithm can be
efficiently computed to allow online cache optimization and
management with time-varying arrival rates.
The proposed algorithm is an iterative algorithm, which
converges within a few iterations in our conducted experiments
and it was validated by the numerical results. For 1000 files,
we find that the algorithm converges within 20 iterations. The
file latency decreases as a convex function as the cache size
increases thus showing diminishing returns for the increase
in cache size. We also find that it is suboptimal in general to
have all k chunks of an (n, k) coded file in the cache. Further,
the cache placement depends heavily on the file arrival rates,
storage server service time distributions, as well as the content
placement on the files. If a high arrival-rate file is placed on
servers which have a less overall load, this file may not have
any contribution in the cache. Thus, the proposed algorithm
accounts for all the system parameters to effectively find the
cache placement. The proposed algorithm is prototyped using
Ceph, an open-source erasure-coded storage system [29] and
tested on a real-world storage testbed with an emulation of
real storage workload. We find that caching helps improve the
latency significantly.
The key contributions of our paper include:
•
•
•
•
•
We propose a new functional caching scheme that leverages existing file coding in erasure-coded storage systems, and quantify its service latency through probabilistic request scheduling.
Based on the service latency analysis, an optimization
problem is formulated which minimizes the average latency of the files. This problem has integer constraints
due to the integer number of chunks in the cache.
An iterative algorithm is developed to optimize cache
content. The proposed algorithm takes file placement, service time distributions, and file arrival rates into account
to find the cache placement which optimizes the service
latency.
The simulation results show that the proposed algorithm
converges within a few iterations.
The prototype implementation of the functional caching
scheme and the cache optimization algorithm using Ceph
are used to validate significant latency reduction on a realworld storage testbed. As compared to the Ceph’s LRU
(least recently used) caching algorithm, the algorithm
proposed in this paper reduces latency by 24.93% on
an average for all tested workloads in the prototype
implementation.
The remainder of this paper is organized as follows. Section II
provides related work for this paper. In Section III, we describe
the system model used in the paper with a description of
functional caching. Section IV formulates the cache optimization problem and develops an iterative algorithmic solution.
Prototype and evaluation results are included in Section V.
Section VI concludes the paper.
II. R ELATED W ORK
Quantifying exact latency for an erasure-coded storage
system is an open problem. Recently, there has been a number
of attempts at finding latency bounds for an erasure-coded
storage system [7–12, 15, 16, 30–32]. In this paper, we utilize
the probabilistic scheduling policy developed in [11, 12] and
analyze the impact of caching on the service latency of
erasure-coded storage. Even though exact characterization of
latency is open, probabilistic scheduling has been shown to
be optimal for latency tail index, when the file size is Pareto
distributed [33].
3
Caches are a critical resource in data centers; however, there
is little work on caching for erasure coded distributed storage.
The problem of online cache management (i.e., decisions for
evicting a data chunk currently in the cache to make room for
a new chunk) has been studied for networks with distributed
caches [34, 35]. Cache replacement strategy called LRU (Least
Recently Used) is widely used in managing buffer cache due to
its simplicity [36–40]. Recently, a steady-state characterization
of various cache policies is developed in [41], and new
coded caching schemes to enable multicast opportunities for
bandwidth minimization are proposed in [19, 42]. Recently,
caching in erasure-coded storage has been studied [43] where
the cache stores the files in their entirety. In contrast, this paper
allows for partial chunks, which are functionally different from
those stored on the servers.
Much of the work on caching in data centers is focused
on specialized application caches, such as Facebook’s photoserving stack [44], Facebook’s social graph store [45], memcached [46], or explicit cloud caching services [47, 48]. However, in the presence of coding, new challenges arise. First, the
chunks are distributed over multiple servers and a part of the
chunks can be in the cache. Thus, it is not necessary for the
complete file to be in the cache. Second, the latency calculation
for a file depends on the placement of the files and the request
of the files from k out of n servers. In this paper, we deal with
these challenges to consider a novel caching strategy. To the
best of our knowledge, this is the first work on caching with
erasure coded files on distributed storage servers accounting
for the latency in file retrievals, based on the estimated arrival
rates.
Coded caching for a single server with multi-cast link to
different users has been considered in [49–52]. This does not
account for multiple distributed storage servers and latency to
get the content. An extension of the approach to distributed
storage systems is considered recently in [53, 54], where
multiple cache-enabled clients connected to multiple servers
through an intermediate network. However, the impact of
coding on the servers, and limited service rate of different
servers is not taken into account. The key idea in this set
of works uses a coded version of different files in the cache
which helps in the case when users request different files with
certain probabilities. The gain of the approach is due to the
model where a message from the server can be received at
multiple nodes and thus combined with coded content in the
cache, one chunk from the server can help give a chunk for
different files at different clients. In this paper, we do not have
multicast links to different clients and thus coding across files
in the cache is not used.
The functional repair was introduced in [3] for repairing
a failed chunk with a new chunk such that the storage
code satisfies the same properties even after repair. Thus,
the replaced content can be different. In this paper, we use
functional caching to have chunks in the cache such that the
file can be recovered from any of k chunks from a combination
of disk and cache contents where (n, k) code is used for the
files in the storage system.
III. S YSTEM M ODEL
We consider a distributed storage system consisting of m
heterogeneous storage nodes, denoted by M = {1, 2, . . . , m}.
To distributively store a set of r files, indexed by i = 1, . . . , r,
we partition each file i into ki fixed-size chunks1 and then
encode it using an (ni , ki ) MDS erasure code to generate ni
distinct chunks of the same size for file i. The encoded chunks
are stored on the disks of ni distinct storage nodes. A set Si
of storage nodes, satisfying Si ⊆ M and ni = |Si | is used
to store file i. Therefore, each chunk is placed on a different
node to provide high reliability in the event of node or network
failures. The use of (ni , ki ) MDS erasure code allows the file
to be reconstructed from any subset of ki -out-of-ni chunks,
whereas it also introduces a redundancy factor of ni /ki .
The files are accessed (to be read in their entirety) by
compute servers located in the same datacenter. A networked
cache of size C is available at each compute server to store a
limited number of chunks of the r files in its cache memory.
File access requests are modeled by a non-homogenous Poisson process. We make the assumption of time-scale separation,
such that system service time is divided into multiple bins,
each with different request arrival rates, while the arrival rates
within each bin remain stationary. This model allows us to
analyze cache placement and latency in steady-state for each
time bin, and by varying arrival rates for different time bins,
to also take into account time-varying service rates during
busy and off-peak hours. Let λi,j,t be the arrival rate of file-i
requests at compute server j in time bin t. These arrival rates
λi,j,t can be estimated using online predictive models [58] or
a simple sliding-window-based method, which continuously
measures the average request arrival and introduces a new
time bin if the arrival rates vary sufficiently. It is easy to see
that estimating arrival rates using a small window is prone to
the dynamics of stochastic request arrivals. However, a large
window size introduces a low-pass filtering effect, causing
higher delay and insensitivity in detecting rate changes. We
also note that more advanced methods such as [59, 60] can
be used to estimate arrival rate changes in non-homogeneous
processes. Since each cache serves a single compute server, we
consider a separate optimization for each cache and suppress
server index j in the notations. Let di ≤ ki (chunks) be the
size of cache memory allocated to storing file i chunks. These
chunks in the cache memory can be both prefetched in an
offline fashion during a placement phase [19] (during hours
of low workload) and updated on the fly when a file i request
is processed by the system.
Functional Caching. Under functional caching, di new coded
data chunks of file i are constructed and cached, so that along
with the existing ni chunks satisfy the property of being an
1 While we make the assumption of fixed chunk size here to simplify the
problem formulation, all results in this paper can be easily extended to variable
chunk sizes. Nevertheless, fixed chunk sizes are indeed used by many existing
storage systems [55–57].
4
(ni + di , ki ) MDS code. In this paper, we use Reed-Solomon
codes [61] to generate MDS codes for arbitrary (ni + ki , ki ).
Thus a subset of the encoded chunks generated using these
codes satisfy the desired property for functional caching. More
precisely, for given erasure coding and chunk placement on
storage nodes and cache, a request to access file i can be
processed using di cached chunks in conjunction with ki − di
chunks on distinct storage nodes. After each file request arrives
at the storage system, we model this by treating the file request
as a batch of ki − di chunk requests that are forwarded to
appropriate storage nodes, as well as di chunk requests that are
processed by the cache. Each storage node buffers requests in a
common queue of infinite capacity and process them in a FIFO
manner. The file request is served when all ki chunk requests
are processed. Further, we consider chunk service time Xj
of node j with arbitrary distributions, whose statistics can
be inferred from existing work on network delay [62, 63] and
file-size distribution [64, 65].
Compute
Server
C1
C2
Compute
Server
F1
Compute
Server
F2
…
…
…
F3
F1
F2
F6
F4
F5
Storage Nodes
Functional Cache
C1=F1+2F2+3F3+4F4+5F5
C2=F1+2F2+4F3+8F4+16F5
Storage Nodes
Storage Nodes
Exact Cache
C1=F1
C2=F2
Fig. 2. An illustration of functional caching and exact caching in an erasurecoded storage system with one file using a (6, 5) erasure code. The file is
split into ki = 5 chunks, denoted by A1 , A2 , A3 , A4 , A5 , and then linearly
encoded to generate ni = 6 coded chunks F1 = A1 , F2 = A2 , F3 = A3 ,
F4 = A4 , F5 = A5 , and F6 = A1 + A2 + A3 + A4 + A5 .
An Illustrative Example. Consider a datacenter storing a
single file using a (6, 5) MDS code. The file is split into
ki = 5 chunks, denoted by A1 , A2 , A3 , A4 , A5 , and then
linearly encoded to generate ni = 6 coded chunks F1 = A1 ,
F2 = A2 , F3 = A3 , F4 = A4 , F5 = A5 and F6 =
A1 + A2 + A3 + A4 + A5 in a finite field of order at-least
5. In this example, we compare the latency of 2 different
cache policies: (i) Exact caching in Rack 3 that stores chunks
F1 and F2 , and (ii) Functional caching in Rack 1 that stores
2 coded chunks C1 and C2 .
Due to the use of (6, 5) MDS code, the file can be
reconstructed using any 5 out of 6 chunks. Two compute
servers in the datacenter access this file and each is equipped
with a cache of size C = 2 chunks as depicted in Figure 2.
The compute server on the right employs an exact caching
scheme and stores chunks F1 , F2 in the cache memory. Thus,
3 out of 4 remaining chunks (i.e., F3 , F4 , F5 or F6 ) must
be retrieved to access the file. Chunks F1 , F2 and thus their
host nodes will not be selected for processing the requests
even if the servers in Rack 1 are least congested. Since file
access latency is determined by the time required to retrieve all
5 selected chunks, caching F1 , F2 may not necessarily reduce
file access latency if other chunks in Rack 2 and 3 are currently
the performance bottleneck.
We show that lower latency can always be achieved in functional caching by exploiting the erasure code structure. More
specifically, the compute server on the left generates di = 2
new coded chunks, i.e., C1 = A1 + 2A2 + 3A3 + 4A4 + 5A5
and C2 = A1 + 2A2 + 4A3 + 8A4 + 16A5 , and saves them
in its cache memory. It is easy to see that chunks F1 , . . . , F6
and C1 , C2 now form a (8, 5) erasure code. Thus, the file
can be retrieved by accessing C1 , C2 in the cache together
with any 3 out of 6 chunks from F1 , . . . , F6 . This allows
an optimal request scheduling mechanism to select the least
busy chunks/nodes among all 5 possible candidates in the
system so that the service latency is determined by the best 3
storage nodes with minimum queuing delay. In order words,
the coded chunks C1 , C2 can always be used to replace the
slowest chunks required under exact caching policy, resulting
in smaller latency. Consider file access latency, which is
determined by the time required to access any 5 chunks on
the storage servers. Under an exact caching of file copies F1
and F2 , any request to access the file must select 3 other
files chunks hosted by storage servers. Thus, access latency in
this case is determined by the minimum time to retrieve any
3 chunks out of F3 , F4 , F5 and F6 . In contrast, under our
proposed functional caching policy, given the availability of
coded chunks C1 and C2 in cache, any 2 chunks from F1 , F2 ,
F3 , F4 , F5 and F6 is sufficient for recovery, which leads to
smaller latency if nodes storing F1 and F2 are less congested
and faster to access. This approach effectively extends to an
(8, 5) erasure code and guarantees lower access latency and
also enables better load balancing due to a higher degree of
flexibility in request scheduling and chunk selection.
In order to have an (n, k) coded file in the storage server,
we can construct chunks by using an (n + k, k) MDS code,
where n chunks are stored in the storage server. The remaining
k out of the n + k coded chunks are assigned to be in part in
the cache based on the contents of the file in the cache. Thus,
irrespective of the value of d ≤ k, we ascertain that (n + d, k)
code, formed with n coded chunks in the storage server and
d coded chunks in the cache will be MDS.
In the proposed system, the placement in each time-bin is
decided based on the predicted arrival rates in the time bin.
The time bin can either be a fixed time or dynamic based
on a significant change of the predicted arrival rates. At the
start of the time-bin, new cache placement is found using the
optimized algorithm. For each file that has a lower number
of chunks in the new time bin, the decreased contents are
removed from the cache. For the files for which the cache
contents increase in a time bin, we wait for the file to be
5
accessed. When the file is accessed, the file contents are
gathered and the required new chunks are generated to be
placed in the cache. Thus, the change of cache content does not
place any additional network overhead and the cache contents
of a file are added only when it is first accessed in the new time
bin. This process can be further improved improving latency
till convergence to the new cache content in the new time bin
by not letting all the chunks which have to be removed all
simultaneously but removing as needed based on the added
chunks.
IV. O PTIMIZED D ISTRIBUTED S TORAGE WITH C ACHE
In this section, we quantify the mean service latency for file
access with functional caching. The result enables us to formulate a cache optimization problem for latency minimization
and develop an efficient algorithm solution.
moment E[X2j ] = Γ2j , and third moment E[X3j ] = Γ̂3j . These
statistics can be readily inferred from existing work on network
delay [62, 63] and file-size distribution [64, 65]. Following
[12], an upper bound on the expected latency is given as
follows.
Lemma 1. The expected latency T̄i,t of file i in time-bin t
under probabilistic scheduling is upper bounded by Ūi,t , given
by
X πi,j,t
Ūi,t = min zi,t +
(E[Qj,t ] − zi,t )
zi,t ∈R
2
j∈Si,t
X πi,j,t q
+
(E[Qj,t ] − zi,t )2 + Var[Qj,t ] (2)
,
2
j∈Si,t
where
A. Formulation of Optimization
At time t, we consider the cache optimization problem,
which decides the optimal number di,t of file-i chunks to
store
Pr in the cache memory, satisfying cache capacity constraint
i=1 di,t ≤ C, in order to minimize mean service latency
of all files. Under functional caching, each file-i request is
served by accessing di,t chunks in the cache, along with
ki −di,t distinct chunks that are selected from ni storage nodes.
Thus, the latency to access file i under functional caching is
determined by the maximum processing (queuing) delay of the
selected ki −di,t storage nodes. Quantifying service latency in
such erasure-coded system is an open problem. In this paper,
we use probabilistic scheduling proposed in [12] to derive an
upper bound on the average file latency.
The key idea is to forward each file-i request to a set
of ki − di,t storage nodes (denoted by Ai,t ⊆ Si ) with
some predetermined probabilities {πi,j,t ∈ [0, 1], ∀i, j, t} for
j ∈ Ai,t . Each node then manages a local queue and process
chunk requests with service rate µj . While the local queues
are not independent due to coupled request arrivals, we can
leverage order statistic analysis to derive an upper bound of
mean service latency in closed-form [12]. The result is then
optimized over probabilities πi,j,t to obtain the tightest bound.
Let Qj,t be the (random) waiting time a chunk request spends
in the queue of node j in time-bin t. Using the functional
caching approach, requests of file i see mean latency T̄i,t given
by
T̄i,t = E EAi,t max {Qj,t } ,
(1)
j∈Ai,t
where the first expectation is taken over system queuing
dynamics and the second expectation is taken over random
dispatch decisions Ai,t . We note that queuing delay Qj,t ’s
are dependent due to coupled request arrivals. Therefore, an
exact queuing analysis of the system is intractable. We use the
technique in [12] to derive an upper bound of (1).
We denote Xj as the service time per chunk at node
j, which has an arbitrary distribution satisfying finite mean
E[Xj ] = 1/µj , variance E[X2j ] − E[Xj ]2 = σj2 , second
E[Qj,t ] =
Var[Qj,t ] = σj2 +
Λj,t Γ2j
1
,
+
µj
2(1 − ρj,t )
Λj,t Γ̂3j
Λ2j,t Γ4j
+
,
3(1 − ρj,t ) 4(1 − ρj,t )2
(3)
(4)
where ρP
j,t = Λj,t /µj is the request intensity at node j, and
r
Λj,t = i=1 λi,t πi,j,t is the mean arrival rate at node j. The
bound is tight in the sense that there exists a distribution of
Qj,t such that (2) is satisfied with exact equality.
Proof. The proof follows on the same lines as in [12]. A brief
detail into the proof idea is provided in Appendix B ( where the
time-index t is omitted) to make the paper more self-contained.
Remark 1. The upper bound of Lemma 1 can be improved if
higher ordered moment statistics of service time are known,
though we do not consider them in this work due to tractability
in latency optimization. This bound, even without the additional moments, has been shown to outperform existing bounds
and is indeed very close the actual latency [12].
Under the assumption of time-scale separation, each time
bin has stationary request arrival rates and is long enough
for cache content to reach a steady state. We consider the
cache optimization in a single time bin to minimize weighted
mean latency quantified by Lemma 1. This optimization will
be performed repeatedly at the beginning of each time bin
to address time-varying request arrival rates during busy and
off-peak hours.
We now formulate the cache optimization in a single timebin. The optimization is over cache content placement di,t ,
scheduling probabilities πi,j,t
Pr, and auxiliary variable zi,t in
the upper bound. Let λ̂t = i=1 λi,t be the total arrival rate,
so λi,t /λ̂ is the fractionP
of file i requests, and average latency
r
of all files is given by i=1 (λi,t /λ̂t )T̄i,t . Our objective is to
minimize an average latency objective, i.e.,
6
min
r
X
λi,t
i=1
s.t.
λ̂t
Ūit
(5)
(2), (3), (4),
m
X
πi,j,t = ki − di,t , πi,j,t , di,t ≥ 0,
j=1
r
X
di,t ≤ C, πi,j,t = 0 for j ∈
/ Si , πi,j,t ≤ 1,
i=1
zi,t ≥ 0, di,t ∈ Z.
πi,j,t , di,t , zi,t ∀i, j, t.
Pm
Here the constraints j=1 πi,j,t = ki − di,t and πi,j,t ≤ 1
ensure that ki − di,t distinct storage nodes (along with di,t
chunks in the cache) are selected to process each file request,
following probabilistic scheduling in [12]. Clearly, storage
nodes without desired chunks cannot be selected, i.e., πi,j,t =
0Pfor j ∈
/ Si . Finally, the cache has a capacity constraint
r
d
i=1 i,t ≤ C.
Solving the cache optimization gives us the optimal cache
content placement and scheduling policy to minimize file
access latency. We note that the constraint zi,t ≥ 0 is not
needed if none of the files is completely in the cache. However,
the latency bound does not hold if the file is completely in
the cache since in that case the bound is zi,t in the above
expression. In order to avoid having indicators representing
the constraint on zi,t = 0 if the file is in the cache, we only
consider zi,t ≥ 0 making the latency bound hold irrespective
of the number of chunks in the cache. This problem can be
rewritten as follows.
Distributed Storage with Caching:
var.
min
r
X
r
X
λi,t zi,t /λ̂t
i=1
+
r X
m
X
λi,t πi,j,t h
i=1 j=1
2λ̂t
s.t.
(7), (8), (9), zi,t ≥ 0
var.
zi,t , ∀i.
Xi,j,t =
2λ̂t
Xi,j,t +
q
i
2
Xi,j,t
+ Yj,t (6)
i
2
Xi,j,t
+ Yj,t (12)
min
r
X
λi,t zi,t /λ̂t
i=1
Λj,t Γ2j
1
− zi,t , ∀j
+
µj
2(1 − ρj,t )
(7)
Λj,t Γ̂3j
Λ2j,t Γ4j
Yj,t =
+
+
, ∀j
3(1 − ρj,t ) 4(1 − ρj,t )2
r
X
ρj,t = Λj,t /µj < 1, Λj,t =
πi,j,t λi,t ∀j
σj2
+
r X
m
X
λi,t πi,j,t h
i=1 j=1
(8)
s.t.
(9)
2λ̂t
πi,j,t = 0 ∀j ∈
/ Si ,
r
X
πi,j,t = 0 ∀j ∈
/ Si ,
di,t ≤ C, di,t ∈ Z
(11)
i=1
zi,t , di,t , πi,j,t , ∀i, j.
B. Proposed Algorithm
The proposed cache optimization problem in (6)-(11) is
an integer optimization problem, since the number di,t of
functional chunks in the cache must be integers. To solve this
problem, we propose a heuristic algorithm, which iteratively
r
X
i=1
var.
+
q
i
2
Xi,j,t
+ Yj,t (13)
(14)
j=1
πi,j,t = ki − di,t , πi,j,t ∈ [0, 1], zi,t ≥ 0,(10)
j=1
Xi,j,t +
(7), (8), (9), πi,j,t ∈ [0, 1],
m
X
KL,i,t ≤
πi,j,t ≤ kU,i,t ,
i=1
var.
q
We define the problem Prob Π for given zi,t , kU,i,t , kL,u,t
r X
m
X
λi,t πi,j,t h
m
X
Xi,j,t +
as
i=1 j=1
s.t.
min
λi,t zi,t /λ̂t
i=1
+
identifies the files that benefit most from caching, and constructs/stores funtional chunks into cache memory accordingly.
We first note that the variable di,t can be absorbed into
scheduling P
decision πi,j,t because of the equality constraint
m
di,t = ki − j=1 πi,j,t . Thus, there are two set of variables zi,t , and πi,j,t - we need to consider. It is easy to show that the
objective function is convex in both
Pm these variables, however
there is an integer constraint on j=1 πi,j,t due to the integer
requirement of di,t .
The algorithm employs an alternating minimization over
two dimensions - the first decides on zi,t given πi,j,t , and the
second decides on πi,j,t given zi,t . The first problem is convex,
and can be easily solved by gradient descent. However, the
second problem has integer constraint. In order to deal with
this, we first remove integer constraint to solve the problem.
Then, a certain percentage of files whose fractional part of
content accessed from the disk is highest are added a part in the
disk to make the part in the diskP
as integers. The optimization
m
over πi,j,t keeps running until j=1 πi,j,t for all files is an
integer. In particular, we derive the two sub-problems that need
to be solved as follows.
We define the problem Prob Z for given πi,j,t as
(ki −
m
X
πi,j,t ) ≤ C, (15)
j=1
πi,j,t , ∀i, j.
The problem Prob Z optimizes over zi,t given πi,j,t . This
problem is convex with only one linear constraint zi,t ≥ 0.
In order to solve this problem, we can use standard gradient
descent, with making zi,t as zero if the solution is negative in
each iteration. The problem Prob Π assumes that the number
of total chunks of a file i accessed from the disk is between
kL,i,t and kU,i,t . As we decide the number of chunks in the
cache for each file, these two bounds will become equal. This
problem is also convex, and can be solved using projected
7
Initialize c = 0 and feasible (zi,t , πi,j,t ∀i, j)
Compute current objective value B (0)
Initialize c = 0 and feasible (zi,t , πi,j,t ∀i, j)
Compute current objective value B (0)
do
Solve Convex Problem Prob Z to get zi,t for given
πi,j,t for all i
Set kL,i,t = 0, kU,i,t = ki
do
Solve Convex Problem Prob Π to get πi,j,t for
given zi,t , kL,i,t , kU,i,t for all i, j
Pm
Let i1 = arg max (fractional
Pm part of j=1 πi,j,t )
kL,i1P
= ceil( j=1 πi,j,t )
,t = kU,i1 ,t P
r
m
while i=1 f rac( j=1 πi,j,t ) > 0
Compute new objective value B (c+1) , Update c = c+1
while B (c) − B (c−1) >
gradient descent. With algorithmic solution to these two subproblems, the algorithm for Distributed Storage with Caching
is given in Algorithm 1.
Lemma 2. The 2 sub-problems, Prob Z and Prob Π, are both
convex.
Remark 2. The proposed algorithm 1 is guaranteed to converge. Since the inner loop that solves convex problem Prob Π
iteratively determines the placement of chunks in cache oneby-one, until the entire cache space is filled up, it runs at most
r iterations. The outer loop that solves convex optimization
Prob Z generates a monotonically decreasing sequence of
objective values and is also guaranteed to converge within
finite number of iterations.
We note that the inner do-while logic to deal with integer
optimization runs at most r times. Since r may be large, rather
than choosing one index ii , we choose a ceiling of certain
fraction of file indices among those which have fractional
content in the cache. This makes the loop run in O(log r).
Thus, each outer loop runs O(log r) convex problems. The
algorithm will be solved repeatedly for each time bin to guide
the update of cache content for service latency minimization.
0.000104/sec for every five out of the 1000 files of each size.
It gives an aggregate arrival rate of all files to be 0.1416/sec.
The inverse of mean service times for the 12 servers are set
based on measurements of real service time in a distributed
storage system, which we obtained from a previous work [12],
and they are {0.1, 0.1, 0.1, 0.0909, 0.0909, 0.0667, 0.0667,
0.0769, 0.0769, 0.0588, 0.0588} for the 12 storage servers
respectively. The placement of files on the servers is chosen at
random, unless explicitly specified. The major objective of our
simulation is to validate our latency analysis of erasure-coded
storage with caching in those areas that are hard to implement
in a real test-bed or are difficult to measure in experiments.
B. Numerical Results
Convergence of Algorithm: We implemented our cache
optimization algorithm using MOSEK, a commercial optimization solver, to project the gradient descent solution to the
feasible space for Prob Π. For 12 distributed storage servers
in our testbed, Figure 3 demonstrates the convergence of our
algorithm in one time-bin, which optimizes the average latency
of all files over request scheduling πi,j,t and cache placement
di,t for each file. We assume 1000 files, each using a (7,4)
erasure code and of size 100 MB, divided evenly into five
groups with the arrival rates of the five groups as mentioned
above. The convergence of the proposed algorithm is depicted
in Fig. 3 for cache size C×25 MB. For C = 4000, four chunks
of each file can be in the cache. A random initialization is
chosen for C = 100, while the converged solution for C = 100
is taken as initialization for C = 200 and so on. We note that
the algorithm converges within a few iterations, and converges
in less than 20 iterations with a threshold of 0.01 on latency
for all cache size values in Figure 3.
30
C=100
C=200
C=300
C=400
C=500
C=600
C=700
28
26
24
Latency
Algorithm 1 Our Proposed Algorithm for Distributed Storage with Caching
22
20
18
16
V. N UMERICAL R ESULTS , I MPLEMENTATION AND
E VALUATION
In this section, we evaluate our proposed algorithm for
functional cache optimization, through both simulation and
implementation in an open-source, distributed filesystem.
A. Numerical Results Setup
We simulated our algorithm in a cluster of m = 12 storage
servers, holding r = 1000 files of size 100 MB each using a
(7,4) erasure code. Unless stated otherwise, cache size remains
as 500 times of the chunk size (i.e., 500 times of 25 MB).
The arrival rate for each file is set at a default value of
λi = 0.000156/sec, 0.000156/sec, 0.000125/sec, 0.000167/sec,
14
12
10
0
5
10
15
20
25
30
Iterations
Fig. 3. Convergence of the proposed algorithm for a system with r = 1000
files each of size 100MB and using a cache size of C × 25M. The algorithm
efficiently computes a solution in less than 20 iterations with a low tolerance
= 0.01 for different cache sizes. A random initialization is chosen for C =
100, while the converged solution for C = 100 is taken as initialization for C
= 200 and so on.
Impact of Cache Size: To validate that our proposed
algorithm can effectively update file content placement in
cache to optimize average latency in the storage system, we
plot average latency for r = 1000 files of size 100MB in
8
one time bin with designated request arrival rate for each file,
while cache size varies from 0 to 4000 times of the chunksize. Fig 4 shows that the average latency decreases as cache
size increases, where average latency is 23 sec when no file
has any content in the cache, and is 0 sec when the cache size
is 4000 chunk-size since 4 chunks of each file can be in the
cache. We note that the latency is a convex decreasing function
of the cache size, depicting that our algorithm is effectively
updating content in the cache and showing diminishing returns
in a decrease of latency after reaching certain cache size.
Evolution of Cache Content Placement: We validate
that our algorithm minimizes service latency by optimizing
the cache content with respect to request arrival rates (the
algorithm is minimizing latency bound Ūi,t , and λi and Λj
is playing an important role in Ūi,t ), i.e., placing more file
chunks that are frequently accessed (with a higher arrival rate)
in the cache. We design a simulation which consists of 10 files,
of size 100MB, and run our algorithm for three time bins,
where each time bin has a different request arrival rate for the
r = 10 files as given in Table I. The arrival rate changes are
detected through a sliding-window-based mechanism, which
continuously measures the average request arrival rates, and
when the difference of arrival rates of an object between two
different time windows is greater than a threshold, a new time
bin is introduced to trigger a new round of optimization. The
arrows in Table I indicate the increase or decrease of the file
arrival rate in the consecutive time bins. In this simulation, we
plot the cache placement for the 10 files in steady state (after
the algorithm converges) for each time bin. From Figure 5,
we see that in the first time bin, as file 4 and file 9 had the
highest request arrival rates, they had the highest numbers of
chunks in the cache in the steady state. In the second time
bin, as the arrival rates of files 4 and 9 decreased, and those
of files 5 and 10 increased, now the files that had the same
and higher request arrival rate were 1, 2, 6, and 7. Thus Fig
5 shows that in the second time bin these four files dominates
the cache’s steady state. In the third time bin, the files that had
the highest arrival rate became files 2, 7, 4, and 9, and Fig.
5 also shows that cache is mainly holding the contents of file
2, 7, and 9. Due to different service rates of the servers and
the randomized placement of contents in the servers, it is not
always the case that files with the highest arrival rate need to
be placed completely in the cache. Rather, it is important to
determine how many chunks of each file should be placed in
the cache.
Impact of Content Placement and Arrival Rate: We note
that in addition to the arrival rates of the files, placement of
content on the storage nodes influence the decision of cache
placement. We consider 10 files, of size 100M, using (7,4)
erasure code are placed on 12 servers as described in the
simulation set up. The first three files are placed on the first
seven servers while the rest of the files are placed on the
last seven servers. Note that servers 6 and 7 host chunks for
all files. For the arrival rates of the files, we fix the arrival
rate for the last eight files such that the arrival rate of the
third and fourth file is .0000962/sec, and of the last six files
is 0.0001042/sec. The arrival rates for the first two files are
assumed to be the same, and due to symmetry of files and
placement, we consider four categories of content placement
in the cache - contents for the first two files, contents for
3rd file, contents for 4th file, and the contents for the last
six files. In the considered arrival rates in Figure 6, there
was no chunk for the third and the fourth file placed in the
cache due to low arrival rates. We note that we always assume
the arrival rates of the first two files as the highest but since
the servers on which they are placed have relatively lower
average load, the arrival rate needs to be significantly higher
for them to increase content in the cache. The six bars in
Figure 6 correspond to arrival rates for the first two files of
0.0001250, 0.0001563, 0.0001786, 0.0002083, 0.0002500, and
0.0002778, respectively. At an arrival rate of .000125/sec for
the first two files, there is no content for these files in the
cache. Thus even though the arrival rate for the first two files
is the largest, they do not need to be placed in the cache due
to the fact that they are placed in the storage servers that are
lightly loaded. As the arrival rate increases to .00015625/sec,
two chunks corresponding to these files start to appear in the
cache. Further increase of the arrival rate leads to more and
more chunks in the cache. Thus, we note that the placement of
chunks on the storage nodes, arrival rates, and service rates all
play a role in allocating chunks of different files in the cache.
Chunk Request Scheduling Evolution: In order to see
how the request scheduling evolves during each time bin, we
run the experiment with r = 1000 objects, each of size 200
MB and using a (7,4) erasure code, and a total cache size
of 62.5 GB. The average file request arrival rate for the two
experiments is chosen as λi = 0.0225 /sec and λi = 0.0384,
respectively. We divide each time bin (100 sec) into 20 time
slots, each with a length of 5 sec, and plot the number of
chunks the client is getting from the cache and from the
storage nodes in each time slot. Fig 7 shows that under both
workloads, the number of chunks retrieved from the cache is
smaller than that from OSDs. As the cache size is 62.5 GB,
with chunk size 50 MB, it has a capacity of 1250 chunks,
which means each file has more chunks in OSDs than that
in cache on an average. Further, since the arrival rate of all
file increases proportionally, the relative percentage of chunks
retrieved from cache over 100s stays almost the same in the
two cases, at about 33%.
C. Ceph Testbed
We have validated the effectiveness of our latency analysis
and the algorithm for joint optimization on average latency
in erasure-coded storage systems with caching through simulation. In the rest of this section, we will evaluate the
performance of our solution by implementing our algorithm
in Ceph (the Jewel version with filestore), which is an opensource, object-based distributed storage system [29]. Details
on Ceph storage API’s and cache tier are given in Appendix
A.
9
24
10
Chunks for the first two files in cache
Chunks for the last six files in cache
8
16
Chunks in Cache
Average File Latency
20
12
8
4
6
4
2
0
0
500
1000 1500 2000 2500 3000 3500 4000
Cache size ( in multiples of chunk-size of 25 MB)
Fig. 4. Average latency as cache size varies from
0 to 4000 chunk-size, each data point represents
average latency for r = 1000 files in one time
bin. This clearly demonstrates that the average
latency decreases as cache size increases when
all other parameters are fixed.
0
Fig. 5. Evolution of cache content in three time
bins, each having different request arrival rates
for the r = 10 files. It shows that our algorithm
is updating content held in the cache according to
the intensity of workload of each file effectively.
1
1.5
2
2.5
Changing Arrival Rate for first two files
3
#10-4
Fig. 6. The chunk placement depends not only on
arrival rates but also on the placement of contents in
the storage nodes and the service rates.
TABLE I
TABLE OF ARRIVAL RATES FOR 10 FILES IN 3 TIME BINS .
Time Bin
1
2
3
File 1
0.000156
0.000156
0.000125 ↓
File 2
0.000156
0.000156
0.00025 ↑
File 3
0.000125
0.000125
0.000125
File 4
0.000167
0.000125 ↓
0.000167
File 5
0.000104
0.000125 ↑
0.000104
Fig. 7. Number of chunk requests that are sent to storage servers or cache
with an average arrival rate λi = 0.0225 /s and λi = 0.0384, respectively,
where a time bin of length 100 s is divided into 20 equal time slots. The
figure shows the dynamic nature of the number of chunks the client obtained
from the cache/nodes.
Our Solution: Our algorithm requires a caching tier that
can store erasure-coded chunks in cache based on the proposed
optimization. Since Ceph cache tier only provides replicated
caching capability, we leverage it to implement an equivalent
version of our functional caching policy. Typically we use a
faster device (SSD) for cache, and by comparing the read
latency from an OSD backed with HDD (shown in Table
IV) and SSD (shown in Table V), we can see that service
latency from cache can be negligible compared to that from
back-end storage pool. As defined in Equation (1), the read
latency of a file in erasure coded storage depends on the largest
retrieval delay of the chunks (this is also true in Ceph). So the
File 6
0.000156
0.000156
0.000125 ↓
File 7
0.000156
0.000156
0.00025 ↑
File 8
0.000125
0.000125
0.000125
File 9
0.000167
0.000125 ↓
0.000167
File 10
0.000104
0.000125 ↑
0.000104
file latency depends only on the retrieval latency of chunks
stored in back-end storage. Then the read process of file with
an (n, k) code and d chunks in cache becomes equivalent to
reading a file with (n, k − d) code (with the same chunk size)
from back-end storage pool, while the latency of the d chunks
in cache is ignored.
This allows us to implement our functional caching policy
using different (n, k − d) codes and measure the resulting
latency. More specifically, as the algorithm provides number
of chunks of each file that are placed in the cache, we
can dynamically adjust the equivalent erasure code in Ceph,
by creating pools with the new equivalent erasure code and
forward the file requests to the new pools. In this evaluation
with Ceph, to be coherent with simulation setup, we create a
Ceph cluster with 12 OSDs, using (7, 4) original erasure code
with r = 1000 files, which means as we put more chunks in
cache, there will be 5 pools in total: (7, 4−d) for d = 0, . . . , 4.
And all of the five pools are backed by the same set of OSDs in
the Ceph cluster. So that this environment provides 12 OSDs
for object storage, each of the 1000 files has 7 chunks in the
storage pool, and 0 to 4 chunks in the cache (changes with
workload/timebin). So when a client tries to access a file, the
file will be accessed from one of the five pools according to
its equivalent code (calculated from the number of chunks in
the cache of each file), from the algorithm at that time.
D. Experiment Setup
Testbed Configuration: Throughout the evaluation, we
compare the performance of our algorithm for optimal caching
with a baseline: Ceph’s caching tier, replicated caching with
LRU. We have a test bed with 3 nodes for the Ceph object
storage cluster, and another four nodes each with one virtual
machine (VM) as 4 Ceph clients, each connected with 10G
Ethernet network interface. All OSDs in this testbed is formatted with an ext4 filesystem. The setup for the two cases
can be described as follows: (also shown in Fig 8)
10
1. Optimal Erasure-Coded Caching: Each node has 4 OSDs,
and each OSD is backed with a HDD drive of size 900G. There
are 5 pools created with the same set of 12 OSDs. Simulated
cache capacity of 2500 chunks of 16 MB files.(1000 files has
4000 chunks in total, cache capacity 10G). Each node has one
SSD drive that provides four 5G partitions for the journal of
each OSD. The number of placement group for each pool is
calculated from Equation (17), which is 256.
2. Ceph Replicated Caching with LRU: Cache tier consists
of 2 OSDs, each backed with a 5G SSD partition, cache
mode write-back with a capacity of 10G. Cache tier uses dual
replication (close to the redundancy provided by (7,4) erasure
code). Storage tier has the same set as that in optimal erasurecoded caching. There is one pool created with (7,4) erasure
code on the 12 OSDs. The SSD partitions come from different
SSD drives on one node. The number of placement group
in cache tier is 128 from Equation (17). Journal settings and
number of placement groups in the storage pool are the same
with the optimal caching case.
Baseline-‐ Ceph Caching
Optimal
Caching
node
2
monitor journal
OS
SAS
280G
node
3
...
...
osd1
osd3
osd4
1 to 1
SSD
20G
node
1
SAS SAS SAS SAS
900G 900G 900G 900G
node
2
...
...
osd2
Number of Containers(pools)
Number of VM Clients
Object Size
Access Pattern
Number of Workers
Rampup
Run time
5 or 1
4
4MB to 1GB
Write, Read
1
30 s
1800 s
to the sizes in the table) for our test, and the average request
arrival rate for each object in each size during the 24 hours is
shown in Table III.
TABLE III
TABLE OF 24- HOUR R EAL S TORAGE W ORKLOAD .
Object Size
4MB
16MB
64MB
256MB
1GB
Average Request Arrival Rate
0.00029868
0.00010824
0.00051852
0.0000078
0.0000024
1 Storage Tier Pool with 12
OSDs : (7,4)
5 Erasure Coded Pools:
(7,0), (7,1), (7,2), (7,3), (7,4)
node
1
TABLE II
TABLE OF COSB ENCH D EFAULT C ONFIGURATION
osd1
node
3
...
osd2
osd3
osd4
1 to 1
...
osd5
osd6
Part
5G
Part
5G
SAS SAS SAS SAS
900G 900G 900G 900G
SSD SSD
446G 446G
Contribute to
Storage Tier Pool
Cache Tier
Pool
Fig. 8. Our Ceph Testbed with 3 nodes for proposed optimal caching and
ceph’s LRU caching as a baseline.
Benchmarking Our Ceph Testbed: We use COSBench
0.4.2 ([66], a benchmarking tool for cloud object storage
developed by Intel) to benchmark the latency performance
(response time in COSBench) in this evaluation. As file
requests can be forward to any of the five storage pools
according to the algorithm dynamically, to ensure that a file
can be accessed from any pools at any time, we perform a
write bench mark without cleaning-up to the pools before we
run any read benchmarks. We modified COSBench to use
existing pools instead of creating new pools (which is by
default). We started 4 COSBench drives on the 4 VM clients.
Each COSBench workload has 5 stages: Initial, Prepare, Main,
Cleanup, and Dispose. For obtaining latency performance, the
number of threads (workers in COSBench) in each driver in
the first three stages is set to 1. The list of parameters is shown
in Table II.
We use COSBench workload generator to emulate a real
distributed storage workload, which is synthesized from actual production traces of a large-scale object storage system
deployed in a commercial service provider. We studied the
24-hour object storage workload and choose the most popular
object sizes (other similar object sizes are round up or down
E. Evaluation
Now we can evaluate the latency performance of our
proposed optimal caching with different types of workloads,
compared to Ceph’s LRU caching as a baseline.
Service Time Distribution: In a Ceph erasure coded pool,
a primary OSD receives all write operations, it encodes the
object into n = k + m chunks and sends them to other
OSDs (according to the value of n). When an object is read
from the erasure coded pool, it first checks the status of all
OSDs that has a chunk of the required object, and then try
to contact all such OSDs for retrieve a chunk. The decoding
function can be called as soon as k chunks are read. While
our service delay bound applies to arbitrary distribution and
works for systems hosting any number of objects, we first run
an experiment to understand actual service time distribution
on our testbed, which is used as an input in the algorithm.
To find the chunk service time distribution for different chunk
sizes, we run a 100% write benchmark for all object sizes in
Table III, into the (7,4) erasure coded pool in baseline testbed
( Ceph Replicated Caching with LRU) without cleaning-up
for 900 seconds. Then, we run a 100% read benchmark for
each object size with an average request arrival rate per object
(among all objects of each size) as shown in III. For the five
set experiments with five different object sizes, we run each set
for 1800 seconds to get enough number of requests to take an
average according to arrival rates in Table III. The chunk sizes
are 1 MB, 4 MB, 16 MB, 64 MB, 256 MB, accordingly. We
collect the service time of the chunks of different sizes at the
OSDs and plot the results. Figure 9 depicts the Cumulative
Distribution Function (CDF) of the chunk service times for
read operations with different chunk sizes. We ignored the
result for 1G object (256 MB chunk size) as in the figure as
11
it is much larger than others and it would be hard to see the
distribution of the first 4 if we add the result. The average
service time at the OSD for read operations with chunk size
256 MB is 6758.06 milliseconds. Using the measured results,
we get the mean and variance of service time of each chunk
size as shown in Table IV. We use the obtained mean, and
TABLE IV
M EAN AND VARIANCE OF S ERVICE T IME FOR D IFFERENT C HUNK S IZES
IN M ILLI S ECONDS .
Chunk Size
1MB
4MB
16MB
64MB
256MB
Mean Service Time
6.6696
35.8800
147.8462
355.0800
6758.06
Variance of Service Time
0.0963
2.6925
388.9872
1256.6100
554180
variance of the service time of chunks with different sizes
to calculate moments of the service time distribution that are
used to find cache placement in our algorithm.
Cumulative Distribution Function for Chunk Service Time
1
1MB chunk
4MB chunk
16MB chunk
64MB chunk
0.9
0.8
Probability
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
50
100
150
200
250
300
350
400
Servie Time (ms)
Fig. 9. Actual service latency distribution for different chunk sizes from the
Ceph testbed with 12 OSDs without cache tier.
Evaluation of the Performance of Our Solution: As
we have validated the algorithm in aspects of convergence
and evolution of dynamic caching, we now evaluate the
effectiveness of the proposed algorithm (in terms of overall
latency improvement) in real storage systems with the real
workload. For experiment in optimal caching, as we introduced
earlier, we have five pools set up with erasure code (7,0), (7,1),
(7,2), (7,3), (7,4), where pool (7,0) means all the four required
chunks are stored in cache, which will typically be some local
SAS or NVMe SSDs in the case of optimal caching. We run a
read test to the SSD drives for all object sizes in Table III and
get the average latency numbers for each object size shown
in Table V. As compared with average service time of the
same chunk size from an OSD (HDD as backend) shown in
Table IV, we can see that the latency of a chunk from cache
can be negligible when the read latency of the whole object is
measured. This is because the latter depends on the chunk read
latency from an OSD in the Ceph cluster, which motivates the
use of equivalent codes in the caching schemes.
In each time bin, given the workload and the total number
of active objects (for simplicity of the algorithm, we set this
number to 1000, which is also very close to the number of
TABLE V
R EAD L ATENCY OF DIFFERENT CHUNK SIZES FROM C ACHE (SAS SSD) IN
OPTIMAL CACHING ( MS ).
Chunk Size
1 MB
4 MB
16 MB
64 MB
256 MB
Read Latency From Cache
1.86619
7.35639
30.4927
97.0968
349.133
objects that have been accessed during workload run time
for 30-min in the real workload ) in the Ceph object storage
cluster, the optimization algorithm provides which object belongs to which pool (an object-pool map), and the read request
arrival rate of each object in each pool. In order to evaluate
the effectiveness of the optimal caching algorithm with various
object sizes, first, we run a 100% write benchmark into the
erasure coded pools according to the object-pool map from the
optimization for each object size in Table III. Then, for each
object size, we run read benchmarks to the erasure-code pools
(according to the object-pool map from the optimization) for
1800 seconds. The average read request arrival rate per object
in each pool can be calculated as the number of objects in the
pool from the algorithm times the average read request arrival
rate for each object size in Table III. After the experiment
is done, we get the average read latency for objects in each
erasure coded pool, and then calculate the average latency over
the 1800 seconds time bin as,
X number of objects in pool i ∗ average latency in pool i
.
total number of objects in all pools
i
(16)
For evaluation of Ceph’s LRU replicated caching as a
baseline, first, we run 100% write benchmark (with the same
number of objects we used in optimal caching) for each object
size in Table III to the (7,4) erasure-code storage pool with a
replicated cache tier overlay. Then for each object size, we
perform read workloads to this pool with an average read
request arrival rate per object as shown in Table III for 1800
seconds. Aggregated request arrival rate in this test for the
(7,4) pool would be the request arrival rate in Table III times
the total number of active objects (1000 in this case). In both
the cases, cache capacity is fixed and set to be 10GB. This
cache can hold 10000 chunks of 4 MB objects, 2500 chunks
of 16 MB objects, 625 chunks of 64 MB objects, 156 chunks
of 256 MB objects, and 39 chunks of 1 GB objects when
using (7,4) erasure code. For example, each object of size 16
MB can have 2500/1000=2.5 chunks on average stored in the
cache in both optimal caching and Ceph cache-tier. However,
as each object of each size has a different read request arrival
rate (value shown in III is an average of all files of the same
size), placing two or three chunks in the cache would help
reduce latency significantly, thus leading to uneven placement
of chunks of the objects in the cache. For this experiment
with baseline, we get the average read latency for objects in
(7,4) erasure coded pool. We compare the average latency to
12
retrieve an object in optimal caching and Ceph’s LRU caching.
Fig 10 shows that average latency increases as the object size
increases, which we can see from our latency bound since
the number of chunks that the cache can hold decreases as
object size increases. The figure also shows that our caching
improves latency as compared to Ceph’s LRU caching as a
baseline, which is using dual replication in its cache tier, by
26% on average. This improvement becomes more and more
significant as the number of objects increase, which shows that
our dynamic optimal chunk placement in the cache is more
effective than traditional LRU cache. The improvement with
increased file size is because the load on the system increases
thus obtaining more latency advantages with caching. For a
lightly-loaded system, the latency is small and the improvement in caching may not be that significant. Fig. 10 shows
that our algorithm with optimal caching significantly improves
latency with a reasonable number of objects and cache sizes.
Fig. 10 also compares the analytically optimized expression of
latency from the simulations and the experimentally observed
latency where we see that the analytical result is an upper
bound and matches well with the experimental results.
Next, we evaluate the performance of optimal caching with
various workload intensities. We fix the object size at 64 MB
(we still use 1000 objects ) and vary the read request arrival
rate for these objects and evaluate the latency performance.
Queuing latency would not be a dominant issue if the workload
is not intense. Thus, we fully load the Ceph cluster with much
higher request arrival rates as compared to those in Table
III. The aggregate read request arrival rates (Single object
request arrival rate times the number of objects) used in this
test are 0.5, 1.0, 2.0, 4.0, and 8.0. In the optimal caching
case, for each workload (request arrival rate), we perform the
read benchmark to the five erasure coded pools according to
the object-pool map from the optimization with this request
arrival rate for 1800 seconds. For Ceph’s LRU caching, we
perform read benchmarks to the (7,4) erasure coded pool with
the same set of arrival rates in the case of optimal caching
for 1800 seconds, and obtain the average access latency per
object. The cache size for both cases is still fixed at 10 GB.
Actual average service latency of objects for each workload is
shown by a bar plot in Figure 11. In this experiment we also
compare the results for the optimal caching scheme and Ceph’s
LRU caching as a baseline. Fig 11 shows that our optimal
caching algorithm outperforms Ceph’s LRU caching in terms
of average latency for all workload intensities in the setup.
The proposed algorithm gives an average 23.86% reduction in
latency. Thus, our algorithm with optimal caching can mitigate
traffic contention and reduce latency very efficiently compared
to Ceph’s LRU caching, this also provides a guideline for
designing caching schemes under very heavy workloads.
VI. C ONCLUSIONS
In this paper, we propose functional caching, a novel
approach to create erasure-coded cache chunks that maintain
MDS code property along with existing data chunks. It outperforms exact caching schemes and provides a higher degree
of freedom in file access and request scheduling. We quantify
an upper bound on the mean service latency in closed-form
for erasure-coded storage systems with functional caching, for
arbitrary chunk placement and service time distributions. A
cache optimization problem is formulated and solved using an
efficient heuristic algorithm. Numerical results and prototype
in an open-source cloud storage validate significant service
latency reduction using functional caching.
This paper assumes that a rate monitoring/prediction oracle
(e.g., an online predictive model or a simple sliding-windowbased method) is available to detect the rate changes. Finding a
robust algorithm that can automatically adjust to such changes
is an open problem and will be considered as future work.
R EFERENCES
[1] V. Aggarwal, Y. F. R. Chen, T. Lan, and Y. Xiang, “Sprout: A functional
caching approach to minimize service latency in erasure-coded storage,”
in 2016 IEEE 36th International Conference on Distributed Computing
Systems (ICDCS), June 2016, pp. 753–754.
[2] H. Weatherspoon and J. Kubiatowicz, “Erasure coding vs. replication: A
quantitative comparison,” in Revised Papers from the First International
Workshop on Peer-to-Peer Systems, ser. IPTPS ’01. Springer-Verlag,
2002.
[3] A. Dimakis, P. Godfrey, Y. Wu, M. Wainwright, and K. Ramchandran,
“Network coding for distributed storage systems,” Information Theory,
IEEE Transactions on, vol. 56, no. 9, pp. 4539–4551, Sept 2010.
[4] M. Sathiamoorthy, M. Asteris, D. Papailiopoulos, A. G. Dimakis,
R. Vadali, S. Chen, and D. Borthakur, “Xoring elephants: Novel erasure
codes for big data,” in Proceedings of the 39th international conference
on Very Large Data Bases., 2013.
[5] C. Huang, H. Simitci, Y. Xu, A. Ogus, B. Calder, P. Gopalan, J. Li, and
S. Yekhanin, “Erasure coding in windows azure storage,” in Proceedings
of the 2012 USENIX Conference on Annual Technical Conference, ser.
USENIX ATC’12. USENIX Association, 2012.
[6] A. Fikes, “Storage architecture and challenges (talk at the google faculty
summit),” http://bit.ly/nUylRW, Tech. Rep., 2010.
[7] L. Huang, S. Pawar, H. Zhang, and K. Ramchandran, “Codes can reduce
queueing delay in data centers,” in Information Theory Proceedings
(ISIT), 2012 IEEE International Symposium on, July 2012, pp. 2766–
2770.
[8] G. Joshi, Y. Liu, and E. Soljanin, “On the delay-storage trade-off in
content download from coded distributed storage systems,” Selected
Areas in Communications, IEEE Journal on, vol. 32, no. 5, pp. 989–997,
May 2014.
[9] S. Chen, Y. Sun, U. Kozat, L. Huang, P. Sinha, G. Liang, X. Liu, and
N. Shroff, “When queuing meets coding: Optimal-latency data retrieving
scheme in storage clouds,” in Proceedings of IEEE Infocom, 2014.
[10] N. Shah, K. Lee, and K. Ramachandran, “The mds queue: analyzing
latency performance of codes and redundant requests,” arXiv:1211.5405,
Nov 2012.
[11] Y. Xiang, T. Lan, V. Aggarwal, and Y. F. R. Chen, “Joint latency and
cost optimization for erasurecoded data center storage,” SIGMETRICS
Perform. Eval. Rev., vol. 42, no. 2, pp. 3–14, Sep. 2014.
[12] ——, “Joint latency and cost optimization for erasure-coded data center
storage,” IEEE/ACM Transactions on Networking, vol. 24, no. 4, pp.
2443–2457, Aug 2016.
[13] V. Aggarwal, J. Fan, and T. Lan, “Taming tail latency for erasure-coded,
distributed storage systems,” in Proc. IEEE Infocom, May 2017.
[14] V. Aggarwal, A. O. Al-Abbasi, J. Fan, and T. Lan, “Taming tail
latency for erasure-coded, distributed storage systems,” arXiv preprint
arXiv:1703.08337, 2017.
[15] Y. Xiang, T. Lan, V. Aggarwal, and Y.-F. Chen, “Multi-tenant latency
optimization in erasure-coded storage with differentiated services,” in
Distributed Computing Systems (ICDCS), 2015 IEEE 35th International
Conference on, June 2015, pp. 790–791.
[16] Y. Xiang, V. Aggarwal, Y.-F. Chen, and T. Lan, “Taming latency in data
center networking with erasure coded files,” in Cluster, Cloud and Grid
Computing (CCGrid), 2015 15th IEEE/ACM International Symposium
on, May 2015, pp. 241–250.
13
Latency in Optimized Function Caching and Baseline
35000
Latency in Functional Caching with different Arrival Rates
140000
39021
Optimized Caching Experimental
Baseline Ceph Cachetier Experimental
Optimized Caching Numerical
Average Access Latency (ms)
Average Access Latency (ms)
40000
30000
25000
21516
20000
15000
11163
10000
7901
5000
0
8 10
4MB
384 430
16MB
120000
256MB
80000
File Size
44679
40000
18379
20000
24179
6510
2055 2800 4730
0.5000
1.0000
2.0000
4.0000
8.0000
Aggregated Arrival Rate
Fig. 10. Average latency in optimal erasure coded caching for different
object sizes during 1800 seconds benchmark run time, compared with
Ceph’s LRU replicated caching as a baseline. Cache capacity fixed at
10 GB in both cases. Figure shows significant latency improvement in
optimal caching.
[17] Y. Xiang, V. Aggarwal, Y. F. Chen, and T. Lan, “Differentiated latency
in data center networks with erasure coded files through traffic engineering,” IEEE Transactions on Cloud Computing, vol. PP, no. 99, pp.
1–1, 2017.
[18] Y. Xiang, T. Lan, V. Aggarwal, and Y. F. Chen, “Optimizing differentiated latency in multi-tenant, erasure-coded storage,” IEEE Transactions
on Network and Service Management, vol. 14, no. 1, pp. 204–216, March
2017.
[19] R. Pedarsani, M. A. Maddah-Ali, and U. Niesen, “Online coded
caching,” in IEEE International Conference on Communications, ICC
2014, Sydney, Australia, June 10-14, 2014, 2014, pp. 1878–1883.
[20] S. J. Nadgowda, R. C. Sreenivas, S. Gupta, N. Gupta, and A. Verma,
Service-Oriented Computing: 12th International Conference, ICSOC
2014, Paris, France, November 3-6, 2014. Proceedings.
Berlin,
Heidelberg: Springer Berlin Heidelberg, 2014, ch. C2P: Co-operative
Caching in Distributed Storage Systems, pp. 214–229.
[21] F. Chang, J. Dean, S. Ghemawat, W. C. Hsieh, D. A. Wallach, M. Burrows, T. Chandra, A. Fikes, and R. E. Gruber, “Bigtable: A distributed
storage system for structured data,” ACM Trans. Comput. Syst., vol. 26,
no. 2, pp. 4:1–4:26, Jun. 2008.
[22] Q. Zhu, A. Shankar, and Y. Zhou, “Pb-lru: A self-tuning power aware
storage cache replacement algorithm for conserving disk energy,” in
Proceedings of the 18th Annual International Conference on Supercomputing, ser. ICS ’04. New York, NY, USA: ACM, 2004, pp. 79–88.
[23] M. Naik, F. Schmuck, and R. Tewari, “Read and write requests to
partially cached files.”
[24] E. Thereska, G. O’Shea, H. Ballani, T. Karagiannis, and A. Rowstron,
“Software-defined caching: Managing caches in multi-tenant data centers,” in ACM Symposium on Cloud Computing (SOCC) 2015, August
2015.
[25] Y.-J. Hong and M. Thottethodi, “Understanding and mitigating the
impact of load imbalance in the memory caching tier,” in Proceedings of
the 4th Annual Symposium on Cloud Computing, ser. SOCC ’13. New
York, NY, USA: ACM, 2013, pp. 13:1–13:17.
[26] Y. Cheng, A. Gupta, and A. R. Butt, “An in-memory object caching
framework with adaptive load balancing,” in Proceedings of the Tenth
European Conference on Computer Systems, ser. EuroSys ’15, 2015, pp.
4:1–4:16.
[27] L. Wang, S. Bayhan, and J. Kangasharju, “Optimal chunking and partial
caching in information-centric networks,” Comput. Commun., vol. 61,
no. C, pp. 48–57, May 2015.
[28] D. Bertsimas, K. Natarajan, and C.-P. Teo, “Tight bounds on expected
order statistics,” Probab. Eng. Inf. Sci., vol. 20, no. 4, pp. 667–686, Oct.
2006.
[29] S. A. Weil, S. A. Brandt, E. L. Miller, D. D. E. Long, and
C. Maltzahn, “Ceph: A scalable, high-performance distributed file
system,” in Proceedings of the 7th Symposium on Operating
Systems Design and Implementation, ser. OSDI ’06. Berkeley, CA,
58917
60000
0
1GB
112172
100000
2182 2833
64MB
135468
Optimized Caching
Baseline Ceph Cachetier
Fig. 11. Comparison of average latency in optimal erasure coded
caching and Ceph’s LRU replicated caching for different intensity of read
workload during 1800 seconds benchmark run time. Cache capacity fixed
at 10 GB in both cases. Figure shows significant latency improvement in
optimal caching.
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
USA: USENIX Association, 2006, pp. 307–320. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1298455.1298485
A. Kumar, R. Tandon, and T. Clancy, “On the latency and energy
efficiency of distributed storage systems,” IEEE Transactions on Cloud
Computing, vol. PP, no. 99, pp. 1–1, 2015.
M. Noori, E. Soljanin, and M. Ardakani, “On storage allocation for
maximum service rate in distributed storage systems,” in 2016 IEEE
International Symposium on Information Theory (ISIT), July 2016, pp.
240–244.
G. Joshi, E. Soljanin, and G. Wornell, “Queues with redundancy:
Latency-cost analysis,” SIGMETRICS Perform. Eval. Rev., vol. 43, no. 2,
pp. 54–56, Sep. 2015.
V. Aggarwal and T. Lan, “Tail index for a distributed storage system
with pareto file size distribution,” CoRR, vol. abs/1607.06044, 2016.
F. Meyer auf der Heide, B. Vöcking, and M. Westermann, “Caching in
networks (extended abstract),” in Proceedings of the Eleventh Annual
ACM-SIAM Symposium on Discrete Algorithms, January 9-11, 2000,
San Francisco, CA, USA., 2000, pp. 430–439.
M. R. Korupolu, C. G. Plaxton, and R. Rajaraman, “Placement algorithms for hierarchical cooperative caching,” in Proceedings of the Tenth
Annual ACM-SIAM Symposium on Discrete Algorithms, ser. SODA ’99,
1999, pp. 586–595.
S. Jiang and X. Zhang, “Making lru friendly to weak locality workloads:
a novel replacement algorithm to improve buffer cache performance,”
Computers, IEEE Transactions on, vol. 54, no. 8, pp. 939–952, Aug
2005.
N. Megiddo and D. S. Modha, “Arc: A self-tuning, low overhead
replacement cache,” in Proceedings of the 2Nd USENIX Conference on
File and Storage Technologies, ser. FAST ’03, 2003, pp. 115–130.
Y. Zhou, J. Philbin, and K. Li, “The multi-queue replacement algorithm
for second level buffer caches,” in Proceedings of the General Track:
2001 USENIX Annual Technical Conference, 2001, pp. 91–104.
S. Liang, K. Chen, S. Jiang, and X. Zhang, “Cost-aware caching
algorithms for distributed storage servers,” in Distributed Computing,
ser. Lecture Notes in Computer Science, A. Pelc, Ed. Springer Berlin
Heidelberg, 2007, vol. 4731, pp. 373–387.
S. Angelopoulos, R. Dorrigiv, and A. López-Ortiz, “On the separation
and equivalence of paging strategies,” in Proceedings of the Eighteenth
Annual ACM-SIAM Symposium on Discrete Algorithms, ser. SODA ’07,
2007, pp. 229–237.
D. S. M. Elisha J. Rosensweig and J. Kurose, “On the steady-state of
cache networks,” in Proceedings of the IEEE INFOCOM, 2013.
J. Zhang, X. Lin, and X. Wang, “Coded caching under arbitrary
popularity distributions,” in 2015 Information Theory and Applications
Workshop, ITA 2015, San Diego, CA, USA, February 1-6, 2015, 2015,
pp. 98–107.
K. Rashmi, M. Chowdhury, J. Kosaian, I. Stoica, and K. Ramchandran,
“Ec-cache: load-balanced, low-latency cluster caching with online erasure coding,” in 12th USENIX Symposium on Operating Systems Design
14
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
and Implementation (OSDI 16). USENIX Association, 2016, pp. 401–
417.
Q. Huang, K. Birman, R. van Renesse, W. Lloyd, S. Kumar, and H. C. Li,
“An analysis of facebook photo caching,” in Proceedings of the TwentyFourth ACM Symposium on Operating Systems Principles, ser. SOSP
’13, 2013, pp. 167–181.
N. Bronson, Z. Amsden, G. Cabrera, P. Chakka, P. Dimov, H. Ding,
J. Ferris, A. Giardullo, S. Kulkarni, H. Li, M. Marchukov, D. Petrov,
L. Puzar, Y. J. Song, and V. Venkataramani, “Tao: Facebook’s distributed
data store for the social graph,” in Presented as part of the 2013 USENIX
Annual Technical Conference (USENIX ATC 13), San Jose, CA, 2013,
pp. 49–60.
B. Fitzpatrick, “Distributed caching with memcached,” Linux J., vol.
2004, no. 124, pp. 5–, Aug. 2004.
G. Chockler, G. Laden, and Y. Vigfusson, “Data caching as a cloud
service,” in Proceedings of the 4th International Workshop on Large
Scale Distributed Systems and Middleware, ser. LADIS ’10, 2010, pp.
18–21.
——, “Design and implementation of caching services in the cloud,”
IBM Journal of Research and Development, vol. 55, no. 6, pp. 9:1–
9:11, Nov 2011.
U. Niesen and M. A. Maddah-Ali, “Coded caching for delay-sensitive
content,” CoRR, vol. abs/1407.4489, 2014.
N. Karamchandani, U. Niesen, M. Maddah-Ali, and S. Diggavi, “Hierarchical coded caching,” in Information Theory (ISIT), 2014 IEEE
International Symposium on, June 2014, pp. 2142–2146.
M. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” Information Theory, IEEE Transactions on, vol. 60, no. 5, pp. 2856–2867,
May 2014.
U. Niesen and M. Maddah-Ali, “Coded caching with nonuniform
demands,” in Computer Communications Workshops (INFOCOM WKSHPS), 2014 IEEE Conference on, April 2014, pp. 221–226.
S. P. Shariatpanahi, S. A. Motahari, and B. H. Khalaj, “Multi-server
coded caching,” IEEE Transactions on Information Theory, vol. 62,
no. 12, pp. 7253–7271, Dec 2016.
T. Luo, V. Aggarwal, and B. Peleato, “Coded caching with distributed
storage,” arXiv preprint arXiv:1611.06591, 2016.
A. Dimakis, V. Prabhakaran, and K. Ramchandran, “Distributed data
storage in sensor networks using decentralized erasure codes,” in Signals,
Systems and Computers, 2004. Conference Record of the Thirty-Eighth
Asilomar Conference on, vol. 2, Nov 2004, pp. 1387–1391 Vol.2.
M. Aguilera, R. Janakiraman, and L. Xu, “Using erasure codes efficiently for storage in a distributed system,” in Dependable Systems and
Networks, 2005. DSN 2005. Proceedings. International Conference on,
June 2005, pp. 336–345.
Q. Lv, P. Cao, E. Cohen, K. Li, and S. Shenker, “Search and replication
in unstructured peer-to-peer networks,” in Proceedings of the 16th
International Conference on Supercomputing, ser. ICS ’02, 2002, pp.
84–95.
D. Shen and J. L. Hellerstein, “Predictive models for proactive network
management: Application to a production web server,” in Network
Operations and Management Symposium, 2000. NOMS 2000. 2000
IEEE/IFIP. IEEE, 2000, pp. 833–846.
W. Massey, G. Parker, and W. Whitt, “Estimating the parameters of a
non-homogeneous poisson-process model with liear rate,” Telecommunication Systems, vol. 5, no. 361, 1996.
M. E. Kuhl, H. Damerdji, and J. R. Wilson, “Least squares estimation
of nonhomogeneous poisson processes,” in Proceedings of the 30th
Conference on Winter Simulation, ser. WSC ’98. Los Alamitos, CA,
USA: IEEE Computer Society Press, 1998, pp. 637–646. [Online].
Available: http://dl.acm.org/citation.cfm?id=293172.293287
W. W. Peterson and E. Weldon, “Error-correcting codes,” in the MIT
Press, March 1972.
A. Abdelkefi and Y. Jiang, “A structural analysis of network delay,” in
Communication Networks and Services Research Conference (CNSR),
2011 Ninth Annual, May 2011, pp. 41–48.
H. Weatherspoon and J. Kubiatowicz, “Erasure coding vs. replication:
A quantitative comparison,” in Peer-to-Peer Systems, ser. Lecture Notes
in Computer Science, P. Druschel, F. Kaashoek, and A. Rowstron, Eds.
Springer Berlin Heidelberg, 2002, vol. 2429, pp. 328–337.
A. Downey, “The structural cause of file size distributions,” in Modeling,
Analysis and Simulation of Computer and Telecommunication Systems,
2001. Proceedings. Ninth International Symposium on, 2001, pp. 361–
370.
[65] F. Paganini, A. Tang, A. Ferragut, and L. Andrew, “Network stability
under alpha fair bandwidth allocation with general file size distribution,”
Automatic Control, IEEE Transactions on, vol. 57, no. 3, pp. 579–591,
March 2012.
[66] Q. Zheng, H. Chen, Y. Wang, J. Duan, and Z. Huang, “Cosbench: A
benchmark tool for cloud object storage services,” in Proceedings of
CLOUD 2012.
[67] J. S. Plank, S. Simmerman, and C. D. Schuman, “Jerasure: A library in
c/c++ facilitating erasure coding for storage applications,” Tech. Rep.,
2007.
[68] T.
Angell,
“The
farkas-minkowski
theorem,”
www.math.udel.edu/∼angell/Opt/farkas.pdf, Tech. Rep., 2002.
Vaneet Aggarwal (S’08 - M’11 - SM15) received
the B.Tech. degree in 2005 from the Indian Institute
of Technology, Kanpur, India, and the M.A. and
Ph.D. degrees in 2007 and 2010, respectively from
Princeton University, Princeton, NJ, USA, all in
Electrical Engineering.
He is currently an Assistant Professor at Purdue
University, West Lafayette, IN. Prior to this, he was
a Senior Member of Technical Staff Research at
AT&T Labs-Research, NJ, and an Adjunct Assistant
Professor at Columbia University, NY. His research interests are in applications of statistical, algebraic, and optimization techniques to distributed storage
systems, machine learning, and wireless systems. Dr. Aggarwal was the
recipient of Princeton University’s Porter Ogden Jacobus Honorific Fellowship
in 2009. In addition, he received AT&T Key Contributor award in 2013, AT&T
Vice President Excellence Award in 2012, and AT&T Senior Vice President
Excellence Award in 2014.
Yih-Farn (Robin) Chen is a Director of Inventive Science, leading the Cloud Platform Software
Research department at AT&T Labs - Research.
His current research interests include cloud computing, software-defined storage, mobile computing,
distributed systems, World Wide Web, and IPTV. He
holds a Ph.D. in Computer Science from University
of California at Berkeley, an M.S. in Computer
Science from University of Wisconsin, Madison,
and a B.S. in Electrical Engineering from National
Taiwan University. Robin is an ACM Distinguished Scientist and a Vice
Chair of the International World Wide Web Conferences Steering Committee
(IW3C2). He also serves on the editorial board of IEEE Internet Computing.
Tian Lan (S’03-M’10) received the B.A.Sc. degree
from the Tsinghua University, China in 2003, the
M.A.Sc. degree from the University of Toronto,
Canada, in 2005, and the Ph.D. degree from the
Princeton University in 2010. Dr. Lan is currently
an Associate Professor of Electrical and Computer
Engineering at the George Washington University.
His research interests include cloud resource optimization, mobile networking, storage systems and
cyber security. Dr. Lan received the 2008 IEEE
Signal Processing Society Best Paper Award, the 2009 IEEE GLOBECOM
Best Paper Award, and the 2012 INFOCOM Best Paper Award.
15
Yu Xiang received the B.A.Sc. degree from Harbin
Institute of Technology in 2010, and the Ph.D.
degree from George Washington University in 2015,
both in Electrical Engineering. She is now a senior inventive scientist at AT&T Labs-Research.
Her current research interest are in cloud resource
optimization, distributed storage systems and cloud
storage charge-back.
A PPENDIX A
C EPH T ESTBED
The underlying foundation of Ceph is the Reliable Autonomic Distributed Object Store (RADOS). Ceph provides
three types of storage API’s on top of RADOS: (i) Object
Storage: the fundamental RADOS object-based storage system. (ii) Block Storage: RADOS Block Device (RBD) that
stores block device images as objects. (iii) Filesystem: this
runs on top of the object storage with a metadata server
that maps the directories and file names of the file system
to objects stored within RADOS clusters. We use Ceph object
storage in our testbed for our joint optimization algorithm with
caching in erasure coded storage because i) Object storage is
a good fit for large-scale cloud storage systems since it stores
data more efficiently. ii) In terms of resilience, Ceph object
storage provides both replicated and erasure-coded storage,
while erasure-coded block storage on top of RADOS is not
supported in Ceph yet.
Ceph Object Storage Cluster: A Ceph object storage
cluster consists of two types of daemons: (i) Ceph monitor,
which keeps a master copy of the cluster map and monitors
the status of daemons in the cluster. (ii) Ceph object storage
device(OSD), which is a physical or logical storage device
used to store data as objects. Each OSD has two components:
journal and data, during a write operation, it writes to the
journal first and then writes data from journal to OSD data
drive. Ceph stores data within pools, a pool consists of a bunch
of OSDs, and Ceph users can specify which bunch of OSDs
forms a certain pool through CRUSH rules. CRUSH is an
algorithm Ceph used to pseudo-randomly store and retrieve
data in OSDs with a uniform distribution of data across the
cluster. CRUSH has a cluster map which contains a list of
OSDs, and each OSD’s physical location, and a list of rules
that tell CRUSH how it should encode/replicate data in pools.
To reduce the amount of metadata when storing data in OSDs,
Ceph maps objects to placement groups, which are fragments
of an object pool that place objects as a group into OSDs. The
number of placement groups in an (n=k+m, k) erasure coded
pool can be calculated as:
Number of OSDs ∗ 100
m
(17)
Erasure-coded Storage in Ceph: When creating an erasure
coded pool in Ceph, users need to specify erasure code
parameters, the number of placement groups, OSDs belong
to this pool, and failure domain of the pool (OSD, or host)
etc., in an erasure-code profile. Ceph object storage is using
Number of Placement Groups =
jerasure plugin (a generic and flexible plugin that encapsulates
the Jerasure code library) for default erasure code set up [67].
For an erasure-code profile with jerasure plugin, we specify
erasure code with k, which is the number of data chunks, and
m, which is the number of coded chunks, thus forming an
(k + m, k) erasure code.
Ceph Cache Tier: Ceph provides a cache tiering service to
improve IO performance for storage devices. Users can create
a Ceph cache tier as an overlay to the existing storage pool.
Usually, the cache tier is a pool with faster but expensive
storage devices (mostly SAS SSDs or Nvme SSDs). The
existing back-end storage pool could be either an erasurecoded or replicated pool composed of slower and cheaper
storage devices. However, the cache tier pool has to be
replicated as Ceph caching does not support erasure-coded
storage yet. For a pool with cache tier, all IO traffic will be
routed to the cache tier first. In the case of a cache miss, the
data will be promoted from the storage tier to the cache tier.
And the tiering agent will evict the least recently used (LRU)
data from cache tier and flushes them to storage tier.
A PPENDIX B
P ROOF OF L EMMA 1
The key idea of probabilistic scheduling is to get the
file i from set Ai , which is chosen probabilistically, and
thus is given byPthe parameters
P(Ai ) ∀Ai ⊆ Si and ∀i,
which involves i nkii decision variables. While it appears
prohibitive computationally, the authors of [12] demonstrated
that the optimization can P
be transformed into an equivalent
form, which only requires i ni variables. The key idea is to
show that it is sufficient to consider the conditional probability
(denoted by πi,j ) of selecting a node j, given that a batch of
ki chunk requests of file i are dispatched. It is easy to see that
for given P(Ai ), we can derive πi,j by
X
πi,j =
P(Ai ) · 1{j∈Ai } , ∀i
(18)
Ai :Ai ⊆Si
where 1{j∈Ai } is an indicator function which equals to 1 if
node j is selected by Ai and 0 otherwise. The equivalence
proof of the probability over sets Ai and the nodes j follows
using Farkas-Minkowski Theorem [68].
Using this result, it is sufficient to study probabilistic
scheduling via conditional probabilities πi,j , which greatly
simplifies the analysis. In particular, it is easy to verify that
under our model, the arrival of chunkP
requests at node j form
a Poisson Process with rate Λj =
i λi πi,j , which is the
superposition of r Poisson processes each with rate λi πi,j , µj
is the service rate of node j. The resulting queuing system
under probabilistic scheduling is stable if all local queues are
stable.
Let Qmax be the maximum of waiting time {Qj , j ∈ Ai }.
We first show that Qmax is upper bounded by the following
inequality for arbitrary z ∈ R:
X
+
+
Qmax ≤ z + [Qmax − z] ≤ z +
[Qj − z] ,
(19)
j∈Ai
16
where [a]+ = max{a, 0} is a truncate function. Now, taking
the expectation on both sides of (19), we have
X
+
E [Qmax ] ≤ z + E
[Qj − z]
j∈Ai
X 1
= z + E
(Qj − z + |Qj − z|)
2
j∈Ai
X 1
(E[Qj ] − z + E|Qj − z|) ,
= z + EAi
2
j∈Ai
X πi,j
=z+
(E[Qj ] − z + E|Qj − z|), (20)
2
j∈Ai
where EAi denotes the expectation over randomly selected
ki storage nodes in Ai ⊆ S according to probabilities
πi,1 , . . . , πi,m . From Cauchy-Schwarz inequality, we have
q
(21)
E|Qj − z| ≤ (E[Zj ] − z)2 + Var[Qj ].
Combining (20) and (21), we obtain
X πi,j
T̄i ≤ min z +
(E[Qj ] − z)
z∈R
2
j∈Si
X πi,j
+
2
j∈Si
q
. (22)
(E[Qj ] − z)2 + Var[Qj ]
Using Pollaczek-Khinchin transform [7] to obtain the mean
and variance of M/G/1 queueing delays, we obtain the result
as in the statement of the Lemma.
| 7cs.IT
|
On Unique Decoding from Insertions and Deletions
arXiv:1611.09073v6 [cs.IT] 28 Sep 2017
Kayvon Mazooji
[email protected]
UCLA, Los Angeles, CA 90095
Abstract—In this paper, we study how often unique decoding
from t insertions or t deletions occurs for error correcting codes.
Insertions and deletions frequently occur in synchronization
problems and DNA, a medium which is beginning to be used
for long term data storage.
We define natural probabilistic channels that make t insertions
or t deletions, and study the probability of unique decoding. Our
most substantial contribution is the derivation of tight upper
bounds on the probability of unique decoding for messages passed
though these channels. We also consider other aspects of the
problem, and derive improved upper bounds for linear codes
and VT-codes.
Index Terms—Insertions and Deletions, Codes, Combinatorics,
Sequence reconstruction, Varshamov-Tenengolts codes
I. I NTRODUCTION
Codes correcting insertions and deletions have historically
been important to problems in synchronization [22]. Recently,
such codes have been useful for DNA storage as well [14].
Researchers have had little luck in finding insertion correcting
codes of optimal cardinality. However, progress has been made
on finding upper and lower bounds on the optimal cardinality
of such codes [10], [11], [13]. The most famous code believed
to be optimal is the VT-code [1], which corrects a single
insertion or deletion [2]. Other codes for insertions are found
in [16], [17], [18], [19]. List decoding from insertions and
deletions is considered in [20], [21].
In this paper we study the probability of uniquely decoding
from insertions and deletions for broad classes of codes. This
problem is particularly relevant to long term storage problems
(e.g. DNA storage [15]). For example, side information may
not be available if a codeword is recovered hundreds of
years in the future, making a list of > 1 decoded codewords
insufficient. It is therefore very desirable for a code to have
a reasonable chance of being uniquely decodable beyond the
error-correction guarantee, especially if no other properties are
compromised.
Consider a channel which makes t insertions into a codeword c and outputs each distinct length n + t received word
with equal probability. The probability of unique decoding for
a particular codeword c is equal to the fraction of length n + t
received words that are unique to c. We refer to this channel
as the uniform t-supersequence channel.
Now consider the channel where t insertions occur, one
at a time. We refer to the temporal list of insertions as a tinsertion history. An insertion is represented by a tuple of the
Some of these results were be presented at the IEEE International Symposium on Information Theory (ISIT) 2017.
form: (position, element). For the codeword 00, an example
insertion history is [(0, 1), (0, 0)], giving the received word
0100.
If for each insertion, the element and position are chosen
uniformly at random, we refer to this channel as the uniform tinsertion channel. If a particular codeword c in a code C ⊆ Fnq
is passed through this channel, then the probability that unique
decoding occurs is equal to the fraction of t-insertion histories
that produce a received word unique to c.
We also define the analogous channels for t deletion errors
in Section II.
Clearly, we can make optimizations involving these probabilities over codes of a particular cardinality, length, and errorcorrection capability. While it is not yet clear how to perform
such optimizations in a non-exhaustive manner, upper bounds
on the probabilities are useful because they help us understand
how effective a code can possibly be for uniquely decoding a
given number of insertions or deletions, without solving the
optimization problems explicitly. The bounds can thus be used
as a reference when designing codes, and can be used to prove
theorems. The purpose of focusing on these channels is to
provide a framework for understanding what happens when t
insertions or deletions occur, regardless of whether the real-life
channel always makes t insertions or deletions.
The remainder of the paper is organized as follows. In
Section II, we present the necessary preliminaries. In Section
III we provide tight upper bounds on the probability of unique
decoding for both insertion channels, discuss the positivity of
the measures, and present improved upper bounds for insertion
channels that apply to VT-codes and linear codes. In Section
IV, we derive tight upper bounds on the probability of unique
decoding for one of the deletion channels, and give improved
upper bounds with additional assumptions. In Section V, we
make observations about the behavior of VT-codes, and raise
open questions. We conclude the paper in Section VI.
II. P RELIMINARIES
Let Fq = {0, 1, . . . , q − 1} be an alphabet containing q ∈ N
symbols. We only consider q ≥ 2 throughout the paper. Fnq is
the set of all length n words over the alphabet Fq . If a ∈ Fq ,
we denote the word
aa
. . . a}
| {z
n a’s
by an . For a sequence c ∈ Fnq , let c[i] be the ith element
in c where i ∈ {0, 1, ..., n − 1}. We use the terms word and
sequence interchangeably.
Let x ∈ Fnq . We define an insertion as the addition of
an element from Fq into some position in x. We define a
deletion as the removal of an element from x. If t insertions
occur, the resulting word is referred to as a t-supersequence
of x. Similarly, if t deletions occur, the resulting word is
referred to as a t-subsequence of x. A substring is a contiguous
subsequence of a word. Given a word, a run is a substring
an , such that the potential elements on both sides of an are
not a. The Levenshtein distance dL (x, y) between two words
x ∈ Fnq 1 , y ∈ Fnq 2 , is defined as the minimum number
of insertions and deletions necessary to transform x into y.
Clearly, we have that dL (x, y) = dL (y, x).
The t-insertion ball of x is the set of all words in Fn+t
q
that are formed by inserting t symbols into x. We denote the
t-insertion
by It (x). It is known that |It (x)| =
ball of x
Pt
n+t
i
(q
−
1)
for
any word x ∈ Fqn . Similarly, the
i=0
i
t-deletion ball of x is the set of all words in Fqn−t formed
by deleting t symbols from x. We denote the t-deletion ball
of x by Dt (x). Unfortunately, a general formula for |Dt (x)|
is not known. However in the t = 1 case, we have that
|D1 (x)| = r(x), where r(x) is the number of runs in x.
Because |It (x)| is independent of the exact length n sequence x, we will sometimes use the expression It (n, q) to
mean the number of sequences in the q-ary t-insertion ball of
a length n sequence.
We define a t-insertion correcting code to be a set of
codewords C ⊆ Fnq such that It (c1 ) ∩ It (c2 ) = ∅ ∀c1 , c2 ∈
C, c1 6= c2 . Similarly, we define a t-deletion correcting code
to be a set of codewords C ⊆ Fnq such that Dt (c1 ) ∩ Dt (c2 ) =
∅ ∀c1 , c2 ∈ C, c1 6= c2 .
It was shown that C is a t-insertion correcting code if
and only if it is a t-deletion correcting code. Furthermore, it
was also shown that C is a t-insertion correcting code if and
only if dL (c1 , c2 ) > 2t ∀c1 , c2 ∈ C, c1 6= c2 . We denote
the minimum Levenshtein distance of an insertion/deletion
correcting code by dmin , and denote the cardinality of the code
by M.
Varshamov-Tenengolts codes introduced in [1] are commonly used single-insertion correcting codes defined as the
set of all words x = (x1 , ..., xn ) ∈ Fn2 such that
n
X
ixi ≡ a ( mod n + 1)
i=1
for some a such that 0 ≤ a ≤ n.
Varshamov-Tenengolts codes are perfect, and are optimal
for n = 1, ..., 9 when a = 0. It is conjectured, though unproven
that Varshamov-Tenengolts codes are optimal for all values of
n when a = 0. We refer to the Varshamov-Tenengolts code of
length n with parameter a as V Ta (n). Some excellent general
resources on insertions and deletions can be found in [6], [7],
and [8].
We focus on two distinct insertion channels, namely the
uniform t-supersequence channel, which we denote by USCt ,
and the uniform t-insertion channel, which we denote by UICt .
These channels are both defined in the introduction.
We consider two deletion channels which make t deletions.
The first is the uniform t-deletion channel, denoted by UDCt .
This channel makes t sequential deletions, where the deletion
at each step is chosen with equal probability. This is equivalent
to saying each t-deletion history occurs with equal probability.
Here, a t-deletion history is represented as temporal list of
deletions of length t, where the ith deletion is represented by
the index of the element in the length n − i + 1 word that is
deleted at step i ∈ {1, ..., t}. For example, if the word 001100
is affected by the 2 deletion history [0, 1], the received word
would be 0100.
The second deletion channel is the uniform t-subsequence
channel, which outputs each distinct t-subsequence of a word
with equal probability. We denote this channel by UBCt .
We define a unique t-supersequence of a codeword c ∈ C ⊆
Fnq as a t-supersequence of c that is not a t-supersequence of
any other codeword in C.
We define a unique t-subsequence of a codeword c ∈ C ⊆
Fnq as a t-subsequence of c that is not a t-subsequence of any
other codeword in C.
We consider two measures of a code’s effectiveness for
uniquely decoding beyond its error-correction guarantee. Let
fK (c, C) be the probability of c ∈ C being uniquely decodable after being passed through channel K. The first
measure is WK (C) = minc∈C fK (c, C). This measure addresses the worst case distribution on the codeword sent
i.e. no matter which codeword is sent, the probability of
unique decoding isPat least WK (C). The second measure
1
is UK (C) = M
c∈C fK (c, C). This measure gives the
probability of unique decoding if each codeword is sent with
equal probability.
A WK optimal code is a code that solves
max
C: {n,q,dmin ,M } fixed
WK (C).
A UK optimal code is a code that solves
max
C: {n,q,dmin ,M } fixed
UK (C).
Finally, we define B(n, q, K) as a tight upper bound on
fK (c, C) over all codes C ⊆ Fnq such that |C| ≥ 2, and over
all codewords c ∈ C.
Similarly, we define Bd (n, q, K) as a tight upper bound on
fK (c, C) over all codes C ⊆ Fnq such that |C| ≥ 2, and over
all codewords c ∈ C such that there is a codeword c0 ∈ C
where dL (c, c0 ) = 2d for 1 ≤ d ≤ n.
III. B OUNDS FOR I NSERTIONS
A. General Bounds
In this subsection we find B(n, q, K) for K ∈
{USCt , UICt }, consider the limiting behavior of the upper
bounds as t or n increases, and establish the positivity of UK
for K ∈ {USCt , UICt } and all t. Recall that q ∈ N, q ≥ 2
throughout the paper.
We begin by recalling a recursion for the intersection cardinality of two insertion balls. This recursion was discovered
by Levenshtein in [4], where he studied what is now known
as Levenshtein’s reconstruction problem.
Proof:
Lemma 1. Let X 0 ∈ Fn+t−k
, Y 0 ∈ Fnq , where n, t, k are
q
positive integers such that k < n + t. Write X 0 = aX and
Y 0 = bY with a, b ∈ Fq . Then, if a = b,
N̈q+ (0, n, n + t, t) =
0
|Ik (X ) ∩ It (Y )| =
t
X
(q − 1)k
k=0
|Ik (X) ∩ It (Y )| + (q − 1)|Ik−1 (aX) ∩ It−1 (aY )|.
If a 6= b,
0
k=0 i=0
=
0
0
|Ik (X ) ∩ It (Y )| = |Ik (X) ∩ It−1 (bY )|+
|Ik−1 (aX) ∩ It (Y )| + (q − 2)|Ik−1 (aX) ∩ It−1 (bY )|.
We define the minimum intersection cardinality as
N̈q+ (n1 , n2 , t1 , t2 ) =
min
X∈Fqn1 ,Y ∈Fqn2
|It1 (X) ∩ It2 (Y )|
where n1 , n2 , t1 , t2 ∈ N, and t1 + n1 = t2 + n2 . Clearly,
N̈q+ (n1 , n2 , t1 , t2 ) = N̈q+ (n2 , n1 , t2 , t1 ). In Theorem 1, we
give a closed form for N̈q+ (n1 , n2 , t1 , t2 ). Levenshtein derived
the maximum intersection analogue of N̈q+ (n1 , n2 , t1 , t2 ) in
his study of sequence reconstruction [4].
Theorem 1. Let n1 , n2 , t1 , t2 ∈ N and t1 + n1 = t2 + n2 .
Then, we have
t2 k−n
X1 k
X
k=n1 i=0
i
(q − 2)i
n2 + t2
.
k
To prove Theorem 1 , we found It1 (0n1 )∩It2 (1n2 ) to equal
the formula in the theorem. We then proved the formula is the
minimum intersection cardinality using an inductive argument
on n1 + t1 . Lemma 1 was used in the inductive step. The
proof is very long, and is given in the appendix. It should
be noted that every pair of sequences at Levenshtein distance
n1 +n2 achieves the minimum intersection cardinality, as given
in Lemma 2.
Lemma 2. Let n1 , n2 , t1 , t2 ∈ N and N = t1 + n1 = t2 +
n2 . Then for any length n1 sequence X and any length n2
sequence Y such that dL (X, Y ) = n1 + n2 , we have that
|It1 (X) ∩ It2 Y )|
=
N̈q+ (n1 , n2 , t1 , t2 )
=
t2 k−n
X
X1 k
k=n1 i=0
i
i
(q − 2)
n2 + t2
.
k
In addition, the following corollary can be proved logically,
or through an application of the binomial theorem as shown
below.
Corollary 1. For n, t ∈ N, we have that
N̈q+ (0, n, n + t, t) = It (n, q).
n+t
k
i
(q − 2)i
n+t
k
= It (n, q).
where the second equality follows from the binomial theorem.
With Theorem 1, we were able to derive B(n, q, USCt ).
Consider any length n code C with cardinality ≥ 2. The
following theorem gives a tight upper bound on the fraction
of unique t-supersequences for any c ∈ C. It is thus an upper
bound on the probability of unique decoding for any codeword
under the uniform t-supersequence channel.
Theorem 2. For n, t ∈ N, we have that
Pt
Pk−n
k
j n+t
j (q − 2)
k
Pt
n+t
i
(q
−
1)
i=0
i
k=n
B(n, q, USCt ) = 1 −
!
j=0
.
Proof: In addition to c, there must exist another codeword
c0 since |C| ≥ 2. The quantity |It (c) ∩ It (c0 )| must be greater
than or equal to
N̈q+ (n, n, t, t)
N̈q+ (n1 , n2 , t1 , t2 ) =
t X
k
X
k
t k−n
X
X k
j n+t
(q − 2)
=
k
j
j=0
k=n
as proved in Theorem 1. Thus, the number of sequences unique
to c in It (c) must be less than or equal to
t
t k−n
X
X
X k
n+t
i
j n+t
(q − 1) −
(q − 2)
i
j
k
i=0
j=0
k=n
The upper bound equals B(n, q, USCt ) because
fUSCt (0n , {0n , 1n } achieves the upper bound as proved
in Lemma 2.
In addition to providing an upper bound on the fraction
of unique t-supersequences for a particular codeword, the
formula above serves as a tight upper bound on WUSCt (C)
and UUSCt (C).
The upper bound on the fraction of unique t-supersequences
approaches zero as t goes to infinity for fixed n in the binary
case as proved in Lemma 3 and exemplified in Figure 1. The
bound is clearly equal to one as n goes to infinity for fixed t
because the minimum intersection cardinality is only positive
for t ≥ n. If t = n + O(na ) for a < .5, our tight upper bound
approaches one in the binary case as n goes to infinity. This
is proved in Lemma 4 and is relevant to the questions raised
in Section V.
Lemma 3. For n, t ∈ N, we have
lim B(n, 2, USCt ) = 0.
t→∞
Proof:
lim B(n, 2, USCt )
Pt Pk−n
t→∞
=
=
k
j n+t
j (q − 2)
k
lim 1 −
Pt
n+t
i
t→∞
(q
−
1)
i=0
i
Pt
n+t
k
1 − lim Pk=n
=0
t
n+t
t→∞
i=0
i
k=n
!
j=0
The first equality follows from plugging in q = 2. The second
equality follows from the fact that the n extra terms in the
denominator are dominated by the other terms as t increases.
Lemma 4. For all values of n ∈ N and a < .5, we have
lim B(n, 2, USCn+O(na ) ) = 1.
n→∞
Proof: Let f : Z → Z be some function. For q = 2 and
letting t = n + f (n),
lim B(n, 2, USCn+f (n) )
Pn+f (n) Pk−n
n→∞
=
=
k
j 2n+f (n)
j=0 j (q − 2)
k
lim 1 −
Pn+f (n) 2n+f (n)
n→∞
i
(q
−
1)
i=0
i
Pn+f (n) 2n+f (n)
k=n
k
1 − lim Pn+f
(n) 2n+f (n)
n→∞
i=0
i
n+f (n)/2
(f (n) + 1) √ 4
3(n+f (n)/2)+1
!
k=n
≥ 1 − lim
2f (n)−1 4n
2(f (n) + 1)
n→∞
= 1 − lim p
n→∞
3(n + f (n)/2) + 1
.
Letting f (n) = O(na ) where a < .5, then
1 − lim p
n→∞
2(O(na ) + 1)
3(n + O(na )/2) + 1
= 1.
The third line follows from plugging in q = 2. To
obtain
the numerator
in the fourth line, we observe that
Pn+f (n)
n+f (n)/2
2n+f (n)
using the
≤
(f (n) + 1) √ 4
k=n
k
3(n+f (n)/2)+1
upper bound on the central binomial coefficient in [23], and
the fact the 2n
obtain the denomk is maximized at k = n. To
Pn+w
inator in the fourth line, we observe that i=0 2n+fi (n) ≥
Pn+f (n)/2 2n+f (n) 1 2n+f (n)
= 22
= 2f (n)−1 4n .
i=0
i
The proof when 2n + f (n) is odd uses the same ideas after
an application of Pascal’s rule to the binomial coefficients in
the third line.
To improve this upper bound for a code at minimum
Levenshtein distance dmin , we could find Bdmin (n, q, USCt ).
This could be approached by finding the minimum insertionball intersection cardinality over all pairs of sequences that
have Levenshtein distance ≤ dmin , and proving that this
measure increases as dmin decreases. Such a formula could
be thought of as the dual of the formula derived by Sala et
al. in their study of Levenshtein’s reconstruction problem [9],
which gives the maximum possible insertion-ball intersection
cardinality over all sequence pairs with Levenshtein distance
≥ dmin . The derivation of this formula in conjunction with
[9] would result in a spectrum of insertion ball intersection
cardinalities according to the Levinshetin distance between the
sequences in the pair. It would also be interesting to derive
a formula for how many sequence pairs have a particular
Levenshtein distance.
To upper bound the probability of unique decoding for
the uniform t-insertion channel, we find the analogue of the
minimum intersection cardinality for insertion histories in
Theorem 3.
Theorem 3. For any distinct sequences X, Y ∈ Fnq with
n, t ∈ N, the fraction of t-insertion histories P
for X giving
t−n
t-supersequences in It (Y ) is lower bounded by i=0 ti (q −
1)i Πti=1 (n + i). The lower bound is tight.
Proof: The number of insertion histories for X where
Y is a subsequence of thePinsertion
pattern is equal to
t−n t
i t
It−n (n, q)Πti=1 (n + i) =
i=0 i (q − 1) Πi=1 (n + i).
t
This is because there are Πi=1 (n + i) ways to choose the
positions for an insertion history. Once the insertion positions
are chosen, the insertions clearly appear in some order in the
resulting sequence. So for each history of insertion positions,
we can assign a length t sequence to the ordered insertion
positions in the resulting sequence. There are It−n (n, q)
sequences that have Y as a subsequence. Thus, there are
It−n (n, q)Πti=1 (n + i) total insertion histories, where Y is
a subsequence of the insertion pattern.
So, the number of t insertion histories forPX giving
tt−n
supersequences in It (Y ) is lower bounded by i=0 ti (q −
1)i Πti=1 (n + i).
The pair X = 0n and Y = 1n achieves this upper bound
because 0n and 1n have no elements in common, so the only
elements in a t-supersequence of X that form a copy of Y,
must be inserted elements. By the same logic, any pair at
Levenshtein distance 2n achieves the upper bound;
Theorem 4. For n, t ∈ N, we have that
Pt
t
B(n, q, UICt ) =
i=t−n+1 i
qt
(q − 1)i
.
Proof: There must be another codeword c0 . The number
of t insertion historiesPfor c giving
t-supersequences in It (c0 )
t−n t
is lower bounded by i=0 i (q − 1)i Πti=1 (n + i).
There are q t Πti=1 (n + i) insertion histories total.
Thus,
Pt−n t
(q − 1)i Πti=1 (n + i)
fUICt (c, C) ≤ 1 − i=0 i t t
q Πi=1 (n + i)
Pt
t
i t
(q − 1) Πi=1 (n + i)
= i=t−n+1 t i t
q Πi=1 (n + i)
Pt
t
i
i=t−n+1 i (q − 1)
=
qt
where the second line follows from the binomial theorem.
fUICt (0n {0n , 1n }) achieves the upper bound as shown in the
proof of Theorem 3, so the upper bound equals B(n, q, UICt ).
B(n, q, UICt ) is also an upper bound on UUICt (C) and
WUICt (C). The following two lemmas give the limiting behavior of the tight upper bound as t and n increase. The behavior
in the lemmas is exemplified in Figure 2. Interestingly the
bound in Figure 2 decays much slower than the bound for
USCt in Figure 1.
Lemma 5. For n, t ∈ N, we have limt→∞ B(n, q, UICt ) = 0.
Proof: We have
Pt
lim B(n, q, UICt ) = lim
t→∞
t→∞
t
i=t−n+1 i
qt
(q − 1)i
Fig. 2.
= 0.
because there are a finite number of terms in the numerator,
each of which grows slower than q t .
Lemma 6. For n ∈ N and constant b ∈ N, we have
limn→∞ B(n, q, UICn+b ) = 1.
Proof:
Pn+b
i=b+1
lim B(n, q, UICn+b ) = lim
n→∞
Pb
n+b
(q − 1)i
= lim 1 − i=0 in+b
n→∞
q
Pb
n+b
i
i=0
i (q − 1)
= 1 − lim
=1
n→∞
q n+b
n→∞
n+b
i
q n+b
(q − 1)i
The second line follows from the binomial theorem, and the
last equality follows because there are a constant number
of terms in the numerator that each grow slower than the
denominator.
In addition to showing upper bounds on the probabilities of
unique decoding, we can show that for both insertion channels
and every code C, there exists some codeword where the
probability of unique decoding is non-zero for all t. This
proves the positivity of UK (C) for all codes C, for both
channels of interest.
Lemma 7. For every code C ⊆ Fnq , and n ∈ N there exists
a codeword c ∈ C such that for all t ∈ N, c has at least one
unique t-supersequence.
Proof: Consider a codeword c with maximal value of
rr (c), where rr (c) is the length of the right-most run in c.
Suppose that for some t ∈ N, c did not have at least one
unique t-supersequence. Let a be the substring of c to the left
of c’s rightmost run. Consider the supersequence s formed by
duplicating the right-most element t times. Then one or more
length n subsequences of s must be codewords in C that are
distinct from c. Any such codeword c0 must be a length n
subsequence of s. c0 therefore must be of the form xy where
x is a strict subsequence of a, and y is the rightmost element
of c repeated n − |x| times. x must be a strict subsequence
of a, because if it was not, then c0 would be the same as c.
Since x is a strict subsequence, y must be of length greater
than rr (c). Thus, c0 has a rightmost run of length greater than
rr (c). This is a contradiction.
Theorem 5. For every code C ⊆ Fnq , and n, t ∈ N, we have
that UK (C) is strictly positive for K ∈ {USCt , UICt }.
B. Bounds with Additional Assumptions
In this subsection, we find upper bounds for the probability
of unique decoding when additional assumptions are made
about the code and the codeword. All of the results here are
applicable to VT-codes of even length with a = 0. Recall that
VT-codes with a = 0 have maximal cardinality over all choice
of a, and are asymptotically optimal single insertion/deletion
correcting codes.
Fig. 1.
Lemma 8. For any binary code C ⊆ Fn2 with n ∈ N, such that
|C| ≥ 3 and {0n , 1n } ⊂ C, we have that fK (c, C) = 0 for any
c∈
/ {0n , 1n }, for K ∈ {USCt , UICt } and t ≥ n − 1, t ∈ N.
Proof: For any codeword c ∈
/ {0n , 1n }, consider In−1 (c).
Suppose c is composed of x1 1’s and x0 0’s. Observe that
1 ≤ x0 , x1 ≤ n − 1.
By proving the statement for t = n − 1, we prove the
statement for t ≥ n − 1. Consider a sequence s in In−1 (c),
and let s1 be the number of ones in s, and s0 be the number
of zeros in s.
If ≥ n − x1 ones are inserted, then s is a supersequence of
1n . If < n − x1 ones are inserted, then > n − 1 − (n − x1 ) =
x1 − 1 zeros are inserted, and thus s0 > x0 + x1 − 1. So, s is
a supersequence of 0n .
Corollary 2. For any binary code C ⊆ Fn2 with n ∈ N, such
that |C| ≥ 3 and {0n , 1n } ⊂ C, we have that WK (C) = 0
for K ∈ {USCt , UICt } and t ≥ n − 1, t ∈ N.
Corollary 3. For any binary code C ⊆ Fn2 with n ∈ N, such
1
that |C| ≥ 3 and {0n , 1n } ⊂ C, we have that UK (C) ≤ |C|
for K ∈ {USCt } and t ≥ n − 1, t ∈ N.
the there are Πtj=1 (n+j) ways to choose the insertion position
history.
Given the i and the insertion position history, there are
t
ways
to choose which insertions are zero insertions, and
i
(q − 1)t−i ways to choose the elements for the remaining t − i
insertions.
There are q t Πtj=1 (n + j) insertion histories total, so the
result follows.
Lemma 9 applies to V T0 (n) of even length. Lemma 10
holds for any linear code since all linear codes contain 0n .
Lemma 10 also holds for V T0 (n) of any length. Provided
there is a codeword c ∈ C of weight w, the corresponding
bounds in Lemmas 9 and 10 clearly serve as upper bounds on
WUICt (C). Given the codeword weight distribution of a code,
Lemmas 9 and 10 can be used to upper bound UUICt (C). In
Figure 3, we plot the bound in Lemma 9 and the probability of
unique decoding for the codewords in V T0 (6) for the uniform
insertion channel. The plot also exemplifies the behavior at
t = n − 1 presented in Lemma 8.
Proof: There are |C| − 2 codewords such that
fUSCt (c, C) = 0. Therefore, at least one codeword c ∈
/
{0n , 1n } has all of its t-supersequences in It (0n ) ∪ It (1n ).
Thus, the number of non-unique t-supersequences in It (0n ) ∪
It (1n ) is ≤ 2It (n, q) − It (n, q) = It (n, q).
V T0 (n) codes of even n always contain 0n and 1n . So the
above results apply.
With knowledge of a codeword’s weight, we are able to find
upper bounds for UICt that drop below 1 before t = n. These
results are also applicable to VT-codes.
Lemma 9. For any binary code C ⊆ Fn2 with n ∈ N, such that
|C| ≥ 2 and {0n , 1n } ⊂ C, we have that for any codeword c ∈
C with
weight 1 P
≤ w ≤ n − 1, it follows that fU ICt (c, C) ≤
Pt
t
t
t
i=w (i)
i=n−w (i)+
1−
for 1 ≤ t ≤ n − 1, t ∈ N.
2t
t
Pt
t
Proof: There are
i=n−w i Πj=1 (n + j) t-insertion
histories for c that give sequences in It (1n ). Call this set of
insertion histories
Pt H1.
There are i=w ti Πtj=1 (n + j) t-insertion histories for c
that give sequences in It (0n ). Call this set of insertion histories
H2 .
Each history in H1 consist of ≥ n − w one insertions, and
thus ≤ t − (n − w) zero insertions. For t ≤ n − 1, this implies
that ≤ w − 1 zero insertions are made. Each history in H2
consists of ≥ w zero insertions, and thus, H1 ∩ H2 = 0.
There are 2t Πtj=1 (n + j) insertion histories total, so the
result follows.
Lemma 10. For any code C ⊆ Fnq with n ∈ N, such that
|C| ≥ 2 and 0n ∈ C, we have that for any codeword
cP∈ C with weight w ≥ 1, it follows that fU ICt (c, C) ≤
w−1 t
t−i
i=0 (i)(q−1)
for t ∈ N.
qt
Pw−1 t
t−i t
Proof: There are
Πj=1 (n + j) ti=0 i (q − 1)
insertion histories for c that give sequences not in It (0n ). This
is because only t-superseqeunces containing ≤ w − 1 inserted
zeros are not in It (0n ). For each number of inserted zeros i,
Fig. 3. Note: The red curves are underneath the purple curves. The blue curve
is underneath the black curve.
IV. B OUNDS FOR D ELETIONS
In this section, we consider the uniform t-deletion channel
UDCt , which is defined in Section II. Specifically, we derive
Bd (n, q, UDCt ), and derive improved upper bounds with
additional assumptions.
Recall that UDCt incurs each t-deletion history with equal
probability. As an example, if the word 001100 is affected by
the 2-deletion history [0, 1], the received word would be 0100.
Clearly, without any assumptions other than |C| ≥ 2, the
tight upper bound on the probability of unique decoding is 1 if
t < n, and 0 if t = n, for both the uniform t-deletion channel
and the uniform t-subsequence channel. This can be seen by
taking the code C = {0n , 1n }.
Our main result for this section is the derivation of
Bd (n, q, UDCt ), a tight upper bound on the probability of
unique decoding for the uniform t-deletion channel, for a
codeword c ∈ C, when there is guaranteed to be a codeword
c0 ∈ C such that dL (c, c0 ) = 2d for 1 ≤ d ≤ n.
To accomplish this, we must first prove a bijection between
the set of all t-deletion histories to the set of all t-deletion
patterns. Here, a t-deletion pattern is a temporal list of indices
there are d! ways to order the deletions. There are then
n−d
t−d ways to choose the remaining indices corresponding to
elements in c that are deleted. Once these are chosen, there are
(t − d)! ways
to order
these deletions. Thus, there are at least
d!(t − d)! dt n−d
t−d insertion patterns that give subsequences
in Dt (c) ∩ Dt (c0 ).
This bound is tight because the codeword 1d 0(n−d) in the
code {1d 0(n−d) , 0n } clearly achieves the bound since there is
only one possible set S of size d. This is S = {0, 1, .., d−1}.
Fig. 4.
of the original elements removed from the codeword. For
example, if the word c = 001100 is affected by the 2deletion pattern [0, 1], the received word would be 1100. This
is because the first element deleted is c[0], and the second
element deleted is c[1].
Interestingly, this bound is alphabet free. The result also
serves as an upper bound on WUDCt (C). Because the upper
bound is tight, the asymptotic analysis of a constant number
of extra insertions is of interest. This is given in the Lemma
12.
Lemma 12. For constant b ∈ N, we have that
lim Bd (n, q, UDCd+b ) = 1.
n−>∞
Lemma 11. There exists a bijection between the set of tdeletion patterns and the set of t-deletion histories.
Proof: We will show there is a bijective mapping from
the set of t-deletion histories to the set of t-deletion patterns.
Given a t-deletion history, we map it to the resulting t-insertion
pattern that occurs.
Any t-deletion pattern clearly has a t-deletion history that
maps to it. Thus the mapping is onto.
Now consider the two distinct t-deletion histories. Consider
the first position k where they differ. After the first k − 1
deletions, the two deletion histories give the same sequence
(with same original codeword elements). Since the kth deletion
is different, a different original element from the codeword
is deleted from the word. Thus the corresponding deletion
patterns differ in the kth element. Thus, the mapping is also
one-to-one, and is thus a bijection.
We now give the central result for the uniform t-deletion
channel.
Proof: We have that
lim Bd (n, q, UDCd+b )
n−d
d!b! d+b
(n − d − b)!
b
d
= lim 1 −
n−>∞
n!
d!b! d+b
(n
−
d)!(n − d − b)!
d
= 1 − lim
n−>∞
(n − d − b)!b!n!
d! d+b
d (n − d)!
= 1 − lim
n−>∞
n!
d! d+b
d
= 1 − lim
= 1.
n−>∞ Πd−1 (n − i)
i=0
n−>∞
One of the difficulties in finding an analogous upper bound
for
the uniform t-subsequence channel lies in the fact that
Theorem 6. For n, t ∈ N and 1 ≤ d ≤ n, we have that
there
is no general formula for the size of a t-deletion ball.
(
1
for 1 ≤ t < d
In contrast to the result for insertions in Lemma 7, there
Bd (n, q, UDCt ) =
.
d!(t−d)!(dt )(n−d
t−d )(n−t)!
exist
codes where all codewords have 0 unique t-subsequences
1−
for d ≤ t ≤ n
n!
for t ≥ b dmin
2 c + 1 deletions e.g. V T0 (6). Thus, the averProof: The total number of t-insertion histories is equal age probability of unique decoding for the uniform deletion
n!
to (n−t)!
. The formula works when t = n because 0! = 1.
channel and the uniform subsequence channel is not always
We proceed to work with the equivalent t-deletion pattern positive. This is an example of how the unique decoding
definition to calculate the numerator in the bound.
of extra deletions differs from the unique decoding of extra
Since the Levenshtein distance between c and c0 is 2d, we insertions.
know that there exists some set S of d positions in c, such
Finally, we can find upper bounds on the probability of
that if we delete the elements at those positions from c, the unique decoding for UDCt which make additional assumptions
resulting word will be a d-subsequence of c0 . We begin by about the code. The following is the analogue of Lemma 9 for
counting the number of ways to select the d steps in the deletions.
t-deletion pattern corresponding
to the deletions at indices
in S. There are dt ways to choose these d steps in the Lemma 13. For any binary code C ⊆ Fn2 with n ∈ N,
deletion pattern. Fixing those steps in the deletion pattern, such that |C| ≥ 2 and {0n , 1n } ⊂ C, we have that for any
codeword c ∈ C with weight 1 ≤ w ≤ n − 1, it follows that
fUDCt (c, C)
(n − t)!w!(t − w)! wt n−w
t−w
n
≤ 1 − Aw t,
n!
t
(n
−
t)!(n
−
w)!(t
− n + w)! n−w
n
− An−w t,
n!
n 0’s
w
t−n+w
for 1 ≤ t < n, t ∈ N, and
fUDCt (c, C) = 0
for t = n, where
Anx (t, e)
(
0
=
e
limn→∞ UK (C) = 0 and limn→∞ WK (C) = 0 for K ∈
{USCb dmin c+1 , UICb dmin c+1 , UDCb dmin c+1 , UBCb dmin c+1 }?
2
2
2
2
On the other hand, for the code
C = {00
. . . 0}, 00
. . . 0} 11
. . . 1}, 11
. . . 1} 00
. . . 0}, 11
. . . 1}},
| {z
| {z
| {z
| {z
| {z
| {z
for 1 ≤ t < x
.
for x ≤ t < n
Proof: The number of t-deletion histories giving
se-
quences in |Dt (c) ∩ Dt (0n )| is equal to w!(t − w)! wt n−w
t−w
by the logic used in the Theorem 6. Similarly, the number of
t-deletion histories giving sequences
inw|Dt(c) ∩ Dt (1n )| is
t
equal to (n − w)!(t − n + w)! n−w
t−n+w . The first set of
insertion histories results in sequences of all 0’s. The second
set of insertion histories results in sequences of all 1’s. Thus,
they are disjoint for 1 ≤ t < n, t ∈ N.
V. O BSERVATIONS
For either insertions or deletions, the most practical value
of t to first consider is b dmin
2 c + 1 i.e. one greater than the
guaranteed number of correctable insertions/deletions. In this
section we discuss the observed behavior of various codes
for a small number of extra insertions and raise some open
questions.
Recall the single insertion correcting VT-code. Given a VTcode of length n, it is known that a = 0 always gives the
highest cardinality, though there are often multiple values of
a that give rise to VT-codes of maximum cardinality.
Among a values that give maximum cardinality, we can select the one which maximizes WK (V Ta (n)), or UK (V Ta (n))
for K ∈ {USCt , UICt , UDCt , UBCt }, for some t > 1 of our
interest. In the case of t = 2 insertions for example, decoding
is efficient. Given any algorithm for decoding 1 insertion that
outputs a subsequence of the received word (e.g. using the
general approach given in [2]), 2 insertions can be decoded
by simply running that decoding algorithm on each of the 1subsequences of the received word and checking whether the
decoding is a codeword.
We observed that by varying a, WK (V Ta (n)) and
UK (V Ta (n)) for K ∈ {USC2 , UIC2 , UDC2 , UBC2 } can
change significantly for smaller n, making this a practical
observation for relatively small n. However, as n increases, we
have observed that WK (V Ta (n)) and UK (V Ta (n)) approach
0 for K ∈ {USC2 , UIC2 , UDC2 , UBC2 }. Through our observations, we ask the following question about the USC2 , UIC2 ,
UDC2 , and UBC2 .
For an optimal length n code C with minimum Levenshtein distance dmin = Θ(1), is it always true that
n
2
0’s
n
2
1’s
n
2
1’s
n
2
0’s
n 1’s
we proved that limn→∞ UK (C) = 1 and limn→∞ WK (C) =
1 for K ∈ {USC n2 , UIC n2 }. The code C = {0n , 1n }
is an example where for limn→∞ UK (C) = 1 and
limn→∞ WK (C) = 1 for K = USCb dmin c+b , UICb dmin c+b
2
2
for constant b ∈ N (see Lemmas 4, 6). Through these
observations, we ask the following question.
For a length n code C with minimum Levenshtein distance
dmin = Θ(n), is it always true that limn→∞ UK (C) = 1 and
limn→∞ WK (C) = 1 for K ∈ {USCb dmin c+b , UICb dmin c+b }
2
2
and constant b ∈ N?
We are also interested in the analogues of these questions
when dmin and b are different functions of n.
VI. C ONCLUSION
In this work, we examined the unique decoding capability
of codes from insertions and deletions beyond the guaranteed
number of correctable insertions/deletions. We defined several
probabilistic channels as a framework to understand what
happens when t insertions or t deletions occur, regardless of
whether the real-life channel always makes t insertions or t
deletions. We computed tight upper bounds on the probability
of unique decoding for the channels, and studied the limiting
behavior of the bounds. We then found upper bounds specific
to particular classes of codes, such as linear codes and VTcodes. We also studied the positivity of the measures, and
raised several open questions.
R EFERENCES
[1] R. R. Varshamov and G. M. Tenengolts, “Codes which correct single
asymmetric errors (in russian),” Automatika i Telemkhanika, vol. 161,
no. 3, pp. 288-292, 1965.
[2] V. I. Levenshtein, “Binary codes capable of correcting deletions, insertions and reversals” (in Russian), Doklady Akademii Nauk SSSR, 163
(No. 4, 1965), 845-848. English translation in Soviet Physics Dokl., 10
(No. 8, 1966), 707-710.
[3] V.I. Levenshtein, “Efficient reconstruction of sequences,” Trans. Info.
Theory, vol. 47, no. 1, pp. 2-22, Jan. 2001.
[4] V.I. Levenshtein, “Efficient reconstruction of sequences from their subsequences or supersequences,” Journal of Combinatorial Theory, vol. 93,
no. 2, pp. 310-332, 2001.
[5] V. Guruswami and M. Sudan, “Improved decoding of Reed-Solomon
and algebraic-geometric codes,” In FOCS ’98, IEEE Computer Society
(1998), pp. 28-39.
[6] N.J.A. Sloane, “On single-deletion-correcting codes.” Codes and Designs:
Proceedings of a Conference Honoring Professor Dijen K. Ray-Chaudhuri
on the Occasion of His 65th Birthday, Ohio State University, May 18-21,
2000, 2002.
[7] P. A. Bours, “Codes for correcting insertion and deletion errors,” PhD
thesis, Eindhoven University of Technology, Jun. 1994.
[8] M. Mitzenmacher, “A survey of results for deletion channels and related
synchronization channels,” Probability Surveys, pp. 1-33, June 2009.
[9] F. Sala, R. Gabrys, C. Schoeny, and L. Dolecek, “Exact reconstruction
from insertions in synchronization codes,” IEEE Trans. Inform. Theory,
vol. 63, no. 4, pp. 2428-2445, Apr. 2017.
[10] F. Sala, R. Gabrys, L. Dolecek, “Gilbert-Varshamov-like Lower Bounds
for Deletion-Correcting Codes,” IEEE Information Theory Workshop
(ITW) 2014.
[11] A. A. Kulkarni and N. Kiyavash, “Nonasymptotic Upper Bounds for
Deletion Correcting Codes,”IEEE Transactions on Information Theory,
vol. 59, no. 8, pp. 5115-5130, 2013.
[12] D. Cullina, N. Kiyavash, and A. A. Kulkarni, “Restricted composition
deletion correcting codes,” IEEE Trans. Inf. Theory, vol. 62, no. 9, pp.
4819-4832, Sept. 2016.
[13] D. Cullina and N. Kiyavash, “An improvement to levenshtein’s upper
bound on the cardinality of deletion correcting codes,” IEEE Transactions
on Information Theory, vol. 60, no. 7, pp. 3862-3870, 2014.
[14] S. Jain, F. Farnoud, M. Schwartz, J. Bruck, “Duplication-Correcting
Codes for Data Storage in the DNA of Living Organisms,” ISIT 2016.
[15] S. M. H. T. Yazdi, Y. Yuan, J. Ma, H. Zhao, and O. Milenkovic.
“A Rewritable, Random-Access DNA-Based Storage System.” Nature
Scientific Reports, 5(14318), 2015.
[16] C. Schoeny, A. Wachter-Zeh, R. Gabrys, and E. Yaakobi, “Codes
correcting a burst of deletions or insertions,” IEEE Trans. Inform. Theory,
vol. 63, no. 4, pp. 1971-1985, Apr. 2017.
[17] K. A. S. Abdel-Ghaffar, F. Paluncic, H. C. Ferreira, and W. A. Clarke,
“On Helberg’s generalization of the Levenshtein code for multiple deletion/insertion error correction,” IEEE Trans. Inf. Theory 58 (2012), No.
3, 1804-1808.
[18] A. S. J. Helberg and H. C. Ferreira, “On multiple insertion/deletion
correcting codes,” IEEE Trans. Inf. Theory, vol. 48, no. 1, 305-308, Jan.
2002.
[19] L. J. Schulman and D. Zuckerman, “Asymptotically good codes correcting insertions, deletions, and transpositions,” IEEE Trans. Inf. Theory, vol.
45, no. 7, Nov. 1999.
[20] A. Wachter-Zeh, “List Decoding of Insertions and Deletions,” Jan. 2017.
[Online]. Available: https://arxiv.org/abs/1701.02612
[21] V. Guruswami and C. Wang, “Deletion codes in the high-noise and highrate regimes,” IEEE Trans. Inform. Theory, vol. 63, no. 4, pp. 1961-1970,
Apr. 2017.
[22] R. Venkataramanan, V. Swamy, and K. Ramchandran, “Low-complexity
interactive algorithms for synchronization from deletions, insertions, and
substitutions,” IEEE Trans. Inf. Theory, vol. 61, no. 10, pp. 5670-5689,
Oct. 2015.
[23] Kazarinoff N.D. Geometric inequalities, New York: Random House,
1961
VII. A PPENDIX
A. Proof of Theorem 1
Theorem 1. Let n1 , n2 , t1 , t2 ∈ N and t1 + n1 = t2 + n2 .
Then, we have
N̈q+ (n1 , n2 , t1 , t2 )
t2 k−n
X1 k
X
i n2 + t2
(q − 2)
. (1)
=
i
k
i=0
k=n1
First, we will show that
|It1 (0n1 ) ∩ It2 (1n2 )| =
Pt2 Proof:
Pk−n
1 k
i n2 +t2
for all n1 , n2 , t1 , t2 , q ∈ N
k=n1
i=0
i (q − 2)
k
where t1 + n1 = t2 + n2 . We will then prove that
min
X∈Fqn1 ,Y ∈Fqn2
|It1 (X) ∩ It2 (Y )| = |It1 (0n1 ) ∩ It2 (1n2 )|
by strong induction on N = n1 + t1 = n2 + t2 .
Without loss of generality, suppose that n1 ≤ n2 . We
observe that for a sequence to be in the intersection of
It1 (0n1 ) and It2 (1n2 ), the sequence must contain at least
n1 zeros and n2 ones. Thus if t2 < n1 , we have that
|It1 (0n1 ) ∩ It2 (1n2 )| = 0.
Let N = n1 + t1 = n2 + t2 . If t2 ≥ n1 , we observe
that It1 (0n1 ) ∩ It2 (1n2 ) consists of all length N sequences
containing N − k ones and p zeros where n1 ≤ k ≤ N − n2
and n1 ≤ p ≤ k.
To count the number of length N sequences with N − k
ones, we first observe that the set S0n1 containing all ksupersequences of 0n1 that contain no ones has cardinality
equal to Ik−n1 (n1 , q − 1). Let Ak (T ) be the set of all
sequences formed by inserting k ones into a sequence T .
Consider T1 , T2 ∈ S0n1 , T1 6= T2 . Then clearly |Ak (T1 ) ∩
Ak (T2 )| = 0. To count the number of sequences in Ak (T1 ),
we can view each of the k + 1 spaces adjacent to sequence
elements in T1 as a labeled bin, and the N − k ones that will
be inserted into the sequence as N − k unlabeled balls that
are being tossed into the k + 1 bins.
Using the formula, for unlabeled balls tossedinto labeled
bins, we see that |Ak (T1 )| = (N −k)+(k+1)−1
= N
=
k
(k+1)−1
n2 +t2
.
k
Thus, the number of supersequences in It1 (0n1 ) ∩It2 (1n2 )
with N − k ones is equal to Ik−n1 (n1 , q − 1) n2 k+t2
Summing over the possible values of k, we obtain
|It1 (0n1 ) ∩ It2 (1n2 )|
t2
X
n2 + t2
Ik−n1 (n, q − 1)
k
k=n1
t2 k−n
X
X1 n1 + (k − n1 )
n2 + t2
=
(q − 2)i
i
k
k=n1 i=0
t2 k−n
X1 k
X
n2 + t2
(q − 2)i
=
i
k
i=0
=
k=n1
We proceed to show that
min
X∈Fqn1 ,Y ∈Fqn2
|It1 (X) ∩ It2 (Y )| = |It1 (0n1 ) ∩ It2 (1n2 )|
by induction on N = n1 + t1 = n2 + t2 .
The property trivially holds for N = 0. For N = 1, |I0 (0)∩
I0 (1)| = 0 while in any other case, the cardinality must be
≥ 0 since set cardinalities are always non-negative.
Suppose the property holds for all N in the range 0 ≤
N ≤ T . For N = T + 1, we consider the pair of sequences
X = 0n1 and Y = 1n2 for any n1 , n2 , t1 , t2 ∈ N such that
t1 + n1 = t2 + n2 = T + 1. From Lemma 1, we see that
|It1 (0n1 ) ∩ It2 (1n2 )| =
|It1 (0n1 −1 ) ∩ It2 −1 (1n2 )| + |It1 −1 (0n1 ) ∩ It2 (1n2 −1 )|
+ (q − 2)|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )|.
Now consider two arbitrary sequences X = aX 0 ∈ Fqn1
and Y = bY 0 ∈ Fqn2 . We see from Lemma 1 that if a 6= b,
then
|It1 (X) ∩ It2 (Y )| =
|It1 (X 0 ) ∩ It2 −1 (bY 0 )| + |It1 −1 (aX 0 ) ∩ It2 (Y 0 )|
+ (q − 2)|It1 −1 (aX 0 ) ∩ It2 −1 (bY 0 )|.
By the inductive hypothesis,
|It1 (0n1 −1 ) ∩ It2 −1 (1n2 )| ≤ |It1 (X 0 ) ∩ It2 −1 (bY 0 )|
and
|It1 −1 (0n1 ) ∩ It2 (1n2 −1 )| ≤ |It1 −1 (aX 0 ) ∩ It2 (Y 0 )|
|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )| + |It1 (0n1 −1 ) ∩ It2 (1n2 −1 )|
tX
2 −1 k−n
X1 k
n2 + t2 − 1
=
(q − 2)i
i
k
k=n1 i=0
k−n
t2
1 +1
X
X
k
n2 − 1 + t2
+
(q − 2)i
i
k
k=n1 −1 i=0
tX
2 −1 k−n
X1
k
n2 − 1 + t2
=
(q − 2)i
i
k
k=n i=0
1
k − 1 n2 − 1 + t2
+
i
k−1
k−n
t2
1 +1
X
X
k
i n1 + t1 − 1
+
(q − 2)
i
k
i=0
and
|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )| ≤ |It1 −1 (aX 0 ) ∩ It2 −1 (bY 0 )|
.
Thus, if a 6= b,
|It1 (0n1 ) ∩ It2 (1n2 )| ≤ |It1 (X) ∩ It2 (Y )|.
Now consider the case where a = b. In this case, we have
|It1 (X) ∩ It2 (Y )| =
k=t2 −1
|It1 (X 0 ) ∩ It2 (Y 0 )| + (q − 1)|It1 −1 (aX 0 ) ∩ It2 −1 (bY 0 )|.
By the inductive hypothesis, we have
The second equality above follows from splitting up the
second term. This expression gives
|It1 (0n1 −1 ) ∩ It2 (1n2 −1 )| ≤ |It1 (X 0 ) ∩ It2 (Y 0 )|
and
t2X
−n1
t2 − 1
n1 + t1 − 1
(q − 2)i
i
t2 − 1
i=0
t2 −n
1 +1
X
t2
i n1 + t1 − 1
+
(q − 2)
i
t2
i=0
t2 −n
1 +1
X
t2
n2 + t2 − 1
= C1 + C2 +
(q − 2)i
i
t2
i=0
C1 +
|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )| ≤ |It1 −1 (aX 0 ) ∩ It2 −1 (bY 0 )|
.
We thus have that
|I t1 (X) ∩ It2 (Y )|
≥ |It1 (0n1 −1 ) ∩ It2 (1n2 −1 )|
+ (q − 1)|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )|
where
We will complete the proof by showing that
tX
2 −1 k−n
X1
k
n2 − 1 + t2
C1 =
(q − 2)
i
k
k=n1 i=0
k − 1 n2 − 1 + t2
+
i
k−1
|I t1 (0n1 −1 ) ∩ It2 (1n2 −1 )| + (q − 1)|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )|
= |It1 (0n1 −1 ) ∩ It2 (1n2 −1 )| + |It1 −1 (0n1 ) ∩ It2 −1 (1n2 )
+ (q − 2)|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )|
≥ |It1 −1 (0n1 ) ∩ It2 (1n2 −1 )| + |It1 (0n1 −1 ) ∩ It2 −1 (1n2 )|
i
+ (q − 2)|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )|
= |It1 (0n1 ) ∩ It2 (1n2 )|
and
To show the above, it suffices to prove that
|It1 −1 (0n1 ) ∩ It2 −1 (1n2 ) + |It1 (0n1 −1 ) ∩ It2 (1n2 −1 )|
C2 =
t2X
−n1
i=0
t2 − 1
i n1 + t1 − 1
(q − 2)
.
i
t2 − 1
≥ |It1 (0n1 −1 ) ∩ It2 −1 (1n2 )| + |It1 −1 (0n1 ) ∩ It2 (1n2 −1 )|
We begin by manipulating the left side of the inequality.
We now manipulate the right side of the inequality.
Thus, if a = b, we have
|It1 (0n1 −1 ) ∩ It2 −1 (1n2 )| + |It1 −1 (0n1 ) ∩ It2 (1n2 −1 )|
tX
k−n
2 −1
1 +1
X
k
n2 + t2 − 1
=
(q − 2)i
i
k
k=n1 −1 i=0
t2 k−n
X
X1 k
n2 + t2 − 1
+
(q − 2)i
i
k
k=n1 i=0
t2 k−n
X1
X
k
n2 − 1 + t2
i
(q − 2)
=
i
k
k=n1 i=0
k − 1 n2 − 1 + t2
+
i
k−1
tX
2 −1 k−n
X1
k
n2 − 1 + t2
i
=
(q − 2)
i
k
k=n1 i=0
k − 1 n2 − 1 + t2
+
i
k−1
t2X
−n1
t2
n2 − 1 + t2
i
+
(q − 2)
i
t2
i=0
t2 − 1 n2 − 1 + t2
+
i
t2 − 1
t2X
−n1
t2
n2 − 1 + t2
(q − 2)i
= C1 +
i
t2
i=0
t2 − 1 n2 − 1 + t2
+
i
t2 − 1
t2X
−n1
t2
n2 − 1 + t2
= C1 +
(q − 2)i
i
t2
i=0
t2X
−n1
n2 − 1 + t2
i t2 − 1
+
(q − 2)
i
t2 − 1
i=0
t2X
−n1
t2
n2 + t2 − 1
(q − 2)i
= C1 + C2 +
i
t2
i=0
Clearly,
t2 −n
1 +1
X
i=0
t2
n2 + t2 − 1
(q − 2)
i
t2
t2X
−n1
t2
n2 + t2 − 1
≥
(q − 2)i
i
t2
i=0
i
Thus, we have
|It1 −1 (0n1 ) ∩ It2 −1 (1n2 ) + |It1 (0n1 −1 ) ∩ It2 (1n2 −1 )|
≥ |It1 (0n1 −1 ) ∩ It2 −1 (1n2 )| + |It1 −1 (0n1 ) ∩ It2 (1n2 −1 )|
This proves,
|I t1 (0n1 −1 ) ∩ It2 (1n2 −1 )| + (q − 1)|It1 −1 (0n1 ) ∩ It2 −1 (1n2 )|
≥ |It1 −1 (0n1 ) ∩ It2 (0n2 −1 )| + |It1 (0n1 −1 ) ∩ It2 −1 (0n2 )|
+ (q − 2)|It1 −1 (0n1 ) ∩ It2 −1 (0n2 )|
|It1 (0n1 ) ∩ It2 (1n2 )| ≤ |It1 (X) ∩ It2 (Y )|
and we have proved the inductive step.
B. Proof of Lemma 2
Lemma 2 is restated below, and follows directly form the
proceeding auxiliary lemmas along with Theorem 1.
Lemma 2. Let n1 , n2 , t1 , t2 ∈ N and N = t1 + n1 =
t2 + n2 . Then for any length n1 sequence X and any length
n2 sequence Y such that dL (X, Y ) = n1 + n2 , we have that
|It1 (X) ∩ It2 Y )|
= N̈q+ (n1 , n2 , t1 , t2 ) =
t2 k−n
X
X1 k
k=n1 i=0
i
(q − 2)i
n2 + t2
.
k
Lemma 14. Let n1 , n2 , t1 , t2 ∈ N and N = t1 + n1 = t2 +
n2 . Then for any length n1 sequence X and any length n2
sequence Y such that dL (X, Y ) = n1 + n2 , we have that Y
only contains elements from some subset B of the alphabet
Fq , and X only contains elements from the subset Fq \ B.
Letting qB = |B|, we have that
|It1 (X) ∩ It2 (Y )|
=
N −n
1 −n2
X
k=0
N
Ik (n1 , q − qB )IN −k−n1 −n2 (n2 , qB )
.
t1 − k
Proof: Clearly, if dL (X, Y ) = n1 + n2 , we have that X
and Y do not have any elements in common. Thus we consider
the set of elements in Y as a new alphabet B, which is clearly
a subset of the alphabet Fq . We then consider the subset Fq \B
as the alphabet for X.
To count the number of sequences in It1 (X) ∩ It2 (Y ), we
partition the sequences based on the number n1 +k of elements
from Fq \ B that the sequence contains. Given some value of
k, there are Ik (n1 , q − qB ) possibilities for the sequences’s
subsequence of elements from Fq \ B.
A sequence in It1 (X) ∩ It2 (Y ) with k additional elements
from Fq \ B must have N − n1 − k elements from B. Given k,
there are IN −k−n1 −n2 (n2 , qB ) possibilities for the sequence’s
subsequence of elements from B.
Given the sequence’s length n1 + k subsequence TX of
elements from Fq \ B and the sequence’s length N − n1 − k
subsequence TY of elements from B, the number of possible
sequences containing both of these subsequences is the number
of ways TY can be interleaved with TX . This can be counted
using a balls and bins approach.
We consider the n1 + k + 1 possible insertion positions
into TX as labeled bins, and the N − k − n1 positions
of elements from TY as unlabeled balls. There are thus
(n1 +k+1)+(N −k−n1 )−1
ways for TY to be interleaved with
N −k−n1
TX .
k can range from 0 to N − n1 − n2 so we have
|It1 (X) ∩ It2 (Y )|
=
N −n
1 −n2
X
Ik (n1 , q − qB )IN −k−n1 −n2 (n2 , qB )
k=0
N
.
t1 − k
Lemma 15. Let n1 , n2 , t1 , t2 ∈ N and N = t1 +n1 = t2 +n2 .
Consider any subset B ⊆ Fq such that 1 ≤ qB ≤ q − 1, where
qB = |B|. Then,
|It1 (0n1 ) ∩ It2 (1n2 )|
=
N −n
1 −n2
X
k=0
N
Ik (n1 , q − qB )IN −k−n1 −n2 (n2 , qB )
.
t1 − k
Proof: Using a similar approach as in Lemma 14, we will
choose a sub-alphabet for the sequence 1n2 , and a sub-alphabet
for 0n1 .
0n1 and 1n2 have no elements in common. So, the alphabet
A for 1n2 can be chosen to be any subset of Fq such that
1 ∈ A, and 0 ∈
/ A, and |A| = qB . Once the alphabet for 1n2
is chosen, the alphabet for 0n1 will be Fq \ A. Thus, we have
that
|It1 (0n1 ) ∩ It2 (1n2 )|
=
N −n
1 −n2
X
k=0
N
Ik (n1 , q − qB )IN −k−n1 −n2 (n2 , qB )
.
t1 − k
by following the proof in Lemma 14
| 7cs.IT
|
1
On Properties of the Support of Capacity-Achieving Distributions for Additive
Noise Channel Models with Input Cost Constraints
Jihad Fahs, Ibrahim Abou-Faycal
Dept. of Elec. and Comp. Eng., American University of Beirut
arXiv:1602.00878v2 [cs.IT] 15 Nov 2017
Beirut 1107 2020, Lebanon
{jjf03, Ibrahim.Abou-Faycal}@aub.edu.lb
Abstract
We study the classical problem of characterizing the channel capacity and its achieving distribution in a generic
fashion. We derive a simple relation between three parameters: the input-output function, the input cost function
and the noise probability density function, one which dictates the type of the optimal input. In Layman terms we
prove that the support of the optimal input is bounded whenever the cost grows faster than a “cut-off” rate equal to
the logarithm of the noise PDF evaluated at the input-output function. Furthermore, we prove a converse statement
that says whenever the cost grows slower than the “cut-off” rate, the optimal input has necessarily an unbounded
support. In addition, we show how the discreteness of the optimal input is guaranteed whenever the triplet satisfy
some analyticity properties. We argue that a suitable cost function to be imposed on the channel input is one that
grows similarly to the “cut-off” rate.
Our results are valid for any cost function that is super-logarithmic. They summarize a large number of previous
channel capacity results and give new ones for a wide range of communication channel models, such as Gaussian
mixtures, generalized-Gaussians and heavy-tailed noise models, that we state along with numerical computations.
Keywords: Channel capacity, memoryless channels, non-linear channels, input cost function, logarithmic
cost, heavy-tailed noise, alpha-stable, Middleton class B, Gaussian mixtures, convex optimization,
Karush-Kuhn-Tucker conditions, discrete inputs.
I. I NTRODUCTION
In communication systems design, a key engineering objective is to build systems that operate close to channel
capacity. Needless to say that this quantity, as defined by Shannon [1], [2] in his pioneering work, is the cutoff
value which delimits the achievable region for “reliable” communications. Clearly, the channel capacity and how
it can be achieved are intimately related to the channel model. Despite the well-known capacity results for discrete
memoryless channels, closed-form capacity expressions are rarely found in the literature for continuous ones. The
most well-understood –and perhaps important– continuous channel is the linear Additive White Gaussian Channel
This work was supported by AUB’s University Research Board, and the Lebanese National Council for Scientific Research (CNRS-L)
November 17, 2017
DRAFT
2
(AWGN) subjected to an average power constraint. This AWGN model was studied by Shannon and may be seen
as an instance of the generic real, deterministic and memoryless discrete-time additive noise model of the form:
Y = f (X) + N,
where Y ∈ R is the channel output, and the channel input X ∈ X ∈ R satisfies an average cost constraint of the
form E [C (|X|)] ≤ A, for some A > 0. Naturally X is assumed to be independent of the additive noise N .
In the literature, multiple instances of such channel models were investigated by making variations to the Shannon
setup in the following aspects:
•
The input-output relationship f (·): While Shannon considered a deterministic linear input-output relationship,
many studies assumed a non-deterministic relationship [3]–[6] or generally a non-linear deterministic one [7].
•
The input constraint or cost function C(·): One of the main reasons of the popularity of the second moment
constraint E X 2 –which corresponds to a cost function C(x) = x2 , is that it represents the average power of
the discrete time transmitted signal which is equal to the average power of the corresponding white continuous
process assuming that the transmitted signals are square integrable. Nevertheless, other input constraints were
studied starting with Smith [8] who considered peak power constraints and a combination of peak and average
power constraints. More recently, the capacity of Gaussian Channels with duty cycle and average power
constraints was studied in [9] .
•
The noise distribution: Though Gaussian statistics of the noise can be motivated by the Central Limit Theorem
(CLT), it also has an appealing property of being the worst case noise from an entropy perspective among
finite second moment Random Variables (RV)s. Nevertheless, Non-Gaussian average power constrained communication channels have some applications and their channel capacities were investigated under a general
setup in the work of Das [10] where the noise is assumed to have a finite second moment, a condition that
was not imposed on the non-Gaussian noise distributions in [11].
•
Combinations of more than one aspect were also considered in the literature. Smith [8] extended his capacity
results for the peak power constrained Gaussian channel to non-Gaussian ones where the noise statistics
are Gaussian like. Later, Tchamkerten [12] considered a scalar additive channel whose input is amplitude
constrained and for which the additive noise is assumed to satisfy some general properties however not
necessarily having a finite second moment. Lately, Fahs and Abou-Faycal [7], [13] investigated non-linear
Gaussian channels under a general setup of input constraints such as even moments, compact support constraints
and a combination of both types. Finally, channel capacity under Fractional Order Moments (FOM) of the form
E [|X|r ] ≤ A, for some A > 0, r > 0 was characterized under a symmetric alpha-stable additive noise [14]
or when the noise has two components, an alpha-stable component and a Gaussian one [15].
Nearly, for all the cited models above, and whenever the noise Probability Density Function (PDF) is assumed
to have an analytical extendability property, the optimal input is proven to be of a discrete nature and in most
cases with a finite number of mass points. Additionally, channel capacity could not be written in closed-form. In
this sense, the linear AWGN channel and some “equivalent” channels [7] seem to be an exception, along with
November 17, 2017
DRAFT
3
a few channel models such as the additive exponential noise channel under a mean constraint with non-negative
inputs [16] and recently the Cauchy channel under a logarithmic constraint [17]. For these channel models, the
optimal input distribution is found to be of the same nature of the noise and capacity is described in closed-form.
One is tempted to study whether there is a general relation between the input-output function f (·), the input
cost function C(·), and the noise PDF pN (·) that governs the type of the capacity-achieving input. In this work,
we conduct this study for general types of the considered channel whereby “general” we mean the input-output
relationship may not be linear but required nevertheless to satisfy some rather mild conditions. Additionally, instead
of formulating the problem in terms of the average power constraint or other moments constraints, we use generic
input cost functions that are also required to satisfy some technical conditions. We emphasize that our results
cover all cost functions which are “super-logarithmic” which is a rather very large set. When it comes to the noise
statistics, the noise is assumed to be absolutely continuous with respect to the Lebesgue measure with positive
and continuous PDFs that are with or without monotonic tails and have a finite logarithmic-type of moments. Two
conditions are however imposed on the noise PDF and are subsequently presented. The first guarantees the finiteness
of the noise differential entropy. The second concerns the tail behavior of a lower envelope to the noise PDF. These
two conditions are “easily satisfied” such as whenever the PDF has a dominant exponential or a dominant polynomial
component. Despite the apparent long list of requirements, we emphasize that the considered functions f (·), input
costs C(·) and noise PDFs cover the vast majority of the known models found in the literature.
Our main contributions are fourfold:
1- Our study provides new capacity results for a multitude of communication channels. We showcase some of
them:
– Gaussian mixtures and generalized Gaussian noise distributions are commonly used in the literature [18]–
[21], however no previous channel capacity studies were conducted for these types of noise models. The
application of our results to such channels is presented in Sections VI-B and VI-C respectively.
– Many communication channels are suitably modeled as impulsive channels where the statistics of the
noise have, for example, an alpha-stable distribution or a composite alpha-stable plus Gaussian such as
telephone noise [22], audio noise signals [23] and Multiple Access Interference (MAI) [24]–[26]. More
recently [27]–[30], the performance of new receivers, mitigation and diversity techniques were investigated
when such impulsive statistics were used as models of additive noise in MAI networks. In Section VI-D.1,
we characterize and compute for a generic cost constraint C(·) the channel capacity for the alpha-stable
and the composite noise channels: alpha-stable plus Gaussian, which is commonly referred to as the
Middleton class B when the stable component is symmetric [31], [32].
2- The results stated in Theorems 1 and 2 generalize those of Das [10], who made similar statements for a linear
channel whenever C(x) = x2 and whenever the noise is restricted, among other things, to have a finite second
moment. The generalization is one to generic possibly non-linear channels, generic cost functions and noise
distributions. They are in line with all the previous channel-capacity results presented earlier. In addition, when
November 17, 2017
DRAFT
4
having an analyticity property of the noise PDF, they recover literally most of the known discreteness results
for cost constrained deterministic channels. These results are stated in Theorem 3.
3- Our methodology also provides capacity results even when the input is subjected only to a support (such
as a peak power constraint), or to a combination of support and cost constraints. In fact, the results stated
in Theorem 4 corroborate those found in [12] for channels whose input is amplitude constrained and where
the noise has a finite r-th moment constraint for some r > 0. However, Theorem 4 is in some sense more
general as it holds whenever the input has a compact support for all noise distributions that have a finite
“super-logarithmic” moment.
4- When applied to monotonically-tailed noise PDFs for example, our main results - stated in Theorems 1
h
i
and 2, imply that whenever C(x) = ω ln pN (f1(x)) ∗ , the support† of the capacity-achieving input is
necessarily bounded. In addition, we state and prove a converse statement that states that whenever C(x) =
i
h
o ln pN (f1(x)) , the optimal input is necessarily unbounded. These results state that –for monotonically
noise PDFs, there exists a threshold growth rate for the cost function which constitutes the transition between
bounded and unbounded optimal inputs. Indeed, for an optimal input to be unbounded, a “necessary condition”
i
h
for the cost function is to be at most Θ ln pN (f1(x)) . This condition is satisfied by the Gaussian channel
under an average power constraint, the exponential channel under a mean moment constraint and the Cauchy
channel under a logarithmic constraint for which a Gaussian, exponentially tailed and a Cauchy input are
respectively optimal [7], [16], [17].
On a related note, we argue that this study does provide insights on what is a suitable measure of signal strength.
Though this question is not crucial when the additive noise has a finite second moment due to the natural power
measure provided by the second moment, it seems of great importance when dealing with heavy tailed noise
distributions having infinite second moments. For these types of channels, since the noise has an infinite second
moment, using the second moment as a measure of signal strength is absurd for evaluating the Signal-to-Noise
Ratio (SNR) for example. In fact, for heavy-tailed noise models, and more specifically for the alpha-stable class,
a general theory of stable signal processing based on Fractional Lower Order Moments (FLOM) (C (|x|) = |x|r ,
0 < r < 2) was presented in [33]; The “stable theory” was in accordance with the fact that second order methods
and linear estimation theory were no longer suitable for infinite variance additive noise channels and new criteria
based on the dispersion of alpha-stable RVs and FLOM were investigated. The stable theory was also used in the
treatment of various detection and estimation problems [34], [35], and the performance of optimum receivers were
investigated in [36].
However, one can argue that moments of the form E [|X|r ], r > 0 do not provide a suitable strength measure for
variables having heavy tailed distributions, simply because no single value of r can be found appropriate [37]. This
∗ In
this work, we say that f (x) = ω (g(x)) if and only if ∀ κ > 0, ∃ c > 0 such that |f (x)| ≥ κ|g(x)|, ∀|x| ≥ c. Equivalently, we say that
g(x) = o (f (x)). We say that f (x) = Ω (g(x)) if and only if ∃ κ > 0, c > 0 such that |f (x)| ≥ κ|g(x)|, ∀|x| ≥ c. Equivalently, we say that
g(x) = O (f (x)). We say that f (x) = Θ (g(x)) if and only if f (x) = O (g(x)) and f (x) = Ω (g(x)).
† We
define the support of a RV as being the set of its points of increase i.e. {x ∈ R : Pr(x − η < X < x + η) > 0 for all η > 0}.
November 17, 2017
DRAFT
5
conclusion is further supported when making the following reasoning on the additive noise channels considered in
this work:
•
Let E [C0 (|X|)] be a measure of the average signal strength where C0 (|x|) is some positive, lower semicontinuous, non-decreasing function of |x| and let pN (x) be the noise PDF which is assumed to have a
monotonic tail.
•
•
h
i
Whenever C0 (|x|) = ω ln pN (f1(x)) , one can always find a cost function C(|x|) such that C(|x|) is both
i
h
and o (C0 (|x|)).
ω ln pN (f1(x))
h
i
Now, since C(|x|) = ω ln pN (f1 (x)) , the channel capacity under an input constraint of the form E [C(|X|)] ≤
A, A > 0 is achieved by a bounded input by virtue of Theorem 1. On the other hand, since C(|x|) = o (C0 (|x|)),
then there exists a distribution function satisfying the cost constraint with “signal strength” E [C0 (|X|)] equal
to ∞.
•
Hence, in the input space of distribution functions, there exist distributions having possibly infinite strength
while the capacity is achieved by a distribution which has a finite one since its support is bounded.
h
i
1
• This non-intuitive conclusion is only possible under the choice of a strength measure that is ω ln p (f (x))
.
N
h
i
By this reasoning, suitable signal strength measures should be at most Θ ln pN (f1(x)) . Said differently,
h
i
depending on the noise, measures of the form Θ ln pN (f1 (x)) are more appropriate. This boils down to Θ(f 2 (x))
under the Gaussian noise and to Θ (ln [f (x)]) for polynomially-tailed additive noise. The latter condition comes
in accordance with the work of Gonzalez et al. [37] who presented a new approach for dealing with heavy-tailed
noise environments. After presenting the shortcomings of the FLOM approach, they presented a “general” unit of
strength-measure based on logarithmic moments where they motivated its usage within the framework of estimation
and filtering under impulsive noise.
The remainder of this paper is organized as follows. Section II presents the generic channel model along
with all the assumptions made in our study. Preliminary lemmas concerning lower and upper bounds on some
quantities of interest are stated and proven in Section III. In Section IV, we discuss the Karush-Kuhn-Tucker
(KKT) theorem and in Section V the main results of this paper are presented as Theorems 1, 2, 3 and 4. Examples
and numerical evaluations of channel capacity and its achieving input distributions are presented in Section VI
where the application of the four theorems is explored for Gaussian mixtures, Generalized Gaussians and impulsive
noise. Finally, Section VII concludes the paper.
II. A
GENERIC CHANNEL MODEL
We consider a generic memoryless real discrete-time noisy communication channel where the noise is additive
and where the input and output are possibly non-linearly related as follows:
Yi = f (Xi ) + Ni ,
(1)
where i is the time index. We denote by Yi ∈ R the channel output at time i. The input at time i is denoted Xi and
is assumed to have an alphabet X ⊆ R. The channel’s input is distorted according to the deterministic and possibly
November 17, 2017
DRAFT
6
non-linear function f (x). Additionally, the communication channel is subjected to an additive noise process that is
independent of the input. The variables {Ni }i are also assumed to be Independent and Identically Distributed (IID)
RVs.
Finally, we subject the input to an average cost constraint of the form: E [C (|Xi |)] ≤ A, for some A ∈ R+∗
where C(·) is some cost function:
C : R+ −→ R.
Accordingly, we define for A > 0
Z
n
o
PA = Probability distributions F of X : C (|x|) dF (x) ≤ A ,
(2)
the set of all distribution functions satisfying the average cost constraint.
Given that the channel model is stationary and memoryless, the capacity-achieving statistics of Xi are also
memoryless (IID), therefore we suppress the time index and write
Y = f (X) + N,
(3)
where the noise is absolutely continuous with respect to the Lebesgue measure and is assumed to have a PDF
pN (·). This implies that the channel transition probability density function is given by
pY |X (y|x) = pN (y − f (x)),
y ∈ R, x ∈ X .
(4)
We characterize the tail behavior of pN (·) by considering the following positive functions which are non-increasing
for x ≥ 0 and non-decreasing for x < 0:
inf pN (t)
0≤t≤x
Tl (x) =
inf pN (t)
x≤t≤0
x≥0
x < 0,
Tu (x) =
sup pN (t)
t≥x
sup pN (t)
x≥0
x < 0.
t≤x
Considering the tail behavior of Tl (x) and Tu (x) instead of pN (x) allows us to include in our analysis PDFs
which do not possess a monotonic tail. For those that do, pN (x), Tl (x) and Tu (x) will be identical for large values
of |x|.
The main results of this work are based on relating the tail behavior of C(·) to that of Tl (·) and Tu (·) in order
to characterize the capacity-achieving input distributions of channel (3). More explicitly we prove that, whenever
h
i
C (|x|) = ω ln Tl (f1(x)) , the optimal input has necessarily a bounded support. Furthermore, we prove a converse
1
, the capacity-achieving input is not bounded.
statement: whenever C (|x|) = o ln
Tu (f (x))
h
i
When the noise PDF has a monotonic tail, our results infer that cost functions which are Θ ln pN (f1(x))
form somehow a “transition” between bounded and unbounded optimal inputs. For example, whenever the noise is
Gaussian, the “transitional” cost is of the form Θ f 2 (x) . The discreteness –and hence the finiteness of the number
of mass points of the optimal input in the bounded case– is a direct consequence of the analyticity properties of
pN (·) and C(·) whenever these properties exist.
November 17, 2017
DRAFT
7
A. Assumptions
In this work, we make the following assumptions:
• The function f (·):
C1- The function is continuous.
C2- The absolute value of the function |f (·)| is a non-decreasing function of |x| and |f (x)| → +∞ as
|x| → +∞.
• The cost function C(·):
C3- The cost function is lower semi-continuous and non-decreasing. Without Loss of Generality (WLOG) we
assume that C(0) = 0: if it were not, define C0 (|x|) = C(|x|) − C(0) and adjust the input space under the
cost C0 (|x|) to PA−C(0) . Note that necessarily A − C(0) ≥ 0.
C4- C (|x|) = ω (ln |f (x)|).
• The noise PDF pN (·):
C5- The PDF is positive and continuous on R. Note that this automatically implies that pN (·) is upper bounded.
C6- There exits a non-decreasing function
CN : R+ −→ R
such that CN (|x|) = ω (ln |x|), and
EN [CN (|N |)] = LN < ∞.
This necessarily implies that EN [ln (1 + |N |)] < ∞. Note that, for example, the above condition holds
true for any noise PDF whose tail is faster than
1
.
x(ln x)3
Since from an information theoretic perspective, the general channel model (1) is invariant with respect to
output scaling, we consider WLOG that the noise PDF is less than “1” for technical reasons. Furthermore, the
boundedness of pN (·) along with the fact that it has a finite logarithmic moment insure that its differential
entropy exists and is finite h(N ) < ∞ (see [38, Proposition 1]).
Restrictions C1 to C6 are “technical” in the sense that they represent sufficient conditions for the existence of
a solution to the capacity problem as defined in [8] and enables the formulation of the Karush Kuhn Tucker
(KKT) conditions as being necessary and sufficient for optimality of the input probability distribution.
• The lower and upper bounds Tl (·) and Tu (·):
Note that by definition, 0 < Tl (x) ≤ pN (x) ≤ Tu (x) ≤ 1 for all x ∈ R. We assume that Tl (·) and Tu (·)
satisfy the following properties:
i
h
1
which is positive, non-decreasing for x ≥ 0 and non-increasing in x < 0,
C7- The function L(x) = ln Tl (x)
satisfies the following inequality:
L(x + y) ≤ κl (L(x) + L(y)) ,
(5)
for some positive constant κl , whenever |x|, |y| are sufficiently large.
November 17, 2017
DRAFT
8
We note that functions that satisfy condition C7 define a convex set. In fact, let f (x), g(x) be two positive,
non-decreasing functions on R+ non-increasing on R−∗ . Let α ∈ [0, 1] and define h = αf + (1 − α)g. The
function h(x) is positive, having the same monotonic properties. Then, whenever there exists κf and κg > 0
for which f and g satisfy condition C7, we have
h(x + y) = αf (x + y) + (1 − α)g(x + y) ≤ κh (h(x) + h(y)),
where κh = max{κf ; κg } > 0.
We clarify that condition C7 is for example satisfied by all noise distribution functions where Tl (x) is any
linear combinations of:
Tl (x) = Θ s(x)er(x)
where
r(x) = |x|a log . . . log(|x|),
|
{z
}
β times
′
Tl (x) = Θ
s(x)
r(x)
,
′
s(x) = |x|a log ... log(|x|),
|
{z
}
β ′ times
and where the parameters a, a′ ∈ R+ , and β, β ∈ N, chosen so that Tl (x) is positive, its total integral is no
greater than one, and conserves its monotonic behavior‡. The fact that these two general types satisfy condition
C7 is based on the following basic identities [39]:
• For all x, y and r ∈ R,
|x + y|r ≤ max{1; 2r−1 } (|x|r + |y|r ) .
• For any x0 > 0, there exist y0 > 0 such that
|x| + |y| ≤ |xy|p , for some p > 1 whenever |x| > x0 , |y| > y0 .
Finally, we also assume that
Z +∞
Tu (x) ln Tl (x) dx exists and is finite.
C8- The integral −
−∞
Note that whenever the tail of pN (·) is monotone, condition C8 is not necessary and boils down to saying that
noise differential entropy is finite which is a byproduct of properties C5 and C6 of the noise PDF.
When it comes to conditions C5 through C8 –and specifically C7 and C8–, they are satisfied by a rather large
class of noise probability functions that includes most of the known probability models such as Gaussian,
generalized Gaussian, generalized t, alpha-stable distributions and all of their possible mixtures.
III. P RELIMINARIES
In this section we establish some preliminary results that are needed in subsequent sections: we derive lower and
upper bounds on the output probability and a quantity of interest presented hereafter.
‡ The
′
values β = 0 and β = 0 imply that respectively r(x) and s(x) have no logarithmic component.
November 17, 2017
DRAFT
9
We start by noting that for channel (3), the existence of a positive, continuous transition PDF such as in (4),
implies the existence for any input distribution F of an induced output probability density function pY (y) = p(y; F )
which is also continuous (hence upper-bounded) [7] and is given by:
Z
pY (y; F ) = p(y; F ) = pN (y − f (x)) dFX (x) ≤ 1.
(6)
Furthermore, equation (6) along with the fact that f (·) is continuous insures that the property that pN (·) is
bounded away from zero on compact subsets of R is conserved as well for pY (y; F ). This in turns implies that
pY (·) is also positive on R.
A. Bounds on p(y; F )
In what follows, we derive upper and lower bounds on the output probability distribution induced by an input
distribution F .
Lemma 1. Let y0 > 0 be sufficiently large. For an input distribution F , the PDF p(y; F ) of the output of channel (3)
is lower bounded by
Tl (y − y0 )
2
p(y; F ) ≥
Tl (y + y0 )
2
y ≤ −y0
y ≥ y0 ,
Proof: Given an input probability distribution F , we define the following:
- We denote by dF a positive constant such that Pr(|X| ≤ dF ) ≥ 21 .
- We denote by fmax = sup|x|≤dF |f (x)|, the existence of which is guaranteed by the assumption that f (·) is
continuous on R.
Let y0 > fmax . In what follows, we only present in detail the case y ≥ y0 as the proof in the other range follows
similar steps.
pY (y; F ) ≥
≥
Z
x:|x|≤dF
Z
x:|x|≤dF
pN y − f (x) dF (x)
Tl (y − f (x)) dF (x)
1
1
≥ Tl (y + fmax ) ≥ Tl (y + y0 ) ,
2
2
(7)
(8)
where equation (7) is due to the fact that Tl (·) is a lower bound on pN (·) by definition and inequalities (8) are
justified since Tl (·) is non-increasing on the considered interval.
We also derive an upper bound on the output law whenever the input is bounded within [−B, B] for some B > 0:
Lemma 2. For an input distribution F that has a bounded support within [−B, B] for some B > 0, the PDF
p(y; F ) of the output of channel (3) is upper bounded by
Tu y + y B
0
p(y; F ) ≤
T y − yB
u
November 17, 2017
0
y ≤ −y0B
y ≥ y0B ,
DRAFT
10
for any large-enough y0B .
B
Proof: Let fmax
= sup[−B;B] |f (x)|, the existence of which is guaranteed by the fact that f (·) is continuous
B
on R. Also let y0B ≥ fmax
. For y ≥ y0B , since Tu (·) is an upper bound on pN (·), we have,
Z
p(y; F ) =
pN (y − f (x)) dF (x)
=
≤
≤
Z
B
−B
Z B
−B
pN (y − f (x)) dF (x)
(9)
Tu (y − f (x)) dF (x)
B
≤ Tu y − y0B
Tu y − fmax
(10)
where equations (10) are due to the fact that Tu (x) is non-increasing on the positive semi-axis. A similar derivation
yields the result for y ≤ −y0B .
We emphasize that this upper bound on p(y; F ) is only possible under the assumption that the support of F is
bounded (as seen in equation (9)).
B. Bounds on i(x; F )
In this section we analyze the function of interest
Z +∞
i(x; F ) = −
pN (y − x) ln pY (y; F ) dy
=−
−∞
Z +∞
(11)
pN (y) ln pY (y + x; F ) dy.
−∞
Lemma 3. For any probability distribution F ,
i(x; F ) = O ln
1
Tl (x)
.
Proof: Consider a large-enough y0 so that Lemma 1 holds, and let x be such that x > y0 . For a probability
distribution F on the input we compute,
Z
i(x; F ) = −
+∞
pN (y) ln pY (y + x; F ) dy = I1 + I2 + I3 ,
−∞
where the interval of integration is divided into three sub-intervals: (−∞, −x − y0 ), [−x − y0 , y0 ], (y0 , +∞).
We study the growth rate in x of the integral terms I1 , I2 and I3 function of the rate of decay of Tl (·).
November 17, 2017
DRAFT
11
Using Lemma 1,
I1 = −
Z
−x−y0
pN (y) ln pY (y + x; F ) dy
−∞
Z −x−y0
Z −x−y0
2
Tl (y + x − y0 )
pN (y) ln
pN (y) ln
dy =
dy
≤−
2
Tl (y + x − y0 )
−∞
−∞
Z −x−y0
1
1
1
+ ln
+ ln
dy
pN (y) ln
≤ ln 2 + κ
Tl (y)
Tl (x)
Tl (−y0 )
−∞
Z +∞
1
1
1
≤ ln 2 + κ ln
pN (y) ln
+ κ ln
+κ
dy
Tl (−y0 )
Tl (x)
Tl (y)
−∞
1
≤ 2κ ln
,
Tl (x)
(12)
(13)
for some positive κ and for x > y0 large-enough. Equation (12) is due to property C7 since both x and y0 are
large enough and so is |y|. The integral term in (13) is finite by property C8 and the last equation is valid since
h
i
1
ln Tl (x)
, which is positive, is increasing to +∞.
Similarly,
I3 = −
Z
∞
pN (y) ln pY (y + x; F ) dy
y0
∞
Z ∞
2
Tl (y + x + y0 )
pN (y) ln
dy =
dy
pN (y) ln
≤−
2
Tl (y + x + y0 )
y0
y0
Z ∞
1
1
1
pN (y) ln
≤ ln 2 + κ
+ ln
+ ln
dy
Tl (y)
Tl (x)
Tl (y0 )
y0
1
≤ 2κ ln
,
Tl (x)
Z
As for I2 ,
I2
= −
= −
≤
≤
≤
≤
Z
y0
pN (y) ln pY (y + x; F ) dy
−x−y0
Z −x+y0
−x−y0
pN (y) ln pY (y + x; F ) dy −
Z
y0
pN (y) ln pY (y + x; F ) dy
−x+y0
Z y0
1
2
sup ln
pN (y) ln
+
dy
pY (y; F )
Tl (y + x + y0 )
|y|≤y0
−x+y0
1
1
sup ln
+ ln 2 + ln
pY (y; F )
Tl (x + 2y0 )
|y|≤y0
1
1
1
+ ln 2 + κ ln
+ κ ln
sup ln
pY (y; F )
Tl (x)
Tl (2y0 )
|y|≤y0
1
2 κ ln
.
Tl (x)
(14)
(15)
The supremum is finite since it is taken over a compact set where pY (y) (which is less than one) is positive,
h
i
continuous and hence positively lower bounded. Equation (14) is due to the fact that ln Tl1(·) is non-decreasing on
the positive axis, equation (15) is given by property C7 since both x and y0 are large enough and the last equation
i
h
1
is increasing to +∞ as |x| → +∞.
is justified since ln Tl (x)
November 17, 2017
DRAFT
12
A similar procedure can be adopted to prove this result when x → −∞ by adjusting the intervals of integration
to the following: (−∞, −y0 ), [−y0 , −x + y0 ], (−x + y0 , +∞) where x < −y0 such that |x| is large enough. This
h
i
1
would imply that for any probability distribution F , i(x; F ) = I1 + I2 + I3 = O ln Tl (x)
.
We also derive a lower bound whenever the input is bounded within [−B, B] for some B > 0:
Lemma 4. For an input distribution F that has a bounded support within [−B, B] for some B > 0,
1
i(x; F ) = Ω ln
.
Tu (x)
Proof: We proceed in a manner akin to the proof of Lemma 3: For an input distribution F that has a bounded
support within [−B, B] for some B > 0, we consider a large-enough y0B so that Lemma 2 holds, and let x be such
that x > y0B .
i(x; F )
= −
≥ −
≥
≥
Z
∞
pN (y) ln p(y + x; F ) dy
−∞
Z +∞
pN (y) ln p(y + x; F ) dy
y0B
"
#
1
dy
pN (y) ln
Tu y + x − y0B
y0B
1
B
> 0.
1 − FN (y0 ) ln
Tu (x)
Z
+∞
In order to write equation (16) we use the upper bound in Lemma 2. Equation (17) is justified since ln
(16)
(17)
h
1
Tu (·)
i
is
non-decreasing on the non-negative semi-axis and the end result is positive since the support of N is R. A similar
analysis may be conducted for the case when x < −y0B < 0.
IV. T HE K ARUSH -K UHN -T UCKER (KKT)
THEOREM
The capacity of channel (1) is the supremum of the mutual information I(·) between the input X and output Y
over all input probability distributions F that meet the constraint PA :
ZZ
pN (y − f (x))
C = sup I(F ) = sup
pN (y − f (x)) ln
dy dF (x).
p(y; F )
F ∈PA
F ∈PA
(18)
Conditions C1 to C6 guarantee that this optimization problem is well-defined and that its solution –the capacity– is
finite and is achievable [40, Theorem 2]. Indeed, the conditions are sufficient for PA to be convex and compact [40,
Theorem 3] and for I(·) to be concave and continuous (in the weak sense [41, Sec.III.7]) [40, Theorems 4,5].
When dealing with constrained optimization problems, the Lagrangian theorem [42] is a useful tool as it transforms
the problem to an unconstrained one when some convexity conditions are satisfied by the objective function and the
constraints. In our problem these conditions are satisfied as the mutual information is concave and the cost is linear
- and hence convex. The theorem states that there exists a non-negative parameter νA such that the optimization
November 17, 2017
DRAFT
13
problem (18) can be written as:
C = sup I(F ) = sup I(F ) − νA EF [C (|X|)] − A
F ∈PA
(19)
F
= I(F ∗ ) − νA EF ∗ [C (|X|)] + νA A,
where the last equality is true since the solution is finite and achievable by an optimal F ∗ . Furthermore,
νA (EF ∗ [C (|X|)] − A) = 0.
For every positive A, denote by C(A) the capacity of the channel under the constraint F ∈ PA , and consider the
function C(A) for A > 0. The significance of the Lagrange parameter νA is addressed in the following Lemma.
Lemma 5. Whenever for some positive A the parameter ν(A) = 0, then C(A′ ) = C(A) for all A′ ≥ A.
Proof: We start by noting that the channel capacity C(A) is a non-decreasing function of A, due to the fact
PA ⊆ PA′ , for 0 < A ≤ A′ . Now assume that ν(A) = 0 for some A > 0. For this value of A, equation (19)
becomes
C = sup I(F ) = sup I(F ) − νA EF [C (|X|)] − A = sup I(F ),
F ∈PA
F
F
which is a maximal value over all probability distributions irrespective of the constraint. This observation along
with the fact that C(A) is non-decreasing establish the result.
In our setup, a value of ν(A) = 0 can be ruled out. Said differently, the cost constraint in equation (18) is
binding. The argument we make is similar to the one used in [3]: we consider a family of input signals composed
N −2
of N discrete levels with equal probabilities at locations {1, L, L2, · · · , L2
}. When L increases, the probability
of error of a minimum probability of error receiver goes to zero, which implies by Fano’s inequality that the mutual
information approaches ln(N ). Therefore, as A → ∞, the achievable rates in our setup are arbitrarily large and
C(A) increases to infinity; a fact that is not possible if ν(A) were equal to zero for some A by Lemma 5. This
conclusion is corroborated by the fact that the capacity for general memoryless continuous-input, continuous-output
channels is achieved by a boundary input for unbounded input cost functions [43].
Whenever weak (Gateaux) differentiability is guaranteed, one can further write necessary and sufficient conditions
on the maximum achieving distribution; conditions that are commonly referred to as the KKT conditions [42]. More
formal statements on the theory of convex optimization are summarized in Appendix C in [7]. The KKT approach
was used previously in many studies [1], [3], [4], [6], [8], [9], [11], [12], [44]–[47] in order to solve the capacity
problem and for the purpose of this work, we follow similar steps. We indeed prove in Appendix I the weak
differentiability of I(·) at any optimal input F ∗ and proceeding as in [3], we write the KKT conditions as being
necessary and sufficient conditions for the optimal input to satisfy. These conditions state that an input RV X ∗
with probability distribution F ∗ achieves the capacity C of an average cost constrained channel if and only if there
exists ν ≥ 0 such that,
ν(C (|x|) − A) + C + H +
November 17, 2017
Z
pN (y − f (x)) ln p(y; F ∗ ) dy = ν(C (|x|) − A) + C + H − i(f (x); F ∗ ) ≥ 0, (20)
DRAFT
14
for all x in R, with equality if x is a point of increase of F ∗ , and where H is the entropy of the noise.
V. M AIN RESULTS
h
i
Theorem 1. Whenever C (|x|) = ω ln Tl (f1(x)) , the support of the capacity-achieving input of channel (3) is
compact.
Proof: We consider the necessary and sufficient conditions of optimality (20), and we study the behavior of
the expression function of the variable x as its magnitude goes to infinity.
These conditions state that for the optimal input X ∗ , condition (20) is satisfied with equality for any point of
increase x0 of the capacity-achieving distribution F ∗ ∈ PA . For such an x0 we obtain,
ν(C (|x0 |) − A) + C + H = i(f (x0 ); F ∗ ).
If these points of increase of X ∗ take arbitrarily large values, |f (xo )| → +∞ since |xo | → +∞. Using Lemma 3,
h
i
i(f (xo ); F ) = O ln Tl (f 1(xo )) , and therefore
1
,
ν(C (|x0 |) − A) + C + H = O ln
Tl (f (x0 ))
h
i
which is a contradiction whenever C (|x|) = ω ln Tl (f1(x))
unless ν = 0. This has been ruled out in Section IV,
which implies that the support of X ∗ is bounded. Finally, we note that the support is always closed, as its complement
is open. Therefore, X ∗ is compactly supported.
A. A converse theorem
Now we make use of the upper bound on the noise PDF. Inthissection, we
state and prove a converse formulation
1
of Theorem 1. Indeed we prove that whenever C (|x|) = o ln
, the capacity-achieving input is not
Tu (f (x))
bounded.
1
, the support of the capacity-achieving input of channel (3)
Theorem 2. Whenever C (|x|) = o ln
Tu (f (x))
is unbounded.
Proof: Suppose that the optimal input X ∗ with distribution function F ∗ has a bounded support within [−B, B]
for some B > 0. The KKT conditions imply that there exists ν ≥ 0 such that,
Z
ν(C (|x|) − A) + C + H + pN (y − f (x)) ln p(y; F ∗ ) dy ≥ 0,
∗
∗
forall x in R, with
equality if x is a point of increase of F . Using Lemma 3, the integral term i(f (x); F ) =
1
Ω ln
and hence, equation (20) necessarily implies that,
Tu (f (x))
1
ν(C (|x|) − A) + C + H = Ω ln
,
Tu (f (x))
1
.
which is impossible whenever C (|x|) = o ln
Tu (f (x))
November 17, 2017
DRAFT
15
B. Discreteness
In what follows, we further characterize the capacity-achieving input statistics when the cost function, the noise
PDF and the channel distortion function have an additional analyticity property. This property guarantees the type
of the optimal bounded input to be a discrete one, and hence with a finite number of mass points by virtue of
compactness. This characterization permits to proceed to numerical computations in order to compute channel
capacity and find the achieving input.
In this section, let η > 0 denote a positive scalar and let Sη = {z ∈ C : |ℑ(z)| < η} be a horizontal strip in the
complex domain. We adopt in this section an alternative definition of Tu (x):
sup
|pN (ζ)|
x≥0
ζ∈Sη :ℜ(ζ)≥x
Tu (x) =
sup
|pN (ζ)|
x < 0,
(21)
ζ∈Sη :ℜ(ζ)≤x
and we assume that the following condition holds: The integral −
that this condition is similar to C8 but it is function of a
Z
+∞
Tu (x) ln Tl (x) dx exists and is finite. Note
−∞
redefined Tu ().
One may think of the condition as more
restrictive. However, this strengthened condition is needed only to establish discreteness. In the remainder of this
document we will refer to this condition as “the strengthened-C8”. We present hereafter, a lemma that guarantees
the analyticity of i(·; F ) on Sη :
Lemma 6. Whenever there exists an η > 0 such that pN (·) admits an analytic extension on Sη , the function
i(·; F ) : Sη → C defined by:
z → i(z; F ) = −
Z
∞
−∞
pN (y − z) ln p(y; F ) dy,
(22)
is analytic.
Proof: To prove this lemma, we will make use of Morera’s theorem:
a) We start first by proving the continuity of i(·; F ). In fact, let ρ > 0, z0 and z ∈ Sη such that |z − z0 | ≤ ρ,
Z
lim i(z; F ) = − lim
pN (y − z) ln p(y; F ) dy
z→z0
z→z0
Z
(23)
=−
lim pN (y − z) ln p(y; F ) dy
z→z0
Z
= − pN (y − z0 ) ln p(y; F ) dy = i(z0 ; F ).
(24)
Equation (24) is justified by pN (y − z) being a continuous function of z on Sη by virtue of its analyticity
and equation (23) by Lebesgue’s Dominated Convergence Theorem (DCT). Indeed, in what follows we find an
integrable function r(y) such that,
pN (y − z) ln p(y; F ) = − pN (y − z) ln p(y; F ) ≤ r(y),
for all z ∈ Sη such that |z − z0 | ≤ ρ and for all y ∈ R. We upper bound first |pN (y − z)|: let y0 be large enough
so that Lemma 1 holds
November 17, 2017
DRAFT
16
• If y ≤ −(y0 + |ℜ(z0 )| + ρ), then y ≤ −y0 + ℜ(z0 ) − ρ (where y0 has been defined in Lemma 1) and
|pN (y − z)| ≤ Tu (y − ℜ(z)) ≤
max
ζ∈Sη :|ζ−z0 |≤ρ
Tu (y − ℜ(ζ)) = Tu (y − ℜ(z0 ) + ρ) ,
where the last equality is due to the fact that for x ≤ 0, Tu (x) is non-decreasing, and for ζ ∈ Sη ; |ζ − z0 | ≤ ρ,
(y − ℜ(ζ)) ≤ (y − ℜ(z0 ) + ρ) < 0.
• Similarly, for y ≥ (y0 + |ℜ(z0 )| + ρ) ≥ (y0 + ℜ(z0 ) + ρ),
pN (y − z) ≤ Tu (y − ℜ(z0 ) − ρ) .
Next, using Lemma 1 we also upper bound − ln p(y; F ) to obtain:
2
y ≤ −(y0 + |ℜ(z0 )| + ρ)
Tu (y − ℜ(z0 ) + ρ) ln
Tl (y − y0 )
′
r(y) =
−M ln M
|y| < y0 + |ℜ(z0 )| + ρ
2
y ≥ y0 + |ℜ(z0 )| + ρ,
Tu (y − ℜ(z0 ) − ρ) ln
Tl (y + y0 )
where
M=
max
max
{|y|≤(y0 +|ℜ(z0 )|+ρ)} {ζ∈Sη :|ζ−z0 |≤ρ}
|pN (y − ζ)|
&
′
M =
min
{|y|≤(y0 +|ℜ(z0 )|+ρ)}
pY (y; F ).
′
Note that M is finite since pN (·) is analytic and the maximization is taken over a compact set, and 0 < M < 1,
since pY (·; F ) is positive, continuous and less than 1. Properties C7 and strengthened-C8 insure the integrability
of r(y) which concludes the proof of continuity of i(z; F ).
b) To continue the proof of analyticity, we need to integrate i(·; F ) on the boundary ∂∆ of a compact triangle
∆ ⊂ Sη . We denote by |∆| its perimeter, η0 = minz∈∂∆ ℜ(z), η1 = maxz∈∂∆ ℜ(z) and φ = y0 + max{|η0 |, |η1 |}.
By similar arguments as above, we have
Z Z
|pN (y − z)|| ln p(y; F )|dz dy
Z
Z
′′
ln p(y; F ) dy + |∆| Tu (y − η0 ) ln
≤ |∆| M
R
∂∆
|y|≤φ
y≤−φ
Z
2
2
dy + |∆| Tu (y − η1 ) ln
dy < ∞,
Tl (y − y0 )
Tl (y − y0 )
y≥φ
where
′′
M = max max |pN (y − ξ)| < ∞.
y:|y|≤φ ξ∈∂∆
Using Fubini’s theorem to interchange the order of integration,
Z
Z Z
Z Z
i(z; F )dz = −
pN (y − z) ln p(y; F ) dy dz = −
pN (y − z) ln p(y; F ) dz dy
∂∆
R ∂∆
Z∂∆ R
Z
= − ln p(y; F )
pN (y − z) dz dy = 0,
R
(25)
∂∆
where (25) is justified by the fact that pN (y − z) is analytic for all z ∈ Sη and y ∈ R. Equation (25) in addition
to the continuity of i(·; F ) insure its analyticity on Sη .
November 17, 2017
DRAFT
17
Theorem 3. Assume there exists an η > 0 such that pN (x) is analytically extendable on Sη , and let I be an
unbounded closed interval of R§ . The capacity-achieving input of channel (3) is compactly supported and discrete
with finite number of mass points on I, whenever the following conditions hold:
i
h
1
.
• C (|x|) = ω ln T (f (x))
l
•
The restrictions of f (x) and C (|x|) on I admit analytic extensions to I × R, denoted fI (·) and CI (·)
respectively.
•
The inverse map fI−1 (·) of fI (·) conserves connectedness.
Before we prove the theorem, we note that a necessary condition for analytical extendability is to have C (|x|)
an explicit function of the variable x on I which can be possibly realized when I is for example a subset of either
R+ or R− .
Proof: We start by setting some notation and making a few remarks:
•
Define J to be the image of interval I by fI (·).
Since by analyticity fI (·) is continuous, then J is an interval of R because fI (·) is identical to f (·) on I,
and is real valued.
•
Let Jη = {z ∈ Sη : ℜ(z) ∈ J } and define Iη = fI−1 (Jη ), the inverse image of Jη by fI (·).
Note that since J is an interval, Jη is connected and so is Iη by virtue of the properties of fI−1 (·). Additionally,
since fI (I) = J then I ⊂ Iη .
In what follows, we work using the induced topology on Iη . Under this topology, Iη is both open and closed.
We proceed with the proof and assume that the optimal input X ∗ with distribution function F ∗ has at least one
point of increase in I for otherwise the result becomes trivial. Assume that the points of increase of F ∗ in I are
accumulating, and let
s(z) = ν (CI (z) − A) + C + H − i(fI (z); F ∗ ).
By the result of Lemma 6, i(fI (z); F ∗ ) is analytic on Iη since it is the composition of two analytic functions:
fI (·) on Iη and i(·; F ∗ ) on Jη = fI (Iη ) ⊂ Sη . This implies that the function s(z) is analytic on Iη . Since
by assumption the points of increase of F ∗ have an accumulation point on I then by the KKT conditions, s(z)
has accumulating zeros on I ⊂ Iη , which necessarily implies by the identity Theorem [48, sec. 66] that s(·) is
identically null on Iη , since Iη is open and connected. Therefore,
Z
ν(C (|x|) − A) + C + H = − pN (y) ln p(y − f (x); F ∗ ) dy,
∀x ∈ I.
Since I is unbounded, this equality is impossible for large values of x by the result of Theorem 1 unless ν = 0
which is non sensible. This leads to a contradiction and rules out the assumption of having an accumulation point
on I. Since R is Lindelof, X ∗ is necessarily discrete on I. Additionally, since the support of X ∗ is compact and
I is closed in R, X ∗ has necessarily a finite number of mass points on I.
§ We
consider that R is both closed and open.
November 17, 2017
DRAFT
18
Theorem 4. Assume there exists an η > 0, such that pN (x) is analytically extendable on Sη , and let I be an
unbounded closed interval of R. Whenever the input is constrained to have a compact support X , the capacityachieving input is discrete with a finite number of mass points on X ∩ I if the following holds:
•
•
The restriction of f (x) on X ∩ I admits an analytic extension to I × R, denoted fI (·).
The inverse map fI−1 (·) of fI (·) conserves connectedness.
Before proving the theorem, we note that the condition that the support of X is compact is a generalization of
the peak power constraint. Also it makes sense to consider sets X that are not discrete, for otherwise the problem
is ill defined.
Proof: We first note that the KKT conditions are valid under the setup of the compactly-supported input
constraint: Indeed, the input space is compact in the weak topology and convex (see [7], [8]). Also note that there
exists a cost function C (|x|) the tail of which is ω (ln |f (x)|) and such that supX C (|x|) = A, for some A > 0.
R
Now, for any F ∈ PX –the set of all input distributions having a compact support X , we have C (|x|) dF (x) ≤ A
which implies that PX ⊂ PA . Since the mutual information is finite, continuous and weakly differentiable on PA
whenever C (|x|) = ω (ln |f (x)|) (see Appendix I) then it is as such on PX . Under this setup, the KKT conditions
state that an input RV X ∗ with CDF F ∗ achieves the capacity C of a compact-support constrained channel if and
only if,
C+H +
Z
pN (y − f (x)) ln p(y; F ∗ ) dy ≥ 0,
∀x ∈ X ,
with equality if x is a point of increase of F ∗ , and where H is the entropy of the noise. By virtue of the
R
analyticity conditions, the function s(z) = C + H + pN (y − fI (z)) ln p(y; F ∗ ) dy would also be analytic on I.
The assumption that the points of increase of X ∗ on X ∩I have an accumulation point is impossible since it will lead
R
by the identity theorem to s(x) = 0 on I which is impossible since i(x; F ∗ ) = − pN (y − f (x)) ln p(y; F ∗ ) =
i
h
(see the proof of Theorem 2), which increases to ∞. Therefore, X ∗ is necessarily discrete on
Ω ln Tu (f1(x))
X ∩ I. The finiteness of the number of mass points is a direct consequence of the compactness of X ∩ I.
Note. A similar statement to that of Theorem 4 may be made whenever, in addition to a compact support constraint,
i
h
or
there is also a cost constraint satisfying the conditions of Theorem 3 with tail behavior either ω ln Tl (f1(x))
i
h
1
o ln Tu (f (x)) .
Before moving to giving some concrete examples to our general theorems, we would like to state that some
conditions were only considered for either the sake of the clarity of the proofs, or for conserving the general aspect
of the results. Many such conditions could be relaxed while conserving some or all of the found conclusions. For
example,
•
The notions of ω, Ω, o and O used in this document are defined as |x| → +∞, i.e., in such a way to capture
a symmetric rate of decay for both tails. However, one can only consider left or right tail behaviors separately.
The results of boundedness and discreteness could be given in terms of each tail where for example for nonsymmetric noise PDFs or non-symmetric cost functions, the optimal input could only be bounded on one of
November 17, 2017
DRAFT
19
the semi-axis.
•
For Theorems 1 and 2, the assumption that pN (·) is positive could be relaxed to one sided noise PDFs. These
theorems are still valid on one side of the axis.
•
The proven theorems –stated in terms of Tl (x) and Tu (x)– could be stated in terms of any two functions
having the same properties and providing lower and upper bounds on pN (x) for large values of |x|.
VI. A PPLICATIONS
OF THE
T HEOREMS
AND
N UMERICAL R ESULTS
In this section we apply our results to a variety of specific channels of interest that fit under the general framework
presented previously. For those channels that have been previously studied in the literature, we verify our results
–in the form of Theorems 1, 2, 3 and 4, and for the other models we state some new results. We note that in all the
examples presented subsequently the considered functions f (·) and the cost constraints satisfy the general conditions
C1 through C4 in Section II-A. The noise distributions are absolutely continuous with positive, continuous PDFs
with tails that have “at least” a polynomial decay and hence satisfying the assumptions C5 and C6. Finally, in
all the provided examples the noise PDFs possess a monotonic tail and a finite differential entropy and therefore,
condition C8 is satisfied. It remains to check for each example condition C7 and possibly the strengthened-C8.
For the purpose of verifying condition C7, we note that one can use pN (x) instead of Tl (x) since they are
identical at large values of |x|. When it comes to discreteness, whenever |x| is large enough the function Tu (x)
defined in (21) becomes
Tu (x) =
sup
{z:
ℜ(z)=x & |ℑ(z)|<η}
|pN (z)| ,
because |pN (z)| is decreasing with |ℜ(z)| at large values for all the given examples.
For each model we consider in what follows, we will check whether the appropriate conditions are satisfied, state
the results –specialized to the channel at hand, and compare with the known results in the literature.
A. The Gaussian Model
For a Gaussian noise distribution with mean zero and variance σ 2 , the PDF is pN (x) =
write N ∼ N 0, σ 2 .
2
x
√ 1
e− 2σ2
2πσ2
and we
Checking the conditions: Condition C7 is validated as follows: for large values of |x| and |y|,
√
1
(x + y)2
= ln 2πσ 2 +
L(x + y) = ln
pN (x + y)
2σ 2
√
√
√
x2
y2
2
2
≤ 2 ln 2πσ + 2 + ln 2πσ + 2 − 3 ln 2πσ 2 = κl (L(x) + L(y)) ,
2σ
2σ
where κl > 2. When it comes to discreteness, let pN (z) =
z2
√ 1
e− 2σ2
2πσ2
, be an analytic extension of pN (x) to the
complex plane, where z = x + jy. The magnitude of pN (z) is
x2 −y2
z2
1
1
|pN (z)| = √
e− 2σ2 ,
e− 2σ2 = √
2πσ 2
2πσ 2
and is decreasing in x = ℜ(z). Therefore, Tu (x) =
November 17, 2017
√ 1
e−
2πσ2
x2 −η2
2σ2
η2
= e 2σ2 pN (x).
DRAFT
20
Checking for the strengthened-C8, the integral −
the noise differential entropy h(N ) is finite.
R +∞
−∞
η2
Tu (x) ln Tl (x) dx = e 2σ2 h(N ) which is finite because
The following theorem is a specialization of Theorems 1 and 2 for this specific Gaussian case:
Theorem 5. Whenever C (|x|) = o f (x)2 , the support of the capacity-achieving input of channel (3) when
N ∼ N 0, σ 2 is unbounded.
Whenever C (|x|) = ω f (x)2 , the support of the capacity-achieving input of channel (3) when N ∼ N 0, σ 2
is compact. Furthermore, the optimal input is discrete with finite number of mass points whenever C(·) and f (·)
satisfy the analyticity and connectedness conditions of Theorem 3.
Under a compact support constraint, the optimal input is also discrete with finite number of mass points whenever
f (·) satisfies the conditions of Theorem 4.
Previous work: A possibly non-linear (f (x) = xn , n ∈ N∗ ) Gaussian channel under an even moment constraint
(C (|x|) = x2k ) was considered in [7] as a core channel model from which results on multiple non-linear channel
models were derived. The authors applied a standard Hilbert space decomposition using Hermite polynomials as
bases and proved that, for n < 2k, the capacity-achieving distribution has the following behavior:
•
Whenever n = k
– if n is odd, the optimal input F ∗ is absolutely continuous.
– if n is even, F ∗ is discrete with no accumulation points.
•
•
Whenever n < k, F ∗ is discrete with finite number of mass points.
Whenever k < n < 2k, F ∗ is discrete with no accumulation points.
We point out that while the results stated in Theorem 5 do not cover the limiting case n = k –which corresponds
to the case C (|x|) = θ f 2 (x) , the result for the case “n < k” is identical. Whenever k < n, Theorem 5 states
that the support of F ∗ is not bounded; a conclusion that could not be reached in [7].
B. Gaussian Mixtures
Gaussian mixtures are widely used as more tractable models to some non-Gaussian noise statistics [18], [19].
One approach in dealing with such distributions is based on the observation that in limiting cases Gaussian mixtures
are nearly Gaussian and they are simplified accordingly. The PDF of a Gaussian mixture RV is:
pN (x) =
n
X
αi pNi (x),
i=1
where n ∈ N∗ and for 1 ≤ i ≤ n,
• Ni ∼ N µi , σi2 are Gaussian RVs with mean µi and variances σi2 6= 0. We assume WLOG that σ1 ≥ · · · ≥ σn .
P
• 0 ≤ αi ≤ 1, and ni=1 αi = 1.
November 17, 2017
DRAFT
21
Before proceeding, we note that the rate of decay of this noise PDF is
√ 2 (x−µ1 )2
2πσ1
2
2σ1
e
1
α1
= ln
ln
(x−µi )2 (x−µ1 )2
pN (x)
Pn
−
+
2σ2
2σ2
i σ1
1
i
1 + i=2 α
e
α1 σi
2
#
#
"
"p
n
µ
µ2
X
2πσ12
αi σ1 −x2 2σ12 − 2σ12 +x σµ2i − µσ12 − 2σi2 − 2σ12
(x − µ1 )2
1
1
1
i
i
i
+
− ln 1 +
e
= ln
α1
2σ12
α σ
i=2 1 i
2
!
#
"p
n
µ
µ2
X
(x − µ1 )2
αi σ1 −x2 2σ12 − 2σ12 +x σµ2i − σµ21 − 2σi2 − 2σ12
2πσ12
1
1
1
i
i
i
+
−Θ
e
= ln
α1
2σ12
α
σ
1
i
i=2
=
Θ(x2 ).
Checking the conditions: Since in Section II-A, we proved that condition C7 defines a convex set of functions,
then by the results of the Gaussian model, each pNi (·) satisfies condition C7 and so does pN (·). To study discreteness,
Pn
we let pN (z) = i=1 αi pNi (z), be an analytic extension of pN (x) on the complex plane. Since
|pN (z)| ≤
then, for a large-enough |x|,
n
X
i=1
αi |pNi (z)| =
n
X
i=1
η2
Tu (x) =
sup
{z:
ℜ(z)=x & |ℑ(z)|<η}
|pN (z)| ≤ e σn2
2
2
(x−µi ) −y
−
αi
2σ2
i
p
,
e
2πσi2
n
X
i=1
2
2
(x−µi )
η
−
α
2
p i e 2σi = e σn2 pN (x),
2
2πσi
which implies that strengthened C8 is valid of the finiteness of
Z +∞
−
Tu (x) ln Tl (x) dx,
−∞
as h(N ) is finite by virtue of the fact that N has a finite variance σ 2 =
Pn
i=1
αi σi2 .
Specializing the results to the channel at hand, we can state the following:
Theorem 6. Whenever C (|x|) = o f (x)2 , the support of the capacity-achieving input of channel (3) when N is
a Gaussian mixture is unbounded.
Whenever C (|x|) = ω f (x)2 , the support of the capacity-achieving input of channel (3) when N is a Gaussian
mixture is compact. Furthermore, the optimal input is discrete with finite number of mass points whenever C(·) and
f (·) satisfy the analyticity and connectedness conditions of Theorem 3.
Under a compact support constraint, the optimal input is also discrete with finite number of mass points whenever
f (·) satisfies the conditions of Theorem 4.
Previous work: To our knowledge, a formal analysis of Gaussian mixtures channels has not been conducted
before, and hence Theorem 6 states a new previously unknown result. We note that since the transitional rate
of decay is θ f 2 (x) , the capacity of the linear (f (x) = x) Gaussian mixtures channel under an average power
constraint (C (|x|) = x2 ) is not in the scope of this work. However, in [11] it was shown that, except for Gaussian
noise, the capacity of the linear average power constrained channel is achieved by discrete statistics for all noise
distributions satisfying certain conditions, ones that are indeed satisfied by Gaussian mixtures.
November 17, 2017
DRAFT
22
In Figure 1, we plot the numerically-computed capacity of a sample Gaussian mixture channel. The results of [11]
were used and an optimal discrete input distribution that satisfies the necessary and sufficient KKT condition was
sought. The numerical computations were conducted using Matlab.
0.7
0.6
Capacity in Nats
Per Channel Use
0.5
0.4
0.3
0.2
0.1
0
−12
−10
−8
−6
−4
−2
0
2
4
6
8
Input Second Moment in dB
Fig. 1. Capacity of the linear channel under the Gaussian mixture noise pN (x) = 0.5 pN1 (x)+0.5 pN2 (x) where N1 ∼ N (0, 1)
and N2 ∼ N (0, 4).
C. Generalized Gaussian
Generalized Gaussians [49] are viewed as a class of distributions generalizing the well-known Laplacian and
Gaussian distributions. Additive noise is often assumed to be a generalized Gaussian RV in order to model the
impulsive¶ nature of noise in communication channels [50]–[53]. In other instances, these models were considered
for the ultra-wide band multiple access interference plus noise [20], [21].
Generalized Gaussians have exponentially decaying PDFs given by:
pN (x) =
|x−µ| a
a
−( b ) ,
1 e
2bΓ a
(26)
where Γ(·) is the Gamma function, a ∈ R+∗ is a shape parameter, b ∈ R+∗ is a scale parameter and µ ∈ R is a
location parameter. In the remainder of this section, we will assume WLOG that the location parameter µ is equal
to zero.
¶ By
impulsive it is meant that extreme values of the noise signal are observed very frequently (i.e., with notable amount of probability).
November 17, 2017
DRAFT
23
Checking the conditions: Condition C7 is satisfied. Indeed,
"
#
2bΓ a1
1
|x + y|a
L(x + y) = ln
= ln
+
pN (x + y)
a
b
#
"
2bΓ a1
|x|a + |y|a
+ max{1; 2a−1}
≤ ln
a
b
a−1
} [L(x) + L(y)] + min{0; 1 − 2
=
max{1; 2
≤
κl [L(x) + L(y)] ,
a−1
"
2bΓ
} ln
a
1
a
#
for some κl > max{1; 2a−1} for large-enough values of |x| and |y|.
One can therefore state the following theorem:
a
Theorem 7. Whenever C (|x|) = o (|f (x)| ), the support of the capacity-achieving input of channel (3) when N is
a generalized Gaussian RV (26) is unbounded.
a
Whenever C (|x|) = ω (|f (x)| ), the support of the capacity-achieving input of channel (3) when N is a
generalized Gaussian RV (26) is compact.
Previous work: To our knowledge, no previous information theoretic work has appeared regarding this channel
model. For the linear channel under an average power constraint for instance, the optimal input of channel (3) is
bounded whenever the noise is a generalized Gaussian with parameter a < 2.
D. Polynomially-Tailed Distributions
Gaussian mixtures and generalized Gaussians are considered by many researchers to fail to capture the “impulsiveness” of the noise. This failure is due to several reasons, the most important of which is that they do not possess
the algebraic behavior of heavy-tailed noise distributions encountered in typical communication channels [54]. One
family of such distributions, the “generalized Cauchy” [49], is found to be reasonable in modeling the amplitude
of atmospheric impulse noise [55]. In this document, we refer by “polynomially-tailed” noise distributions to all
distributions satisfying
pN (x) = Θ
1
|x|1+α
,
for some α > 0,
which include among others: the Gamma, Pareto (one sided) and alpha-stable distributions.
Checking the conditions: In order to proceed, we use the “obvious” lower and upper bounds on pN (x) for large
values of |x| instead of pN (x) itself and we state the corresponding theorems accordingly. These bounds are of the
form
ζl
|x|1+α
and
November 17, 2017
ζu
|x|1+α ,
for some ζl and ζu > 0. We prove now that condition C7 is satisfied; Let
1+α
|x|
L(x) = ln
= (1 + α) ln |x| − ln ζl ,
ζl
DRAFT
24
which implies that for large-enough |x| and |y|,
L(x + y) = (1 + α) ln |x + y| − ln ζl
≤ (1 + α) ln [|x| + |y|] − ln ζl
≤ p(1 + α) [ln |x| + ln |y|] − ln ζl
= p [(1 + α) ln |x| − ln ζl + (1 + α) ln |y| − ln ζl ] + (2p − 1) ln ζl
≤ 2p [(1 + α) ln |x| − ln ζl + (1 + α) ln |y| − ln ζl ]
= 2p [L(x) + L(y)] ,
where p > 1. Consequently, the following holds:
Theorem 8. Whenever C (|x|) = ω (ln |f (x)|), the support of the capacity-achieving input of channel (3) when N
is polynomially-tailed is compact.
For example, for a linear channel subjected to an additive polynomially-tailed noise, the optimal input has a
bounded support for any cost function that is super logarithmic (i.e., ω (ln |x|)) such as the average power constraint.
Note that the other “range” C (|x|) = o (ln |f (x)|) is outside the scope of this work as condition C4 will not be
satisfied. When it comes to discreteness and strengthened-C8, it depends on the analyticity property of the specific
pN (·) under consideration.
The remaining part of this Section is dedicated to two important types of polynomially decaying distributions,
for which we prove that the discreteness results of Theorem 3 apply.
1) Non-Totally Skewed Alpha-Stable and their Mixtures: The term “stable” is used because, under some constraints, these distributions are closed under convolution. The stable distributions, which are a subset of that of
infinitely divisible distributions, are the only laws that have the captivating property of being the resultant of a limit
of normalized sums of IID RVs. This result is referred to as the Generalized Central Limit Theorem (GCLT), a
property that constitutes one of the main reasons behind the adoption of Gaussian statistics for noise models in
communication channels.
Though the Gaussian distribution is one of the stable laws, it represents the exception: it is unique in the sense
that it is the only one that has a finite variance and an exponential tail; All others have an infinite variance and
a polynomial tail. A complete literature on the theory of stable distributions can be found in [56]–[59]. In this
document we use the term “alpha-stable” to refer to stable variables other than the Gaussian. Although only few
alpha-stable RVs have closed form densities (namely the Cauchy and the Lévy laws), these distributions are well
characterized in the Fourier domain: The characteristic functionk of an alpha-stable RV is given by:
α
+∗
α
0<α<2 −1≤β ≤1 γ ∈R
δ∈R ,
φ(t) = exp iδt − γ 1 − iβ sgn(t)Φ(t) |t| ,
k The
characteristic function φ(t) of a distribution function F (x) is defined by:
Z
eitx dF (x).
φ(t) =
R
November 17, 2017
DRAFT
25
where sgn(t) is the sign of t, and the function Φ(·) is given by:
tan πα
α 6= 1
2
Φ(t) =
− 2 ln |t|
α = 1.
π
The constant α is called the “characteristic exponent”, β is the “skewness” parameter, γ is the “scale” parameter
(γ α is often called the “dispersion”) and δ is the “location” parameter. Such a RV will be denoted N ∼ S(α, β, γ, δ).
In what follows, we limit our analysis to non-totally skewed alpha-stable variables, i.e., ones for which |β| 6= 1.
Checking the conditions: For non-totally skewed laws, both the right and the left tails are polynomially decaying
1
(see [60, Th.1.12, p.14]), and Theorem 8 holds. Furthermore, whenever α ≥ 1 the alpha-stable
as Θ |x|α+1
variables are analytically extendable, to the whole complex plane when α > 1 and to some horizontal strip when
α = 1 [61, theorem 2.3.1 p. 48 and remark 1 p. 49]. We check in what follows the strengthened-C8. We derive in
Appendix II a novel bound on the rate of decay of the complex extension of the alpha-stable PDF when α ≥ 1:
For small-enough η > 0, there exist κ > 0 and n0 > 0 such that
|pN (z)| ≤
κ
,
|ℜ(z)|α+1
∀ z ∈ Sη : |ℜ(z)| ≥ n0 .
(27)
This bound insures the validity of Theorems 3 and 4 whenever the conditions on C(·) and f (·) are satisfied, and
hence the following theorem is valid:
Theorem 9. Whenever C (|x|) = ω (ln |f (x)|), the support of the capacity-achieving input of channel (3) when
N ∼ S(α, β, γ, δ) is a non-totally skewed alpha-stable variable is compact.
Whenever α ≥ 1, the optimal input is discrete with finite number of mass points whenever C(·) and f (·) satisfy
the analyticity and connectedness conditions of Theorem 3.
Under a compact support constraint, the optimal input is also discrete with finite number of mass points whenever
f (·) satisfies the conditions of Theorem 4.
Note that by virtue of the fact that condition C7 defines a convex set, the results presented here for one alpha-stable
variable are valid for any convex combinations of them.
Previous work: The capacity of the additive linear channel was considered in [14], where the noise is modeled as
r
symmetric alpha-stable (β = 0) for the range α ≥ 1. Subjected to a fractional r-th moment constraint, E [|X| ] ≤ a,
a > 0 and r > 1, the optimal input was found to be achieved by discrete statistics. Theorem 9 generalizes this result
to cover the non totally-skewed alpha-stable family and generic input cost functions that are “super-logarithmic”.
As a direct application of Theorem 9, it can be seen that the result of [14] also holds for |β| 6= 1 (not necessarily
equal to 0) and for the range r ≤ 1.
We use a specialized numerical MATLAB package [7] to search for the positions of the optimal points and their
respective probabilities whenever the optimal input is discrete. In Figure 2, we plot the capacity of channel (3)
whenever f (x) = x, C(|x|) = x2 and N ∼ S(α, 0, 1, 0) for α = 1, 1.2, 1.5 and 1.8. The capacity curves clearly
shows that as α gets bigger the capacity is higher. This is in accordance with the fact that the lower the value of
α, the distribution becomes heavier.
November 17, 2017
DRAFT
26
0.9
1
1.2
1.5
1.8
0.8
Capacity in Nats
Per Channel Use
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
−15
−10
−5
0
5
10
15
Input Second Moment in dB
Fig. 2.
Capacity of the linear channel subject to symmetric “standard” alpha-stable noise N ∼ S(α, 0, 1, 0) for various values
of the characteristic exponent α.
2) Composite noise: Gaussian + Alpha-Stable: Recently, a compound noise model was adopted to capture
potentially different sources of noise: a Gaussian model for the thermal noise and an alpha-stable model for the
potential MAI, as is the case for ad-hoc self configuring networks with applications in CDMA networks [29],
and in the general context of ultra wideband technologies [26]. Further information on the subject can be found
in [62]–[64]. This noise model is widely known as the Middleton class B model [31], [32]. We consider hence the
following additive noise N = N1 + N2 , where
•
•
N1 ∼ S(α, β, γ, δ), which represents the effect of the MAI, assumed a non totally-skewed alpha-stable RV.
N2 ∼ N µ, σ 2 is a Gaussian RV that models the effect of thermal noise.
Checking the conditions: It has been proved in [15, Appendix I] that pN (x) is polynomially-tailed which implies
that Theorem 8 holds for the compound noise model. In order to apply Theorems 3 and 4 for the channels impaired
by the composite noise N , we use the fact that its PDF is analytically extendable on C (for all values of 0 < α < 2)
and therefore on Sη [15, Appendix I], and check the strengthened-C8:
Z
(z−t)2
1
e− 2σ2 pN1 (t) dt
Tu (x) = sup |pN (z)| ≤
sup √
2πσ 2
|ℑ(z)|<η
|ℑ(z)|<η
Z
(x−t)2
η2
η2
1
e 2σ2
e− 2σ2 pN1 (t) dt = e 2σ2 pN (x),
≤ √
2
2πσ
which implies
−
Z
+∞
−∞
η2
Tu (x) ln Tl (x) dx ≤ −e σ2
Z
+∞
−∞
η2
pN (x) ln pN (x) dx = e σ2 h(N ) < ∞.
The following theorem therefore holds:
November 17, 2017
DRAFT
27
Theorem 10. Whenever C (|x|) = ω (ln |f (x)|), the support of the capacity-achieving input of channel (3) when
N = N1 + N2 is compact. The optimal input is discrete with finite number of mass points whenever C(·) and f (·)
satisfy the analyticity and connectedness conditions of Theorem 3.
Under a compact support constraint, the optimal input is also discrete with finite number of mass points whenever
f (·) satisfies the conditions of Theorem 4.
Previous work: The capacity of the additive channel subjected to the compound noise N = N1 + N2 was
characterized in [15] by the authors.
We plot in Figure 3 the capacity of the linear channel under an input second-moment constraint whenever
N1 ∼ N (0, 1) and N2 ∼ S(α, 0, 1, 0) for the values of α = 1 and 1.5 where the optimal input at 7.27dB was
found to have 16 and 18 mass points respectively.
We note that the composite noise channel cannot be approximated by a Gaussian channel because the overall noise
will be heavy tailed whenever the stable noise is present. Indeed, the composite noise here has infinite variance.
If one where to ignore the presence of a “mild” stable noisecomponentsuch as N ∼ S(1, 0, 0.1, 0) and assume
E [X 2 ]
= 0.356 nats at 0.16 dB. This is to be
the additive noise to have only a Gaussian component, 12 ln 1 + σ2
compared with the capacity of the composite channel which is only 0.298 nats/channel-use.
0.45
0.4
1
1.5
Capacity in Nats
Per Channel Use
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
−10
−8
−6
−4
−2
0
2
4
6
8
Input Second Moment in dB
Fig. 3. Capacity of the linear channel under the composite noise: a standard Gaussian & a standard alpha-stable for α = 1 & 1.5.
VII. C ONCLUSION
We studied the problem of characterizing the capacity and its achieving distributions for additive noise channels of
the form Y = f (X)+N , where the input is subjected to an input cost constraint of the form E [C (|X|)] ≤ A, A > 0.
November 17, 2017
DRAFT
28
We proved that the type of the optimal input is intimately related to the growth rate at infinity of the functions
f (x), C (|x|) and
1
pN (x)
through a simple relationship. Indeed, for monotonically tailed noise density functions
h
i
whenever C (|x|) = ω ln pN (f1(x)) , the support of the optimal input is necessarily bounded. Conversely, if
i
h
C (|x|) = o ln pN (f1(x)) , the support is unbounded. Similar statements are true for non-monotonically tailed
PDFs with replacing pN (·) by well chosen lower and upper envelopes whose tails are monotone. Furthermore,
whenever some analyticity properties are satisfied by the triplet, the discrete nature of the optimal distribution is
guaranteed. Discreteness holds also if additional input compact support constraints are imposed.
These results are very broad; They are consistent with a multitude of previously know capacity results, and
provide solutions for a multitude of new channel models. The generalization is one to many: generic input-output
functions, generic cost functions and generic noise PDFs which include a large number of well-known noise models
such as the Gaussian, generalized-Gaussian, alpha-stable and their mixtures. Interestingly, the results hold for all
cost functions that are ω (ln |x|) where it is guaranteed that the channel capacity exists and is finite.
The main idea behind the proofs of the theorems is the characterization of the behavioral pattern of the KKT
equation at infinity after providing lower and upper bounds on some quantities of interest. A key property is the
subadditivity of the logarithm of the inverse of a lower bound on the noise PDF at large values. This was referred to
as property C7 and is satisfied by all noise PDFs whose tail has a dominant polynomial or exponential component.
A direct implication of the results concerns the question on what are suitable power measures of the input signals
of a communication channel. Though the question seems to be absurd when dealing with noise models with finite
second moment where the natural power measure would be the standard average power –which corresponds the
cost function C(|x|) = x2 , defining a power measure when the noise second moment is infinite is deemed crucial.
This is due to the fact that the natural signal-to-noise ratio would be equal to zero. Based on our results, suitable
i
h
average power measures should correspond to cost functions which are “at most” Θ ln pN (f1(x)) since otherwise
capacity will be achieved by an input having a finite power while the input space contains distributions having an
infinite one. Hence, each channel has its suitable average power measure resulting from a suitable cost function.
For example, a suitable cost behaves “at most” like Θ(x2 ) for the linear Gaussian channel and has a logarithmic
growth for channels with polynomially tailed additive noise.
A PPENDIX I
W EAK D IFFERENTIABILITY
OF
I(·) AT F ∗
Theorem 11. Let F ∗ be an optimal input distribution. Under a cost constraint
R
C (|X|) dF (x) ≤ A, A > 0, the
mutual information I(F ) between the input and the output of channel (3) is weakly differentiable at F ∗ .
Before proceeding to the proof, we note that the existence of an optimal F ∗ and the finiteness of the solution
are insured as per the discussion in Section IV.
Proof: Let θ be a number in [0, 1], (F ∗ , F ) ∈ PA × PA and define Fθ = (1 − θ)F ∗ + θF . The weak derivative
November 17, 2017
DRAFT
29
of I(.) at F ∗ in the direction of F is defined as,
I ′ (F ∗ , F ) , lim+
θ→0
I(Fθ ) − I(F ∗ )
,
θ
whenever the limit exists. For simplicity, we denote by
t(x) = i(x; F ∗ ),
where i(x; F ) is given by equation (11), and we prove
Z
′
∗
I (F , F ) = − p(y; F ) ln p(y; F ∗ ) dy − hY (F ∗ )
Z
=
t(f (x)) dF (x) − hY (F ∗ ),
where by Tonelli, the interchange is valid as long as the integral term is finite which we prove next. Using L’Hôpital’s
rule,
I(Fθ ) − I(F ∗ )
hY (Fθ ) − hY (F ∗ )
= lim
θ
θ
θ→0+
θ→0+
Z
′
= lim+ −
p(y; Fθ ) ln p(y; Fθ ) dy ,
I ′ (F ∗ , F ) = lim
(28)
θ→0
where the derivative is with respect to θ. In order to evaluate
the derivative
R
p(y; Fθ ) ln p(y; Fθ ) dy
′
we use the definition of
′
p(y; Fθ ) ln p(y; Fθ ) dy
R
R
p(y; Fθ ) ln p(y; Fθ ) dy
p(y; Fθ+h ) ln p(y; Fθ+h ) dy
,
−
= lim
h→0
h
h
Z
where by the limit we mean that both, the limit as h goes to 0+ and the limit as h goes to 0− exist and are equal.
In what follows, we only provide detailed evaluations as h goes to 0+ since those when h goes to 0− are similar.
Using the mean value theorem, for some 0 ≤ c(h) ≤ h,
R
R
p(y; Fθ ) ln p(y; Fθ ) dy
p(y; Fθ+h ) ln p(y; Fθ+h ) dy
−
lim
h→0+
h
h
Z
′
= lim+ [p(y; Fθ ) ln p(y; Fθ )]|θ+c(h) dy.
h→0
Now, since p(y; Fθ ) = p(y; F ∗ ) + θ [p(y; F ) − p(y; F ∗ )],
Z
′
lim+ [p(y; Fθ ) ln p(y; Fθ )]|θ+c(h) dy
h→0
Z
Z
∗
= lim+ [p(y; F ) − p(y; F )] ln p(y; Fθ+c(h) ) dy + [p(y; F ) − p(y; F ∗ )] dy
h→0
Z
=
lim [p(y; F ) − p(y; F ∗ )] ln p(y; Fθ+c(h) ) dy
h→0+
Z
= [p(y; F ) − p(y; F ∗ )] ln p(y; Fθ ) dy,
November 17, 2017
(29)
(30)
DRAFT
30
where (30) is due to the fact that c(h) → 0 as h → 0 and that p(y; Fθ ) is continuous in θ by virtue of its linearity,
and (29) is due to DCT. Indeed,
[p(y; F ) − p(y; F ∗ )] ln p(y; Fθ+c(h) ) ≤ p(y; F ) + p(y; F ∗ ) ln p(y; Fθ+c(h) ) ,
and
p(y; Fθ+c(h) ) = [1 − θ − c(h)] p(y; F ∗ ) + [θ + c(h)] p(y; F )
≥ [1 − θ − c(h)] p(y; F ∗ ) ≥
1
p(y; F ∗ ),
2
whenever θ + c(h) ≤ 12 , which is true since both θ and c(h) are arbitrarily small. Therefore, since 0 < p(y; F ) < 1
for all F
1
∗
p(y; F ) .
[p(y; F ) − p(y; F )] ln p(y; Fθ+c(h) ) ≤ − p(y; F ) + p(y; F ) ln
2
R
Since hY (F ) = − p(y; F ) ln p(y; F ) dy is finite for all F in PA [40, Theorem 2], −p(y; F ∗ ) ln p(y; F ∗ ) is
∗
∗
integrable. It remains to prove that −p(y; F ) ln p(y; F ∗ ) is integrable to justify (29) and hence (30). To this end,
we will proceed by choosing first a specific F (·), namely
Bs
Bs
Fs (x) = 1 −
u(x)¶ +
u(x − xs ),
C(xs )
C(xs )
for some xs > 0 such that C(xs ) > 0 and where (0 <) Bs < min {A; C (xs )}. We note that Fs ∈ PA since
R
C(0) = 0 and hence C(|x|)dFs = Bs ≤ A. If Fs were the input distribution, it would induce the following output
Bs
Bs
pN (y) +
p(y; Fs ) = 1 −
pN (y − f (xs )).
(31)
C(xs )
C(xs )
Equation (31) along with lemma 1 and properties C7 and C8 show that −p(y; Fs ) ln p(y; F ∗ ) is integrable and (30)
is justified for F ≡ Fs . Hence,
′
∗
I (F , Fs )
=
=
Z
p(y; Fθ∗ ) ln p(y; Fθ∗ ) dy
′
lim −
Z
Z
∗
∗
lim
[p(y; Fs ) − p(y; F )] ln p(y; Fθ ) dy = t(f (x))dFs (x) − hY (F ∗ ).
θ→0+
θ→0+
where the interchange between the limit and integral sign is justified in an identical fashion as done to validate (30).
Now, since F ∗ is optimal, necessarily I ′ (F ∗ , Fs ) ≤ 0 (see Appendix C in [7]), which implies that
Z
t(f (x)) dFs (x) ≤ hY (F ∗ ).
Plugging in the expression of Fs (x) yields,
Bs
Bs
t(f (0)) +
1−
t(f (xs )) ≤ hY (F ∗ ) ⇔
C(xs )
C(xs )
t(f (xs )) ≤
hY (F ∗ ) − t(f (0))
C(xs ) + t(f (0)).
Bs
(32)
The above equation is valid for any xs > 0 (such that C(xs ) > 0) and therefore for all |x| ≥ xs since C(|x|) is
non-decreasing in |x|. we proceed by writing
Z
Z
t(f (x)) dF =
|x|≤xs
¶ where
t (f (x)) dF +
Z
t(f (x)) dF.
|x|>xs
u(x) denotes the Heaviside unit step function.
November 17, 2017
DRAFT
31
As for the first integral term, we have:
Z
t (f (x)) dF
|x|≤xs
=−
=−
−
Z
Z
|x|≤xs
Z
Z
|x|≤xs
|x|≤xs
pN (y − f (x)) ln p(y; F ∗ ) dy dF
Z
Z
|y|≥y0
|y|≤y0
pN (y − f (x)) ln p(y; F ∗ ) dy dF
pN (y − f (x)) ln p(y; F ∗ ) dy dF
(33)
Using lemma 1 and property C7, the first term of equation (33) is finite. As for the second term, it is finite by the
fact that p( y; F ∗ ) is positive and continuous hence achieves a positive minimum on compact subsets of R. When
it comes to the range |x| > xs , we use the upper bound in (32) which gives:
Z
Z
hY (F ∗ ) − t(f (0))
t(f (x)) dF ≤
C(|x|) + t(0) dF
Bu
|x|>xs
|x|>xs
≤
hY (F ∗ ) − t(f (0))
A + t(f (0)),
Bu
which is finite.
In conclusion,
−
and I ′ (F ∗ , F ) =
R
Z
∗
p(y; F ) ln p(y; F ) dy =
t(f (x)) dF − hY (F ∗ ), ∀ F ∈ PA .
Z
t(f (x)) dF < ∞,
Cost
The mapping from F to R:
T (F ) =
Z
C (|x|) dF − A
is weakly differentiable on PA as well. In fact,
T ′ (F ∗ , F ) = T (F ) − T (F ∗ ),
which is finite, since −A < T (F ) ≤ 0 for all F ∈ PA .
A PPENDIX II
R ATE OF D ECAY
OF
S(α, β, γ, δ) ON
THE
H ORIZONTAL S TRIP
We study in this appendix the rate of decay of alpha-stable distributions S(α, β, γ, δ) on the horizontal strip
Sη = {z ∈ C : |ℑ(z)| < η} where η is a small-enough positive number.
1
We prove in this appendix that |pN (z)| = O |ℜ(z)+δ|
as |ℜ(z)| → ∞, whenever N ∼ S(α, β, γ, δ) and
α+1
z ∈ Sη . The study is limited to the case: α ∈ [1, 2), β ∈] − 1, 1[, γ ∈ R+∗ and δ ∈ R.
Before we proceed, we first prove the following Lemma:
November 17, 2017
DRAFT
32
Lemma 7. Whenever N ∼ S(α, β, γ, δ), where α ∈ [1, 2), β ∈] − 1, 1[, γ ∈ R+∗ and δ ∈ R, pN (·) can be formally
extended on Sη = {z ∈ C : |ℑ(z)| < η} as
Z
1
pN (z) =
2π
Proof: By definition,
pN (x) =
1
2π
e−izt φ(t)dt.
(34)
R
Z
∞
e−ixt φ(t)dt,
−∞
where
exp iδt − γ α 1 − iβ sgn(t)Φ(t) |t|α
tan πα
α 6= 1
2
Φ(t) =
− 2 ln |t|
α = 1.
π
Let pN (z) be the extension of pN (x) on C. It is known that pN (z) is analytic on Sη (see [58] for example) .
φ(t)
=
Now, define
1
q(z) =
2π
Z
∞
e−izt φ(t)dt,
−∞
for all z = (x + iy) ∈ C. If we establish that q(z) is analytic on Sη then by the identity theorem, pN (z) = q(z),
for all z ∈ Sη . We start by proving the continuity of q(z):
Z ∞
1
e−izt φ(t)dt
lim q(z) = lim
z→z0
z→z0 2π −∞
Z
1
=
lim e−izt φ(t)dt
2π z→z0
Z
1
e−iz0 t φ(t)dt = q(z0 ).
=
2π
(35)
where the interchange in (35) is justified by DCT since:
α
e−izt φ(t) ≤ eyt−|γt| ,
which is integrable on Sη since η is small-enough and chosen so that |y| < η ≤ γ α . Now, let ∆ ⊂ Sη be a compact
triangle and denote by ∂∆ its boundary and |∆| its perimeter. We obtain
Z
Z Z
1
q(z)dz =
e−izt φ(t) dt dz
2π
∂∆
Z∂∆Z R
1
=
e−izt φ(t) dz dt
2π R ∂∆
Z
Z
=
φ(t)
e−izt dz = 0,
R
(36)
∂∆
−izt
where the last equation is due to the fact that e
is entire. The interchange in (36) is valid by Fubini since
Z Z
Z Z
Z
α
α
|∆|
1
1
eyt−|γt| dt dz <
eyt−|γt| | dt < ∞.
e−izt φ(t) dt dz ≤
2π ∂∆ R
2π ∂∆ R
2π R
By applying Morera’s Theorem [48, sec. 53], q(z) is analytic on Sη and the result is established.
November 17, 2017
DRAFT
33
Note that equation (34) shows that pN (z) = pN ′ (z − δ) where N ′ ∼ S(α, β, γ, 0). Therefore, and without loss
of generality, we restrict our analysis in the remainder of this section to pN (z), for N ∼ S(α, β, γ, 0).
For z = (x + iy),
pN (z) =
1
2π
1
=
2π
=
1
2π
Z
∞
−∞
Z
∞
α
e−izt−γ
−∞
yn
n!
n=0
Z
∞
−∞
1−iβ sgn(t)Φ(t) |t|α
dt =
1
2π
Z
∞
e−ixt+yt−γ
α
−∞
∞
X
yn n
t dt
n!
n=0
α
α
tn e−ixt−γ 1−iβ sgn(t)Φ(t) |t| dt.
e−ixt−γ
∞
X
α
1−iβ sgn(t)Φ(t) |t|α
dt
1−iβ sgn(t)Φ(t) |t|α
(37)
The interchange in (37) is justified by DCT. Indeed,
N
∞
X
X
α
y n n −ixt−γ α 1−iβ sgn(t)Φ(t) |t|α
|y|n n −|γt|α
t e
≤
|t| e
= e|y||t|−|γt| ,
n!
n!
n=0
n=0
which is integrable for |y| < η (≤ γ α ) and α ≥ 1. Now we proceed to studying the rate of decay in two separate
cases.
A. Rate of Decay for 1 < α < 2:
πα
2
In this case Φ(t) is a constant and it is equal to Φ(t) = Φ = tan
the change of variable u = γt
. Then, using equation (37), we obtain by
n Z ∞
∞
α
x
1 X 1 y
pN (z) =
tn e−i γ t−[1−iβ sgn(t)Φ]|t| dt
2πγ n=0 n! γ
−∞
∞
n
1 X 1 y
x
=
Tn − ; β ,
(38)
2πγ n=0 n! γ
γ
Z ∞
α
where Tn (x; β) is a function defined as Tn (x; β) =
ˆ
tn eixt−[1−iβ sgn(t)Φ]|t| dt.∗∗ Define k1 = (1 − iβΦ) and
−∞
denote by k 1 = (1 + iβΦ) its conjugate. In what follows, we study the behavior of the function Tn (x; β).
For n ≥ 1 and x > 0, we have
Z ∞
Z ∞
α
α
xn+α+1 Tn (x; β) = xn+α
xtn eixt−k1 t dt + (−1)n
xtn e−ixt−k1 t dt
0
0
Z ∞
Z ∞
′
α
α
tn+α−1 eixt−k1 t dt
dt + k1 α
= − i xn+α
tn eixt−k1 t
0
0
Z ∞
α
+ (−1)n−1
tn e−ixt−k1 t
tn+α−1 e−ixt−k1 t dt
dt + (−1)n−1 k 1 α
0
0
Z ∞
Z ∞
n+α
n−1 ixt−k1 tα
n−1
n−1 −ixt−k1 tα
dt
= inx
t
e
dt + (−1)
t
e
0
0
Z ∞
Z ∞
α
n+α−1 −ixt−k1 tα
n+α
n+α−1 ixt−k1 t
n−1
k1
k1
t
e
dt
(39)
− iαx
t
e
dt + (−1)
0
0
= i n xn+α Tn−1 (x; β) − iα k1 Sn (x; k1 ) + (−1)n−1 k 1 Sn (x; k 1 ) ,
(40)
Z
∗∗ Note
∞
′
α
(n)
n
1 (−i)
2π γ n+1
that Tn (−x; β) = (−1)n Tn (x; −β) and that pN (x) =
November 17, 2017
; β) =
Tn (− x
γ
in
1
2π γ n+1
; −β), n ∈ N∗ .
Tn ( x
γ
DRAFT
34
where equation (39) is obtained by integration by parts and regrouping, and where Sn (·; ·) is the complex conjugate
of Sn (·; ·) defined as,
n+α
Sn (x; k1 ) = x
Z
∞
t
n+α−1 ixt−k1 tα
e
0
where c =
1
n+α
dt = c
Z
∞
eiv
c
−k1 ζv αc
dv,
0
(> 0), ζ = x−α (> 0) and the change of variable is v = (xt)n+α . As x → ∞, ζ → 0+ and hence
Z ∞
Z ∞
c
αc
c icθ
αc iαcθ
lim Sn (x; k1 ) = c lim+
eiv −k1 ζv dv = c lim+
lim eiv e −k1 ζv e +iθ dv
x→+∞
θ→0
ζ→0
ζ→0
0
0
Z ∞
c icθ
αc iαcθ
= c lim+ lim
eiv e −k1 ζv e +iθ dv
(41)
ζ→0 θ→0 0
Z ∞
c icθ
αc iαcθ
= c lim lim
eiv e −k1 ζv e +iθ dv
(42)
+
θ→0 ζ→0
0
Z ∞
c icθ
= c lim
eiv e +iθ dv
(43)
θ→0 0
Z
c
= c lim
lim
eiz dz,
θ→0 R→∞,ρ→0
where z = ve
iθ
L1
iθ
and L1 = {z ∈ C : z = ve , 0 < ρ ≤ v ≤ R}. Equation (41) is justified by DCT since:
eiv
c icθ
e
−k1 ζv αc eiαcθ +iθ
≤ e−v
c
sin(cθ)−ζv αc [cos(αcθ)+βΦ sin(αcθ)]
ζ
αc
≤ e− 2 v ,
for small-enough θ, and the upper-bound is integrable since c and ζ are positive. The last inequality is justified by
1
virtue that sin(cθ) > 0 and cos(αcθ) + β tan απ
2 sin(αcθ) > 2 for small positive θ. Similarly, (43) is justified
because the integrand in (42) is O(e−v
c
sin cθ
) as ζ → 0+ which is also integrable. The interchange between the
two limits in (42) is valid by the preceding argument as long as the result in (43) is finite. To evaluate the limit of
R
iz c
dz as R → ∞, ρ → 0, we use contour integration over C shown in Figure 4.
L1 e
✻
L2
C1
✮
C2
Fig. 4.
✶
L1
✲
The contour C.
△ π
2c
The arcs C1 and C2 are of radius R, and ρ respectively and are between angles θ and ϕ =
mod 2π. Note
that since we are interested in the limit as θ goes to zero, we can always choose it small enough in order to have
the contour counter-clockwise. Finally, L2 is a line connecting the extremities of the arcs.
c
Now since f (z) = eiz is analytic on and inside C (by choosing an appropriate branch cut in the plane), by
November 17, 2017
DRAFT
35
Cauchy’s Theorem [65, p.111 Sec.2.2],
Z
Z
Z
Z
I
f (z).
f (z) +
f (z) +
f (z) +
f (z) dz =
0=
C
On C1 , we have:
Z
Z
lim
f (z)dz = lim
R→∞
R→∞
C1
ϕ
iReiφ eiR
c icφ
e
dφ ≤ lim
R→∞
θ
c
C2
L2
C1
L1
Z
ϕ
Re−R
c
sin(cφ)
dφ =
Z
ϕ
lim Re−R
θ
θ
c
sin(cφ)
R→∞
dφ = 0,
where the interchange is valid because Re−R sin(cφ) is decreasing as 0 < cθ ≤ cφ ≤ π2 . Similarly, on C2 ,
Z
Z ϕ
Z ϕ
Z ϕ
c icφ
c
c
lim
f (z)dz = lim
iρeiφ eiρ e dφ ≤ lim
lim ρe−ρ sin(cφ) dφ = 0,
ρe−ρ sin(cφ) dφ =
ρ→0
ρ→0
C2
ρ→0
θ
θ
θ
c
where we justify the interchange by virtue of the fact that ρe−ρ
sin(cφ)
ρ→0
is bounded for small values of ρ. It remains
π
i 2c
to evaluate the integral on L2 where z = te ,
Z
Z ∞
Z
c iπ
π
π
lim
f (z)dz = −
ei 2c eit e 2 dt = −ei 2c
R→∞,ρ→0
0
0
L2
∞
In conclusion,
lim
R→∞,ρ→0
Z
f (z) dz = e
π
i 2c
L1
1
Γ
c
c
π
e−t dt = −ei 2c
1
Γ
c
1
.
c
1
,
c
which implies that
π
lim Sn (x; k1 ) = ei 2 (n+α) Γ(n + α),
x→+∞
and by (40), we can write for n ≥ 1
lim
x→+∞
n+α+1
x
Tn (x; β) − i nxn+α Tn−1 (x; β)
iπ
(n+α)
n−1
−i π
(n+α)
2
2
,
= Wn (β) =
ˆ − iαΓ(n + α) k1 e
+ (−1)
k1 e
which implies that Un (β) =
ˆ
lim xn+α+1 Tn (x; β) is a well defined quantity because
x→+∞
U0 (β) = lim
x→+∞
xα+1 T0 (x; β) = 2πγ lim xα+1 pN (−γx) ,
x→+∞
exists –and is non zero for β 6= 1 and U0 (1) = 0, and
"
n
Un (β) = inUn−1 (β) + Wn (β) = n! i U0 (β) +
Furthermore, for n ≥ 0,
|Un (β)|
"
≤ n! |U0 (β)| +
"
n−1
X
k=0
|Wn−k (β)|
(n − k)!
n−1
X
#
n−1
X
k=0
"
#
ik
Wn−k (β) .
(n − k)!
≤ n! |U0 (β)| + 2α|k1 |
#
n−1
X
k=0
Γ(n + α − k)
(n − k)!
#
Γ(n + 2 − k)
(n − k)!
k=0
"
#
n−1
X
|U0 (β)|
2
= n! |U0 (β)| + 4|k1 |
(n + 1 − k) = 2 n! |k1 |n + 3|k1 |n +
,
2
≤ n! |U0 (β)| + 4|k1 |
(44)
k=0
where equation (44) is justified using the fact that 0 < α < 2 and Γ(α + l) is increasing in α > 0 for l ∈ N∗ .
November 17, 2017
DRAFT
36
Now using equation (38),
n
∞
X
x
1 y
1
α+1
lim x
|pN (z)| =
Tn − ; β
lim x
x→∞
2πγ x→∞
n! γ
γ
n=0
n
∞
x
1 X 1 y
lim xα+1 Tn − ; β
=
x→∞
2πγ n=0 n! γ
γ
∞
n
x
1 X 1 y
α+1
; −β
lim x
Tn
≤
2πγ n=0 n! γ x→∞
γ
∞
∞
n
1 X 1 y
x
1 X 1 y
n+α+1
≤
; −β =
lim x
Tn
2πγ n=0 n! γ x→∞
γ
2πγ n=0 n! γ
∞
|U0 (−β)|
γα X n
,
≤
|y| |k1 |n2 + 3|k1 |n +
π n=0
2
α+1
(45)
n
γ n+α+1 |Un (−β)|
which is finite because |y| < η which is small-enough (and assumed to be less than one), and where we used the
fact that f (x) = |x| is continuous. The interchange in (45) is valid because the end result is finite.
In conclusion, lim xα+1 |pN (z)| < ∞ which concludes our proof.
x→+ ∞
B. Rate of Decay for α = 1:
In this case, Φ(t) = − π2 log |t| is a function of t. According to equation (37) and for z = x + iy,
Z
∞
1 X y n ∞ n −ixt−γ 1−iβ sgn(t)Φ(t) |t|
t e
dt.
(46)
pN (z) =
2π n=0 n! −∞
R∞
Once more, we study the behavior of the integral In (x) = −∞ tn e−ixt−γ(1−iβ sgn(t)Φ)|t| dt as x → ∞ for n ≥ 0.
November 17, 2017
DRAFT
37
We note that I0 (x) = 2πpN (x; 1, β, γ, 0) which is Θ
Z +∞
In (x) =
tn e−ixt−γ(1−iβ sgn(t)Φ)|t| dt
−∞
+∞
=
Z
tn e−ixt−γ (1+i π β log(t))t dt +
2
0
=
Z
+∞
0
1
x2
Z
. For n ≥ 1,
0
tn e−ixt−γ (1−i π β log(−t))(−t) dt
−∞
e−ixt tn e−γ (1+i π β log(t))t dt + (−1)n
2
2
Z
+∞
eixt tn e−γ (1−i π β log(t))t dt
2
0
+∞
+∞
1 ixt n −γ (1−i π2 β log(t))t
1 −ixt n −γ (1+i π2 β log(t))t
n
+ (−1)
= − e
t e
e t e
ix
ix
0
0
Z +∞
2
1
2
2
+
e−ixt ntn−1 − γtn − i βγtn − i βγtn log(t) e−γ (1+i π β log(t))t dt
ix 0
π
π
Z
2
2
2
(−1)n+1 +∞ ixt n−1
nt
− γtn + i βγtn + i βγtn log(t) e−γ (1−i π β log(t))t dt
e
+
ix
π
π
0
Z
2
1 +∞ −ixt n−1
2
2
=
nt
− γtn − i βγtn − i βγtn log(t) e−γ (1+i π β log(t))t dt
e
ix 0
π
π
Z
2
2
2
(−1)n+1 +∞ ixt n−1
nt
− γtn + i βγtn + i βγtn log(t) e−γ (1−i π β log(t))t dt
e
+
ix
π
π
0
+∞
2
2
2
1
= 2 e−ixt ntn−1 − γtn − i βγtn − i βγtn log(t) e−γ (1+i π β log(t))t
x
π
π
0
Z +∞
1
e−ixt gn (t) dt
− 2
x 0
+∞
2
1
2
2
+ (−1)n+1 − 2 eixt ntn−1 − γtn + i βγtn + i βγtn log(t) e−γ (1−i π β log(t))t
x
π
π
0
n+1 Z +∞
(−1)
+
eixt hn (t) dt
x2
0
Z +∞
Z +∞
1
= 2 (−1)n+1
e−ixt gn (t) dt
eixt hn (t) dt −
x
0
0
(47)
(48)
(49)
where equations (47) and (48) are due to integration by parts. The functions gn (·) and hn (·), n ≥ 1 are defined on
R+∗ and are given by:
4
4
4
8
gn (t) = n(n − 1)tn−2 − 2nγtn−1 + (γ 2 − 2 β 2 γ 2 + i βγ 2 )tn + (− 2 β 2 γ 2 + i βγ 2 )tn log(t)
π
π
π
π
2
2
4
4 2 2 n
2
n−1
n−1
−i (2n + 1)βγt
− i nβγt
log(t) − 2 β γ t log (t) e−γ (1+i π β log(t))t .
π
π
π
(50)
The term n(n − 1)tn−2 is equal to zero when n = 1 and hn (t) is deduced from gn (t) by replacing β by −β. The
functions gn (t), hn (t) are L1 (R+ ) functions and hence by Riemann-Lebesgue [66, p.3 sec.2 th.1] their L1 (R+ )
Fourier transforms are o(1). Therefore equation (49) is o( x12 ). Equivalently, In (x) = o( x12 ) as x → ∞ for all
November 17, 2017
DRAFT
38
n ≥ 1. Now using equation (46) we obtain:
lim 2πx2 |pN (z)|
x→∞
= lim
x→∞
∞
X
yn 2
x In (x)
n!
n=0
Z +∞
Z +∞
∞
X
yn
(−1)n+1
eixt hn (t) dt −
e−ixt gn (t) dt
x→∞
n!
0
0
n=1
Z
Z
∞
+∞
+∞
X |y|n
|hn (t)| dt +
|gn (t)| dt
≤ lim 2πx2 pN (x) + lim
x→∞
x→∞
n!
0
0
n=1
Z
∞
+∞
n
X
|y|
= lim 2πx2 pN (x) +
(|hn (t)| + |gn (t)|) dt
x→∞
n! 0
n=1
Z +∞ X
∞
|y|n
(|hn (t)| + |gn (t)|) dt
= lim 2πx2 pN (x) +
x→∞
n!
0
n=1
= lim 2πx2 pN (x) +
(51)
The interchange in (51) is valid since:
∞
X
|y|n
(|hn (t)| + |gn (t)|)
n!
n=1
≤
∞
X
|y|n
A1 n(n − 1)tn−2 + A2 tn + A3 tn | log(t)| + A4 (2n + 1)tn−1
n!
n=1
+A5 ntn−1 | log(t)| + A6 tn log2 (t) e−γt
∞
X
|y|n
A1 n(n − 1)tn−2 + (A2 + A3 | log(t)| + A6 log2 (t))tn
n!
n=1
+n(3A4 + A5 | log(t)|)tn−1
i
h
≤ e−γt A1 y 2 e|y|t + (A2 + A3 | log(t)| + A6 log2 (t))(e|y|t − 1) + |y|(3A4 + A5 )e|y|t
≤ e−γt
(52)
≤ e−(γ−|y|)t A2 + (3A4 + A5 )|y| + A1 y 2 + A3 | log(t)| + A6 log2 (t)
which is integrable on [0, +∞[ since |y| < η (< γ). The Ai s, 1 ≤ i ≤ 6 are positive constants function of β, γ and
can be derived from the expression of gn (t) (equation 50) and from that of hn (t) accordingly after taking the norm
of each term in those expressions. To write equation (52), we used the obvious inequality 2n + 1 ≤ 3n whenever
n ≥ 1. Back to (51),
lim 2πx2 |pN (z)| ≤ lim 2πx2 pN (x) +
x→∞
x→∞
≤ lim 2πx2 pN (x) +
x→∞
Z
0
Z
∞
+∞ X
|y|n
(|hn (t)| + |gn (t)|) dt
n!
n=1
+∞
l(t) dt
0
where l(t) = e−(γ−|y|)t A2 + (3A4 + A5 )|y| + A1 y 2 + A3 | log(t)| + A6 log2 (t) . Since limx→∞ 2πx2 pN (x) and
R +∞
l(t) dt are both finite and non zero when |y| < η (< γ), then 0 ≤ lim 2πx2 |pN (z)| < ∞ and |pN (z)| =
0
x→∞
1
O |ℜ(z)|2 as ℜ(z) → ∞ whenever z ∈ Sη .
November 17, 2017
DRAFT
39
ACKNOWLEDGMENTS
The authors would like to thank Professor Aslan Tchamkerten for suggesting the idea of the converse.
R EFERENCES
[1] C. E. Shannon, “A mathematical theory of communication, part i,” Bell Syst. Tech. J., vol. 27, pp. 379–423, 1948.
[2] ——, “A mathematical theory of communication, part ii,” Bell Syst. Tech. J., vol. 27, pp. 623–656, 1948.
[3] I. Abou-Faycal, M. D. Trott, and S. Shamai, “The capacity of discrete-time memoryless Rayleigh-fading channels,” Information Theory,
IEEE Transactions on, vol. 47, no. 4, pp. 1290–1301, May 2001.
[4] M. Katz and S. Shamai, “On the capacity-achieving distribution of the discrete-time noncoherent and partially coherent AWGN channels,”
Information Theory, IEEE Transactions on, vol. 50, no. 10, pp. 2257–2270, October 2004.
[5] R. Nuriyev and A. Anastasopoulos, “Capacity and coding for the block-independent noncoherent AWGN channel,” IEEE Transactions on
Information Theory, vol. 51, no. 3, pp. 866–883, March 2005.
[6] C. Luo, “Communication for wideband fading channels: on theory and practice,” Ph.D. dissertation, Massachusetts Institute of Technology,
February 2006.
[7] J. Fahs, I. Abou-Faycal, “Using Hermite bases in studying capacity-achieving distributions over AWGN channels,” Information Theory,
IEEE Transactions on, vol. 58, no. 8, August 2012.
[8] J. G. Smith, “The information capacity of peak and average power constrained scalar Gaussian channels,” Inform. Contr., vol. 18, pp.
203–219, 1971.
[9] L. Zhang and D. Guo, “Capacity of Gaussian Channels with Duty Cycle and Power Constraints,” in IEEE International Symposium on
Information Theory, Saint Petersburg, Russia, 2011, pp. 424–428.
[10] A. Das, “Capacity-achieving distributions for non-Gaussian additive noise channels,” in Proc. IEEE International Symposium on Information
Theory, p. 432, June 2000, sorrento, Italy.
[11] J. Fahs, N. Ajeeb, and I. Abou-Faycal, “The capacity of average power constrained additive non-Gaussian noise channels,” in IEEE
International Conference on Telecommunications.
Beirut, Lebanon, April 2012.
[12] A. Tchamkerten, “On the Discreteness of Capacity-Achieving Distributions,” IEEE Transactions on Information Theory, vol. 50, no. 11,
pp. 2773–2778, November 2004.
[13] I. Abou-Faycal, J. Fahs, “On the capacity of some deterministic non-linear channels subject to additive white Gaussian noise,” in IEEE
17th International Conference on Telecommunications (ICT), Doha, Qatar, April 2010, pp. 63–70.
[14] J. Fahs and I. Abou-Faycal, “On the capacity of additive white alpha-stable noise channels,” in IEEE International Symposium on
Information Theory, Cambridge, MA, USA, 2012, pp. 294–298.
[15] J. Fahs, I. Abou-Faycal, “On the single-user capacity of some multiple access channels,” in The Eleventh International Symposium on
Wireless Communication Systems, Barcelona, Spain, August, 26-29 2014.
[16] V. Anantharam and S. Verdu, “Bits through queues,” IEEE Transactions on Information Theory, vol. 42, no. 1, pp. 4–18, January 1996.
[17] J. Fahs, I. Abou-Faycal, “A Cauchy input achieves the capacity of a Cauchy channel under a logarithmic constraint,” in IEEE International
Symposium on Information Theory, Honolulu, HI, USA, June 29 - July 4 2014.
[18] R. S. Blum, R. J. Kozick, and B. M. Sadler, “An adaptive spatial diversity receiver for non-Gaussian interference and noise,” Signal
Processing, IEEE Transactions on, vol. 47, no. 8, pp. 2100–2111, Aug. 1999.
[19] A. Nasri, and R. Schober, “Performance of BICM-SC and BICM-OFDM systems with diversity reception in non-Gaussian noise and
interference,” Communications, IEEE Transactions on, vol. 57, no. 11, pp. 3316–3327, Nov. 2009.
[20] J. Fiorina, “A simple IR-UWB receiver adapted to multi-user interferences,” in IEEE Globecom, San Francisco, CA, 27 November - 1
December 2006, pp. 1–4.
[21] N. Beaulieu, H. Shao, and J. Fiorina, “P-order metric UWB receiver structures with superior performance,” Communications, IEEE
Transactions on, vol. 56, no. 10, pp. 1666–1676, October 2008.
[22] B. W. Stuck and B. Kleiner, “A statistical analysis of telephone noise,” Bell Syst. Tech. J., vol. 53, no. 7, pp. 1263–1320, 1974.
[23] P. G. Georgiou, P. Tsakalides, and C. Kyriakakis, “Alpha-stable modeling of noise and robust time-delay estimation in the presence of
impulsive noise,” Multimedia, IEEE Transactions on, vol. 1, no. 3, pp. 291–301, 1999.
November 17, 2017
DRAFT
40
[24] E. S. Sousa, “Performance of a spread spectrum packet radio network link in a Poisson field of interferers,” Information Theory, IEEE
Transactions on, vol. 38, no. 6, pp. 1743–1754, Nov. 1992.
[25] J. Ilow and D. Hatzinakos, “Analytic alpha-stable noise modeling in a Poisson field of interferers or scatterers,” Signal Processing, IEEE
Transactions on, vol. 46, no. 6, pp. 1601–1611, Jun. 1998.
[26] M. Win, P. Pinto, and L. Shepp, “A mathematical theory of network interference and its applications,” Proceedings of the IEEE, vol. 97,
no. 2, pp. 205 –230, February 2009.
[27] N. Beaulieu and D. Young, “Designing time-hopping ultrawide bandwidth receivers for multiuser interference environments,” Proceedings
of the IEEE, vol. 97, no. 2, pp. 255 –284, February 2009.
[28] M. Nassar, K. Gulati, A. Sujeeth, N. Aghasadeghi, B. Evans, and K. Tinsley, “Mitigating near-field interference in laptop embedded
wireless transceivers,” in IEEE International Conference on Acoustics, Speech and Signal Processing, Las Vegas, NV, 30 March - 4 April
2008, pp. 1405 –1408.
[29] H. El Ghannudi, L. Clavier, N. Azzaoui, F. Septier, and P.-a. Rolland, “Stable interference modeling and Cauchy receiver for an IR-UWB
ad hoc network,” Communications, IEEE Transactions on, vol. 58, no. 6, pp. 1748 –1757, Jun. 2010.
[30] A. Rajan and C. Tepedelenlioglu, “Diversity combining over Rayleigh fading channels with symmetric alpha-stable noise,” Communications,
IEEE Transactions on, vol. 9, no. 9, pp. 2968–2976, Sep. 2010.
[31] D. Middleton and A. D. Spaulding, “A tutorial review of elements of weak signal detection in non-Gaussian EMI environments,” U.S.
Dept. of Commerce,” NTIA Rep. 86-194, 1986.
[32] Y. Kim and G. T. Zhou, “The Middleton class B model and its mixture representation,” Center for Signal and Image Processing, Georgia
Institute of Technology, Atlanta, GA, Tech. Rep. CSIP TR-98-01, May 1998.
[33] M. Shao and C. Nikias, “Signal processing with fractional lower order moments: stable processes and their applications,” Proceedings of
the IEEE, vol. 81, no. 7, pp. 986 –1010, Jul. 1993.
[34] P. Tsakalides and C. L. Nikias, “Maximumm likelihood localization of sources in noise modeled as a stable process,” Signal Processing,
IEEE Transactions on, vol. 43, no. 11, pp. 2700–2713, Nov. 1995.
[35] X. Ma and C. L. Nikias, “Joint estimation of time delay and frequency delay in impulsive noise using fractional lower order statistics,”
Signal Processing, IEEE Transactions on, vol. 44, no. 11, pp. 2669–2687, Nov. 1996.
[36] G. A. Tsihrintzis and C. L. Nikias, “Performance of optimum and suboptimum receivers in the presence of impulsive noise modeled as
an alpha-stable process,” Communications, IEEE Transactions on, vol. 43, no. 2/3/4, pp. 904–914, Feb./Mar./Apr. 1995.
[37] J. G. Gonzalez, J. L. Paredes, and G. R. Arce, “Zero-order statistics: A mathematical framework for the processing and characterization
of very impulsive signals,” Signal Processing, IEEE Transactions on, vol. 54, no. 10, pp. 3839–3851, Nov. 2006.
[38] O. Rioul, “Information theoretic proofs of entropy power inequality,” IEEE Transactions on Information Theory, vol. 57, no. 1, pp. 33–55,
January 2011.
[39] B. V. Bahr and C. Esseen, “Inequalities for the rth absolute moment of a sum of random variables, 1 ≤ r ≤ 2,” The Annals of Mathematical
Statistics, vol. 36, no. 1, pp. 299–303, February 1965.
[40] J. Fahs and I. Abou-Faycal, “On the finiteness of the capacity of continuous channels,” IEEE Transactions on Communications, vol. 64,
no. 1, pp. 166–173, January 2016.
[41] A. N. Shiryaev, Probability, 2nd ed.
Springer-Verlag, 1996.
[42] D. G. Luenberger, Optimization By Vector Space Methods.
New York: Wiley, 1969.
[43] E. Agrell, “Conditions for a monotonic channel capacity,” IEEE Transactions on Communications, vol. 63, no. 3, pp. 738–748, March
2015.
[44] R. Gallager, Information Theory and Reliable Communication.
John Wiley & Sons, Nov. 1968.
[45] W. Hirt and J. Massey, “Capacity of the Discrete-Time Gaussian Channel with Intersymbol Interference,” Information Theory, IEEE
Transactions on, vol. 34, no. 3, pp. 380–388, May 1988.
[46] J. Fahs and I. Abou-Faycal, “On the detrimental effect of assuming a linear model for non-linear AWGN channels,” in IEEE International
Symposium on Information Theory, Saint Petersburg, Russia, 2011, pp. 1693–1697.
[47] Y. Tsai, C. Rose, R. Song, and I. S. Mian, “An Additive Exponential Noise Channel with a Transmission Deadline,” in IEEE International
Symposium on Information Theory, Saint Petersburg, Russia, 2011, pp. 598–602.
[48] R. V. Churchill, J. W. Brown and R. F. Verhey, Complex Variables and Applications, 3rd ed.
November 17, 2017
McGraw-Hill, 1976.
DRAFT
41
[49] J. H. Miller and J. B. Thomas, “Detectors for discrete-time signals in non-Gaussian noise,” Information Theory, IEEE Transactions on,
vol. 18, no. 2, pp. 241–250, Mar. 1972.
[50] M. P. Shinde and S. N. Gupta, “Signal detection in the presence of atmospheric noise in tropics,” Communications, IEEE Transactions
on, vol. 22, pp. 1055–1063, Aug. 1974.
[51] M. Bouvet and S. C. Schwartz, “Comparison of adaptive and robust receivers for signal detection in ambient underwater noise,” Acoustics,
Speech and Signal Processing, IEEE Transactions on, vol. 37, pp. 621–626, May 1989.
[52] K. L. Blackard, T. S. Rappaport, and C. W. Bostian, “Radio frequency noise measurments and models for indoor wireless communications
at 918 MHz, 2.44 GHz, AND 4.0 GHz,” in IEEE International Conference on Communications, vol. 1, Denver, CO, June 1991, pp. 28–32.
[53] J. G. Gonzalez, “Robust Techniques for Wireless Communications in non-Gaussian Environments,” Ph.D. dissertation, University of
Delaware, 1997.
[54] M. Shao and C. Nikias, “Signal processing with fractional lower order moments: stable processes and their applications,” in Proceedings
of the IEEE, vol. 81, July 1993, pp. 986 –1010.
[55] S. A. Kassam, Signal Detection in Non-Gaussian Noise.
Springer-Verlag, 1988.
[56] W. Feller, An Introduction to Probability Theory and Its Applications.
Wiley, New York, 1966, vol. 2.
[57] B. V. Gnedenko and A. N. Kolmogorov, Limit Distributions for Sums of Independent Random Variables.
Reading Massachusetts:
Addison-Wesley Publishing Company, 1968.
[58] V. M. Zolotarev, One-dimensional Stable Distributions.
American Mathematical Society, 1983, vol. 65.
[59] V. V. Uchaikin and V. M. Zolotarev, CHANCE and STABILITY: Stable Distributions and their Applications.
Utrecht, Netherlands: VSP,
1999.
[60] J. P. Nolan, Stable Distributions - Models for Heavy Tailed Data.
Boston: Birkhauser, 2012, in progress, Chapter 1 online at
academic2.american.edu/∼jpnolan.
[61] I. A. Ibragimov and Y. V. Linnik, Independent and Stationary Sequences of Random Variables.
Wolters-Noordhoff, Groningen: J.F.C.
Kingman, 1971.
[62] B. L. Hughes, “Alpha-stable models of multiuser interference,” in IEEE International Symposium on Information Theory, Sorrento, Italy,
2000.
[63] K. Gulati, B. L. Evans, J. G. Andrews, and K. R. Tinsley, “Statistics of co-channel interference in a field of Poisson and Poisson-Poisson
clustered interferers,” Signal Processing, IEEE Transactions on, vol. 58, no. 12, pp. 6207–6222, Dec. 2010.
[64] A. Chopra, “Modeling and mitigation of interference in wireless receivers with multiple antennas,” Ph.D. dissertation, University of Texas
at Austin, December 2011.
[65] J. E. Marsden and M. J. Hoffman, Basic Complex Analysis, 3rd ed.
W. H. Freeman and Company, 1999.
[66] S. Bochner and K. Chandrasekharan, Fourier Transforms, ser. Annals of mathematics studies. Princeton University Press, 1949. [Online].
Available: http://books.google.fr/books?id=zsfbTJkyp90C
November 17, 2017
DRAFT
| 7cs.IT
|
arXiv:1711.07459v1 [cs.NE] 20 Nov 2017
SquishedNets: Squishing SqueezeNet further for edge
device scenarios via deep evolutionary synthesis
Mohammad Javad Shafiee
Dept. of Systems Design Engineering
University of Waterloo, DarwinAI
[email protected]
Francis Li
Dept. of Systems Design Engineering
University of Waterloo, DarwinAI
[email protected]
Brendan Chwyl
Dept. of Systems Design Engineering
University of Waterloo, DarwinAI
[email protected]
Alexander Wong
Dept. of Systems Design Engineering
University of Waterloo, DarwinAI
[email protected]
Abstract
While deep neural networks have been shown in recent years to outperform other
machine learning methods in a wide range of applications, one of the biggest challenges with enabling deep neural networks for widespread deployment on edge devices such as mobile and other consumer devices is high computational and memory requirements. Recently, there has been greater exploration into small deep
neural network architectures that are more suitable for edge devices, with one of
the most popular architectures being SqueezeNet, with an incredibly small model
size of 4.8MB. Taking further advantage of the notion that many applications of
machine learning on edge devices are often characterized by a low number of target classes, this study explores the utility of combining architectural modifications
and an evolutionary synthesis strategy for synthesizing even smaller deep neural
architectures based on the more recent SqueezeNet v1.1 macroarchitecture (considered state-of-the-art in efficient architectures) for applications with fewer target classes. In particular, architectural modifications are first made to SqueezeNet
v1.1 to accommodate for a 10-class ImageNet-10 dataset, and then an evolutionary
synthesis strategy is leveraged to synthesize more efficient deep neural networks
based on this modified macroarchitecture. The resulting SquishedNets possess
model sizes ranging from 2.4MB to 0.95MB (∼5.17X smaller than SqueezeNet
v1.1, or 253X smaller than AlexNet). Furthermore, the SquishedNets are still able
to achieve accuracies ranging from 81.2% to 77%, and able to process at speeds of
156 images/sec to as much as 256 images/sec on a Nvidia Jetson TX1 embedded
chip. These preliminary results show that a combination of architectural modifications and an evolutionary synthesis strategy can be a useful tool for producing
very small deep neural network architectures that are well-suited for edge device
scenarios without the need for compression or quantization.
1 Introduction
One can consider deep neural networks [5] to be one of the most successful machine learning methods, outperforming many state-of-the-art machine learning methods in a wide range of applications
ranging from image categorization [4] to speech recognition. A very major factor to the great recent successes of deep neural networks has been the availability of very powerful high performance
computing systems due to the great advances in parallel computing hardware. This has enabled
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
researchers to significantly increase the depth and complexity of deep neural networks, resulting in
greatly increased modeling capabilities. As such, the majority of research in deep neural networks
have largely focused on designing deeper and more complex deep neural network architectures for
improved accuracy. However, the increasing complexity of deep neural networks has become one
of the biggest obstacles to the widespread deployment of deep neural networks on edge devices
such as mobile and other consumer devices, where computational, memory, and power capacity is
significantly lower than that in high performance computing systems.
Given the proliferation of edge devices and the increasing demand for machine learning applications in such devices, there has been an increasing amount of research exploration on the design of
smaller, more efficient deep neural network architectures that can both infer and train faster, as well
as transfer faster onto embedded chips that power such edge devices. One commonly employed approach for designing smaller neural network architectures is synaptic precision reduction, where the
number of bits used to represent synaptic strength is significantly reduced from 32-bit floating point
precision to fixed-point precision [12], 2-bit precision [6, 13, 7], or 1-bit precision [1, 8]. While this
approach leads to greatly reduced model sizes, the resulting deep neural networks often require specialized hardware support for accelerated deep inference and training on embedded devices, which
can limit their utility for wide range of applications.
Another approach to designing smaller deep neural network architectures is to take a principled
approach and employ architectural design strategies to achieve more efficient deep neural network
macroarchitectures [2, 3]. An exemplary case of what can be achieved using such an approach
is SqueezeNet [3], where three key design strategies where employed: 1) decrease the number of
3x3 filters, 2) decrease the number of input channels to 3x3 filters, and 3) downsample late in
the network. This resulted in a macroarchitecture composed of Fire modules that possessed an
incredibly small model size of 4.8MB, which is 50X smaller than AlexNet with comparable accuracy
on ImageNet for 1000 classes. The authors further introduced SqueezeNet v1.1, where the number
of filters as well as the filter sizes are further reduced, resulting in 2.4X less computation than
the original SqueezeNet without sacrificing accuracy and thus can be considered state-of-the-art in
efficient network architectures.
Inspired by the incredibly small macroarchitecture of SqueezeNet, we are motivated to take it one
step further by taking into account that the majority of applications of machine learning on edge
devices such as mobile and consumer devices are quite specialized and often require a much fewer
number of target classes (typically less than 10 target classes). As such, this study explores the utility of combining architectural modifications and an evolutionary synthesis strategy for synthesizing
even smaller deep neural architectures based on the SqueezeNet v1.1 macroarchitecture for applications with fewer target classes. We will refer to these smaller deep neural network architectures as
SquishedNets.
2 Architectural modification for fewer target classes
The first and simplest strategy taken in this study is to perform some simple architectural modifications to the macroarchitecture of SqueezeNet v1.1 to accommodate for scenarios where we are
dealing with much fewer target classes. For this study, we explored a classification scenario where
the number of target classes is reduced to 10, as fewer target classes is quite common for many machine learning applications on edge devices such as mobile and other consumer devices. Given the
reduced number of target classes, we modify the conv10 layer to a set of 10 1x1 filters. Given that
the conv10 layer contains ∼40% of all parameters in SqueezeNet [3], this architectural modification
resulted in a significant reduction in model size. While not a leap of the imagination and a rather
trivial modification, this illustrates that SqueezeNet v1.1 is a great macroarchitecture that can be
modified to be even more efficient for edge scenarios where there are fewer target classes.
3 Evolutionary synthesis of more efficient network architectures
The second strategy taken in this study is to employ an evolutionary synthesis strategy [10, 11, 9] to
synthesize even more efficient deep neural network architectures than can be achieved through principled macroarchitecture design strategies. First proposed by [10] and subsequently extended [11, 9],
the idea behind evolutionary synthesis is to automatically synthesize progressively more efficient
deep neural networks over successive generations in a stochastic synaptogenesis manner within a
probabilistic framework. More specifically, synaptic probability models are used to encode the ge2
Table 1: ImageNet-10 dataset
wnid
n02783161
n03085013
n04557648
n04037443
n03793489
Class Name
pen
keyboard
water bottle
car
computer mouse
wnid
n03584254
n04548362
n07930864
n03782006
n04409515
Class Name
cell phone
wallet
cup
monitor
tennis ball
netic information of a deep neural network, thus mimic the notion of heredity. Offspring networks
are then synthesized in a stochastic manner given the synaptic probability models and a set of computational environmental constraints for influencing synaptogenesis, thus forming the next generation
of deep neural networks, thus mimicking the notions of random mutation and natural selection. The
offspring deep neural networks are then trained and this evolution synthesis process is performed
over generations until the desired traits are met.
While a more detailed description of the evolutionary synthesis strategy can be found in [10], a
brief mathematical description is provided as follows. Let the genetic encoding of a network be
formulated as P (Hg |Hg−1 ), where the network architecture Hg at generation g is influenced by the
network architecture Hg−1 of generation g − 1. An offspring network is synthesized in a stochastic
manner via a synthesis probability P (Hg ), which combines the genetic encoding P (Hg |Hg−1 ) with
the environmental factor model R being imposed:
P (Hg ) ≈ P (Hg |Hg−1 ) · R.
(1)
Since the underlying goal is to influence the synaptogenesis behaviour of offspring deep neural networks to be progressively more efficient generation after generation, the environmental factor model
R is set to a value less than 1 to further enforce a resource-starved environment on the deep neural
networks from generation to generation. In this study, the aforementioned modified macroarchitecture is used as the ancestral precursor for the evolutionary synthesis process to produce deep neural
networks with even smaller network architectures. In this study, 15 generations of evolutionary
synthesis was performed to produce the final SquishedNets.
Table 2: Performance results of SquishedNets.
Model
Name
Model
size
SquishedNet-1
SquishedNet-2
SquishedNet-3
SquishedNet-4
2.4MB
2.0MB
1.3MB
0.95MB
Reduction in
model size
vs. SqueezeNet v1.1
2.04X
2.45X
3.77X
5.17X
Reduction in
model size
vs. AlexNet
100X
120X
184X
253X
Runtime
speed
(images/sec)
156.09
174.86
225.35
256.00
Top-1
accuracy
(ImageNet-10)
81.2%
79.6%
78.6%
77.0%
4 Preliminary Results and Discussion
To study the utility of a combination of architectural modifications and evolutionary synthesis on
synthesizing very small deep neural network architectures based on the SqueezeNet v1.1 macroarchitecture that are well-suited for edge device scenarios, we examine the top-1 accuracies and runtime speeds (on an Nvidia Jetson TX1 embedded chip with batch size of 32) of our synthesized
SquishedNets on the 10-class ImageNet-10 dataset. The ImageNet-10 dataset used in this study is
a subset of the ImageNet dataset composed of the following ten target classes reported in Table 1.
The performance results of four different SquishedNets (produced at four different generations of
the evolutionary synthesis process) are shown in Table 1. A number of observations can be made
based on Table 2. First, it can be observed that leveraging both architectural modifications to account
for fewer target classes as well as evolutionary synthesis results in the generation of even more efficient network architectures, as evident by the SquishedNets having model sizes range from 2.4MB
to just 0.95MB. Therefore, the smallest SquishedNet is 5.17X smaller than SqueezeNet v1.1 (or
∼253X smaller than AlexNet). Second, not only was there a significant model size reductions, the
SquishedNets were able to process at speeds of 156 images/sec to as much as 256 images/sec on a
3
Nvidia Jetson TX1 embedded chip, which is significant particularly for edge scenarios with mobile
and other consumer devices. Therefore, the ability to not only achieve very small model sizes but
also fast runtime speeds has great benefits when used in resource-starved environments with limited
computational, memory, and energy requirements. In terms of top-1 accuracy, the SquishedNets are
able to still achieve accuracies ranging from 81.2% to 77.0%, which is high enough for many edge
applications. These preliminary results show that a combination of architectural modifications and
an evolutionary synthesis strategy can be a useful tool for producing very small deep neural network architectures that are well-suited for edge device scenarios without the need for compression
or quantization.
Acknowledgment
The authors thank NSERC, the Canada Research Chairs program, Nvidia, and DarwinAI.
References
[1] M. Courbariaux, Y. Bengio, and J.-P. David. Binaryconnect: Training deep neural networks with binary
weights during propagations. NIPS, 2015.
[2] A. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam.
Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861,
2017.
[3] Forrest N. Iandola, Song Han, Matthew W. Moskewicz, Khalid Ashraf, William J. Dally, and Kurt
Keutzer. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <0.5mb model size.
arXiv:1602.07360, 2016.
[4] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional
neural networks. In NIPS, 2012.
[5] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 2015.
[6] Fengfu Li, Bo Zhang, and Bin Liu. Ternary weight networks. arXiv:1605.04711, 2016.
[7] W. Meng, Z. Gu, M. Zhang, and Z. Wu. Two-bit networks for deep learning on resource-constrained
embedded devices. arXiv:1701.00485, 2017.
[8] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor-net: Imagenet classification using binary
convolutional neural networks. ECCV, 2015.
[9] M. Shafiee, F. Li, and A. Wong. Exploring the imposition of synaptic precision restrictions for evolutionary synthesis of deep neural networks. In https://arxiv.org/abs/1707.00095, 2016.
[10] M. Shafiee, A. Mishra, and A. Wong. Deep learning with darwin: Evolutionary synthesis of deep neural
networks. arXiv:1606.04393, 2016.
[11] M. Shafiee and A. Wong. Evolutionary synthesis of deep neural networks via synaptic cluster-driven
genetic encoding. In NIPS Workshop, 2016.
[12] S. Shin, Y. Boo, and W. Sung. Fixed-point optimization of deep neural networks with adaptive step size
retraining. arXiv:1702.08171, 2017.
[13] P. Yin, S. Zhang, J. Xin, and Y. Qi. Training ternary neural networks with exact proximal operator.
arXiv:1612.06052, 2017.
4
| 2cs.AI
|
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
68
Timing and Code Size Optimization on
Achieving Full Parallelism in Uniform Nested
Loops
Y. Elloumi, M.Akil and M.H. Bedoui
Abstract- Multidimensional Retiming is one of the most important optimization techniques to improve timing parameters of nested
loops. It consists in exploring the iterative and recursive structures of loops to redistribute computation nodes on cycle periods, and thus
to achieve full parallelism. However, this technique introduces a large overhead in a loop generation due to the loop transformation. The
provided solutions are generally characterized by an important cycle number and a gr eat code size. It represents the most limiting
factors while implementing them in embedded systems.
In this paper, we present a new Multidimensional Retiming technique, called “Optimal Multidimensional Retiming” (OMDR). It reveals
the timing and dat a dependency characteristics of nodes, to minimize the overhead. The experimental results show that the average
improvement on t he execution time of the nested loops by our technique is 19.31% compared to the experiments provided by an
existent Multidimensional Retiming Technique. The average code size is reduced by 43.53% compared to previous experiments.
Index Terms— Graph Theory, Multidimensional Applications, Optimization, Parallelism and concurrency.
—————————— ——————————
1 INTRODUCTION
T
he design of real time systems should respect many
constraints such as the execution time and code size,
which require using optimization techniques. The
Retiming presents one of these techniques which can be used
to add and remove registers in order to provide a more
efficient circuit [1].
The increased complexity of such application leads to the
frequent use of a nested iterative and recursive loops. Such
applications can be modeled as “Multidimensional Data
Flow Graph” (MDFG). The standard software pipelining
techniques can only be used to optimize a one-dimensional
loop. When they are applied to optimize nested loops, the
performance improvement is very limited [8].
Other works are proposed to offer an optimization
technique taking advantage of the multiple nested loops,
which is called “Multi-Dimensional Retiming” (MDR). It
aims achieving full parallelism of uniform nested loops. It
consists in scheduling the MDFG with the minimum cycle
period and modifying the execution order of nodes, such as
each one is executed in a separate cycle.
But, Achieving full parallelism requires adding a large
code overhead [2],[3]. It dramatically increases the whole
code size of the provided MDFG. Furthermore, this extra
code requires a significant cycle number to be executed
outside the loop body. Thus, the provided solution does not
allow achieving an application with an adequate execution
time and a code size. It represents a limiting factor to
implement the provided MDFG in a real-time embedded
system.
We propose in this paper a new technique of MDR,
called “Optimal Multidimensional Retiming”. It allows
————————————————
• Y. Elloumi is a P.H.D. student in Paris-Est University, 93162 Noisy le
Grand Cedex, France.
• M. Akil is Professor at computer science department, ESIEE, Paris, 93162
Noisy le Grand Cedex, France.
• M.H. Bedoui is Professor in faculty of medicine of Monastir, University of
Monastir, 5019, Monastir, Tunisia,
redistributing optimally the nodes on cycle periods, while
scheduling the MDFG with the minimal cycle period. This
technique significantly allows optimizing the number of
period cycles (notably the execution time) and the code size,
by exploring the execution time and data dependency
between nodes belonging to the MDFG. Thus, it provides
enhanced solutions, compared to the existent techniques.
The rest of the paper is organized as follow. In section 2,
we give an overview of MDFG formalism. In section 3, we
list the existent MDR techniques and their constraints and
limits. In section 4, we present the theory of the “Optimal
Multidimensional Retiming” technique by describing the
principles and basics concepts, and proposing the
correspondent algorithms. Experimental results are
presented in section 5, followed by concluding remarks in
section 6.
2 MULTIDIMENSIONAL DATA FLOW GRAPH
The Multidimensional Data Flow Graph (MDFG) is an
extension of the classic data flow graph that allows to
represent a nested iterative and recursive structures. It is
modeled by a node-weighted and edge-weighted directed
graph such as 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡), where V is the set of
computation nodes, 𝐸𝐶 𝑉 × 𝑉 is the set of edges, and 𝑑(𝑒𝑖 ) is
a function from E to Z n , representing the multidimensional
delay between two nodes, where n is the number of
dimensions (loops), and 𝑡(𝑣𝑗 ) is a function from V to the
positive integers, representing the computation time of the
node 𝑣𝑗 .
For a MDFG with n dimensions, each edge 𝑒 ∶ 𝑣𝑖 → 𝑣𝑗 is
characterized by a delay where 𝑑(𝑒) = (𝑐1 , 𝑐2 , … , 𝑐𝑛 ). The
value 𝑐𝑘 represents the difference between the execution
iteration of 𝑣𝑗 and the execution iteration of 𝑣𝑖 of the loop 𝑘.
We show in Fig.1.a a two-dimensional Data Flow Graph
(2DFG) corresponding the Wave Digital Filter described in
Algorithm 1, which is composed of two nested loops. The
execution of each node in V exactly represents one iteration,
which is the execution of one instance of the loop body. Each
edge belonging to the 2DFG shown in Fig.1.a is labeled by a
delay 𝑑(𝑒) = (𝑑. 𝑥, 𝑑. 𝑦). Both terms « 𝑑. 𝑥 » and « 𝑑. 𝑦 »
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
represent the difference between the iteration number
executing 𝑣𝑗 and the iteration number executing 𝑣𝑖 , in the
outermost loop as well as in the innermost loop [9].
For an edge 𝑒 : 𝑣𝑖 → 𝑣𝑗 , the delay 𝑑(𝑒) = (0, 𝑥) consists
in the execution of 𝑣𝑖 and 𝑣𝑗 in the same iteration of the
outermost loop. For the innermost loop, if the node vi is
executed in the iteration 𝑘, the node 𝑣𝑗 is executed in the
ALGORITHM 1
WAVE DIGITAL FILTER
0: For i from 0 to m do
1: For j from 0 to n do
2:
D(i,j)= B(i -1 , j+1) × C(i -1 , j-1)
3:
A(i,j)= A(i,j) × 5
4:
B(i,j)= A(i,j) + 1
5:
C(i,j)= A(i,j) + 2
6: End for
7: End for
G illustrates the dependencies between copies of nodes
representing the MDFG G, such as the CDG shown in Fig.2.b
which corresponds to the MDFG G in Fig.1.a. A node in CDG
is a computational cell that represents a complete iteration.
The CDG of a nested loop is bounded by the loop indexes.
A schedule vector s defines a sequence of execution in
the cell dependency graph. The CDG shown in Fig.2.b, can
be executed by a row-wise execution sequence, i.e., the
schedule vector 𝑠 = (1,0). A legal MDFG 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡) is
realizable if there exists a schedule vector s for the cell
dependency graph in respect to G; i.e., 𝑠 × 𝑑(𝑒) ≥ 0, 𝑒 ∈ 𝐸,
and no cycle exists in its corresponding CDG. Note that delay
vectors (0,1) and (0, −1) are both legal in respect to the
schedule vector 𝑠 = (1,0), but they create cycles in cell
dependency graph.
3 MULTIDIMENSIONAL RETIMING
iteration(𝑘 − 𝑥). An edge with zero delay 𝑑(𝑒) = (0,0)
represents a data dependency in the same iteration, such as
the edges 𝐷 → 𝐴, 𝐴 → 𝐵 and 𝐴 → 𝐶 as shown in Fig. 1.a.
𝑒
69
3.1 Principles
The retiming technique consists in redistributing delays in
We use the notation 𝑣𝑖 → 𝑣𝑗 to indicate that 𝑒 is an edge
𝑝
from 𝑣𝑖 node to 𝑣𝑗 node, and 𝑣𝑖 ⇒ 𝑣𝑗 to mean that 𝑝 is a path
from 𝑣𝑖 node to 𝑣𝑗 node. The delay vector of a path 𝑝: 𝑣𝑖
𝑒𝑚
𝑒𝑚+1
𝑒𝑛
�� 𝑣𝑖+1 �⎯⎯� … → 𝑣𝑗 is 𝑑(𝑝) = ∑𝑘=𝑛
and the total
𝑘=𝑚 𝑑(𝑒𝑘 )
𝑘=𝑗
computation time of a path 𝑝 is 𝑡(𝑝) = ∑𝑘=𝑖 𝑡(𝑣𝑘 ). The period
during which all computation nodes in iteration are
executed, according to existing data dependencies and
without resource constraints, is called a cycle period. The
cycle period C(G) of an MDFG is the maximum computation
time among paths that have a zero delay. For example,
assuming that each node is executed in one time unit
𝑡(𝐴) = 𝑡(𝐵) = 𝑡(𝐶) = 𝑡(𝐷) = 1, the MDFG of Fig.1.a
has C(G) = 3. It can be measured through the paths 𝑝: 𝐷 →
𝐴 → 𝐵 or 𝑝: 𝐷 → 𝐴 → 𝐶, as shown in the iteration scheduling
illustrated in Fig.1.b. Each set of nodes belonging to the same
iteration are modeled by a different motif.
The execution pattern of a nested loop can be illustrated
by iteration space as shown in Fig.2.a. Each cell in the
iteration space is a copy of the MDFG. The marked cell,
labeled by (0,0), is the first iteration to be executed. This
…….
B
C
A
e2
(0,0)
B
C
(1,-1)
e4
e5
(1,1)
(a)
D
B
C
A
D
B
(0,0)
e3
D
T
im
e
A
(0,0)
e1
C
A
D
(b)
Fig.1. (a) MDFG of Wave Digital Filter; (b) Iteration scheduling of the
MDFG in Fig.1.a.
graph is transformed on an acyclic graph, called cell
dependency graph (CDG), allowing to show clearly the
execution sequence of a nested loop. The CDG of an MDFG
(a)
(b)
Fig.2. (a) Iteration space of the MDFG in Fig.1; (b) The cell
dependency graph.
the graph. This technique can be applied on a data flow
graph to minimize the cycle period in a polynomial time. The
delays are moved around in the graph in the following way:
a delay unit is drawn from each of the incoming edges of 𝑣,
and then added to each of the outgoing edges of 𝑣, or vice
versa [1]. In the case of MDFG, it consists in redistributing
the execution of nodes on the iterations. The retiming vector
𝑟(𝑢) of a node 𝑢 ∈ 𝐺 represents the offset between the
original iteration containing 𝑢, and the one after retiming.
Note that the retiming technique preserves data
dependencies of the original MDFG. Therefore, we have
𝑑𝑟 (𝑒) = 𝑑(𝑒) + 𝑟(𝑢) − 𝑟(𝑣) for every edge and 𝑑𝑟 (𝑙) = 𝑑(𝑙)
for every cycle 𝑙 ∈ 𝐺. After retiming, the execution of the
node u in the iteration i is moved to the iteration 𝑖 − 𝑟(𝑢).
We show in Fig.3.a the MDFG 𝐺𝑟 = (𝑉, 𝐸, 𝑑𝑟 , 𝑡) of the
wave digital filter after applying the retiming function
𝑟(𝐷) = (0,1). When a delay is pushed through node 𝐷 to its
outgoing edge as shown in Fig.3.a, the actual effect on the
Algorithm 2 of the new MDFG is that the 𝑖 th copy of 𝐷 is
shifted up and is executed with (𝑖 − (0,1)) th copy of nodes 𝐴,
𝐵, and 𝐶. The original zero-delay edge 𝐷 → 𝐴 in Fig.1.a now
has a delay (0,1) after retiming as shown in Fig.3.a. Node 𝐷
in the new loop body has not any data dependency with
other nodes executed in the same cycle. So, node 𝐷 can be
executed in parallel to node 𝐴, as shown in the iteration
scheduling of Fig.3.b. Thus, the cycle period is reduced
from three to two time units.
P
P
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
70
In fact, every retiming operation corresponds to a
software pipelining operation. When 𝑟(𝑢) delay units are
pushed through a node u, every copy of this node is moved
by 𝑟(𝑢) iterations. Hence, a new iteration consists in
redistributing the execution nodes into different iterations.
ALGORITHM 2
WAVE DIGITAL FILTER AFTER RETIMING BY THE FUNCTION
R(D)=(0,1)
0: For i from 0 to m do
1: D(i,0) = B(i-1 , 1) × C(i-1 , -1)
2: For j from 0 to n-1 do
3:
D(i,j+1) = B(i-1 , j+2) × C(i-1 , j)
4:
A(i,j) = D(i,j) × 5
5:
B(i,j) = A(i,j) + 1
6:
C(i,j) = A(i,j) + 2
7: End for
8: A(i,n) = D(i,n) × 5
9: B(i,n) = A(i,n) + 1
10: C(i,n) = A(i,n) + 2
11:End for
(a)
Fig.4. (a) The iteration space the retimed MDFG in Fig.3; (b) The cell
dependency graph.
Hence, the MDR technique aims to transform a realizable
MDFG G on MDFG 𝐺𝑟 in a way that 𝐺𝑟 is still realizable.
Using such concepts, the basic conditions for legal
multidimensional retiming are defined in the following
e2
(0,0)
C
(1,-1)
e4
e5
(1,1)
(0,1)
e3
(a)
D
C
(1,-3)
e4
e5
(1,-1)
D
B
C
A
D
B
C
A
D
B
C
A
D
A
D
D
(0,1)
e3
(a)
Prologue
(b)
Fig.5. (a) Fully parallelized graph of the MDFG in Fig.1; (b) Iteration
scheduling of the MDFG in Fig.5.a
lemma [2].
B
C
A
D
Lemma 1. Let 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡) be a realizable MDFG, r a
multidimensional retiming, and s a schedule vector for the
retimed graph 𝐺 𝑟 = (𝑉, 𝐸, 𝑑𝑟 , 𝑡), then
T
im
e
(0,0)
e1
e2
(0,1)
B
…….
…….
r(D)=(0,1)
A
A
(0,1)
e1
r(D)=(0,2)
r(A)=(0,1)
T
im
e
Some nodes are shifted out of the loop body to provide the
necessary data for the iterative process, which is called
prologue. Correspondingly, some nodes will be executed
after the loop body to complete the process, which is called
epilogue.
Using MD retiming function r, we can trace the pipelined
nodes and also measure the size of the prologue and
epilogue. For node v with retiming 𝑟(𝑣) = (𝑖, 𝑗), there are i
copies of node v appearing in the prologue of outer loop, and
j copies of node v in the prologue of the innermost loop. The
B
(b)
B
C
A
D
B
C
A
D
1.
2.
D
Prologue
(b)
Fig.3. (a) MDFG of algorithm 2; (b) Iteration scheduling of the MDFG in
Fig.3.a
number of copies of a node in the epilogue can also be
derived in a similar way. The iteration space of the retimed
MDFG shown in Fig.4.a with retiming 𝑟(𝐷) = (0,1) clearly
shows that one copy of node D is pushed out of the loop
body on j-dimension, and becomes prologue for the
innermost loop. The corresponding cell dependency graph is
shown in Fig.4.b.
It is known that an MDFG can always be fully
parallelized by applying successively the MD retiming
functions 𝑟(𝐷) = (0,2) and 𝑟(𝐴) = (0,1), which is illustrated
in Fig.5.a. We note that the retimed MDFG has non-zerodelay on each edge. It implies that the nodes belonging to the
same iteration in the original loop body are distributed into
three cycle periods. The MDFG is then scheduled with the
minimal cycle period equal to one time unit, as schematized
the iteration scheduling of Fig.5.b.
To achieve a realizable MDFG after retiming, the legality
condition,𝑠 × 𝑑(𝑒) ≥ 0, has to be satisfying, and there should
not exist any cycle in the cell dependency graph of the
MDFG.
3.
4.
for any path , we have 𝑑𝑟 (𝑝) = 𝑑(𝑝) + 𝑟(𝑢) − 𝑟(𝑣)
for any cycle 𝑙 ∈ 𝐺 we have 𝑑𝑟 (𝑙) = 𝑑(𝑙)
𝑒
for any edge 𝑢 → 𝑣, 𝑑𝑟 (𝑒) × 𝑠 ≥ 0
there is no cycle in the DG equivalent to the MDFG G.
The selection of a legal multidimensional Retiming
function is based on the edge delay of the MDFG. The
approach proposed in [2],[3] consists in defining a
scheduling subspace S for a realizable MDFG 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡).
It represents the space region where there exist schedule
vectors that realize 𝐺; i.e., if schedule 𝑠 ∈ 𝑆 then 𝑠 × 𝑑(𝑒) ≥ 0
for any 𝑒 ∈ 𝐸. In fact, the multidimensional retiming
technique means to decrease zero-delay edges. Thus, a
strictly positive scheduling subspace 𝑠 + is the set al all
vectors 𝑠 ∈ 𝑆 where 𝑠 × 𝑑(𝑒) > 0 for every 𝑑(𝑒) ≠ (0,0, … ,0).
The method of predicting a legal multidimensional retiming
is introduced in the next theorem.
Theorem 1. let 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡) be a realizable MDFG, 𝑠 + a
strictly positive scheduling sub-space of G, s a scheduling vector
in 𝑠 + , and 𝑢 ∈ 𝑉 a node with all incoming edge having nonzero
delay. A legal MD retiming r of u is any vector orthogonal to s.
3.2 Multidimensional Retiming techniques
We describe in this section the existent multidimensional
retiming techniques. They are characterized by achieving full
parallelism by providing the MDFG with no zero-delay edge
[2],[3],[7].
a) Incremental Multidimensional Retiming
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
This technique is based on selecting a set of nodes that
can be retimed by the same multidimensional retiming
function, as described in the following corollary.
Corollary 1. Given a realizable MDFG 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡), 𝑠 + a
strictly positive scheduling sub-space of G, s a scheduling vector
ALGORITHM 5
CHAINED MULTIDIMENSIONAL RETIMING
Input : a realizable MDFG G =(V,E,d,t),
Output : a realizable MDFG Gr =(V,E,dr,t) without d(e)=(0,0,
… ,0)
0: Begin
1: Find a legal MDR function r as describedin steps 2 and 3
in algorithm 3
2: Provide the multi -chain graph and the maximal length
of chain K, as indicated in algorithm 4
3: For each node v with label i do
4:
Apply the MDR function (k-i)× (r)
5: End for
6: End
71
nodes assigned by 𝐴𝑗 . Fig.7 illustrates the full parallelized
MDFG of the IIR filter after applying the Incremental
Multidimensional Retiming.
The steps of algorithm 3 are repeated four times, where
in each one a different MD retiming function is applied. The
fully parallelized MDFG is showed in Fig.7 where all edges
are non-zero delay.
a) Chained Multidimensional Retiming [2],[3]
This technique allows obtaining the full parallelism
solution by defining just one MDR function. It is based on the
following corollary.
in 𝑠 + , and an MD retiming function r orthogonal to s, if a set
𝑋𝐶𝑉has all incoming edges nonzero, then 𝑟(𝑋) is a legal MD
retiming.
Thus, this technique consists in defining a schedule
vector s as described in definition 1, and chooses an MDR
function orthogonal to s. This chosen function is applied to
each node respecting the previous corollary. Those steps are
ALGORITHM 3
INCREMENTAL MULTIDIMENSIONAL RETIMING
Input : a realizable MDFG G =(V,E,d,t)
Output : a realizable MDFG Gr=(V,E,dr ,t) without d(e) =
(0,0, … ,0)
0: Begin
1: While exist zero-delay edge in the graph Do
2: Find a scheduling vector s=(s.x,s.y), that s.x+s.y is
minimum
3: Choose a MDR function
4: Apply the selected MDR function to any nodes that
has all incoming edges with nonzero delays and at
least one outgoing edge with zero delay
5: End while
6: End
repeated incrementally, until all zero-delay edges are
transformed, as described in algorithm 3.
We apply the algorithm above to the 2DFG of Infinite
Fig.6. MDFG of IIR Filter.
Impulse Response filter (IIR) that is illustrated in Fig.6. It is
composed by multiplier nodes assigned by 𝑀𝑖 and adder
Fig.7. IIR Filter MDFG after Incremental MDR.
Corollary 2. [2] Given 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡), 𝑆 + a strict positive
scheduling subspace for G,s a scheduling vector 𝑆 + , a MD
retiming function orthogonal to s, a set 𝑋𝐶𝑉 which all incoming
edges nonzero, and an integer value 𝑘 > 1, then (𝑘 × 𝑟) (𝑋) is
a legal MD retiming.
Thus, it applies the MD retiming to successive nodes in a
path where each node has a retiming function multiple of the
selected retiming value smaller than its predecessor nodes.
ALGORITHM 4
MULTI-CHAIN GRAPH CONSTRUCTION
Input : a realizable MDFG G =(V,E,d,t)
Output : Labeled Multi -chain Graph
0: Begin
1: Remove all non-zero delay edges from the MDFG
2: For each chainCH do
3: Compute the length L of CH
4: For each node starting from the last to the first do
5:
Labeled the node by L
6:
L=L-1
7: End For
8: End For
9: End
This technique starts by transforming the MDFG on
Multi-chain Graph as described in Algorithm 4. Each chain
represents a node succession where all interconnected edges
between them are zero-delay. Each node is labeled by a level
whose the value is greater than its predecessor node and
smaller than its successor ones.
In the case of MDFG of IIR filter, the red integers above each
node of Fig.8 represent the level values that are labeled after
executing algorithm 4. Therefore, the multi-chain maximum
length of 𝐺 is 4.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
So, The technique proceeds to retime all the labeled
nodes by a MD retiming function (𝑘 − 𝑖) × (𝑟), as described
in algorithm 5.
We present in Fig.8 the full parallelized graph of the IIR
filter after applying the chained multidimensional retiming.
The algorithm starts by finding the MDR function which is
equal to (1, −1). We note that all zero delay edges in the
original MDFG are assigned by a delay vector equal to the
MDR function.
Until now, no research works have been interested in
comparing results provided by the techniques described
above. The random choice of the scheduling vector does not
allow defining the technique providing the optimal solution.
However, based on their approaches, Chained MDR is
generally more performing than the Incremental MDR. The
first one consists in defining just one scheduling vector,
which is executed in 𝑂(|𝐸|); while the second requires
defining scheduling vector for each iteration of algorithm 4,
which is executed in 𝑂(|𝑉|).
b) SPINE (Software PIpelining of NEsted loops)
Multidimensional Retiming
This technique tries to provide a more optimal MDFG in
terms of execution time and code size than those described
above. It proceeds to remove all delays such as (0, 𝑘) by
merging them in a delay such as (𝑖, 𝑗). This modification is
applicable only if the MDFG contains at least one edge with a
72
loop. Since this limit is directly associated to the size of the
iteration space, it is called spatial constraint and it is formally
defined as follows.
ALGORITHM 6
SPINE-FULL ALGORITHM
Input : a realizable MDFG G =(V,E,d,t)
Output : a realizable MDFG
Gr =(V,E,dr,t) fully
parallelized with minimum code size
1: Begin
2: If s=(1,0) is legal then
3:
Apply s=(1,0)
4: Else If s=(0,1) is legal and d(e)×(0,1)>=0 then
5:
Apply s=(0,1)
6: Else If s=(1,1) is legal and d(e)×(1,1)>=0 then
7:
Apply s=(1,1)
8: Else
9: Choose a legal scheduling vector s such as d(e)×s>=0,
for any edge and | sx| +| sy| is minimal
10: End
Definition 1. Let a MDFG G contains k-level nested loop N,
controlled by the set of indices 𝐼 = {𝑖0 , 𝑖1 , … , 𝑖𝑘 }, whose values
vary, in unitary increments, in the range 𝐿 = {𝑙0 , 𝑙1 , … , 𝑙𝑘 } to
𝑈 = {𝑢0 , 𝑢1 , … , 𝑢𝑘 } where L is the set of lower boundaries for
the indices and U is the set of maximum values, such as
𝑙𝑗 ≤ 𝑖𝑗 ≤ 𝑢𝑗 , then the spatial constraint Sc of the loop is defined
as:
𝑆𝑐 = [(𝑢0 − 𝑙0 + 1), (𝑢1 − 𝑙1 + 1), … , (𝑢𝑘 − 𝑙𝑘 + 1)]
This definition allows establishing the relation between
the maximum retiming operation and the spatial constraint
according to the following lemma.
Lemma 2. Given a k-level loop N with spatial constraint 𝑆𝑐 =
[𝑠0 , 𝑠1 , … , 𝑠𝑘 ]. The multi-dimensional retiming technique will be
able to achieve full parallelism of the loop body instructions if
the maximum retiming vector r applied to any node u, 𝑟(𝑢) =
(𝑟0 , 𝑟1 , … , 𝑟𝑘 ) satisfies the following condition:
𝑟𝑗 < 𝑠𝑗 𝑠𝑢𝑐ℎ 𝑎𝑠 0 ≤ 𝑗 ≤ 𝑘
Fig.8. IIR Filter MDFG after Chained MDR.
delay equal to (𝑖, 𝑗) such as 𝑖 > 0.
It consists in finding a scheduling vector 𝑠 and a retiming
function 𝑟 orthogonal to 𝑠, as described in algorithm 6, to
provide a minimal overhead.
3.3 Multidimensional Retiming Constraints [5]
We describe in this paragraph the algorithmic constraints
which must be taken into account to achieve full parallelism.
These constraints come from the ratio between loop bounds
and a number of MDR functions to apply.
For example, consider the multi-dimensional data flow
graph in Fig.1; it is easy to verify that if 𝑑(𝑒4) = (0, 𝑘) where
𝑘 < 3, retiming 𝐷 and 𝐴 by some vector (0, 𝑝) will not
satisfy the goal of re-distributing the delays among all edges
in the graph. The same will happen if 𝑑(𝑒5) = (𝑚, 0)
where 𝑚 < 3 for any retiming vectors of the form (𝑞, 0).
Thus, if the loop has only one occurrence, i.e., the loop
boundaries are both 1, then no parallelism can be obtained.
This last constraint is equally applicable to a software or
hardware implementation of the retimed loop.
This study begins by evaluating the constraints imposed
by the limitation on the number of iterations comprising the
3.4 Limitations of existing techniques
We have shown that nested loops can always be fully
parallelized using MD retiming. The presented techniques of
MD retiming are a polynomial time algorithm that fully
parallelizes a given MDFG 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡) by selecting a legal
schedule vector s with 𝑠 × 𝑑(𝑒) > 0, 𝑒 ∈ 𝐸, and a retiming
vector 𝑟 where 𝑟 is orthogonal to 𝑠. Each MD retiming
techniques presented above shows that the selected 𝑠 is a
legal schedule vector for the retimed graph where 𝑑𝑟 (𝑒) ≠
𝑑(𝑒) ± 𝑘. 𝑟.
However, Multidimensional retiming techniques imply a
large overhead of the generated code. It is caused by several
aspects of loop transformation. First, the code size is
increased because of the large code sections of the prologue
and epilogue produced in all the loop dimensions. Second,
the computation of the new loop bounds and loop indexes
need to be recomputed [10].
Moreover, the execution of the prologue and epilogue
section is not fully parallel, which requires a considerable
period cycle number compared to that required by the loop
body. Those disadvantages are aggravated in terms of the
retiming vector value, and the number of the
multidimensional retiming function.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
Each Multidimensional Retiming techniques have a
specific approach to choose the retiming function in order to
decrease the overhead size. Chained and incremental
Multidimensional retiming techniques proceed to chose a
scheduling vector 𝑠 = (𝑠. 𝑥, 𝑠. 𝑦), where 𝑠. 𝑥 + 𝑠. 𝑦 is
minimum. The SPINE technique tries to modify the MDFG
with the intention of applying an MDR function that skews
the minimum column-wise or/and row-wise. Despite
providing an optimal solution, it is reliable only in the
particular case of MDFG. In opposite cases, it applies the
same approach than other techniques.
However, all existent techniques consist in retiming each
node of the MDFG having an out-coming edge with zerodelay: if a path p as 𝑑(𝑝) = (0,0, … ,0) is composed by n
nodes, any technique applies (𝑛 − 1) MDR function to
achieve full parallelism. But, overhead consequences are
sudden after applying each MDR function: the more the
number of MDR function increases, the more the
consequences are dramatic.
As a result, the provided solution becomes very
complicated and not sufficient to be implemented in
embedded systems. Therefore, the existing MD retiming
techniques, although achieving full parallelism, are not
suitable for software nested loops.
4 THEORY OF OPTIMAL MULTIDIMENSIONAL RETIMING
In this section, we present the theoretical foundation of our
proposal MDR technique “Optimal Multidimensional
Retiming”. It aims at minimizing MDR functions by
exploring execution times and data dependency of nodes,
while achieving full parallelism.
3.1 Principle
Multidimensional retiming techniques allow scheduling the
MDFG with a minimal cycle period. For any path p 𝑝: 𝑣𝑖
𝑒𝑚+1
body is executed in 5 cycles. We illustrate in Fig.9 the static
schedule of such an iteration after achieving full parallelism
by the chained multidimensional retiming which is retimed
by the function 𝑟 = (−1,1). The nodes belonging to the same
iteration are modeled by gray circles.
Each gray node has not any data dependency with any
other node executed in the same cycle period. However,
these scheduling shows that provided data by some nodes
are not consumed immediately. For example, nodes 𝐴5 and
𝐴7 are executed in just one time unit and their provided
values are consumed two times units later. We conclude that
cycle periods are not exploited optimally: (more than 66% of
the cycle periods (𝑖 − 2, 𝑗 + 2), (𝑖 − 3, 𝑗 + 3) and (𝑖 − 4, 𝑗 + 4)
are not used). However, they allow executing more than one
node, due to the difference between execution time.
Let us try to execute nodes 𝐴5 and 𝐴7 in the same cycle
period as nodes 𝐴1, 𝐴2 and 𝐴6, as schematized in Fig.10. The
correspond MDFG with new delay values is illustrated in
Fig.11. This transformation results in a legal MDFG that
respects all conditions of lemma 1. Furthermore, it still keeps
a fully parallelized execution, while preserving data
dependency for the whole application. Compared to Fig.8,
this transformation can be considered as decreasing the
number of MDR functions by depriving nodes 𝐴5 and 𝐴7 to
be retimed.
In fact, minimizing the number of MDR functions
implies decreasing the overhead of the generated code. It
consists in reducing the correspondent prologue, epilogue
M5
M6
M7
M8
M1
M2
M3
A3
M5
M6
M7
M8
M1
M2
M3
Theorem 2. Let 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡) a MDFG, the minimal value of
cycle 𝑐𝑚𝑖𝑛 for the G graph is:
𝑐𝑚𝑖𝑛 = 𝑚𝑎𝑥{𝑡(𝑣𝑖 ), 𝑣𝑖 ∈ 𝑉}
We choose to model the MDFG of IIR filter with different
execution time of nodes such as of 𝑡(𝑀𝑖 ) = 3 and 𝑡(𝐴𝑗 ) = 1.
The minimal cycle 𝑐𝑚𝑖𝑛 of this graph is equal to the execution
time of the multiplication node.
Applying any MDR techniques to the IIR filter results in
the fact that each iteration belonging to the original loop
A4
A1
A2
A6
A5
A7
A8
Cycle
(i-3 , j+3)
M4
𝑒𝑛
�� 𝑣𝑖+1 �⎯⎯� … → 𝑣𝑗 of MDFG, they proceed to execute each
node 𝑣𝑘 where 𝑖 ≤ 𝑘 ≤ 𝑗 in a period cycle separately. These
approaches can be generalized into general-time cases [2] [4].
In fact, the computation nodes belonging to a data flow
graph have not generally the same execution times. These
depend on the kind of operation to be done; for example, a
multiplication node needs usually more clock period than an
adder node.
In this case, the minimal cycle period should be fixed
differently. A cycle period represents a time interval leading
to execute computation nodes. The minimal value of a period
cycle can be defined as the smaller time interval that allows
executing any node belonging to the MDFG. Thus, the
minimal cycle period should be equal to the maximal
execution time of node, as described in theorem 2 [1].
Cycle
(i-4 , j+4)
M4
A3
A4
A1
A2
A6
A5
A7
A8
T
im
e
𝑒𝑚
73
M5
M6
M7
M8
M1
M2
M3
Cycle
(i-2 , j+2)
M4
A3
M5
M6
M7
M8
M1
M2
M3
M6
M7
M8
M1
M2
M3
A1
A2
A6
A5
A7
A8
Cycle
(i-1 , j+1)
M4
A3
M5
A4
A4
A1
A2
A6
A5
A7
A8
Cycle
(i , j)
M4
A3
A4
A1
A2
A6
A5
A7
A8
Fig.9. Iteration scheduling after chained MDR.
and the loop bound and index instructions. Moreover, it
results in decreasing the number of cycle periods required to
execute any iteration from 5 to 4, while respecting a fully
parallelized execution. This minimization of cycle periods
implies a similar minimization on the execution time of the
whole application. Thus, this minimization of MDR functions
leads to improve the performance of the provided full
parallel solution.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
Our technique proceeds to share the MDFG on zerodelay paths. It leads to maximize the node number in such a
A8
M5
M6
M7
M8
M1
M2
M3
A3
A4
A2
A1
Cycle
(i-2 , j+2)
A7
A5
M4
A6
A8
M5
M6
M7
M8
M1
M2
M3
Cycle
(i-1 , j+1)
A7
A5
M4
T
im
e
Such modification can be defined as applying the MDR
to path p that can be composed of several nodes where 𝑡(𝑝)
is smaller than 𝑐𝑚𝑖𝑛 . The more the paths contain nodes, the
more the MDR function number decreases. Thus, we propose
a new MDR approach which consists in applying MDR
function to a path of nodes, which can be executed in the
same period cycle, instead of applying the MDR to each node
separately. Our approach is based on selecting the path with
the maximal nodes to achieve full parallelism with a minimal
number of MDR functions. We start by computing the
74
A3
A4
A2
A1
A6
A8
M5
M6
M7
M8
M1
M2
M3
A5
M4
A3
M5
M6
M7
M8
M1
M2
M3
A4
A5
M4
A4
A6
A2
A1
A6
M5
M6
M7
M8
M1
M2
M3
M4
A5
A8
A3
A4
A2
A1
Cycle
(i-1 , j+1)
A7
A6
M6
M7
M8
M1
M2
M3
M4
A5
A8
A3
A4
A1
Cycle
(i , j)
A7
A2
A6
A8
Fig.10. Iteration scheduling after collecting nodes in cycle (i-2, j+2).
Fig.11. MDFG of Fig.10.
minimal period cycle 𝑐𝑚𝑖𝑛 , extracting the paths with maximal
nodes from the MDFG while keeping their execution in 𝑐𝑚𝑖𝑛 ,
and applying the MDR function to the extracted paths.
3.2 Basics concepts
Our technique consists in retiming a path of nodes that can
be executed in the same cycle period. It means that those
nodes are executed in the same iteration of the original loop
body; i.e., all edges belonging to the path have zero-delay.
We call such a kind of path “zero-delay path”, which is
indicated in theorem 3.
𝑒𝑚+1
M1
M2
M3
A4
A1
Cycle
(i , j)
A7
A5
M4
A2
A6
Definition 2. Let 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡), 𝑐𝑚𝑖𝑛 the minimal period cycle
of G. The optimal multidimensional retiming consists in
retiming a set of path p follows :
2.
3.
𝑒𝑛
Theorem 3. Let 𝑝: 𝑣𝑖 �� 𝑣𝑖+1 �⎯⎯� … → 𝑣𝑗 . p is zero-delay 𝑑(𝑝) =
(0, … ,0),if and only if :
∀ 𝑒𝑙 : 𝑣𝑘 → 𝑣𝑘+1 , 𝑑(𝑒𝑙 ) = (0, … ,0); 𝑠𝑢𝑐ℎ 𝑎𝑠 𝑒𝑙 ∈ 𝐸, 𝑣𝑘 ∈ 𝑉, 𝑘
∈ [𝑖, (𝑗 − 1)], 𝑙 ∈ [𝑚, 𝑛]
𝑚𝑎𝑥 (𝑐𝑎𝑟𝑑(𝑝))
𝑑(𝑝) = (0, … ,0)
𝑡(𝑝) ≤ 𝑐𝑚𝑖𝑛
𝑒𝑚
𝑒𝑚+1
𝑒𝑛
Where 𝑝: 𝑣𝑖 �� 𝑣𝑖+1 �⎯⎯� … → 𝑣𝑗 as 𝑣𝑘 ∈ 𝑉 ∀ 𝑘 ∈ [𝑖, 𝑗]
and card(p) is the number of nodes belonging to p such as
𝑐𝑎𝑟𝑑(𝑝) = 𝑗 − 𝑖 + 1.
To preserve data dependency of MDFG, each p path
should not contain any cycle in a way that no cycle between
vp and vm where ∀ 𝑝, 𝑚 ∈ [𝑖, 𝑗] for each selected path
𝑒𝑚
𝑒𝑚
M8
path, while being executed in the minimal cycle period. We
define our idea of optimal multidimensional retiming as
described in definition 2.
1.
M5
M7
Fig.15. Iteration scheduling of MDFG shown in Fig.14.
Cycle
(i-2 , j+2)
A7
M6
A3
A8
T
im
e
A3
Cycle
(i-3 , j+3)
A7
A2
A1
M5
𝑒𝑚+1
𝑒𝑛
Furthermore,
multidimensional
𝑝: 𝑣𝑖 �� 𝑣𝑖+1 �⎯⎯� … → 𝑣𝑗 .
retiming consists in executing several paths in the same cycle
period. Thus, such a path should not have any cycle between
nodes that they belong to.
Our technique is based on retiming a zero-delay path. It
means that all edges belonging to this path preserve the same
delay (zero-delay). Only the delay of in-coming and outcoming edges of the whole path are changed. Thus, it means
retiming the last node that it belongs to.
Referring to [2] and [3], the multidimensional retiming
function is defined from edges having a non-zero delay of
the MDFG. However, there is not constraint that requires
applying the MDR function to nodes with non-zero incoming edges. In the case of zero delay paths, it can be
applied to any node belonging to them, as defined in
theorem 4.
Theorem
4.
Let 𝐺 = (𝑉, 𝐸, 𝑑, 𝑡), and a zero delay path 𝑝: 𝑣𝑖
𝑒(0,…,0)
𝑒(0,..,0)
�⎯⎯⎯� … �⎯⎯� 𝑣𝑗 . If r is a legal MDR function of 𝑣𝑖 ,then r is a
legal MDR function of vk, where 𝑣𝑘 ∈ 𝑃 and 𝑘 ∈ [(𝑖 + 1) , 𝑗].
Proof. a strict positive scheduling sub-space S + contains all
scheduling vectors s where 𝑑(𝑒) × 𝑠 > 0 for each 𝑑(𝑒) ≠
(0,0, … ,0), such as described in definition 2. This implies
that 𝑣𝑖 and vk have the same sub-space 𝑆 + . But, a legal MDR
𝑟 of 𝑣𝑖 is any orthogonal vector to 𝑠, where 𝑠 ∈ 𝑆 +, as
indicated in theorem 1. This means that a legal MDR 𝑟 of 𝑣𝑖
is a legal MDR of 𝑣𝑘 .
So, we provide an MDR function as indicated in theorem
2. This function is applied to the last node of the zero delay
path to retime.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
We conclude in section 3 that it is more sufficient to
select an MDR function to apply it for all nodes to be retimed
to achieve full parallelism, as described in corollary 1. Thus,
we proceed by selecting every last node of a zero-delay path
that will be retimed. Afterwards, we find a legal MDR
function to be applied successively to the selected nodes
while respecting data dependency of the MDFG.
ALGORITHM 8
OPTIMAL MULTIDIMENSIONAL RETIMING
Input : a realizable MDFG G =(V,E,d,t)
Output : a realizable MDFG Gr =(V,E,dr,t)
0: Begin
1: Find a legal MDR function r
2: Provide the LMDFG and the maximal length M , as
described in algorithm 7
3: For i from M down to 1 do
4:
Select nodesv k that are labeled by i
5:
For each v k do
6:
Apply the MDR Retiming (i × r)
7:
End for
8: End for
9: End
3.3 Path Extraction
This step is based on exploring nodes belonging to the
MDFG by testing their data dependency and execution time,
to share them onto paths. We proceed by sweeping
ALGORITHM 7
LABELED MULTIDIMENSIONAL DATA FLOW GRAPH
CONSTRUCTION
Input : a realizable MDFG G =(V,E,d,t)
Output : Labeled Multidimensional Data Flow Graph
(LMDFG), maximal length M
0: Begin
1: Compute cmin
2: Extract all nonzero delay edges from G
3: i =1
4: Add the elements (v j, t(v j)) to R, such as vj is a node
without outcoming edge
5: While R is not empty do
6: For each v j of R do
7:
Collect all predecessor ofv j
8:
For each predecessorv p of v j do
9:
If t(v p)+t(v j) <= cmin and respect data dependency
then
10:
Add (v p, t(v p)+t(v j)) to R
11:
Else
12:
Add (v p, t(v p)) to N E
13:
End If
14:
End for
15: End for
16: Label all nodes of N E by i
17: I= i+1
18: R=N E
19: End while
20: M = i
21: End
75
incrementally the MDFG from the opposite direction of edge.
For each node, we verify that it respects the previous
conditions, to execute it in the suitable cycle.
Our process consists on extracting the last node of each
zero delay path that will be retimed, and labeled by an
increasing order starting from 1. The result is illustrated in a
“Labeled Multidimensional Data Flow Graph” (LMDFG)
taken from the MDFG, as described in algorithm 7.
We proceed by exploring the MDFG in the opposite
direction of data dependency. We start by extracting the nonzero delay edges from the MDFG and collecting all nodes 𝑣𝑘
without outcoming edge in 𝑅 list. For each node belonging to
𝑅, we define all the predecessor nodes 𝑣𝑝 to verify that can be
executed in the same cycle time as nodes in 𝑅. For each 𝑣𝑝 , if
𝑡�𝑣𝑝 � + 𝑡(𝑣𝑘 ) < 𝑐𝑚𝑖𝑛 and if 𝑣𝑝 respects data dependency
conditions described above, we consider 𝑣𝑝 as a node
belonging to the path and we add it to the 𝑅 list. Else, 𝑣𝑘 is
the last node of the previous path that should be retimed,
and then we add then the node 𝑣𝑘 to 𝑁𝐸 list to label it. This
test is repeated for all predecessor nodes of 𝑅 list element.
The next step consists in labeling all the nodes of 𝑁𝐸 list by
1(first value of 𝑖), before replacing 𝑅 elements by 𝑁𝐸
elements. These steps are repeated until testing all nodes of
MDFG.
We show in Fig.12 the LMDFG of IIR filter with 𝑡(𝑀𝑖 ) =
3 and 𝑡(𝐴𝑗 ) = 1. The first iteration of algorithm 7 consists in
labeling the nodes 𝑀1, 𝑀2, 𝑀3, 𝑀4, 𝐴3 and 𝐴4 by 1. The last
iteration assigns nodes 𝑀5, 𝑀6, 𝑀7 and 𝑀8 by 2. Therefore,
the multi-chain maximum length is 2.
3.4 Optimal Multidimensional Algorithm
Our technique starts by finding a legal MDR r of MDFG, as
indicated in theorem 2. After that, it provides the LMDFG
and the maximal label by running algorithm 7. Then, it
selects nodes with maximal label, and applies the MDR
function (𝑖 × 𝑟). These steps are repeated by decreasing the
label of the selected node until achieving retiming all the
labeled nodes, as described in algorithm 8.
Fig.12. Labeled Multidimensional Data Flow Graph (LMDFG) of IIR
filter.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
As an example, we apply this algorithm to the MDFG of
IIR filter. It starts by defining an MDR function 𝑟 = (1, −1),
and providing the LMDFG shown in Fig.12. The first
iteration of the algorithm retimes the nodes that are labeled
by 2, by applying the retiming function 𝑟 = (2, −2) as
illustrated in Fig.13. The second iteration retimes the other
labeled nodes by 𝑟 = (1, −1), to provide the fully
parallelized MDFG of IIR filter as shown in Fig.14. The nodes
interconnected by a zero-delay edge are executed in the same
cycle period. It means that nodes belonging to the same
iteration in the original MDFG are executed in three cycle
periods, as illustrated in the scheduling iteration of the fully
parallelized MDFG in Fig.15.
76
TABLE 4
EVOLUTION OF THE CODE SIZE IN TERMS OF CYCLE PERIOD
MDFG
G1
G2
G3
G4
G5
G6
Optimal MDR
144
144
64
144
60
60
Code size
Chained MDR
400
400
400
400
72
72
Improve
48%
48%
84%
48%
16.66%
16.66%
5 EXPERIMENTAL RESULTS
In this section, we validate our MDR technique by comparing
its provided MDFGs to those generated by the chained MDR.
Four parameters are compared in our experimentation: the
cycle period, the number of MDR functions, the execution
time and the code size. We choose as an application the IIR
filter graph and the wave digital filter after applying the
Fettweis transformation [2], as illustrated in Fig.16. These
two graphs are both composed of addition and
multiplication nodes.
Our experiments consists in modeling each MDFG
frequently in different cases of the execution time 𝑡(𝑣𝑖 ) where
𝑣𝑖 𝜖 𝑉, whose values are indicated in Table.1. We present in
the last column the minimum cycle period whose values are
defined as theorem 2.
Fig.14. Full parallelized IIR filter after applying optimal MDR.
required to achieve full parallelism, for each MDFG, using
both techniques.
After providing the fully parallelized MDFGs, we
generate their respective algorithms, to extract their time and
code size parameters. We present in Table.3 the values of the
period cycle number and the execution time of each MDFG
For each MDFG in Table.1, we apply both chained MDR
and optimal MDR techniques. Each one is characterized by a
number of MDR functions applied to achieve full parallelism.
The chained MDR generates the same fully parallelized
MDFG with the same MDR function number, whatever the
Fig.16. MDFG of wave digital filter
illustrated in Table.1 in terms of MDR techniques. The
column “improve” presents the improvement of the
execution time of result generated by our approach
compared to those generated by the chained MDR, which
accounts for an average improvement of 19.31%.
The code size of each MDFG provided in term of both of
TABLE 1
CYCLE PERIOD IN TERMS OF NODE EXECUTION TIMES
Application
IIR Filter
Fig.13. Retiming labeled nodes by 2.
set of node execution time is.
Contrariwise, optimal MDR provides a specific MDFG
for each MDFG of Table.1, with different MDR function
numbers and retimed nodes. To guaranty a reliable
comparison, we use the same MDR function for both
techniques, which we apply the functions 𝑟 1 = (1, −1) and
𝑟2 = (0,1) respectively to the IIR filter and wave digital
filter. Table.2 illustrates the numbers of MDR functions
WD filter
MDFG
G1
G2
G3
G4
G5
G6
t(A i )
1
1
1
2
1
1
t(M j)
2
3
4
5
2
3
Cmin
2
3
4
5
2
3
the two MDR techniques are shown in Table.4. Each value
presents the instructions number of the loop body and the
overhead caused by an MDR transformation. The code size
values mention that our technique proposes an average
improvement equal to 43.53% of the code size.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 7, JULY 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
6 CONCLUSION
In this paper, we have proposed a new Multidimensional
Retiming technique to achieve full parallelism of MDFGs. It
allows providing an optimized MDFG compared to those
provided by the existent techniques. It allows minimizing
Multidimensional functions by exploring the execution times
and data dependencies between the nodes.
In the section above, we have applied our technique and
the chained MDR on different cases of MDFG. The results
have shown that our technique generates a more efficient
solution MDFG, than those generated by the chained MDR,
in terms of cycle number, execution time and code size. We
have concluded that our technique provides a more efficient
solution, which allows respecting timing and code
constraints while implementing the nested loop in embedded
TABLE 2
EVOLUTION OF MDR FUNCTION NUMBER IN TERMS OF CYCLE
PERIOD
MDFG
G1
G2
G3
G4
G5
G6
MDR function number
Chained MDR
Optimal MDR
2
4
2
4
1
4
2
4
1
2
1
2
systems.
As an optimization technique, we try in our future works
to study using our MDR technique with other optimization
approaches such as unrolling, loop fusion ... It consists in
defining the applying order and the evolution of the
performance parameter in terms of both approaches.
However, MDR techniques are based on scheduling the
MDFG with a minimum cycle period. This period cycle value
does not mean providing an MDFG with minimum
execution time. We will be interested to extending our MDR
technique to provide the adequate period cycle and a
TABLE 3
EVOLUTION OF CYCLE NUMBER AND EXECUTION TIME IN TERMS
OF CYCLE PERIOD
Execution time
Cycle number
MDFG Optimal Chained Optimal Chained
MDR
MDR
MDR
MDR
G1
258
316
516
632
G2
258
316
774
948
G3
229
316
916
1264
G4
258
316
1290
1580
G5
500
600
1000
1200
G6
500
600
1500
1800
Improve
18.35%
18.35%
27.53%
18.35%
16.66%
16.66%
scheduling approach which allows providing MDFG with
minimum execution time.
Also, in the case of real-time embedded system, the
design consists in respecting the code size constraint, which
should not be exceeded. This principle implies reducing
execution time while achieving a limit value of the code size.
Thus, based on the opposite evolution of the timing
parameters and the code size in terms of MDR functions, we
77
will be interested on proposing an optimization approach
using the MDR technique: it requires finding the set of MDR
functions that provide a retimed MDFG with a best ratio
between the mentioned parameters.
REFERENCES
[1] E. Leiserson & B. Saxe, “retiming synchronous circuitry”,
algorithmica, Springer New York, volume 6, numbers 1-6, juin 1991
[2] N.L. Passos, E.H.M. Sha, “achieving full parallelism using multidimensional retiming”, IEEE Transactions on Parallel and Distributed
Systems , Volume 7 , Issue 11 (November 1996)
[3] N.L. Passos, E.H.M. Sha, ”Full Parallelism In Uniform Nested Loops
Using Multi-Dimensional Retiming”, International Conference on
Parallel Processing, 1994.
[4] Q. Zhuge, C. Xue, M. Qiu, J. Hu and E. H.-M. Sha, "Timing
Optimization via Nest-Loop Pipelining Considering Code Size", Journal
of Microprocessors and Microsystems, Volume 32, Issue 7, October 2008.
[5] by N. L. Passos , D. C. Defoe, R. J. Bailey, R. H. Halverson, and R. P.
Simpson, “Theoretical Constraints on Multi-Dimensional Retiming
Design Techniques", Proceedings of the AeroSense-Aerospace/Defense
Sensing, Simulation and Controls, Orlando, FL, April, 2001.
[6] Qingfeng Zhuge, Chun Jason Xue, Meikang Qiu, Jingtong Hu, Edwin
H.-M. Sha, ”timing optimization via nested loop pipelinig considering
code size”, journal of microprocessors and Microsystems 32, 351-363,
2008.
[7] Qingfeng Zhuge, Zili Shao, Edwin H.-M. Sha, “timing optimization of
nested loops considering code size for DSP applications”, Proceeding of
the 2004 international conference on parallel processing, 2004.
[8]Q. Zhuge, E.H.-M. Sha, C. Chantrapornchai, CRED: code size
reduction technique and implementation for software-pipelined
applications, in Proceedings of the IEEE Workshop of Embedded System
Codesign (ESCODES), September, 2002, pp. 50–56.
[9] T. C. Denk and K. K. Parhi, "Two-Dimensional Retiming," in the IEEE
Transactions on VLSI, Vol. 7, No. 2, June, 1999, pp. 198-211.
[10] N. L. Passos and E. H.-M. Sha, "Scheduling of Uniform MultiDimensional Systems under Resource Constraints," in the IEEE
Transactions on VLSI Systems, December, 1998, Volume 6, Number 4,
pages 719-730.
Yaroub Elloumi is a PHD student since September 2009, registered in
both Paris-Est university (France) and Sfax university (Tunisia). He is a
member of Institut Gaspard-Monge, unité mixte de recherche CNRSUMLPE-ESIEE, UMR 8049. His research interests are High-level design
of real-time system, optimization techniques, and H igh-level parameter
estimation.
Mohamed Akil received his PhD degree from the Montpellier university
(France) in 1981 and hi s doctorat d'état from the Pierre et Marie curie
University (Paris, France) in 1985. He currently teaches and does
research with the position of Professor at computer science department,
ESIEE, Paris. He is a member of Institut Gaspard-Monge, unité mixte de
recherche CNRS-UMLPE-ESIEE, UMR 8049. His research interests are
Architecture for image processing, Image compression, Reconfigurable
architecture and FPGA, High-level Design Methodology for multi-FPGA,
mixed architecture (DSP/FPGA), System on C hip (SoC) and parallel
programming of 2D/3D topological operators. Dr. Akil has more than 80
research papers in the above areas.
Mohamed Hedi Bedoui received his PhD degree from Lille University in
1993. He currently teaches with the position of Professor of biophysics in
the Faculty of Medicine of Monastir (FMM), Tunisia. He is a member of
Medical Technology and i mage processing team (TIM), UR 08-27. His
research interests are real-time and embedded systems, image & signal
processing and har dware/software design in medical field, electronic
applications in biomedical instrumentation. He is the president of the
Tunisian Association of Promotion of Applied Research.
| 6cs.PL
|
REGULAR BOARDGAMES
arXiv:1706.02462v1 [cs.AI] 8 Jun 2017
JAKUB KOWALSKI, JAKUB SUTOWICZ, AND MAREK SZYKUŁA
Abstract. We present an initial version of Regular Boardgames general game description language. This stands as an extension of Simplified Boardgames language. Our language is designed
to be able to express the rules of a majority of popular boardgames including the complex rules
such as promotions, castling, en passant, jump captures, liberty captures, and obligatory moves.
The language describes all the above through one consistent general mechanism based on regular
expressions, without using exceptions or ad hoc rules.
1. Introduction
1.1. Simplified Boardgames. Simplified Boardgames is the class of fairy chess-like games introduced by Björnsson in [1]. The language describes turn-based, two player, zero-sum chess-like games
on a rectangular board with piece movements described by regular languages and independent on
the move history. It was slightly extended in [2], and used as a comparison class for assessing the
level of Stanford’s GGP programs.
The language can describe many of the fairy chess variants in a concise way. Unlike Metagame [4],
Simplified Boardgames include games with asymmetry and position-dependent moves (e.g. Chess
initial double pawn move). The usage of finite automata for describing pieces’ rules, and thus
to move generation, allows fast and efficient computation of all legal moves given a board setup,
however, it causes some important limitations. For example it is impossible to express actions like
castling, en-passant, or promotions.
Here we follow the class formalization from [3] to provide a shortened necessary introduction.
The game is played between the two players, black and white, on a rectangular board of size
width×height. White player is always the first to move. Although it may be asymmetric, the initial
position is given from the perspective of the white player, i.e. forward means “up” for white, and
“down” for black.
During a single turn, the player has to make a move using one of his pieces. Making a move is
done by choosing the piece and changing its position according to the specified movement rule for
this piece. At any time, at most one piece can occupy a square, so finishing the move on a square
containing a piece (regardless of the owner) results in removing it (capturing). No piece addition
is possible. After performing a move, the player gives control to the opponent.
For a given piece, the set of its legal moves is defined as the set of words described by a regular
expression over an alphabet Σ containing triplets (∆x, ∆y, on), where ∆x and ∆y are relative
column/row distances, and on ∈ {e, p, w} describes the content of the destination square: e indicates
an empty square, p a square occupied by an opponent piece, and w a square occupied by an
own piece. A positive ∆y means forward, which is a subjective direction, and differs in meaning
depending on the player.
Institute of Computer Science, University of Wrocław, Wrocław, Poland
E-mail addresses: [email protected], [email protected] , [email protected].
1
2
REGULAR BOARDGAMES
Consider a piece and a word w ∈ Σ∗ that belongs to the language described by the regular
expression in the movement rule for this piece. Let w = a1 a2 . . . ak , where each ai = (∆xi , ∆yi , on i ),
and suppose that the piece stands on a square hx, yi. Then, w describes a move of the piece, which
is applicable in the current board position if and only if, for every i such that 1 ≤ i ≤ k, the content
Pi
Pi
condition on i is fulfilled by the content of the square hx + j=1 ∆xj , y + j=1 ∆yj i. The move of
Pk
Pk
w changes the position of the piece piece from hx, yi to hx + i=1 ∆xi , y + i=1 ∆yi i.
For example, the movement rule of the rook in chess is expressed by:
R (0,1,e)^*((0,1,e)+(0,1,p)) + (0,-1,e)^*((0,-1,e)+(0,-1,p)) +
(1,0,e)^*((1,0,e)+(1,0,p)) + (-1,0,e)^*((-1,0,e)+(-1,0,p)) &
There are three types of terminal conditions that can be defined. The first one is based on the
turnlimit, whose exceedance automatically causes a draw if no other terminal condition is fulfilled.
A player can win by moving a certain piece to a fixed set of squares (e.g. move a pawn to the
opponent backrank). Alternatively, a player can lose if the number of his pieces of a certain type
reaches a given amount (e.g. 0 king type pieces). The terminal conditions can be asymmetric. A
player automatically loses the game when he has no legal moves at the beginning of his turn.
1.2. Regular Boardgames. The construction of Regular Boardgames aims to remove most of
the Simplified Boardgames restrictions. It emerged from an analysis of popular boardgames (not
only chess variants and checkers but also go, tic-tac-toe, exchange games and games for more than
two players), and a try to generalize the concepts they use. Our goal was to provide maximal
expressiveness under the one common set of rules, without introducing special cases. The resulting
language allows to describe many non-standard behaviors, although, as no shortcutting exceptions
are provided, they may require non-trivial usage of the language constructions. Of course still there
are some concepts that are impossible to express, e.g. counting the board repetitions, as it requires
to store the game history. Summarizing, the main features of this extension are:
(1) A possibility to remove existing pieces or produce new pieces during a move (to express
promotions in chess or sequential capture in checkers).
(2) Priorities, forcing to perform a move with the highest possible priority (to express that
a sequential capture must take a maximal number of pieces in checkers; also helpful in
expressing complex behaviors like en passant).
(3) Regular expressions as terminal conditions (to express e.g three in a row rule in tic-tac-toe).
(4) A lot of minor syntax and semantic issues, like the arbitrary selection of the next player
after a move.
(5) C-like macros to make writing game descriptions easier.
In the following sections we present detailed language syntax with the descriptions of semantics,
and an example of the fully defined game.
2. Syntax and Semantics
In this section we provide the formal grammar of Regular Boardgames in the EBNF format.
The grammar is introduced in steps, alongside with the description of each part, to make it more
convenient to the reader.
The start non-terminal symbol is “rbg”. Tokens can be separated using any whitespace characters
forming strings of arbitrary lengths. C-like comments can appear anywhere in the game definition:
“//” starts a line comment and every next character in the line is ignored. “/*” starts a multiline
comment and every character is ignored until the first occurrence of “*/”.
REGULAR BOARDGAMES
3
The main sections of the game, defined by “rbg” symbol, can be defined in any order. Moreover,
macrodefinitions (described in details in Section 2.1), can appear anywhere inside the sections, as
they are in some sense not a part of the final language. The actual game code is the one after
unwrapping all defined macros.
rbg ::= permutation( game, board, players, goals, order );
game ::= ‘#game‘’ ‘"’ alphanumspace {alphanumspace} ‘"’ ;
alphanumspace ::= any alphanumeric character or space ;
alphanum ::= any alphanumeric character;
letter ::= any letter character;
string ::= sequence of alphanumeric characters or spaces or ‘~’ or any other valid token such
as ‘+’,‘-’ etc.;
nat ::= any natural number ;
int ::= [‘-’], nat ;
2.1. Macros. Macros are construction that allows to reduce the repetitive code. They are defined
in a similar way to functions in standard programming languages, having name list of arguments
and body. It is possible to overload macros when they have different, non-zero arity.
macro ::= ‘#def’, identifier, arguments, ‘=’ string
| ‘#def’, identifier, ‘=’ string;
identifier ::= alpha {alphanum} ;
arguments ::= ‘(’, identifier { ‘,’, identifier}, ‘)’ ;
When the macro is called, the occurrences of its arguments (surrounded by whitespaces or tilde)
are replaced by their values (defined as list of tokens). If two tokens are separated by tilde after
macro unwrapping and their concatenation is also a valid token, they are merged. Afterwards, the
tilde characters are removed.
Below, the example showing the basic behavior of the macro construction.
#def
#def
#def
#def
#def
//
//
//
//
//
M1(x) = x~5
M2(x,y,z) = M1(x) M1(y)~z
M3(x) = x (a,b,c)
M4(x,y) = x~y (5)
M5(x,y,z) = a b~M1(x x)~y~z~c d
M1(4) => 45
M1(a) => a5
M1(^) => invalid; ’^’ and ’5’ cannot be concatenated into one token
arguments can consist of more than one token:
M1(* a) => * a5
4
REGULAR BOARDGAMES
//
//
//
//
//
//
//
//
//
M2(a,b,6) => a5 b56
some or all arguments can be ommited:
M2(a,b,) => a5 b5
M2(,,) => 5 5
macro names created dynamically are not interpreted as macro calls:
M3(M2) => M2 (a,b,c)
M4(M,1) => M1 (5)
merging arguments with ’~’ works through other macro calls:
M5(e,f,g) => a be e5fgc d
Lastly, it has to be pointed out that all constructions, to be valid, have to be syntactically correct
after applying the macros. That is – it is valid to have e.g. expressions and macros with unmaching
brackets, as long as they match after the full expansion of all contained macros.
2.2. Board initialization. The initial game position is given in the #board definition consisting
of board’s width and height followed by the pieces names. Pieces can be named as any string
containing alphanumeric characters.
board ::= ‘#board’, width, height, piecename, { piecename } ;
width ::= nat ;
height ::= nat ;
piecename ::= alpha, {alphanum} ;
The ordering of given pieces works as follows: first given piece is in the top row, leftmost column,
the second one is top row, second column, etc. It is not possible to skip any square. If one wants to
leave the square empty, some dummy piece has to be put there. The bottom row, leftmost column
is (0, 0) square. Below, the example of initializing the board of simple chess Silverman 4 × 5 variant1
(using macros).
#def PIECES(color) = color~Rook color~King color~Queen color~Rook
#def COPY5(piece) = piece piece piece piece piece
#board 4 5 PIECES(Black) COPY5(BlackPawn) COPY5(empty) COPY5(WhitePawn)
PIECES(White)
2.3. Players definition. The definition of the player rules, is the core element of the game. The
order of the players is defined in the #order list, while the set of legal moves of each player is
described by the proper player definition.
Legal moves are defined in the sets described in the descending order of priority. If there is any
move legal in the higher-priority set, it has to made by the player. Only in the opposite case the
moves from the next lower-priority set can be tried.
Which player turn should be next, is determined by the move the current player will choose. For
every set of moves described by the regular language given in single movexpr different effect can
be set. There are two ways to specify the next player. First is to explicitly put the player name,
1https://en.wikipedia.org/wiki/Minichess#4.C3.974.2C_4.C3.975_and_4.C3.978_chess
REGULAR BOARDGAMES
5
while the second is to put the number indicating offside given the established player order (i.e. +1
is always the next player, and the first player after the last one).
If next player is not specified, the so-called semimove occurs. The turn counter does not increase,
and the current player can perform another move.
player ::= ‘#player’, playername, movexpr, {‘+>’, movexpr} ;
playername ::= alpha, {alphanum} ;
order ::= ‘#order’, playername, {playername} ;
2.4. Movement rules. Here, we present movexpr non-terminal, which encodes the regular language of possible piece movements.
This language is the set of words over the alphabet Σ, containing quadruples (∆x, ∆y, on, off )
and all players names. ∆x and ∆y are relative column/row distances, on is the set of pieces which
can stay on the square to enter (i.e. the precondition), and off is the set of pieces which can be
left after the passing. We assume that x ∈ {−width + 1, . . . , width − 1}, and y ∈ {−height +
1, . . . , height − 1}, and so Σ is finite.
movexpr ::= step, [power]
| movexpr, movexpr
| movexpr, ‘+’, movexpr
| ‘(’, movexpr, ‘)’, [power]
| nextplayer
nextplayer ::= ‘[’, playername, ‘]’ | ‘[’, int, ‘]’ ;
power ::= ‘^’ nat | ‘^’ ‘*’
step ::= ‘(’, deltax, ‘,’, deltay, ‘,’, ons, ‘/’, offs, ‘)’
| ‘(’, deltax, ‘,’, deltay, ‘,’, ons, ‘)’
| ‘(’, deltax, ‘,’, deltay, ‘/’, offs ‘)’
| ‘(’, deltax, ‘,’, deltay, ‘)’
| ‘(’, ons, ‘/’, offs ‘)’
| ‘(’, ons ‘)’
| ‘(’, ‘/’, offs ‘)’ ;
deltax ::= int ;
deltay ::= int ;
ons ::= pieceset ;
offs ::= pieceset ;
pieceset ::= ‘{’, piecename, {‘,’, piecename}, ‘}’
| piecename ;
It should be noted that nextplayer can appear on every level of movexpr tree as long as it fulfills
one condition. Square brackets enclosed player name or offset can appear only at the end of every
word generated by movexpr expression. It is allowed to not appear at all.
6
REGULAR BOARDGAMES
// valid:
#player white
#player white
#player white
#player white
// invalid:
#player white
#player white
#player white
(1,1,e)(2,2,e)([white]+[black])
(3,2,e)([white]+(0,3,e))
(0,1,e)[black]
(1,1,e)(3,3,e)
(1,1,e)[white](2,2,e)
(2,2,e)[black]^*
(2,0,e)([white]+(1,1,e))(1,1,e)
A number of special cases can be defined for a single move step. If not specified, ∆x, ∆y are
equal to 0. Empty on axis accepts any square. Empty off is automatically equal to the current
content of the square.
Consider a word w ∈ Σ∗ that belongs to the language described by the regular expression in the
movement rule. The move can be performed from any square, as long as the formula is fulfilled.
Let w = a1 a2 . . . ak , where each ai = (∆xi , ∆yi , on i ), and we consider a square hx, yi.
Then, w describes a move, which is applicable in the current board position if and only if, for
every i such that 1 ≤ i ≤ k, the content condition on i is fulfilled by the content of the square
Pi
Pi
hx + j=1 ∆xj , y + j=1 ∆yj i. Note that the content of any intermediate square may depend on
the piece left on that square by the postcondition (off ) on the previous visit on that square.
A single capturing jump by checkers man piece can be described as follows (assuming the macro
empty indicates an empty square, and opponents a set of opponent pieces):
#def MAN_CAPTURE = (man/empty)((1,1,{opponents}/empty)(1,1,empty/man) +
(-1,1,{opponents}/empty)(-1,1,empty/man))
2.5. Goals and terminal conditions. Lastly, the terminal conditions for each player are defined
in goal sections. The goals can be different for each player, and they are boolean formulas evaluating
to true or false after every move along with a positive number which represents player’s potential
score. The endgame and score are computed in the following way:
• If no player has his goal expression evaluated to true – the game continues.
• If any player have his goal formula fulfilled, the game ends. The score of every player whose
formula is fulfilled is the natural number specified in his goal section. The other players
get no points.
This way, the draw is expressed by the simultaneous occurrence of winning conditions for multiple
players.
goal ::= ‘#goal’, playername, nat, goalexpr ;
goalexpr ::= goalexpr, ‘and’, goalexpr
| goalexpr, ‘or’, goalexpr
| ‘not’, goalexpr
| ‘(’, goalexpr, ‘)’
| value, cmp, value
| ‘@’, piecename, nat, nat
| ‘move(’, movexpr, ‘)’ ;
REGULAR BOARDGAMES
7
value ::= ‘$’, piecename | ‘$turn’ | int ;
cmp ::= ‘<’ | ‘>’ | ‘<=‘ | ‘>=’ | ‘==’ | ‘!=’ ;
Goal expressions can consist of the following components
• Comparison between some constant values, number of onboard pieces of given name (e.g.
$BlackPawn), and the current turn. Turn counter starts with 1 and increases each time the
move with specified ‘nextplayer’ is used.
• Square occupancy check, which evaluates to true iff. a given piece stands on a given square.
The squares are identified using the absolute coordinates – (0, 0) being the bottom-left
corner, and (height − 1, 0) being the top-left corner in the initial position.
• Movement patterns. Iff. there exist a square for which a given move is legal, the it is
evaluated to true. Movement pattern cannot contain any off.
Goal conditions are checked after the player’s turn and are not checked after semimoves. It
means, that that no player can win or lose if current move does not specify a next player. The
exception is when the current player cannot perform legal move. In this case he automatically loses
regardless of the next player specification .
The goal section for the Chess variant without promotions, where the turnlimit is set to 100 and
reaching the opponent backrank with the pawn wins the game, is presented below.
#def OR8(p,r) = p 0 r or p 1 r or p 2 r or p 3 r or
p 4 r or p 5 r or p 6 r or p 7 r
#def BACKRANK(col,row) = OR8(@ col~pawn,row)
#goal white 1 ($bkingu < 1 and $bkingm < 1) or BACKRANK(w,8) or $turn > 100
#goal black 1 ($wkingu < 1 and $wkingm < 1) or BACKRANK(b,1) or $turn > 100
Another example, terminal conditions for Tic-Tac-Toe:
#def
LINE(mark) = (mark)(
(0,1,mark)(0,1,mark) + (1,1,mark)(1,1,mark) + (1,0,mark)(1,0,mark))
#goal xs 1 move(LINE(x)) or $turn > 9
#goal os 1 move(LINE(o)) or $turn > 9
3. Complete Examples
3.1. Chess.
1
2
3
4
5
6
7
8
9
10
# game " Chess ␣ ( alpha ) "
# def INITP ( u p p e r p a w n , s , l o w e r p a w n ) =
upperpawn upperpawn upperpawn upperpawn upperpawn upperpawn upperpawn upperpawn
s∼rooku s∼knight s∼bishop s∼queen s∼kingu s∼bishop s∼knight s∼rooku
lowerpawn lowerpawn lowerpawn lowerpawn lowerpawn lowerpawn lowerpawn lowerpawn
# def INITE = empty empty empty empty empty empty empty empty
# def INITZ = zone zone zone zone zone zone zone zone
8
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
REGULAR BOARDGAMES
# board 8 10 INITZ INITP ( ,b,bpawn ) INITE INITE INITE INITE INITP ( wpawn,w, ) INITZ
# def PIECES ( s ) = s∼rookm,s∼rooku,s∼knight,s∼bishop,s∼queen,s∼kingm,s∼kingu,s∼pawn,s∼pawnd
# def ANYSQUARE = { PIECES ( w ) ,PIECES ( b ) ,empty }
# def P R O M O T E P I E C E S ( s ) = { s∼rookm,s∼knight,s∼bishop,s∼queen }
# def RIDE ( x,y,o,fig ) = (/ empty )( x,y, { empty })^*( x,y, { e m p t y , P I E C E S ( o )}/ fig )
# def ROOK ( s,o,fig ) = RIDE (1 ,0,o,fig ) + RIDE ( -1 ,0,o,fig ) + RIDE (0 ,1,o,fig ) + RIDE (0 , -1 ,o,fig )
# def ROOK ( s,o )
= ROOK ( s,o,s∼rookm ) // indicate as ’ moved ’
# def BISHOP ( s,o,fig ) = RIDE (1 ,1,o,fig ) + RIDE ( -1 ,1,o,fig ) + RIDE (1 , -1 ,o,fig ) + RIDE ( -1 , -1 ,o,fig )
# def BISHOP ( s,o )
= BISHOP ( s,o,s∼bishop )
# def QUEEN ( s,o ) = ROOK ( s,o,s∼queen ) + BISHOP ( s,o,s∼queen )
# def HOP ( x,y,o,fig )
= (/ empty )( x,y, { e m p t y , P I E C E S ( o )}/ fig )
# def HOPAROUND ( x,y,o,fig ) = HOP ( x,y,o,fig ) + HOP ( x, - y,o,fig ) + HOP ( - x,y,o,fig ) + HOP ( - x, - y,o,fig )
# def KNIGHT ( s,o ) = HOPAROUND (1 ,2,o,s∼knight ) + HOPAROUND (2 ,1,o,s∼knight )
# def KING ( s,o ) = HOP (1 ,0,o,s∼kingm ) + HOP ( -1 ,0,o,s∼kingm ) + HOP (0 ,1,o,s∼kingm ) + HOP (0 , -1 ,o,s∼kingm )
+ HOPAROUND (1 ,1,o,s∼kingm )
# def CASTLING ( s ) = ( s∼rooku / empty )(1 ,0,empty )(1 ,0,empty / s∼kingm )(1 ,0,empty / s∼rookm )
(1 ,0,s∼kingu / empty )
+ ( s∼rooku / empty )( -1 ,0,empty / s∼kingm )( -1 ,0,empty / s∼rookm )( -1 ,0,s∼kingu / empty )
# def PAWNPROMO T E ( s,g ) = ((0 ,g 1 ,ANYSQUARE ) + (/ P R O M O T E P I E C E S ( s ))(0 ,g 1 ,zone ))
# def PAWN ( s,o,g,d )
= (/ empty )((0 ,g 1 ,empty / s∼pawn ) + ( -1 ,g 1 , { PIECES ( o )}/ s∼pawn ) +
(1 ,g 1 , { PIECES ( o )}/ s∼pawn )) PAWNPROMO TE ( s,g ) // standard move
+ (/ empty )(( -1 ,0,o∼pawnd / empty ) + (1 ,0,o∼pawnd / empty ))(0 ,g 1 ,empty / s∼pawn ) // en - passant
+ (/ empty )(0 ,g 1 ,empty )(0 ,g 6 ,zone )(0 ,d 5 ,empty / s∼pawnd ) // initial double move
# def R E M O V E D O U B L E P A W N ( s ) = ( s∼pawnd / s∼pawn )
# def LEGALS ( s,o,g,d ) = ({ s∼rooku,s∼rookm })( ROOK ( s,o )) + ( s∼bishop )( BISHOP ( s,o ))
+ ({ s∼kingu,s∼kingm })( KING ( s,o )) + ( s∼queen )( QUEEN ( s,o ))
+ ( s∼knight )( KNIGHT ( s,o )) + ({ s∼pawn,s∼pawnd })( PAWN ( s,o,g,d ))
+ CASTLING ( s )
# player white R E M O V E D O U B L E P A W N ( w ) +> ( LEGALS ( w,b,, -))[ black ]
# player black R E M O V E D O U B L E P A W N ( b ) +> ( LEGALS ( b,w, - , ))[ white ]
# order white black
REGULAR BOARDGAMES
9
61
62 # goal white 1 ( $ bkingu < 1 and $ bkingm < 1) or $ turn > 100
63 # goal black 1 ( $ wkingu < 1 and $ wkingm < 1) or $ turn > 100
References
[1] Y. Björnsson. Learning Rules of Simplified Boardgames by Observing. In European Conference on Artificial
Intelligence, volume 242 of FAIA, pages 175–180. 2012.
[2] J. Kowalski and A. Kisielewicz. Testing General Game Players Against a Simplified Boardgames Player Using
Temporal-difference Learning. In IEEE Congress on Evolutionary Computation, pages 1466–1473, 2015.
[3] J. Kowalski, J. Sutowicz, and M. Szykuła. Simplified Boardgames. arXiv:1606.02645 [cs.AI], 2016.
[4] B. Pell. METAGAME: A New Challenge for Games and Learning. In Heuristic Programming in Artificial Intelligence: The Third Computer Olympiad., 1992.
10
REGULAR BOARDGAMES
This figure "examplechess.png" is available in "png" format from:
http://arxiv.org/ps/1706.02462v1
| 2cs.AI
|
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
Bradly C. Stadie 1 Ge Yang 2 Rein Houthooft 3 Xi Chen 4 Yan Duan 4 Yuhuai Wu 5 Pieter Abbeel 4 6
Ilya Sutskever 3
arXiv:1803.01118v1 [cs.AI] 3 Mar 2018
Abstract
We consider the problem of exploration in meta
reinforcement learning. Two new meta reinforcement learning algorithms are suggested: EMAML and E-RL2 . Results are presented on a
novel environment we call ‘Krazy World’ and a
set of maze environments. We show E-MAML
and E-RL2 deliver better performance on tasks
where exploration is important.
1. Introduction
Supervised learning algorithms typically have their accuracy measured against some held-out test set that the algorithm was not privy too at training time. A supervised
learning model generalizes well if it maintains its accuracy
under data that has non-trivial dissimilarity from the training data. This approach to evaluating algorithms based on
their generalization ability contrasts the approach in reinforcement learning (RL), wherein there is usually no distinction between training and testing environments. Instead, an agent is trained on one environment, and results
are reported on this same environment. Most RL algorithms thus favor mastery over generalization.
Meta RL is the suggestion that RL should take generalization seriously. In Meta RL, agents are evaluated on their
ability to quickly master new environments at test time.
Thus, a meta RL agent must not learn how to master the
environments it is given, but rather it must learn how to
learn so that it can quickly train at test time. Recent advances in meta RL have introduced algorithms that are capable of generating large policy improvements at test time
with minimal sample complexity requirements (Finn et al.,
2017; Duan et al., 2016; Wang et al., 2016).
One key question for meta RL that has been inadequately
1
UC Berkeley, Department of Statistics 2 University of
Chicago, Department of Physics 3 OpenAI 4 UC Berkeley, Department of EECS 5 University of Toronto 6 International Computer Science Institute. Correspondence to: Bradly C. Stadie
<[email protected]>.
Preprint. Code will be released with conference version of paper.
Code not available with preprint version on ArXiv.
considered by (Finn et al., 2017; Duan et al., 2016) is that
of exploration (see (Kompella et al., 2002) for an overview
of exploration in this context). A crucial step in learning
to solve many environments is an initial period of exploration and system identification. Furthermore, we know
that real-life agents become better at this exploratory phase
with practice. Unfortunately, we have found existing algorithms deficient in this area. We hypothesize that this
failure can at least partially be attributed to the tendency of
existing algorithms to do greedy, local optimization at each
step of meta-training, as discussed further below.
To address the problem of exploration in meta RL, we introduce two new algorithms: E-MAML and E-RL2 . It
should come as no surprise that these algorithms are similar
to their respective namesakes MAML and RL2 . The algorithms are derived by reformulating the underlying metalearning objective to account for the impact of initial sampling on post-meta-updated returns. We show that our algorithms achieve better results than MAML and RL2 on two
environments: a Krazy World environment we developed
to benchmark meta RL, and a set of maze environments. In
section 6.1 of this work, we consider a more general form
of our E-MAML derivation. This more general derivation
suggests several promising directions for future work on
exploration in meta-learning and highlights the novelty of
our contributions.
2. Related Work
Recently, a flurry of new work in Deep Reinforcement
Learning has provided the foundations for tackling RL
problems that were previously thought intractable. This
work includes: 1) (Mnih et al., 2015; 2016; Koutnk et al.,
2013), which allow for discrete control in complex environments directly from raw images. 2) (Schulman et al., 2015;
Mnih et al., 2016; Schulman et al., 2017; Lillicrap et al.,
2015), which have allowed for high-dimensional continuous control in complex environments from raw state information.
Although these algorithms offer impressive capabilities,
they still often falter when faced with environments where
exploration presents a significant challenge. This is not
surprising, as there exists a large body of RL work ad-
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
dressing the exploration problem (Kearns & Singh, 2002;
Brafman & Tennenholtz, 2002; Kolter & Ng, 2009). In
practice, these methods are often not used due to difficulties with high-dimensional observations, difficulty in implementation on arbitrary domains, and lack of promising
results (though recent work such as (Schmidhuber, 2015a)
has seen progress in this regard). This resulted in most deep
RL work utilizing epsilon greedy exploration (Mnih et al.,
2015), or perhaps a simple scheme like Boltzmann exploration (Carmel & Markovitch, 1999). As a result of these
shortcomings, a variety of new approaches to exploration in
deep RL have recently been suggested (Tang et al., 2016;
Houthooft et al., 2016; Stadie et al., 2015; Osband et al.,
2016; Bellemare et al., 2016; Ostrovski et al., 2017). Further, there has been extensive research on the idea of artificial curiosity. For example (Ngo et al., 2012; Graziano
et al., 2011; Schmidhuber, 1991; 2015b; Storck et al., 1995;
Sun et al., 2011) all deal with teaching agents to generate
a curiosity signal that aids in exploration. The work (Kompella et al., 2002) even considers the problem of curiosity
in the context of meta-RL via Success Story Algorithms
(SSA). These algorithms use dynamic programming to explicitly partition experience into checkpoints and serve as
a precursor to MAML and RL2 , which are less focused on
explicitly partitioning episodes and more focused on handling high dimensional input (Schmidhuber et al., 1997). In
spite of the numerous cited efforts in the above paragraph,
the problem of exploration in RL remains difficult.
This paper will consider the problem of exploration in meta
RL. However, it is important to note that many of the problems in meta RL can alternatively be addressed with the
field of hierarchical reinforcement learning. In hierarchical RL, a major focus is on learning primitives that can
be reused and strung together. These primitives will frequently enable better exploration, since they’ll often relate
to better coverage over state visitation frequencies. Recent
work in this direction includes (Vezhnevets et al., 2017;
Bacon & Precup, 2015; Tessler et al., 2016; Rusu et al.,
2016; Barto & Mahadevan, 2003; Wiering & Schmidhuber,
1997). The primary reason we consider meta learning over
hierarchical RL is that we find hierarchical RL tends to focus more on defining specific architectures that should lead
to hierarchical behavior, whereas meta learning instead attempts to directly optimize for these behaviors.
As for meta RL itself, the literature is spread out and goes
by many different names. There exist relevant literature on
life-long learning, learning to learn, continual learning, and
multi-task learning. Consider for instance (Schmidhuber,
2006; 1987). These papers deal with self modifying learning machines or genetic programs. In particular, they propose that if you have a genetic program that itself modifies
the underlying genetic program, you will obtain a meta-GP
approach (see 6.1 for some clarification of this approach to
the the algorithms presented in this paper). These methods describe an extremely broad class of algorithms which
could be said to encompass most meta-learning approaches
with the correct interpretation. In addition to these references, we encourage the reviewer to look at the review articles (Silver et al., 2013; Taylor & Stone, 2009; Thrun,
1996) and their citations. Most attempts to address exploration in these contexts has focused on curiosity or on a
free learning phrase during training. Curiosity is addressed
above, and is quite different from this work because it focuses on defining an intrinsic motivation signal, whereas
we only consider better utilization of the existing reward
signal. This paper’s methods share more similarities with
free learning phases, though our algorithm makes a more
explicit connection between this phase and the more exploitative phase of training (directly incorporating this information into the gradient updates).
This paper will seek to capture insights from this prior literature and apply them to recently developed algorithms.
In particular, we will consider considering E-MAML and
E-RL2 , which build upon the existing meta learning algorithms MAML (Finn et al., 2017) and RL2 (Duan et al.,
2016). In MAML, tasks are sampled and a policy gradient
update is computed for each task with respect to a fixed initial set of parameters. Subsequently, a meta update is performed where a gradient step is taken that moves the initial
parameter in a direction that would have maximally benefited the average return over all of the sub-updates. Results
are presented on an environment wherein a simulated ant
must determine its goal direction, which changes each time
a new task is sampled. In RL2 , an RNN is made to ingest
multiple episodes from many different MDPs and then perform a policy gradient update through the entire temporal
span of the RNN. The hope is that the RNN will learn a
faster RL algorithm in its memory weights.
3. Reinforcement Learning Notation
Let M = (S, A, P, r, ρ0 , γ, T ) represent a discrete-time
finite-horizon discounted Markov decision process (MDP).
The elements of M have the following definitions: S is
a state set, A an action set, P : S × A × S → R+
a transition probability distribution, r : S × A → R a
reward function, ρ0 : S → R+ an initial state distribution, γ ∈ [0, 1] a discount factor, and T the horizon.
We will sometimes speak of M having a loss function L
rather than reward function r. All we mean here is that
L(s) = −r(s) In a classical reinforcement learning setting, we optimize to obtain a set of parameters θ which
maximize the expected discounted return under the policy πθ : S × A → R+ . That is, we optimize to obPT
tain θ that maximizes η(πθ ) = Eπθ [ t=0 γ t r(st )], where
s0 ∼ ρ0 (s0 ), at ∼ πθ (at |st ), and st+1 ∼ P(st+1 |st , at ).
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
4. Problem Formulation and Algorithms
This provides an unbiased expression for computing the
gradient which correctly accounts for the dependence on
the initial sampling distribution.
4.1. The Meta Reinforcement Learning Objective
In meta reinforcement learning, we consider a family of
MDPs M = {Mi }N
i=1 which comprise a distribution of
tasks. The goal of meta RL is to find a policy πθ and paired
update method U such that, given Mi ∼ M, we have πU (θ)
solves Mi quickly. The word quickly is important here, so
let us elaborate. By quickly, we mean orders of magnitude
more sample efficient than simply solving Mi with policy
gradient or value iteration methods. In fact, an ideal meta
RL algorithm would solve these tasks by collecting 1-10
trajectories from the environment for tasks where policy
gradients require 100,000 or more trajectories. The thinking here goes that if an algorithm can solve a problem with
so few samples, then it might be ‘learning to learn.’ That
is, the agent is not learning how to master a given task but
rather how to quickly master new tasks. We can write this
objective cleanly as
min
θ
X
E(πU (θ) ) [LMi ]
In any case, we find ourselves wishing to find a tractable
expression for the gradient of (2). This can be done quite
smoothly by applying the product rule under the integral
sign and going through the REINFORCE style derivation
twice to arrive at a two term expression, one of which encourages exploitation and one of which encourages exploration.
ZZ
∂
R(τ )πU (θ) (τ )πθ (τ̄ )dτ̄ dτ
∂θ
ZZ
∂
=
R(τ ) πθ (τ̄ ) πU (θ) (τ )+
(3)
∂θ
∂
πU (θ) (τ ) πθ (τ̄ ) dτ̄ dτ
∂θ
≈
T
∂
1X
R(τ i ) log πU (θ) (τ i )+
T i=1
∂θ
(1)
T
∂
1X
R(τ i ) log πθ (τ̄ i )
T i=1
∂θ
Mi
Note that this objective is similar to the one that appears in
MAML (Finn et al., 2017), which we will discuss further
below. In MAML, U is chosen to be the stochastic gradient
descent.
4.2. E-MAML
We can expand the expectation from (1) into the integral
Z
R(τ )πU (θ) (τ )dτ
(2)
The objective (1) can be optimized by taking a derivative
of this integral with respect to θ and carrying out a
standard REINFORCE style analysis to obtain a tractable
expression for the gradient (Williams, 1992).
One issue with this direct optimization is that it will
not account for the impact of the original sampling
distribution πθ on the future rewards R(τ ), τ ∼ πU (θ) .
We would like to account for this, because it would allow
us to weight the initial samples τ̄ ∼ πθ by the expected
future returns we expect after the sampling update R(τ ).
More importantly, this extra-term reinforces particular
explorative trajectories if the updated agent learn well in
the end. On the contrary, when this term is abscent, good
exploratory behaviors will not be encouraged.
Including this dependency can be done by writing the modified expectation as
ZZ
R(τ )πU (θ) (τ )πθ (τ̄ )dτ̄ dτ
(4)
i
τ̄ ∼ πθ
i
τ ∼ πU (θ)
Note that, if we only consider the term on the left, we arrive at the original MAML algorithm (Finn et al., 2017).
This term encourages the agent to take update steps U that
achieve good final rewards. It is exploitative. The second term encourages the agent to take actions such that
the eventual meta-update yields good rewards (crucially, it
does not try and exploit the reward signal under its own
trajectory τ̄ ). This term allows the policy to be more exploratory, as it will attempt to deliver the maximal amount
of information useful for the future rewards R(τ ) without
worrying about its own rewards R(τ̄ ). Since this algorithm
augments MAML by adding in an exploratory term, we call
it E-MAML. For the experiments presented in this paper,
we will assume that U utilized in MAML and E-MAML is
stochastic gradient descent. However, we direct the reader
to section 6.1 for an insightful discussion on the choice of
this operator U .
4.3. E-RL2
RL2 optimizes (1) by feeding multiple rollouts from multiple different MDPs into an RNN. The hope is that the RNN
hidden state update ht = C(xt , ht−1 ), will learn to act as
the update function U . Then, performing a policy gradient
update on the RNN will correspond to optimizing the meta
objective (1).
We can write the RL2 update rule more explicitly in the
following way. Suppose L represents an RNN. Let Envk (a)
be a function that takes an action, uses it to interact with the
MDP representing task k, and returns the next observation
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
o 1 , reward r, and a termination flag d. Then we have
xt = [ot−1 , at−1 , rt−1 , dt−1 ]
(5)
[at , ht+1 ] = L(ht , xt )
(6)
[ot , rt , dt ] = Envk (at )
(7)
To train this RNN, we sample N MDPs from M and obtain
k rollouts for each MDP by running the MDP through the
RNN as above. We then compute a policy gradient update
to move the RNN parameters in a direction which maximizes the returns over the k trials performed for each MDP.
To make this algorithm more exploratory, we can take inspiration from the insights shone by E-MAML. For each
MDP Mi , we will sample p exploratory rollouts and k − p
non-exploratory rollouts. During an exploratory rollout,
the forward pass through the RNN will receive all information. However, during the backwards pass, the rewards
contributed during exploratory episodes will be set to zero.
For example, if there is one exploratory episode and one exploitative episode, then we would proceed as follows. During the forward pass, the RNN will receive all information
regarding rewards for both episodes. During the backwards
pass, the returns will be computed as
R(xi ) =
T
X
γ j rj · χE (xj )
(8)
j=i
Where χE is an indicator that returns 0 if the episode is
exploratory and 1 otherwise. This return, and not the standard sum of discounted returns, is what is used to compute the policy gradient. The hope is that zeroing the return
contributions from exploratory episodes will encourage the
RNN to treat its initial rollouts as exploratory, taking actions which may not lead to immediate rewards but rather to
the RNN hidden weights performing better system identification, which will lead to higher rewards in later episodes.
5. Experiments
5.1. Krazy World Environment
To test the ability of meta RL algorithms to explore, we
introduce a new environment known as Krazy World.
This environment has the following features:
1. 8 different tile types: Goal squares provide +1 reward when retrieved. The agent reaching the goal does
not cause the episode to terminate, and there can be
multiple goals. Ice squares will be skipped over in
the direction the agent is transversing. Death squares
RL2 works well with POMDP’s because the RNN is good at
system-identification. This is the reason why we chose to use o as
in “observation” instead of s for “state” in this formulation.
Figure 1: Three example worlds drawn from the task distribution. The agent must first perform an exploratory stage
of system identification before exploiting. For example, in
the leftmost grid the agent must first identify that the orange squares give +1 reward and the blue squares replenish
its limited supply of energy. Further, it will need to identify
that the gold squares block progress and the black square
can only be passed by picking up the pink key. The agent
may also want to identify that the brown squares will kill
it and that it will slide over the purple squares. The other
center and right worlds show these assignments will change
and need to be re-identified every time a new task is drawn.
will kill the agent and end the episode. Wall squares
act as a wall, impeding the agent’s movement. Lock
squares can only be passed once the agent has collected the corresponding key from a key square. Teleporter squares transport the agent to a different teleporter square on the map. Energy squares provide the
agent with additional energy. If the agent runs out of
energy, it can no longer move. The agent proceeds
normally across normal squares.
2. Ability to randomly swap color palette: The color
palette for the grid can be randomly permuted, changing the color that corresponds to each of the different
tile types. The agent will thus have to identify the new
system to achieve a high score. Note that in representations of the gird wherein basis vectors are used
rather than images to describe the state space, each
basis vector will correspond to a tile type, and permuting the colors will correspond to permuting the types
of tiles these basis vectors represent. We prefer to use
the basis vector representation in our experiments, as
it is more sample efficient.
3. Ability to randomly swap dynamics: The game’s
dynamics can be altered. The most naive alteration
simply permutes the player’s inputs and corresponding actions (issuing the command for down moves the
player up etc). More complex dynamics alterations
allow the agent to move multiple steps at a time in arbitrary directions, making the movement more akin to
that of chess pieces.
1
4. Local or Global Observations: The agent’s observation space can be set to some fixed number of squares
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
around the agent, the squares in front of the agent, or
the entire grid. Observations can be given as images
or as a grid of basis vectors. For the case of basis vectors, each element of the grid is embedded as a basis
vector that corresponds to that tile type. These embeddings are then concatenated together to form the
observation proper. We will use local observations.
Figure 2: A comparison of local and global observations
for the Krazy World environment. In the local mode, the
agent only views a 3×3 grid centered about itself. In global
mode, the agent views the entire environment.
A successful meta learning agent will first need to explore
the environment, identifying the different tile types, color
palette, and dynamics. A meta learning algorithm that cannot learn to incorporate exploration into its quick updating
strategy will ultimately have insufficient data to navigate
the game. We see this empirically in our experiments below.
5.2. Mazes
A collection of maze environments. The agent is placed at a
random square within the maze and must learn to navigate
the twists and turns to reach the goal square. A good exploratory agent will spend some time learning the maze’s
layout in a way that minimizes repetition of future mistakes. The mazes are not rendered, and consequently this
task is done with state space only. The mazes are 20 × 20
squares large.
5.3. Results
In this section we present the following experimental results.
1. Learning curves on Krazy World and mazes.
2. The gap between the agent’s initial performance on
new environments and its performance after updating.
A good meta learning algorithm will have a large gap
after updating. A standard RL algorithm will have virtually no gap after only one update.
3. Three heuristic metrics for how well the algorithms
explore.
Figure 3: One example maze environment rendered in human readable format. The agent attempts to find a goal
within the maze.
For both Krazy World and mazes, training proceeds in
the following way. First, we initialize 32 training environments and 64 test environments. Every initialized environment has a different seed. Next, we initialize our
chosen meta-RL algorithm by uniformly sampling hyperparameters from predefined ranges. Data is then collected
from all 32 training environments. The meta-RL algorithm
then uses this data to make a meta-update, as detailed in the
algorithm section of this paper. The meta-RL algorithm is
then allowed to do one training step on the 64 test environments to see how fast it can train at test time. These test
environment results are recorded, 32 new tasks are sampled from the training environments, and data collection
begins again. For MAML and E-MAML, training at test
time means performing one VPG update at test time (see
figure 8 for evidence that taking only one gradient step is
sufficient). For RL2 and E-RL2 , this means running three
exploratory episodes to allow the RNN memory weights
time to update and then reporting the loss on the fourth and
fifth episodes. For both algorithms, meta-updates are calculated with PPO (Schulman et al., 2017). The entire process
from the above paragraph is repeated from the beginning
64 times and the results are averaged to provide the final
learning curves featured in this paper.
When plotting learning curves in Figure 4 and Figure 5, the
Y axis is the reward obtained after training at test time on a
set of 64 held-out test environments. The X axis is the total
number of environmental time-steps the algorithm has used
for training. Every time the environment advances forward
by one step, this count increments by one. This is done to
keep the time-scale consistent across meta-learning curves.
For Krazy World, learning curves are presented in Figure 4.
We see that E-MAML and E-RL2 have the best final performance. E-MAML has the steepest initial gains for the
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
Figure 4: Meta learning curves on Krazy World. We see
that -RL2 is at times closest to the optimal solution of
0.6 (obtained via value iteration), but has the highest initial variance. E-MAML converges faster than MAML, although both algorithms do manage to converge. RL2 has
relatively poor performance and high variance. A random
agent achieves a score of around 0.05 on this task.
first 10 million time-steps. Since meta-learning algorithms
are often very expensive, having a steep initial ascent is
quite valuable. Around 14 million training steps, E-RL2
passes E-MAML for the best performance. By 25 million
time-steps, E-RL2 is somewhat close to the optimal solution (obtained via value iteration). However, it appears to
have converged below this value. MAML delivers comparable final performance to E-MAML, suggesting that perhaps MAML also learns to explore sufficiently well on this
task. However, it takes it much longer to obtain this level
of performance. Finally, RL2 has comparatively poor performance on this task and very high variance. When we
manually examined the RL2 trajectories to figure out why,
we saw the agent consistently finding a single goal square
and then refusing to explore any further. The exploration
metrics presented below seem consistent with this finding.
Learning curves for mazes are presented in Figure 5. Here,
the story is different than Krazy World. RL2 and E-RL2
both perform better than MAML and E-MAML. We suspect the reason for this is that RNNs are able to leverage
memory, which is more important in mazes than in Krazy
World. This environment carries a penalty for hitting the
wall, which MAML and E-MAML discover quickly. However, it takes E-RL2 and RL2 much longer to discover this
penalty, resulting in worse performance at the beginning
of training. MAML delivers worse final performance and
typically only learns how to avoid hitting the wall. RL2
and E-MAML sporadically solve mazes. E-RL2 manages
to solve many of the mazes.
When examining meta learning algorithms, one important
metric is the update size after one learning episode. Our
meta learning algorithms should have a large gap between
Figure 5: Meta learning curves on mazes. The environment
gives the agent a penalty for hitting the wall. Interestingly,
E-RL2 and RL2 take much longer to learn how to avoid the
wall. However, they also deliver better final performance.
E-RL2 learns faster than RL2 and offers superior final performance. Since MAML and E-MAML utilize MLPs instead of RNNs and have no memory, their worse returns
are not surprising. The optimal return for this task would
be around 1.1 (obtained via value iteration) and a random
agent would achieve a return of around -2.2 (mostly from
hitting the wall). Note that this figure appears somewhat
busy due to having all four algorithms on one graph. Figure
6 shows each curve in isolation, making it easier to discern
their individual characteristics.
their initial policy, which is largely exploratory, and their
updated policy, which should often solve the problem entirely. For MAML, we look at the gap between the initial
policy and the policy after one policy gradient step (see
figure 8 for information on further gradient steps). For
RL2 , we look at the results after three exploratory episodes,
which give the RNN hidden state h sufficient time to update. Note that three is the number of exploratory episodes
we used during training as well. This metric shares similarities with the Jump Start metric considered in prior literature (Taylor & Stone, 2009). These gaps are presented in
figure 6.
Finally, in Figure 7 we consider three heuristic metrics
for measuring the exploratory ability of the meta-learners.
First, we consider the fraction of tile types visited by the
agent at test time. Our intuition is that good exploratory
agents will visit and identify many different tile types as
this should aid in optimally solving the game. Second, we
consider the number of times an agent visits a death tile at
test time. Agents that are efficient at exploring should visit
this tile type exactly once and then avoid it. We often find
that more naive agents will run into these tiles repeatedly,
causing their death many times over and a sense of pity
in any onlookers. Finally, we look at how many goals the
agent reaches. We see that RL2 tends to visit fewer goals
and instead tend to stick with a strategy of finding one goal
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
Figure 6: Looking at the gap between initial performance and performance after one update for RL2 , E-RL2 , MAML, and
E-MAML. All algorithms show some level of improvement after one update, which suggests the meta-learning is working.
On Krazy World, the gap between the initial policy performance and updated policy performance is fairly large for all
algorithms. Interestingly, the initial performance of all algorithms on gridworld is similar, converging to values around 0.1.
However, E-MAML and RL2 tend to have the highest variance. Meanwhile, for mazes, we see the gap between initial and
update policy returns for MAML and E-MAML is unfortunately small. Perhaps this is because these algorithms do not
have a memory, which makes the maze task more difficult. For RL2 and E-RL2 the gap is bigger, with E-RL2 delivering
the best improvement.
and exploiting it. Overall, we see that the exploratory algorithms achieve better performance under these metrics.
6. Discussion
(12)
· π̂@θ0 @τ̄ dτ dτ̄ .
6.1. Exploration as a choice
During our derivation of E-MAML, we made an implicit
choice in that the inner update operator U takes the form of
SGD. However, this is but one of many eligible operators.
One interesting avenue for the future is considering other
possible choices for U . These choices are best elucidated
by introducing another layer of abstraction on top the EMAML derivation. Consider the path-integral formulation
of the expected reward in E-MAML
ZZ
L(θ0 ) = −
R(τ ) · π(ai |si , U (θ0 , R(τ̄ ),
(9)
τ̄ ∼ π(a|θ0 ))) · π(āi |s̄i , θ0 ) dτ dτ̄ .
We can make it easier to tell the role of U by re-writing the
action probability π and U in type notation where
π̂(θ) : ⇒ {(τ ) : R+ ⇒ π(at |st , θ)}
and Û (θ0 ) : Θ ⇒ U θ0 , R(τ ), τ ∼ π(a|θ0 ) .
left-to-right, the surrogate loss could be written as
ZZ
L(θ0 ) = −
R(τ ) · π̂ ◦ Û @θ0 @τ̄
(10)
In this notation, it becomes clear that you could pick any arbitrary function for the inner-update operator Û . Under this
context, the update operator of choice in E-MAML (and
similarly MAML) is an instance from a more general class
of exploratory operators O:
ÛMAML (θ0 ) : Θ ⇒ SGD LPPO , τ (π(θ0 )), θ0
(13)
where the class of operators it belongs to is ÛMAML ∈ O :
{(Θ) :⇒ Θ}
So what is this PPO/VPG update operator? One could
argue that UMAML is a “reward-maximizing” exploration
strategy. The reward collected during the exploration phase
is used to compute the gradient direction for θ0 , which is
in-turn used to update θ0 → θ.
Then it becomes immediately obvious there are a few other
operators that could also be plugged in:
(11)
π̂ returns a function on trajectories τ ∈ T , whereas Û takes
in two more variables from it’s closure, and returns an updated θ ∈ Θ. Using the Polish notation @ applied from-
• random exploration on a single task, where θ is perturbed randomly.
• natural gradient on a single task, where such perturbation is scaled with the inverse fisher information.
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
• perpendicular to gradient: Another extreme is to update the initial parameter θ0 perpendicular to the gradient direction.
Carmel, D. and Markovitch, S. Exploration strategies
for model-based learning in multi-agent systems. Autonomous Agents and Multi-Agent Systems, 1999.
• ε-greedy: A middle ground between these extremes is
a ε-greedy exploration strategy, where for ε of the time
the per-task exploration is done randomly, whereas for
the rest of the time (1 − ε) the gradient direction is
picked.
Duan, Y., Schulman, J., Chen, X., Bartlett, P. L., Sutskever,
I., and Abbeel, P. Rl2: Fast reinforcement learning via slow reinforcement learning. arXiv preprint
arXiv:1611.02779, 2016.
• Meta-learned Exploratory Operator where the innerupdate operator U is learned to optimality to a specific
class of tasks. See (Schmidhuber, 2006; 1987).
• Distraction-free Learning where the sign of the explorative log probability is flipped. This should discourage exploratory behavior. This could be useful in a
life-long learning agent during test time, where exploration is extremely costly.
We are planning on exploring these directions in the future.
6.2. Closing Remarks
In this paper, we considered the problem of exploration in
meta reinforcement learning. Two new algorithms were derived and their properties were analyzed. In particular, we
showed that these algorithms tend to learn more quickly
and explore more efficiently than existing algorithms. It
is likely that future work in this area will focus on metalearning a curiosity signal which is robust and transfers
across tasks, or learning an explicit exploration policy. Another interesting avenue for future work is learning intrinsic
rewards that communicate long-horizon goals, thus better
justifying exploratory behavior. Perhaps this will enable
meta agents which truly want to explore rather than being
forced to explore by mathematical trickery in their objectives.
References
Bacon, P. and Precup, D. The option-critic architecture. In
NIPS Deep RL Workshop, 2015.
Barto and Mahadevan. Recent advances in hierarchical reinforcement learning. Discrete Event Dynamic Systems,
2003.
Bellemare, M., Srinivasan, S., Ostrovski, G., Schaul, T.,
Saxton, D., and Munos, R. Unifying count based exploration and intrinsic motivation. NIPS, 2016.
Brafman, R. I. and Tennenholtz, M. R-max - a general
polynomial time algorithm for near-optimal reinforcement learning. Journal of Machine Learning Research,
2002.
Finn, C., Abbeel, P., and Levine, S. Model-agnostic metalearning for fast adaptation of deep networks. ICML,
2017.
Graziano, Vincent, Glasmachers, Tobias, Schaul, Tom,
Pape, Leo, Cuccu, Giuseppe, Leitner, Jurgen, and
Schmidhuber, Jürgen. Artificial Curiosity for Autonomous Space Exploration. Acta Futura, 1(1):41–
51, 2011. doi: 10.2420/AF04.2011.41. URL http:
//dx.doi.org/10.2420/AF04.2011.41.
Houthooft, R., Chen, X., Duan, Y., Schulman, J., De Turck,
F., and Abbeel, P. Vime: Variational information maximizing exploration. NIPS, 2016.
Kearns, M. J. and Singh, S. P. Near-optimal reinforcement
learning in polynomial time. Machine Learning, 2002.
Kolter, J. Z. and Ng, A. Y. Near-bayesian exploration in
polynomial time. ICML, 2009.
Kompella, Stollenga, Luciw, and Schmidhuber. Exploring
the predictable. Advances in Evolutionary Computing,
2002.
Koutnk, Jan, Cuccu, Giuseppe, Schmidhuber, Jrgen, and
Gomez, Faustino. Evolving large-scale neural networks
for vision-based reinforcement learning. GECCO 2013,
15:1061–1068, 2013.
Lillicrap, T., Hunt, J., Pritzel, A., Heess, N., Erez, T., Tassa,
Y., Silver, D., and Wierstra, D. Continuous control with
deep reinforcement learning. arXiv:1509.02971, 2015.
Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David,
Rusu, Andrei A, Veness, Joel, Bellemare, Marc G,
Graves, Alex, Riedmiller, Martin, Fidjeland, Andreas K,
Ostrovski, Georg, et al. Human-level control through
deep reinforcement learning. Nature, 518(7540):529–
533, 2015.
Mnih, Volodymyr, Badia, Adria Puigdomenech, Mirza,
Mehdi, Graves, Alex, Lillicrap, Timothy P, Harley,
Tim, Silver, David, and Kavukcuoglu, Koray. Asynchronous methods for deep reinforcement learning.
arXiv:1602.01783, 2016.
Ngo, Hung, Luciw, Matthew, Forster, Alexander, and
Schmidhuber, Juergen. Learning skills from play: Artificial curiosity on a Katana robot arm. Proceedings of the
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
Figure 7: Three heuristic metrics for exploration on Krazy World: Fraction of tile types visited during test time, number of
times killed at a death square during test time, and number of goal squares visited. We see that E-MAML is consistently the
most diligent algorithm at checking every tile type during test time. Beyond that, things are fuzzy with RL2 and MAML
both checking a majority of tile types at test time and E-RL2 being sporadic in this regard. As for the number of times
the death tile type was visited, we see that most algorithms start by dying in all three test episodes, and subsequently
decrease to between one and two by the time they have converged. As mentioned above, RL2 suffers from finding one
goal and exploiting it, whereas the other algorithms regularly explore to find more goals. For the most part, the exploratory
algorithms consistently deliver the best performance on these metrics. Performance on the death heuristic and the tile types
found heuristic seem to indicate the meta learners are learning how to do at least some system identification.
Figure 8: MAML on the right and E-MAML on the left. A look at the number of gradient steps at test time vs reward on
the Krazy World environment. Both MAML and E-MAML do not typically benefit from seeing more than one gradient
step at test time. Hence, we only perform one gradient step at test time for the experiments in this paper.
International Joint Conference on Neural Networks, pp.
10–15, 2012. ISSN 2161-4393. doi: 10.1109/IJCNN.
2012.6252824.
Schmidhuber. Evolutionary principles in self-referential
learning, or on learning how to learn: The meta-meta... hook. Diploma thesis, TUM, 1987.
Osband, I., Blundell, C., Pritzel, A., and Van Roy, B. Deep
exploration via bootstrapped dqn. NIPS, 2016.
Schmidhuber. Gdel machines: Fully self-referential optimal universal self-improvers. Artificial General Intelligence, 2006.
Ostrovski, G., Bellemare, M. G., Oord, A. v. d., and Munos,
R. Count-based exploration with neural density models.
arXiv:1703.01310, 2017.
Rusu, A. A., Rabinowitz, N. C., Desjardins, G., Soyer,
H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., and
Hadsell, R. Progressive neural networks. CoRR, vol.
abs/1606.04671, 2016.
Schmidhuber, J. Continual curiosity-driven skill acquisition from high-dimensional video inputs for humanoid
robots. Artificial Intelligence, 2015a.
Schmidhuber, J., Zhao, J., and Schraudolph, N. Reinforcement learning with self-modifying policies. Learning to
learn, Kluwer, 1997.
Some Considerations on Learning to Explore via Meta-Reinforcement Learning
Schmidhuber, Juergen. On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models. pp. 1–36, 2015b. ISSN 10450823.
doi: 1511.09249v1. URL http://arxiv.org/
abs/1511.09249.
Schmidhuber, Jürgen. A Possibility for Implementing Curiosity and Boredom in Model-Building Neural Controllers. Meyer, J.A. and Wilson, S.W. (eds) : From Animals to animats, pp. 222–227, 1991.
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and
Klimov, O. Proximal policy optimization algorithms.
arXiv:1707.06347, 2017.
Schulman, John, Levine, Sergey, Moritz, Philipp, Jordan,
Michael I., and Abbeel, Pieter. Trust region policy optimization. Arxiv preprint 1502.05477, 2015.
Silver, Yand, and Li. Lifelong machine learning systems:
Beyond learning algorithms. DAAAI Spring SymposiumTechnical Report, 2013., 2013.
Stadie, Bradly C., Levine, S., and Abbeel, P. Incentivizing
exploration in reinforcement learning with deep predictive models. arXiv:1507.00814, 2015.
Storck, Jan, Hochreiter, Sepp, and Schmidhuber, Jürgen.
Reinforcement driven information acquisition in nondeterministic environments. Proceedings of the International . . . , 2:159–164, 1995.
Sun, Yi, Gomez, Faustino, and Schmidhuber, Jürgen. Planning to be surprised: Optimal Bayesian exploration in
dynamic environments. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 6830
LNAI:41–51, 2011. ISSN 03029743. doi: 10.1007/
978-3-642-22887-2 5.
Tang, H., Houthooft, R., Foote, D., Stooke, A., Chen, X.,
Duan, Y., Schulman, J., De Turck, F., and Abbeel, P.
exploration: A study of count-based exploration for deep
reinforcement learning. arXiv:1611.04717, 2016.
Taylor and Stone. Transfer learning for reinforcement
learning domains: A survey. DAAAI Spring SymposiumTechnical Report, 2013., 2009.
Tessler, C. Givony, S., Zahavy, T., Mankowitz, D.J., and
Mannor, S. A deep hierarchical approach to lifelong
learning in minecraft. arXiv:1604.07255, 2016.
Thrun. Is learning the n-th thing any easier than learning
the first? NIPS, 1996.
Vezhnevets, A. S., Osindero, S., Schaul, T., Heess, N.,
Jaderberg, M., Silver, D., and Kavukcuoglu, K. Feudal
networks for hierarchical reinforcement learning. arXiv
preprint arXiv:1703.01161, 2017.
Wang, J. X., Kurth-Nelson, Z., Tirumala, D., Soyer, H.,
Leibo, J. Z., Munos, R., Blundell, C., Kumaran, D.,
and Botvinick, M. Learning to reinforcement learn.
arXiv:1611.05763, 2016.
Wiering and Schmidhuber. Hq-learning. Adaptive Behavior, 1997.
Williams, Ronald J. Simple statistical gradient-following
algorithms for connectionist reinforcement learning.
Machine Learning 8, 3-4 (1992), 229256, 1992.
| 2cs.AI
|
The Backtracking Survey Propagation Algorithm
for Solving Random K-SAT Problems
Raffaele Marino1 , Giorgio Parisi2 , and Federico Ricci-Tersenghi2
arXiv:1508.05117v4 [cs.CC] 6 Oct 2016
1
NORDITA and AlbaNova University Centre, Dept. of Computational Biology, KTH-Royal Institute
of Technology and Stockholm University, Roslagstullsbacken 23, SE-10691 Stockholm, Sweden,
email: [email protected]
2
Dipartimento di Fisica, Sapienza Università di Roma and Istituto Nazionale di Fisica Nucleare,
Sezione di Roma1 and CNR-Nanotec, Unità di Roma, P.le Aldo Moro 5, I-00185, Roma, Italy,
emails: [email protected], [email protected]
October 7, 2016
Abstract
Discrete combinatorial optimization plays a central role in many scientific disciplines, however for hard
problems we lack linear time algorithms that would allow us to solve very large instances. Moreover it is still
unclear what are the key features that make a discrete combinatorial optimization problem hard to solve. Here
we study random K-satisfiability problems with K = 3, 4 which are known to be very hard close to the SATUNSAT threshold, where problems stop having solutions. We show that the Backtracking Survey Propagation
algorithm, in a time practically linear in the problem size, is able to find solutions very close to the threshold, in
a region unreachable by any other algorithm. All solutions found have no frozen variables, thus supporting the
conjecture that only unfrozen solutions can be found in linear time, and that a problem becomes impossibile to
solve in linear time when all solutions contain frozen variables.
Optimization problems with discrete variables are
widespread among scientific disciplines and often
among the hardest to solve. The K-satisfiability (KSAT) problem is a combinatorial discrete optimization problem of N Boolean variables, x = {xi }i=1,N ,
submitted to M constraints. Each constraint, called
clause, is in the form of an OR logical operator of K
literals (variables and their negations): the problem is
solvable when there exists at least one configuration of
the variables, among the 2N possible ones, that satisfies all constraints. The K-SAT problem for K ≥ 3 is a
central problem in combinatorial optimization: it was
among the first problems shown to be N P -complete
[1, 2, 3] and is still very much studied. A growing
collaboration between theoretical computer scientists
and statistical physicists has focused on the random
K-SAT ensemble [4, 5], where each formula is generated by randomly choosing M = αN clauses of K literals. Formulas from this ensemble become extremely
hard to solve when the clause to variable ratio α grows
[6]: nevertheless, even in this region, the locally treelike structure of the factor graph [7], representing the
interaction network among variables, makes the random K-SAT ensemble a perfect candidate for analytic
computations. The study of random K-SAT problems
and of the related solving algorithms is likely to shed
light on the origin of the computational complexity
and to allow for the development of improved solving
algorithms.
Both numerical [8] and analytical [9, 10] evidence
suggest that a threshold phenomenon takes place in
random K-SAT ensembles: in the limit of very large
formulas, N → ∞, a typical formula has a solution for
α < αs (K), while it is unsatisfiable for α > αs (K). It
has been very recently proved in Ref. [11] that for K
large enough the SAT-UNSAT threshold αs (K) exists
in the N → ∞ limit and coincides with the prediction from the cavity method of statistical physics [12].
A widely accepted conjecture is that the SAT-UNSAT
threshold αs (K) exists for any value of K. Finding
solutions close to αs is very hard, and all known algorithms running in polynomial time fail to find solutions when α > αa , for some αa < αs . Actually, each
algorithm ALG has it own algorithmic threshold αaALG ,
such that the probability of finding a solution vanishes
for α > αaALG in the large N limit. For most algorithms
αaALG is well below αs . We define αa = maxALG αaALG
the threshold beyond which no polynomial-time algo1
rithm can find solutions. There are two main open
questions: to find improved algorithms having a larger
αaALG , and to understand what is the theoretical upper
bound αa . Here we present progress on both issues.
The best prediction about the SAT-UNSAT threshold comes from the cavity method [12, 13, 14, 15]: for
example, αs (K = 3) = 4.2667 [14] and αs (K = 4) = 9.931
[15]. Actually the statistical physics study of random
K-SAT ensembles also provides us with a very detailed
description of how the space of solutions changes when
α spans the whole SAT phase (0 ≤ α ≤ αs ). Let us
consider typical formulas in the large N limit and the
vast majority of solutions in these formulas (i.e. typical solutions), we know that, at low enough α values,
the set of solutions is connected, so that they form a
single cluster. In SAT problems we say 2 solutions are
neighbors if they differ in the assignment of just one
variable; in other problems (e.g. XOR-SAT [16]) this
definition of neighbor needs to be relaxed, because a
pair of solutions differing in just one variable are not
allowed by the model definition. As long as the notion
of neighborhood is relaxed to Hamming distances o(N )
all the picture of the solution space based on statistical
physics remains unaltered.
As α increases, not only the number of solutions
decreases, but at αd the random K-SAT ensemble undergoes a phase transition: the space of solutions shatters into an exponentially large (in the problem size N )
number of clusters; two solutions belonging to different
clusters have a Hamming distance O(N ). If we define
the energy function E(x) as the number of unsatisfied
clauses in configuration x, it has been found [12] that
for α > αd the energy E(x) has exponentially many (in
N ) local minima of positive energy, which may trap algorithms that look for solutions by energy relaxation
(e.g. Monte Carlo simulated annealing).
Further increasing α, each cluster loses solutions
and shrinks, but the most relevant change is in the
number of clusters. The cavity method allows us to
count clusters of solutions as a function of the number of solutions they contain [17]: using this very detailed description several other phase transitions have
been identified [18, 15]. For example, there is a value
αc where a condensation phase transition takes place,
such that for α > αc the vast majority of solutions
belong to a sub-exponential number of clusters, leading to effective long-range correlations among variables
in typical solutions, which are hard to approximate
by any algorithm with a finite horizon. In general
αd ≤ αc ≤ αs holds. Most of the above picture of
the solution space has been proven rigorously in the
large K limit [19, 20].
Moving to the algorithmic side, a very interesting
question is whether such a rich structure of the solution
space affects the performance of searching algorithms.
While clustering at αd may have some impact on algorithms that sample solutions uniformly [21], many algorithms exist that can find at least one solution with
α > αd [12, 22, 23].
A solid conjecture is that the hardness of a formula is
related to the existence of a subset of highly correlated
variables, which are very hard to assign correctly altogether; the worst case being a subset of variables that
can have a unique assignment. This concept was introduced with the name of backbone in Ref. [24]. The
same concept applied to solutions within a single cluster lead to the definition of frozen variables (within
a cluster) as those variables taking the same value in
all solutions of the cluster [25]. It has been proven in
Ref. [26] that the fraction of frozen variables in a cluster is either zero or lower bounded by (αe2 )−1/(K−2) ;
in the latter case the cluster is called frozen.
According to the above conjecture, finding a solution in a frozen cluster is hard (in practice it should
require a time growing exponentially with N ). So the
smartest algorithm running in polynomial time should
search for unfrozen clusters as long as they exist. Unfortunately counting unfrozen clusters is not an easy
job, and indeed a large deviation analysis of their number has been achieved only very recently [27] for a different and simpler problem (bicoloring random regular
hypergraphs). For random K-SAT only partial results
are known, that can be stated in terms of two thresholds: for α > αr (rigidity) typical solutions are in
frozen cluster (but a minority of solutions may still
be unfrozen), while for α > αf (freezing) all solutions are frozen. It has been rigorously proven [28, 29]
that αf < αs holds strictly for K > 8. For small K,
which is the interesting case for benchmarking solving algorithms, we know αr = 9.883(15) for K = 4
from the cavity method [15], while for K = 3 the estimate αf = 4.254(9) comes from exhaustive enumerations in small formulas (N ≤ 100) [30] and is likely
to be affected by strong finite size effects. In general
αd ≤ αr ≤ αf ≤ αs holds.
The conjecture above implies that no polynomial
time algorithm can solve problems with α ≥ αf , but
also finding solutions close to the rigidity threshold αr
is expected to be very hard, given that unfrozen solutions becomes a tiny minority. And this is indeed what
happens for all known algorithms. Since we are interested in solving very large problems we only consider
algorithms whose running time scales almost linearly
with N and we measure performance of each algorithm
in terms of its algorithmic threshold αaALG .
Solving algorithms for random K-SAT problems can
be roughly classified in two main categories: algorithms that search for a solution by performing a bi2
Fraction of solved instances by BSP
ased random walk in the space of configurations and
algorithms that try to build the solutions by assigning variables, according to some estimated marginals.
WalkSat [31], focused Metropolis search (FMS) [22]
and ASAT [23] belong to the former category; while in
the latter category we find Belief Propagation guided
Decimation (BPD) [21] and Survey Inspired Decimation (SID) [32]. All these algorithms are rather effective in finding solutions to random K-SAT problems:
e.g. for K = 4 we have αaBPD = 9.05, αaFMS ' 9.55
and αaSID ' 9.73 to be compared with a much lower
algorithmic threshold αaGUC = 5.54 achieved by Generalized Unit Clause, the best algorithm whose range
of convergence to a solution can be proven rigorously
[33]. Among the efficient algorithms above, only BPD
can be solved analytically [21] to find the algorithmic
threshold αaBPD ; for the others we are forced to run
extensive numerical simulations to measure αaALG .
At present the algorithm achieving the best performance on several constraint satisfaction problems is
SID, which has been successfully applied to the random K-SAT problem [12] and to the coloring problem
[34]. The statistical properties of the SID algorithm for
K = 3 have been studied in details in Refs. [35, 32]. Numerical experiments on random 3-SAT problems with
a large number of variables, up to N = 3 × 105 , show
that in a time that is approximately linear in N the
SID algorithm finds solutions up to αaSID ' 4.2525
[35], that is definitely smaller, although very close to,
αs (K = 3) = 4.2667. In the region αaSID < α < αs the
problem is satisfiable for large N , but at present no
algorithm can find solutions there.
To fill this gap we study a new algorithm for finding solutions to random K-SAT problems, the Backtracking Survey Propagation (BSP) algorithm. This
algorithm (fully explained in the Methods section) is
based, as SID, on the survey propagation (SP) equations derived within the cavity method [12, 35, 32] that
provide an estimate on the total number of clusters
Nclus = exp(Σ). The BSP algorithm, like SID, aims at
assigning gradually the variables such as to keep the
complexity Σ as large as possibile, i.e. trying not to
kill too many clusters [35]. While in SID each variable is assigned only once, in BSP we allow unsetting
variables already assigned such as to backtrack on previous non-optimal choices. In BSP the r parameter is
the ratio between the number of backtracking moves
(unsetting one variable) and the number of decimation
moves (assigning one variable). r < 1 must hold and
for r = 0 we recover the SID algorithm. The running
time scales as N/(1 − r), with a slight overhead for
maintaining the data structures, making the running
time effectively linear in N for any r < 1.
The idea supporting backtracking [36] is that a
1
0.8
αr
0.6
0.4
0.2
0
9.5
4-SAT
r=0.0 N=5e3
r=0.5 N=5e3
r=0.9 N=5e3
r=0.0 N=5e4
r=0.5 N=5e4
r=0.9 N=5e4
9.55
9.6
9.65
9.7
α
9.75
9.8
9.85
9.9
Figure 1: Fraction of random 4-SAT instances
solved by BSP as a function of the constraints per
variable ratio α. The average is computed over 100
instances with N = 5 000 (solid symbols) and N =
50 000 (empty symbols) variables. The vertical line is
the best estimate for αr and the shaded region is the
statistical error on this estimate. For each instance, the
algorithm has been run once; on instances not solved
on the first run, a second run rarely (< 1%) finds a
solution. The plot shows that the backtracking (r > 0)
definitely makes the BSP algorithm more efficient in
finding solutions. Although data become sharper by
increasing the problem size N , a good estimation of the
algorithmic threshold from these datasets is unfeasible.
choice made at the beginning of the decimation process, when most of the variables are unassigned, may
turn to be suboptimal later on; if we re-assign a variable that is no longer consistent with the current best
estimate of its marginal probability, we may get a better satisfying configuration. We do not expect the
backtracking to be essential when correlations between
variables are short ranged, but approaching αs we
know that correlations become long ranged and thus
the assignment of a single variable may affect a huge
number of other variables: this is the situation when
we expect the backtracking to be crucial.
This idea may look similar in spirit to the survey propagation reinforcement (SPR) algorithm [37],
where variables are allowed to change their most likely
value during the run, but in practice BSP works much
better. In SPR, once reinforcement fields are large,
the re-assignment of any variable becomes unfeasible,
while in BSP variables can be re-assigned to better values until the very end, and this is a major advantage.
3
Results
0.035
(a)
0.03
Probability of finding a SAT assignment The
standard way to study the performance of a solving
algorithm is to measure the fraction of instances it can
solve as a function of α. We show in Fig. 1 such a fraction for BSP run with three values of the r parameter
(r = 0, 0.5 and 0.9) on random 4-SAT problems of two
different sizes (N = 5 000 and N = 50 000). The probability of finding a solution increases both with r and
N , but an extrapolation to the large N limit of these
data is unlikely to provide a reliable estimation of the
algorithmic threshold αaBSP .
In each plot having α on the abscissa, the right end
of the plot coincides with the best estimate of αs , in
order to provide an immediate indication of how close
to the SAT-UNSAT threshold the algorithm can work.
Σres/Nres
0.025
αr
0.02
0.015
0.01
4-SAT r=0.9
N=1e4
N=2e4
N=5e4
N=5e5
0.005
0
9
9.1
9.2
9.3
9.4 9.5
α
9.6
9.7
9.8
9.9
0.035
(b)
0.03
0.025
Σres/Nres
Order parameter and algorithmic threshold In
order to obtain a reliable estimate of αaBSP we look for
an order parameter vanishing at αaBSP and having very
little finite size effects. We identify this order parameter with the quantity Σres /Nres , where Σres and Nres
are respectively the complexity (i.e. log of number of
clusters) and the number of unassigned variables in the
residual formula. As explained in Methods, BSP assigns and re-assigns variables, thus modifying the formula, until the formula simplifies enough that the SP
fixed point has only null messages: the residual formula
is defined as the last formula with non-null SP fixed
point messages. We have experimentally observed that
the BSP algorithm (as the SID one [35]) can simplify
the formula enough to reach the trivial SP fixed point
only if the complexity Σ remains strictly positive during the whole decimation process. In other words, on
every run where Σ becomes very close to zero or negative, SP stops converging or a contradiction is found.
This may happen either because the original problem
was unsatisfiable or because the algorithm made some
wrong assignments incompatible with the few available solutions. Thanks to the above observation we
have that Σres ≥ 0 and thus a null value for the mean
residual complexity signals that the BSP algorithm is
not able to find any solution, and thus provides a valid
estimate for the algorithmic threshold αaBSP . From the
statistical physics solution to random K-SAT problems
we expect Σres to vanish linearly in α.
As we see in panel (a) of Fig. 2 the mean value
of the intensive mean residual complexity Σres /Nres is
practically size-independent and a linear fit provides
a very good data interpolation: tiny finite size effects
are visible in the largest N datasets only close to the
dataset right end. The linear extrapolation predicts
αaBSP ≈ 9.9 (for K = 4 and r = 0.9), which is slightly
αr
0.02
0.015
0.01
4-SAT
r=0.0 N=5e3
r=0.5 N=5e3
r=0.0 N=5e4
r=0.5 N=5e4
0.005
0
9
9.1
9.2
9.3
9.4 9.5
α
9.6
9.7
9.8
9.9
Figure 2: BSP algorithmic threshold on random
4-SAT problems. The residual complexity per variable, Σres /Nres , goes to zero at the algorithmic threshold αaBSP . (a) The very small finite size effects, mostly
producing a slight downward curvature at the right
end, allow for a very reliable estimate of αaBSP via a
linear fit. For random 4-SAT problems solved by BSP
with r = 0.9 we get αaBSP ≈ 9.9, slightly beyond the
rigidity threshold αr = 9.883(15), marked by a vertical
line (the shaded area being its statistical error). (b)
The same linear extrapolation holds for other values of
r (red dotted line for r = 0.5 and blue dashed line for
r = 0). The black line is the fit to r = 0.9 data shown
in panel (a). SID without backtracking (r = 0) has a
much lower algorithmic threshold, αaSID ≈ 9.83. Error
bars in both panels are the standard error on the mean
(sem).
above the rigidity threshold αr = 9.883(15) computed
in Ref. [15] and reported in the plot with a shaded region corresponding to its statistical error (the value of
αf in this case is not known, but αaBSP < αf ≤ αs should
hold). Although for the finite sizes studied no solution
has been found beyond αr , Fig. 2 suggests that in the
large N limit BSP may be able to find solutions in a
4
region of α where the majority of solutions is in frozen
clusters and thus very hard to find. We show below
that BSP actually finds solutions in atypical unfrozen
clusters, as it has been observed for some smart algorithms solving other kind of constraint satisfaction
problems [38, 39].
The effectiveness of the backtracking can be appreciated in panel (b) of Fig. 2, where the order parameter
Σres /Nres is shown for r = 0 and r = 0.5, together with
linear fits to these datasets and to the r = 0.9 dataset
(black line). We observe that the algorithmic threshold for BSP is much larger (on the scale measuring
the relative distance from the SAT-UNSAT threshold)
that the one for SID (i.e. r = 0 dataset).
For random 3-SAT the algorithmic threshold of BSP,
run with r = 0.9, practically coincide with the SATUNSAT threshold αs (see Fig. 3), thus providing a
strong evidence that BSP can find solutions in the entire SAT phase. The estimate for the freezing threshold αf = 4.254(9) obtained in Ref. [30] from N ≤ 100
data is likely to be too small and affected by strong finite size effects, given that all solutions found by BSP
for N = 106 are unfrozen, even beyond the estimated
αf . Moreover we have estimated αr = 4.2635(10)
improving the data of Ref. [15] and the inequality
αr ≤ αf ≤ αs makes the above estimate for αf not
very meaningful.
Computational complexity As explained in
Methods, the BSP algorithm performs f −1 (1 − r)−1
steps roughly, where at each step f N variables are either assigned [with prob. 1/(1 + r)] or released [with
prob. r/(1 + r)]. At the beginning of each step, the algorithm solves the SP equations with a mean number
η of iterations. The average η is computed only on instances where SP always converges, as is usually done
for incomplete algorithms (on the remaining problems
the number of iterations reaches the upper limit set
by the user, and then BSP exit, returning failure).
Fig. 4 shows that η is actually a small number changing
mildly with α and N both for K = 3 and K = 4. The
main change that we observe is in the fluctuations of η
that become much larger approaching αs . We expect
η to eventually grow as O(log(N )), but for the sizes
studied we do not observe such a growth.
After convergence to a fixed point, the BSP algorithm just need to sort local marginals, thus the total
number of elementary operations to solve an instance
grows as f −1 (1 − r)−1 (a1 ηN + a2 N log N ), where a1
and a2 are constants. Moreover, given that the sorting
of local marginals does not need to be strict (i.e. a partial sorting [40] running in O(N ) time can be enough),
we have that in practice the algorithm runs in a time
almost linear in the problem size N .
Whitening procedure Given that the BSP algo-
0.004
0.0035
0.003
Σres/Nres
0.0025
αf
αr
0.002
0.0015
0.001
0.0005
3-SAT r=0.9
N=1e6
0
4.225 4.23 4.235 4.24 4.245 4.25 4.255 4.26 4.265
α
Figure 3: BSP algorithmic threshold on random
3-SAT problems. Same as Fig. 2 for K = 3. The
estimate for the freezing threshold αf = 4.254(9) measured on small problems in Ref. [30] is not very meaningful, given our new estimate for the rigidity threshold
αr = 4.2635(10), and the observation that all solutions
found by BSP are not frozen. Shaded areas are the
statistical uncertainties on the thresholds. A linear fit
to the residual complexity (brown line) extrapolates
to zero slightly beyond the SAT-UNSAT threshold, at
αaBSP ≈ 4.268, strongly suggesting BSP can find solutions in the entire SAT phase for K = 3 in the large
N limit. The black line is a linear fit vanishing at αs .
Error bars are sem.
13
(a)
3-SAT
r=0.9 N=1e6
η
11
9
7
α
4.23
4.24
4.25
4.26
7
(b)
4-SAT
r=0.9 N=5e4
r=0.9 N=5e3
η
6
5
4
3
α
9.65
9.7
9.75
9.8
9.85
9.9
Figure 4: BSP convergence time. The mean number η of iterations to reach a fixed point of SP equations
grows very mildly with α and N , both for K = 3 (a)
and K = 4 (b). Error bars are standard deviations.
rithm is able to find solutions even very close to the
rigidity threshold αr , it is natural to check whether
these solutions have frozen variables or not. We concentrate on solutions found for random 3-SAT problems with N = 106 , since the large size of these prob5
lems makes the analysis very clean.
On each solution found we run the whitening procedure (first introduced in [41] and deeply discussed in
[42, 26]), that identifies frozen variables by assigning
the joker state ? to unfrozen (white) variables, i.e. variables that can take more than one value without violating any clause and thus keeping the formula satisfied.
At each step of the whitening procedure, a variables is
considered unfrozen (and thus assigned to ?) if it belongs only to clauses which either involve a ? variable
or are satisfied by another variable. The procedure is
continued until all variables are ? or a fixed point is
reached: non-? variables at the fixed point correspond
to frozen variables in the starting solution.
We uncover that all solutions found by BSP are converted to all-? by running the whitening procedure,
thus showing that solutions found by BSP have no
frozen variables. This is somehow expected, according
to the conjecture discussed in the Introduction: finding
solutions in a frozen cluster would take an exponential
time, and so the BSP algorithm actually finds solutions at very large α values by smartly focusing on the
sub-dominant unfrozen clusters.
The whitening procedure leads to a relaxation of the
number of non-? variables as a function of the number of iterations t that follows a two steps relaxation
process [25] with an evident plateau, see panel (a) in
Fig. 5, that becomes longer increasing α towards the algorithmic threshold. The time for leaving the plateau,
scales as the time τ (c) for reaching a fraction c on non? variables (with c smaller than the plateau value).
The latter has large fluctuations from solution to solution, as shown in panel (b) of Fig. 5 for c = 0.4 (very
similar, but shifted, histograms are obtained for other
c values). However, after leaving the plateau, the dynamics of the whitening procedure is the same for each
solution. Indeed plotting the mean fraction of non-?
variables as a function of the time to reach the all? configuration, τ (0) − t, we see that fluctuations are
strongly suppressed and the relaxation is the same for
each solution (see panel (c) in Fig. 5).
Critical exponent for the whitening time divergence In order to quantify the increase of the
whitening time approaching the algorithmic threshold,
and inspired by critical phenomena, we check for a
power law divergence as a function of (αaBSP −α) or Σres ,
which are linearly related. In Fig.6 we plot in a double
logarithmic scale the mean whitening time τ (c) as a
function of the residual complexity Σres , for different
choices of the fraction c of non-? variables defining the
whitening time. Data points are fitted via the power
law τ (c) = A(c)+B(c)Σ−ν
res , where the critical exponent
ν is the same for all the c values. Joint interpolations
return the following best estimates for the critical ex-
Mean fraction of non-* variables
3-SAT
α=4.262
α=4.254
α=4.240
(a)
1
0.8
0.6
0.4
0.2
0
0
5
10
15
20
t
25
30
0.8
35
40
45
3-SAT
α=4.262
α=4.254
α=4.240
(b)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
5
10
15
20 25
τ(0.4)
30
35
40
45
1
Mean fraction of non-* variables
(c)
0.8
0.6
0.4
0.2
0
3-SAT
α=4.262
α=4.254
α=4.240
0
5
10
15
20 25
τ(0)-t
30
35
40
45
Figure 5: Whitening random 3-SAT solutions.
(a) The mean fraction of non-? variables during the
whitening procedure applied to all solutions found by
the BSP algorithm goes to zero, following a two steps
process. The relaxation time grows increasing α towards the algorithmic threshold. The horizontal line
is the SP prediction for the fraction of frozen variables
in typical solutions at α = 4.262 and the comparison
with the data shows that solutions found by BSP are
atypical. (b) Histograms of the whitening times, defined as the number of iterations required to reach a
fraction 0.4 of non-? variables. Increasing α both the
mean value and the variance of the whitening times
grow. (c) Averaging the fraction of non-? variables
at fixed τ (0)−t, i.e. fixing the time to the all-? fixed
point, we get much smaller errors than in panel (a),
suggesting that the whitening procedure is practically
solution-independent once the plateau is left. In all the
panels, error bars are sem.
6
100
son for its superiority is the fact that variables can be
re-assigned at any time during the run, even at the
very end. In other solving algorithms that may look
similar, as e.g. survey propagation reinforcement [37],
re-assignment of variables actually takes place mostly
at the beginning of the run, and this is far less efficient
in hard problems. Even doing a lot of helpful backtracking, the BSP running time is still O(N log N ) in
the worst case, and thanks to this it can be used on
very large problems with millions of constraints.
For K = 3 the BSP algorithm finds solutions practically up to the SAT-UNSAT threshold αs , while for
K = 4 a tiny gap to the SAT-UNSAT threshold still
remains, but the algorithmic threshold αaBSP seems to
be located beyond the rigidity threshold αr in the large
N limit. Beating the rigidity threshold, i.e. finding solutions in a region where the majority of solutions belongs to clusters with frozen variables, is hard, but not
impossible (while going beyond αf should be impossible). Indeed, even under the assumption that finding
frozen solutions takes an exponential time in N , very
smart polynomial time algorithms can look for a solution in the sub-dominant unfrozen clusters [38, 39].
BSP belongs to this category, as we have shown that
all solutions found by BSP have no frozen variables.
One of the main questions we tried to answer with
our extensive numerical simulations is whether BSP is
reaching (or approaching closely) the ultimate threshold αa for polynomial time algorithms solving large
random K-SAT problems. Under the assumption that
frozen solutions cannot be found in polynomial time,
such an algorithmic threshold αa would coincide with
the freezing transition at αf (i.e. when the last unfrozen solution disappears). Unfortunately for random
K-SAT the location of αf is not known with enough
precision to allow us to reach a definite answer to this
question. It would be very interesting to run BSP
on random hypergraph bicoloring problems, where the
threshold values are known [43, 44] and a very recent
work has shown that the large deviation function for
the number of unfrozen clusters can be computed [27].
It is worth noticing that the BSP algorithm is easy
to parallelize, since most of the operations are local
and do not require any strong centralized control. Obviously the effectiveness of a parallel version of the algorithm would largely depend on the topology of the
factor graph representing the specific problem: if the
factor graph is an expander, then splitting the problem on several cores may require too much inter-core
bandwidth, but in problems having a natural hierarchical structure the parallelization may lead to further
performance improvements.
The backtracking introduced in the BSP algorithm
helps a lot in correcting errors made during the partial
τ
3-SAT τ(0)
3-SAT τ(0.4)
3-SAT τ(0.6)
4-SAT τ(0)
4-SAT τ(0.4)
4-SAT τ(0.6)
10
100
Σres
1000
Figure 6: Critical exponent for the whitening
time divergence. The whitening time τ (c), defined
as the mean time needed to reach a fraction c of non-?
variables in the whitening procedure, is plotted in a
double logarithmic scale as a function of Σres for random 3-SAT problems with N = 106 (upper dataset)
and random 4-SAT problems with N = 5 × 104 (lower
dataset). The whitening time measured with different
c values seems to diverge at the algorithmic threshold, where the residual complexity Σres vanishes. The
lines are power law fits with exponent ν = 0.281(6) for
K = 3 and ν = 0.269(5) for K = 4. Error bars are sem.
ponent: ν = 0.281(6) for K = 3 and ν = 0.269(5) for
K = 4, where the uncertainties are only fitting errors.
The two estimate turn out to be compatible within errors, thus suggesting a sort of universality for the critical behavior close to the algorithmic threshold αaBSP .
Nonetheless a word of caution is needed since the solutions we are using as starting points for the whitening
procedure are atypical solutions (otherwise they would
likely contain frozen variables and would not flow to
the all-? configuration under the whitening procedure).
So, while finding universal critical properties in a dynamical process is definitely a good news, how to relate
it to the behavior of the same process on typical solutions it is not obvious (and indeed for the whitening
process starting from typical solutions one would expect the naive mean field exponent ν = 1/2, which is
much larger than the one we are finding).
Discussion
We have studied the Backtracking Survey Propagation (BSP) algorithm for finding solutions in very large
random K-SAT problems and provided numerical evidence that it works much better than any previously
available algorithm. That is, BSP has the largest algorithmic threshold known at present. The main rea7
assignment of variables and this allows the BSP algorithm to reach solutions at large α values. Clearly we
pay the price that a too frequent backtracking makes
the algorithm slower, but it seems worth paying such
a price to approach the SAT-UNSAT threshold closer
than any other algorithm.
A natural direction to improve this class of algorithms would be to used biased marginals focusing on
solutions which are easier to be reached by the algorithm itself. For example in the region α > αr the
measure is concentrated on solutions with frozen variables, but these can not be really reached by the algorithm. The backtracking thus intervenes and corrects
the partial assignment until a solution with unfrozen
variables is found by chance. If the marginals could be
computed from a new biased measure which is concentrated on the unfrozen clusters, this could make the
algorithm go immediately in the right direction and
much less backtracking would be hopefully needed.
number of clusters can be estimated via the complexity
X
X
Σ = log Nclus =
Σi +
(1 − Ka )Σa ,
(4)
a
i
Σa = log 1−
Y
m?j→a ,
Σi = log(1− πi+ πi− ) (5)
j∈∂a
with
πi±
=1−
Y
(1 − m̂?b→i ) ,
(6)
b∈∂i±
where Ka is the length of clause a (initially Ka = K)
and ∂i+ (resp. ∂i− ) is the set of clauses satisfied by
setting xi = 1 (resp. xi = −1). The SP fixed point
messages also provide information about the fraction
of clusters where variable xi is forced to be positive
(wi+ ), negative (wi− ) or not forced at all (1 − wi+ − wi− )
wi± =
πi± (1 − πi∓ )
.
1 − πi+ πi−
(7)
The SID algorithm then proceed by assigning variables
(decimation step). According to SP equations,
Methods
assigning a variable xi to its most probable value (i.e.,
+
−
Survey Inspired Decimation (SID) A detailed setting xi = 1 if wi > wi and viceversa), the number
description of the SID algorithm can be found in of clusters gets multiplied by a factor, called bias
Refs. [12, 13, 32]. The SID algorithm is based on the
bi = 1 − min(wi+ , wi− ) .
(8)
survey propagation (SP) equations derived by the cavity method [12, 13], that can be written in a compact
With the aim of decreasing the lesser the number of
way as
cluster and thus keeping the largest the number of soY
m̂a→i =
mj→a ,
(1) lutions in each decimation step, SID assigns/decimate
variables with the largest bi values. In order to keep
j∈∂a \i
the algorithm efficient, at each step of decimation a
−
+
πia
(1 − πia
)
mi→a =
(2) small fraction f of variables is assigned, such that in
+ − ,
1 − πia
πia
O(log N ) steps of decimation a solution can be found.
Y
±
After each step of decimation, the SP equations are
with πia = 1 −
(1 − m̂b→i ) ,
(3)
±
solved again on the subproblem, which is obtained by
b∈∂ia
removing satisfied clauses and by reducing clauses con+
where ∂a is the set of variables in clause a, and ∂ia taining a false literal (unless a zero-length clause is
−
(resp. ∂ia
) is the set of clauses containing xi , excluding generated, and in that case the algorithm returns a
a itself, satisfied (resp. not satisfied) when the variable failure). The complexity and the biases are updated
according to the new fixed point messages, and a new
xi is assigned to satisfy clause a.
The interpretation of the SP equations is as follows: decimation step is performed.
m̂a→i represents the fraction of clusters where clause a
The main idea of the SID algorithm is that fixing
is satisfied solely by variable xi (that is, xi is frozen by variables which are almost certain to their most probclause a), while mi→a is the fraction of clusters where able value, one can reduce the size of the problem withxi is frozen to an assignment not satisfying clause a.
out reducing too much the number of solutions. The
The SP equations impose 2KM self-consistency con- evolution of the complexity Σ during the SID algoditions on the 2KM variables {mi→a , m̂a→i } living on rithm can be very informative [35]. Indeed it is found
the edges of the factor graph [7], that are solved in that, if Σ becomes too small or negative, the SID alan iterative way, leading to a message passing algo- gorithm is likely to fail, either because the iterative
rithm (MPA) [4], where outgoing messages from a fac- method for solving the SP equations no longer contor graph node (variable or clause) are functions of verges to a fixed point or because a contradiction is
the incoming messages. Once the MPA reaches a fixed generated by assigning variables. In these cases the
point {m?i→a , m̂?a→i } that solves the SP equations, the SID algorithm returns a failure. On the contrary, if
8
are reassigned on average 1/(1 − r) times each before
the BSP algorithm reaches the end, but its complexity
remains at most O(N log N ) in the problem size.
The BSP algorithm can stop for the same reasons the
SID algorithm does: either the SP equations can not
be solved iteratively or the generated subproblem has
a contradiction. Both cases happen when the complexity Σ becomes too small or negative. On the contrary if
the complexity remain always positive the BSP eventually generate a subproblem where all SP messages
are null and on this subproblem WalkSat is called.
Σ always remains well positive, the SID algorithm reduces so much the problem, that eventually a trivial
SP fixed point, m?i→a = m̂?a→i = 0, is reached. This
is a strong hint that the remaining subproblem is easy
and the SID algorithm tries to solve it by WalkSat [31].
A careful analysis of the SID algorithm for random
3-SAT problems of size N = O(105 ) shows that the algorithmic threshold achievable by SID is αaSID = 4.2525
[35], which is close, but definitely smaller than the
SAT-UNSAT threshold αs = 4.2667.
The running time of the SID algorithm experimentally measured is O(N log(N )) [32].
Backtracking Survey Propagation (BSP) Willing to improve the SID algorithm to find solutions also
in the region αaSID < α < αs , one has to change the
way variables are assigned. The fact the SID algorithm assigns each variable only once is clearly a strong
limitation, especially in a situation where correlations
between variables becomes extremely strong and longranged. In difficult problems it can easily happen that
one realizes that a variable is taking the wrong value
only after having assigned some of its neighbours variables. However the SID algorithm is not able to solve
this kind of frustrating situations.
The BSP algorithms [36] tries to solve this kind
of problematic situations by introducing a new backtracking step, where a variable already assigned can
be released and eventually re-assigned in a future decimation step. It is not difficult to understand when it is
worth releasing a variable. The bias bi in terms of the
SP fixed point messages {m̂?a→i }a∈∂i arriving in i can
be computed also for a variable xi already assigned:
if the bias bi , that was large at the time the variable
xi was assigned, gets strongly reduces by the effect of
assigning other variables, then it is likely that releasing the variable xi may be beneficial in the search for
a solution. So both the variables to be fixed in the
decimation step and the variables to be released in the
backtracking step are chosen according to their biases
bi : the variables to be fixed have the largest biases and
the variables to be released have the smallest biases.
The BSP algorithm then proceeds similarly to SID,
by alternating the iterative solution to the SP equations and a step of decimation or backtracking on a
fraction f of variables in order to keep the algorithm
efficient (in all our numerical experiments we have used
f = 10−3 ). The choice between a decimation or a
backtracking step is taken according to a stochastic
rule (unless there are no variables to unset), where
the parameter r ∈ [0, 1) represents the ratio between
backtracking steps to decimation steps. Obviously for
r = 0 we recover the SID algorithm, since no backtracking step is ever done. Increasing r the algorithm
becomes slower by a factor 1/(1 − r), because variables
Data Availability Statement
The numerical codes used in this study and the data
that support the findings are available from the corresponding author upon request.
References
[1] S. A. Cook, The complexity of theorem proving procedures, Proc. 3rd Ann. ACM Symp. on Theory of
Computing, Assoc. Comput. Mach., New York, pp.
151-158 (1971).
[2] M. Garey and D. S. Johnson, Computers and
Intractability; A guide to the theory of NPcompleteness , Freeman, San Francisco (1979).
[3] C. H. Papadimitriou, Computational Complexity,
Addison-Wesley (1994).
[4] M. Mézard and A. Montanari, Information,
Physics, and Computation, Oxford University Press
(2009).
[5] C. Moore and S. Mertens, The Nature of Computation, Oxford University Press (2011).
[6] S. A. Cook and D. G. Mitchell, Finding Hard Instances of the Satisfiability Problem: A Survey, In:
Satisfiability Problem: Theory and Applications.
J. Du, D. Gu and P. Pardalos (Eds). DIMACS Series
in Discrete Mathematics and Theoretical Computer
Science, Volume 35 (1997).
[7] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger,
Factor Graphs and the Sum-Product Algorithm,
IEEE Trans. Infor. Theory 47, 498-519 (2001).
[8] S. Kirkpatrick and B. Selman, Critical Behaviour
in the satisfiability of random Boolean expressions,
Science 264, 1297-1301 (1994).
[9] O. Dubois, Y. Boufkhad, J. Mandler, Typical random 3-SAT formulae and the satisfiability threshold,
9
in Proc. 11th ACM-SIAM Symp. on Discrete Algo- [22] S. Sakari, M. Alava, and P. Orponen, Focused local search for random 3-satisfiability, J. Stat. Mech.
rithms, 124-126 (San Francisco, CA, 2000).
P06006 (2005).
[10] O. Dubois, R. Monasson, B. Selman and
R. Zecchina (Eds.), Phase Transitions in Combina- [23] J. Ardelius and E. Aurell, Behavior of heuristics
torial Problems, Theoret. Comp. Sci. 265 (2001).
on large and hard satisfiability problems, Phys. Rev.
E 74, 037702 (2006).
[11] J. Ding, A. Sly, and N. Sun, Proof of the satisfiability conjecture for large k, Proceedings of the [24] R. Monasson, R. Zecchina, S. Kirkpatrick, B. SelForty-Seventh Annual ACM on Symposium on Theman and L. Troyansky, Determining computational
ory of Computing (2015).
complexity from characteristic phase transitions,
Nature 400, 133-137 (1999).
[12] M. Mézard, G. Parisi and R. Zecchina, Analytic and algorithmic solution of random satisfiabil[25] G. Semerjian, On the freezing of variables in ranity problems, Science 297, 812-815 (2002).
dom constraint satisfaction problem, J. Stat. Phys.
130, 251-293 (2008).
[13] M. Mézard and R. Zecchina, The random Ksatisfiability problem: from an analytic solution to
an efficient algorithm, Phys. Rev. E 66, 056126 [26] E. Maneva, E. Mossel, and M. J. Wainwright, A
new look at survey propagation and its generaliza(2002).
tions, Journal of the ACM (JACM) 54, 17 (2007).
[14] S. Mertens, M. Mézard and R. Zecchina, Threshold Values of Random K-SAT from the Cavity [27] A. Braunstein, L. Dall’Asta, G. Semerjian, and
L. Zdeborova, The large deviations of the whitening
Method, Random Struct. Alg. 28, 340-373 (2006).
process in random constraint satisfaction problems,
[15] A. Montanari, F. Ricci-Tersenghi and G. SemerJ. Stat. Mech. P053401 (2016).
jian, Clusters of solutions and replica symmetry
breaking in random k-satisfiability, J. Stat. Mech. [28] D. Achlioptas and F. Ricci-Tersenghi, On the
Solution-Space Geometry of Random Constraint
P04004 (2008).
Satisfaction Problems, STOC ’06: Proceedings of
[16] M. Mézard, F. Ricci-Tersenghi, and R. Zecchina,
the 38th annual ACM symposium on Theory of
Two solutions to diluted p-spin models and XORComputing (Seattle, WA, USA), 130-139 (2006).
SAT problems, J. Stat. Phys. 111, 505-533 (2003).
[29] D. Achlioptas and F. Ricci-Tersenghi, Random
[17] F. Krzakala, A. Montanari, F. Ricci-Tersenghi,
Formulas Have Frozen Variables, SIAM J. Comput.
G. Semerjian, and L. Zdeborova, Gibbs States and
39, 260-280 (2009).
the Set of Solutions of Random Constraint Satisfaction Problems, PNAS 104, 10318-10323 (2007).
[30] J. Ardelius and L. Zdeborova, Exhaustive enumer[18] L. Zdeborova and F. Krzakala, Phase transitions
in the coloring of random graphs, Phys. Rev. E 76,
031131 (2007).
ation unveils clustering and freezing in the random
3-satisfiability problem, Phys. Rev. E 78, 040101
(2008).
[19] D. Achlioptas and A. Coja-Oghlan, Algorithmic [31] B. Selman, H. A. Kautz, and B. Cohen, Noise
strategies for improving local search, in Proc. AAAIBarriers from Phase Transitions, in FOCS ’08: Pro94, 337-343 (1994).
ceedings of the 49th annual IEEE symposium on
Foundations of Computer Science, 793-802 (2008).
[32] A. Braunstein, M. Mézard, and R. Zecchina,
Survey propagation: An algorithm for satisfiability,
[20] D. Achlioptas, A. Coja-Oghlan, and F. RicciRandom Struct. Alg. 27, 201-226 (2005).
Tersenghi, On the Solution-Space Geometry of Random Constraint Satisfaction Problems, Random
[33] A. Frieze and S. Suen, Analysis of two simple
Struct. Alg. 38, 251-268 (2011).
heuristics on a random instance of k-SAT, J. Al[21] F. Ricci-Tersenghi and G. Semerjian, On the cavgor. 20, 312-355 (1996).
ity method for decimated random constraint satisfaction problems and the analysis of belief propagation [34] R. Mulet, A. Pagnani, M. Weigt, and R. Zecchina,
Coloring Random Graphs, Phys. Rev. Lett. 89,
guided decimation algorithms, J. Stat. Mech. P09001
268701 (2002).
(2009).
10
[35] G. Parisi, Some remarks on the survey deci- Additional information
mation algorithm for K- satisfiability, preprint at
Competing financial interests: The authors dehttp://arxiv.org/abs/cs/0301015 (2003).
clare no competing financial interests.
[36] G. Parisi, A backtracking survey propagation algorithm for K-satisfiability, preprint at
http://arxiv.org/abs/cond-mat/0308510 (2003).
[37] J. Chavas, C. Furtlehner, M. Mézard, and
R. Zecchina, Survey-propagation decimation through
distributed local computations, J. Stat. Mech.
P11016 (2005).
[38] L. Dall’Asta, A. Ramezanpour, and R. Zecchina,
Entropy landscape and non-Gibbs solutions in constraint satisfaction problems, Phys. Rev. E 77,
031118 (2008).
[39] L. Zdeborová and M. Mézard, Constraint satisfaction problems with isolated solutions are hard, J.
Stat. Mech. P12004 (2008).
[40] J. M. Chambers, Algorithm 410: Partial Sorting,
Commun. ACM 14, 357-358 (1971).
[41] G. Parisi, On local equilibrium equations for clustering states, preprint at http://arxiv:cs/0212047
(2002). G. Parisi, On the survey-propagation equations in random constraint satisfiability problems, J.
Math. Phys. 49, 125216 (2008).
[42] A. Braunstein and R. Zecchina, Survey Propagation as local equilibrium equations, J. Stat. Mech.
P06007 (2004).
[43] T. Castellani, V. Napolano, F. Ricci-Tersenghi,
and R. Zecchina, Bicolouring random hypergraphs,
J. Phys. A 36, 11037 (2003).
[44] A. Coja-Oghlan and L. Zdeborová, The condensation transition in random hypergraph 2-coloring,
Proceedings of the twenty-third annual ACM-SIAM
symposium on Discrete Algorithms (2012).
Acknowledgements
We thank K. Freese, R. Eichhorn and E. Aurell for useful discussions. This research has been supported by
the Swedish Science Council through grant 621-20122982 and by the European Research Council (ERC)
under the European Unions Horizon 2020 research and
innovation programme (grant agreement No [694925]).
Author contributions
All authors contributed to all aspects of this work.
11
| 8cs.DS
|
M. Abuella and B. Chowdhury, “Solar Power Forecasting Using Support Vector Regression,” in Proceedings of the American Society for Engineering Management International
Annual Conference, 2016.
Solar Power Forecasting Using Support Vector
Regression
Mohamed Abuella, Student Member, IEEE
Badrul Chowdhury, Senior Member, IEEE
Energy Production and Infrastructure Center
Department of Electrical and Computer Engineering
University of North Carolina at Charlotte
Charlotte, USA
Email: [email protected]
Energy Production and Infrastructure Center
Department of Electrical and Computer Engineering
University of North Carolina at Charlotte
Charlotte, USA
Email: [email protected]
Abstract— Generation and load balance is required in the
economic scheduling of generating units in the smart grid. Variable
energy generations particularly from wind and solar energy
resources are witnessing a rapid boost, and, it is anticipated that
with a certain level of their penetration, they become noteworthy
sources of uncertainty. As in the case of load demand, energy
forecasting can also be used to mitigate some of the challenges that
arise from the uncertainty in the resource. While wind energy
forecasting research is considered mature, solar energy forecasting
is witnessing a steadily growing attention from the research
community. This paper presents a support vector regression model
to produce solar power forecasts on a rolling basis for 24 hours
ahead over an entire year, to mimic the practical business of energy
forecasting. Twelve weather variables are brought from a highquality benchmark dataset and new variables are extracted. The
added value of heat index and wind speed as additional variables to
the model is studied across different seasons. The support vector
regression model’s performance is compared with artificial neural
networks, multiple linear regression models for energy forecasting.
Keywords—Solar power forecasting, support vector regression,
weather variables.
I.
INTRODUCTION
The wind and solar energy resources have created
operational challenges for the electric power grid due to the
uncertainty involved in their output in the short term. The
intermittency of these resources may adversely affect the
operation of the electric grid when the penetration level of
these variable generations is high. Thus, wherever the variable
generation resources are used, it becomes highly desirable to
maintain higher than normal operating reserves and efficient
energy storage systems to manage the power balance in the
system. The operating reserves that use fossil fuel generating
units should be kept as low as possible to get the highest
benefit from the deployment of the variable generations.
Therefore, the forecast of these renewable resources becomes a
vital tool in the operation of the power systems and electricity
markets [1].
As in wind power forecasting the solar power also consists
of a variety of methods based on the time horizon being
forecasted, the data available to the forecaster and the
particular application of the forecast. The methods are broadly
categorized according to the time horizon in which they
generally show value. Methods that are common in solar
forecasting include Numerical Weather Prediction (NWP) and
Model Output Statistics (MOS) to produce forecasts, as well as
hybrid techniques that combine ensemble forecasts, and
Statistical Learning Methods [2]. In solar power forecasting up
to 2-hours ahead, the most important input is the available
observations of solar power, while for longer horizons NWP
model becomes crucial for accurate forecasts [3]. Applying
machine learning techniques directly to historical time-series of
PV production associated with NWP outcomes have been
placed among the top models in the last global competition of
energy forecasting, GEFCom 2014 [4]. Moreover, the
empirical study in [5] presents that errors of the solar power
forecasting follow a non-parametric distribution, therefore,
using non-parametric methods such as machine learning
techniques can yield a better accuracy.
Support Vector Regression (SVR) is extended from support
vector machines, a popular tool in machine learning. Timeseries predictions by using SVM includes the load forecasting
are explored in [6]. The SVM utilization in wind forecasting
with severe ramp events is explained by a good extent in [7].
The prediction of solar irradiance from SVM and other
machine learning methods are presented in [8], it concludes
that SVM gives the best forecasts. The study [9] is for using
SVR to forecast the solar power of a 1-MW PV power plant
with the weather variables including the cloudiness, it shows
that SVR and the cloudiness are improving the accuracy. The
authors of [10] propose an approach including support vector
machine and a weather classification methods to a PV-system
of 20kW for a day ahead forecasting, the approach shows
promising results. A benchmark study [11] for short-term wind
and solar power forecasts over different sites in Europe, shows
SVM models bring fairly good results. In [12] two separated
models of SVR are constructed based on cloud cover to
achieve more accurate forecasts of the solar irradiance.
Adding new variables to NWP models is not always
valuable. By [13] found out that the accuracy of the solar
irradiance forecasts based on aerosol chemical transport model
depends on the sky conditions, whether it’s clear or cloudy.
A recent review of a dozen of solar power forecasting
studies are presented in [14], where the authors indicate that
the majority of related publishing literature is focusing on solar
irradiance instead of solar power due to the lack of the
observed data of the solar power production. The solar
irradiance forecasts need an additional step of a conversion to
Note: This is a pre-print of the full paper that published in American Society for Engineering Management, International Annual Conference, 2016, which can be
referenced as below:
M. Abuella and B. Chowdhury, “Solar Power Forecasting Using Support Vector Regression,” in Proceedings of the American Society for Engineering Management 2016
International Annual Conference, 2016.
zones are not different a lot since they are adjacent sola
systems. Note, the order of months as it is provided from th
competition benchmark data and appearsin the box plot doe
not necessary mean the same order of actual year months.
the solar power and this might not be as accurate as the direct
solar power forecasts.
Where: T is the air temperature (F°), and R is the relative
humidity (percentage).
II. THE DATA
All available data are listed monthly as in Fig.1. a. The box
plot of the annual distribution of the historical data of solar
power is shown in Fig.1. b. Since the data from Australia, the
distribution of the
solar
power over
the
Fig.
2. Flowcha
rt dia
grayear
m of months
data preobviously
paration
follows the southern hemisphere seasons, where the winter and
summer seasons opposite to the northern hemisphere.
A. Data Source
The data is derived from the Global Energy Forecasting
Competition 2014 (GEFCom, 2014) which also including
forecasting in the domains of electric load, wind power, solar
power, and electricity prices. The solar data is for small solar
power systems and their location is Australia [4].
B. Data Description
The target variable is the solar power. There are 12
independent weather variables available from the European
Centre for Medium-Range Weather Forecasts (ECMWF) that
are used to produce solar forecasts. These are:
•
•
•
•
•
•
•
•
•
•
•
•
Total column liquid water of cloud (tclw) - (kg/m2).
Total column ice water of cloud (tciw) - (kg/m2)
Surface pressure (SP) - (Pa).
Relative humidity at 1000 mbar (r) - (%).
Total cloud cover (TCC) - (0-1)
10 meter u-wind component (U) - (m/s).
10 meter v-wind component (V) - (m/s).
2 meter temperature (2T◦) - (K)
Surface solar radiation down (SSRD) - (J/m2)accumulated field.
Surface thermal radiation down (STRD) - (J/m2)accumulated field.
Top net solar radiation (TSR) - (J/m2) -accumulated
field
Total precipitation (TP) - (m) - accumulated field.
The last four weather variables (i.e. solar and thermal
radiations besides the precipitation) are given in accumulated
field values, and not average values. They are increasing for
every hour until the end of the day and then start again in
accumulation. The wind variables are given as two
components u and v representing wind directional
components. The u-component of wind is positive for the west
to east direction, while the v-component is positive for the
south to north direction. The resultant vector of both wind
components is the wind speed vector [15]. The polar form of
wind speed (|𝑉|∠𝜃°) in (1) is used later as an additional
weather variable.
|𝑉| = √(𝑢2 + 𝑣 2 ) ,
𝜃 = tan−1
𝑣
𝑢
(1)
Heat index is also an additional weather variable, which is
extracted from the relative humidity and the air temperature
variables. There are some methods to find the heat index, such
as the regression method as follows [16].
Heat Index = -42.379 + 2.04901523T + 10.14333127R 0.22475541TR - 6.83783*10-3T2 - 5.481717*10-2R2 +
1.22874*10-3T2R + 8.5282*10-4TR2 - 1.99 *10-6T2R2
No.
Month
Year
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
April
May
June
July
August
September
October
November
December
January
February
March
April
May
June
July
August
September
October
November
December
January
February
March
April
May
2012
2012
2012
2012
2012
2012
2012
2012
2012
2013
2013
2013
2013
2013
2013
2013
2013
2013
2013
2013
2013
2014
2014
2014
2014
2014
(a)
(b)
Fig.1. Fig.
a) The
Box plot of of
the obse
distribution
of r power
3.entire
Boxavailable
plot ofdata.
the(b)
distribution
rved sola
observed solar power in 2012 months.
The scatte
also useMful
to get sense of the relation
III.r Splot
OLARis
FORECAST
ODELING
ships
betwe
en theispre
variable
(theforecasting
weather variables
Machine
Learning
thedictor
basic skeleton
ofs the
a
nd
re
s
pons
e
va
ria
ble
(the
s
ola
r
powe
r).
Fig.
4 presents th
model in this paper, where the support vector regression (SVR)
isa
deployed
forof
short-term
solar
shows
dvantage
plotting
thepower
dataforecasting.
since theFig.2
scatte
r plots for th
the
flowchart
of
the
machine
learning,
where
the
hypothesis
observed power with respect to the solar irradiance as it i
block embodies the solar forecasting process, inputs (X) are the
also ca
lled
surfa
ce sola
r radia
tion
down
(Sand
SRD)
weather
and
past
historical
actual
solar
power
data,
the for zone1
outputs
are
the solar
forecasts.
training
Thes(Y)
catte
r plot
on power
theleft
in Fig.Whereas
4 is forthe
the
solar irradianc
setSS
isRD169.128
a part of the weather
past actual
solar the
powerNWP
data, which i
as it and
is the
provide
d from
and the learning algorithm is SVR model.
in accumulated values (J/ m2), while the plot on the righ
for the average values of solar irradiance SSRD (W/ m2)
The relationships between the variables on the right plot i
more obvious and it can tell it is a positive relationshi
with relatively high positive correlation coefficient. The las
four given weather variables (i.e. solar and thermal radiation
besides the pre
cipita
tion)
are Flowchart
in accumulated field values no
Fig.2.
Machine
Learning
average values. They are increasing in every hour until th
end of the day and then start again in accumulation [13]
A. Data Preparation
For getting the average values of these data by applying th
It is always a good idea to get the analysis of the used data
formula
in equation 2.
before setting up the forecasting model. The historical data
contains the actual solar power and 12 weather variables. The
various steps of the data preparation are shown in Fig.3.
Fig.3. Flowchart diagram of data preparation
(2)
M. Abuella and B. Chowdhury, “Solar Power Forecasting Using Support Vector Regression,” in Proceedings of the American Society for Engineering Management International
Annual Conference, 2016.
B. Variables Selection
For a kernel type of a Radial Basis Function (RBF):
Selection of the predictor variables by plotting is
cumbersome since there are many weather variables to choose
from. So the greedy forward and backward search approaches
besides the correlation analysis are carried out for the historical
data to investigate the most effective variables. Fig.4 is the
outcome of the variable selection. It is obvious that the solar
irradiance, in addition to the time (hours), the surface
irradiance and net top solar irradiance variables and their
second order polynomial or quadratic terms have the highest
correlation with solar power. The relative humidity and the air
temperature also have a noticeable impact compared to other
less effective variables.
2
𝐾(𝑥𝑖 , 𝑥𝑗 ) = 𝑒 −𝛾(‖𝑥𝑖 −𝑥𝑗 ‖ ) ,
Where 𝛾 (Gamma) is a kernel parameter.
(4)
D. Grid Search
The grid-search is a heuristic approach to find good SVR’s
hyper-parameters (C, γ), as they appear in (3), and (4). Many of
methods are iterative processes, such as walking along a path
until finding the optimal pair of these hyper-parameters. Since
doing a complete grid-search may still be time-consuming, it’s
recommended to use the exponentially growing sequences of C
and γ to identify good parameters (for example, C = 2 −5, 2−3, . .
. , 215, γ = 2−15, 2−13, . . . , 23) [19].
Fig.5. The Boundary margin and loss setting for a linear SVR [18]
III. MODEL RESULTS AND EVALUATION
Fig.4. The order of most effective variable for solar power forecast
C. SVR Model
In general, support vector machine (SVM) is a machinelearning algorithm for classification applications. It’s usually
used with binary classes even if they are non-linearly separable
in 2-D space, that is achieved by introducing a trick that
transforms the classes into a higher dimensional space where
the classes become linearly separable, and this trick is called
kernel trick. The basic kernels are linear, polynomial, radial
basis function [17]. Thereby, the support vector regression
model SVR inherits some of the properties of the SVM. In
SVR, the classification instead of being between the classes, it
is basically a classification of the regression errors that are
greater or less than a certain threshold as shown in Fig.5. For
more details of SVR and its related mathematics refer to [18].
The main optimization and the kernel that are utilized with the
SVR forecasting model here, are presented by (3) and (4).
The SVR requires the solution of the following
optimization problem:
𝑙
min
𝑤,𝑏,𝜉
subject to
1 𝑇
𝑊 𝑊 + 𝐶 ∑ 𝜉𝑖
2
𝑖=1
(3)
𝑦𝑖 (𝑊 𝑇 𝜙(𝑥𝑖 ) + 𝑏) ≥ 1 − 𝜉𝑖
Since (𝑥𝑖 , 𝑦𝑖 ) training set pairs i = 1, . . ., l. W is a normal unit
vector that is perpendicular to the boundary margin, b is a slack
variable, ε a threshold parameter, ξ deviation larger than ε. C >
0 is the penalty parameter. Training vectors 𝑥𝑖 are mapped into
a higher dimensional space by the function 𝜙, this is the kernel
trick 𝐾(𝑥𝑖 , 𝑥𝑗 ) ≡ 𝜙(𝑥𝑖 )𝑇 (𝑥𝑗 ).
The following measures are used to evaluate the accuracy
of the forecasts and the model performance: plots and graphs,
Root Mean Square Error (RMSE) are represented in (5), and a
comparison with other models. For comparison purposes, the
Multiple Linear Regression (MLR) Analysis model, and
Artificial Neural Networks model (ANN) are used [20], [21].
The Library LIBSVM is used for simulation of SVR model
[22].
1
𝑛
2
𝑅𝑀𝑆𝐸 = √ ∑(𝑌𝑖 − 𝑌̂ 𝑖 )
𝑛
(5)
𝑖=1
Where 𝑌̂ is the forecast of the solar power and Y is the
observed value of the solar power. 𝑌̂ and Y are normalized
values of the nominal power capacity of the solar power
system. 𝑛 is the number of hours, it can be day hours or the
month hours. The objective is to minimize RMSE for all
forecasting hours to yield more accurate forecasts. If the
training and testing of the model are carried out for just the
daylight hours and filtering out the night hours (which have
zero solar power generation), the RMSE should also be
determined for these day hours only without including the
night hours.
The grid search of the SVR’s hyper-parameters is
conducted to find out the optimal C and Gamma. In other
words, searching for SVR’s parameters that give a minimum
RMSE. Fig.6 shows the RMSE contour, which results from the
grid search of the SVR’s hyper-parameters. The search grid is
done by two cases: Firstly by using a cross-validation
technique over arbitrary 30 days of the year and these resulted
SVR’s parameters are fixed and then used with all days [19].
While the other case, the adaptive hyper-parameters for each
Note: This is a pre-print of the full paper that published in American Society for Engineering Management, International Annual Conference, 2016, which can be
referenced as below:
M. Abuella and B. Chowdhury, “Solar Power Forecasting Using Support Vector Regression,” in Proceedings of the American Society for Engineering Management 2016
International Annual Conference, 2016.
month throughout the year are searched out and used in SVR to
forecast the solar power of the days of each corresponding
month adaptively.
the accuracy drops to 16% from the case of not including the
heat index and wind speed. The main reason behind this is the
seasonal trend of the weather data.
TABLE I summarizes the evaluation results of the solar
power forecasts of the daily forecasts on a rolling basis over a
year. The evaluation (monthly-RMSE) of both cases, i.e., the
fixed SVR’s parameters and the adaptive parameters from June
2013 to May 2014 are presented. In February 2014 days, the
adaptive SVR’s hyper-parameters provide the best
improvement 7% from using the fixed parameters case.
To see this seasonal variation in a clear picture, the
correlation analysis is conducted between the actual solar
power and six related weather variables, as shown in Fig.7. In
this figure, the correlation results of three cases are presented.
The first case is for training data of SVR, which includes all
days from first day to the forecasted day, as shown in Fig.1. a.
The other two cases of the correlation analysis are for February
and September months’ data when the most change of using
RMSE Contour
0.085
TABLE II
0.08
25
0.075
THE RESULT OF USING HEAT INDEX AND WIND SPEED
Gamma (g)
20
0.07
RMSE = 0.0421
15
0.065
0.06
10
0.055
5
0.05
0
0
20
40
60
80
100
0.045
C
Fig.6.
Grid Search of SVR’s Hyper-Parameters, C, and
Gamma
Thus, the SVR model is doing well even by using the fixed
parameters (C=16, Gamma=1) comparing to the adaptive
parameters case. Because the fixed parameters are validated by
a cross-validation technique over random 30 days and seem
this validation data is good enough to capture almost all
information variation for the SVR model.
TABLE I
THE FORECASTS OF SVR MODEL OVER THE ENTIRE YEAR
Month
Year
RMSE at C=16
& Gamma=1
C
Gamma
June
July
August
September
October
November
December
January
February
March
April
May
2013
2013
2013
2013
2013
2013
2013
2014
2014
2014
2014
2014
0.0758
0.0872
0.0827
0.0751
0.0743
0.0670
0.0583
0.0557
0.0739
0.0817
0.0644
0.0553
55
100
56
16
75
75
10
99
5
11
61
16
1.5
1.0
1.5
1.0
0.5
0.5
0.5
1.4
1.0
1.0
1.0
1.0
RMSE at diff. Imporvement
C & Gamma
(%)
0.0734
0.0851
0.0818
0.0751
0.0712
0.0663
0.0574
0.0534
0.0684
0.0810
0.0635
0.0553
3%
2%
1%
0%
4%
1%
2%
4%
7%
1%
1%
0%
The added value of the additional weather variables, the
heat index (HI) and wind speed, is investigated. These new
variables are extracted from the available weather variables as
in (1) and (2). The heat index is used with its two components
variables; the humidity and the air temperature are also
included. While for wind speed in the polar form, the wind
components (u, v) are excluded from the weather data. The
forecasts are generated for all days of the year, as shown in
TABLE II. The forecasts without using heat index and wind
speed are the similar of those in SVR’s adaptive parameters
case. Adding heat index and wind speed brings some
improvement with September 2014 equals 4% and
insignificant improvement in some months. Whereas in other
months, adding these new variables reduces the accuracy of the
model, as in its worst case with February 2014, the decline of
Month
Year
RMSE w/o
HI & Wind
June
July
August
September
October
November
December
January
February
March
April
May
2013
2013
2013
2013
2013
2013
2013
2014
2014
2014
2014
2014
0.0734
0.0851
0.0818
0.0751
0.0712
0.0663
0.0574
0.0534
0.0684
0.0810
0.0635
0.0553
RMSE w.
RMSE
HI & Wind Difference (%)
0.0746
0.0845
0.0834
0.0722
0.0698
0.0665
0.0572
0.0562
0.0793
0.0846
0.0681
0.0549
-2%
1%
-2%
4%
2%
0%
0%
-5%
-16%
-4%
-7%
1%
heat index and wind speed occur. From Fig.7, it’s obvious in
September case; the correlation analysis is showing an almost
similar trend of that in the training data of SVR model. While
for February the trend of the heat index and wind speed
variables is opposite of that in the training data of SVR model.
Hence, in February 2014 and by including heat index and wind
speed, SVR model suffers from a learning pattern change from
its training data and that leads to the big reduction of forecasts
accuracy.
Fig.7. The solar power correlation plots with training and two months’ data
The comparison of with other models is evaluated for SVR
model in comparison with Multiple Linear Regression (MLR)
and Neural Networks (ANN) models. For this purpose, the
daily RMSE of last month days, May 2014 is shown in Fig.8.
By the aggregation of daily RMSEs over the whole month,
ANN and SVR models produce more accurate forecasts than
MLR model. However, in some days MLR can have a lower
RMSE as in 9th, 16th, 26th, and 30th days. The fluctuation in
daily RMSE can be seen and it’s a normal trend in the
forecasting models. Since it appears that for a certain day a
forecasting model could do a good job, but with another day
the same model doesn’t do well.
M. Abuella and B. Chowdhury, “Solar Power Forecasting Using Support Vector Regression,” in Proceedings of the American Society for Engineering Management International
Annual Conference, 2016.
REFERENCES
[1]
J. M. Morales, A. J. Conejo, H. Madsen, P. Pinson, and M. Zugno,
“Integrating renewables in electricity markets - Operational
problems,” Springer, vol. 205, p. 429, 2014.
[2]
A. Tuohy, J. Zack, S. E. Haupt, J. Sharp, M. Ahlstrom, S. Dise, E.
Grimit, C. Mohrlen, M. Lange, M. G. Casado, J. Black, M. Marquis,
and C. Collier, “Solar Forecasting: Methods, Challenges, and
Performance,” IEEE Power Energy Mag., vol. 13, no. 6, pp. 50–59,
2015.
[3]
J. Kleissl, Solar Energy Forecasting and Resource Assessment.
Elsevier, 2013.
[4]
T. Hong, P. Pinson, S. Fan, H. Zareipour, A. Troccoli, and R. J.
Hyndman, “Probabilistic energy forecasting: Global energy
forecasting competition 2014 and beyond,” Int. J. Forecast., 2016.
[5]
S. Sayeef and S. West, “Very short-term solar forecasting using
inexpensive fisheye camera sky-imagery,” 2014.
[6]
N. Sapankevych and R. Sankar, “Time series prediction using
support vector machines: A survey,” IEEE Comput. Intell. Mag.,
vol. 4, no. 2, pp. 24–38, 2009.
[7]
L. Yang, M. He, J. Zhang, and V. Vittal, Spatio-Temporal Data
Analytics for Wind Energy Integration. Springer, 2014.
[8]
N. Sharma, P. Sharma, D. Irwin, and P. Shenoy, “Predicting solar
generation from weather forecasts using machine learning,” in
Smart Grid Communications (SmartGridComm), 2011 IEEE
International Conference on, 2011, pp. 528–533.
[9]
J. G. Silva Fonseca, T. Oozeki, T. Takashima, G. Koshimizu, Y.
Uchida, and K. Ogimoto, “Use of support vector regression and
numerically predicted cloudiness to forecast power output of a
photovoltaic power plant in Kitakyushu, Japan,” Prog.
photovoltaics Res. Appl., vol. 20, no. 7, pp. 874–882, 2012.
[10]
J. Shi, W. J. Lee, Y. Liu, Y. Yang, and P. Wang, “Forecasting
power output of photovoltaic systems based on weather
classification and support vector machines,” IEEE Trans. Ind. Appl.,
vol. 48, no. 3, pp. 1064–1069, 2012.
[11]
S. Sperati, S. Alessandrini, P. Pinson, and G. Kariniotakis, “The
‘Weather Intelligence for Renewable Energies’ Benchmarking
Exercise on Short-Term Forecasting of Wind and Solar Power
Generation,” Energies, vol. 8, no. 9, pp. 9594–9619, 2015.
[12]
H.-Y. Cheng, C.-C. Yu, and S.-J. Lin, “Bi-model short-term solar
irradiance prediction using support vector regressors,” Energy, vol.
70, pp. 121–127, 2014.
[13]
H. Breitkreuz, M. Schroedter-Homscheidt, T. Holzer-Popp, and S.
Dech, “Short-range direct and diffuse irradiance forecasts for solar
energy applications based on aerosol chemical transport and
numerical weather modeling,” J. Appl. Meteorol. Climatol., vol. 48,
no. 9, pp. 1766–1779, 2009.
[14]
L. Nonnenmacher, A. Kaur, and C. F. M. Coimbra, “Day-ahead
resource forecasting for concentrated solar power integration,”
Renew. Energy, vol. 86, pp. 866–876, 2016.
[15]
“European Centre for Medium-Range Weather Forecasts,” 2016.
[Online]. Available: http://www.ecmwf.int/en/what-directionconvention-u-and-v-components-winds.
[16]
“The National Oceanic and Atmospheric Administration,” 2016.
[Online]. Available: http://www.srh.noaa.gov/ama/?n=heatindex.
[17]
T. Hastie, R. Tibshirani, and J. Friedman, The Elements of
Statistical Learning, 2 Edition. Springer-Verlag New York, 2009.
[18]
A. J. Smola, B. Sch, and B. Schölkopf, “A Tutorial on Support
Vector Regression,” Stat. Comput., vol. 14, no. 3, pp. 199–222,
2004.
Fig.8. Daily RMSE of Different Models for May 2014
To get a broader evaluation of SVR’s performance, the
comparison with MLR and ANN models over a complete year
is done as well, as shown in Fig.9. It is clear that ANN has
lower monthly RMSEs. For some months SVR model can be
as good as ANN. Therefore, SVR model has more accurate
forecasts than MLR and less than ANN. The SVR has an
advantage that cannot be shown illustratively, that it’s more
robust than ANN. Since SVR doesn’t suffer from the local
minima issue as ANN does. Due to the local minima issue in
the fitting algorithm of ANN, thereby each time ANN runs it
gives a different result for the same forecasting step (hour), so
to overcome this, ANN forecasts are resulted from combining
10 forecasting values for each step [23].
Fig.9. Monthly RMSE of Different Models
V. CONCLUSION
Whereas artificial neural networks suffer from multiple
local minima, the support vector regression model’s solution is
global and unique. Furthermore, SVR is less prone to
overfitting issue; these features make SVR more practical.
Thereby, the SVR forecasting model is more accurate than
multiple linear regression model and at the same time, it’s
more robust than ANN.
The added value study of including heat index and wind
speed as additional weather variables leads to that these
variables can improve the models’ performance in some
months. On the other hand, for other months that have a data
pattern different from that in SVR training data, adding heat
index and wind speed decreases the forecasts accuracy.
In the clear sky hours, the model produces more accurate
forecasts than cloudy hours. The more accurate weather
forecasts we use, the more accurate solar power forecasts will
be produced. With additional historical data, the model
performance will improve.
Although in a big picture, a particular forecasting model
gives better forecasts, this forecasting model in some cases
could not do well. Therefore, the combining models can be the
efficient option to get the best from the different models, the
authors are going to take the combining models into account in
further work.
Note: This is a pre-print of the full paper that published in American Society for Engineering Management, International Annual Conference, 2016, which can be
referenced as below:
M. Abuella and B. Chowdhury, “Solar Power Forecasting Using Support Vector Regression,” in Proceedings of the American Society for Engineering Management 2016
International Annual Conference, 2016.
[19]
and C.-J. L. Chih-Wei Hsu, Chih-Chung Chang, “A Practical Guide
to Support Vector Classification,” BJU Int., vol. 101, no. 1, pp.
1396–400, 2008.
[20]
M. Abuella and B. Chowdhury, “Solar Power Probabilistic
Forecasting by Using Multiple Linear Regression Analysis,” in
IEEE Southeastcon Proceedings, 2015.
[21]
M. Abuella and B. Chowdhury, “Solar Power Forecasting using
Artificial Neural Networks,” NAPS2015.
[22]
C. Chang and C. Lin, “LIBSVM : A Library for Support Vector
Machines,” ACM Trans. Intell. Syst. Technol., vol. 2, pp. 1–39,
2011.
[23]
H. T. C. Pedro and C. F. M. Coimbra, “Assessment of forecasting
techniques for solar power production with no exogenous inputs,”
Sol. Energy, vol. 86, no. 7, pp. 2017–2028, 2012.
| 5cs.CE
|
Optimal Temporal Logic Planning in Probabilistic Semantic Maps
arXiv:1510.06469v1 [cs.RO] 22 Oct 2015
Jie Fu, Nikolay Atanasov, Ufuk Topcu, and George J. Pappas
Abstract—This paper considers robot motion planning under
temporal logic constraints in probabilistic maps obtained by
semantic simultaneous localization and mapping (SLAM). The
uncertainty in a map distribution presents a great challenge
for obtaining correctness guarantees with respect to the linear
temporal logic (LTL) specification. We show that the problem
can be formulated as an optimal control problem in which
both the semantic map and the logic formula evaluation are
stochastic. Our first contribution is to reduce the stochastic
control problem for a subclass of LTL to a deterministic shortest
path problem by introducing a confidence parameter δ. A
robot trajectory obtained from the deterministic problem is
guaranteed to have minimum cost and to satisfy the logic
specification in the true environment with probability δ. Our
second contribution is to design an admissible heuristic function
that guides the planning in the deterministic problem towards
satisfying the temporal logic specification. This allows us to
obtain an optimal and very efficient solution using the A*
algorithm. The performance and correctness of our approach
are demonstrated in a simulated semantic environment using a
differential-drive robot.
I. I NTRODUCTION
This paper addresses robot motion planning in uncertain
environments with tasks specified by linear temporal logic
(LTL) co-safe formulas. A map distribution, obtained from
a semantic simultaneous localization and mapping (SLAM)
algorithm [25, 1, 30, 3], facilitates natural robot task specifications in terms of objects and landmarks in the environment.
For example, we can require a robot to “go to a room where
there is a desk and two chairs” instead of giving it exact target
coordinates. One could even describe tasks when the entire
map is not available but is to be obtained as the robot explores
its environment. Meanwhile, temporal logic allows one to
specify rich, high-level robotic tasks. Hence, a meaningful
question we aim to answer is the following. Given a semantic
map distribution, how does one design a control policy that
enables the robot to efficiently accomplish temporal logic
tasks with high probability, despite the uncertainty in the
true environment?
This question is motivated by two distinct lines of work,
namely, control under temporal logic constraints and multitask SLAM. Control synthesis with temporal logic specifications has been studied for both deterministic [15, 13, 4] and
J. Fu, N. Atanasov, and G. Pappas are with the Department
of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104, USA, {jief, atanasov,
pappasg}@seas.upenn.edu.
U. Topcu is with the Department of Aerospace Engineering and Engineering Mechanics, University of Texas, Austin, TX 78712, USA,
[email protected].
This work is supported by AFRL # FA8650-15-C-2546, ONR #
N000141310778, NSF # 1550212, and TerraSwarm, one of six centers of
STARnet, a Semiconductor Research Corporation program sponsored by
MARCO and DARPA.
stochastic systems [19, 7]. A recent line of work focuses
on the design problem in the presence of unknown and
uncertain environments. In general, three types of uncertainty
are considered: sensor uncertainty [12], incomplete environment models [16, 11, 22, 23], or uncertainty in the robot
dynamics [31, 9]. Johnson and Kress-Gazit [12] employ a
model checking algorithm to evaluate the fragility of the
control design with respect to temporal logic tasks when
sensing is uncertain. To handle unexpected changes in the
environment and incompleteness in the environment model,
Kress-Gazit et al. [16] develop a sensor-based reactive motion
planning method that guarantees the correctness of the robot
behaviors under temporal logic constraints. Livingston et al.
[22, 23] propose a way to efficiently modify a nominal
controller through local patches for assume-guarantee LTL
formulas. Guo et al. [11] develop a revision method for online
planning in a gradually discovered environment. Probabilistic
uncertainty is studied in [31, 9]. Wolff et al. [31] develop
a robust control method with respect to temporal logic
constraints in a stochastic environment modeled as an interval
Markov decision process (MDP). Fu and Topcu [9] develop
a method that learns a near-optimal policy for temporal logic
constraints in an initially unknown stochastic environment.
However, existing work abstracts the system and its environment into discrete models, such as, MDPs, two-player
games, and plans in the discrete state space. In this work, the
environment uncertainty is represented by a continuous map
distribution which makes methods for discrete systems not
applicable. Moreover, control design with map distributions
obtained by uncertain sensor and semantic SLAM has not
been addressed in literature.
While temporal logic is expressive in specifying a wide
range of robot behaviors, recent advances in SLAM motivate the integration of task planning with simultaneously
discovering an initially unknown environment using SLAM
algorithms. Multi-task SLAM is proposed in Guez and
Pineau [10]. The authors consider a planning problem in
which a mobile robot needs to map an unknown environment,
while localizing itself and maximizing long-term rewards.
The authors formulate the decision-making problem as a
partially observable Markov decision process and plan with
both the mean of the robot pose and the mean of the map
distribution. Bachrach et al. [2] develop a system for visual
odometry and mapping using an RGB-D camera. The authors
employ the Belief Roadmap algorithm [24] to generate the
shortest path from the mean robot pose to a goal state, while
propagating uncertainties along the path. It is difficult, however, to extend these approaches to temporal logic planning
with probabilistic semantic maps. Unlike reachability and
reward maximization, the performance criteria induced by
LTL formulas require a rigorous way to reason about the uncertainty in the map distribution. To tackle these challenges,
our method brings together the notions of robustness and
probabilistic correctness to satisfy quantitative temporal logic
specifications in the presence of environment uncertainty
given by map distributions.
This work makes the following contributions:
• We formulate a stochastic optimal control problem for
planning robot motion in a probabilistic semantic map
under temporal logic constraints.
• For a subclass of LTL we reduce the stochastic problem
to a deterministic shortest path problem that can be
solved very efficiently. We prove that for a given confidence parameter δ, the robot trajectory obtained from
the deterministic problem, if it exists, satisfies the logic
specification with probability δ in the true environment.
• We design an admissible heuristic for A* to compute the
optimal solution of the deterministic problem efficiently.
II. P ROBLEM F ORMULATION
In this section, we introduce models for the robot and
its uncertain environment, represented by a semantic map
distribution. Using temporal logic as the task specification
language, we formulate a stochastic optimal control problem.
A. Robot and environment models
Consider a mobile robot whose dynamics are governed by
the following discrete-time motion model:
xt+1 = f (xt , ut )
(1)
(xpt , xat )
∈ X is the robot state, containing
where xt =
its pose xpt and other variables xat such as velocity and
acceleration and ut ∈ U is the control input, selected from a
finite space of admissible controls. A trajectory of the robot,
for t ∈ N ∪ {∞}, is a sequence of states x0:t := x0 x1 . . . xt ,
where xk ∈ X is the state at time k.
The robot operates in an environment modeled by a semantic map M := {l1 , . . . , lM } consisting of M landmarks.
Each landmark li := (lip , lic ) ∈ M is defined by its pose
lip and class lic ∈ C, where C is a finite set of classes (e.g.,
table, chair, door, etc.). The robot does not know the true
landmark poses but has access to a probability distribution
P over the space of all possible maps. Such a distribution
can be produced by a semantic SLAM algorithm [30, 3]
and typically consists of a Gaussian distribution over the
landmark poses and a discrete distribution over the landmark
classes. More precisely, we assume P is determined by
p
¯p p
parameters (¯lp , Σp , {ρci }M
i=1 ) such that l ∼ N (l , Σ ) and
c
li is generated by the probability mass function ρci . In this
work, we suppose that the class of each landmark is known
and leave the case of uncertaint landmark classes for future
work.
B. Temporal logic specifications
We use linear temporal logic (LTL) to specify the robot’s
task in the environment. LTL formulas [29] can describe
temporal ordering of events along the robot trajectories and
are defined by the following grammar: φ := p | ¬φ | φ1 ∨φ2 |
φ | φ1 Uφ2 , where p ∈ AP is an atomic proposition,
and
and U are temporal modal operators for “next” and
“until”. Additional temporal logic operators are derived from
basic ones: ♦ϕ := true Uϕ (eventually) and ϕ := ¬♦¬ϕ
(always). We assume that the robot’s task is given by an LTL
co-safe formula [17], which allows checking its satisfaction
using a finite-length robot trajectory.
The LTL formula is specified over a finite set of atomic
propositions that are defined over the robot state space X and
the environment map M. Examples of atomic propositions
include:
αip (r) : d(xp , lip ) ≤ r
for r ∈ R, i ∈ {1, . . . , M },
αic (Y
for Y ⊆ C, i ∈ {1, . . . , M }.
):
lic
∈Y
(2)
Proposition αip (r) evaluates true when the robot is within
r units distance from landmark i, while proposition αic (Y )
evaluates true when the class of the i-th landmark is in the
subset Y of classes. In order to interpret an LTL formula
over the trajectories of the robot system, we use a labeling
function to determine which atomic propositions hold true
for the current robot pose.
Definition 1 (Labeling function1 ). Let AP be a set of atomic
propositions and M be the set of all possible maps. A labeling
function L : X × M → 2AP maps a given robot state x ∈ X
and map M ∈ M to a set L(x, M) of atomic propositions
that evaluate true.
For robot trajectory x0:t and map M ∈ M, the label
sequence of x0:t in M, denoted L(x0:t , M), is such that
L(x0:t , M) = L(x0 , M)L(x1 , M)L(x2 , M) . . . L(xt , M).
Given an LTL co-safe formula ϕ, one can construct
a deterministic finite-state automaton (DFA) Aϕ =
(Q, 2AP , T, q0 , F ) where Q, 2AP , q0 , F are a finite set of
states, the alphabet, the initial state, and a set of final states,
respectively. T : Q × 2AP → Q is a transition function
such that T (q, a) is the state that is reached with input a at
state q. We extend the transition function in the usual way2 :
T (q, uv) = T (T (q, u), v) for u, v ∈ (2AP )∗ . A word w is
accepted in Aϕ if and only if T (q0 , w) ∈ F . The set of words
accepted by Aϕ is the language of Aϕ , denoted L(Aϕ ).
We say that a robot trajectory x0:∞ satisfies the LTL
formula ϕ in the map M if and only if there is k ≥ 0 such
that L(x0:k , M) ∈ L(Aϕ ). Then, x0:k is called a good prefix
for the formula ϕ. Furthermore, Aϕ accepts exactly the set
of good prefixes for ϕ and for any state q ∈ F , it holds that
T (q, a) ∈ F for any a ∈ 2AP .
We are finally ready for a formal problem statement.
Problem 1. Given an initial robot state x0 ∈ X, a semantic map distribution P, and an LTL co-safe formula ϕ
1 When the map is fixed, our labeling function definition reduces to the
commonly-used definition in robotic motion planning under temporal logic
constraints [7].
2 Notation: Let A be a finite set. Let A∗ , Aω be the set of finite and
infinite words over A. Let λ = A0 be the empty string. Abusing notation
slightly, we use ∅ and λ interchangeably. For w ∈ Aω , if there exist u ∈ A∗
and v ∈ Aω such that w = uv then u is a prefix of w and v is a suffix of
w.
represented by a DFA Aϕ , choose a stopping time τ and a
sequence of control policies ut ∈ U for t = 0, 1, . . . , τ that
maximizes the probability of the robot satisfying ϕ in the true
environment M while minimizing its motion cost:
min
τ,u0 ,u1 ,...,uτ
s.t.
τ
X
E
c(xt , xt+1 ) + κP(qτ +1 ∈
/ F)
t=0
xt+1 = f (xt , ut ),
qt+1 = T (qt , L(xt+1 , M)), ∀0 ≤ t < τ,
where c is a positive-definite motion cost function, which
satisfies the triangle inequality, and κ ≥ 0 determines the
relative importance of satisfying the specification versus the
total motion cost.
Remark: The optimal cost of Problem 1 is bounded below
by 0 due to the assumptions on c, κ and bounded above by κ,
obtained by stopping immediately without satisfying ϕ, i.e.,
τ = 0.
III. P LANNING TO BE P ROBABLY C ORRECT
The map uncertainty in Problem 1 leads to uncertainty
in the evaluation of the atomic propositions and hence to
uncertainty in the robot trajectory labeling. In the meantime,
the automaton state cannot be observed. Rather than solving
the resulting optimal control problem with partial observability, we propose an alternative solution that generates a
near-optimal plan with a probabilistic correctness guarantee
for the temporal logic constraints. The main idea is to convert
the original semantic map distribution to a high-confidence
deterministic representation and solve a deterministic optimal
control problem with this new representation. The advantage
is that we can solve the deterministic problem very efficiently
and still provide a correctness guarantee. This avoids the
need for sampling-based methods in the continuous space of
map distributions, which become computationally expensive
when planning in large environments. To this end, we use a
confidence region around the mean M̄ := {(¯lip , lic )}M
i=1 of
the semantic map distribution P to extend the definition of
the labeling function.
Proposition 1. Given a robot trajectory x0:t and a map
distribution P, Lδ (x0:t , P) is a subsequence of L(x0:t , M)
with probability δ.
Proof. See Appendix A
Intuitively, a label L(xk , M) is preserved at the k-th
position of Lδ (x0:t , P) if for any two sample maps m, m0 in
the δ-confidence region, L(xi , m) = L(xi , m0 ). Otherwise, it
is replaced with ∅. Next, we show that when the LTL formula
ϕ satisfies a particular property, if Lδ (x0,t , P) is accepted
by the DFA Aϕ , then with probability δ, x0:t satisfies the
LTL specification ϕ. The required property is that the formal
language characterization of the logic formula translates to
a simple polynomial [27]. An ω-regular language L over an
alphabet A is simple monomial if and only if it is of the form
A∗ a1 A∗ a2 A∗ . . . A∗ ak A∗ (A∗ b1 A∗ b2 A∗ . . . A∗ b` A∗ )ω
where a1 , a2 , . . . , ak , b1 , b2 , . . . , b` ∈ A, k ≥ 0, and ` ≥
0. A finite union of simple monomials is called a simple
polynomial.
Theorem 1. If the the language L(Aϕ ) of Aϕ is a simple polynomial, then Lδ (x0:τ , P) ∈ L(Aϕ ) implies that
P(L(x0:τ , M) ∈ L(Aϕ )) = δ.
Proof. See Appendix B.
The significance of Thm. 1 is that it allows us to reduce the
stochastic control problem with an uncertain map (Problem
1) to a deterministic shortest path problem. We introduce the
following product system to facilitate the conversion.
Definition 3 (δ-Probably correct product system). Given the
robot system in (1), the map distribution P, the automaton
Aϕ , and a parameter δ ∈ (0, 1), a δ-probably correct product
system is a tuple G δ = hS, U, ∆, s0 , SF i defined as follows.
• S = X × Q is the product state space.
• ∆ : S × U → S is a transition function such that
∆((x, q), u) = (x0 , q 0 ) where x0 = f (x, u) and q 0 =
T (q, Lδ (x0 , P)). It is assumed that T (q, ∅) = q, ∀q ∈
Q.
• s0 = (x0 , q0 ) is the initial state.
• SF = X × F is the set of final states.
Definition 2 (δ-Confident labeling function). Given a robot
state x ∈ X, a map distribution P, and a parameter δ ∈
(0, 1), a δ-confident labeling function is defined as follows:
For the subclass of LTL co-safe formulas whose languages
are
simple polynomials, Thm. 1 guarantees that the projection
L(x,
M̄)
if
L(x,
M̄)
=
L(x,
m)
for
all
maps
δ
on
X
of any trajectory s0:t of G δ that reaches SF in the δL (x, P) :=
m in the δ-confidence region of P,
confident
map has probability δ of satisfying the specification
∅
otherwise.
in the true map. The implications are explored in Sec. IV.
Before we proceed, however, it is important to know to
We now explain the intuition for defining the δ-confident
labeling function as in Def. 2. For a given robot trajectory what extent the expressiveness of LTL is limited by restricting
x0:t , rather than maintaining a distribution over the possible it to the subclass of simple polynomials. In Appendix C, we
label sequences, the robot keeps only a sequence of labels show that such LTL formulas can express reachability and
that, with probability δ, is a subsequence3 of the label sequencing properties. Moreover, with a slight modification
sequence L(x0:t , M) in the true environment. This statement of Def. 3, we can also ensure the correctness of plans with
respect to safety constraints.
is made precise in the following proposition.
Consider safety constraints in the following form φsaf e
with
φsaf e being a propositional logic formula over AP.
3 For a word u ∈ Aω , u is a subsequence of w if u can be obtained from
For example, an obstacle avoidance requirement is given by
w by replacing symbols with the empty string λ.
d(x, xo ) ≥ r where xo are the coordinates of an obstacle.
When the LTL formula includes such safety constraints,
we need to modify the transition function in Def. 3 in the
following way. For any state s ∈ S and any input u ∈ U , let
s0 = (x0 , q 0 ) = ∆(s, u). Then, if there exists at least one m
in the δ-confidence region of P such that the propositional
logic formula corresponding to L(x0 , m) implies4 ¬φsaf e , let
∆(s, u) = sink, where sink is a non-accepting sink state that
satisfies ∆(sink, u) = sink for any u ∈ U . Thus, the state
sink will not be visited by any trajectory of G δ that reaches
SF , which means the safety constraint will be satisfied with
probability δ in the true environment. The following toy
example illustrates the concepts.
Example 1. In Fig. 1, a mobile robot is tasked with visiting at least one landmark in an uncertain environment.
Formally, the LTL specification of the task is ϕ := ♦p where
p := d(x, lip ) ≤ 1 for any landmark i. Given the robot
trajectory x0:t represented by the dashed line in the figure,
when the robot traverses the 95%-confidence region of l1 ’s
pose distribution, it cannot confidently (with confidence level
δ = 0.95) decide the value of p in the true map because
for some map realizations, d(x, l1p ) > 1. On the other hand,
when the robot is near l2 , p evaluates true because a unit
ball around the robot covers the entire 95%-confidence region
of l2 ’s pose distribution. Let Br (l) a ball centered at l with
radius r. The label sequence of x0:t in the true environment is
L(x0:t , M) = ∅k1 {p}k2 ∅k3 {p} = {p}k2 +1 where k1 , k2 , k3
are the numbers of steps before reaching B1 (l1p ), in B1 (l1p ),
and after leaving B1 (l1p ) but before reaching B2 (l2p ). The label sequence L0.95 (x0:t , P) = ∅k1 +k2 +k3 {p} = {p}. Clearly,
L0.95 (x0:t , P) is a subsequence of L(x0:t , M). Moreover,
the trajectory satisfies the LTL specification which is a
reachability constraint.
Thus, in any run that is safe in G δ , the robot is able to safely
avoid both l1 and l2 with probability δ.
IV. R EDUCTION TO D ETERMINISTIC S HORTEST PATH
For a fixed confidence δ, due to Thm. 1, we can convert
Problem 1 to a deterministic shortest path problem within the
probably correct product system G δ . In this case, P(qτ +1 ∈
/
F ) ∈ {0, 1} and the optimal solution to Problem 1 is either
to stop immediately (τ = 0), incurring cost κ, or to find a
robot trajectory satisfying ϕ with cumulative motion cost less
than κδ. The latter corresponds to the following deterministic
problem.
Problem 2. Given an initial robot state x0 ∈ X, a semantic
map distribution P, a confidence δ ∈ (0, 1), and an LTL cosafe formula ϕ represented by a DFA Aϕ , choose a stopping
time τ and a sequence of control inputs ut ∈ U for t =
0, 1, . . . that minimize the motion cost of a trajectory that
satisfies ϕ:
min
τ
X
u0 ,u1 ,...,uτ
s.t.
c(xt , xt+1 )
t=0
xt+1 = f (xt , ut ),
qt+1 = T (qt , Lδ (xt , P)), ∀0 ≤ t < τ,
qτ +1 ∈ F,
τ
X
c(xt , xt+1 ) ≤ κδ.
t=0
If Problem 2 is infeasible, it is best in Problem 1 to stop
immediately (τ = 0), incurring cost κ; otherwise, the robot
should follow the control sequence u∗0:τ computed above and
the corresponding trajectory x∗0:τ +1 to incur cost:
τ
X
c(x∗t , x∗t+1 ) + κ(1 − δ) ≤ κ
t=0
`1
`1
x
`2
x
`2
in the original Problem 1. Since Problem 2 is a deterministic
shortest path problem, we can use any of the traditional motion planning algorithms, such as RRT [20], RRT* [14, 13]
or A* [21] to solve it. We choose A* due to its completeness
guarantees5 [26] and because the automaton Aϕ can be used
to guide the search as we show next.
A. Admissible Heuristic
Fig. 1: Example: robot with a reachability objective. The red
diamonds are the means of the landmarks’ pose distributions and
the ellipsoids are the associated 95%-confidence regions. The robot
is represented by the blue dot. The green squares represent the true
locations of landmarks l1 and l2 .
In the case of a safety constraint, e.g., ϕsaf e = p where
p := d(x, lip ) > 1 for any landmark i, once the robot
gets close to l1 the δ-probably correct product system will
transition to the non-accepting sink because there exists a
sample map m such that the safety constraint is violated.
4 Given a label L(x0 , m) ⊆ AP, the corresponding propositional logic
V
V
formula is αi ∈L(x0 ,m) αi ∧ αj ∈AP\L(x0 ,m) ¬αj .
The efficiency of A* can be increased dramatically by
designing an appropriate heuristic function to guide the
search. Given a state s := (x, q) in the product system (Def.
3), a heuristic function h : S → R provides an estimate
of the optimal cost h∗ (s) from s to the goal set SF . If the
heuristic function is admissible, i.e., never overestimates the
cost-to-go (h(s) ≤ h∗ (s), ∀s ∈ S), then A* is optimal [26].
Lacerda et al. [18] propose a distance metric to evaluate
the progression of an automaton state with respect to an
LTL co-safe formula. We use a similar idea to design an
admissible heuristic function. We partition the state space
5 To guarantee completeness of A* for Problem 2, the robot state space
X needs to be assumed bounded and compact and needs to discretized.
Q of Aϕ into level sets as follows. Let Q0 := F and for
Si
i ≥ 0 construct Qi+1 := {q ∈ Q \ k=0 Qk | ∃q 0 ∈ Qi , a ∈
2AP , such that T (q, a) = q 0 }. The generation of level sets
stops when Qi = ∅ for some i. Further, we denote the set
of all sink states by Q∞ . Thus, given q ∈ Q one can find a
unique level set Qi such that q ∈ Qi . We say that i is the
level of q and denote it by Level(q) = i.
Proposition 2. Let s0:t be a trajectory of the product system
G δ that reaches SF , i.e., st ∈ SF . Then, for any 0 ≤ k < t,
given sk = (xk , qk ) and sk+1 = (xk+1 , qk+1 ), it holds that
Level(qk ) ≤ Level(qk+1 ) + 1.
Proof. Since T (qk , Lδ (xk+1 , P)) = qk+1 , if qk+1 ∈ Qi for
some level i, then, S
by construction of the level sets, either
i
qk ∈ Qi+1 or qk ∈ j=0 Qj .
By construction of the level sets, the automaton states q0:t ,
associated with any trajectory s0:t of the product system that
reaches a goal state (st ∈ SF ), have to pass through the level
sets sequentially. In other words, if Level(q0 ) = i, then there
0
exists a subsequence q0:i
of q0:t such that Level(q10 ) = i − 1,
0
Level(q2 ) = i − 2, . . . , Level(qi0 ) = 0. Thus, we can construct
a heuristic function that underestimates the cost-to-go from
some state s := (x, q) ∈ S with Level(q) = i by computing
the minimum cost to to reach a state s0 := (x0 , q 0 ) such that
Level(q 0 ) ∈ {i − 1, i} and q 6= q 0 . To do so, we determine all
the labels that trigger a transition from q to q 0 in Aϕ and then
find all the robot states B that produce those labels. Then,
h(x, q) is the minimum distance from x to the set B. The
details of this construction and other functions needed for A*
search with LTL specifications, are summarized in Alg. 2.
Proposition 3. The heuristic function in Alg. 2 is admissible.
Proof. See Appendix D.
Prop. 3 guarantees that A* will either find the optimal
solution to Problem 2 or will report that Problem 2 is
infeasible. In the latter case, the robot cannot satisfy the logic
specification with confidence δ and it should either reduce δ
or stop planning.
Note that while the heuristic function is admissible, it is not
guaranteed that it is also consistent. Consider two arbitrary
states (x, q) and (x0 , q 0 ) with Level(q) = n and Level(q 0 ) =
n + 1. It is possible that the cost to get from (x, q) to a place
in the environment, where a transition to level n − 1 occurs,
is very large, i.e., h(x, q) is large, but it might be very cheap
to get from (x0 , q 0 ) to (x, q) and vice versa. In other words,
it is possible that the following inequalities hold:
c(x, x0 ) + h(x0 , q 0 ) ≤ c(x, x0 ) + c(x0 , x) < h(x, q),
which makes the heuristic in Alg. 2 inconsistent. We emphasize that, even with an inconsistent heuristic, A∗ can be
very efficient if a technique such as bi-directional pathmax
is employed to propagate heuristics between neighboring
states [32].
B. Accelerating A* Search using Motion Primitives
If willing to sacrifice completeness, one can significantly
accelerate the A* search by using motion primitives for
the robot in order to construct a lattice-based graph [28].
The advantage of such a construction is that the underlying
graph is sparse and composed of dynamically-feasible robot
trajectories that can incorporate a variety of constraints. We
present motion primitives for a differential-drive robot in Fig.
3 but much more general models can be handled [6].
A motion primitive is similar to the notion of macro-action
[8, 5] and consists of a collection of control inputs σ [T ] :=
(u0 , u2 , . . . , uT −1 ) that are applied sequentially to a robot
state xt so that:
xt+1+k = f (xt+k , σ [T ] (k)), k = 0, . . . , T − 1.
Instead of using the original control set U , we can plan with
[T ]
a set Ū := {σj j } of motion primitives (see G ET S UCCESSORS
in Alg. 2). In our experiments, the motion primitives were
designed offline. Twenty locations with outward facing orientations were chosen on the perimeter of a circle of radius
10 m. A differential-drive controller was used to to generate
a control sequence of length 5 that would lead a robot at
the origin to each of the selected locations. Fig. 3 shows the
resulting set of motion primitives. They are wavy because
the controller tries to follow a straight line using a discrete
set of velocity and angular-velocity inputs.
Summary. We formulate the temporal logic planning with
a semantic map distribution as a stochastic optimal control
problem (Problem 1). Since Problem 1 is intractable, we
reduce it to a deterministic shortest path problem (Problem 2)
with probabilistic correctness guaranteed by Thm 1. We can
solve Problem 2 optimally using A* because the heuristic
function proposed in Alg. 2 is admissible (Prop. 3). The
obtained solution is partial with respect to Problem 1 because,
rather than a controller that trades off the probability of
satisfying the specification and the total motion cost, it
provides the optimal controller in a subspace of deterministic
controllers that guarantee that the probability of satisfying the
specification is δ.
V. E XAMPLES
In this section, we demonstrate the method for LTLconstrained motion planning with a differential-drive robot
with state xt := (xt , yt , θt )T ∈ SE(2), where (xt , yt ) and θt
are the 2D position and orientation of the robot, respectively.
The kinematics of the robot are discretized using a sampling
period τ as follows:
τ ν cos(θt + τ ω/2)
τ ν sin(θt + τ ω/2) , |τ ω| < 0.001,
xt+1
xt
τω
yt+1 = yt +
ν
(sin(θ
+
τ
ω)
−
sin
θ
)
t
t
ω
θt+1
θt
ν
ω (cos θt ) − cos(θt + τ ω)) , else.
τω
The mobile robot is controlled by motion primitives in Fig.
3, whose segments are specified by ν = 1 m/s, τ = 2 s, and
ω ∈ [−3, 3] rad/s.
The LTL constraints were specified over the two
types of atomic propositions for object classes C =
{Sq, Cir, Hex, Dia, Tri}. Proposition αic (h) means the class
of i-th landmark is h for h ∈ C. Proposition αip (r) :
d(xp − lip ) ≤ r means the robot is r-close to landmark li .
The following LTL specification was given to the robot:
problem can be solved very efficiently and yet the planned
robot trajectory is guaranteed to have minimum cost and to
satisfy the logic specification with probability δ.
This work takes an initial step towards integration of
semantic SLAM and motion planning under temporal logic
ϕ : (♦(φ1 ∧ ♦(φ2 ∧ ♦φ3 )) ∧ ♦φ4 ) ∧ φsaf e
constraints. In future work, we plan to extend this method to
where φi , i = 1, . . . , 4 are the following propositional logic handle the following: 1) landmark class uncertainty, 2) robot
motion uncertainty, 3) a more general class of LTL specificaformulas:
tions, 4) map distributions that are changing online. Our goal
i ∈ {1, . . . , M }is to develop a coherent approach for planning autonomous
φ1 : = αip (1) ∧ αic (Tri),
φ2 : = αip (2) ∧ αic (Dia) ∧ αjp (2) ∧ αjc (Cir), i, j ∈ {1, . . . , M }robot behaviors that accomplish high-level temporal logic
φ : = αp (2) ∧ αc (Sq) ∧ αp (2) ∧ αc (Cir), i, j ∈ {1, . . . , M }tasks in uncertain semantic environments.
3
φ4 : =
i
αip (1)
∧
i
αic (Hex),
j
j
i ∈ {1, . . . , M }
A PPENDIX
and for i, j ∈ {1, . . . , M } the safety constraint is:
A. Proof of Prop. 1
φsaf e := ¬(αip (2) ∧ αic (Sq) ∧ αjp (2) ∧ αjc (Sq)).
Let xk be the state at time k. If for all samples m
in the δ-confidence region of P, L(xk , m) = L(xk , M̄),
then Lδ (xk , P) = L(xk , M̄), which equals L(xk , M) with
probability δ. Otherwise, Lδ (xk , P) = λ (empty string).
Since k is arbitrary, we can conclude that Lδ (x0:t , P) is a
subsequence of L(x0:t , M) with probability δ.
In other words, the robot needs to first visit a triangle, then go
to a region where it is close to both a circle and a diamond,
and finally visit a region where it is close to both a circle
and a square, while visiting a hexagon at some point and
avoiding getting stuck between any two squares.
Several case studies were carried out using a simulated
semantic map distribution. Robot trajectories with least cost
that satisfy the LTL specification with confidence δ = 0.95
for different initial conditions were computed with A* and
are shown in Fig. 4. Optimal paths with the same initial
conditions but different confidence parameters δ = 0.95 and
0.5 are shown in Fig. 5. As expected, we observe a trade-off
between the probability of satisfying the LTL formula and
the total cost of the path. With a lower confidence (δ = 0.5),
the total cost for satisfying the LTL formula is also lower
than that of a path which satisfies the formula with a high
confidence (δ = 0.95). Particularly, the uncertainty in the
pose of triangle l1 with pose distribution ¯l1p = (2.5, 1.19) is
the main reason for the difference in the planned trajectories.
With δ = 0.95, even though the robot can reach the vicinity
of triangle l1 , it does not have enough confidence to ensure
that the triangle would be visited. Instead, it plans to visit
another triangle l2 for which the uncertainty in the pose
distribution ¯l2p = (−5.18, 12.04) is smaller. Reducing the
confidence requirements allows the robot to plan a path that
visits triangle l1 and has a lower total cost compared to that
of visiting triangle l2 .
VI. C ONCLUSION
This paper proposes an approach for planning optimal
robot trajectories that probabilistically satisfy temporal logic
specifications in uncertain semantic environments. By introducing a δ-confident labeling function, we show that the
original stochastic optimal control problem in the continuous
space of semantic map distributions can be reduced to a
deterministic optimal control problem in the δ-confidence
region of the map distribution. Guided by the automaton
representation of the LTL co-safe specification, we develop
an admissible A* algorithm to solve the deterministic problem. The advantage of our approach is that the deterministic
B. Proof of Thm. 1
Since the language L(Aϕ ) is a simple polynomial, the
following upward closure [27] property holds: For any word
uv ∈ L(Aϕ ) and any a ∈ 2AP , it holds that uav ∈ L(Aϕ ).
In other words, if any empty string in a word from a simple
polynomial language is replaced by a symbol in the alphabet,
then the resulting word is still in the language.
For a given robot trajectory x0:τ , let Lδ (x0:τ , P) =
b0 b1 . . . bτ ∈ L(Aϕ ) be the δ-confident label sequence
and L(x0:τ , M) = a0 a1 . . . aτ be the true label sequence.
According to Prop. 1, for each 0 ≤ i ≤ τ , either ai = bi or
ai 6= bi and bi = ∅. Thus, if b0 b1 . . . bτ belongs to L(Aϕ ),
a0 a1 . . . aτ must be in L(Aϕ ) because it is obtained by
replacing each empty string in b0 b1 . . . bτ with some symbol
in 2AP and the language L(Aϕ ) is upward closed.
C. Characterization of LTL co-safe formulas that translate
to simple polynomials
Formally, the subset of LTL formulas is defined by the
grammar
ϕ := ϕreach | ϕseq | ϕ ∧ ϕ | ϕ ∨ ϕ,
(3)
where ϕreach = ♦φ | ϕreach ∧ ϕreach | ϕreach ∨ ϕreach
represents the reachability, ϕseq := φ | ♦ϕseq | ♦(ϕseq ∧
♦ϕseq ) is a set of formulas describing sequencing constraints.
Here, φ is a propositional logic formula.
D. Proof of Prop. 3
We proceed by induction on the levels in Aϕ . In the base
case, q ∈ Q0 ≡ F and h(x, q) = 0 for any x ∈ X. Suppose
that the proposition is true for level n and let (x, q) be some
state with x ∈ X and Level(q) = n + 1. As before, let
h∗ (x, q) be the optimal cost-to-go. Due to Prop. 2, there are
only three possibilities for the next state (x0 , q 0 ) along the
optimal path starting from (x, q):
•
Level(q 0 ) = n: By construction of h:
∗
0
∗
0
0
h (x, q) = c(x, x ) + h (x , q ) ≥ h(x, q) + 0.
•
•
Level(q 0 ) = n + 1: Same conclusion as above.
Level(q 0 ) = k > n + 1: In this case, there exists another
state (x00 , q 00 ) later along the optimal path such that
Level(q 00 ) = n + 1 (otherwise the optimal path cannot
reach the goal set). Then, by the triangle inequality for
the motion cost c:
h∗ (x, q) = c(x, x0 ) + c(x0 , x00 ) + h∗ (x00 , q 00 )
≥ c(x, x00 ) + 0 ≥ h(x, q)
Thus, we conclude that h∗ (x, q) ≥ h(x, q) ≥ 0 for all x ∈ X
and q ∈ Q.
R EFERENCES
[1] N. Atanasov, M. Zhu, K. Daniilidis, and G. Pappas. Semantic
Localization Via the Matrix Permanent. In Robotics: Science
and Systems, 2014.
[2] A. Bachrach, S. Prentice, R. He, P. Henry, A. S. Huang,
M. Krainin, D. Maturana, D. Fox, and N. Roy. Estimation,
planning, and mapping for autonomous flight using an rgbd camera in GPS-denied environments. The International
Journal of Robotics Research, 31(11):1320–1343, 2012.
[3] S. Bao and S. Savarese. Semantic Structure from Motion.
In IEEE Conf. on Computer Vision and Pattern Recognition
(CVPR), pages 2025–2032, 2011.
[4] A. Bhatia, L. E. Kavraki, and M. Y. Vardi. Sampling-based
motion planning with temporal goals. In IEEE International
Conference on Robotics and Automation, pages 2689–2696.
IEEE, 2010.
[5] A. Botea, M. Enzenberger, M. Muller, and J. Schaeffer. MacroFF: Improving AI planning with automatically learned macrooperators. Journal of Artificial Intelligence Research, 24:581–
621, 2005.
[6] B. Cohen, G. Subramanian, S. Chitta, and M. Likhachev. Planning for Manipulation with Adaptive Motion Primitives. In
IEEE International Conference on Robotics and Automation,
pages 5478–5485, 2011.
[7] X. Ding, S. Smith, C. Belta, and D. Rus. Optimal control
of Markov decision processes with linear temporal logic
constraints. IEEE Transactions on Automatic Control, 59(5):
1244–1257, May 2014.
[8] R. Fikes and N. Nilsson. Strips: A new approach to the
application of theorem proving to problem solving. Artificial
Intelligence, 2(3):189–208, 1971.
[9] J. Fu and U. Topcu. Probably approximately correct mdp
learning and control with temporal logic constraints. In
Proceedings of Robotics: Science and Systems, Berkeley, USA,
July 2014.
[10] A. Guez and J. Pineau. Multi-tasking SLAM. In IEEE
International Conference on Robotics and Automation, pages
377–384, 2010.
[11] M. Guo, K. H. Johansson, and D. V. Dimarogonas. Revising
motion planning under linear temporal logic specifications in
partially known workspaces. In IEEE International Conference
on Robotics and Automation, pages 5025–5032, 2013.
[12] B. Johnson and H. Kress-Gazit. Analyzing and revising
synthesized controllers for robots with sensing and actuation
errors. The International Journal of Robotics Research, 34(6):
816–832, 2015.
[13] S. Karaman and E. Frazzoli. Sampling-based algorithms
for optimal motion planning with deterministic µ-calculus
specifications. In American Control Conference, pages 735–
742, 2012.
[14] S. Karaman and E. Frazzoli. Sampling-based algorithms
for optimal motion planning. The International Journal of
Robotics Research, 30(7):846–894, 2011.
[15] M. Kloetzer and C. Belta. A fully automated framework
for control of linear systems from temporal logic specifications. IEEE Transactions on Automatic Control, 53(1):287–
297, 2008.
[16] H. Kress-Gazit, G. E. Fainekos, and G. J. Pappas. Temporallogic-based reactive mission and motion planning. IEEE
Transactions on Robotics, 25(6):1370–1381, 2009.
[17] O. Kupferman and M. Y. Vardi. Model checking of safety
properties. Formal Methods in System Design, 19(3):291–314,
2001.
[18] B. Lacerda, D. Parker, and N. Hawes. Optimal policy generation for partially satisfiable co-safe LTL specifications. In
International Joint Conference on Artificial Intelligence, pages
1587–1593, 2015.
[19] M. Lahijanian, J. Wasniewski, S. B. Andersson, and C. Belta.
Motion planning and control from temporal logic specifications with probabilistic satisfaction guarantees. In Robotics
and Automation (ICRA), 2010 IEEE International Conference
on, pages 3227–3232. IEEE, 2010.
[20] S. LaValle. Rapidly-Exploring Random Trees A New Tool for
Path Planning. Technical Report Computer Science Dept., TR
98-11, Iowa State University, 1998.
[21] M. Likhachev, G. Gordon, and S. Thrun. ARA* : Anytime
A* with Provable Bounds on Sub-Optimality. In Advances in
Neural Information Processing Systems, pages 767–774. 2004.
[22] S. C. Livingston, R. M. Murray, and J. W. Burdick. Backtracking temporal logic synthesis for uncertain environments. In
IEEE International Conference on Robotics and Automation,
pages 5163–5170, 2012.
[23] S. C. Livingston, P. Prabhakar, A. B. Jose, and R. M. Murray. Patching task-level robot controllers based on a local
µ-calculus formula. In IEEE International Conference on
Robotics and Automation, pages 4588–4595, 2013.
[24] P. E. Missiuro and N. Roy. Adapting probabilistic roadmaps
to handle uncertain maps. In IEEE International Conference
on Robotics and Automation, pages 1261–1267, 2006.
[25] A. Nüchter and J. Hertzberg. Towards Semantic Maps for
Mobile Robots. Robotics and Autonomous Systems, 56(11):
915–926, 2008.
[26] J. Pearl. Heuristics: Intelligent Search Strategies for Computer
Problem Solving. Addison-Wesley, 1984.
[27] J.-É. Pin. Mathematical foundations of automata theory.
Lecture notes LIAFA, Université Paris, 7, 2010.
[28] M. Pivtoraiko and A. Kelly. Generating near minimal spanning
control sets for constrained motion planning in discrete state
spaces. In IEEE/RSJ International Conference on Intelligent
Robots and Systems, pages 3231–3237, 2005.
[29] A. Pnueli. The temporal semantics of concurrent programs.
Theoretical Computer Science, 13(1):45 – 60, 1981.
[30] R. Salas-Moreno, R. Newcombe, H. Strasdat, P. Kelly, and
A. Davison. SLAM++: Simultaneous Localisation and Mapping at the Level of Objects. In IEEE Conf. on Computer
Vision and Pattern Recognition (CVPR), pages 1352–1359,
2013.
[31] E. M. Wolff, U. Topcu, and R. M. Murray. Robust control
of uncertain markov decision processes with temporal logic
specifications. In IEEE Annual Conference on Decision and
Control, pages 3372–3379, 2012.
[32] Z. Zhang, N. Sturtevant, R. Holte, J. Schaeffer, and A. Felner.
A* search with inconsistent heuristics. In International Joint
Conference on Artificial Intelligence, pages 634–639, 2009.
18
16
14
12
y [m]
10
8
6
4
2
0
-2
-10
-5
0
5
10
15
20
x [m]
(a) x0 = [−5, 14, π/6]T , total motion cost 24 m.
18
16
14
12
Fig. 2: Functions needed for LTL-constrained A* search
y [m]
10
8
6
4
2
0
-2
-10
-5
0
5
10
15
20
x [m]
(b) x0 = [15, 0, π/6]T , total motion cost 44 m.
Fig. 4: Robot trajectories obtained by the planning algorithm with
different initial conditions. The green triangle represents the mobile
robot. The blue dash circle indicates the unit ball and the red dash
circle indicates a ball of radius 2 m. The red sold ellipses represent
the 0.95-confidence region of the landmark pose distributions. The
landmark coordinates are drawn according to the mean of the pose
distribution.
Fig. 3: The set of motion primitives used for LTL-constrained A*
planning. Each segment contains 5 robot poses indicated by the red
triangles. Each robot pose represents a different motion primitive.
18
16
14
12
y [m]
10
8
6
4
2
0
-2
-10
-5
0
5
10
15
20
x [m]
(a) The optimal plan with 0.95-confidence and motion cost 38 m.
18
16
14
12
y [m]
10
8
6
4
2
0
-2
-10
-5
0
5
10
15
20
x [m]
(b) The optimal plan with 0.5-confidence and motion cost 34 m.
Fig. 5: Robot trajectories obtained by the planning algorithm
with different confidence parameters. The initial state is x0 =
[10, 16, 0.77π]T .
| 3cs.SY
|
PERFORMABILITY ASPECTS OF THE ATLAS VO;
USING LMBENCH SUITE
Fotis Georgatos
Department of Computer Science, University of Cyprus, Cyprus
[email protected]
John Kouvakis
Department of Mathematics, University of the Aegean, Greece
[email protected]
John Kouretis
Department of Physics, National Technical University of Athens, Greece
[email protected]
Abstract
1.
The ATLAS Virtual Organization is grid’s largest Virtual Organization which is
currently in full production stage. Hereby a case is being made that a user working within that VO is going to face a wide spectrum of different systems, whose
heterogeneity is enough to count as “orders of magnitude” according to a number
of metrics; including integer/float operations, memory throughput (STREAM)
and communication latencies. Furthermore, the spread of performance does not
appear to follow any known distribution pattern, which is demonstrated in graphs
produced during May 2007 measurements.
It is implied that the current practice where either “all-WNs-are-equal” or,
the alternative of SPEC-based rating used by LCG/EGEE is an oversimplification which is inappropriate and expensive from an operational point of view,
therefore new techniques are needed for optimal grid resources allocation.
Acknowledgements
We hereby wish to thank our respective supervising professors, M. Dikaiakos, I. Gkialas, T. Alexopoulos as well as Dr. G. Tsouloupas for their support
and constructive feedback at various opportunities during this work.
2
2.
Introduction and outline
Grid computing emphasizes on sharing of heterogeneous resources, which
might be based on different hardware or software architectures. In correspondence with this diversity of the infrastructure, the execution time of any single
job, as well as the total grid performance can both be affected substantially.
The objective of the current work is to document this effect with real-world
data.
A microbenchmarking technique using lmbench [4]has been applied in order to explore if this assumption is true within the context of the ATLAS Virtual
Organization; Indeed, results show that performance can vary, up to an order a
magnitude and the effects are even more apparent within this larger VO, than
what we found during earlier work within SEE VO [8][9].
Real grid characteristics serve as a proof that metrics-guided resource selection is nearly imperative, if not to optimally select resources, at least to
specifically avoid ones which are known a priori that they don’t perform as
good as required. Our results hint in favour of a more intelligent matchmaking
process which involves performance metrics.
3.
Related Work
A similar approach was taken by the developers of the GridBench platform
[1], which is a tool for evaluating the performance of Grids and Grid resources
through benchmarking. Measurements were taken from CrossGrid and the
LCG testbed for the purpose of resource characterization. Lately more tests
have been done on the EGEE infrastructure with more conclusive results [2].
Indeed, such benchmarking techniques have already been demonstrated to be
of interest [3]. The current static Information System-based practices should
be augmented by dynamic characterization of grid resources. There exists evidence that performance can be improved for a few application categories, by
applying metrics-correlation techniques on a case-by-case basis [1].
4.
Issues and Methodology
Benchmarks are standardized programs or detailed specifications of algorithms designed to investigate well-defined performance properties of computer systems in alignment with a widely demonstrated set of methods and
procedures. For many years, benchmarking has been used to characterize a
large variety of systems ranging from CPU architectures and caches to filesystems, databases, parallel systems, Internet infrastructures and middleware.
Computer benchmarking provides a commonly accepted basis for comparing
the performance of different computer systems in a highly repeatable and consistent manner, so it appears appealing to use it for providing resource metrics.
Performability aspects of the ATLAS VO; using lmbench suite
3
LCG/EGEE grid currently provides insufficient information about sites’ characteristics, which results in longer queue and job execution times and, indirectly, to more failures. The data available in the Information System is total
memory of a node, Operating System distribution name and version, processor
model and total CPUs per site. This information is not always complete and
the commonly used dataset that provides sites’ technical specifications, such as
processor model, total memory and total cpus per site are often inaccurate or
inconsistent, since they are manually edited, turning it far from optimal choices
as selection criteria, while finding the most appropriate ones among sites.
The basic job submission framework of this endeavour is already available
as an open source python code package, which is able to submit self-compiling
lmbench sources along with the related scripts that gather other system information -software & hardware- and collect back their reports. It is available
along with the rest of the code at: http://GetNRunBench.sourceforge.net
What is very important to specifically clarify, is that the benefits of applying the benchmarking technique and grid resource characterization can greatly
outnumber the measurement system’s overhead in itself; typically it can be run
at a rate of 1 test/site/day consuming less than 0.5% of the ATLAS Virtual Organization capacity, traded for a benefit which can be much larger, or at least
this is what we have conjectured from these first results.
5.
Results
The third release of lmbench (lmbench-3.0-a4) has been selected for its
better benchmarking techniques and more detailed and documented microbenchmarking tools [5]. During the execution of this activity a Technical Report has been produced which includes the full set of processed results [7].
In parallel to the lmbench activity scripts were run which collected data on
the grid environment, gathering information like: linux distribution and kernel, memory and swap configuration, CPU Model and Vendor as well as grid
middleware. This information, once collected, can both document the heterogeneity of the grid, as well as provide input for processes that have particular
needs for their execution environment. Please refer to the Technical Report for
further details or the raw dataset available in the grid manner at
’lfn:/grid/see/fotis/ATLAS lmbench 2007 Taurus dataset reduced-by-fotis.tgz’.
We hereby provide a subset of the data in pie and histogram graphs.
The SizeOfSite distribution graph Figure 1 presents the amount of sites that
exist within a given “size category”. Sites range in size, and their capacity can
be from a few jobs up to thousands. It is worthy to notice that the large majority
of resources, about 60% is provided by the top 10% of sites. Also: about 75%
by the top 20%, a 90% by just 40% of sites, 95% by the top 60% and 99% by
the top 80% of sites. The last 20% of sites contribute just 1% of resources.
4
Figure 1.
SizeOfSite Distribution
In order to better visualize results and instead of providing only minima and
maxima, we employed a “projection” technique: we attempted to derive histograms from our measurements combining data from the Information System
(BDII & GSTAT), attempting to correlate our multiple measurements and site
size, “pro-rata”. Projection was unavoidable since benchmarking all of EGEE
is impossible; The resulting graphs should be considered reliable on the horizontal axis, since all measurements have indeed taken place within the ATLAS
VO, but with a grain of salt on the vertical axis since large sites can contribute
a larger error margin. For the conclusions of this work, this effect is irrelevant;
We are looking into ways of minimizing the error margin or find bounds for it,
though.
As show in Figure 4, the most popular linux distribution is Scientific Linux
CERN v3.0.8, which can be explained due to the endorsement by the LHC
Computing Grid project: all large Tier-1 sites (many 100s or 1000s of CPUs)
and Tier-2 sites have to use it: Any system administrator of an LCG site has
to update her site promptly, in order to remain compatible with CERN experiments’ software. The next most popular distribution is Scientific Linux v3.0.x.
It is obvious that their dispersion is much higher, perhaps a by-product of their
system administrators following a more conservative upgrade policy, combined
with the lack of a centrally coordinated driving force, or just being driven by
more diverse needs.
The amount of Linux distributions that are concurrently in use on the grid
is extraordinary and include versions of CentOS, RedHat, SL, SLC, Ubuntu,
Debian and some more customized RedHat clones; it should be considered
fortunate if there are no applications failing due to that very reason: there are
many software tools that expect to find software in specific path locations, esp.
scripting languages like perl, python or php, or system libraries, so this variety
of systems could very well increase job failure rates.
Performability aspects of the ATLAS VO; using lmbench suite
Figure 2.
Environmentals
5
6
Figure 3.
Linux Kernel and CPU Model
Performability aspects of the ATLAS VO; using lmbench suite
Figure 4.
Linux Distributions
7
8
Figure 5.
LMbench measurements - Numerical operations
Performability aspects of the ATLAS VO; using lmbench suite
Figure 6.
LMbench measurements - Stream, Stream2
9
10
It is worthy to note that certain benchmark histograms manifest interesting
patterns. It is unlikely that they are poisson processes, at least as can be seen
from a grid user point of view:
Multiplication of 64bit integers “int64 mul” is a great example of grid
performance effects: the measurements can be split into two areas, those
of “fast” and “slow” clusters, thereby allowing us to apply during job
submission two distinct strategies: optimize for throughput (use both
sets of resources) or latency (use only fast ones). The values cluster between 1.5-8.5 nanoseconds and 10.0-15.0 nanoseconds; preliminary investigation hints that the effect is caused by architectural aspects, namely
the support for EM64T/x86-64 operations, which were introduced during year 2004 for 32bit processors.
Multiplication of simple precision floating point numbers (Float Mul)
excibits a cluster effect in the area 1.5-3.5 nanoseconds, while there are
peaks of performability at 5.2, 6.0 and 7.2 nanoseconds. For example,
if we execute a DAG workflow on the grid for which we want to minimize the makespan, it could prove beneficial if we exclude resources
with Float-Mul metric >3.5nanoseconds, trading-off a minor part of resources, assuming a float multiplication-dominant operation. Similar
patterns and conclusions could be seen with double precision multiplication (Double Mul) operation.
Addition of 64bit Integers is an operation that appears to happen in a
pattern which does not follow any known distribution, between 0.3-8.1
nanoseconds. This in effect is a ratio greater than 1:25, which implies
that the execution of any relevant workflow on the grid could easily endup being last-job-bound instead of CPU-bound: the processes of a certain group would have to wait before the last one of them finishes, and
this is quite easy to happen if this kind of operation is dominant. Excluding resources would be a difficult decision as well, because there is
no “optimal sweetspot”, so there is no obvious strategy, optimal strategy
has to be considered on a case-by-case basis.
As seen in Figure 7, doing two BIT operations of 32 bit data is apparently not the same as doing one 64 bit operation; in fact, the latter is
slower, which is perhaps counter-intuitive. This kind of information can
help grid algorithm developers to take the optimal decisions while constructing software to be run on the grid, at least in an instance of the grid
as it has right now. The idea can be generalized further in other similar
dilemmas, eg. when to trade a division operation for modulo and rotate
operations etc.
Performability aspects of the ATLAS VO; using lmbench suite
Figure 7.
Bit operations
(a) Integer 32 bits
6.
11
(b) Integer 64 bits
Analysis and Discussion
Timings of the various microbenchmarks are mostly being presented in
nanoseconds. Even though they appear very small, according to these measurements, if they occupy a repetitive part of a grid job they can have an important and highly-impacting factor: A job that could be executed in a site in
a time period T, in some other site could be executed in double that time, T*2;
ignoring at this point any communication overheads. In fact, it is possible to
find on the graphs ratios that are as high as 1:10, All STREAM benchmarks
show this effect quite well, while the same holds true in particular with multiplication operations, or the operations involving many bits (more than 32, 64
bit operations plus simple and double floating point).
It is important to realize that extended job lengths can result in higher job
failure rates, due to time window limits in sites’ queues, which is standard
practice.
In the special case that a grid job can be further parallelized in multiple
components, each part could be sent in the most suitable site depending on the
nature of the subprocess, and then the time of individual subprocesses will also
be decreased, and consequently the total execution time as well (we assume no
I/O overheads).
Finally, it is impossible for each user to know or measure the characteristics
of each site. Therefore some mechanism must exist that allows the matchmaking to happen in an automatic way. There is some ongoing discussion if the
best way to implement this would be through a job description technique (ie.
in the .jdl file), or at the global scheduling stage (ie. RB or WMS), or both. The
latter of course is advantageous, if it is combined with an Information System
that can provide such benchmarking results; then it becomes possible for the
middleware to identify the sites that are best for a specific job, assuming all
other issues equal. For one thing, resource ranking is deemed necessary.
12
7.
Future work
In order for resource ranking to be efficient, we have to remove the heterogeneity at the queue interface which is the optimal, most fine-grained, resource
handle. Having multiple queues within a site, CE in gLite parlance, should be
considered required if we are to support differentiated “subclusters”, where
Worker Nodes of similar performance aspects are grouped together. It is contemplated that if these metrics are consistent over time, it would be possible
to optimize grid performance further, simply by profiling the current infrastructure, storing the results, then using submission frameworks that can understand this information. Alternatively, we could create and store ensembles of
performance models [6], so that we are able to capture heterogeneity behind
submission queues, and trigger mechanisms which provide for a “heterogeneity” metric: this will allow the user to trade bulk performance and uncertainty
level for latency.
We hope that this information will be used for further cluster performance
research and that it will help future system administrators choose better hardware and/or software components during the deployment of new clusters. In
fact, a new era begins where instead of "brute-force" usage of resources, we
will be able to load-balance grids according to their true capabilities, and purchase new Grid hardware upgrades according to real user needs, just as is envisaged in power, transportation and communication systems. For instance,
should we buy more 64bit or, 32bit CPUs to upgrade a VO optimally?
We easily concluded during the analysis of raw data, that there areas for
improvement, since this study is unusually sensitive to various deficiencies or
shortcomings of data as they are provided by the Information System:
heizenbugs; transient faulty data on the Information System
dubious entries; permanent faulty data on the Information System
duplicate Computing Elements or plainly wrong number of CPUs; there
is difficulty in automatically extracting “safe” information
lack of SubClusters feature support; by far the most serious issue, and
the one for which we can do little about, yet.
Note also that once someone has to cope with a very high failure rate reaching nearly 50%, as we had, she will have to face overheads and unavoidably
narrower findings. As said earlier though, the robustness of the technique is
already severely limited by the native capabilities of the Information System
itself, at least as we know its implementation right now with gLite-based grids
(May 2007)
Performability aspects of the ATLAS VO; using lmbench suite
+
+
+
+
-
13
better understanding of performance-related job-failures
better job scheduling with substancial throughput gains
capability to deliver prompt service for urgent computing
archival of performance aspects of the systems
system deployment overhead; but submission framework is ready
consumption of resources; can be optimized for continuous use
the technique is sensitive in resource heterogeneity (queue level)
Table 1.
Continuous Grid Benchmarking: The arguments of a debate
As a final point in this discussion, we would to report some arguments for
and against continuous grid benchmarking, which are shown hereby.
8.
Conclusions
The tool used in this report has been lmbench, due to its rich set of microbenchmarks, as well as its availability and compatibility with Linux. It includes
latency, bandwidth and timing measurements. Furthermore, information has
been retrieved and reported about sites characteristics, such as linux kernel
version and distribution, grid middleware, cpu and memory size.
The value of this work is to provide some initial real-world experimental
data, in order to be able to evaluate possible strategies for future implementations.
The preliminary conclusion is that any typical grid can largely benefit from
even trivial resource characterization and match-making techniques, assuming
advantage is taken of this information early during job scheduling. In other
words we could obtain knowledge about sites’ performability with very little overhead and then take advantage of this information during scheduling to
minimize total execution time considerably.
One more conclusion is the internal heterogeneity of large clusters is alone
enough to drive much of the grid away from optimal performance: Once we
can’t differentiate among fast and slow Worker Nodes, we will be forced to
cope with the uncertainty of makespan for any single job. This can seriously
affect low-latency jobs which are characteristic of Urgent Computing applications, eg. earthquake and other natural hazard signal analysis, and impact badly
the total workflow execution time, even prevent us from certain categories of
domains, eg. a hurricane or Tsunami early warning system. The problem can
be corrected, if the “subclusters” concept is put to use: in fact, we claim that if
every single queue is always addressing a homogeneous resource pool, then we
would be able to do queue-based characterization of availability, performance,
stability; many aspects of the grid as we know it would be vastly improved.
14
We wish at this point to summarize our own understanding of the results:
a) Resource characterization and metrics driven-scheduling is imperative, if
we seek optimal grid latency & throughput within a VO like ATLAS
b) Resource characterization is effective, as long as collected metrics correspond to true performance of grid sites: “heterogeneity brings uncertainty”
and “uncertainty is expensive”, so “heterogeneity is expensive”,
Our results are quite conclusive in the aforementioned directions. We are eager to compare the results with those of other teams and make a bold statement
on the need for resource auditing and ranking within the Information System
itself. We claim that such an option can assist any Grid Infrastructure tremendously both in terms of throughput and latency and it is an implied requirement
for any generic large scale grid system, if efficiency is actively seeked for.
References
[1] George Tsouloupas, Marios D. Dikaiakos. GridBench: A Tool for the Interactive Performance Exploration of Grid Infrastructures. Journal of Parallel and Distributed Computing,
2007, Elsevier.
[2] G. Tsouloupas, M. D. Dikaiakos. GridBench: A Workbench for Grid Benchmarking. In
Advances in Grid Computing - EGC 2005. European Grid Conference, Amsterdam, The
Netherlands, February 14-16, 2005, Revised Selected Papers, Lecture Notes in Computer
Science, vol. 3470, pages 211-225, Springer, June 2005.
[3] Eamonn Kenny, Brian A. Coghlan, George Tsouloupas, Marios D. Dikaiakos, John Walsh,
Stephen Childs, David O’Callaghan, Geoff Quigley. Heterogeneous Grid Computing: Issues and Early Benchmarks. Lecture Notes in Computer Science, Computational Science,
ICCS 2005, pp. 870-874.
[4] Carl Staelin, Larry McVoy. LMBench : Portable tools for performance analysis In Proceedings USENIX Winter Conference, San Diego, CA, January 1996, pp. 279-284.
http://www.usenix.org/publications/library/proceedings/sd96/full papers/mcvoy.pdf
[5] Carl
Staelin,
Hewlett-Packard
Laboratories.
lmbench3:
scalability
Technical
Report,
HPL-2002-313,
November
http://www.hpl.hp.com/techreports/2002/HPL-2002-313.html
measuring
8,
2002,
[6] Zhang, S., Cohen, I., Goldszmidt, M., Symons, J. and Fox, A. DSN ’05. Ensembles of Models for Automated Diagnosis of System Performance Problems. Proceedings of the 2005 International Conference on Dependable Systems and Networks (DSN’05), pages 644-653,
2005. IEEE Computer Society. http://www.hpl.hp.com/techreports/2005/HPL-2005-3.pdf
[7] John Kouvakis, Fotis Georgatos. A Technical Report On Grid Benchmarking. TRMOD-07-001. University of the Aegean, March 2007. Available from arxiv.org with id
cs.PF/0705.1915v1 http://arxiv.org/abs/cs/0703086
[8] John Kouvakis, Fotis Georgatos. A Technical Report On Grid Benchmarking using ATLAS
V.O. TR-MOD-07-002. University of the Aegean, March 2007. Available from arxiv.org
with id cs.PF/0705.1915v1 http://arxiv.org/abs/0705.1915
[9] John Kouvakis, Fotis Georgatos. A Report on the Effect of Heterogeneity of the Grid Environment on a Grid Job; SEE VO. LSSC07, Sozopol, June 2007.
| 5cs.CE
|
MPC on manifolds with an application to the control of spacecraft attitude
on SO(3)
Uroš V. Kalabića,1, Rohit Guptaa , Stefano Di Cairanob,2, Anthony M. Blochc , Ilya V. Kolmanovskya
a Department
of Aerospace Engineering, University of Michigan, Ann Arbor, MI 48109
Mitsubishi Electric Research Laboratories, Cambridge, MA 02139
c Department of Mathematics, University of Michigan, Ann Arbor, MI 48109
arXiv:1509.08567v3 [math.OC] 23 Sep 2016
b Mechatronics,
Abstract
We develop a model predictive control (MPC) design for systems with discrete-time dynamics evolving on
smooth manifolds. We show that the properties of conventional MPC for dynamics evolving on Rn are
preserved and we develop a design procedure for achieving similar properties. We also demonstrate that for
discrete-time dynamics on manifolds with Euler characteristic not equal to 1, there do not exist globally
stabilizing, continuous control laws. The MPC law is able to achieve global asymptotic stability on these
manifolds, because the MPC law may be discontinuous. We apply the method to spacecraft attitude control,
where the spacecraft attitude evolves on the Lie group SO(3) and for which a continuous globally stabilizing
control law does not exist. In this case, the MPC law is discontinuous and achieves global stability.
Keywords: Model predictive control; Geometric control; Manifolds; Lie groups; Spacecraft attitude.
1. Introduction
Conventional model predictive control (MPC) [2]
is developed for and usually applied to systems
whose discrete-time dynamics evolve on the “flat”
normed vector space Rn . However, the configuration spaces of some systems are smooth manifolds that are not diffeomorphic to Rn . To design
the prediction dynamics for such systems, finitedimensional manifolds may be embedded in Rn ,
and standard integration schemes employed to derive the discrete-time update equation, enforcing
their evolution on the manifold through the use of
equality constraints. However, standard integration
techniques do not preserve symmetries for systems
evolving on manifolds, and this results in the integration not correctly representing the actual dyEmail addresses: [email protected] (Uroš
V. Kalabić), [email protected] (Rohit Gupta),
[email protected] (Stefano Di Cairano),
[email protected] (Anthony M. Bloch), [email protected]
(Ilya V. Kolmanovsky)
1 Preliminary results related to this research were presented in [1]. Corresponding author U. V. Kalabić. Tel.
+1-617-621-7548. Fax +1-617-621-7550.
2 This work was not supported by Mitsubishi Electric
Co. or any of its subsidiaries.
Preprint submitted to Automatica
namics. Because of this, specific methods for integrating system dynamics that evolve on manifolds
have been developed in, for example, [3, 4, 5].
As a motivating example, consider the attitude
dynamics of a rigid body evolving on the manifold SO(3). This manifold is a 3-dimensional space
which can be embedded in Rn for n ≥ 5 but is
not diffeomorphic to Rn for any n. The group
SO(3) is a particular case of a compact Lie group;
for mechanical systems whose configuration space
is a compact Lie group, the Lie group variational
integrator (LGVI) [6] has been developed to obtain discrete-time update equations that preserve
the underlying group structure. Unlike ordinary
integration schemes, the LGVI also preserves the
conserved quantities of motion [7] and is therefore
a more realistic prediction model. Also unlike ordinary integration schemes, the LGVI equations of
motion are significantly different from conventionally used, discrete-time dynamics. For instance, the
LGVI for spacecraft attitude gives an equation of
the form [8],
gk+1 = gk fk ,
fk+1 J −
T
Jfk+1
= Jfk −
(1a)
fkT J
2
+ h uk ,
(1b)
September 26, 2016
of spacecraft attitude, in which we apply our MPC
scheme to the LGVI dynamics evolving on SO(3).
The paper is organized as follows. Section 2
develops the MPC scheme and the conditions for
closed-loop stability. Section 3 describes the design of the terminal penalty and terminal set constraint achieving closed-loop stability by exploiting
a specifically designed local control law. Section
4 proves the non-existence of a globally stabilizing, continuous control law under certain assumptions, and hence the discontinuity of the MPC law.
Section 5 presents simulation results for dynamics
evolving on the Lie group SO(3) and Section 6 summarizes the conclusions.
where gk ∈ SO(3) is the orientation, fk ∈ SO(3)
is the change in gk , uk is related to the external
torques, h is the length of the integration step.
Indeed, (1b) is an implicit equation, and implicit
equations are seldom used in conventional MPC
prediction models.
The control of dynamics evolving on manifolds
present additional challenges. On manifolds with
Euler characteristic not equal to 1, there exist topological obstructions that imply the non-existence of
globally-stabilizing, continuous control laws. This
has been shown for the continuous-time setting [9],
and is a consequence of the Poincaré-Hopf theorem
[10]. In this paper, we appeal to the Lefschetz-Hopf
theorem and derive a similar result for discrete-time
dynamics.
In this paper we investigate the design of MPC
for systems evolving in manifolds, with the aim of
retaining properties of conventional MPC applied
to flat spaces. The applications of MPC to systems
whose dynamics evolve on manifolds have scarcely
been considered, with the exceptions of [11, 12],
which focus on computational issues, and have limited analysis of closed-loop properties.
In this paper3 , we first show that MPC on smooth
manifolds achieves the same properties of conventional MPC in Rn , namely, recursive feasibility and
asymptotic stability of the equilibrium. Then we
describe the design of the MPC terminal cost and
terminal set that achieves such properties.
As an additional contribution of this paper, we
show that when the manifold is compact, our MPC
is able to achieve global closed-loop asymptotic stability in the state-unconstrained case. In particular,
this property holds for manifolds with Euler characteristic not equal to 1, for which a continuous,
globally stabilizing control law does not exist. Because MPC can produce discontinuous control laws
[13], the topological obstruction for such manifolds
is not restrictive in this case. Thus this paper highlights the capability to design possibly discontinuous, stabilizing feedback control laws with global
stability properties, by employing a systematic and
unified design procedure in the MPC framework.
A practical contribution of this paper is the presentation of a control law for the constrained control
1.1. Notation
The notation is standard with a few notable exceptions. The set ZN denotes the set of the first N
nonnegative integers and Z+ denotes the set of all
nonnegative integers. For a set A, its interior is denoted by int A, its closure by cl A, and its boundary
by bd A. The set AN represents N copies of A. The
identity matrix is denoted In ∈ Rn×n and an n-bym zero matrix is denoted 0n×m ∈ Rn×m . For two
functions f, g : R → R, f ◦ g is their composition,
and g(t) = o(f (t)) implies that limt→∞ g(t)/f (t) =
0. A function α : [0, a) → [0, ∞) is said to be
class K if it is strictly increasing and α(0) = 0; furthermore α is said to be class K∞ if a = ∞ and
α(x) → ∞ as x → ∞. Finally, for a sequence
{vk , vk+1 , . . . , vk+N }, its predicted value at time k
is denoted by {vk|k , vk+1|k , . . . , vk+N |k }.
2. MPC on manifolds
We begin by developing a general MPC law for
application to smooth manifolds. In the construction of the control law, we follow ideas inspired by
conventional nonlinear MPC laws [2], for systems
whose dynamics evolve on Rn . For instance, in order to ensure recursive feasibility of the finite horizon optimal control problem and to enlarge the domain of attraction, we utilize a terminal set and terminal cost function [2]. We thereby obtain a result
on system stability that rigorously generalizes MPC
stability results to a more generous class of systems
whose dynamics evolve on manifolds. In this setting, we obtain a remarkable result that when the
manifold is compact and there are no state constraints, this MPC law may provide global asymptotic stability.
3 This work is a significant extension of the authors preliminary investigation in [1], which focused only on SO(3),
and did not present detailed proofs. The results in this paper
apply to general manifolds, and the theoretical results are
rigorously derived, including the non-existence of discretetime continuous control laws on certain types of manifolds.
2
Let M be an n-dimensional smooth manifold,
which is endowed with a metric d, and let U be a
compact subset of an m-dimensional smooth manifold. Note that the fact that all smooth manifolds
are metrizable follows from Whitney’s embedding
theorem [10]. Furthermore, all smooth manifolds
admit a Riemannian metric [14]. Consider the dynamic update equation,
xk+1 = f (xk , uk ),
(2a)
f : M × U → M,
(2b)
introduce a control law κ : M → U , which we refer
to as the local control law satisfying,
κ(xk ) ∈ U,
f (xk , κ(xk )) ∈ XT ,
(6b)
F (f (xk , κ(xk ))) − F (xk ) ≤ −L(xk , κ(xk )), (6c)
for all xk ∈ XT .
The MPC control law is obtained through the
solution to the following problem,
(7a)
min
VN xk ; {uk+i|k }i∈ZN ,
{uk+i|k }i∈ZN
where xk ∈ M and uk ∈ U . The evolution on manifold is highlighted by using M in the domain and
co-domain of (2b); the function f is continuously
differentiable and satisfies f (xe , ue ) = xe for some
xe ∈ M , which we refer to as the equilibrium of f ,
and some ue ∈ U .
The system is subject to state and control constraints,
xk ∈ X , uk ∈ U,
(3)
subject to xk+i+1|k = f (xk+i|k , uk+i|k ),
= F (xk+N |k ) +
X
L(xk+i|k , uk+i|k ),
(7b)
xk+i|k ∈ X ,
uk+i|k ∈ U, ∀i ∈ ZN ,
(7c)
(7d)
xk+N |k ∈ XT .
(7e)
When the solution to (7) exists, it is denoted by
∗
VN
(xk ) and the control sequence solving it is denoted by {u∗k+i|k }i∈ZN .
The input obtained from the model predictive
control law at time k is the first element in the
sequence solving (7),
where X and U are compact and connected subsets
of M and U , respectively, that satisfy xe ∈ int X
and ue ∈ int U.
We introduce a cost function VN : M × U N → R
satisfying,
VN xk ; {uk+i|k }i∈ZN
(6a)
uk = u∗k|k .
(8)
The MPC law defined above can be used on differentiable manifolds. We will now show that the
domain of attraction of the equilibrium xe of the
closed-loop system defined by (2) and (8) coincides
with the set of initial conditions that can be steered
using open-loop control to the target set without violating the constraints. Define,
(4)
i∈ZN
where xk+i+1|k = f (xk+i|k , uk+i|k ) is the predicted
state at time k + i + 1 for i ∈ ZN . The functions L :
M × U → R and F : M → R are twice continuously
differentiable and have the following properties,
L(xe , ue ) = F (xe ) = 0,
(5a)
DN = {xk ∈ M : ∃{uk+i|k }i∈ZN ∈ U
s.t. (7b)-(7e) are satisfied}. (9)
L(xk , uk ) ≥ L(xk , ue ) ≥ γ(d(xk , xe )),
F (xk ) ≥ α(d(xk , xe )),
(5b)
(5c)
We state the main result below and provide the
proof in the appendix.
for all xk ∈ M and uk ∈ U , where γ is a class K∞ ,
α is a class K function.
We now introduce the target set and terminal
feedback law. These are required in order to enforce recursive feasibility, ensuring that if the state
at the end of the prediction horizon is in the target
set, it and the associated feedback control satisfy all
constraints and are therefore feasible solutions to
the finite horizon optimal control problem. Specifically, we introduce the target set XT ⊂ X , which
is compact and contains xe in its interior. We also
Theorem 1. Let x0 ∈ DN . Then (i) the finite
horizon optimal control problem (7) is feasible for
all k ∈ Z+ and (ii) xe ∈ M is the asymptotically
stable equilibrium for closed-loop system defined by
(2) and (8) with a basin of attraction DN , i.e., for
any x0 ∈ DN and ε > 0, there exists δ > 0 such
that d(x0 , xe ) < δ implies that d(xk , xe ) < ε for all
k > 0, and furthermore, d(xk , xe ) → 0 as k → ∞.
Note that whenever DN = M , Theorem 1 implies
that the control law (8) is globally stabilizing. In
3
general, this may not be possible because DN is a
subset of X and X is a subset of M . However, if M
is a compact manifold, then X may be equal to M ;
in this case, a large enough N can be chosen so that
DN = X . Therefore, based on the following proposition, it is possible to guarantee global asymptotic
stability in the case where M is compact.
around the equilibrium [2]. In our approach, we
utilize diffeomorphisms to obtain local coordinates,
on which we construct a stabilizing controller; we
then transform the local control law to the coordinates of the original system.
To begin, because M is an n-dimensional manifold and U is a compact subset of an m-dimensional
manifold that contains ue in its interior, there exist
local diffeomorphisms at xe and ue [15, 10],
Proposition 2. Suppose M is compact and let D∞
be the set of initial conditions x0 such that there
exists a sequence of control inputs steering x0 to
xe ,
φ : V → M, ψ : W → U,
(11)
where V ⊂ Rn and W ⊂ Rm are open neighborhoods of xe and ue , respectively. Furthermore, because the equilibrium is in the interior of
both V and W , f is continuously differentiable,
and f (xe , ue ) = xe , there exists an open neighborhood V ′ × W ′ ⊂ V × W of (0, 0) such that
f (φ(V ′ ) × ψ(W ′ )) ⊂ φ(V ). Therefore,
D∞ = {xk ∈ M : ∃{uk+i|k }i∈Z+ ∈ U,
lim xk+i|k = xe }. (10)
k→∞
Suppose the system (2) is state-unconstrained so
that X = M and suppose D∞ = M . Then there
exists a finite N ∗ such that DN ∗ = M and therefore the closed-loop system defined by (2) and (8) is
globally asymptotically stable.
f ′ = φ−1 ◦ f ◦ (φ × ψ),
(12)
and the derivative of f ′ at (0, 0) is,
Proof. It is clear that for any finite N ≥ 0,
DN ⊂ D∞ = M and that the set cl(M \ XT ) =
cl(D∞ \ XT ) is a subset of D∞ . Let N (x) = inf{i :
xk = x, ∃{uk+i|k }i∈Z+ ∈ U, xk+i|k ∈ XT } be the
minimum number of control steps required to guide
an initial condition x ∈ D∞ = M to the set XT . According to the definition of D∞ and the fact that
xe ∈ int XT , N (x) is finite for all x ∈ cl(M \ XT ).
Let N ∗ = sup{N (x) : x ∈ cl(M \ XT )}. Since
cl(M \ XT ) is compact, N ∗ is finite.
Let x ∈ M . Therefore there exists a control
sequence {uk+i|k }k∈Z+ guiding xk = x to xe and
xk+N ∗ |k ∈ XT . The sequence {uk+i|k }k∈ZN ∗ is feasible for (7) because it satisfies both state and control constraints. Therefore, DN ∗ ⊃ D∞ = M .
By Theorem 1, M is the domain of attraction of
the closed-loop system defined by (2) and (8), and
therefore the closed-loop system is globally asymptotically stable at xe .
′
= dφ−1
df(0,0)
0 ◦ df(xe ,ue ) ◦ (dφ0 × dψ0 ).
(13)
′
Let ξk ∈ Rn and ηk ∈ Rm and let A = df(0,0)
◦
′
(In , 0m×n ) and B = df(0,0) ◦ (0n×m , Im ). Define a
linear update equation,
ξk+1 = Aξk + Bηk .
(14)
1
L ◦ (φ × ψ),
λ
(15)
Let,
L′ =
where 0 < λ < 1 is a scalar parameter. Consider
the discrete-time algebraic Riccati equation,
0 = AT P A − P + Q
T
(16)
T
−1
− (A P B + N )(BP B + R)
T
(A P B + N )T ,
where,
3. Local control law
The development of the control law (8) depends
on the design of target set XT and the local control
law κ : M → U with the properties enumerated in
(6). In this section, we describe a procedure to construct XT and κ that satisfy the properties in (6)
for dynamics (2) evolving on manifolds. For conventional MPC, the terminal set and terminal cost
are designed from a linearization of the dynamics
Q
NT
N
= Hess L′ (0, 0).
R
(17)
is the Hessian of L′ at (0, 0). A positive-definite
solution P to the algebraic Riccati equation (16)
exists if and only if the pair (A, B) is stabilizable
and the associated symplectic pencil does not have
eigenvalues on the unit circle [16]. A sufficient condition implying the latter is that Hess L′ (0, 0) is
positive-definite.
4
Note that,
Suppose a positive-definite solution P exists for
(16). Let κ′ : Rn → Rm be a stabilizing feedback
control law for (14) where,
f ′ (ξk , ηk ) = Aξk + Bηk + o(k(ξk , ηk )k),
Q N ξk
L′ (ξk , ηk ) = ξkT ηkT
N T R ηk
κ′ (ξk ) = −(BP B T + R)−1 (AT P B + N )T ξk . (18)
+ o(k(ξk , ηk )k2 ),
n
′
We introduce a set Pc ⊂ R and a function F :
Rn → R where,
Pc = {ξ ∈ Rn : F ′ (ξ) ≤ c},
′
T
F (ξ) = ξ P ξ.
F ′ (ξk ) = ξkT P ξk .
Therefore, F ′ (ξk+1 ) − F ′ (ξk ) + L′ (ξk , κ′ (ξk )) =
F ′ ◦ f ′ (ξk , κ′ (ξk )) − F ′ (ξk ) + L′ (ξk , κ′ (ξk )) =
o(kξk k2 ).
This implies that for xk ∈ Pc′ ,
F (xk+1 ) − F (xk ) + L(xk , κ(xk ))
=
F ◦
f (xk , κ(xk )) − F (xk ) + L(xk , κ(xk ))
=
F ′ ◦ f ′ (φ−1 (xk ), κ(φ−1 (xk ))) − F ′ (φ−1 (xk )) +
λL′ (φ−1 (xk ), κ(φ−1 (xk ))) + o(kφ−1 (xk )k2 ), and
therefore,
(19)
(20)
Because P is the solution to the algebraic Riccati
equation (16), the design of the control law (18)
implies that the set Pc is compact and invariant
with respect to the closed loop dynamics (14), (18).
Our goal is to use Pc and κ′ in order to design the
target set XT and local control law κ. The following
result is a proof of the existence of XT and κ.
F (xk+1 ) − F (xk ) + L(xk , κ(xk ))
= −(1 − λ)L′ (φ−1 (xk ), κ(φ−1 (xk ))) + o(kxk k2 ).
Because L′ is positive definite, for any choice of
0 < λ < 1, there exists c′′′ ≥ 0 such that (6b) and
(6c) are satisfied. Let c = min(c′′ , c′′′ ) to complete
the proof.
Proposition 3. Let F = F ′ ◦ φ−1 . Suppose there
exists a solution P > 0 to the algebraic Riccati equation (16). Then there exists c > 0 such that the set
XT ⊂ φ(Pc ) and control law κ = κ′ ◦ φ−1 satisfy the
assumptions of (6).
4. Globally stabilizing control law
The results of Section 2 show that the MPC law
(8) is globally asymptotically stabilizing under certain assumptions. We will now show that this implies that the control law generated by MPC is necessarily discontinuous for certain classes of manifolds. Specifically, we will show that there does
not exist a globally stabilizing, continuously differentiable control law for compact manifolds M
with Euler characteristic χ(M ) not equal to 1. The
Euler characteristic is a topological invariant that
restricts the possible combinations of sinks and
sources that can exist for a continuous vector field
defined on the manifold [10]. Generally, only a
manifold with Euler characteristic of 1 admits a
continuous vector field with one sole sink and no
other equilibrium. As an example, the normed vector space Rn has an Euler characteristic of 1, the
sphere S 2 has an Euler characteristic of 2, and all
compact Lie groups have an Euler characteristic of
0. Therefore, the sphere and compact Lie groups
do not admit continuous vector fields with one sink
and no other equilibrium.
We begin by noting that there exists no globally
stabilizing, continuously differentiable, continuoustime control law on compact manifolds M [9]. The
Proof.
Firstly, we show that F satisfies the
assumptions made in the design of the MPC
law. Because M is a smooth manifold, φ is
smooth and this, along with the smoothness of
F ′ , implies that F is smooth.
Furthermore,
F (xe ) = F ′ (φ−1 (xe )) = F ′ (0) = 0 which satisfies (5a). Finally, due to the equivalence of
norms on Rn and (5b), there exists r > 0 such
that F (x) = F ′ (φ−1 (x)) = φ−1 (x)T P φ−1 (x) ≥
rφ−1 (x)T Qφ−1 (x) = rL′ (φ−1 (x), 0) = rL(x, ue ) ≥
r · γ(d(x, xe )). Since r · γ is a class K function, (5c)
is satisfied.
If c > 0, then by definition (19), the set Pc
has a non-empty interior, which contains the origin. Because V ′ contains the origin, there exists
c′ > 0 such that Pc′ ⊂ V ′ and κ′ (Pc′ ) ⊂ U ′
and therefore φ(Pc′ ) = {x : F (x) ≤ c′ } because
F ′ (φ−1 (x)) = F (x).
Furthermore, because (xe , ue ) is in the interior
of X × U and κ′ is C 1 , there exists c′′ > 0 such
that (φ(ξ), κ(φ(ξ))) ∈ X × U for all ξ ∈ Pc′′ . It
remains to show that there exists a c > 0 such that
XT = φ(Pc ) satisfies conditions (6b) and (6c).
5
proof of the continuous-time result appeals to a
consequence of the Poincaré-Hopf theorem, which
states that the sum of fixed point indexes on M
must be equal to the Euler characteristic. Since
the index of a sink is equal to 1, if χ(M ) 6= 1, the
set of fixed points on M cannot consist of solely a
sink. In discrete-time, the analogous result to the
Poincaré-Hopf theorem is the Lefschetz-Hopf theorem, of which we use a strong version below.
systems, the assumption is generally valid because,
for such systems, the dynamics (2) correspond to
a discretization of a continuous-time system, i.e.,
if x(0) = xk and uk = 0, then xk+1 = φ(x(0), h),
where h is the discretization step and φ is the continuous flow map.
Corollary 6. As a result of Proposition 2 and
Theorem 5, if there exists a control sequence that
is capable of steering any initial condition to xe ,
then there exists a finite N such that the MPC law
is globally asymptotically stabilizing and therefore
the MPC law is discontinuous.
Theorem 4 (Lefschetz-Hopf [17]). Let M be a
compact smooth manifold and let f0 : M → M be
a continuous map that is homotopic to the identity
map id : M → M . Suppose the set of points x ∈ M
satisfying x = f0 (x) is finite. Then,
X
(21)
if0 (x0 ) = χ(M ),
Proof of Theorem 5. Assume κ is continuous.
This implies that f1 is continuous. By assumption, limk→∞ f1 (xk ) = xe for all x0 ∈ M . This
fact, along with the continuity of f1 implies that
M is connected. All connected manifolds are also
path-connected, so M is path-connected. Define a
continuous path pxk : [0, 1] → M from xe to xk .
Since f0 is homotopic to the identity, there exists a
homotopy H0 from idM to f0 . Define the map H1 :
M ×[0, 1] → M such that H1 (xk , t) = H0 (xk , 2t) for
t ∈ [0, 21 ) and H1 (xk , t) = f (xk , κ(pxk (2t − 1))) for
t ∈ [ 21 , 1]. Then, H1 (xk , 0) = xk and H1 (xk , 1) =
f1 (xk ). Because κ(xe ) = ue , H1 is continuous and
is therefore a homotopy between idM and f1 , i.e.,
f1 is homotopic to the identity.
The solePequilibrium is xe and if1 (xe ) = 1.
Therefore,
x0 ∈{x:x=f1 (x)} if1 (x0 ) = if1 (xe ) = 1.
However, χ(M ) 6= 1, and, as a consequence of the
Lefschetz-Hopf theorem, this leads to a contradiction. Therefore, κ cannot be continuous.
x0 ∈{x:x=f0 (x)}
where if0 (x0 ) is the fixed point index of f0 at x0 .
The fixed point index is a property characterizing
an equilibrium. Note that when the fixed point is an
asymptotically stable equilibrium, i.e., a sink, then
its fixed point index is equal to 1. This fact is used
in the proof of the discrete-time analogue to Brockett’s necessary condition [18] for systems where
M = Rn but, to the authors’ knowledge, an equivalent result for M with χ(M ) 6= 1 does not appear
elsewhere in the literature. Hence we present below a theorem proving that there exists no globallystabilizing, continuous, discrete-time control law on
manifolds M with χ(M ) 6= 1 whenever the closedloop dynamics are homotopic to the identity.
Theorem 5. Let M be a compact smooth manifold. Consider the dynamics (2) and assume the
unforced dynamics map f0 : x 7→ f (x, ue ) is homotopic to identity. Further assume that there exists
a control law uk = κ(xk ) with globally asymptotically stable equilibrium x = xe , and define the resulting closed-loop dynamics by xk+1 = f1 (xk ) =
f (xk , κ(xk )). If χ(M ) 6= 1, then κ : M → U is not
continuous.
5. Application to SO(3) and spacecraft attitude control
In this section, we apply the previously developed techniques to the constrained control of spacecraft attitude.4 The orientation of a spacecraft can
be uniquely represented by an element of the Lie
group SO(3), which is the group of all orthogonal
matrices whose determinant is equal to 1. Physically, the first, second, and third columns of an
element g ∈ SO(3) represent the direction of the
x, y, and z axes, respectively, when viewed from
a fixed frame in R3 . This allows us to derive the
Remark 1. The homotopy assumption is crucial
since it ensures that the unforced dynamics correspond to a continuous deformation of the system
from an equilibrium configuration. A case where
this assumption is violated is when the unforced
map satisfies f0 (xk ) = xe for all xk ∈ M . In this
case, κ(xk ) = 0 is an asymptotically stabilizing continuous control law. Nevertheless, for mechanical
4 This example originally appeared in [1] and in this paper,
we report additional details, derivations, and include further
discussion relevant to this example
6
discrete-time dynamics using Lie group variational
integrator techniques, resulting in a dynamic equation that ensures that state updates are elements of
SO(3) within a guaranteed numerical tolerance.
The LGVI spacecraft dynamics are given by (1).
In (1), gk ∈ SO(3) represents the spacecraft orientation, fk ∈ SO(3) is a one time-step change in
gk ∈ SO(3), and uk ∈ so(3) is related to the applied
torque τk by the equation uk = τk× , where so(3) is
the set of skew-symmetric 3-by-3 matrices5 . Note
that the dynamics (1) of [1] are equivalent to the
dynamics (1) in this paper. In [1], the dynamics
are presented in Hamiltonian form, whereas here
we use the Lagrangian form (see [6], which presents
dynamics in both forms, for details).
To solve the implicit equation (1b), we use the
procedure from [19]. At time k, the quantities fk
and uk are known, so we let Mk = Jfk −fkT J +h2 uk
T
so that fk+1 J − Jfk+1
= Mk . Then,
fk+1 = (Mk /2 + Sk ) J −1 ,
The linearized dynamics (14) corresponding to
(1) evolve on R3 × R3 and, according to [8], are
given by,
ζ
ζk+1
(26)
= A k + Bτk ,
ωk
ωk+1
where ζk , ωk ∈ R3 satisfy ζk× = Log gk and
(hωk )× = Log fk , where Log is the standard branch
of the matrix logarithm function. The matrices A
and B are,
0
I
h
.
(27)
, B=
A= 3
hI3
0 I3
In the language of Section 3, the composition of
the map ·× and the matrix exponential exp from ζk
to gk and from hωk to fk give the diffeomorphism
φ. The diffeomorphism ψ is the map ·× . We now
define L′ according to (15). Specifically, we choose
0 < λ < 1 so that,
(22)
λL′ (ζk , ωk , τk ) = tr(Qg (I3 − exp(ζk× )))
1
1
+ 2 tr(Qf (I3 − exp(hωk )× )) + tr(τk×T Rτk× ),
h
2
where S is the solution to the algebraic Riccati
equation,
− (Mk /2)T Sk − Sk (Mk /2) − Sk2 + (J 2 + Mk2 /4) = 0.
(23)
Note that (23) is solvable if and only if the term
J 2 + Mk2 /4 is positive semi-definite, i.e.,
J 2 + Mk2 /4 0,
which implies that,
λL′ (ζk , ωk , τk ) =
(24)
which is a convex constraint that can be enforced
by the MPC law [1].
1 T
1
1
ζk Q̃g ζk + ωkT Q̃f ωk + τkT R̃τk
2
2
2
+ o(k(ζk , ωk )k2 ), (28)
where Q̃g = tr(Qg )I3 − Qg , Q̃f = tr(Qf )I3 − Qf ,
and R̃ = tr(R)I3 − R.
The equation (28) can be used to define the Hessian matrix of (17). We use (28) in order to construct the terminal cost F and target set XT ⊂ Pc
according to the steps outlined in (16)-(20).
Thus, the MPC law (8) has been defined by the
finite horizon optimal control problem (7), which
has been formulated by using the constraint sets X
and U together with the cost functions F and L and
the terminal set XT designed above.
We now consider rest-to-rest controllability for
the dynamics (1) subject to constraints. We show
that, given a long enough prediction horizon, the
MPC control law is able to stabilize to the origin
any spacecraft attitude gk that is near to rest. In
the case that we consider, there is no constraint
on the spacecraft rotation gk ; however there may
be limitations on the rate of change fk or allowed
torque τk .
5.1. MPC law
In this section, we develop an MPC-based spacecraft attitude controller. We begin by choosing an
appropriate cost function of the form (4) with,
L(gk , fk , uk ) = tr(Qg (I3 − gk ))
1
1
+ 2 tr(Qf (I3 − fk )) + tr(uT
k Ruk ), (25)
h
2
and with positive-definite symmetric matrices Qg ,
Qf , and R. Next we construct a locally stabilizing
control law.
5 ·× is the map from elements of R3 to elements of so(3)
which preserves the cross product under multiplication, i.e.,
if C = a× , then a × b = Cb for any b ∈ R3 , where × is the
cross-product.
7
Corollary 7. Let the set DN be defined according
to the definition in (9) and the dynamics in (1).
Assume SO(3) × {I3 } ⊂ X , where X satisfies the
properties given in Section 2. Then there exists a
finite number N such that SO(3) × V ′ ⊂ DN , where
V ′ ⊂ SO(3) is an open set containing I3 .
60
40
τ3 (Nm)
20
The proof of the corollary is available in [1].
0
-20
5.2. Simulation results
We present a numerical simulation of the MPC
law on SO(3). In the simulation, in order to provide
an example of discontinuity in the MPC law, we
simulate a rest-to-rest rotation from a rotation of
180 degrees to I3 .
For the cost function (25), we choose the following matrices: Qg = I3 , Qf = J, R = 2I3 , and
λ = 0.1. In both cases, the terminal constraint
set is chosen as XT = Pc , where Pc is defined as in
(19) and is chosen to be as large as possible. We set
N = 10 and, U is large enough that the constraints
are not active during our simulation.
Theoretically, a control law that is globally stabilizing should exhibit a discontinuity, so our goal
for the simulation is to confirm that this is true
of our controller. The results for two simulations
are presented in Figs. 1-4. In the first simulation,
the initial rotation is 180 degrees about the z-axis,
and in the second case, it is close but not equal to
−180 degrees about the z-axis. The torque and angular velocity on the z-axis are shown in Figs. 1-2,
respectively, where we can observe that, although
the initial conditions are very close to each other,
the trajectories are almost opposite in sign. This
shows that there is a discontinuity in the control
law. In fact, the discontinuity occurs exactly at the
location of the 180 degree rotation or at any resting
initial condition g0 for which tr(g0 ) = −1, because
these points lie on the branch cut of the Log function.
In Figs. 3-4, the direction of rotation is marked
with an arrow. We see that in one case, the control
rotates the spacecraft counter-clockwise whereas, in
the other case, the rotation is clockwise.
-40
-60
0
1
2
3
4
5
6
7
8
9
10
time (s)
Figure 1: τk,3 for initial rotation of 180 degrees (solid) and
−0.99 · 180 (dotted)
4
3
2
0
3
ω (rad/s)
1
-1
-2
-3
-4
0
1
2
3
4
5
6
7
8
9
10
time (s)
Figure 2: ωk,3 for initial rotation of 180 degrees (solid) and
−0.99 · 180 (dotted)
2
z
1
0
1
−1
0.5
0
6. Conclusion
−1
This paper presented a general MPC theory for
dynamics that evolve on smooth manifolds. An
MPC scheme was developed for manifolds that generalizes the results of conventional MPC in the case
of Rn . In the case of MPC on manifolds, a general
−0.5
0
−0.5
0.5
1
−1
y
x
Figure 3: Orientation maneuver corresponding to the initial
rotation of 180 degrees plotted at 2s increments
8
Note that if f (xe ) = xe , then f k (xe ) = xe , for all
k ∈ Z+ , where f k+1 = f ◦ f k for all k ∈ Z+ and
f 0 = idM .
Definition 3. An equilibrium point xe ∈ M is said
to be Lyapunov stable if for any open neighborhood U ⊂ M of xe , there exists an open neighborhood W ⊂ M of xe such that for all x0 ∈ W ,
f k (x0 ) ∈ U , for all k ∈ Z+ . It is said to be locally
asymptotically stable if it is Lyapunov stable and for
all x0 ∈ W , limk→∞ f k (x0 ) = xe or, equivalently,
limk→∞ d(f k (x0 ), xe ) = 0.
2
z
1
0
1
−1
0.5
−2
0
−1
−0.5
−0.5
0
0.5
1
−1
y
x
Definition 4. Let xe be an equilibrium point of
(A.1). A function V : M → R is a Lyapunov
function if there exist two class-K functions α1 , α2 ,
that satisfy α1 (de (x)) ≤ V(x) ≤ α2 (de (x)) for
all x in some neighborhood of xe , and ∆V(x) :=
V(f (x)) − V(x) is negative semi-definite in a neighborhood of xe . If ∆V(x) is negative definite in a
neighborhood of xe , then V is a strict Lyapunov
function.
Figure 4: Orientation maneuver corresponding to the initial
rotation of −0.99 · 180 degrees plotted at 2s increments
construction for a locally valid control law on manifolds was derived, which was used in the design of
the MPC terminal cost and target set.
Results were presented that showed the globally
stabilizing properties of the MPC scheme as well as
the nonexistence of continuous discrete-time control laws on manifolds whose Euler characteristic
is not equal to 1. Thus, MPC achieves stability by
producing a discontinuous control law, even in cases
where a continuous globally stabilizing law does not
exist. The authors believe that the ability to generate, possibly discontinuous, stabilizing feedback
control laws for systems whose dynamics evolve on
manifold is appealing for practical applications.
Finally, an application of the results to a constrained spacecraft attitude control problem was
presented in the case of the matrix Lie group SO(3).
The simulation results that were presented showed
the expected discontinuity in the MPC law.
More formally, the above definition implies that
V is a Lyapunov function if there exists a neighborhood V of xe such that V(x) > 0, ∆V(x) ≤ 0
for all x ∈ V \ {xe }, and V(xe ) = 0. If, instead of
∆V(x) ≤ 0, the condition ∆V(x) < 0 holds for all
x ∈ V \ {xe }, then the Lyapunov function is a strict
Lyapunov function.
Theorem 8. Let xe be an equilibrium point of
(A.1).
(i) If there exists a Lyapunov function V, then xe
is Lyapunov stable.
(ii) If there exists a strict Lyapunov function V,
then xe is locally asymptotically stable.
Appendix A. Proof of Theorem 1
Proof. The proof follows arguments similar to
the one given for the continuous-time case in [20].
In order to prove Theorem 1, we rely on discretetime Lyapunov stability theory for smooth manifolds. Therefore, we begin by presenting the
Lyapunov stability analysis. Let M be an ndimensional smooth manifold with metric d. Consider the following discrete-time dynamical system,
xk+1 = f (xk ),
(i) Let U ⊂ M be a neighborhood of xe . As
a consequence of Lemma 2.23 in [14], M admits a locally finite cover by precompact sets,
implying that there exists a neighborhood
U ′ ⊂ M of xe that is a subset of the union
of finitely many precompact sets, and therefore is itself precompact. Let U ′′ ⊂ U ∩ U ′
be an open neighborhood of xe on which
α1 (de (x)) ≤ V(x) ≤ α2 (de (x)) for all x ∈ U ′′ .
Let Vε := {x ∈ M : de (x) ≤ ε} and choose
ε > 0 so that Vε is a subset of U ′′ . By
(A.1)
where xk ∈ M and f : M → M is a continuous
function.
Definition 2. A point xe ∈ M is called an equilibrium point of (A.1) if f (xe ) = xe .
9
construction, Vε is a closed subset of a precompact set and is therefore compact. Let
Vε (xe ) ⊂ Vε be the connected component of Vε
containing xe . The interior of Vε (xe ) is nonempty and contains xe . Let x0 ∈ int Vδ (xe )
where δ := α−1
2 (α1 (ε)) ≤ ε. Since V(x0 ) ≤
α2 (de (x0 )) ≤ α2 (δ) and V(f k (x0 )) − V(x0 ) =
Pk
j
j−1
(x0 ))) ≤ 0, it folj=1 (V(f (x0 )) − V(f
k
lows that V(f (x0 )) ≤ V(x0 ) ≤ α2 (δ), which
k
implies that de (f k (x0 )) ≤ α−1
1 (V(f (x0 ))) ≤
α−1
(α
(ε))
=
ε
for
all
k
∈
Z
.
Because
Vε is
1
+
1
compact and V is positive definite in a neighborhood of xe , there exists a constant ε′ ≤ ε
such that Vε′ consists of only one connected
component, so that Vε′ = Vε′ (xe ). As a consequence, for any neighborhood U , there exists W = int Vε′ such that for any x0 ∈ W ,
f k (x0 ) ∈ U for all k ∈ Z+ .
(ii) Let U ⊂ M be a neighborhood of xe . Choose
ε′ > 0 so that V is positive definite, ∆V is
strictly negative definite on Vε′ , and Vε′ is a
compact set with only one connected component. Let W = int Vε′ as in the proof of (i),
implying that xe is Lyapunov stable. For all
x0 ∈ W , the sequence {α1 (de (f k (x0 )))}k∈Z+
is a non-increasing sequence bounded below
by 0, and hence limk→∞ α1 (de (f k (x0 ))) =
β ≥ 0. We can show that β = 0 by contradiction. Assume β > 0, implying that the
set Vε′ \ int Vα−1 (β) does not contain xe . Be1
cause this set is compact and does not contain xe , the strictly negative definite, continuous function ∆V attains a negative maximum on this set. Denote the maximum by
−σ < 0. For any x0 ∈ W , V(f k (x0 )) =
P
V(x0 ) + kj=1 (V(f j (x0 )) − V(f j−1 (x0 ))) ≤
V(x0 ) − kσ. For any k > V(x0 )/σ, it follows that α1 (de (f k (x0 ))) ≤ V(f k (x0 )) < 0,
which contradicts the fact that α1 ◦ de is positive definite. Therefore β = 0 and, because
α1 ◦ de is continuous and positive definite,
limk→∞ f k (x0 ) = xe .
Assume xk ∈ DN so that there exists an optimal sequence {u∗k|k , . . . , u∗k+N −1|k } solving (7). Together, xk+1 = f (xk , u∗k|k ) and x∗k+N |k ∈ XT imply
that {u∗k+1|k , . . . , u∗k+N −1|k , κ(x∗k+N |k )} is feasible
for (7) at time k +1, which implies that xk+1 ∈ DN .
Since x0 ∈ DN , then by induction, xk ∈ DN for all
k ∈ Z+ , proving (i).
For xk ∈ XT , let xk+i+1|k = f (xk+i|k , κ(xk+i|k ))
for i ∈ ZN and xk|k = xk .
Let Jκ :
XT → R be a function such that Jκ (xk ) =
VN (xk , {κ(xk+i|k )}k∈ZN ). The function Jκ is positive definite on XT and continuous, because κ and
VN are continuous.
Let J : DN → R be a function such that J(xk ) =
∗
VN
(xk ). Let J ∗ = J|XT be the restriction of J to
the domain XT ⊂ DN . Let α1 = γ + N α. To prove
that J ∗ is a Lyapunov function we use the following
result.
Lemma 9. Let Jκ : M → R be a function that
is continuous and positive definite on a compact
neighborhood V ⊂ M of xe . Then there exists a
class-K function α2 such that Jκ (x) ≤ α2 (de (x))
for all x ∈ V .
Proof. Let Jm (δ) = maxx∈V {Jκ (x) : de (x) ≤
δ}. By construction, Jm is a continuous and nondecreasing function with domain [0, aNa ], where 0 <
aNa < ∞, and satisfies Jm (0) = 0. This implies
that the domain can be decomposed into finitely
many closed intervals [0, aNa ] = ∪i∈ZNa [ai , ai+1 ],
where Jm (δ) is strictly increasing on [ai , ai+1 ] when
i is even and Jm (δ) is constant on [ai , ai+1 ] when i
is odd.
Choose a small, positive scalar value σ > 0.
Let α2 (0) = 0, α2 (δ) = α2 (ai ) − Jm (ai ) + Jm (δ)
for δ ∈ [ai , ai+1 ] when i is even, and α2 (δ) =
δ−ai
α2 (ai ) + σ ai+1
−ai for δ ∈ [ai , ai+1 ] when i is
odd. By construction, α2 is a class-K function
and Jm (δ) ≤ α2 (δ) for all δ ∈ [0, aNa ]. Because
Jκ (x) ≤ Jm (de (x)) ≤ α2 (de (x)) for all x ∈ V , the
proof is complete.
According to the lemma, there exists a function α2 such that α1 (de (x)) ≤ J ∗ (x) ≤ Jκ (x) ≤
α2 (de (x)) for all x ∈ XT , where α1 and α2 are classK functions.
By optimality, J(xk+1 ) − J(xk ) ≤ −F (x∗k+N |k ) −
L(xk , uk ) + F (x∗k+N +1|k ) + L(x∗k+N |k , κ(x∗k+N |k )) ≤
−L(xk , uk ) ≤ −γ(d(xk , xe )) for any xk ∈ DN . This
implies that γ(de (xk )) → 0 as k → ∞, and therefore xk → xe as k → ∞ for any xk ∈ DN . Therefore
Remark 5. Note that, because f is a function
whose domain and codomain are equal to M , f k ,
where k ∈ Z+ , is also a function whose domain
and codomain are equal to M . Therefore, the existence and uniqueness of the sequence {f k (x0 )}k∈Z+
is guaranteed for any x0 ∈ M .
Proof of Theorem 1. Given xk , denote by x∗k+i|k
the predicted state at time k + i by (7) at time k.
10
the equilibrium of xe is asymptotically stable and
its domain of attraction is DN , proving (ii).
[20] F. Bullo, A. D. Lewis, Geometric Control of Mechanical
Systems, Springer, New York, 2005.
References
[1] U. Kalabić, R. Gupta, S. Di Cairano, A. Bloch, I. Kolmanovsky, Constrained spacecraft attitude control on
SO(3) using reference governors and nonlinear model
predictive control, in: Proc. American Control Conf.,
Portland, OR, 2014, pp. 5586–5593.
[2] J. B. Rawlings, D. Q. Mayne, Model Predictive Control:
Theory and Design, Nob Hill, Madison, WI, 2009.
[3] H. Munthe-Kass, High order Runge-Kutta methods on
manifolds, Appl. Numer. Math. 29 (1) (1999) 115–127.
[4] J. E. Marsden, M. West, Discrete mechanics and variational integrators, Acta Numerica 10 (2001) 357–514.
[5] E. Hairer, C. Lubich, G. Wanner, Geometric Numerical
Integration, Springer-Verlag, Berlin, Germany, 2006.
[6] T. Lee, Computational geometric mechanics and control of rigid bodies, Ph.D. thesis, University of Michigan, Ann Arbor (2008).
[7] T. Lee, N. H. McClamroch, M. Leok, A Lie group
variational integrator for the attitude dynamics of a
rigid body with applications to the 3D pendulum, in:
Proc. IEEE Int. Conf. Control Applicat., Toronto, ON,
2005, pp. 962–967.
[8] T. Lee, M. Leok, N. H. McClamroch, Optimal attitude
control of a rigid body using geometrically exact computations on SO(3), J. Dynamical Control Syst. 14 (4)
(2008) 465–487.
[9] S. P. Bhat, D. S. Bernstein, A topological obstruction to
continuous global stabilization of rotational motion and
the unwinding phenomenon, Systems & Control Letters
39 (1) (2000) 63–70.
[10] V. Guillemin, A. Pollack, Differential Topology, Prentice Hall, Englewood Cliffs, NJ, 1974.
[11] Z. Nagy, R. Findeisen, M. Diehl, F. Allgöwer, H. Georg
Bock, S. Agachi, J. P. Schlöder, D. Leineweber, Realtime feasibility of nonlinear predictive control for large
scale processes - a case study, in: Proc. American Control Conf., Chicago, IL, 2000, pp. 4249–4253.
[12] S. Gros, M. Zanon, M. Vukov, M. Diehl, Nonlinear MPC and MHE for mechanical multi-body systems with application to fast tethered airplanes, in:
Proc. Nonlin. Model Predictive Control, Noordwijkerhout, Netherlands, 2012, pp. 86–93.
[13] E. S. Meadows, M. A. Henson, J. W. Eaton, J. B. Rawlings, Receding horizon control and discontinuous state
feedback stabilization, Int. J. Control 62 (5) (1995)
1217–1229.
[14] J. M. Lee, Introduction to Smooth Manifolds, Springer,
New York, 2013.
[15] J. M. Milnor, Topology from the Differential Viewpoint,
Princeton University Press, Princeton, NJ, 1997.
[16] W. F. Arnold III, A. J. Laub, Generalized eigenproblem
algorithms and software for algebraic Riccati equations,
Proc. IEEE 72 (12) (1984) 1746–1754.
[17] A. Granas, J. Dugundji, Fixed Point Theory, Springer,
New York, 2003.
[18] W. Lin, C. I. Byrnes, Design of discrete-time nonlinear control systems via smooth feedback, IEEE T. Automat. Control 39 (11) (1994) 2340–2346.
[19] J. R. Cardoso, F. S. Leite, The Moser-Veselov equation,
Linear Algebra Applicat. 360 (2003) 237–248.
11
| 3cs.SY
|
arXiv:1708.09149v2 [cs.IT] 19 Oct 2017
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER
EXPECTED EMERGENT OPEN-ENDEDNESS
FELIPE S. ABRAHÃO
KLAUS WEHMUTH
ARTUR ZIVIANI
National Laboratory for Scientific Computing (LNCC)
25651-075 – Petropolis, RJ – Brazil
Abstract. This article investigates emergence and complexity in complex
systems that can share information on a network. To this end, we use a theoretical approach from information theory, computability theory, and complex
networks. One key studied question is how much emergent complexity arises
when a population of computable systems is networked compared with when
this population is isolated. First, we define a general model for networked
theoretical machines, which we call algorithmic networks. Then, we narrow
our scope to investigate algorithmic networks that optimize the average fitnesses of nodes in a scenario in which each node imitates the fittest neighbor
and the randomly generated population is networked by a time-varying graph.
We show that there are graph-topological conditions that make these algorithmic networks have the property of expected emergent open-endedness for
large enough populations. In other words, the expected emergent algorithmic
complexity of a node tends to infinity as the population size tends to infinity.
Given a dynamic network, we show that these conditions imply the existence
of a central time to trigger expected emergent open-endedness. Moreover, we
show that networks with small diameter compared to the network size meet
these conditions. We also discuss future research based on how our results are
related to some problems in network science, information theory, computability
theory, distributed computing, game theory, evolutionary biology, and synergy
in complex systems.
E-mail addresses: [email protected], [email protected], [email protected].
A preliminary version of some of the results in this article was orally presented in the the 8th
International Workshop on Guided Self-Organization at the Fifteenth International Conference on
the Synthesis and Simulation of Living Systems (ALIFE) under the title “Emergent algorithmic
creativity on networked Turing machines” with an extended abstract available at http://guidedself.org/gso8/program/index.html .
Authors acknowledge the partial support from CNPq through their individual grants: F. S.
Abrahão (313.043/2016-7), K. Wehmuth (312599/2016-1), and A. Ziviani (308.729/2015-3). Authors also acknowledge the INCT in Data Science – INCT-CiD (CNPq 465.560/2014-8).
1
2ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
emergence; complexity; open-endedness; networked Turing machines; time centrality; small diameter
[2010]68Q30; 68Q05; 05C82
1. Introduction
The general context of our research is to mathematically investigate emergence
and complexity when a population of complex systems is networked. That is,
we study emergence and complexity in a system composed of interacting complex
systems. As supported by [9], the pursuit of a universal framework for the problem of complexity in complex networks is paramount for a wide range of topics
in complex systems. The present work relies on an intersection between information theory, computability theory, game theory, distributed computing, multiagent
systems, communication complexity, adaptive complex systems, and biology. In
fact, we discuss in further detail in Section 10 that the investigated problem is
connected to questions ranging, for instance, from the problem of symbiosis [50],
cooperation [7, 8, 42], and integration [32, 56] to biological [75], economic [7, 66, 67],
and social [52] networks.
In particular, we study the problem of how a group of randomly generated computable systems give rise to emergent phenomena through the exchange of information and how it would affect the overall performance, fitness, or payoff. We
create a theoretical toy model from which one can investigate how much emergent
complexity the whole group has on average when they are networked compared to
when they are isolated. As we will formally show later in this paper, the randomly
generated algorithmic network NBB that plays the Busy Beaver imitation game
(BBIG) — see Section 3 — is a mathematical object useful to prove fruitful theorems using well-known results from statistics, computability theory, information
theory, and graph theory.
Thus, how does one define a measure of average emergent information/complexity
for a randomly generated population of Turing machines? Within the context of
this abstract toy model, we answer this question by generalizing the results for
open-ended evolutionary systems under algorithmic information theory (AIT), as
presented in [1–4, 21–23, 43], in order to present a new mathematical phenomenon of open-endedness. We will show that there are graph-topological conditions
that trigger expected emergent open-endedness (EEOE), i.e., that trigger an unlimited increase of expected emergent complexity as the population size goes to
infinity (see Section 4). It is an akin — but different — phenomenon to evolutionary open-endedness: Instead of achieving an unbounded quantity of algorithmic
complexity over time (or successive mutations), an unbounded quantity of emergent algorithmic complexity is achieved as the population size increases (see Section
4.2).
We start by defining a general encompassing mathematical model that we call
algorithmic networks in Section 2. This definition relies on a population of arbitrarily chosen theoretical machines and relies on a multiaspect graph [76], making
the nodes in the respective multiaspect graph to correspond to this population of
systems/programs (see Definition 2.2.3) such that edges are communication channels that nodes/systems can use to send and receive information. Thus, as it was
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS3
our intention described in the first paragraph, an algorithmic network is a network composed of algorithms as nodes, with each node representing a computable
system.
Then, we introduce in Section 3 a particular model of synchronous dynamic
algorithmic networks NBB that is based on simple imitation of the fittest neighbor:
a type of algorithmic network that plays the BBIG. A network Busy Beaver game
is a game in which each player is trying to calculate the biggest integer1 it can
using the information shared by its neighbors. The BBIG is a variation of the
Busy Beaver game in which every node can only propagate the biggest integer,
taking into account the one produced by itself and the ones from their neighbors.
It configures a simple imitation-of-the-fittest procedure. Thus, these algorithmic
networks NBB can be seen as playing an optimization procedure where the whole
pursues the increase of the average fitness/payoff through diffusing on the network
the best randomly generated solution (see discussion on Section 3.1).
We present our main Theorem 8.1 proving that there is a lower bound on the
expected emergent algorithmic complexity for algorithmic networks NBB . Additionally, we prove Corollary 8.1.1 showing that this lower bound can be calculated
from a diffusion measure like cover time [27]. Further, from this corollary, we also
prove in Theorem 8.2 that there are asymptotic conditions on the increasing power
of the cover time as a function of the population size such that they ensure that
there is a central time to trigger EEOE. Then, we introduce in Section 9 a small
modification on the family of multiaspect graphs of NBB with the purpose of investigating what would happen if the time-varying networks present a small diameter
compared to the network size, i.e. D(Gt , t) = O(lg(N )). Indeed, in this case, we
show in Corollary 9.1 that a small diameter is sufficient to ensure the existence
of a central time2 to trigger EEOE for algorithmic networks NBB — even in a
computably bigger number of cycles compared to the temporal diffusion diameter.
Additionally, we discuss in Section 10 future research from how our results are
related to problems in network science, statistical (or probabilistic) information
theory, computability theory, distributed computing, game theory, evolutionary
biology, and synergy in complex systems. We also choose to add discussions in the
Sections in order to improve the readability and explanation of the new definitions
and new models that we will introduce in the present article. Finally, Section 11
concludes the paper. Complementarily, the appendix shows extended versions of
the proofs of Lemmas 8.1, 8.2, 8.3, 8.4, 8.5, and 8.6, as well as of Theorems 8.1
and 8.2.
2. Algorithmic Networks
2.1. Discussion on algorithmic networks. In this section we will define a general mathematical model for the study of networked machines which can share
information with each other across their respective network while performing their
computations. We want to define it in a general sense in order to allow future variations, to add specificities and to extend the model presented in Section 3, while
still being able to formally grasp a mathematical analysis of systemic features like
the emergence of information and complexity along with its related phenomena:
1 As established as our measure of fitness or payoff. See Section 3.
2 Or just to trigger EEOE, if the graph is static. See 9.1.
4ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
for example, the average emergent open-endedness in our case (see discussion 3.1
and Section 4).
Since, following this general approach, one can see these mathematical models as
a merger of algorithmic (and statistical) information theory and complex networks
theoretically combining both distributed computing (or multiagent systems) and
game theory, we refer to it as ‘algorithmic networks’. The main idea is that a
population of formal theoretical machines can use communication channels over the
graph’s edges. Thus, the graph topology makes this population be networked. Once
the elements of the population start to exchange information, it forms a overarching
model for a system composed of interacting subsystems. So, note that algorithmic
networks will be networks of algorithms — which is the reason of its chosen name
—, should each theoretical machine represent a computable system. Indeed, in
the present article we will consider each node as a program of a universal Turing
machine (see definitions 3.5 and 3.13), which justifies calling either the nodes or
the elements of the population of an algorithmic network as nodes/programs
hereafter.
The term ‘algorithmic network’ is also employed to visually represent systems
or processes with circuits of algorithms [45, 47, 51] in order to tackle the problem of
computationally modeling these processes. Differently, we employ the expression
with the purpose of mathematically representing a model for networked populations
of computable systems from which one can investigate systemic properties and
prove theorems. Thus, a possible disambiguation may be using the expression
‘circuit-modeling algorithmic network’ for the former case and ‘populationsystemic algorithmic network’ for our present approach — on which, for the sake
of simplifying our notation, we will use only the expression ‘algorithmic network’
in this article. Although the usage in the present article shares the general goal
of mathematically representing models for systems through networked algorithms,
it may be seen as a generalization3 of the circuit-modeling approach. Note that,
as we will discuss in Subsection 2.4, there may be several different aspects of the
graph corresponding to different properties of the population such that a “circuit of
algorithms” would not grasp. For example, there may be different nodes being the
same Turing machine (that is, the population of programs may contain repeated
elements), there may be no need of a ‘delay operator’ for multi-step loops, and the
graph may not be static. In fact, as we will present in Section 3, these and other
features will occur in NBB .
2.2. Discussion on graphs and complex networks. One can have populations
with very different properties and several different graphs linking them. Thus, first
a general mathematical representation for graphs is paramount — see Definition
2.1 . Second, we need a way to make abstract aspects of these graphs correspond
to properties of the population of theoretical machines, which will be formalized in
Definition 2.2.
Aiming a wider range of different network configurations, as mentioned in Sections 1 and 2.1, we ground our formalism about graph representations on MultiAspects graphs (MAG) as presented in [76]. Therefore, one can mathematically
represent abstract aspects that could appear in complex networks. For example,
dynamical (or time-varying) networks, multicolored nodes and multilayer networks.
3 If one accepts a formalization of algorithms using theoretical machines (e.g. Turing machines).
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS5
Moreover, it facilitates network analysis by showing that their aspects can be isomorphically mapped into a classical directed graph. Thus, MAG abstraction has
proven to be crucial to establish connections between the characteristics of the network and the properties of the population composed of theoretical machines (see
Definition 2.2.3).
2.3. Definition.
Definition 2.1. As defined in [27, 76], let G = (A , E ) be a graph, where E is the
set of edges of the graph and A is a class of sets, each of which is an aspect.
Note 2.1.1. Note that E determines the (dynamic or not) topology of G.
Note 2.1.2. Each aspect in A determines which variant of a graph G will be (and
how the set E will be defined). As in [27], we will deal only with time-varying
graphs Gt hereafter, so that there will be only two aspects (|A | = 2): the set
of nodes (or vertices) V(Gt ) and the set of time instants T(Gt ). An element in
V (Gt ) × T (Gt ) is a composite vertex (or composite node). The graphs Gt will
be better explained in Section 3 .
2.4. Discussion on the general model. One can think of an algorithmic network in a broad sense as a theoretical distributed computing representation model
in which each node (or vertex) computes using network’s shared information, returning a final output4 or not5 . The computation of each node may be seen in a
combined point of view or taken as individuals. Respectively, nodes/programs may
be computing using network’s shared information to solve a common purpose —
as the classical approach in distributed computing — or, for example, nodes may
be competing6 with each other. For the present purposes, we are interested in the
average fitness (or payoff), and its related emergent complexity that may arise from
a process that increases the average fitness.
An algorithmic network may have several different configurations, so that the
following Definition 2.2 will not specify how a particular algorithmic network works.
Instead, our formalism enables one to represent every (or most) variation of algorithmic networks with the purpose of modeling a particular problem that may arise
from a networked complex system. For example, the networked population may be
synchronous 3.8 or asynchronous, have a set of information sharing protocols 3.9
or none, a randomly generated population 3.6 or a fixed one, with communication
costs or without them7 etc. In addition, the network topology that determines the
communication channels may be dynamical 3.1 , with weighted edges, multilayer
etc. In fact, in accordance with the goal of this article, we will prove theorems
on a particular model of algorithmic networks which will be defined in Section
3 . However, in this section we limit ourselves to present a general definition of
algorithmic networks. Also the reader is invited to note that a circuit-modeling
4 Which is our current case.
If the maximum number of cycles is finite, then every
node/program must return a final output.
5 An algorithmic network may have a non limited number of cycles, so that each node may
remain returning an endless number of partial outputs. This is also the case for circuit-modeling
algorithmic networks. However, we do not tackle this problem in the present article.
6 This game theoretical approach will be discussed further in Section 10.
7 The current model of this article does not consider any cost of communication or of using
the communication channels to send information.
6ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
algorithmic network [45, 47, 51] is one of these possible different configurations of
population-systemic algorithmic network (see Subsection 2.1).
2.5. Definitions.
Definition 2.2. We define an algorithmic network N = (G, P, b) upon a population of theoretical machines P, a graph G = (A , E ) and a function b that makes
aspects of G correspond8 to properties of P, so that a node in V (G) corresponds
one-to-one to an element of P. The graph G was previously defined in 2.1, and we
will define P and b in definitions 2.2.1 and 2.2.3 , respectively.
Definition 2.2.1. Let the population P be a subset of L, where L is the language
on which the chosen theoretical machine U are running.9 Each member of this
population may receive inputs and return outputs through communication channels.
Note 2.2.1.1. The choice of L and U determines the class of nodes/systems. For
example, one may allow only time-bounded Turing machines in the population. In
the present work, L will be a self-delimiting universal programming language for a
simple extended universal Turing machine U’ (see Definition 3.5 ) — i.e., an oracle
Turing machine — that returns zero whenever a non-halting computation occur.
Definition 2.2.1.1. Let C be a set of the maximum number of cycles that any
node/program o in the population P can perform in order to return a final output. A node cycle in an algorithmic network N is defined as a node/program10
returning a partial output (which, depending on the language and the theoretical
machine the nodes are running on, is equivalent to a node completing a halting computation)11 and sharing (or not) this partial output with its neighbors (accordingly
to a specific information sharing protocol or not — see Definition 3.11).
Note 2.2.1.1.1. So, if the algorithmic network is asynchronous, a cycle can be seen
as an individual communication round that doesn’t depend on whether its neighbors
are still running or not, while if the network is synchronous a cycle can be seen as
the usual communication round in synchronous distributed computing. Also note
that one may also refer to a network cycle, which denotes when all nodes of the
algorithmic network have returned their final outputs (if it is the case). Thus, a
network cycle must not be confused with a node cycle.
Definition 2.2.2. A communication channel between a pair of elements from P
is defined in E by an edge (whether directed or not) linking this pair of nodes/programs.
Note 2.2.2.1. A directed edge (or arrow) determines which node/program sends
output to another node/program that takes this information as input. An undirected edge (or line) may be interpreted as two opposing arrows.
Definition 2.2.3. Let
b : Y ⊆ A (G) → X ⊆ P r(P)
a
7→ b(a) = pr
8 See Definition 2.2.3 .
9 A population of machines P is a set of programs p ∈ L (i.e., more formaly, Turing machines
or other variants on language L).
10 Once there is a mapping of the set of nodes into the population of programs, the expression
“node/program” becomes well-defined within the theory of algorithmic networks.
11 In the present article for example — see Definition 3.5 .
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS7
be a function that maps a subspace of aspects Y in A into a subspace of properties
X in the set of properties P r(P) of the respective population in the graph G =
(A , E ) such that there is an bijective function fV P such that, for every (v, x) ∈
Y ⊆ A (G) where b(v, x) = (o, b|Y |−1 (x)) ∈ X, v is a vertex, o is a element of P
and
fV P : V (G) →
P
v 7→ fV P (v) = o
Definition 2.3. We say a element oi ∈ P is networked when there is N =
(G, P, b) such that oi is running on it.
Definition 2.3.1. We say oi is isolated otherwise. That is, it is only functioning
as a element of P.
Definition 2.4. We say that an input w ∈ L is a network input iff it is the only
external source of information every node/program receives and it is given to every
node/program before the algorithmic network begins any computation.
3. The Busy Beaver Imitation Game (BBIG)
3.1. Discussion on the Busy Beaver imitation model, emergence, complexity, optimization, and synchronicity. In this section, we will formalize a
toy model for comparing the emergence of algorithmic complexity from one of the
simplest forms of information sharing that percolates through the network: imitation of the fittest. As a toy model, our theoretical simple object of investigation
must be general, abstract and highly tractable, while enabling further variations
and extensions with the purpose of studying more properties of networked complex
systems for example (see Section 9). In the present section we choose to give a
full discussion on how one characterizes the algorithmic network NBB (N, f, t, τ, j)
(formerly denoted only as NBB ) before listing any formal definition. This way,
the reader might get a full picture of underlying ideas and motivations around this
model of algorithmic network that plays the Busy Beaver imitation game.
Take a randomly generated set of programs. They are linked, constituting a
network which is represented by a graph. Each node/program is trying to return
the “best solution” it can. And eventually one of these nodes/programs end up
being generated carrying beforehand a “best solution” for the problem in question.
This “best solution” is spread through the network by a diffusion process in which
each node imitates the fittest neighbor if, and only if, its shared information is
“better” than what the very node can produce. The question is: how much more
complexity can this diffusion process generates on the average compared to the best
nodes/programs could do if isolated?
A comparison between the complexity of what a node/program can do when
networked and the complexity of the best a node/program can do when isolated
will give the emergent algorithmic complexity of the algorithmic network. In
the present case, the networked “side of the equation” relies only on the simple
imitation of the fittest neighbor. Since this kind of imitating procedure is one of
the simplest or “worst” ways to use neighbors’ partial outputs to get closer to a
8ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
best solution12 , we are interested in obtaining the emergent algorithmic complexity
that arises from a “worst” networked case compared to the best isolated case.
Indeed, a possible interpretation of the diffusion described to the above is average optimization through diffusion in a random sampling. While optimization
through selection in a random sampling may refer to evolutionary computation or
genetic algorithms for example [34], in which the best solution eventually appears
and remains sufficiently stable over time, in our model optimization is obtained in
a manner that a best solution also eventually appears, but is diffused over time in
order to make every individual as averagely closer to the best solution as they can.
Therefore, the underlying goal of this process would be to optimize the average
fitness of the population using the least amount of diffusion time — we will come
back to this in Section 10. This type of optimization would particularly be better
suited for the cases when adding new nodes/programs (or subparts) is cheaper than
adding new cycles (or computational resources).
But how does one measure complexity of what a node/program computes? We
ground our complexity analysis on algorithmic information theory (AIT). As a
well-established mathematical field in theoretical computer science and information
science it has proven to be a powerful tool to achieve analytical results, proving
sound lemmas and theorems, in order to investigate and build theories on how
complexity changes over time. See [1, 3, 4, 21–23, 43, 44]. These works present
mathematical results on evolutionary open-endedness for computable complex
systems. That is, a process in which systems that could be fully simulated on a
theoretical Turing machine gain a unlimited amount of complexity over time as
random mutations and natural selection13 apply on them.
Therefore, as mentioned before, a different form of open-endedness plays the
central role in the fundamental characteristics and consequences of the results we
will present here (see Section 4 ): the average emergent open-endedness. Instead of asking about how complex systems become over time, as in evolutionary
open-endedness, we are focusing another akin question: how complex systems (in
fact, systems composed of interacting systems [67] [7]) become when the number of
its subparts increases? Or, more specifically in our case, how much more emergent
complexity arises on the average when the number of networked systems increases?
In other words, we are interested in how synergy [13, 38, 39, 62] (see also discussion
10) among interacting systems may have an impact on the emergence of complexity.
In order to tackle this problem we choose to first work under the framework
of classical algorithmic complexity, as in [1–4, 21–23, 43]. Furthermore, it will give
a direct way to measure fitness of each node/program as explained below. Despite being a theoretical and abstract model for complex systems, it gives the solid
foundation and fruitful framework to develop further quasi-isomorphic extensions
from resource-bounded — with a closer application to computer simulations — and
hipercomputable versions, as analogously did in [1,3,4]. However, in this article the
population of nodes/programs will be composed of arbitrary programs running on
12 Not necessarily given by the randomly generated node/program with the higher fitness,
since a node/program could use its neighbor’s partial output to calculate a even bigger integer for
example. Thus, this is the reason we say non formally that simply imitating the fittest neighbor
is one of the “worst”. Nevertheless, we leave the actual mathematical investigation of what would
be the less effective way to get averagely closer to a better solution for future research.
13 Which can be mathematically described by a fitness function and a selection procedure.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS9
an arbitrarily chosen universal Turing machine. It makes the definition of algorithmic complexity 3.4 straightforwardly applicable. But, analogously in [21], it forces
the need of a simple and restrict hipercomputable procedure to deal with eventual
non-halting programs (see definitions 3.5 and 3.7).
The second toy-modeling property that our algorithmic network NBB (N, f, t, τ, j)
will have is synchronicity, which will be defined as a population property in
3.8 . As in synchronous distributed computing, communication rounds are forced
to happen at the same time, so that only at the end of each cycle 2.2.1.1 each
node/program is allowed to exchange information by sending partial outputs and
receiving partial outputs from its neighbors as input for the next cycle. Indeed, an
asynchronous version of our results and how it relates to one presented here are
paramount for future research.
3.2. Discussion on dynamic graphs and measures of diffusion. As in Section 2, we will start defining topological properties of networks. Following a pursuit
of overarching mathematical theorems, we choose to deal with time-varying (or
dynamical) directed graphs [57] [41] [27]. The static case is covered by a particular
case of dynamical networks in which topology does not change over time — that is,
each temporal snapshot Gt is the equal [41]. And the undirected case can be seen
as a graph in which each undirected edge (or line) represents two opposing directed
edges (or arrows). Moreover, the dynamical case sets proper theoretical foundations
for future research, should one be interested in studying systemic emergent properties of algorithmic networks with a cycle-varying population for example (i.e.,
an algorithmic network that changes its population size as the number of cycles
increases).
Since we aim a formalization of a model for optimizing diffusion with the purpose of investigating emergent complexity, we need a diffusion measure that can
be applied on dynamical networks. Cover time is a diffusion measure that gives
the average time intervals in which a fraction τ of the nodes is “infected” [41] [27].
Besides being useful in order to measure diffusion on dynamical networks, it offers some advantages on domain conditions that other measures like the average
geodesic distance (or average shortest path length) does not. See notes 3.4.2 and
3.4.2.1. The algorithmic networks that we will define below get their graph topologies from a family of dynamical graphs that has a certain cover time function as a
common feature. This function has the domain on population sizes, time instants
and fractions τ of nodes (see Definition 3.4).
Next we will define the properties of the population composed of nodes/programs.
Remember in Section 2 that both aspects of a graph and properties of a population
on a theoretical machine (as well as a function bj that maps aspects into properties)
are necessary in order to properly formalize an algorithmic network. As mentioned
in the beginning of the present section, the diffusion process must rely on an imitation of the fittest. So one does not only need to define how fitness14 is measured but
also to mathematically state the exact procedure from which each diffusion step is
determined.
3.3. Discussion on the population playing the BBIG. As in [1–4, 21–23], we
use the Busy Beaver function as our fitness function. Naming bigger integers
relates directly to increasing algorithmic complexity [21], which will allow us to
14 Or payoff in a game-theoretical approach.
10
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
establish crucial probabilistic and statistical properties of a randomly generated
population in Lemma 8.1. Remember that there is a diffusion of the “best solutions”
during the cycles. Now the “best solution” assumes a formal interpretation of fittest
final output (or payoff). The choice of the word “solution” for naming bigger
integers now strictly means a solution for the Busy Beaver problem. Also note
that several uncomputable problems are equivalently reduced to the Busy Beaver,
including the halting problem. Thus, these mathematical features supports the
Busy Beaver function as a sound and meaningful choice for a fitness function for
a toy model [1, 22]. Not only on a resource-boundless case (e.g. Turing machines)
in which finding the best solution might be reducible to a first order uncomputable
problem (in the Turing hierarchy) — which is the case presented in this article
—, but also for more realistic resource-bounded versions of the Busy Beaver, as
shown in [1, 3, 4, 30, 81]. Such resource-bounded versions may be useful to model
optimization problems in which finding the best solution falls under a higher time
complexity class15 and, as we will mention in Section 10, we leave for future research.
Thus, with a fixed fitness function that works as a universal16 parameter for every node/program’s final (and partial) output it makes sense to have an interpretation of these running algorithmic networks NBB (N, f, t, τ, j) as playing a network
Busy Beaver game: during the cycles each node is trying to use the information
shared by its neighbors to return the biggest integer it can. The bigger the final
output integer the better the payoff (or fitness). However, only after we define
the language and the theoretical machine in 3.5 the notions of partial output and
algorithmic complexity — necessary for a fitness measure — will be well-defined.
Furthermore, a definition of the imitation part of the game is still required. Under what circumstances should a node/program imitate a neighbor? How the best
partial outputs are diffused through the network? We will give a formal definition of the information sharing protocols that each node/program must follow
when networked in 3.9. But before we will define the language and universal Turing
machine for a population of programs. Doing so, it makes the algorithmic complexity — as mentioned in open-endedness — well-defined as a direct consequence
and, moreover, its correspondent algorithmic probabilities ground our definition of
randomly generated population.
We use the term protocol as an abstraction of its usage in distributed computing
and telecommunications. A protocol is understood as a set of rules or algorithmic
procedures that nodes/program must follow at the end of each cycle when communicating. For example, it can be seen as the “rules for the communications” under
a game-theoretical perspective. So there must be a computable procedure determining what each node/program do when receiving inputs and sending outputs to
its neighbors. In fact, as our model of algorithmic network works under a simple
imitation of the fittest neighbor, the protocol which must be followed by every
node/program determines the exact procedure in doing this imitation. Hence, we
call this global information sharing protocol as imitation-of-the-fittest protocol (IFP). We will define these simple algorithmic procedures in 3.9 . The main
15 Note that time complexity refers to computational complexity in which computation is
limited by a certain amount of time.
16 As a universal parameter it may not work as a measure of how well adapted a system is in
respect to its respective environment or in respect to generate more offsprings. It is not in our
present scope to discuss the problem of measuring fitness and adaptation in adaptive complex
systems.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
11
idea is that each node/program oi compares its neighbors’ partial output (that is,
the integer they have calculated in the respective cycle) and runs the program of
the neighbor that have output the biggest integer if, and only if, this integer is
bigger than the one that the node/program oi has output. Since NBB (N, f, t, τ, j)
is playing the Busy Beaver game on a network while limited to simple imitation
performed by a randomly generated population of programs, we say it is playing a
Busy Beaver imitation game.
Finally, we end this section by using these previous definitions and the definition of the function bj , which “binds together” network and programs, in order
to define NBB (N, f, t, τ, j) in 3.13 . It is an algorithmic network populated by N
nodes/programs (that constitutes population PBB (N )) such that, after the first (or
c0 cycles) cycle, it starts a diffusion17 process of the biggest partial output (given
at the end of the first cycle) determined by a time-varying graph Gt that belongs
to a family of graphs G(f, t, τ ). Then, at the last time instant diffusion stops and
one cycle (or more) is spent in order to make each node return a final output.
3.4. Definitions.
Definition 3.1. As defined in [27], let Gt = (V, E , T) be a time-varying graph
(TVG), where V is the set of nodes, T is the set of time instants, and E ⊆ V × T ×
V × T is the set of edges.
Notation 3.1.1. Let V(Gt ) denote the set of nodes (or vertices) of Gt .
Notation 3.1.2. Let |V(Gt )| be the size of the set of nodes in Gt .
Notation 3.1.3. Let T(Gt ) denote the set of time instants in Gt = (V, E , T).
Notation 3.1.4. Let Gt (t) denote the graph Gt at time instant t ∈ T (Gt )
Definition 3.1.1. We define the set of time instants of the graph Gt as T (Gt ) =
{t0 , t1 , . . . , t|T (Gt )|−1 }.
Note 3.1.1.1. For the sake of simplifying our notations in the theorems below one
can take a natural ordering for T (Gt ) such that
∀i ∈ N ( 0 ≤ i ≤ |T (Gt )| − 1 =⇒ ti = i + 1 )
.
Definition 3.2. Let dt (Gt , ti , u, τ ) be the number of time instants (steps, time
intervals [57] or, in our case, cycles) for a diffusion starting on node u at time
instant ti to reach a fraction τ of nodes in the graph Gt .
Definition 3.3. We define the cover time for time-varying graphs as
X
1
dt (Gt , ti , u, τ )
if τ reached;
|V(G
t )|
CT (Gt , ti , τ ) =
u∈V(Gt )
∞
otherwise;
17 There are other diffusions too, since it is possible that two randomly generated neighbors are
not close to the node/program with the highest fitness and have different integers as first partial
outputs. However, only the one from the biggest partial output is independent of neighbor’s
partial outputs, so that it disseminates in any situation.
12
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Notation 3.3.1. Let D(Gt , t) denote the temporal diffusion diameter of the
graph Gt taking time instant t as the starting time instant of the diffusion process.
That is,
(
max{x | x = dt (Gt , t, u, 1) ∧ u ∈ V (Gt )};
D(Gt , t) =
∞
if ∀x ∈ N∀u ∈ V (Gt ) x 6= dt (Gt , t, u, 1) ;
Definition 3.4. Let
G(f, t, τ ) = {Gt | i = |V (Gt )| ∧ f (i, t, τ ) = CT (Gt , t, τ ) ∧
∧ ∀i ∈ N∗ ∃!Gt ∈ G(f, t, τ )( |V (Gt )| = i )}
where
f : N∗ × X ⊆ T (Gt ) × Y ⊆]0, 1] → N
(x, t, τ )
7→ y
be a family of unique time-varying graphs which shares f (i, t, τ ) = CT (Gt , t, τ ),
where i is the number of nodes, as a common property. The number of nodes of
each graph in this family may vary from 1 to ∞.
Note 3.4.1. The results in this article can be equally achieved with a weaker assumption within this family: one can instead define a family of graphs upon an
arbitrary function f such that f (i, t, τ ) = CT (Gt , t, τ ) only when N → ∞.
Note 3.4.2. Note that a well-defined family G(f, t, τ ) for every t ∈ T (Gt ) and τ ∈
]0, 1] implies that CT (Gt , t, τ ) 6= ∞. Hence, in this case, the graphs Gt ∈ G(f, t, τ )
must be strongly temporal-connected for every time instant (and, thus, also cyclic)
— See [76] —, so that every composite node can reach any other composite node
given a sufficient amount of time intervals.
Note 3.4.2.1. For our main results to hold one may weaken conditions on τ and
t, because our theorems assume arbitrary values as long as they belong to their
respective intervals or set and the cover time has a function f well-defined for that
respective domain. In fact, this is one of the advantages of using the cover time as
a diffusion metric for dynamic graphs.
Notation 3.1. Let lg(x) denote the binary logarithm log2 (x).
Notation 3.2. Let U(x) denote the output of a universal Turing machine U when
x is given as input in its tape. So, U(x) denote a partial recursive function ϕU (x)
that is a universal partial function [65]. If x is a non-halting program on U, then
this function U(x) is undefined for x.
Notation 3.3. Let LU be a binary self-delimiting universal programming language
for a universal Turing machine U such that there is a concatenation of strings
w1 , ..., wk in the language LU , which preserves18 the self-delimiting (prefix-free)
property of the resulting string, denoted by
w1 ◦ ... ◦ wk ∈ LU
Notation 3.4. Let LU be a binary self-delimiting universal programming language
for a universal Turing machine U. The (prefix) algorithmic complexity (Kolmogorov complexity, program-size complexity or Solomonoff-Komogorov-Chaitin
18 For example, by adding a prefix to the entire concatenated string w w ...w that encodes
1 2
k
the number of concatenations. Note that each string was already self-delimiting.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
13
complexity) of a string w ∈ LU , denoted by A(w), is the size of the smallest program p∗ ∈ LU such that U(p∗ ) = w.
Note 3.4.1. The reader may also find in the literature the prefix algorithmic complexity denoted by H(w) or — more frequently used — K(w). As introduced in
Section 1 and presented in Section 10, there might be several intersections with
other fields in future work. Thus, we choose a self-explaining approach on notation
in order to avoid ambiguity and notation conflicts in future work, such that we
would denote the (prefix) algorithmic complexity/information19 by IA (w). However, for the sake of simplifying our notation, we choose to denote it only by A(w)
in the present article.
Definition 3.5. Given a binary self-delimiting universal programming language
LU for a universal Turing machine U, where there is a constant ǫ ∈ R, with
0 < ǫ ≤ 1, and a constant 0 ≤ CL ∈ N such that, for every N ∈ N,
A(N ) ≤ lg(N ) + (1 + ǫ) lg(lg(N )) + CL
we then define an oracle20 Turing machine U’ such that for every w ∈ LU :
(
U(w) + 1
if U halts on w
U’(w) =
0
if U does not halt on w
Note 3.5.1. The oracle Turing machine is basically (except for a trivial bijection)
the same as the chosen universal Turing machine. The oracle is only triggered to
know whether the program halts or not in first place. Also note that U′ (w) is a
total function, and not a partial function as U(w) — see Definition 3.2.
Note 3.5.2. Note that from algorithmic information theory (AIT) we know that
the algorithmic complexity (see next Definition 3.4 ) A(U’(w)) only differs from
A(U(w)) by a constant if U halts on w. This constant is always limited by the size
of the smallest program that always adds 1 to any other halting computation. This
is the reason why the algorithmic complexity of the final outputs of node/programs
in NBB (N, f, t, τ, j) only differ by a constant, should nodes be halting programs of
U instead of U’. And it also makes Lemma 8.2 sound.
Definition 3.6. We say a a population P ⊆ LU is randomly generated iff P is
a sample generated by |P| trials accordingly to an algorithmic probability where,
for a constant C,
p∈P
if f
P[p] = C
1
2|p|
Note 3.6.1. The constant C is important for us because it allows us to characterise
population PBB (N ) in Definition 3.11 as randomly generated, taking into account
that there are global information sharing protocols.
19 That is, the algorithmic information contained in a object about itself [48].
20 Or any hypercomputer with a respective Turing degree higher than or equal to 1.
14
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Definition 3.7. We say a population P is sensitive to oracles iff whenever an
oracle is triggered during any cycle in order to return a partial output the final
output of the respective node/program is also 0.21 Or more formally:
Let pnetU be a program such that U(pnetU ◦ oi ◦ c) computes on machine U
cycle-by-cycle what a node/program oi ∈ P does on machine U until cycle c when
networked. Let poi ,c be the partial output sent by node/program oi at the end
of cycle c to every neighbor (at the end of the arrow or directed edge) of oi . Also,
poi ,max{c|c∈C} denotes the final output of the node/program oi . Then, for every
oi ∈ P,
∄pnetU =⇒ poi ,max{c|c∈C} = 0
Definition 3.8. In a synchronous population of an algorithmic network each
node/program is only allowed to receive inputs from its incoming neighbors and
to send information to its outgoing neighbors at the end of each cycle (or communication round22), except for the last cycle. Each cycle always begins and ends at
the same time even if the computation time of the nodes/programs is arbitrarily
different. Or more formally:
Let Gt be a time-varying graph such that each time interval corresponds to nodes
sending information to their neighbors at the same time. Then, there is an injective
function
f : T (Gt ) →
C
t 7→ f (t) = t
Note 3.8.1. If the population is isolated, then only each partial output counts in
respect to the respective element of the population.
Definition 3.8.1. In the last cycle, every node only returns its final output.
Definition 3.9. We say a population P follows an imitation-of-the-fittest protocol (IFP) iff each node/program always obeys protocols defined in 3.9.1 , 3.9.2
and 3.9.3 when networked. Or more formally:
Let Xneighbors (oj , c) be the set of incoming neighbors of node/program oj that
have sent partial outputs to it at the end of the cycle c. Let {poi ,c | oi ∈
Xneighbors (oj , c) ∧ i ∈ Y ⊂ N ∧ c ∈ C}, where Y is finite, be the set of partial
outputs relative to Xneighbors (oj , c). Let w be the network input as defined in 2.4.
Let ◦ denote a recursively determined concatenation of finite strings. Then, for
every oj , oi ∈ P and c, c − 1 ∈ C,
(1) if max{c | c ∈ C} = 1, then
poj ,c = U′ (oj ◦ w)
(2) if c = 1 and c 6= max{c | c ∈ C}, then
poj ,c = w ◦ oj ◦ U′ (oj ◦ w)
(3) if c 6= 1 and c 6= max{c | c ∈ C}, then
poj ,c = w ◦ oi ◦ max{x | poj ,c−1 = w ◦ oi ◦ x ∨ w ◦ oi ◦ x ∈ {poi ,c−1 | oi ∈
Xneighbors (oj , c − 1) ∧ i ∈ Y ⊂ N ∧ c − 1 ∈ C}}
(4) if c = max{c | c ∈ C} and poj ,c−1 = w ◦ oi ◦ x, then
poj ,c = x
21 Since we are assuming 0 as the assigned non-halting output for U′ in relation to the machine
U.
22 As in distributed computing.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
15
Note 3.9.1. Since we will be working with synchronous algorithmic networks, these
global sharing protocols applies at the end of each cycle (or communication round)
— see Definition 2.2.1.1 . So, after the first cycle the diffusion of the biggest
partial output works like a spreading in time-varying networks [41] [27]. And the
last cycle (or more cycles — see Definition 3.13 ) is spent in order to make each
node/program return a number — from which we measure the complexity of the
respective node/program as discussed in Subsection 3.1 .
Note 3.9.2. In order to simplify our notation we let w ◦ U′ (x) denote the prefix
preserving concatenation ◦ (see Notation 3.3) of the string w ∈ LU with the string
y ∈ LU such that y represents the number U′ (x) in the language LU .
Definition 3.9.1. We call a Busy Beaver contagion protocol as a global information sharing protocol in which every node/program runs the node/program of
the neighbor that have output — a partial output — the biggest integer instead of
its own program iff the partial output of this neighbor is bigger than the receiver’s
own partial output.
Note 3.9.1.1. Note that a node/program only needs to take into account the biggest
partial output that any of its neighbors have sent. If more than one sends the biggest
integer as partial output, the receiver node/program choose one of these respective
neighbors accordingly to an arbitrary rule. Then, this partial output is the one
that will be compared to the partial output from the receiver node/program.
Definition 3.9.2. In a maximally cooperative protocol every node/program
shares its own program and latest partial output with all its neighbors at the end
of each cycle and before the next cycle begins.
Note 3.9.2.1. In the model defined in 3.11 sharing only the last partial output turns
out to be equivalent23 to the definition of maximal cooperation.
Note 3.9.2.2. Remember that NBB (N, f, t, τ, j) only lets its nodes/programs perform computation in the first cycle (see Definition 3.9.3). This is the reason why
only the network input w matters in its respective maximally cooperative protocol.
Note 3.9.2.3. As in [27] this diffusion process may be interpreted as following a
Breadth-First Search (BFS), in which each node starts a diffusion by sending the
specified information in Definition 3.9 to all of its adjacent nodes. Then, these
adjacent nodes relay information for their own adjacent nodes in the next time
instant, and so on.
Note 3.9.2.1. Typically, one can set a collection of time instants in which the diffusion of information is limited to 1 step. However, the presented final results 8.1
, 8.1.1 and 8.2 also hold if more than one step per time instant of diffusion (in our
case, the application of the global sharing protocols) is allowed at the end of each
′
cycle. In fact, it can only make τE(max) (N, f, ti , τ )|tt bigger — see Lemma 8.6 .
Definition 3.9.3. We call a contagion-only protocol as a global information
sharing protocol in which every node/program only plays the Busy Beaver contagion
and does not perform any other computation after the first cycle when networked
in some N = (G, P, b).
23 However, at the expense of using more computation time.
16
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Note 3.9.3.1. This is the condition that allows us to investigate the “worst” (see
discussion 3.1) case in which no node/program spends computational resources
other than playing its global sharing protocols.24 In other words, it forces the
algorithmic network to rely on a diffusion process only.
Note 3.9.3.2. In the main model presented in this article the first time instant
occurs after the first cycle — see 3.13 .
Definition 3.10. Let LBB ⊂ LU be a language of programs of the form Pprot ◦ p
where p ∈ LU and the prefix Pprot is any program that always makes Pprot ◦ p
obey when running on U’ the imitation-of-the-fittest protocol when it is
networked. Otherwise, if it is isolated, then U’(Pprot ◦ p) = U’(p) and every
subsequent cycle works like a reiteration of partial outputs as immediate next input
for the same node/program.25
Definition 3.11. Let PBB (N ) ⊆ LBB be a randomly generated population
of N elements that is synchronous26, sensitive to oracles and with randomly
generated27 suffixes p ∈ LPBB (N ) such that28
p ∈ LPBB (N )
if f
Pprot ◦ p ∈ PBB (N ) ⊆ LBB
Note 3.11.1. Note that all conditions and protocols in 3.11 define the set of properties P r(PBB (N )) of the population PBB (N ) as in 2.2.3 .
Note 3.11.2. There is a misplaced usage of the operator ⊆ in PBB (N ) ⊆ LBB here.
Since PBB (N ) is a population, it may contain repeated elements of LBB . However,
for the sake of simplicity, we say a population P is contained in a language L iff
∀pi , 1 ≤ i ≤ |P| pi ∈ P =⇒ pi ∈ L
Definition 3.12. For the sake of simplifying our notation, we denote the set of
the size-ordered29 N smallest
p ∈ lim LPBB (N )
N →∞
as
LU (N )
24 In our case, generating expected emergent algorithmic complexity of a node.
25 Thus, the isolated case may be represented (and is equivalent to) by same algorithmic net-
work built on a population in language LU that does not follow any information sharing protocol
and the topology of the multiaspect graph is composed by one-step loops on each node/program
only.
26 The procedure responsible for performing the synchronization may be abstract-hypothetical
or defined on an underlying oracle machine that makes each individual cycle start at the same
time (or after every node/program returns its partial output in the respective cycle).
27 Note that we are dealing with self-delimiting languages, so that one can always define
algorithmic probabilities.
28 Hence, L
PBB (N) is defined as the set of suffix programs that were randomly generated in
order to constitute PBB (N ).
29 That is, an ordering from the smallest to the biggest size. Also note that ordering members
with the same size may follow a chosen arbitrary rule.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
17
Definition 3.13. Let
NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj )
be an algorithmic network where f is an arbitrary well-defined function such that
f : N∗ × X ⊆ T (Gt ) × Y ⊆]0, 1] → N
(x, t, τ )
7→ y
and Gt ∈ G(f, t, τ ), |V (Gt )| = N , |T (Gt )| > 0 and there are arbitrarily chosen30
c0 , n ∈ N where c0 + |T (Gt )| + 1 ≤ n ∈ N such that bj is a injective function
bj : V (Gt ) × T (Gt ) →
PBB (N ) × N|n1
(v, tc−1 )
7→ bj (v, tc−1 ) = (o, c0 + c)
such that31 , since one has fixed the values of c0 and n,
bj bj : V (Gt ) × T (Gt ) →
PBB (N ) × N|n1
≤ NN
(v, tc−1 )
7→ bj (v, tc−1 ) = (o, c0 + c)
Note 3.13.1. Thus, NBB (N, f, t, τ, j) is an algorithmic network populated by N
nodes/programs from PBB (N ) such that, after the first (or c0 cycles) cycle, it
starts a diffusion32 process of the biggest partial output (given at the end of the
first cycle) determined by a time-varying graph Gt that belongs to a family of graphs
G(f, t, τ ). Then, at the last time instant diffusion stops and one cycle (or more)
is spent33 in order to make each node return a final output. Each node returns as
final output its previous partial output determined at the last time instant — see
Definition 3.9.3 .
Note 3.13.2. Note that in this model the aspects of the graphs in the family G
that are mapped into the properties of the population PBB (N ) by functions bj are
nodes and time instants.
Note 3.13.3. The reader is invited to note that the main results presented in this
paper also hold for only one function bj per graph in the family G (see 8.1.3 ).
Definition 3.13.1. We denote N|n1 as CBB .
4. Average Emergent Open-Endedness (AEOE)
4.1. Discussion on emergence and complexity. In this section we will introduce a formal definition of average emergent open-endedness (AEOE). Some of the
current discussion is already presented in 3.1 34 . Thus, we will abstain from addressing the problem of how AEOE relates to the BBIG. Instead, we will discuss
general aspects and the idea that the former mathematically grasps.
As discussed in Section 2 , on a general approach one can understand the theory
of algorithmic networks as a mathematical framework for studying interacting complex systems. However, we are limiting ourselves to deal with computable complex
30 Since they are arbitrarily chosen, one can take them as minimum as possible in order to
minimize the number of cycles for example. That is, c0 = 0 and n = |T (Gt )| + 1 for example.
31 See Definition 2.2.3 .
32 There other diffusions too. However, only the one from the biggest partial output is independent of neighbor’s partial outputs.
33 This condition is necessary to make this algorithmic network defined even when |T (G )| = 1.
t
34 And we recommend its reading before the current subsection.
18
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
systems that sends and receives information from neighbors (during cycles) in order
to return final outputs. This way, as mentioned in discussion 3.1 , we can apply
the tools from theoretical computer science and information theory — specially
algorithmic information theory — on the study of how a systemic feature like complexity “behaves” as this system of interacting subsystems runs (see discussion 3.1
). So, our central inquiry falls under investigating how much complexity emerges
when subsystems are interacting compared to when they are running completely
isolated from each other.
We follow a consensual notion of emergence [26, 31, 35, 43, 55, 60] as a systemic
feature or property that appears only when the system is analyzed (theoretically
or through computer simulations) as a “whole”. Thus, the algorithmic complexity
(i.e., an irreducible number of bits of information) of a node/program’s final output when networked35 minus the algorithmic complexity of a node/program’s final
output when isolated formally defines an irreducible quantity36 of information that
emerges in respect to a node/program37 that belongs to an algorithmic network.
We call it as emergent algorithmic complexity (EAC) of a node/program.
Note that if a system is analyzed as a separated — the subparts do not need to be
necessarily apart from each other, but each part in this case would be taken as an
object of investigation where no information enters or exits anyway — collection of
“subparts”, the EAC of a node/program will be always 0.
An important distinction is crucial: the EAC of a node/program must be not confused with EAC of the entire algorithmic network. Measuring the emergent algorithmic complexity of the algorithmic network taking into account every node/program
“at the same time” is — as our intuition demands to be — mathematically different
from looking at each individual final output’s algorithmic complexity. For example, one may consider the algorithmic information of each node/program combined
(in a non-trivial way) with the algorithmic information of the network’s topology
[17, 53, 80]. Within the framework of algorithmic networks, this “whole” emergent algorithmic complexity can be formally captured by the joint algorithmic
complexity of each node/program’s final output when networked minus the joint
algorithmic complexity of each node/program’s final output when isolated. However, analyzing this systemic property is not of the scope of the present article and
it will be a necessary future research. Specially in investigating different emergent
complexity phase transitions as discussed in Subsection 4.2.
Therefore, instead of investigating the joint EAC of an algorihtmic network, one
may look for a mean value of EAC for all nodes/programs. The average emergent
algorithmic complexity (AEAC) is defined by the mean on all nodes/programs’
(and possible network’s topologies) EAC (see Definition 4.3 ). It gives the average
complexity of the nodes/programs’ respective fitnesses (or, in a game-theoretical
interpretation, payoff) in a networked population, once there is a fitness function
that evaluates final outputs. For further discussions, see Subsection 3.1 .
As the main object of mathematical investigation of this article is an algorithmic network NBB (N, f, t, τ, j) in which the population is randomly generated from
a stochastic process of independent and identically distributed random variables
35 That is, interacting with other parts of the system.
36 Note that this quantity of bits may be 0 or negative.
37
More precisely, in respect to a node’s final output, i.e. the complexity of its fitness/payoff
as discussed in Subsection 3.3.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
19
under an algorithmic probability distribution, we refer to the average EAC in
NBB (N, f, t, τ, j) as expected emergent algorithmic complexity (EEAC), denoted
here 4.4 by
net
ENBB (N,f,t,τ ) ∆ A(oi , c)
iso
4.2. Discussion on open-endedness. Definition 4.3 will tell us how much emergent algorithmic complexity emerges on the average in an algorithmic network N.
The question is: which algorithmic networks produces more AEAC? And an immediate question would be: how does AEAC increase as the population grows in size?
More AEAC means that a node/program needs more irreducible information on
the average than it already contains, should it try to compute isolated what it does
networked. A system with a bigger AEAC “informs” or “adds” more information
to its parts on the average. Thus, the investigation of these questions are directly
related to important topics in complex systems, like integration [56], synergy (see
Subsection 10 ), and emergence [26, 31, 35, 43, 55, 60].
Another important concept mentioned in discussion 3.1 is open-endedness. It
is commonly defined in evolutionary computation and evolutionary biology as the
inherent potential of a evolutionary process to trigger an endless increase in complexity [1,5,43]. That means that in the long run eventually will appear an organism
that is as complex as one may want.38 Given a certain complexity value as target, one would just need to wait a while in order to appear an organism with a
bigger complexity than or equal to the target value, no matter how big this value
is. In fact, within the framework of algorithmic information theory as shown in
[1, 21–23], a cumulative39 evolution reaches N bits of complexity after — realistic fast — O(N 2 (log(N ))2 ) successive mutations on one organism at the time —
whether your organisms are computable, sub-computable or hyper-computable (see
[1, 3, 4]).
What we have found is that within the theory of algorithmic networks openendedness is a mathematical phenomenon closely related to the core questions in
the first paragraph of this subsection. However, it emerges as an akin — but
different — phenomenon to evolutionary open-endedness: instead of achieving an
unbounded quantity of algorithmic complexity over time (or successive mutations),
an unbounded quantity of emergent algorithmic complexity is achieved as the population size increases. Since it is a property that emerges depending on the amount of
parts of a system, it arises as an emergent property. So, we refer to it as emergent
open-endedness.
The main results of this paper (see 8.1, 8.1.1, 8.2 and 9.1) prove that there
are conditions of the network and protocols of the population that trigger a type
of emergent open-endedness as the randomly generated populations grows. They
make the AEAC increases as one may want, should the population size increases
sufficiently. Since it is an increase in average complexity (see discussion 4.1), we
refer to it as average emergent open-endedness (AEOE) (see Definition 4.5).
Moreover, since the population is randomly genarated so that AEAC turns to EEAC
in the case of algorithmic networks NBB (N, f, t, τ, j), one may refer to AEOE as
38 As a consequence, an infinite number of different organisms tends to appear in the evolutionary line after an infinite time, or after an infinite amount of successive mutations.
39 Which allows organisms to recall its predecessors
20
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
expected emergent open-endedness (EEOE) (see Definition 4.5.1). For example, even if the AEAC is 0 or negative for some algorithmic networks with finite
size populations, for bigger enough population sizes the probability that these algorithmic networks have a bigger AEAC tends to 1 — see note 8.1.1. In other
words, we will show that there are conditions for algorithmic networks such that
a phase transition in AEAC eventually occurs, should the AEAC be non-positive
for small populations. Thus, one may call this transition as average emergent
complexity phase transition.
4.3. Definitions.
Definition 4.1. The emergent algorithmic complexity (EAC) of a node/program
oi in c cycles is given in an algorithmic network that always produces partial and
final outputs by
net(b)
∆ A(oi , c) = A(U(pbnet (oi , c)) − A(U(piso (oi , c))
iso
where:
(1) oi ∈ L;
(2) pbnet is the program that computes cycle-per-cycle the partial outputs of oi
when networked assuming the position v, where b(v, x̄) = (oi , b(x̄)), in the
graph G in the specified number of cycles c with network input w;
(3) piso is the program that computes cycle-per-cycle the partial outputs of oi
when isolated in the specified number of cycles c with network input w;
Note 4.1.1. Note that:
(1) A(U(pbnet (oi , c))) is the algorithmic complexity of what the node/program
oi does when networked;
(2) A(U(piso (oi , c)) is the algorithmic complexity of what the node/program
oi does when isolated;
Note 4.1.2. While program piso may be very simple, since it is basically a program
that reiterates partial outputs of oi as inputs to itself at the beginning of the
next cycle (up to c times), program pbnet may also comprise giving the sent partial
outputs from oi ’s incoming neighbors at the end of each cycle as inputs to oi at the
beginning of the respective next cycle, so that it may be only described by a much
more complex procedure.
Note 4.1.3. Note that the algorithmic complexity of piso or pbnet may be not directly
linked to A(U(piso (oi , c)) or A(U(pbnet (oi , c))) respectively, since the A(U(piso (oi , c))
and A(U(pbnet (oi , c))) are related to the final outputs (if any) of each node’s computation.
Note 4.1.4. Remember definitions 3.7 and 3.5 which states that even when a
node/program does not halt in some cycle, machine U′ was defined in order to
assure that there is always a partial output for every node/program for every cycle.
If population P is defined in a way that eventually a partial or final output is not
obtained when running on the respective theoretical machine (see Definition 2.2.1 ),
Definition 4.1 would be inconsistent. This is the reason we stated in its formulation
that there always is partial and final outputs. However, it is not necessary in the
case of NBB (N, f, t, τ, j) (see Definition 3.11 and 4.1.1 ).
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
21
Note 4.1.5. If one defines the emergent creativity of a node/program as
A(U(pbnet (oi , c))|U(piso (oi , c))
our results also hold for replacing the expected emergent algorithmic complexity
(EEAC) with expected emergent algorithmic creativity (EEACr). Since we are
estimating lower bounds, note that from AIT we have that
A(U(pbnet (oi , c))|U(piso (oi , c)) ≥ A(U(pbnet (oi , c)) − A(U(piso (oi , c)) + O 1
Definition 4.1.1. More specifically, one can denote the emergent algorithmic
complexity of a node/program oi in an algorithmic network NBB (N, f, t, τ, j)
(EACBB ) during c cycles as
net(bj )
b
j
(oi , c))) − A(U(piso (pi , c)))
∆ A(oi , c) = A(U(pnet
iso
where:
(1) oi = Pprot ◦ pi ∈ PBB (N ) ⊆ LBB ;
bj
(2) pnet
is the program that computes cycle-per-cycle what a program oi does
when networked assuming the position v, where bj (v) = (oi ), in the
graph Gt in c cycles with network input w;
(3) piso is the program that computes cycle-per-cycle what a program pi does
when isolated40 in c cycles with network input w;
Note 4.1.1.1. In order to check if Definition 4.1.1 is well-defined, remember the
definitions 3.5 and 3.10 .
Definition 4.2 (Notation). For the sake of simplifying our notation, let {bj } denote
bj bj : V (Gt ) × T (Gt ) →
PBB (N ) × N|n1
(v, tc−1 )
7→ bj (v, tc−1 ) = (o, x + c)
P
denote
and bj in the sum
bj
bj : V (Gt ) × T (Gt ) →
PBB (N ) × N|n1
(v, tc−1 )
7→ bj (v, tc−1 ) = (o, x + c)
Definition 4.3. We denote the average emergent algorithmic complexity of
a node/program (AEAC) for an algorithmic network N = (G, P, b), as
net
EN ∆ A(oi , c) =
iso
P
=
X oi ∈P
b
net(b)
∆ A(oi ,c)
iso
N
|{b}|
Note 4.3.1. As in note 3.13.3 , if only one function b exists per population, then
there is only one possible network’s topology linking each node/program in the
40 Which is a redundancy since we are refering to p instead of o here.
i
i
22
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
population. So, in this case,
net(b)
EN
net
∆ A(oi , c) =
P
∆ A(oi , c)
iso
oi ∈P
iso
N
Definition 4.4. We denote the expected emergent algorithmic complexity
of a node/program for algorithmic networks NBB (N, f, t, τ, j) (EEACBB ) with
network input w, where 0 < j ≤ |{bj }| 41, as
ENBB (N,f,t,τ )
∆ A(oi , c) =
net
iso
P
=
∆ A(oi ,c)
X oi ∈PBB (N )
iso
N
=
|{bj }|
bj
P
=
net(bj )
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c)))−A(U(piso (pi ,c)))
N
bj
|{bj }|
Note 4.4.1. As in note 3.13.3 and 4.3.1 , if only one function bj exists per population, then only one possible network’s topology linking each node/program in the
population. So, in this case,
ENBB (N,f,t,τ )
net
∆ A(oi , c) =
P
oi ∈PBB (N )
iso
net(bj )
∆ A(oi , c)
iso
N
Definition 4.5. We say an algorithmic network N with a population of N nodes
has the property of average emergent open-endedness (AEOE) for a given
network input w in c cycles iff
net
lim EN ∆ A(oi , c) = ∞
N →∞
iso
Definition 4.5.1. We say an algorithmic network NBB (N, f, t, τ ) has the property
of expected emergent open-endedness (EEOE) for a given network input w
in c cycles iff
net
lim ENBB (N,f,t,τ ) ∆ A(oi , c) = ∞
N →∞
iso
41
For
the
sake
of
simplifying our
notation,
b b : V (Gt ) × T (Gt ) →
PBB (N ) × N|n
and
bj
in
1
j j
(v, tc−1 )
7→ bj (v, tc−1 ) = (o, x + c)
bj : V (Gt ) × T (Gt ) →
PBB (N ) × N|n
.
1
(v, tc−1 )
7→ bj (v, tc−1 ) = (o, x + c)
let
{bj }
denote
the
sum
denote
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
23
5. Cycle-bounded conditional halting probability
5.1. Discussion on halting probabilities. As presented in [11, 18–20, 48], the
number Ω is the probability of randomly picking a halting program given an algorithmic probability distribution for a self-delimiting language. So, a conditional
probability Ω(w) is the probability of randomly picking a halting program with w
as its input. A variation of halting probabilities on self-delimiting universal programming languages is the resource-bounded halting probability: the probability
of randomly picking a halting program on universal machine with limited computation time or memory [1, 6, 48]. However, for algorithmic networks composed
of populations of arbitrary self-delimiting programs, another variation is needed.
The “scarce” resource is not time or memory, but cycles. Since nodes/programs
are computing during cycles (see 2.2.1.1 ) which determine the maximum number
of times a node/program may reuse (through successive halting computations) its
partial output before returning a final output, the cycle-bounded conditional
halting probability Ω(w, c) is the probability of randomly picking a program that
halts for every cycle until c with w as its initial input.
In Subsection 3.1 we have discussed that we want to compare the average algorithmic complexity of an algorithmic network playing the BBIG with its population
when nodes/programs are completely isolated from each other. Thus, in order to
prove the central Theorem 8.1 one will need to calculate an upper bound on the expected algorithmic complexity of what a node/program does when isolated during
the cycles. The cycle-bounded conditional halting probability Ω(w, c) is used with
the purpose of normalizing the probabilities of halting nodes/programs in Lemmas
8.3 and 8.5 . Remember that there may be programs that halt on every cycle and
programs that do not. Since we need to calculate probability in each case and
the population is randomly generated, Ω(w, c) necessarily arises as an important
constant in Theorem 8.1.
5.2. Definitions.
Definition 5.1. Let L be an arbitrary population. Let NBB (N, f, t, τ, j) be an
algorithmic network. We denote the set of pi ∈ L, where 1 ≤ i ≤ N and Pprot ◦ pi ∈
PBB (N ), such that pi always halts on network input w in every cycle until c when
Pprot ◦ pi is isolated as
Haltiso (L, w, c)
Definition 5.1.1. Analogously, we denote the set of pi ∈ L, where 1 ≤ i ≤ N and
Pprot ◦ pi ∈ PBB (N ) with NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ), such that pi does
not halt on network input w in at least one cycle until c when Pprot ◦ pi is isolated
as
Haltiso (L, w, c)
Definition 5.2. We denote the cycle-bounded conditional halting probability42 of a program in a language LU that always halts for an initial input w in c
42 That is, a conditional Chaitin’s Omega number for isolated programs in a population of an
algorithmic network
24
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
cycles as
Ω(w, c) =
1
X
pi ∈Haltiso (LU ,w,c)
= lim
N →∞
2|pi |
X
= lim
N →∞
pi ∈Haltiso (LPBB (N ) ,w,c)
X
pi ∈Haltiso (LU (N ),w,c)
1
2|pi |
=
1
2|pi |
Note 5.2.1. Since LU is self-delimiting, the algorithmic probability of each program
is well-defined. Hence, one can define the halting probability Ω 43 for LU . Further,
the same holds for conditional halting probability Ω(w), i.e. the probability that
a program halts when w is given as input. Then, the set of programs that always
halt on initial input w in c cycles is a proper subset of the set of programs that
halt, so that
Ω(w, c) ≤ Ω(w) = Ω(w, 1) < 1
In fact, one can prove that
Ω(w, c′ ) ≤ Ω(w, c)
when c′ ≥ c > 0. On the other hand, one can also build programs that always halts
for every input and for every number of cycles. So, for every c > 0
0 < Ω(w, c) < 1
6. Average diffusion density
6.1. Discussion on diffusion in the BBIG. On the “networked side of the
equation” in Theorem 8.1, as discussed in Subsection 3.1, we are looking for a lower
bound on the expected emergent algorithmic complexity of what a node/program
does when networked playing the Busy Beaver imitation game (BBIG). This is the
value from which will be subtracted the expected algorithmic complexity of what
a node/program does when isolated.
Remember that in the BBIG once nodes/programs start sharing their partial
outputs the only relevant feature that matters from then on is who has the biggest
integer. In other words, the neighbor with a bigger partial output will “infect” a
node/program with a smaller partial output. Note that it may be analogous to
a multi-“disease” spreading, since there are several different graph arrangements
of randomly generated nodes/programs in which there might be separated clusters
with different respective “diseases” (in our case, different integers) spreading at the
same time for a while. However, the BBIG rules make one type of partial output
spreads over any other partial output: the biggest partial output of the randomly
generated population. As we will stress in proof of Theorem 8.1, this is one of the
mathematical shortcuts to reach a desired proof of a lower bound on the expected
emergent algorithmic complexity of what a node/program does when networked
playing the BBIG. It is given using the average density of nodes/programs infected
by a single source in Definition 6.2.1.1 — in the case, the source is the randomly
generated node/program that returns the biggest integer. Thus, we call it as average singleton diffusion density. Note that as there is a probability of two or
more nodes/programs being generated which return the biggest integer, the average singleton diffusion density is actually a lower bound on the average diffusion
43 Note that the Greek letter Ω here does not stand for an asymptotic notation opposed to
the big O notation.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
25
density (which might be multisource). Although Lemma 8.6 is proved using the
average singleton diffusion density, it also holds for the average diffusion density,
since it works like a lower bound. The average diffusion density is related to prevalence in [58, 59], although there is no “curing” process in our case. And it might be
also related to flooding time in [25] — see also Section 10.
Indeed, this analogy with diffusion or spreading in complex networks is not only
a coincidence but is also responsible for bringing the aspects and proprieties studied
in complex networks to our current results in algorithmic networks. The EEAC in
the BBIG is lower bounded by a value that depends on how powerful the diffusion
is on a network. We address more specifically some of theses relations with complex
networks in Section 10.
6.2. Definitions.
Definition 6.1. Let Amax (N, c) denote the algorithmic complexity of the biggest
final output returned by a member of the population P in a maximum number
number of cycles c, where N = (G, P, b).
Definition 6.1.1. In the case of NBB (N, f, t, τ, j) and c = 1, for the sake of
simplifying our notation, we will just denote it as Amax .
Definition 6.2. Let ti , t, t′ ∈ T (Gt ) with ti ≤ t ≤ t′ . We denote the fraction to
which the node/program with the “best” partial output on time instant ti in the
algorithmic network N = (G, P, b) diffuses its partial output during time instants
t until t′ , given that this diffusion started at time instant ti , as
′
τmax (N, ti )|tt =
Xτmax (N,ti )|t′
t
|V (G)|
or
′
τmax (G, P, b, ti )|tt =
Xτmax (N,ti )|t′
t
|V (G)|
Note 6.2.1. The notion of what is the “best” partial output may vary on how the
algorithmic network N is defined — in some algorithmic networks the notion of
“best” partial output may even be not defined. We consider the “best” partial
output as being the one that always affects the neighbors to which it is shared by
making them return a final result that is at least as “good” as the one that the
node with the “best” partial output — that is, the one that started this diffusion
— initially had. How good is a final result also depends on how N is defined and on
how one defines what makes a result better than another (e.g. a fitness function).
While this general definition is not formally stated, in the particular case of the
present paper, these matters become formal and precise in Definition 6.2.1 .
Note 6.2.2. Note that a node/program does only belong to Xτmax (N,ti )|t′ iff its
t
partial outputs between time instants t and t′ historically depends on the information diffused by the node/program with the “best” partial output on time instant
ti .
Note 6.2.3.
′
0 ≤ τmax (N, ti )|tt ≤ 1
26
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Definition 6.2.1. We denote the fraction to which the node/program with the
biggest partial output at the first cycle in the algorithmic network NBB (N, f, ti , τ, j)
diffuses this partial output during time instants t until t′ , given that this diffusion
started at time instant ti , as
′
τmax (N, f, ti , τ, j)|tt =
Xτmax (N,f,ti ,τ,j)|t′
t
N
Definition 6.2.1.1. In the average case for all possible respective node mappings
into the population, we denote the average (or expected) singleton diffusion
density as
′
τE(max) (N, f, ti , τ )|tt =
X τmax (N, f, ti , τ, j)|t′
t
bj
|{bj }|
7. Time centralities
7.1. Discussion on how to trigger EEOE. We have previously mentioned that
our model of algorithmic networks playing the BBIG was built on time-varying
graphs (see Section 3 ), so that the static case is also covered as an instance of
time-varying graphs which every time-instant snapshot Gt (see [41]) is equal — see
also note 9.1. In fact, a preliminar result for static graphs was presented in [2].
The reader is also invited to note that a model in which the population size varies
during the cycles is a particular case of NBB (N, f, ti , τ, j) where N = ω 44 and
each new edge is added in Gt in the exact cycle (or time instant) that an edge
links the older network’s population to a newborn node/program 45 — and we
leave for future research proving the present results for graphs with |V (Gt )| = ω.
Within the framework of algorithmic dynamic networks in which one is interested
in investigating conditions that trigger AEOE, a natural question is when is the
best time instant (or cycle) to allow communication between nodes/programs with
the purpose of generating more and more AEAC. Note that there may be cases
where there is a cost in using a communication channel (i.e., an edge) for example.
Since the information sharing relies only on diffusion processes in the BBIG played
by algorithmic networks NBB (N, f, ti , τ, j), we tackle this problem following the
approach and metrics developed in [27, 41]. The more powerful the diffusion, the
more NBB (N, f, ti , τ, j) can be optimized (see discussion 3.1). So, in the present
case, finding the central time to start a diffusion process is directly related to finding
the central time to trigger expected emergent open-endedness.
We will present two related types of time centralities on NBB (N, f, ti , τ, j) (see
7.2.1). Both relies on cover time 3.3 3.4 as our diffusion measure as discussed in
Subsection 3.2 and both depends on how time instants correspond to cycles (see
function c(x) in 7.1 and 7.2) . First, we define the central time to trigger expected
emergent open-endedness in 7.1. Second, we define the central time to trigger
maximum expected emergent algorithmic complexity in 7.2 . One can think of
these time centralities as the minimum value t for which the function f (N, t, τ )
minimizes the number of cycles, that depends on t + f (N, t, τ ), in order to generate
44 Note that ω is the first ordinal number and it an ordered set with the same cardinality of
the natural numbers N.
45 And analogously a similar procedure applies when nodes are being “killed” during the cycles.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
27
increasing expected emergent algorithmic complexity of a node when N tends to
∞, given a fraction τ of the nodes/programs as a diffusion “yardstick”.
7.2. Definitions.
Definition 7.1. Let w ∈ LU be a network input. Let 0 < N ∈ N. Let c(x) be a
non-decreasing total computable function where
c: N →
N∗
x 7→ c(x) = y
Let NBB (N, f, tz , τ, j) = (Gt , PBB (N ), bj ), where 0 ≤ z ≤ |T (Gt )| − 1, be welldefined, where there is tz0 ∈ T (Gt ) such that46
net
lim ENBB (N,f,tz0 ,τ ) ∆ A(oi , c(z0 + f (N, tz0 , τ ) + 2)) = ∞
N →∞
iso
We define the central time tcen1 in generating unlimited expected emergent algorithmic complexity of a node on a diffusion process over a fraction τ as the population grows that minimizes the number of cycles47 as a function48 of the topological
diffusion measure f (N, t, τ ) as
(
n
tcen1 (c) = min tj j + f (N, tj , τ ) = min ti + f (N, ti , τ ) ti + f (N, ti , τ ) ≤
)
net
o
≤ z0 + f (N, tz0 , τ ) ∧ lim ENBB (N,f,ti ,τ ) ∆ A(oi , c(i + f (N, ti , τ ) + 2)) = ∞
N →∞
iso
Definition 7.2. Let w ∈ LU be a network input. Let 0 < N ∈ N. Let c(x) be a
non-decreasing total computable function where
c: N →
N∗
x 7→ c(x) = y
Let NBB (N, f, tz , τ, j) = (Gt , PBB (N ), bj ), where 0 ≤ z ≤ |T (Gt )| − 1, be welldefined, where there is tz0 ∈ T (Gt ) such that49
|T (Gt )|−1
∀x ∈ N|0
lim ENBB (N,f,tx ,τ )
N →∞
≤ lim ENBB (N,f,tz0 ,τ )
N →∞
net
∆ A(oi , c(x + f (N, tx , τ ) + 2)) ≤
net
iso
∆ A(oi , c(z0 + f (N, tz0 , τ ) + 2))
iso
!
We define the central time tcen2 in generating the maximum expected emergent
algorithmic complexity of a node on a diffusion process over a fraction τ as the population grows that minimizes the number of cycles as a function of the topological
diffusion measure f (N, t, τ ) as
46 This condition directly assures that this definition of time centrality is well-defined.
47 And communication rounds.
48 A computable function.
49 This condition directly assures that this definition of time centrality is well-defined.
28
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
(
n
tcen1 (c) = min tj j + f (N, tj , τ ) = min ti + f (N, ti , τ ) i + f (N, ti , τ ) ≤
net
≤ z0 + f (N, tz0 , τ ) ∧ ∀x lim ENBB (N,f,tx ,τ ) ∆ A(oi , c(x + f (N, tx , τ ) + 2)) ≤
iso
N →∞
o)
net
≤ lim ENBB (N,f,ti ,τ ) ∆ A(oi , c(i + f (N, ti , τ ) + 2))
N →∞
iso
Note 7.2.1. Note that, by definition, since |T (Gt )| ≤ ∞, if tcen1 is well-defined,
then tcen2 = tcen1 .
8. An EEOE Phenomenon in the BBIG
In this section we will prove lemmas, theorems and corollaries with the purpose
of showing that NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ) is an algorithmic network
capable of EEOE (see Definition 4.5.1 ) under certain topological conditions of the
graph Gt (see Corollary 8.1.1 ). We will show that there is a trade-off between the
average diffusion density and the number of cycles (see Theorem 8.1). Moreover,
once these topological properties are met, the concept of central time to trigger
EEAC within the minimum number of cycles becomes well-defined (see Theorem
8.2).
We split the proof of the main Theorem 8.1 in six Lemmas 8.1, 8.2, 8.3, 8.4,
8.5 and 8.6. Corollary 8.1.1 does not only enables one to apply Theorem 8.1 to a
network diffusion measure like cover time 3.3 but also bridges EEAC to time centrality by linking Theorems 8.1 and 8.2 . The main idea behind the construction of
the proof of the main Theorem 8.1 comes from combining an estimation of a lower
bound for the average algorithmic complexity of a networked node/program and
a estimation of the expected algorithmic complexity of a isolated node/program.
While the estimation of the former comes from the very BBIG dynamics, the estimation of the latter comes from the strong law of large numbers, Gibb’s inequality
and algorithmic information theory applied on the randomly generated population
PBB (N ). Thus, calculating the former estimation minus the latter gives directly
the EEAC (see defnition 4.4). Corollary 8.1.1 follows directly from Theorem 8.1.
The proof of Theorem 8.2 is a mathematical adjustment of definitions 7.1 and 7.2
into Corollary 8.1.1.
In appendix 12 we present extended versions of these proofs, except for the
Corollary 8.1.1 and 9.1, which will be presented in the next section.
Lemma 8.1 (12.1.1). Let NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ) be an algorithmic
network. Let N be the size of a randomly generated population PBB (N ). Then,
on the average as N grows, we will have that there is a constant C4 such that50
Amax ≥ lg(N ) − C4
Note 8.1.1. Let P [ X = a ] be the usual notation for the probability of a random
variable X assuming value a. Or P [ statement S ] denote the probability of a
50 This result — on random samples following algorithmic probabilities — is inspired by a
similar argument found in [21] with the purpose of building a true open-ended evolutionary model
for software — see also [1, 23].
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
29
statement S be true. Thus, the conclusion of this theorem can be also stated
using the strong law of large numbers [14] as: there is a constant C4 such that
h
i
P lim Amax − lg(N ) + C4 ≥ 0 = 1
N →∞
In fact, the main theorems and Corollaries 8.1, 8.1.1, 8.2 and 9.1 presented in this
article can be translated into such probabilistic form by putting their last statements
(or conclusions) into the square brackets like we just did to the above. Note that
they all originate from Lemma 8.1. For the sake of simplifying the notation and
shortening the formulas, we have chosen to state our results without using this
probabilistic form.
Short proof. Let φN (p) be the frequency that p occurs in a random sample of size
N . From the strong law of large numbers and algorithmic information
theory (AIT) we have that
1
(1)
P lim φN (p) = |p| = 1
N →∞
2
Let BB(k) be the Busy Beaver value for an arbitrary k ∈ N on machine U. From
AIT, we know there are constants CΩ , CBB ≥ 0 and a program pBB(k) such that
for every w ∈ LU
U(pBB(k) ◦ w) = U(pBB(k) ) = BB(k)
and
k − CΩ ≤ A(BB(k)) ≤ |pBB(k) | ≤ k + CBB
and
∀x > BB(k) A(x) + CBB ≥ A(BB(k))
Since k was arbitrary, let k = lg(N ) − CBB . From step 1 we will have that, when
N is big enough, one should expect that pBB(lg(N )−CBB ) occurs
N
N
≥ lg(N ) = 1
2
2|pBB(lg(N )−CBB ) |
times within N random tries. Let C4 = CΩ + 2 CBB . Thus, from definitions 3.10
and 6.1.1, since any node/program count as isolated from the network when c = 1,
we will have that, when N → ∞,
Amax ≥ A(U(Pprot ◦ pBB(lg(N )−CBB ) )) − CBB = A(U(pBB(lg(N )−CBB ) )) − CBB =
= A(BB(lg(N ) − CBB )) − CBB ≥ lg(N ) − CBB − CΩ − CBB = lg(N ) − C4
Lemma 8.2 (12.1.2). Given a population PBB (N ) defined in 3.11 , where pi ∈
Haltiso (LPBB (N ) , w, c) and Pprot ◦ pi ∈ PBB (N ) and N ∈ N∗ is arbitrary, there is
a constant C1 such that
A(U(piso (pi , c))) ≤ C1 + |pi | + A(w) + A(c)
Short proof. We have that there is at least one program p′ such that, for every
pi ∈ Haltiso (LPBB (N ) , w, c),
U(piso (pi , c)) = U(p′ ◦ pi ◦ w ◦ c)
30
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Take the smallest such program p′ and let C1 = |p′ | + 5 − |p+1 |, where p+1 is the
smallest program that adds 1 to any computation (see Definition 3.5 ). Then, from
AIT and Definition 3.5 we will have that
A(U(piso (pi , c))) = A(U(p′ ◦ pi ◦ w ◦ c)) ≤ C1 + |pi | + A(w) + A(c)
Lemma 8.3 (12.1.3). Given a population PBB (N ) defined in 3.11 , where pi ∈ LU
and Pprot ◦ pi ∈ PBB (N ), we will have that
X
|p
|
1
i
lim
+ lg(Ω(w, c)) ≤ lim lg(Ω(w, c)N )
N →∞
Ω(w, c) N →∞
2|pi |
pi ∈Haltiso (LU (N ),w,c)
Short proof. Since LU is self-delimited, from AIT and the Definition 5.2 , we will
have that
X
1
(2)
lim
=1
N →∞
2|pi |+lg(Ω(w,c))
pi ∈Haltiso (LU (N ),w,c)
From step 2 ,
(3)
X
lim
N →∞
pi ∈Haltiso (LU (N ),w,c)
1
lim
=
Ω(w, c) N →∞
|pi | + lg(Ω(w, c))
=
2|pi |+lg(Ω(w,c))
X
pi ∈Haltiso (LU (N ),w,c)
|pi |
+ lg(Ω(w, c))
2|pi |
Since step 2 holds, from Gibb’s (or Jensen’s) inequality [28, 49], we will have that
(4)
X
pi ∈Haltiso (LU (N ),w,c)
|pi | + lg(Ω(w, c))
≤ lg(|Haltiso (LU (N ), w, c)|)
2|pi |+lg(Ω(w,c))
We also have by Definition 5.2 that
(5)
lim lg(|Haltiso (LU (N ), w, c)|) = lim lg(Ω(w, c)N )
N →∞
N →∞
Therefore, from steps 3 , 4 and 5 we will have that
X
1
|p
|
i
lim
+ lg(Ω(w, c)) ≤ lim lg(Ω(w, c)N )
N →∞
Ω(w, c) N →∞
2|pi |
pi ∈Haltiso (LU (N ),w,c)
Lemma 8.4 (12.1.4). Given a population PBB (N ) defined in 3.11 , where pi ∈ LU
and Pprot ◦ pi ∈ PBB (N ), we will have that
P
lim
N →∞
P pi ∈Haltiso (LPBB (N ) ,w,c)
bj
|pi |
N
|{bj }|
≤ lim Ω(w, c) lg(N )
N →∞
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
31
Short proof. From the definition of language LBB in 3.10 , as in 8.1, we will have
that
P
|pi |
P pi ∈Haltiso (LPBB (N ) ,w,c)
N
bj
(6)
X
= lim
|{bj }|
N →∞
pi ∈Haltiso (LU (N ),w,c)
|pi |
2|pi |
And from Lemma 8.3 we will have that
X
|pi |
≤ lim Ω(w, c) lg(N )
(7)
lim
N →∞
2|pi | N →∞
pi ∈Haltiso (LU (N ),w,c)
Then, from steps 6 and 7
P
N →∞
=
|{bj }|
X
= lim
N →∞
N
bj
lim
|pi |
P pi ∈Haltiso (LPBB (N ) ,w,c)
pi ∈Haltiso (LU (N ),w,c)
|pi |
≤ Ω(w, c) lg(N )
2|pi |
Lemma 8.5 (12.1.5). Given a population PBB (N ) defined in 3.11 , where pi ∈ LU
and Pprot ◦ pi = oi ∈ PBB (N ), there is a constant C0 such that
P
lim
N →∞
P
A(U(piso (pi ,c)))
pi ∈Haltiso (LP
,w,c)
BB (N )
N
bj
= C0 (1 − Ω(w, c))
|{bj }|
Short proof. As in step 6 , A(U(piso (pi , c))) is independent of any topology, so that,
from the Definition 5.1.1 as in step 6 , we will have that
(8)
P
A(U(piso (pi ,c)))
lim
N →∞
P pi ∈Haltiso (LPBB (N ) ,w,c)
N
bj
X
= lim
|{bj }|
N →∞
pi ∈Haltiso (LU (N ),w,c)
A(U(piso (pi , c)))
2|pi |
Let A(0) = C0 . Since pi ∈ Haltiso (LU (N ), w, c) and PBB (N ) is sensitive to
oracles 3.11 , by the definition of piso in 4.1.1 we will have that, for every pi ∈
Haltiso (LU , w, c),
(9)
A(U(piso (pi , c))) = A(0) = C0
Thus, by the definition of Ω(w, c) in 5.2 ,
X
A(U(piso (pi , c)))
= C0 (1 − Ω(w, c))
(10)
lim
N →∞
2|pi |
pi ∈Haltiso (LU (N ),w,c)
And we conclude from steps 8 , 9 and 10 that
P
lim
N →∞
A(U(piso (pi ,c)))
P pi ∈Haltiso (LPBB (N ) ,w,c)
bj
N
|{bj }|
= C0 (1 − Ω(w, c))
32
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Lemma 8.6 (12.1.6). Let PBB (N ) be a population in an arbitrary algorithmic
network NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ) as defined in 3.13 and 3.11 . Let
t0 ≤ t ≤ t′ ≤ t|T (Gt )|−1 . Let c ∈ CBB be an arbitrary number of cycles where
c0 + t′ + 1 ≤ c. Then, there is a constant C2 such that
P
bj
P pi ∈PBB (N ) A(U(pnet (oi ,c)))
N
bj
|{bj }|
′
≥ (Amax − C2 ) τE(max) (N, f, t, τ )|tt + C2
Short proof. Let C2 = min{A(w) | ∃x ∈ LU (U(x) = w)}. From the definitions
6.2.1, 6.2 , 3.10 and 6.1.1 we will have that
P
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c)))
N
bj
|{bj }|
=
P
bj
+
≥
b
j
A(U(pnet
(oi ,c)))
P oi ∈Xτmax (N,f,t,τ,j)|tt′
′
τmax (N,f,t,τ,j)|t N
bj
(11)
P
P
bj
A(U(pnet
(oi ,c)))
oi ∈X
D(Gt ,t)
τmax (N,f,t,τ,j)| ′
t
D(G ,t)
τmax (N,f,t,τ,j)|t′ t N
bj
′
τmax (N, f, t, τ, j)|tt
t
|{bj }|
P
=
D(Gt ,t)
′
D(Gt ,t)
Amax τmax (N, f, t, τ, j)|tt + C2 τmax (N, f, t, τ, j)|t′
|{bj }|
=
τmax (N, f, t, τ, j)|t′
|{bj }|
P
′
(Amax − C2 )τmax (N, f, t, τ, j)|tt + C2
bj
|{bj }|
+
≥
=
Theorem
8.1 (12.1.1). A lower bound on EEACBB from an arbitrary number of
cycles . Let w ∈ LU be a network input. Let 0 < N ∈ N. Let NBB (N, f, t, τ, j) =
(Gt , PBB (N ), bj ) be well-defined. Let t0 ≤ t ≤ t′ ≤ t|T (Gt )|−1 . Let
c : N → CBB
x 7→ c(x) = y
be a total computable function where c(x) ≥ c0 + t′ + 1. Then, we will have that:
net
′
lim ENBB (N,f,t,τ ) ∆ A(oi , c(x)) ≥ lim τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N )−
N →∞
iso
N →∞
− Ω(w, c(x)) lg(x) − Ω(w, c(x)) lg(lg(x)) − A(w) − C5
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
33
Note 8.1.1. Thus, note that for bigger enough values of c(x) compared to lg(N )
one can make this lower bound always negative for example. One of the main ideas
behind forthcoming results in this paper is to find optimal conditions where this
lower bound is not only positive, but also goes to ∞.
Note 8.1.2. Note that this lower bound on the expected emergent algorithmic complexity is dependent on the value in the domain of the function c and not on function
c itself, even if it grows fast. In fact, one may want to obtain this theorem for fixed
values of c in which it is not a function but an arbitrary value. And the same result
also hold in this case. The reader is invited to check that a simple substitution of
c(x) for c and of x inside the logarithms for c is enough.
Note 8.1.3. The same result also holds if only one possible function bj is defined
for each member of the family G(f, t, τ ). However, the notion of expected emergent
algorithmic complexity (EEAC) gains some redundancy (see note 4.4.1 ), as only
one function bj will be taken into account within the sum in order to give the mean.
Thus, in this case one can replace τE(max) with τmax not only in Theorem 8.1 but
also in 8.1.1 and 8.2 . Such variation of these theorems becomes useful when one
has algorithmic networks NBB (N, f, t, τ, j) built upon a historical population size
growth in which each new node/program is linked (or not) to the previous existing
algorithmic network.
Short proof. We have from AIT that there is Cc ∈ N for every x ∈ N such that
(12)
A(c(x)) ≤ Cc + A(x)
Let C5 = Cc + CL + C1 + C4 . From Definition 4.4, we have that the expected emergent algorithmic complexity of a node/program for NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ),
where 0 < j ≤ |{bj }| is given by
(13)
P
ENBB (N,f,t,τ )
net
∆ A(oi , c(x))
=
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c(x))))−A(U(piso (pi ,c(x))))
N
bj
|{bj }|
And from definitions 5.1 , 5.1.1 , 5.2 , 6.2.1.1 , 3.5 and Lemmas 8.2 , 8.5 , 8.4 , 8.6
, 8.1 and step (12) we will have that
iso
P
(14)
lim
N →∞
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c(x))))−A(U(piso (pi ,c(x))))
N
bj
|{bj }|
P
≥ lim
P
b
j
−
P
N →∞
≥
b
P oi ∈PBB (N )
j
A(U(pnet
(oi ,c(x))))
bj
N
|{bj }|
C1 +|pi |+A(w)+A(c(x))
pi ∈Haltiso (LP
,w,c(x))
BB (N )
N
|{bj }|
−
+ C0 (1 − Ω(w, c(x)))
≥
34
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
′
≥ lim (Amax − C2 ) τE(max) (N, f, t, τ )|tt + C2 −
N →∞
− Ω(w, c(x)) lg(N ) + Ω(w, c(x)) C1 + A(w) + A(c(x)) + C0 1 − Ω(w, c(x)) ≥
′
≥ lim (lg(N ) − C4 − C2 ) τE(max) (N, f, t, τ )|tt + C2 −
N →∞
− Ω(w, c(x)) lg(N ) + Ω(w, c(x)) C1 + A(w) + A(c(x)) + C0 1 − Ω(w, c(x)) ≥
≥ lim
N →∞
′
τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N )−
−Ω(w, c(x)) lg(x) − Ω(w, c(x)) lg(lg(x)) − C5 − A(w)
Corollary 8.1.1. A
lower bound on EEACBB from a diffusion process as a function of cover time . Let w ∈ LU be a network input. Let 0 < N ∈ N. Let
NBB (N, f, tz , τ, j) = (Gt , PBB (N ), bj ) be well-defined. Let tz ∈ T (Gt ). Let
τ ∈ ]0, 1]. Let c : N → CBB be a total computable function where
x 7→ c(x) = y
c(z + f (N, tz , τ ) + 2) ≥ c0 + z + f (N, tz , τ ) + 2
Then, we will have that:
net
lim ENBB (N,f,tz ,τ ) ∆ A(oi , c(z + f (N, tz , τ ) + 2)) ≥
iso
N →∞
t
≥ lim τE(max) (N, f, tz , τ )|tzz+f (N,tz ,τ ) − Ω(w, c(z + f (N, tz , τ ) + 2)) lg(N )−
N →∞
− Ω(w, c(z + f (N, tz , τ ) + 2)) lg(z + f (N, tz , τ ) + 2)−
− Ω(w, c(z + f (N, tz , τ ) + 2)) lg(lg(z + f (N, tz , τ ) + 2)) − A(w) − C5
Proof. Remember the Definition 3.13 . Note that z + f (N, tz , τ ) goes as an index of
tz+f (N,tz ,τ ) , so that one must add 1 on the right side of the condition c(x) ≥ c0 +t′ +1
in 8.1 in order to make c0 + tz+f (N,tz ,τ ) + 1 = c0 + z + f (N, tz , τ ) + 2. Thus, the
proof follows directly from replacing t with tz , t′ with tz+f (N,tz ,τ ) and x with
z + f (N, tz , τ ) + 2 in 8.1 .
Theorem 8.2 (12.1.2). Central time in reaching unlimited
EEAC from the BBIG
played by a randomly generated algorithmic population . Let w ∈ LU be a network
input. Let 0 < N ∈ N. Let NBB (N, f, tz , τ, j) = (Gt , PBB (N ), bj ) be well-defined.
If there is 0 ≤ z0 ≤ |T (Gt )| − 1 and ǫ, ǫ2 > 0 such that
C
N
z0 + f (N, tz0 , τ ) = O
lg(N )
where
tz
−1−ǫ ≤ C =
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2) − ǫ
Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2)
Then, for every non-decreasing total computable function c : N → CBB where
x 7→ c(x) = y
tz0 ∈ T (Gt ) and c(z0 + f (N, tz0 , τ ) + 2) ≥ c0 + z0 + f (N, tz0 , τ ) + 2 we will have
that there are tcen2 (c) and tcen1 (c) such that
tcen2 (c) = tcen1 (c) ≤ tz0
≤
1
ǫ2
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
35
Short proof. Suppose that there is tz0 ∈ T (Gt ), where 0 ≤ z0 ≤ |T (Gt )| − 1, and a
small enough ǫ > 0 such that
C
N
(15)
z0 + f (N, tz0 , τ ) = O
lg(N )
where
tz
C=
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2) − ǫ
Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2)
Let
tz
τE(max) (N, f, tz0 , τ )|tz00
′
C =
+f (N,tz0 ,τ )
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) − ǫ
Ω(w, c(z0 + f (N, tz0 , τ ) + 2))
Thus, since we are assuming c(z0 + f (N, tz0 , τ ) + 2) ≥ c0 + z0 + f (N, tz0 , τ ) + 2, for
tz
fixed values of τE(max) (N, f, tz0 , τ )|tz00
tz
τE(max) (N, f, tz0 , τ )|tz00
′
(16) C =
+f (N,tz0 ,τ )
and ǫ, we will have that
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) − ǫ
Ω(w, c(z0 + f (N, tz0 , τ ) + 2))
tz
≥
+f (N,tz0 ,τ )
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2) − ǫ
Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2)
Hence, from steps 15 and 16 we will have that there is ǫ2 such that
!
′
NC
(17)
z0 + f (N, tz0 , τ ) = O
lg(N )
≥
=C
where
tz
′
−1−ǫ ≤ C =
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) − ǫ
Ω(w, c(z0 + f (N, tz0 , τ ) + 2))
≤
C′
N
And since z0 + f (N, tz0 , τ ) is now assymptotically dominated by lg(N
) , then by
definition we will have from definitions 5.2 and 6.2.1.1 and Corollary 8.1.1 and step
17 that there is a constant C6 such that
net
(18)
lim ENBB (N,f,tz0 ,τ ) ∆ A(oi , c(z0 + f (N, tz0 , τ ) + 2)) ≥
N →∞
≥ lim
N →∞
iso
tz
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(N )−
′
NC
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(C6
)−
lg(N )
′
NC
)) − A(w) − C5 ≥
lg(N )
1
≥ lim (ǫ) lg(N ) − lg(C6 ) − lg(2) − lg( ) − A(w) − C5 = ∞
N →∞
ǫ2
Then, directly from the definitions 7.1 and 7.2 and step 18 , since tz0 satisfy these
definitions, we will have that
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(lg(C6
tcen2 (c) = tcen1 (c) ≤ tz0
1
ǫ2
36
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
9. EEOE from a small diameter
We have previously established definitions and theorems for an abstract toy
model of a simple optimization of average fitness through diffusion performed by a
randomly generated population of interacting systems. It is shown in Theorem 8.1
that there are conditions (e.g. in Theorem 8.2) on the average diffusion density 6
and the number of cycles 2.2.1.1 that are sufficient to make the lower bound on the
EEAC goes to infinity when the populations size goes to infinity, giving rise to a
phenomenon we call expected emergent open-endedness (EEOE).
Following this purpose of theoretical investigation, we present in this section an
application of our results using an important property in complex networks that
fulfills these conditions mentioned to the above. Characterizing a small-diameter
network, in which either the average shortest path length (also called as average
geodesic distance) or the diameter (i.e., the maximum shortest path length) is
“small” compared to the network size (i.e., when it is dominated by O(lg(N ))), is a
relevant topic in the current literature of complex networks [9,10,15,36]. Therefore,
in this section we aim to exemplify and reinforce the relevance of the study of
models like the BBIG within the theory of algorithmic networks in relation to
small-diameter networks (see Sections 1 and 2 for more discussions). In fact, we
will show that a small maximum shortest path length (that is, a small diameter)
gives rise to an emergent phenomenon: it increases the very potential51 of average
(or, in our case, expected) emergent open-endedness (AEOE or EEOE), should
the respective network link a population of systems that mathematically behave
like NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ) (see Definition 3.13 ). Thus, we not
only present a model in which the small-diameter phenomenon (SD) plays an
central role in achieving the EEOE, but also suggest in a general way that AEOE in
algorithmic networks might be related to why the SD would appear in real networks
— in which emergent complexity might be somehow advantageous to some degree.
A family Gsm of graphs will be defined in the same way as G in Definition 3.4.
However, with two additional properties: first, there is a time instant on which every node can reach all nodes in a finite number of time intervals; second, taking this
time instant as the starting point of diffusion, the graphs have a temporal diffusion
diameter (see Definition 3.3.1) small compared to the size of the graph. The main
idea is to ensure that a small-diameter phenomenon for dynamic networks, which
relies on diffusion processes, eventually appears as the population grows. Then, one
can directly define an algorithmic network NBB sm (N, f, t, 1, j) that is totally analogous to NBB (N, f, t, τ, j) except for replacing family G(f, t, τ ) with Gsm (f, t, 1).52
Thus, we will prove in Corollary 9.1, using in Theorem 8.2 the flexibility of Theorem
8.1 and Corollary 8.1.1, that these conditions on Gsm (f, t, 1) are sufficient for, in a
computably bigger number of cycles compared to the temporal diffusion diameter,
existing a central time to trigger EEOE.
51 In fact, a lower bound on the EEAC.
52 Further, as it is straightforward, we choose to omit a formal definition of N
BB sm (N, f, t, 1, j)
in 9.2 .
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
37
Definition 9.1. Let
Gsm (f, t, 1) = {Gt | i = |V (Gt )| ∧ f (i, t, 1) = D(Gt , t) = O lg(i) ∧
∧ ∀i ∈ N∗ ∃!Gt ∈ G(f, t, τ )( |V (Gt )| = i ) ∧ ∀u ∈ V (Gt )∃x ∈ N x = dt (Gt , t, u, 1) }
where
f : N∗ × X ⊆ T (Gt ) × Y ⊆]0, 1] → N
(x, t, τ )
7→ y
be a family
of
unique
time-varying
graphs
which
shares
f (i, t, 1) = D(Gt , t) =
O lg(i) , where i is the number of nodes, as common properties. The number of
nodes of each graph in this family may vary from 1 to ∞.
Note 9.1.1. Note thatthe diameter might be much
smaller indeed. For example,
lg(N )
or
D(G
,
t)
=
O
lg(lg(N
))
are covered by Definition 9.1 .
D(Gt , t) = O lg(lg(N
t
))
Note 9.1.2. As pointed in 3.4.1 and 12.1.1.1 , Corollary 9.1 also holds with a weaker
assumption on family Gsm (f, t, 1) in which
h
i
=1
P lim N = |V (Gt )| =⇒ D(Gt , t) = O lg(N )
N →∞
Definition 9.2. Let the algorithmic network NBB sm (N, f, t, 1, j) denote the same
algorithmic network NBB (N, f, t, τ, j) defined in 3.13 except for replacing family
G(f, t, τ ) with family Gsm (f, t, 1).
Corollary 9.1. Let w ∈ LU be a network input. Let 0 < N ∈ N.
Let NBB sm (N, f, tz0 , 1, j) = (Gt , PBB (N ), bj ) be well-defined. Then, for every
non-decreasing total computable function c : N → CBB where tz0 ∈ T (Gt ) and
x 7→ c(x) = y
c(z0 + f (N, tz0 , 1) + 2) ≥ c0 + z0 + f (N, tz0 , 1) + 2 we will have that there are tcen2 (c)
and tcen1 (c) such that
tcen2 (c) = tcen1 (c) ≤ tz0
Proof. The proof is analogous to Theorem 8.2. However, the average singleton
diffusion density for a number of time intervals equal to the temporal diffusion
diameter assumes its maximum value 1. And the small-diameter condition on
family Gsm (f, t, 1) will assure that the conditions of Theorem 8.2 are met. First,
remember that Corollary 8.1.1 from Theorem 8.1 only depends on function f (i, t, τ )
and not on it being equal to CT (Gt , t, τ ). Second, the same is also valid for Theorem
8.2 that only depends on Corollary 8.1.1 and also not on f (i, t, τ ) = CT (Gt , t, τ ).
Thus, we just replace τ with 1 in Theorem 8.2 and the same proposition will also
holds for
(19)
f (i, t, 1) = D(Gt , t) = O lg(i)
Since Gt ∈ Gsm (f, t, 1), we will have from Definition 9.1 that
(20)
z0 + f (N, tz0 , 1) = z0 + D(Gt , tz0 ) = z0 + O (lg(N ))
where 0 ≤ z0 ≤ |T (Gt )| − 1. As in steps 15 and 17, we will have from definitions
3.3.1, 5.2, 6.2.1.1 and 9.1 and step 20 that there are constants ǫ, ǫ2 > 0 such that
(21)
tz
−1−ǫ<0<C =
τE(max) (N, f, tz0 , 1)|tz00
+f (N,tz0 ,1)
− Ω(w, c0 + z0 + f (N, tz0 , 1) + 2) − ǫ
Ω(w, c0 + z0 + f (N, tz0 , 1) + 2)
=
38
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
=
We have that if
(22)
1
ǫ2
1 − Ω(w, c0 + z0 + f (N, tz0 , 1) + 2) − ǫ
1
≤
Ω(w, c0 + z0 + f (N, tz0 , 1) + 2)
ǫ2
≥ C > 0, then
D(Gt , tz0 ) = O (lg(N )) =⇒ D(Gt , tz0 ) = O
NC
lg(N )
Then, from steps 21 and 22 and 19 we will have that there are tcen2 (c) and tcen1 (c)
such that
tcen2 (c) = tcen1 (c) ≤ tz0
Note 9.1. As mentioned in discussions 3.2 and 7.1, note that the result 9.1 to the
above also holds for static graphs. These can be considered an instance of timevarying graphs in which every time snapshot Gt is equal. In this case, for strongly
connected static graphs G the usual diameter of graphs diam(G) [16] becomes
equivalent to D(G, t0 ). Also, the notion of time centrality will be vacuous, since
the graph’s topology is always the same. Hence, for static graphs, the Corollary
9.1 states only that the respective algorithmic network has the property of EEOE.
10. Related Work and Future Research
The results presented in this article link algorithmic information properties with
network topological ones. As the reader might notice, there may be graph particular
aspects that trigger AEOE (see Theorems 8.1 and 8.2 and Corollaries 8.1.1 and 9.1).
These are embedded in the average singleton diffusion density (Definition 6.2.1.1)
and function c that takes some time interval as input (see Corollaries 8.1.1 and
9.1 and Theorem 8.2). In fact, the AEOE phenomenon benefits from topologies
capable of faster spreading or flooding [25]. Although building variations of our
models to fit real complex networks measurements for scale-free networks in [58, 59]
(through a susceptible-infected-susceptible (SIS) variation of the BBIG model and
the correspondence between prevalence and average diffusion density) will be a
fruitful application of our models in network science, there would be no necessary
advantage in some particular heterogeneity on the network structure. Note that
our result 9.1 only uses the temporal diffusion diameter. We suggest that our
mathematical results might indeed corroborate with computational experimental
approaches in complex networks in [69] [68] [71]. If an average optimization through
diffusion corresponds somehow to average emergent complexity, graph sparseness
(also discussed in [25]) and the small-diameter property might play a central role
in achieving this optimization with minimum cost. A more accurate theoretical
analysis of this potential corroborations is paramount for future research.
With the same focus on complex networks analysis, an statistical or probabilistic information-theoretic counterpart of our work might be developed with further
investigation on [29, 53, 68]. The case more specific of an experimental approach
centered on the relation between living systems and graph properties using statistical information is presented in [64, 75].
Establishing relations between algorithmic information and statistical information like in [28, 40, 72, 81] within the context of algorithmic networks is crucial to
approximate or find equivalences between measures from algorithmic information
and measures from statistical information. In the particular case of an approach
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
39
centered on emergence and self-organization, we suggest further investigation from
[33, 61, 73]. Another interesting topic is whether the average (and the joint) emergent algorithmic complexity is sufficient to deal with synergistic and integrated
information [38, 39, 56, 62, 77, 78] or not, as it directly gives a measure of irreducible
information [18, 20, 48]. If one assumes the definition of synergy as the general
phenomenon in which the whole system is irreducibly better in solving a common
problem than the sum (or the union) of its parts taken separately, as is widely used
in statistical information-theoretic approaches cited in the previous sentence, then
there may be a possible interpretation of (joint or average) emergent algorithmic
complexity as (joint or average) synergistic algorithmic information. For this, note
that one may consider for example that the common goal of the algorithmic network
is to increase the average fitness of each node/program (see Subsection 3.1).
The present model requires a hypercomputable function to deal with eventual
nodes/programs that do not halt (see Definition 3.5). With the purpose of bringing
this Busy Beaver imitation game to computer simulations or just to a fully computable model, the recursive relative incompressibility (or sub-incompressibility) as
in [1, 3, 4] can be used to define a measure for resource-bounded algorithmic complexity [48] [6]. Or one can use a measure through asymptotic approximations to
algorithmic complexity as in [30, 70, 81] in order to develop further investigations
on algorithmic network complexity as an extension of an approach for graphs presented in [17, 79, 80]. This way one can aim to make the relative halting problem
tractable while keeping the fitness function still relatively uncomputable.
There are further necessary investigations from a game-theoretical perspective.
For example, in [54] it is shown a relation for reaching global consensus between
a specific type of connectiveness of graphs and a local strategy (in our notation,
protocols) for nodes. A transposition of this problem into an algorithmic network
model would help to mathematically understand global certification of best solutions (see Subsection 3.1) from local ones for example. If fitness or payoff [42] is
somehow connected to the complexity of the player’s strategy, algorithmic networks
would enable one to investigate game-theoretical consequences of randomly generated strategies for players without interaction or for networked players [7, 67] with
global strategies like the one in the BBIG.
Another variation of the BBIG model in which the problem of consensus may
be investigated is from synchronous distributed systems of failure-prone processors
connected by small-diameter graphs as presented in [24]. Adding the possibility of
failure or crash to the BBIG model will require modifications on the information
sharing protocols (see Section 3 and Definition 3.9) in order to solve the respective
problem of consensus [24] of choosing the fittest solution. A crash-prone system
would introduce new difficulties on the plain diffusion that we have presented in this
article; And a limit to the maximum expected number of failures in relation to the
population size seems to be necessary to keep the phenomenon of expected emergent
open-endedness in such algorithmic networks with crash-prone node/programs.
We suggest that other relations to the fields we have mentioned in this Section
and in Section 1 may configure a long term horizon for future research. For example, one may relate the problem of evolvability versus robustness in evolutionary
biology [12, 74] with intentional attack vulnerability and resilience to random failure in complex networks [37]. To this end, one can investigate a variation of the
BBIG model (see Section 3) in which the diffused best solution may correspond
40
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
to an increase in the distributed redundancy [63]. Following this direction in the
context of algorithmic networks, one may study why a network topology and an
information-sharing protocol become relevant from an emergent open-ended evolutionary point of view that takes into account complexity and computational power
in solving problems [43,44]. While it is conjectured in complex networks theory that
different types of real-world networks are indeed related by graph properties (e.g.
the small-world phenomenon [46]), the theoretical approach we are developing may
suggest (e.g. see Corollary 9.1) that they may be indeed [2], when one assumes that
problem solving, complexity increasing and networks are deeply connected through
the concept of computation.
11. Conclusion
In this article, we have presented definitions and theorems in order to investigate
the problem of emergence and complexity when a population of computable systems
is networked. In particular, we have investigated conditions that enable the triggering of emergent open-endedness, that is, these conditions trigger an unlimited
increase of emergent complexity as the population size goes to infinity.
First, we have introduced the general definition of an algorithmic network N.
This definition relies on a population of systems that runs on an arbitrarily chosen
theoretical machine and relies on a multiaspect graph. This makes the population
of systems to correspond to nodes in the respective multiaspect graph such that
edges are communication channels that nodes/systems can use to send and receive
information.
Second, we have defined a particular mathematical model of algorithmic networks
NBB based on the simple imitation of the fittest neighbor: A type of algorithmic
network that plays the Busy Beaver imitation game. In this model, the randomly
generated population of nodes/programs of a universal Turing machine is synchronous and non-halting nodes/programs always return the “worst” fitness/payoff.
This population is networked by a time-varying graph. Hence, topological measures like cover time and temporal diffusion diameter can be promptly well-defined.
Moreover, we have defined a graph-topological measure of diffusion (the average
diffusion density) and the cycle-bounded conditional halting probability. These algorithmic networks NBB can be seen as playing an optimization procedure where
the whole is “trying” to increase the average fitness/payoff by diffusing through the
network the best randomly generated solution.
Furthermore, we have presented the average (or expected) emergent algorithmic
complexity of a node/program as the average (or expected) value obtained from the
algorithmic complexity of what a node/program can do when running networked
minus the algorithmic complexity of what the same node/program can do when
running isolated. Thus, when this difference goes to infinity as the population size
goes to infinity, this configures a phenomenon we have called as the average (or
expected) emergent open-endedness, i.e., unlimited emergent complexity happens as
an emergent phenomenon that occurs as the population grows. This is a phenomenon akin to evolutionary open-endedness, but in the latter unlimited complexity
occurs as successive mutations and natural selection apply over time.
Since our network playing the Busy Beaver imitation game is linked by a timevarying graph, a centrality for dynamic networks is defined for the optimum time
instant to trigger average emergent open-endedness in a way that it minimizes the
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
41
amount of cycles (or communication rounds) needed to do this triggering. Moreover, we presented our main theorem proving that there is a lower bound on the
expected emergent algorithmic complexity for algorithmic networks NBB such that
it depends on how much larger is the average diffusion density in a given time interval compared to the cycle-bounded conditional halting probability. This lower
bound also depends on the parameter for which one is calculating the number of cycles. In fact, we prove a corollary showing that this parameter can be calculated for
example from the cover time, so that our results hold even in the case of spending
a computably bigger number of cycles compared to the cover time. Furthermore,
from this corollary, we proved that there are asymptotic conditions on the increasing power of the cover time as a function of the population size such that they
ensure that there is a central time to trigger expected emergent open-endedness.
Additionally, we have made a small modification on the family of time-varying
graphs of NBB with the purpose of investigating what would happen if the networks were small-diameter. We replaced the cover time with the temporal diffusion
diameter (i.e., the number of time intervals to reach every node from any node) —
or the classical diameter (i.e., the maximum shortest path length) in the static case
— in the definition of this family of graphs. Indeed, we proved that, in this case,
a small-diameter (i.e., dominated by O(lg(N )), where N is the network size) is
sufficient for existing a central time to trigger expected emergent open-endedness.
Finally, we have suggested future research and discussed that our results might
be related to problems in network science, information theory, computability theory,
distributed computing, game theory, evolutionary biology, and complex systems.
References
[1] Felipe S. Abrahão, Metabiologia, Subcomputação e Hipercomputação: em direção a uma teoria geral de evolução de sistemas, Ph.D. thesis, 2015.
, Emergent algorithmic creativity on networked Turing machines, The 8th interna[2]
tional workshop on guided self-organization at the fifteenth international conference on the
synthesis and simulation of living systems (alife), 2016.
[3]
, Relativizing an incompressible number and an incompressible function through subrecursive extensions of Turing machines, arXiv preprint (2016dec), available at 1612.05221.
[4]
, The “paradox” of computability and a recursive relative version of the Busy Beaver
function, Information and complexity, 2016jan, pp. 3–15.
[5] Alyssa Adams, Hector Zenil, Paul C. W. Davies, and Sara Imari Walker, Formal Definitions
of Unbounded Evolution and Innovation Reveal Universal Mechanisms for Open-Ended Evolution in Dynamical Systems, Scientific Reports 7 (2017dec), no. 1, 997.
[6] Eric Allender, Michal Koucký, Detlef Ronneburger, and Sambuddha Roy, The pervasive reach
of resource-bounded Kolmogorov complexity in computational complexity theory, Journal of
Computer and System Sciences 77 (2011), no. 1, 14–40.
[7] Robert Axelrod, The Complexity of Cooperation: Agent-Based Models of Competition and
Collaboration, Vol. 1, Princeton University Press, 1997.
, The Evolution of Cooperation: Revised Edition, Basic Books, 2006.
[8]
[9] Albert-László Barabási, Scale-Free Networks: A Decade and Beyond, Science 325 (2009jul),
no. 5939, 412–413, available at 0803973233.
[10] Albert-László Barabási and Eric Bonabeau, Scale-Free Networks, Scientific American 288
(2003may), no. 5, 60–69.
[11] George Barmpalias, Nan Fang, and Andrew Lewis-Pye, Optimal asymptotic bounds on the
oracle use in computations from Chaitin’s Omega, Journal of Computer and System Sciences
82 (2016dec), no. 8, 1283–1299.
[12] Patrick Bateson and Peter Gluckman, Plasticity, Robustness, Development and Evolution,
Cambridge University Press, Cambridge, 2011.
42
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
[13] Nils Bertschinger, Johannes Rauh, Eckehard Olbrich, Jürgen Jost, and Nihat Ay, Quantifying
unique information, Entropy 16 (2014apr), no. 4, 2161–2183, available at 1311.2852.
[14] Patrick Billingsley, Probability and Measure, Anniversar, John Wiley & Sons, Hoboken, 2012.
[15] Béla Bollobás and Oliver Riordan, The diameter of a scale-free random graph, Combinatorica
24 (2004jan), no. 1, 5–34.
[16] Béla Bollobás, Random Graphs, Cambridge University Press, Cambridge, 2001.
[17] Harry Buhrman, Ming Li, John Tromp, and Paul Vitányi, Kolmogorov Random Graphs and
the Incompressibility Method, SIAM Journal on Computing 29 (1999jan), no. 2, 590–599.
[18] Cristian S. Calude, Information and Randomness: An algorithmmic perspective, 2nd ed.,
Springer-Verlag, 2002.
[19] Gregory Chaitin, A Theory of Program Size Formally Identical to Information Theory, Journal of the ACM 22 (1975), no. 3, 329–340.
, Algorithmic Information Theory, 3rd ed., Cambridge University Press, 2004.
[20]
, Life as Evolving Software, A computable universe, 2012dec, pp. 277–302.
[21]
[22]
, Proving Darwin: making biology mathematical, Vintage Books, 2013.
[23] Gregory Chaitin, Virgı́nia Chaitin, and Felipe S. Abrahão, Metabiologı́a: los orı́genes de la
creatividad biológica, Investigación y Ciencia (2014), 74–80.
[24] Bogdan S. Chlebus and Dariusz R. Kowalski, Robust gossiping with an application to consensus, Journal of Computer and System Sciences 72 (2006dec), no. 8, 1262–1281.
[25] A. E. F. Clementi, C. Macci, A. Monti, F. Pasquale, and R. Silvestri, Flooding time of edgeMarkovian evolving graphs, SIAM Journal on Discrete Mathematics 24 (2010), 1694–1712.
[26] S. Barry Cooper, Emergence as a computability-theoretic phenomenon, Applied Mathematics
and Computation 215 (2009oct), no. 4, 1351–1360.
[27] Eduardo Chinelate Costa, Alex Borges Vieira, Klaus Wehmuth, Artur Ziviani, and Ana
Paula Couto da Silva, Time Centrality in Dynamic Complex Networks, Advances in Complex
Systems 18 (2015), no. 07n08, available at 1504.00241.
[28] Thomas M. Cover and Joy A. Thomas, Elements of Information Theory, John Wiley & Sons,
Inc., Hoboken, NJ, USA, 2005.
[29] Matthias Dehmer, Frank Emmert-Streib, and Alexander Mehler (eds.), Towards an Information Theory of Complex Networks, Birkhäuser Boston, Boston, 2011.
[30] Jean Paul Delahaye and Hector Zenil, Numerical evaluation of algorithmic complexity for
short strings: A glance into the innermost structure of randomness, Applied Mathematics
and Computation 219 (2012sep), no. 1, 63–77.
[31] Ítala Maria Loffredo D’Ottaviano and Ettore Bresciani, Sistêmica, Auto-Organização e
Criação, MultiCiência: revista interdisciplinar dos Centros e Núcleos da Unicamp 3 (2004),
1–23.
[32] Bernard Feltz, Marc Crommelinck, and Philippe Goujon, Self-organization and emergence in
life sciences (Bernard Feltz, Marc Crommelinck, and Philippe Goujon, eds.), Kluwer Academic Publishers, Dordrecht, 2006.
[33] Nelson Fernandez, Carlos Maldonado, and Carlos Gershenson, Information Measures of Complexity, Emergence, Self-organization, Homeostasis, and Autopoiesis, 2013, pp. 19–51.
[34] David B. Fogel, Evolutionary Computation: Toward a New Philosophy of Machine Intelligence, IEEE Press Series on Computational Intelligence, John Wiley & Sons, Inc., Hoboken,
NJ, USA, 2005.
[35] Carlos Gershenson and Nelson Fernández, Complexity and information: Measuring emergence, self-organization, and homeostasis at multiple scales, Complexity 18 (2012nov), no. 2,
29–44, available at 1205.2026.
[36] Carlos Gershenson and Mikhail Prokopenko, Complex Networks, Artificial Life 17 (2011oct),
no. 4, 259–261.
[37] Sam F. Greenbury, Iain G. Johnston, Matthew A. Smith, Jonathan P K Doye, and Ard A.
Louis, The effect of scale-free topology on the robustness and evolvability of genetic regulatory
networks, Journal of Theoretical Biology 267 (2010nov), no. 1, 48–61, available at 1005.4342.
[38] Virgil Griffith, Quantifying synergistic information, Ph.D. Thesis, 2014.
[39] Virgil Griffith and Christof Koch, Quantifying synergistic mutual information, 2012, pp. 159–
190.
[40] Peter Grunwald and Paul Vitanyi, Shannon Information and Kolmogorov Complexity, CoRR
cs.IT/0410 (2004oct), 54, available at 0410002.
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
43
[41] Abraão Guimarães, Alex B Vieira, Ana Paula C Silva, and Artur Ziviani, Fast CentralityDriven Diffusion in Dynamic Networks, Proceedings of the 22nd international conference on
world wide web companion, 2013, pp. 821–828.
[42] Peter Hammerstein and Reinhard Selten, Game theory and evolutionary biology, Handbook
of game theory with economic applications, 1994, pp. 929–993.
[43] Santiago Hernández-Orozco, Francisco Hernández-Quiroz, and Hector Zenil, The Limits of
Decidable States on Open-Ended Evolution and Emergence, Proceedings of the artificial life
conference 2016, 2016, pp. 200–207.
[44] Santiago Hernández-Orozco, Hector Zenil, and Narsis A Kiani, Algorithmically probable mutations reproduce aspects of evolution such as convergence rate, genetic memory, modularity, diversity explosions, and mass extinction, arXiv preprint (2017), 1–13, available at
1709.00268v3.
[45] Vyacheslav V. Ivanishev, The technology of modeling on a basis of visual representation, Proceedings of the international conference on informatics and control (ici&c’97), 1997, pp. 448–
450.
[46] Jon Kleinberg, The small-world phenomenon, Proceedings of the thirty-second annual acm
symposium on theory of computing - stoc ’00, 2000mar, pp. 163–170.
[47] Oleg Korolev, Vladimir Parfenov, and Semyon Potryasaev, Program System for Solving Problems of Monitoring Social and Economic Development of Saint-Petersburg, 2016, pp. 347–
358.
[48] Ming Li and Paul Vitányi, An Introduction to Kolmogorov Complexity and Its Applications,
2nd ed., Springer Science & Business Media, New York, 1997.
[49] David J. C. MacKay, Information Theory, Inference, and Learning Algorithms David J.C.
MacKay, 11th ed., Vol. 100, Cambridge University Press, 2005.
[50] Margulis Lynn, Symbiosis in Cell Evolution: Life and its Environment on the Early Earth.,
W.H Freeman & Co. (1981), 419.
[51] V. E. Marley, Operations over algorithmic networks, Proceedings of the international conference on informatics and control (ici&c’97), 1997, pp. 451–457.
[52] J H Miller and S E Page, Complex Adaptive Systems: An introduction to computational
models of social life, Princeton University Press, 2007.
[53] Abbe Mowshowitz and Matthias Dehmer, Entropy and the complexity of graphs revisited,
Entropy 14 (2012mar), no. 3, 559–570.
[54] Nabil H. Mustafa and Aleksandar Pekec, Listen to Your Neighbors: How (Not) to Reach a
Consensus, SIAM Journal on Discrete Mathematics 17 (2004jan), no. 4, 634–660.
[55] Gregoire Nicolis and Catherine Nicolis, Foundations of complex systems, Vol. 17, World Scientific Publishing Co. Pte. Ltd., 2009.
[56] Masafumi Oizumi, Larissa Albantakis, and Giulio Tononi, From the Phenomenology to the
Mechanisms of Consciousness: Integrated Information Theory 3.0, PLoS Computational Biology 10 (2014may), no. 5, e1003588.
[57] Raj Kumar Pan and Jari Saramäki, Path lengths, correlations, and centrality in temporal
networks, Physical Review E - Statistical, Nonlinear, and Soft Matter Physics 84 (2011jul),
no. 1, 016105, available at 1101.5913.
[58] Romualdo Pastor-Satorras and Alessandro Vespignani, Epidemic spreading in scale-free networks, Physical Review Letters 86 (2001apr), no. 14, 3200–3203, available at 0010317.
, Immunization of complex networks, Physical Review E - Statistical, Nonlinear, and
[59]
Soft Matter Physics 65 (2002feb), no. 3, 036104, available at 0107066.
[60] Mikhail Prokopenko (ed.), Guided Self-Organization: Inception, Emergence, Complexity and
Computation, vol. 9, Springer Berlin Heidelberg, Berlin, Heidelberg, 2014.
[61] Mikhail Prokopenko, Fabio Boschetti, and Alex J. Ryan, An information-theoretic primer
on complexity, self-organization, and emergence, Complexity 15 (2009sep), no. 1, 11–28,
available at arXiv:1103.4838v1.
[62] Rick Quax, Omri Har-Shemesh, and Peter Sloot, Quantifying Synergistic Information Using
Intermediate Stochastic Variables, Entropy 19 (2017feb), no. 2, 85.
[63] Martin Randles, David Lamb, E. Odat, and A. Taleb-Bendiab, Distributed redundancy and
robustness in complex systems, Journal of Computer and System Sciences 77 (2011mar),
no. 2, 293–304.
[64] N. Rashevsky, Life, information theory, and topology, The Bulletin of Mathematical Biophysics 17 (1955sep), no. 3, 229–235.
44
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
[65] Hartley Rogers Jr., Theory of Recursive Functions and Effective Computability, MIT Press,
Cambridge, MA, USA, 1987.
[66] Frank Schweitzer, Giordio Fagiolo, Didier Sornette, Fernando Vega-Redondo, and Douglas R.
White, Economic Networks: What Do We Know and What Do We Need To Know?, Advances
in Complex Systems 12 (2009aug), no. 04n05, 407–422.
[67] Yoav Shoham and Kevin Leyton-Brown, Multiagent Systems: Algorithmic, Game-Theoretic,
and Logical Foundations, Cambridge University Press, New York, NY, USA, 2008.
[68] R Solé and Sergi Valverde, Information theory of complex networks: On evolution and architectural constraints, Complex networks, 2004aug, pp. 189–207.
[69] Ricard V. Solé, Ramon Ferrer-Cancho, Jose M. Montoya, and Sergi Valverde, Selection, tinkering, and emergence in complex networks, Complexity 8 (2002sep), no. 1, 20–33.
[70] Fernando Soler-Toscano, Hector Zenil, Jean-Paul Delahaye, and Nicolas Gauvrit, Calculating
Kolmogorov Complexity from the Output Frequency Distributions of Small Turing Machines,
PLoS ONE 9 (2014may), no. 5, e96223.
[71] Fernanda S H Souza, Geraldo R. Mateus, and Alexandre Salles Da Cunha, Optimization in
designing complex communication networks, Handbook of optimization in complex networks,
2012, pp. 3–37.
[72] Andreia Teixeira, Armando Matos, André Souto, and Luı́s Antunes, Entropy measures vs.
Kolmogorov complexity, Entropy 13 (2011mar), no. 3, 595–611.
[73] Bernard Testa and Lemont B. Kier, Emergence and dissolvence in the self-organisation of
complex systems, Entropy 2 (2000feb), no. 1, 1–25.
[74] A. Wagner, Robustness and evolvability: a paradox resolved, Proceedings of the Royal Society
B: Biological Sciences 275 (2008jan), no. 1630, 91–100.
[75] Sara Imari Walker, Hyunju Kim, and Paul C. W. Davies, The informational architecture
of the cell Subject Areas : Author for correspondence :, Philosophical Transactions A 374
(2016mar), no. 2063, 20150057.
[76] Klaus Wehmuth, Éric Fleury, and Artur Ziviani, On MultiAspect graphs, Theoretical Computer Science 651 (2016), 50–61, available at 1408.0943.
[77] Michael Wibral, Joseph T. Lizier, and Viola Priesemann, Bits from Brains for Biologically
Inspired Computing, Frontiers in Robotics and AI 2 (2015mar), no. March, 1–25, available
at 1412.0291.
[78] Paul L. Williams and Randall D. Beer, Nonnegative Decomposition of Multivariate Information, arXiv preprint (2010apr), available at 1004.2515.
[79] Hector Zenil, Narsis A. Kiani, Francesco Marabita, Yue Deng, Szabolcs Elias, Angelika
Schmidt, Gordon Ball, and Jesper Tegner, An Algorithmic Information Calculus for Causal
Discovery and Reprogramming Systems, bioRxiv (2017).
[80] Hector Zenil, Fernando Soler-Toscano, Kamaludin Dingle, and Ard A. Louis, Correlation of
automorphism group size and topological properties with program-size complexity evaluations
of graphs and complex networks, Physica A: Statistical Mechanics and its Applications 404
(2014jun), 341–358, available at 1306.0322.
[81] Hector Zenil, Fernando Soler-Toscano, Narsis A. Kiani, Santiago Hernández-Orozco, and Antonio Rueda-Toicen, A Decomposition Method for Global Evaluation of Shannon Entropy and
Local Estimations of Algorithmic Complexity (2016sep), 1–48 pp., available at 1609.00110.
12. Appendix
12.1. Lemmas and Theorems.
Lemma 12.1.1 (8.1).
Let NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ) be an algorithmic network.
Let N = |LPBB (N ) |.
Thus, on the average as N grows, we will have that there is a constant C4 such that
Amax ≥ lg(N ) − C4
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
45
Note 12.1.1.1. Let P [ X = a ] be the usual notation for the probability of a random variable X assuming value a. Or P [ statement S ] denote the probability of
a statement S be true. Thus, this theorem is formaly given by the strong law of
large numbers [14] as: there is a constant C4 such that
P
h
i
lim Amax − lg(N ) + C4 ≥ 0 = 1
N →∞
In fact, the main results 12.1.1 and 12.1.2 presented in this article can be translated into such probabilistic form by putting their last statements into the square
brackets like we just did to the above. For the sake of simplifying the notation
and shortening the formulas, we have chosen to state our results without using this
probabilistic form.
Proof.
From AIT, we know that the algorithmic probability of occurring a program p ∈ LU
is
1
2|p|
(23)
Let φN (p) be the frequency that p occurs in a random sample of size N . In the
case, this random sample is the randomly generated population P.
Define a Bernoulli trial on a random variable X that assumes value 1 iff program
p occurs and assumes value 0 iff otherwise. Since this random sample is identically
∞
P
V ar[Xn ]
distributed and/or define a binomial distribution where
< ∞, we will
n2
have from the strong law of large numbers that
(24)
1
P lim φN (p) = |p|
N →∞
2
n=1
=1
Thus, when N is big enough, one expects that p occurs 2N
|p| times within N random tries. That is, since p was arbitrary, the probability distribution in N random
tries tends to match the algorithmic probability distribution on LU when N goes
to ∞.
Let BB(k) be the Busy Beaver value for an arbitrary k ∈ N on machine U.
From AIT [21], we know there are constants CΩ , CBB ≥ 0 and a program pBB(k)
such that for every w ∈ LU
U(pBB(k) ◦ w) = U(pBB(k) ) = BB(k)
(25)
and
46
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
k − CΩ ≤ A(BB(k)) ≤ |pBB(k) | ≤ k + CBB
and
∀x > BB(k) A(x) + CBB ≥ A(BB(k))
Since k was arbitrary, let k = lg(N ) − CBB .
From 24 we will have that, when N is big enough, one should expect that
pBB(lg(N )−CBB ) occurs
N
(26)
≥P
≥
N
=1
2
times within N random tries. That is, from conditional probabilities,
|pBB(lg(N )−CBB ) |
2lg(N )
P
lim φN (pBB(lg(N )−CBB ) )N =
lim φN (pBB(lg(N )−CBB ) )N ≥
N →∞
N →∞
N
2lg(N )
=1 ≥
N
2|pBB(lg(N )−CBB ) |
=1
Let C4 = CΩ + 2 CBB .
From definitions 3.10 and 6.1.1 and step 25 , since any node/program count as
isolated from the network when c = 1, we will have that, when N → ∞:
if
lim φN (pBB(lg(N )−CBB ) )N ≥ 1
N →∞
(27)
then
Amax ≥ A(U(Pprot ◦ pBB(lg(N )−CBB ) )) − CBB = A(U(pBB(lg(N )−CBB ) )) − CBB =
= A(BB(lg(N ) − CBB )) − CBB ≥ lg(N ) − CBB − CΩ − CBB = lg(N ) − C4
Thus, from conditional probabilities, we will have that
h
i
P lim Amax − lg(N ) + C4 ≥ 0 ≥
N →∞
h
i
≥ P lim φN (pBB(lg(N )−CBB ) )N ≥ 1 = 1
N →∞
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
47
Lemma 12.1.2 (8.2). Given a population PBB (N ) defined in 3.11 , where pi ∈
Haltiso (LPBB (N ) , w, c) and Pprot ◦ pi ∈ PBB (N ) and N ∈ N∗ is arbitrary, there is
a constant C1 such that
A(U(piso (pi , c))) ≤ C1 + |pi | + A(w) + A(c)
Note 12.1.2.1. Note that from the Definition 4.1.1 this result is independent of any
topology in which PBB (N ) could be networked.
Proof.
Let N ∈ N∗ be arbitrary. Remember the definition of LBB in 3.10 . And note that
pi is a program in LU .
Then, there is at least one program p′ such that
U(piso (pi , c)) = U(p′ ◦ pi ◦ w ◦ c)
is a well-defined value for every pi ∈ Haltiso (LPBB (N ) , w, c).
Take the smallest such program p′ and let C1 = |p′ | + 5 − |p+1|, where p+1 is the
smallest program that adds 1 to any computation (see Definition 3.5 ).
Then, from AIT and Definition 3.5 we will have that
A(U(piso (pi , c))) = A(U(p′ ◦ pi ◦ w ◦ c)) ≤ C1 + |pi | + A(w) + A(c)
Lemma 12.1.3 (8.3). Given a population PBB (N ) defined in 3.11 , where pi ∈ LU
and Pprot ◦ pi ∈ PBB (N ), from AIT and Gibb’s (or Jensen’s) inequality, we will
have that
1
lim
Ω(w, c) N →∞
X
pi ∈Haltiso (LU (N ),w,c)
|pi |
+ lg(Ω(w, c)) ≤ lim lg(Ω(w, c)N )
N →∞
2|pi |
Proof.
Since LU is self-delimited, from AIT and the Definition 5.2 , we will have that
X
1
= Ω(w, c) < 1
(28)
lim
N →∞
2|pi |
pi ∈Haltiso (LU (N ),w,c)
48
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Hence,
(29)
lim
N →∞
X
pi ∈Haltiso (LU (N ),w,c)
1
= lim
N →∞
Ω(w, c)2|pi |
X
pi ∈Haltiso (LU (N ),w,c)
1
2|pi |+lg(Ω(w,c))
=1
Thus, from step 29 ,
(30)
N →∞
pi ∈Haltiso (LU (N ),w,c)
X
= lim
N →∞
X
lim
pi ∈Haltiso (LU (N ),w,c)
=
|pi |
|p
|
i
2 Ω(w, c)
1
lim
Ω(w, c) N →∞
|pi | + lg(Ω(w, c))
=
2|pi |+lg(Ω(w,c))
X
+ lim
N →∞
X
pi ∈Haltiso (LU (N ),w,c)
pi ∈Haltiso (LU (N ),w,c)
lg(Ω(w, c))
=
2|pi | Ω(w, c)
|pi |
+ lg(Ω(w, c))
2|pi |
Since 29 holds, from Gibb’s (or Jensen’s) inequality [28] [49], we will have that
(31)
X
pi ∈Haltiso (LU (N ),w,c)
|pi | + lg(Ω(w, c))
≤ lg(|Haltiso (LU (N ), w, c)|)
2|pi |+lg(Ω(w,c))
We also have by Definition 5.2 that
(32)
lim lg(|Haltiso (LU (N ), w, c)|) = lim lg(Ω(w, c)N )
N →∞
N →∞
Therefore, from steps 30 , 31 and 32 we will have that
1
lim
Ω(w, c) N →∞
X
pi ∈Haltiso (LU (N ),w,c)
|pi |
+ lg(Ω(w, c)) ≤ lim lg(Ω(w, c)N )
N →∞
2|pi |
Lemma 12.1.4 (8.4). Given a population PBB (N ) defined in 3.11 , where pi ∈ LU
and Pprot ◦ pi ∈ PBB (N ), from Lemma 12.1.3 we will have that
P
lim
N →∞
P pi ∈Haltiso (LPBB (N ) ,w,c)
bj
|pi |
N
|{bj }|
≤ lim Ω(w, c) lg(N )
N →∞
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
49
Note 12.1.4.1. This theorem gives an upper bound on the algorithmic complexity
of the randomly generated part of the elements of the population PBB (N ). And
it will be crucial to prove a lower bound on the emergent algorithmic complexity.
However, the upper bound of Lemma 12.1.4 is overestimated since an algorithmic
probability distribution is far from being uniform, which is the case where Gibb’s
inequality applies on entropies.
Proof.
From the definition of language LBB in 3.10 we have that pi is independent of
any topology, so that
P
|pi |
P pi ∈Haltiso (LPBB (N ) ,w,c)
N
bj
(33)
=
|{bj }|
P
=
|pi |
pi ∈Haltiso (LPBB (N ) ,w,c)
N
And in the limit the ramdonly generated population LPBB (N ) tends to include
all programs in LU , so that
(34)
P
|pi |
pi ∈Haltiso (LPBB (N ) ,w,c)
lim
N
N →∞
= lim
P
|pi |
pi ∈Haltiso (LU (N ),w,c)
N
N →∞
Since one can define algorithmic probabilities in LU , by the Strong Law of Large
Numbers, as in 12.1.1 , we have that in the limit LU (N ) tends to follow the same
distribution, so that 53
1
N →∞ N
(35)
lim
X
1
N →∞ N
X
|pi | = lim
pi ∈Haltiso (LU (N ),w,c)
X
= lim
N →∞
pi ∈Haltiso (LU (N ),w,c)
pi ∈Haltiso (LU (N ),w,c)
N |pi |
=
2|pi |
|pi |
2|pi |
From Lemma 12.1.3 we will have that
(36)
lim
N →∞
X
pi ∈Haltiso (LU (N ),w,c)
|pi |
≤ lim Ω(w, c) lg(Ω(w, c)N ) − Ω(w, c) lg(Ω(w, c))
N →∞
2|pi |
53 Note that, since |Halt (L (N ), w, c)| < N , then
iso
U
P
pi ∈Haltiso (LU ,w,c)
P
|pi |
2|pi |
N→∞ p ∈Halt
(L
(N),w,c)
i
iso
U
Thus, lim
is not a proper Shannon entropy.
1
2|pi |
= Ω(w, c) < 1.
50
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
And
(37)
Ω(w, c) lg(Ω(w, c)N ) − Ω(w, c) lg(Ω(w, c)) =
= Ω(w, c) lg(Ω(w, c)) + Ω(w, c) lg(N ) − Ω(w, c) lg(Ω(w, c)) =
= Ω(w, c) lg(N )
So, from steps 33 , 34 , 35 , 36 and 37
P
N
bj
lim
N →∞
= lim
|{bj }|
P
=
N
X
= lim
pi ∈Haltiso (LU (N ),w,c)
=
|pi |
pi ∈Haltiso (LU (N ),w,c)
N →∞
N →∞
|pi |
P pi ∈Haltiso (LPBB (N ) ,w,c)
|pi |
≤ Ω(w, c) lg(N )
2|pi |
Lemma 12.1.5 (8.5). Given a population PBB (N ) defined in 3.11 , where pi ∈ LU
and Pprot ◦ pi = oi ∈ PBB (N ), there is a constant C0 such that
P
lim
N →∞
A(U(piso (pi ,c)))
P pi ∈Haltiso (LPBB (N ) ,w,c)
N
bj
= C0 (1 − Ω(w, c))
|{bj }|
Note 12.1.5.1. Note that every lemma until here deals with the behavior of A(U(piso (pi , c))),
so that they give tools to obtain an upper bound on the algorithmic complexity of
what each node can do when isolated. Thus, these results are independent of any
topological feature that the algorithmic network NBB (N, f, t, τ, j) might have.
Proof.
As in step 33 , A(U(piso (pi , c))) is independent of any topology, so that
(38)
lim
N →∞
P
A(U(piso (pi ,c)))
P pi ∈Haltiso (LPBB (N ) ,w,c)
bj
N
|{bj }|
= lim
N →∞
P
A(U(piso (pi , c)))
pi ∈Haltiso (LPBB (N ) ,w,c)
N
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
51
From the Definition 5.1.1 , as in step 35 , we will have that
(39)
lim
P
A(U(piso (pi , c)))
pi ∈Haltiso (LPBB (N ) ,w,c)
N →∞
X
= lim
N
N →∞
pi ∈Haltiso (LU (N ),w,c)
A(U(piso (pi , c)))
2|pi |
Now, let A(0) = C0 .
Since pi ∈ Haltiso (LU (N ), w, c) and PBB (N ) is sensitive to oracles 3.11 , by the
definition of piso in 4.1.1 we will have that, for every pi ∈ Haltiso (LU , w, c),
(40)
A(U(piso (pi , c))) = A(0) = C0
Thus, by the definition of Ω(w, c) in 5.2 ,
(41)
X
lim
N →∞
pi ∈Haltiso (LU (N ),w,c)
A(U(piso (pi , c)))
= lim
N →∞
2|pi |
X
= lim C0
N →∞
1
pi ∈Haltiso (LU (N ),w,c)
2|pi |
X
pi ∈Haltiso (LU (N ),w,c)
C0
=
2|pi |
= C0 (1 − Ω(w, c))
And we conclude from steps 38 , 39 and 41 that
P
lim
N →∞
A(U(piso (pi ,c)))
P pi ∈Haltiso (LPBB (N ) ,w,c)
N
bj
= lim
|{bj }|
N →∞
X
pi ∈Haltiso (LU (N ),w,c)
A(U(piso (pi , c)))
=
2|pi |
= C0 (1 − Ω(w, c))
Lemma 12.1.6 (8.6). Let PBB (N ) be a population in an arbitrary algorithmic
network NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ) as defined in 3.13 and 3.11.
Let t0 ≤ t ≤ t′ ≤ t|T (Gt )|−1 .
Let c ∈ CBB be an arbitrary number of cycles where c0 + t′ + 1 ≤ c.
Then, there is a constant C2 such that
P
bj
P pi ∈PBB (N ) A(U(pnet (oi ,c)))
bj
N
|{bj }|
′
≥ (Amax − C2 ) τE(max) (N, f, t, τ )|tt + C2
52
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Proof.
Let Xτmax (N,f,t,τ,j)|t′ denote the set of nodes/programs that belong to fraction
′
t
τmax (N, f, t, τ, j)|tt as defined in 6.2.1. Hence,
′
|Xτmax (N,f,t,τ,j)|t′ | = N τmax (N, f, t, τ, j)|tt
t
Let C2 = min{A(w) | ∃x ∈ LU (U(x) = w)}.
54
′
From the Definition 6.2.1 of τmax (N, f, ti , τ, j)|tt we will have that
P
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c)))
N
bj
|{bj }|
=
P
bj
+
P
b
j
A(U(pnet
(oi ,c)))
P oi ∈Xτmax (N,f,t,τ,j)|tt′
′
τmax (N,f,t,τ,j)|t N
bj
(42)
=
t
|{bj }|
P
bj
A(U(pnet
(oi ,c)))
oi ∈X
D(Gt ,t)
τmax (N,f,t,τ,j)| ′
t
D(G ,t)
τmax (N,f,t,τ,j)|t′ t N
′
τmax (N, f, t, τ, j)|tt
+
D(Gt ,t)
τmax (N, f, t, τ, j)|t′
|{bj }|
Note that in the case the temporal diameter D(Gt , t) is not well-defined one can
D(G ,t)
replace τmax (N, f, t, τ, j)|t′ t with fraction
PBB (N ) \ Xτmax (N,f,t,τ,j)|t′
t
N
in the rest of this proof below without loss of generality.
′
Since fraction τmax (N, f, t, τ, j)|tt is centered on a node from which Amax is
diffused — see Definition 6.2 —, by definitions 3.10 and 6.1.1 we will have that
(43)
oi ∈Xτ
P
b
j
A(U(pnet
(oi , c)))
t′
max (N,f,t,τ,j)|t
′
τmax (N, f, t, τ, j)|tt N
′
′
τmax (N, f, t, τ, j)|tt ≥ Amax τmax (N, f, t, τ, j)|tt
54 Note that depending on the choice of the programming language one may have C ≤ A(0)
2
for example .
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
53
D(Gt ,t)
and, analogously, the following always holds despite on which node fraction τmax (N, f, t, τ, j)|t′
is centered and whenever it starts to diffuse
(44)
b
P
oi ∈X
j
(oi , c)))
A(U(pnet
D(Gt ,t)
τmax (N,f,t,τ,j)| ′
t
D(Gt ,t)
D(Gt ,t)
τmax (N, f, t, τ, j)|t′
τmax (N, f, t, τ, j)|t′
N
′
D(Gt ,t)
≥ C2 τmax (N, f, t, τ, j)|t′
D(Gt ,t)
Thus, since we have that τmax (N, f, t, τ, j)|tt +τmax (N, f, t, τ, j)|t′
by steps 42 , 43 and 44
P
b
j
A(U(pnet
(oi ,c)))
P oi ∈Xτmax (N,f,t,τ,j)|tt′
′
τmax (N,f,t,τ,j)|t N
bj
(45)
t
|{bj }|
P
bj
+
≥
P
P
bj
′
τmax (N, f, t, τ, j)|tt
b
D(Gt ,t)
′
D(Gt ,t)
Amax τmax (N, f, t, τ, j)|tt + C2 τmax (N, f, t, τ, j)|t′
|{bj }|
=
τmax (N, f, t, τ, j)|t′
|{bj }|
P
+
j
A(U(pnet
(oi ,c)))
oi ∈X
D(Gt ,t)
τmax (N,f,t,τ,j)| ′
t
D(G ,t)
τmax (N,f,t,τ,j)|t′ t N
= 1, then
′
(Amax − C2 )τmax (N, f, t, τ, j)|tt + C2
bj
|{bj }|
≥
=
Theorem 12.1.1 (8.1).
Let w ∈ LU be a network input.
Let 0 < N ∈ N.
Let NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ) be well-defined.
Let t0 ≤ t ≤ t′ ≤ t|T (Gt )|−1 .
54
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Let c : N → CBB be a total computable function where c(x) ≥ c0 + t′ + 1.
x 7→ c(x) = y
Then, we will have that:
lim ENBB (N,f,t,τ )
N →∞
net
∆ A(oi , c(x))
iso
≥ lim
N →∞
′
τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N )−
− Ω(w, c(x)) lg(x) − Ω(w, c(x)) lg(lg(x)) − A(w) − C5
Note 12.1.1.1. Thus, note that for bigger enough values of c(x) compared to lg(N )
one can make this lower bound always negative for example. One of the main ideas
behind forthcoming results in this paper is to find optimal conditions where this
lower bound is not only positive, but also goes to ∞.
Note 12.1.1.2. Note that this lower bound on the expected emergent algorithmic
complexity is dependent on the value in the domain of the function c and not on
function c itself, even if it grows fast. In fact, one may want to obtain this theorem
for fixed values of c in which it is not a function but an arbitrary value. And the
same result also hold in this case. The reader is invited to check that a simple
substitution of c(x) for c and of x inside the logarithms for c is enough.
Note 12.1.1.3. The same result also holds if only one possible function bj is defined
for each member of the family G(f, t, τ ). However, the notion of expected emergent
algorithmic complexity τE(max) becomes redundant, as only one function bj will be
taken into account within the sum in order to give the mean. Thus, in this case one
can replace τE(max) with τmax not only in Theorem 12.1.1 but also in 8.1.1 and 8.2 .
Such variation of these theorems becomes useful when one has algorithmic networks
NBB (N, f, t, τ, j) built upon a historical population size growth in which each new
node/program is linked (or not) to the previous existing algorithmic network.
Proof.
We have from AIT that there is Cc ∈ N for every x ∈ N such that
(46)
A(c(x)) ≤ Cc + A(x)
Let C5 = Cc + CL + C1 + C4 .
From Definition 4.4, we have that the expected emergent algorithmic complexity
of a node/program for NBB (N, f, t, τ, j) = (Gt , PBB (N ), bj ), where 0 < j ≤ |{bj }|
is given by
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
55
(47)
P
ENBB (N,f,t,τ )
net
∆ A(oi , c(x))
iso
=
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c(x))))−A(U(piso (pi ,c(x))))
N
bj
|{bj }|
And from definitions 5.1 , 5.1.1 , 5.2 , 6.2.1.1 , 3.5 and Lemmas 12.1.2 , 12.1.5 ,
12.1.4 , 12.1.6 , 12.1.1 and step (46) we will have that
P
lim
(48)
N →∞
b
P oi ∈PBB (N )
j
A(U(pnet
(oi ,c(x))))−A(U(piso (pi ,c(x))))
N
bj
=
|{bj }|
P
= lim
N →∞
P
b
j
−
P
−
P
A(U(piso (pi ,c(x))))
+
N
A(U(piso (pi ,c(x))))
pi ∈Haltiso (LP
,w,c(x))
BB (N )
P
bj
N
−
|{bj }|
P
A(U(piso (pi ,c(x))))
N
|{bj }|
P
N →∞
b
j
−
j
A(U(pnet
(oi ,c(x))))
pi ∈Haltiso (LP
,w,c(x))
BB (N )
≥ lim
b
P oi ∈PBB (N )
bj
+ C0 (1 − Ω(w, c(x)))
≥
P oi ∈PBB (N ) A(U(pnet (oi ,c(x))))
N
bj
|{bj }|
P
C1 +|pi |+A(w)+A(c(x))
pi ∈Haltiso (LP
,w,c(x))
BB (N )
N
|{bj }|
=
N
|{bj }|
b
j
−
P
|{bj }|
P
N →∞
N
bj
pi ∈Haltiso (LP
,w,c(x))
BB (N )
= lim
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c(x))))
−
+ C0 (1 − Ω(w, c(x)))
=
56
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
P
= lim
N →∞
P
bj
j
A(U(pnet
(oi ,c(x))))
N
P
|pi |
N
|{bj }|
+ Ω(w, c(x)) C1 + A(w) + A(c(x)) + C0 1 − Ω(w, c(x))
≥
P
≥ lim
N →∞
−
|{bj }|
pi ∈Haltiso (LP
,w,c(x))
BB (N )
b
j
−
b
P oi ∈PBB (N )
bj
P oi ∈PBB (N ) A(U(pnet (oi ,c(x))))
bj
N
|{bj }|
−
− Ω(w, c(x)) lg(N ) + Ω(w, c(x)) C1 + A(w) + A(c(x)) + C0 1 − Ω(w, c(x)) ≥
′
≥ lim (Amax − C2 ) τE(max) (N, f, t, τ )|tt + C2 −
N →∞
− Ω(w, c(x)) lg(N ) + Ω(w, c(x)) C1 + A(w) + A(c(x)) + C0 1 − Ω(w, c(x)) ≥
′
≥ lim (lg(N ) − C4 − C2 ) τE(max) (N, f, t, τ )|tt + C2 −
N →∞
− Ω(w, c(x)) lg(N ) + Ω(w, c(x)) C1 + A(w) + A(c(x)) + C0 1 − Ω(w, c(x)) =
′
= lim τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N ) − Ω(w, c(x))A(c(x))−
N →∞
′
−(C4 +C2 )τE(max) (N, f, t, τ )|tt −Ω(w, c(x))C1 +Ω(w, c(x))C0 +C2 −C0 −Ω(w, c(x))A(w) ≥
′
≥ lim τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N ) − Ω(w, c(x))A(c(x))−
N →∞
−(C4 + C2 ) − C1 + C0 + C2 − C0 − A(w) =
′
= lim τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N ) − Ω(w, c(x))A(c(x))−
N →∞
−C4 − C1 − A(w) ≥
≥ lim
N →∞
′
τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N )−
−Ω(w, c(x)) lg(x) − Ω(w, c(x))(1 + ǫ) lg(lg(x)) − Ω(w, c(x))CL −
−Cc − C4 − C1 − A(w)
≥ lim
N →∞
′
τE(max) (N, f, t, τ )|tt − Ω(w, c(x)) lg(N )−
−Ω(w, c(x)) lg(x) − Ω(w, c(x)) lg(lg(x)) − C5 − A(w)
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
57
Theorem 12.1.2 (8.2).
Let w ∈ LU be a network input.
Let 0 < N ∈ N.
Let NBB (N, f, tz , τ, j) = (Gt , PBB (N ), bj ) be well-defined.
If there is 0 ≤ z0 ≤ |T (Gt )| − 1 and ǫ, ǫ2 > 0 such that
z0 + f (N, tz0 , τ ) = O
where
tz
−1−ǫ ≤ C =
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
NC
lg(N )
− Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2) − ǫ
Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2)
Then, for every non-decreasing total computable function c : N → CBB where
x 7→ c(x) = y
tz0 ∈ T (Gt ) and c(z0 + f (N, tz0 , τ ) + 2) ≥ c0 + z0 + f (N, tz0 , τ ) + 2 we will have
that there are tcen2 (c) and tcen1 (c) such that
tcen2 (c) = tcen1 (c) ≤ tz0
Proof.
We know from Corollary 8.1.1 that
(49)
lim ENBB (N,f,tz ,τ )
N →∞
≥ lim
N →∞
∆ A(oi , c(z + f (N, tz , τ ) + 2)) ≥
net
iso
t
τE(max) (N, f, tz , τ )|tzz+f (N,tz ,τ ) − Ω(w, c(z + f (N, tz , τ ) + 2)) lg(N )−
− Ω(w, c(z + f (N, tz , τ ) + 2)) lg(z + f (N, tz , τ ) + 2)−
− Ω(w, c(z + f (N, tz , τ ) + 2)) lg(lg(z + f (N, tz , τ ) + 2)) − A(w) − C5
Suppose that there is tz0 ∈ T (Gt ), where 0 ≤ z0 ≤ |T (Gt )| − 1, and ǫ > 0 such
that
(50)
where
z0 + f (N, tz0 , τ ) = O
NC
lg(N )
≤
1
ǫ2
58
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
tz
C=
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2) − ǫ
Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2)
From the Definition 5.2 we have that, for every y ∈ N, if y ≥ c0 +z0 +f (N, tz0 , τ )+
2, then
Ω(w, y) ≤ Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2)
(51)
Thus, since we are assuming c(z0 + f (N, tz0 , τ ) + 2) ≥ c0 + z0 + f (N, tz0 , τ ) + 2,
tz
for fixed values of τE(max) (N, f, tz0 , τ )|tz00
that
tz
τE(max) (N, f, tz0 , τ )|tz00
(52)
and ǫ we will have from step 51
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) − ǫ
Ω(w, c(z0 + f (N, tz0 , τ ) + 2))
tz
≥
+f (N,tz0 ,τ )
+f (N,tz0 ,τ )
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2) − ǫ
Ω(w, c0 + z0 + f (N, tz0 , τ ) + 2)
≥
=C
Let
tz
′
C =
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) − ǫ
Ω(w, c(z0 + f (N, tz0 , τ ) + 2))
Remember that for every x > 0 and t, t′ ∈ T (Gt ) there is ǫ2 such that55
(53)
0 < ǫ2 ≤ Ω(w, x) ≤ 1
and thus, from the Definition 6.2.1.1 , we will also have that
′
(54)
−1 − ǫ ≤
τE(max) (N, f, t, τ )|tt − Ω(w, x) − ǫ
1
≤
Ω(w, x)
ǫ2
Hence, from steps 52 and 54 we will have that
′
z0 + f (N, tz0 , τ ) = O
NC
lg(N )
!
where
55 Remember that one can always have a program that halts for every input, so that it will
also halts for every partial output and, hence, halt on every cycle — see 5.2 .
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
59
tz
′
−1−ǫ ≤ C =
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) − ǫ
Ω(w, c(z0 + f (N, tz0 , τ ) + 2))
≤
′
And since z0 + f (N, tz0 , τ ) is now assymptotically dominated by
definition we will have that there is a constant C6 such that
NC
lg(N ) ,
then by
(55)
lim
N →∞
tz
τE(max) (N, f, tz0 , τ )|tz00
+f (N,tz0 ,τ )
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(N ) −
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(z0 + f (N, tz0 , τ ) + 2)−
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(lg(z0 + f (N, tz0 , τ ) + 2)) − A(w) − C5 ≥
tz +f (N,tz ,τ )
0
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(N )−
≥ lim τE(max) (N, f, tz0 , τ )|tz00
N →∞
′
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(C6
NC
)−
lg(N )
′
NC
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(lg(C6
)) − A(w) − C5 ≥
lg(N )
tz +f (N,tz ,τ )
0
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg(N )−
≥ lim τE(max) (N, f, tz0 , τ )|tz00
N →∞
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) (lg(C6 ) + C ′ lg(N ) − lg(lg(N )))) −
′
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg lg(C6 ) + lg(N C ) − lg(lg(N ))) − A(w) − C5 ≥
≥ lim (ǫ) lg(N ) − Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) (lg(C6 ) − lg(lg(N )))) −
N →∞
′
− Ω(w, c(z0 + f (N, tz0 , τ ) + 2)) lg lg(C6 ) + lg(N C ) − lg(lg(N ))) − A(w) − C5 ≥
′
≥ lim (ǫ) lg(N ) − lg(C6 ) − lg(lg(N )) − lg lg(C6 ) + lg(N C ) − lg(lg(N )) −
N →∞
− A(w) − C5 ≥
′
≥ lim (ǫ) lg(N ) − lg(C6 ) + lg(lg(N )) − lg 2 lg(N C ) − A(w) − C5 ≥
N →∞
≥ lim (ǫ) lg(N ) − lg(C6 ) + lg(lg(N )) − lg(2) − lg(C ′ lg(N )) − A(w) − C5 ≥
N →∞
≥ lim (ǫ) lg(N ) − lg(C6 ) + lg(lg(N )) − lg(2) − lg(C ′ ) − lg(lg(N )) − A(w) − C5 ≥
N →∞
≥ lim (ǫ) lg(N ) − lg(C6 ) − lg(2) − lg(
N →∞
1
) − A(w) − C5 = ∞
ǫ2
Thus, from steps 49 and 55, we will have that
(56)
lim ENBB (N,f,tz0 ,τ )
N →∞
∆ A(oi , c(z0 + f (N, tz0 , τ ) + 2)) = ∞
net
iso
1
ǫ2
60
ALGORITHMIC NETWORKS: CENTRAL TIME TO TRIGGER EXPECTED EMERGENT OPEN-ENDEDNESS
Then, directly from the definitions 7.1 and 7.2 and step 56 , since tz0 satisfy
these definitions, we will have that
tcen2 (c) = tcen1 (c) ≤ tz0
| 7cs.IT
|
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
FOR RANDOM GRAPHS WITH COMPLEX TOPOLOGICAL
STRUCTURES
arXiv:1702.01812v3 [math.ST] 3 Feb 2018
B Y M ICHAEL S CHWEINBERGER AND J ONATHAN S TEWART
Rice University
Statistical inference for exponential-family random graphs with complex
topological structures is challenging. We stress the importance of additional
structure and show that additional structure facilitates statistical inference.
A simple and common form of additional structure, which is observed in
multilevel networks, is a random graph with neighborhoods and local dependence within neighborhoods. We develop the first concentration results
for maximum likelihood and M -estimators of a wide range of canonical and
curved exponential-family random graphs with local dependence. All results
are non-asymptotic and cover random graphs of fixed and finite size, provided
the neighborhoods are small relative to the size of the random graph. We discuss extensions to larger random graphs with more neighborhoods along with
concentration results for subgraph-to-graph estimators. As applications, we
consider canonical and curved exponential-family random graphs, with local
dependence induced by sensible forms of transitivity and parameter vectors
whose dimensions depend on the number of nodes.
1. Introduction. Models of network data have witnessed a surge of interest in
statistics and related areas [e.g., 25]. Such data arise in the study of social networks,
terrorist networks, epidemics, and other areas.
Since the work of Holland and Leinhardt in the 1970s [e.g., 16], it is known that
network data exhibit a wide range of dependencies induced by transitivity and other
complex topological features of networks [e.g., 33]. A large statistical framework
for modeling dependencies induced by complex topological features of networks
is given by discrete exponential-family random graphs [e.g., 13, 48, 14, 45, 19,
27, 33, 30]. Such models are popular among network scientists for the same reason
Ising models are popular among physicists: both classes of models enable scientists
to model a wide range of dependencies of scientific interest [e.g., 33].
Despite the appeal of the discrete exponential-family framework and its relationship to other discrete exponential-family models for dependent random variables [e.g., Ising models and Ising graphical models, 7, 36, 2], statistical inference
for discrete exponential-family random graphs is challenging. One reason is that
∗
Supported by NSF award DMS-1513644.
MSC 2010 subject classifications: curved exponential families, exponential families,
exponential-family random graph models, M -estimators, multilevel networks, social networks.
1
2
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
some exponential-family random graphs are ill-behaved [e.g., the so-called triangle model, 46, 22, 14, 39, 1, 8, 5], but it is worth noting that well-behaved alternatives have been developed, among them curved exponential-family random graphs
[45, 19]. A second reason is that in most applications of exponential-family random
graphs statistical inference is based on a single observation of a dependent random
graph. Establishing desirable properties of estimators, such as consistency of estimators, is non-trivial when no more than one observation of a dependent random
graph is available. While some consistency results have been obtained under independence assumptions [11, 49, 37, 43, 34, 29, 50] and dependence assumptions
[49, 43, 34]—as discussed in Section 7—the existing consistency results do not
cover the models of primary interest to network scientists: that is, canonical and
curved exponential-family random graphs, with dependence induced by sensible
forms of transitivity and other complex topological features of networks [33].
We stress the importance of additional structure and show that additional structure facilitates statistical inference. We consider here a simple and common form
of additional structure called multilevel structure. Network data with multilevel
structure are popular in network science, as the recent monograph of Lazega and
Snijders [31] and a growing number of applications demonstrate [e.g., 47, 52, 32,
44, 17, 18]. A simple form of multilevel structure is given by a partition of a set of
nodes into subsets of nodes, called neighborhoods. In applications, neighborhoods
may correspond to school classes within schools, departments within companies,
and units of armed forces. It is worth noting that in multilevel networks the partition
of the set of nodes is observed and models of multilevel networks attempt to capture dependencies within and between neighborhoods [e.g., 47, 52, 32, 44, 17, 18],
whereas the well-known class of stochastic block models [35] assumes that the
partition is unobserved and that edges are independent conditional on the partition.
Multilevel structure offers outstanding opportunities in terms of statistical inference:
• First, an exponential-family random graph with multilevel structure can be
extended to a larger exponential-family random graph with more neighborhoods [40], whereas many exponential-family random graphs without multilevel structure cannot be extended to larger exponential-family random
graphs [43, 10, 30].
• Second, as long as the dependence is local in the sense that it is restricted
to neighborhoods and the neighborhoods are small, the overall dependence
is weak, which facilitates finite-graph concentration and consistency results
for exponential-family random graphs.
We take advantage of these opportunities and develop the first statistical theory
which shows that statistical inference for many exponential-family random graphs
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
3
with complex topological structures is possible, provided multilevel structure is
available. To do so, we develop the first finite-graph concentration and consistency results for a wide range of canonical and curved exponential-family random
graphs with local dependence. The finite-graph concentration results cover sparse
and dense random graphs with countable support and local dependence and may
be of independent interest. The finite-graph consistency results cover maximum
likelihood and M -estimators under correct and incorrect model specifications. All
results are non-asymptotic and cover random graphs of fixed and finite size, provided the neighborhoods are small relative to the size of the random graph. We
discuss extensions to larger random graphs with more neighborhoods along with
finite-graph consistency results for subgraph-to-graph estimators. As applications,
we consider canonical and curved exponential-family random graphs, with local
dependence induced by sensible forms of transitivity and parameter vectors whose
dimensions depend on the number of nodes.
These finite-graph concentration and consistency results have important implications:
• The most important implication is that statistical inference for transitivity
and other complex topological features of networks is possible, provided
sensible specifications of exponential-family random graphs are used and
multilevel structure is available. To date, it has been widely believed that
statistical inference for transitivity is challenging [e.g., 43, 8], but multilevel
structure facilitates statistical inference for sensible forms of transitivity.
• In practice, researchers should take advantage of multilevel structure when
it is available and collect more network data with multilevel structure when
it is possible to do so.
We note that we focus here on sensible specifications of exponential-family random
graphs, because statistical inference for models that are known to be ill-posed [e.g.,
the so-called triangle model, 46, 22, 14, 39, 1, 8, 5] is problematic.
The paper is structured as follows. Section 2 introduces models. Section 3 describes finite-graph concentration results for dependent random graphs while Sections 4 and 5 describe finite-graph consistency results for maximum likelihood and
M -estimators, respectively. Section 6 discusses extensions to larger random graphs
with more neighborhoods along with finite-graph consistency results for subgraphto-graph estimators. A comparison with existing asymptotic consistency results can
be found in Section 7. Section 8 presents simulation results.
2. Discrete exponential-family random graphs with multilevel structure.
We introduce discrete exponential-family random graphs with multilevel structure.
A simple and common form of multilevel structure is a partition of a set of
nodes into K ≥ 2 non-empty subsets of nodes A1 , . . . , AK , called neighborhoods.
4
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
We note that in multilevel networks the partition of the set of nodes is observed
[e.g., 47, 52, 32, 44, 17, 18] and that some neighborhoods may be larger than others. We consider undirected random graphs with countable sample spaces, which
covers binary and non-binary network data, including network count data. Extensions to directed random graphs are straightforward. Let X = (Xk )K
k=1 and Y =
K
(Yk,l )k<l be sequences of within- and between-neighborhood edge variables based
on a sequence of neighborhoods A1 , . . . , AK , where Xk = (Xi,j )i∈Ak < j∈Ak and
Yk,l = (Yi,j )i∈Ak , j∈Al (k < l) correspond to within- and between-neighborhood
edge variables Xi,j ∈ Xi,j and Yi,j ∈ Yi,j , taking on values in countable sets
Xi,j and Yi,j , respectively. We exclude self-edges, assuming that
Q Xi,i = 0 holds
with probability 1 (i ∈ Ak , k = 1, . . . , K), and write Xk = i∈Ak < j∈Ak Xi,j ,
Q
QK Q
Q
X= K
k<l
i∈Ak , j∈Al Yi,j .
k=1
i∈Ak < j∈Ak Xi,j , and Y =
We assume that within-neighborhood edges X are independent of betweenneighborhood edges Y , i.e.,
P(X ∈ X, Y ∈ Y) = P(X ∈ X) P(Y ∈ Y) for all X × Y ⊆ X × Y,
where P denotes a probability distribution with support X × Y. We do not assume
that edges are independent, neither within neighborhoods nor between neighborhoods.
While in principle both within-neighborhood edge variables X and betweenneighborhood edge variables Y may be of interest, we focus on withinneighborhood edge variables, which are of primary interest in applications [e.g.,
31, 47, 52, 32, 44, 17, 18]. We therefore restrict attention to the probability law
of X and do not make assumptions about the probability law of Y . We assume
that the parameter vectors of the probability laws of X and Y are variationindependent, i.e., the parameter space is a product space, so that statistical inference
concerning the parameter vector of the probability law of X can be based on X
without requiring knowledge of Y .
The within-neighborhood edge variables X are assumed to be governed by
an exponential family with countable support X, having densities with respect to
counting measure of the form
pη (x) = exp (hη, s(x)i − ψ(η)) ν(x)
(1)
!
K
X
= exp
hηk , sk (xk )i − ψ(η) ν1 (x1 ) · · · νK (xK ),
k=1
where
ψ(η) = log
X
x1 ∈X1
···
X
xK ∈XK
exp
K
X
k=1
!
hηk , sk (xk )i
ν1 (x1 ) · · · νK (xK ).
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
5
P
Here, hη, s(x)i = K
k=1 hηk , sk (xk )i is the inner product of a natural parameter vector η ∈ Rdim(η) and a sufficient statistic vector s : X 7→ Rdim(η) while
ηk ∈ Rdim(ηk ) and sk : Xk 7→ Rdim(ηk ) denote the natural parameter vector and
sufficient statistic vector of neighborhood Ak , respectively (k = 1, . . . , K). The
functions ν : X 7→ R+ ∪ {0} and νk : Xk 7→ R+ ∪ {0} (k = 1, . . . , K) along
with the sample space X determine the reference measure of the exponential family. A careful discussion of the choice of reference measure for both binary and
non-binary, network count data can be found in Krivitsky [27]. We note that an exponential family of the form (1) induces local dependence in the sense that edges
may depend on other edges in the same neighborhood, but do not depend on edges
in other neighborhoods [40].
We consider here a wide range of exponential families of the form (1). Examples of exponential families of the form (1) with local dependence within neighborhoods induced by sensible forms of transitivity can be found in Sections 4.1 and
4.2. We assume, without loss of generality, that exponential families of the form (1)
are minimal in the sense that neither the sufficient statistic vector s(X) is contained
in a hyperplane in Rdim(η) with probability 1 nor the natural parameter vector η is
contained in a hyperplane in Rdim(η) . It is well-known that all non-minimal exponential families can be reduced to minimal exponential families [4, Theorem 1.9,
p. 13]. We consider both full and non-full exponential families of the form (1).
An exponential family {Pη , η ∈ Ξ} is full if Ξ = N and non-full if Ξ ⊂ N,
where N = {η ∈ Rdim(η) : ψ(η) < ∞} is the natural parameter space, i.e., the
largest set of possible values the natural parameter vector η can take on. While
full exponential families may be more convenient on mathematical grounds, nonfull exponential families—the most important example being curved exponential
families [e.g., 12, 23]—offer parsimonious parameterizations of high-dimensional
exponential families with dim(ηk ) → ∞ (k = 1, . . . , K) as K → ∞ and have
turned out to be useful in practice [45, 19]. A simple approach to generating nonfull exponential families is to assume that η : Θ 7→ int(N) is a known function
of a parameter vector θ ∈ Θ, where Θ ⊆ {θ ∈ Rdim(θ) : ψ(η(θ)) < ∞}. The
map η : Θ 7→ int(N) may be affine or non-affine. If η(θ) is an affine function of
θ of the form η(θ) = A θ + b, we call the exponential family canonical, because
it can be reduced to a canonical exponential family with natural parameter vector
η(θ) = θ by absorbing the matrix A ∈ Rdim(η)×dim(θ) and the vector b ∈ Rdim(η)
into the sufficient statistic vector and reference measure, respectively. If η(θ) is a
non-affine function of θ, we call the exponential family curved.
Throughout, we assume that the neighborhoods are of the same order of magnitude and that the neighborhood-dependent natural parameters ηk,i (θ) are of the
form ηk,i (θ) = ηi (θ) (i = 1, . . . , dim(ηk ), k = 1, . . . , K). We define neighborhoods of the same order of magnitude as follows.
6
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
Definition. Neighborhoods of the same order of magnitude. A sequence of neighborhoods A1 , . . . , AK is of the same order of magnitude if there exist A1 > 0 and
A2 > 0 and a non-decreasing function ρ : N 7→ N such that A1 ρ(K) ≤ |Ak | ≤
A2 ρ(K) (k = 1, . . . , K, K = 1, 2, . . . ), where N = {0, 1, 2, . . . }.
In other words, the sizes of neighborhoods may not be identical, but are assumed
to be similar. We note that when the neighborhoods are not of the same order of
magnitude, the natural parameters of neighborhoods may have to depend on the order of magnitude of neighborhoods [e.g., 28, 29, 6], because there are good reasons
to believe that small and large within-neighborhood subgraphs are not governed by
the same natural parameters [43, 10, 30]. However, despite recent progress on sizedependent parameterizations of simple models [e.g., 28, 29, 6], more research on
size-dependent parameterizations of more complex models is needed, which is an
important topic in its own right but is beyond the scope of our paper.
The assumption ηk,i (θ) = ηi (θ) (i = 1, . . . , dim(ηk ), k = 1, . . . , K) implies
that the exponential families considered here can be reduced to exponential families
with natural parameter vectors of the form
η(θ) = (η1 (θ), . . . , ηd∞ (θ))
and sufficient statistic vectors of the form
s(x) = (s1 (x), . . . , sd∞ (x)) ,
P
where si (x) = K
k=1 sk,i (xk ) (i = 1, . . . , d∞ ) and d∞ = max1≤k≤K dim(ηk ).
The dimensions dim(ηk ) of the neighborhood-dependent natural parameter vectors
ηk (θ) may be functions of the sizes |Ak | of neighborhoods Ak (k = 1, . . . , K),
which implies that d∞ = max1≤k≤K dim(ηk ) may be a function of kAk∞ =
max1≤k≤K |Ak |. An example is given by curved exponential families with geometrically weighted model terms, as described in Section 4.2. In addition, in both
canonical and curved exponential families, we allow the dimension of parameter
vector θ to depend on kAk∞ .
Notation. To prepare the ground for the finite-graph concentration and consistency results in Sections 3, 4, 5, and 6, we introduce mean-value parameterizations of exponential families along with additional notation. Mean-value parameterizations facilitate finite-graph concentration and consistency results, because
concentration inequalities [3] bound probabilities of deviations from means and
the mean-value parameters of an exponential family are the means of the sufficient
statistics, defined by µ(η(θ)) = Eη(θ) s(X) ∈ rint(M) [4, p. 2 and p. 75]. Here,
Eη(θ) s(X) denotes the expectation of s(X) with respect to exponential-family
distributions Pη(θ) having densities of the form (1), M denotes the convex hull of
the set {s(x) : x ∈ X}, and rint(M) denotes the relative interior of M. We denote
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
7
the data-generating parameter vector by θ ? ∈ int(Θ) and write P ≡ Pη(θ? ) and
E ≡ Eη(θ? ) , where int(Θ) denotes the interior of Θ. An open ball in Rdim(θ) with
center θ ? ∈ Rdim(θ) and radius > 0 is denoted by B(θ ? , ). We denote by k.k1 ,
k.k2 , and k.k∞ the `1 -, `2 -, and `∞ -norm of vectors, respectively. Throughout,
d : X × X 7→ R+ ∪ {0} denotes the Hamming metric defined by
d(x1 , x2 ) =
K
X
X
1(x1,i,j 6= x2,i,j ), (x1 , x2 ) ∈ X × X,
k=1 i∈Ak < j∈Ak
where 1(x1,i,j 6= x2,i,j ) is an indicator function, which is 1 if x1,i,j 6= x2,i,j and is
0 otherwise.
3. Finite-graph concentration results for dependent random graphs. To
obtain finite-graph concentration and consistency results for sparse and dense
exponential-family random graphs with countable support and local dependence,
we need finite-graph concentration results for dependent random graphs.
Finite-graph concentration results for dependent random graphs are non-trivial
for at least three reasons. First, exponential families of the form (1) may induce
strong dependence within neighborhoods and the sizes of neighborhoods need not
be identical. Second, exponential families of the form (1) can induce a wide range
of dependencies within neighborhoods. Therefore, we need general-purpose concentration inequalities that cover a wide range of dependencies. Third, we consider
exponential families with countable support and some distributions with countable
support do not have nice Subgaussian tails.
The following general-purpose concentration inequality addresses the three challenges discussed above. It shows that the dependence induced by exponential families of the form (1) may be strong within neighborhoods but is sufficiently weak
overall to obtain concentration results as long as the neighborhoods are not too
large.
Proposition 1. Consider an exponential family with countable support X and
local dependence. Let f : X 7→ R be Lipschitz with respect to the Hamming metric
d : X × X 7→ R+ ∪ {0} with Lipschitz coefficient kf kLip > 0 and assume that
E f (X) exists. Then there exists C > 0 such that, for all t > 0,
!
t2
P(|f (X) − E f (X)| ≥ t) ≤ 2 exp − PK |A |
.
C k=1 2k kAk4∞ kf k2Lip
It is worth noting that there is a large body of literature in probability theory and
related areas on concentration results for random graphs [e.g., 9], but almost all of
them assume that edge variables are independent and almost surely bounded, i.e.,
8
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
Subgaussian. In contrast, we consider dependent edge variables with countable
sample spaces having distributions with either Subgaussian or non-Subgaussian
tails. Examples of dependent edge variables having distributions with
non-Subgaussian tails (e.g., Poisson-referenced exponential-family random graphs
with transitivity) can be found in Krivitsky [27]. The proof of Proposition 1 relies on concentration inequalities for dependent random variables [e.g., 26, 3] and
bounds mixing coefficients, which quantify the strength of dependence induced by
exponential families with local dependence, in terms of kAk∞ .
Proposition 1 paves the way for finite-graph concentration and consistency results for sparse and dense exponential-family random graphs with countable support and local dependence. We present here an example of special interest for
maximum likelihood estimation and note that additional concentration results for
\? )) = s(X) and note that
M -estimation can be found in Section 5. Let µ(η(θ
\? )) = s(X) is an estimator of the data-generating mean-value parameter
µ(η(θ
vector µ(η(θ ? )) = Eη(θ? ) s(X) ∈ rint(M). The following concentration result
\? )) is close to µ(η(θ ? )) ∈ rint(M) with high probability as long
shows that µ(η(θ
as the neighborhoods and the dimension d∞ of µ(η(θ ? )) are small relative to the
number of neighborhoods K.
Proposition 2. Consider a full or non-full, canonical or curved exponential family with countable support X and local dependence. Assume that there exists A > 0
such that, for all (x1 , x2 ) ∈ X × X,
ks(x1 ) − s(x2 )k∞ ≤ A d(x1 , x2 ) kAk∞ .
(2)
Then there exists C > 0 such that, for all deviations of the form t = δ
with δ > 0 and 0 ≤ α ≤ 1,
\? )) − µ(η(θ ? ))k∞ ≥ t
P kµ(η(θ
\? )) − µ(η(θ ? ))k2 ≥ t
P kµ(η(θ
≤ 2 exp −
δ2 C K
4 (2−α)
kAk∞
≤ 2 exp −
PK
|Ak | α
2
!
+ log d∞
δ2 C K
4 (2−α)
d∞ kAk∞
k=1
!
+ log d∞
,
where d∞ may be a function of kAk∞ .
The smoothness condition (2) of Proposition 2 is satisfied as long as changing a within-neighborhood edge cannot change the within-neighborhood sufficient
statistics by more than a multiple of kAk∞ . It is verified in Corollaries 1 and 2
in Section 4. The finite-graph concentration results for estimators of mean-value
parameters in Proposition 2 cover both sparse and dense random graphs. We note
that the conventional definition of the terms sparse and dense random graphs is
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
9
based on the scaling of the expected number of edges of Bernoulli random graphs
with independent edges, i.e., the expectation of the sufficient statistic of Bernoulli
random graphs. We use these terms here to refer to the scaling of the expectations
of all sufficient statistics of exponential-family random graphs. The power α of
|Ak | α
controls the level of sparsity: if α = 1, the within-neighborhood subgraphs
2
are said to be dense in the sense that the expectations of within-neighborhood suffi
cient statistics are non-negligible fractions of the number of edge variables |A2k | in
neighborhood Ak (k = 1, . . . , K), otherwise the within-neighborhood subgraphs
are said to be sparse (0 ≤ α < 1). Between-neighborhood subgraphs can be sparse
or dense.
Remark 1. Sharpness. The concentration results discussed above are not sharp.
We are not interested in sharp bounds in special cases, because the main appeal of
the exponential-family framework is that it can capture a wide range of dependencies. Therefore, we are interested in concentration results that cover a wide range
of exponential families with local dependence.
4. Finite-graph concentration results for maximum likelihood estimators.
In exponential families, maximum likelihood estimation of natural parameters is
related to maximum likelihood estimation of mean-value parameters and, therefore, the finite-graph concentration results for estimators of mean-value parameters
in Proposition 2 pave the way for finite-graph concentration results for estimators
of natural parameters and functions of natural parameters. All of the following results are non-asymptotic and cover random graphs of fixed and finite size, provided
the neighborhoods are small relative to the size of the random graph. Throughout,
a random graph of fixed and finite size is taken to be a random graph with a finite number of neighborhoods and a finite number of nodes. Extensions to larger
random graphs with more neighborhoods are discussed in Section 6.
To cover a wide range of full and non-full, canonical and curved exponential
families under weak assumptions on the map η : Θ 7→ int(N), we consider estimating functions of the form
(3)
\? ))) = kµ(η(θ
\? )) − µ(η(θ))k2 , θ ∈ Θ,
g(θ; µ(η(θ
which are approximations of
g(θ; µ(η(θ ? ))) = kµ(η(θ ? )) − µ(η(θ))k2 , θ ∈ Θ.
\? ))) is an approximation of g(θ; µ(η(θ ? ))) follows from
The fact that g(θ; µ(η(θ
the triangle inequality, which shows that
\? ))) − g(θ; µ(η(θ ? )))
g(θ; µ(η(θ
\? )) − µ(η(θ ? ))k2 , θ ∈ Θ,
≤ kµ(η(θ
10
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
\? )) − µ(η(θ ? ))k2 is
along with the fact that, under suitable conditions, kµ(η(θ
small with high probability by Proposition 2.
Estimating function (3) has at least three advantages. First, estimating function (3) admits finite-graph concentration results under weak assumptions on the
map η : Θ 7→ int(N). Indeed, the main results of Section 4 assume that the
map η : Θ 7→ int(N) is one-to-one and satisfies a weak identifiability assumption, but the map η : Θ 7→ int(N) need not be differentiable. The weakness of
these assumptions implies that the main results of Section 4 cover a wide range
of full and non-full exponential families—including, but not limited to curved exponential families—and it is possible to verify these assumptions in some of the
most challenging applications, as demonstrated by Corollaries 1 and 2. Second,
estimating function (3) is natural, because maximum likelihood estimation of the
data-generating natural parameter vector η ? of an exponential family with natural parameter vector η and sufficient statistic vector s(x) can be based on the
? ) − µ(η) provided
\
gradient of the loglikelihood function ∇η log pη (x) = µ(η
? ) = s(x) [4, Lemma 5.3, p. 146]. Therefore, maxi\
η ∈ int(N), where µ(η
mum likelihood estimation of η ? can be based on estimating functions of the form
? ) − µ(η)k . By the parameterization-invariance of maximum likelihood es\
kµ(η
2
timators, maximum likelihood estimation of functions of η ? , such as θ ? , can be
\? )) − µ(η(θ))k2 provided the map η : Θ 7→ int(N) is one-tobased on kµ(η(θ
one. We note that estimating function (3) is chosen for mathematical convenience,
facilitating finite-graph concentration results for maximum likelihood estimators
of a wide range of full and non-full, curved exponential families under weak assumptions on the map η : Θ 7→ int(N), and is not chosen for computational
convenience. Last, but not least, the simple form of estimating function (3) helps
determine when minimizers of (3) exist and are unique, and how the minimizers
are related to each other when there is more than one minimizer. These advantages
are most useful in non-full exponential families, in particular curved exponential
families.
In the following, we assume that the map η : Θ 7→ int(N) is one-to-one. A
natural class of estimators is hence given by
\? ))) = inf g(θ̇; µ(η(θ
\? ))) .
θb =
θ ∈ Θ : g(θ; µ(η(θ
θ̇∈Θ
If the set θb is non-empty, it may contain one element (e.g., in full exponential families) or more than one element (e.g., in non-full exponential families). If the set
θb contains more than one element, then all elements of the set θb map to mean\? ))
b that have the same `2 -distance from µ(η(θ
value parameter vectors µ(η(θ))
b
by construction of estimating function (3). In addition, if the set θ is non-empty,
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
11
\? ))) are approximations of the minimizer of
then all minimizers θb of g(θ; µ(η(θ
?
g(θ; µ(η(θ ))) under suitable conditions. The minimizer of g(θ; µ(η(θ ? ))) is
unique and is given by the data-generating parameter vector θ ? provided
µ(η(θ ? ))) ∈ rint(M):
?
?
?
θ =
θ ∈ Θ : g(θ; µ(η(θ ))) = inf g(θ̇; µ(η(θ ))) .
θ̇∈Θ
The data-generating parameter vector θ ? is the unique minimizer of g(θ; µ(η(θ ? )))
= kµ(η(θ ? )) − µ(η(θ))k2 , because θ ? ∈ int(Θ) and the map η : Θ 7→ int(N) is
one-to-one by assumption, while the map µ : int(N) 7→ rint(M) is one-to-one by
classic exponential-family theory [4, Theorem 3.6, p. 74]. Therefore, kµ(η(θ ? )) −
µ(η(θ))k2 = 0 holds if and only if θ = θ ? .
In the remainder of the section, we show that the estimator θb is close to the datagenerating parameter vector θ ? with high probability as long as the neighborhoods
are small relative to the size of the random graph. We discuss finite-graph concentration results for full and non-full, curved exponential families in Sections 4.1 and
4.2, respectively. More general finite-graph concentration results for M -estimators
can be found in Section 5. Finite-graph concentration results for subgraph-to-graph
estimators are discussed in Section 6.
4.1. Finite-graph concentration: full exponential families. The main concentration result for full exponential families is stated in Theorem 1. It covers some of
the most interesting exponential-family random graphs used in practice, including
exponential-family random graphs with sensible forms of transitivity, as Corollary
1 demonstrates. We note that the term full exponential-family random graph in
Theorem 1 refers to the fact that the exponential family is full, i.e., the parameter
space of the exponential family is as large as possible, and does not refer to a full
graph with all possible edges.
Theorem 1. Consider a full exponential-family random graph of fixed and finite
size with countable support X and local dependence. Let η(θ) = θ ∈ Θ, where
Θ = {θ ∈ Rdim(θ) : ψ(η(θ)) < ∞}.
Assume that θ ? ∈ int(Θ) and that, for all > 0 small enough so that B(θ ? , ) ⊆
int(Θ), there exist δ() > 0 and A > 0 such that, for all θ ∈ Θ \ B(θ ? , ),
K
X
|Ak | α
?
(4) kµ(η(θ )) − µ(η(θ))k2 ≥ δ()
for some 0 ≤ α ≤ 1
2
k=1
and that, for all (x1 , x2 ) ∈ X × X,
(5)
ks(x1 ) − s(x2 )k∞ ≤ A d(x1 , x2 ) kAk∞ .
12
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
Then, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exist κ() > 0
and C > 0 such that θb exists, is unique, and
!
2 C K
κ()
P θb ∈ B(θ ? , ) ⊆ int(Θ) ≥ 1 − 2 exp −
+ log dim(θ) ,
4 (2−α)
dim(θ) kAk∞
where dim(θ) may be a function of kAk∞ and K ≥ 2.
Theorem 1 is non-asymptotic in the sense that it covers all finite random graphs
with K ≥ 2 neighborhoods. It excludes random graphs with K = 1 neighborhood,
because random graphs with K = 1 neighborhood are equivalent to random graphs
with unrestricted dependence, which can possess undesirable properties [46, 22,
14, 39, 1, 8]. An important implication of Theorem 1 is that θb is close to the datagenerating parameter vector θ ? with high probability when the neighborhoods and
the dimension of θ ? are small relative to the number of neighborhoods K.
Conditions (4) and (5) of Theorem 1 are verified in Corollary 1. Condition (4) is
an identifiability assumption and covers both sparse (0 ≤ α < 1) and dense (α =
1) within-neighborhood subgraphs, as explained in Section 3. Theorem 1 shows
that sparsity comes at a cost, because the probability of event θb 6∈ B(θ ? , ) decays
slower when the within-neighborhood subgraphs are sparse rather than dense. The
fact that sparsity weakens concentration results is well-known in the literature on
concentration results for Bernoulli random graphs with independent edges [e.g.,
21, 24]. Condition (5) is satisfied as long as changing a within-neighborhood edge
cannot change the within-neighborhood sufficient statistics by more than a multiple of kAk∞ . Both conditions are satisfied by exponential-family random graphs
capturing sensible forms of transitivity and other complex topological features of
networks.
Example: full, canonical exponential family. As an example, we consider
PK
|Ak |
exponential-family random graphs with support X = {0, 1} k=1 ( 2 ) and local dependence induced by sensible forms of transitivity. Transitivity refers to the
tendency of nodes i and j with shared partners (i.e., xi,h = xj,h = 1 for at least
one node h 6= i, j) to be connected (i.e., xi,j = 1) [33]. A sensible form of transitivity is induced by exponential-family random graphs with within-neighborhood
edge and transitive edge terms [20], with neighborhood-dependent natural parameters ηk,1 (θ) = θ1 and ηk,2 (θ) = θ2 and sufficient statistics sk,1 (xk ) and sk,2 (xk )
(k = 1, . . . , K) given by
X
sk,1 (xk ) =
xi,j
i∈Ak < j∈Ak
sk,2 (xk ) =
X
i∈Ak < j∈Ak
xi,j
max
h∈Ak , h6=i,j
xi,h xj,h .
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
13
Here, as elsewhere in the paper, we assume that the neighborhoods are of the same
order of magnitude, as defined in Section 2.
Corollary 1. Consider an exponential-family random graph of fixed and finite
size with within-neighborhood edge and transitive edge terms. Let θ ? ∈ int(Θ),
where Θ = {θ ∈ R2 : ψ(η(θ)) < ∞} = R2 . Then conditions (4) and (5) of
Theorem 1 are satisfied and hence, for all > 0, there exist κ() > 0 and C > 0
such that θb exists, is unique, and
κ()2 C K
?
b
P θ ∈ B(θ , ) ⊆ int(Θ)
≥ 1 − 4 exp −
,
kAk4∞
provided |Ak | ≥ 3 (k = 1, . . . , K) and K ≥ 2.
Corollary 1 is the first finite-graph concentration result which shows that consistent estimation of exponential-family random graphs with sensible forms of transitivity is possible. To date, it has been widely believed that statistical inference for
transitivity is challenging [e.g., 43, 8], but the additional structure in the form of
multilevel structure facilitates statistical inference for sensible forms of transitivity.
Remark 2. Sensible forms of transitivity. We focus here on sensible forms of
transitivity based on the number of within-neighborhood transitive edges rather
than triangles. The reason is that triangle terms tend to induce near-degenerate
distributions in large graphs and, therefore, large within-neighborhood subgraphs
[46, 22, 14, 39, 1, 8]. We note that these issues are serious issues in large random
graphs with unrestricted dependence, but are less serious issues in large random
graphs with local dependence as long as the neighborhoods are small relative to
the size of the random graph. Nonetheless, it is prudent to focus on sensible forms
of transitivity. Transitive edge terms are sensible alternatives to triangle terms [20],
because transitive edge terms count the first triangle in which a pair of nodes is
involved but discard all additional triangles, reducing the incentive of pairs of nodes
to be involved in an excessive number of triangles.
4.2. Finite-graph concentration: non-full, curved exponential families. In situations where the dimensions of the natural parameter vectors ηk (θ) of neighborhoods Ak depend on the number of nodes in Ak (k = 1, . . . , K), it is sometimes
convenient to use non-full, curved exponential families rather than full exponential families. The reason is that curved exponential families offer more parsimonious parameterizations, because the neighborhood-dependent natural parameter
vectors ηk (θ) may be non-affine functions of a lower-dimensional parameter vector θ (k = 1, . . . , K). Such curved exponential families have turned out to be
useful in practice [e.g., 45, 19].
The main concentration result for non-full, curved exponential families is stated
14
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
in Theorem 2. It covers the popular class of curved exponential-family random
graphs with geometrically weighted model terms, as Corollary 2 demonstrates.
Theorem 2. Consider a non-full, curved exponential-family random graph of
fixed and finite size with countable support X and local dependence. Let
Θ ⊆ {θ ∈ Rdim(θ) : ψ(η(θ)) < ∞}, Θ open.
Assume that θ ? ∈ int(Θ). Let η : Θ 7→ int(N) be one-to-one and assume that, for
all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exists γ() > 0 such that,
for all θ ∈ Θ \ B(θ ? , ), we have η(θ) ∈ N \ B(η(θ ? ), γ()). In addition, assume
that there exist δ() > 0 and A > 0 such that, for all η(θ) ∈ N \ B(η(θ ? ), γ()),
(6)
kµ(η(θ ? ))
− µ(η(θ))k2 ≥ δ()
K
X
|Ak | α
k=1
2
for some 0 ≤ α ≤ 1
and, for all (x1 , x2 ) ∈ X × X,
(7)
ks(x1 ) − s(x2 )k∞ ≤ A d(x1 , x2 ) kAk∞ .
Then, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exist κ() > 0
and C > 0 such that
!
2CK
κ()
P θb ∈ B(θ ? , ) ⊆ int(Θ) ≥ 1 − 2 exp −
+ log d∞ ,
4 (2−α)
d∞ kAk∞
where d∞ may be a function of kAk∞ and K ≥ 2.
Theorem 2 does not make strong assumptions about the map η : Θ 7→ int(N):
It assumes that the map η : Θ 7→ int(N) is one-to-one and that η(θ) is far from
η(θ ? ) when θ is far from θ ? , but the map η : Θ 7→ int(N) need not be differentiable. As a result, Theorem 2 covers a wide range of non-full exponential families,
including curved exponential families.
It is worth noting that in non-full, curved exponential families there may be more
than one minimizer of estimating function (3). A pleasant feature of estimating
function (3) is that, with high probability, the minimizers of (3) do not give rise to
global minima that are separated by large distances, under the assumptions made.
The reason is that, if the set θb contains more than one element, then all elements of
b whose `2 -distance from
the set θb map to mean-value parameter vectors µ(η(θ))
\? )) is identical and whose `2 -distance from µ(η(θ ? )) is bounded above by
µ(η(θ
\? ))k2 + kµ(η(θ
\? )) − µ(η(θ ? ))k2
b − µ(η(θ ? ))k2 ≤ kµ(η(θ))
b − µ(η(θ
kµ(η(θ))
\? )) − µ(η(θ ? ))k2 ,
≤ 2 kµ(η(θ
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
15
\? )) is close to
as shown in the proof of Theorem 2. By Proposition 2, µ(η(θ
?
µ(η(θ )) with high probability provided the neighborhoods and the dimension
d∞ of µ(η(θ ? )) are small relative to the size of the random graph. Therefore, all
b are close to µ(η(θ ? )) with high probability and hence,
elements of the set µ(η(θ))
by the identifiability conditions of Theorem 2, all elements of the set θb are close to
θ ? with high probability, under the stated assumptions.
Example: non-full, curved exponential family. As an example, we consider one
of the most popular curved exponential-family random graphs with support X =
PK
|Ak |
{0, 1} k=1 ( 2 ) : curved exponential-family random graphs with within-neighborhood edge and geometrically weighted edgewise shared partner terms [45, 19].
These models are based on the number of within-neighborhood edges and pairs of
nodes with i edgewise shared partners:
X
sk,1 (xk )
=
xa,b
a∈Ak < b∈Ak
sk,i+1 (xk ) =
X
xa,b fa,b,i (xk ),
i = 1, . . . , |Ak | − 2,
a∈Ak < b∈Ak
P
where fa,b,i (xk ) = 1( c∈Ak , c6=a,b xa,c xb,c = i) is an indicator function, which
is 1 if nodes a and b have i shared partners in neighborhood Ak and is 0 otherwise (k = 1, . . . , K). The natural parameters of the within-neighborhood edge and
edgewise shared partner terms are of the form
ηk,1 (θ)
= θ1
h
i
ηk,i+1 (θ) = exp(ϑ) 1 − (1 − exp(−ϑ))i ,
i = 1, . . . , |Ak | − 2,
where ϑ > 0 controls the rate of decay of the geometric sequence (1 − exp(−ϑ))i ,
i = 1, 2, . . . (k = 1, . . . , K). For convenience, we consider here the parameterization θ2 = exp(−ϑ) ∈ (0, 1), so that θ = (θ1 , θ2 ) ∈ R × (0, 1). Such
model terms are called geometrically weighted terms, because the natural parameters ηk,i+1 (θ) are based on the geometric sequence (1 − exp(−ϑ))i , i = 1, 2, . . .
The dimensions dim(ηk ) = |Ak | − 1 of the neighborhood-dependent natural parameter vectors ηk (θ) are functions of |Ak | (k = 1, . . . , K) and hence the dimension d∞ = max1≤k≤K dim(ηk ) = kAk∞ − 1 of natural parameter vector η(θ) is
a function of kAk∞ . The resulting model captures transitive closure in neighborhoods and, in the special case ϑ = 0, reduces to the edge and transitive edge model
described in Section 4.1. A full-fledged discussion of these models is beyond the
scope of our paper. We therefore refer the interested reader to Snijders et al. [45]
and Hunter and Handcock [19] and focus here on finite-graph concentration results.
16
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
The following finite-graph concentration result shows that the estimator θb is
close to the data-generating parameter vector θ ? with high probability, provided
kAk∞ is small relative to the number of neighborhoods K; note that θ = (θ1 , θ2 ) ∈
R × (0, 1) since θ2 = exp(−ϑ) and ϑ > 0.
Corollary 2. Consider a curved exponential-family random graph of fixed and
finite size with within-neighborhood edge and geometrically weighted edgewise
shared partner terms. Let Θ = R × (0, 1) and assume that θ ? ∈ int(Θ). Then all
conditions of Theorem 2 are satisfied and hence, for all > 0 small enough so that
B(θ ? , ) ⊆ int(Θ), there exist κ() > 0 and C > 0 such that
κ()2 C K
?
b
P θ ∈ B(θ , ) ⊆ int(Θ)
≥ 1 − 2 exp −
+ log kAk∞ ,
kAk6∞
provided |Ak | ≥ 4 (k = 1, . . . , K) and K ≥ 2.
Corollary 2 is the first concentration result for curved exponential-family random graphs with geometrically weighted model terms and dependence induced by
sensible forms of transitivity. Concentration results for other curved exponentialfamily random graphs with geometrically weighted model terms, such as geometrically weighted degree and dyadwise shared partner terms [45, 19], can be established along the same lines.
While these results are encouraging, it goes without saying that numerous open
questions remain: e.g., it is unknown whether θb is unique, because the mean-value
parameter vectors µ(η(θ)) ∈ rint(M) are not available in closed form and it is not
straightforward to characterize the set of mean-value parameter vectors induced by
Θ. However, if the set θb contains more than one element, then all elements of the
set θb are close to θ ? with high probability according to Theorem 2, as we pointed
out in the discussion of Theorem 2.
4.3. Extensions to other complex topological features. Theorems 1 and 2 can
be used to establish finite-graph concentration results for maximum likelihood estimators of other sensible specifications of exponential-family random graphs capturing complex topological features of networks. In many cases, the conditions of
Theorems 1 and 2 can be verified along the lines of Corollaries 1 and 2.
5. Finite-graph concentration results for M -estimators. The finite-graph
concentration results for maximum likelihood estimators in Section 4 are special
cases of more general results for M -estimators. To demonstrate, we introduce a
natural class of M -estimators in Section 5.1, which includes both likelihood- and
moment-based estimators, and present finite-graph concentration results in Section
5.2 along with an application to misspecified models with omitted covariate terms.
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
17
5.1. M -estimators. A natural class of M -estimators, which includes both
likelihood- and moment-based estimators, can be constructed as follows.
Let b : X 7→ Rdim(b) be a vector of statistics. These statistics might be
• the sufficient statistics of the data-generating exponential family;
• the sufficient statistics of an exponential family other than the data-generating
exponential family, which implies that the model is misspecified;
• statistics motivated by computational considerations.
An example is presented in Corollary 3, where we consider a misspecified exponential family with omitted covariate terms.
A natural extension of the class of likelihood-based estimating functions in Section 4 is given by estimating functions of the form
(8)
g(θ; b(x)) = kb(x) − β(θ)k2 , θ ∈ Θ,
which are approximations of
g(θ; E b(X)) = kE b(X) − β(θ)k2 , θ ∈ Θ,
provided E b(X) exists. Here, E b(X) is the expectation of b(X) under the datagenerating exponential-family distribution. The vector β : Θ 7→ Rdim(b) is a
vector-valued function of θ ∈ Θ defined by β(θ) = Eθ b(X), where Eθ b(X)
is the expectation of b(X) under a distribution parameterized by θ ∈ Θ, provided
Eθ b(X) exists. The distribution may not belong to the data-generating exponential
family or any other exponential family, i.e., the model may be misspecified.
Estimating functions of the form (8) cover both likelihood- and moment-based
estimating functions:
• Likelihood-based estimating functions: The choice b(x) = s(x) gives
g(θ; s(x)) = ks(x) − Eθ s(X)k2 , which is based on the gradient of the
loglikelihood function with respect to the natural parameter vector of the
data-generating exponential family and is therefore based on moments of
the sufficient statistics of the data-generating exponential family.
• Moment-based estimating functions: The choice b(x) 6= s(x) gives
g(θ; b(x)) = kb(x) − Eθ b(X)k2 , which is based on moments of statistics
other than the sufficient statistics of the data-generating exponential family.
M -estimators based on estimating functions of the form (8) are defined by
b
θ =
θ ∈ Θ : g(θ; b(x)) = inf g(θ̇; b(x)) ,
θ̇∈Θ
which are estimators of
θ0 =
θ ∈ Θ : g(θ; E b(X)) =
inf g(θ̇; E b(X)) .
θ̇∈Θ
18
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
If, given an observation x of a random graph X, the estimating function
g(θ; b(x)) is close to g(θ; E b(X)), then we would expect the minimizers θb and
θ0 of g(θ; b(x)) and g(θ; E b(X)) to be close under suitable conditions. To show
that θb is close to θ0 with high probability, we make the following assumptions. In
the following, B denotes the convex hull of the set {b(x) : x ∈ X}.
[C.1 ] The expectation E b(X) ∈ rint(B) exists and there exists A > 0 such that,
for all (x1 , x2 ) ∈ X × X,
kb(x1 ) − b(x2 )k∞ ≤ A d(x1 , x2 ) kAk∞ .
[C.2 ] The parameter space Θ is an open subset of Rdim(θ) , the expectation β(θ) =
Eθ b(X) exists for all θ ∈ Θ, and there exists a unique element θ0 ∈ Θ such
that β(θ0 ) = E b(X) ∈ rint(B).
[C.3 ] For all > 0 small enough so that B(θ0 , ) ⊆ Θ, there exists δ() > 0 such
that, for all θ ∈ Θ \ B(θ0 , ),
kβ(θ0 ) − β(θ)k2 ≥ δ()
K
X
|Ak | α
k=1
2
for some 0 ≤ α ≤ 1.
Conditions [C.1]—[C.3] are verified in Corollary 3. Conditions [C.1] and [C.3]
resemble the conditions of Theorems 1 and 2 and are verified in Corollaries 1 and 2
in the special case b(x) = s(x). Condition [C.2] is a moment-matching condition:
It assumes that the family of distributions parameterized by θ ∈ Θ, which may
not be the data-generating exponential family, is able to match the first moment
E b(X) of b(X) under the data-generating exponential-family distribution. The
unique parameter vector θ0 ∈ Θ that matches the moment E b(X) is the unique
minimizer of g(θ; E b(X)), i.e., θ0 = arg minθ∈Θ g(θ; E b(X)).
5.2. Finite-graph concentration results. We establish finite-graph concentration results for the class of M -estimators introduced in Section 5.1.
To do so, we need to show that the probability mass of statistic vector b(X)
concentrates around its expectation E b(X) under the data-generating exponentialfamily distribution.
Proposition 3. Consider an exponential family with countable support X and
local dependence. Let b : X 7→ Rdim(b) and assume that condition [C.1] is satisfied.
there
exists C > 0 such that, for all deviations of the form t =
P Then
|Ak | α
δ K
with
δ > 0 and 0 ≤ α ≤ 1,
k=1
2
!
δ2 C K
P (kb(X) − E b(X)k2 ≥ t) ≤ 2 exp −
+ log dim(b) ,
4 (2−α)
dim(b) kAk∞
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
19
where dim(b) may be a function of kAk∞ .
Proposition 3 paves the way for finite-graph concentration results for the class of
M -estimators introduced in Section 5.1. The following concentration result shows
that M -estimators θb = arg minθ∈Θ g(θ; b(X)) based on estimating functions of
the form (8) are close to θ0 = arg minθ∈Θ g(θ; E b(X)) with high probability as
long as the neighborhoods and dim(b) are small relative to the number of neighborhoods K.
Theorem 3. Consider an exponential-family random graph of fixed and finite
size with countable support X and local dependence. Assume that conditions [C.1]—
[C.3] are satisfied. Then, for all > 0 small enough so that B(θ0 , ) ⊆ Θ, there
exist κ() > 0 and C > 0 such that
!
2CK
κ()
P θb ∈ B(θ0 , ) ⊆ Θ ≥ 1 − 2 exp −
+ log dim(b) ,
4 (2−α)
dim(b) kAk∞
where dim(b) may be a function of kAk∞ and K ≥ 2.
Conditions [C.1]—[C.3] of Theorem 3 are verified by Corollary 3. It is worth
noting that θb may not be unique, but an argument along the lines of Section 4.2
shows that, with high probability, the minimizers of estimating function (8) do not
give rise to global minima that are separated by large distances, under the stated
assumptions.
Example: misspecified exponential family with omitted covariate term. To demonstrate, we consider an extension of the exponential-family random graph with withinneighborhood edge and transitive edge terms, as described in Section 4.1. The extended model includes an additional same-attribute edge term with natural parameters ηk,3 (θ) = θ3 and sufficient statistics sk,3 (xk ),
X
sk,3 (xk ) =
xi,j 1(ci = cj ),
i∈Ak < j∈Ak
where 1(ci = cj ) is an indicator function, which is 1 if ci = cj and is 0 otherwise,
and ci ∈ {C1 , . . . , CM } (Cm ∈ R, m = 1, . . . , M , M ≥ 2) is a categorical
attribute of node i ∈ Ak (k = 1, . . . , K). Same-attribute edge terms are popular in applications and capture excesses in the expected number of edges among
nodes with the same attribute [e.g., 19]: e.g., students may prefer to befriend other
students of the same region of origin.
Suppose that researchers are unaware that the attributes ci of nodes i are important predictors of edges and estimate the edge and transitive edge parameters
θ1 and θ2 based on the misspecified exponential family with natural parameter
vector θ = (θ1 , θ2 ) and sufficient statistic vector b(x) = (s1 (x), s2 (x)), where
20
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
PK
P
s1 (x) = K
k=1 sk,2 (xk ) are defined in Section 4.1.
k=1 sk,1 (xk ) and s2 (x) =
In other words, suppose that statistical inference is based on the estimating function
g(θ; b(x)) = kb(x) − Eθ b(X)k2 , θ ∈ Θ0 ,
where Θ0 = R2 is the parameter space of the misspecified exponential family,
which is a two-dimensional subspace of the three-dimensional parameter space
Θ? = R3 of the data-generating exponential family. The following concentration
result shows that the M -estimator θb = arg minθ∈Θ0 g(θ; b(X)) is close to θ0 =
arg minθ∈Θ0 g(θ; E b(X)) with high probability as long as the neighborhoods
are small relative to the number of neighborhoods K.
Corollary 3. Consider an exponential-family random graph of fixed and finite size with within-neighborhood edge, transitive edge, and same-attribute edge
terms. Suppose that an observation of the random graph is generated by θ ? ∈ Θ? .
Then all conditions of Theorem 3 are satisfied and hence, for all > 0 small
enough so that B(θ0 , ) ⊆ Θ0 , there exist κ() > 0 and C > 0 such that θb exists,
is unique, and
κ()2 C K
b
P θ ∈ B(θ0 , ) ⊆ Θ0
≥ 1 − 4 exp −
,
kAk4∞
provided |Ak | ≥ 3 (k = 1, . . . , K) and K ≥ 2.
It is worth noting that θ0 ∈ R2 and θ ? ∈ R3 do not have the same dimension and cannot be identical, but the distribution parameterized by θ0 is as close
as possible to the data-generating distribution parameterized by θ ? in terms of
Kullback-Leibler divergence: by construction, θ0 minimizes kE b(X)−Eθ b(X)k2
=
kE ∇θ log pθ (X)k2 and hence maximizes E log pθ (X), where
pθ (x) ∝ exp(hθ, b(x)i) denotes the density of x ∈ X under the misspecified
exponential-family distribution with natural parameter vector θ = (θ1 , θ2 ) and
sufficient statistic vector b(x) = (s1 (x), s2 (x)). Therefore, θ0 satisfies
(9)
θ0 = arg max E log pθ (X) − E log pθ? (X) = arg min KL(Pθ? ; Pθ ),
θ ∈ Θ0
θ ∈ Θ0
where KL(Pθ? ; Pθ ) is the Kullback-Leibler divergence from Pθ? to Pθ and the
expectations E log pθ (X) and E log pθ? (X) are with respect to Pθ? . Owing to the
dependence of within-neighborhood edge variables and sufficient statistics, it is
not straightforward to bound kθ ? − θ0 k2 , but—by the properties of maximum likelihood estimation—we are assured that the distribution parameterized by θ0 is as
close as possible to the distribution parameterized by the data-generating parameter
vector θ ? in terms of Kullback-Leibler divergence, as shown by (9).
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
21
6. Extendability to larger random graphs with more neighborhoods and
subgraph-to-graph estimation. A natural question to ask is whether it makes
sense to extend a model of a random graph of fixed and finite size to a larger
random graph [43, 10, 30]. We discuss the issue of extendability in Section 6.1 and
the related issue of subgraph-to-graph estimation in Section 6.2.
6.1. Extendability to larger random graphs with more neighborhoods. While
many exponential-family random graphs cannot be extended to larger exponentialfamily random graphs [43, 10, 30], an exponential-family random graph with multilevel structure can be extended to a larger exponential-family random graph with
more neighborhoods.
To demonstrate, consider a population graph (XL , YL ) with a set of neighborhoods L = {A1 , . . . , AL }, where XL ∈ XL and YL ∈ YL denote the sequences
of within- and between-neighborhood edge variables based on the set of neighborhoods L, respectively. As before, assume that XL is governed by an exponential family with countable support XL and local dependence, with neighborhooddependent natural parameters ηA,i (θ) = ηi (θ) and sufficient statistics sA,i (xA )
(i = 1, . . . , dim(ηA ), A ∈ L). Therefore, the exponential family can be reduced
to an exponential family with natural parameter vector
η(θ) = (η1 (θ), . . . , ηd∞ (θ))
(10)
and sufficient statistic vector
s(xL ) = (s1 (xL ), . . . , sd∞ (xL )) ,
P
where si (xL ) = A∈L sA,i (xA ) (i = 1, . . . , d∞ ) and d∞ = maxA∈L dim(ηA ).
Consider a subgraph (XK , YK ) induced by a subset of neighborhoods K ⊆ L.
Then the subgraph (XK , YK ) with subset of neighborhoods K is extendable to the
population graph (XL , YL ) with set of neighborhoods L ⊃ K as follows.
Proposition 4. Consider a full or non-full, curved exponential-family random
graph with set of neighborhoods L, countable support XL , and local dependence.
Assume that, for all yL ∈ YL ,
Y
P(YL = yL ) =
P(YC,D = yC,D ),
C∈L, D∈L, C6=D
where YC,D = (Yi,j )i∈C, j∈D . Then, for all θ ∈ Θ ⊆ {θ ∈ Rdim(θ) : ψL (η(θ)) <
∞}, all K ⊆ L, and all xK ∈ XK and yK ∈ YK ,
Pη(θ) (XK = xK , YK = yK , XL\K ∈ XL\K , YL\K ∈ YL\K )
= Pη(θ) (XK = xK , YK = yK ),
22
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where
ψL (η(θ)) =
X
ψA (η(θ)) =
A∈L
X
A∈L
log
X
exp (hη(θ), sA (xA )i) νA (xA ).
xA ∈XA
The marginal density of a subgraph xK ∈ XK of xL ∈ XL induced by K ⊆ L is
an exponential-family density with support XK and local dependence:
X
pη(θ) (xL ) = pη(θ) (xK ) = exp (hη(θ), s(xK )i − ψK (η(θ))) νK (xK ),
xL\K ∈XL\K
P
where
. , ηd∞ (θ)), s(xK ) =
A∈K ψA (η(θ)), η(θ) = (η1 (θ), . .Q
P ψK (η(θ)) = P
( A∈K sA,1 (xA ), . . . , A∈K sA,d∞ (xA )), and νK (xK ) = A∈K νA (xA ).
In other words, the marginal probability mass function of a subgraph induced by
a subset of neighborhoods K is consistent with the joint probability mass function
of the population graph with set of neighborhoods L ⊃ K. Thus, the exponentialfamily random graph induced by subset of neighborhoods K can be extended to
the exponential-family random graph with set of neighborhoods L ⊃ K. A more
restrictive result was proved by Schweinberger and Handcock [40, Theorem 1].
6.2. Finite-graph concentration results for subgraph-to-graph estimators. In
practice, researchers may wish to extend an exponential-family random graph to
a larger exponential-family random graph with more neighborhoods when it is infeasible to observe a population graph with a set of neighborhoods L and therefore
a subgraph induced by a subset of neighborhoods K ⊆ L is sampled.
We consider here population graphs of finite size, because in practice population graphs are of finite size and the concentration results in Sections 3, 4, and 5
enable concentration results for population graphs of finite size. Let L be the set of
neighborhoods of the population graph and assume that xL ∈ XL was generated
by an exponential family with countable support XL and local dependence. Suppose that it is infeasible to observe xL ∈ XL , but it is feasible to sample a subset
of neighborhoods K ⊆ L and collect data on the subgraphs induced by K ⊆ L.
We assume henceforth that the sampling design is ignorable in the sense of Rubin
[38] and Handcock and Gile [15], i.e., the probability of observing subgraphs does
not depend on the unobserved subgraphs. A simple example is a sampling design
that samples neighborhoods at random and collects data on the subgraphs induced
by the sampled neighborhoods.
By Proposition 4 and the ignorability of the sampling design [38, 15], the
observed-data likelihood function based on the observed subgraph xK ∈ XK of
xL ∈ XL is proportional to
X
pη(θ) (xL ) = pη(θ) (xK ),
xL ∈ XL ,
xL\K ∈ XL\K
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
23
where η(θ) is of the form (10). In other words, maximum likelihood estimation
can be based on pη(θ) (xK ). Motivated by the same considerations we detailed in
Section 4, we therefore consider estimating functions of the form
gK (θ; µK\
(η(θ ? ))) = kµK\
(η(θ ? )) − µK (η(θ))k2 ,
where µK\
(η(θ ? )) = s(xK ), µK (η(θ)) = Eη(θ) s(XK ), and s(xK ) is defined in
Proposition 4. The data-generating parameter vector θ ? of the population graph can
hence be estimated by the estimator θbK based on the observed subgraph xK ∈ XK
of xL ∈ XL :
?
?
\
\
b
θK =
θ ∈ Θ : gK (θ; µK (η(θ ))) = inf gK (θ̇; µK (η(θ ))) .
θ̇∈Θ
The following finite-graph concentration result shows that, with high probability,
the estimator θbK based on the observed subgraph induced by K ⊆ L is close to
the data-generating parameter vector θ ? of the population graph as long as the
neighborhoods and the dimensions of neighborhood-dependent natural parameter
vectors are small relative to the number of sampled neighborhoods |K|.
Theorem 4. Consider a full or non-full, curved exponential-family random graph
with set of neighborhoods L, countable support XL , and local dependence. Let
Θ ⊆ {θ ∈ Rdim(θ) : ψL (η(θ)) < ∞}, Θ open.
Assume that θ ? ∈ int(Θ) and that, for all > 0 small enough so that B(θ ? , ) ⊆
int(Θ), there exists γ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ), we have
η(θ) ∈ N \ B(η(θ ? ), γ()), where η(θ) is of the form (10). In addition, assume that there exist δ() > 0 and A > 0 such that, for all K ⊆ L and all
η(θ) ∈ N \ B(η(θ ? ), γ()),
X |A|α
?
for some 0 ≤ α ≤ 1
(11) kµK (η(θ )) − µK (η(θ))k2 ≥ δ()
2
A∈K
and, for all K ⊆ L and all (x1 , x2 ) ∈ XK × XK ,
(12)
ks(x1 ) − s(x2 )k∞ ≤ A d(x1 , x2 ) kKk∞ ,
where kKk∞ = maxA∈K |A|. Then, for all > 0 small enough so that B(θ ? , ) ⊆
int(Θ), there exist κ() > 0 and C > 0 such that
!
2 C |K|
κ()
P θbK ∈ B(θ ? , ) ⊆ int(Θ) ≥ 1 − 2 exp −
+ log d∞ ,
4 (2−α)
d∞ kLk∞
24
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where d∞ = maxA∈L dim(ηA ) may be a function of kLk∞ and |K| ≥ 2. If the exponential family is full, then θbK is unique in the event θbK ∈ B(θ ? , ) ⊆ int(Θ).
Theorem 4 shows that there are costs associated with observing a subset of
neighborhoods K ⊆ L rather than the whole set of neighborhoods L of the population graph: the probability of event θbK 6∈ B(θ ? , ) decays with the number of
observed neighborhoods |K| and is hence lowest when the whole set of neighborhoods L of the population graph is observed.
To demonstrate, consider curved exponential-family random graphs with withinneighborhood edge and geometrically weighted edgewise shared partner terms, as
described in Section 4.2.
Corollary 4. Consider a curved exponential-family random graph with set of
neighborhoods L, countable support XL , and local dependence induced by withinneighborhood edge and geometrically weighted edgewise shared partner terms.
Let Θ = R × (0, 1) and assume that θ ? ∈ int(Θ). Then all conditions of Theorem
4 are satisfied and hence, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ),
there exist κ() > 0 and C > 0 such that
κ()2 C |K|
?
b
P θK ∈ B(θ , ) ⊆ int(Θ) ≥ 1 − 2 exp −
+ logkLk∞ ,
kLk6∞
provided |A| ≥ 4 (A ∈ L) and |K| ≥ 2.
Remark 3. “Bad” subsets of neighborhoods K ⊆ L. Since the neighborhoods
need not have the same size, it is natural to ask whether it is possible to sample
a “bad” subset of neighborhoods K with too small or too large neighborhoods,
which could make it challenging to estimate some of the parameters. However, the
assumptions of Theorem 4 rule out “bad” subsets of neighborhoods K, for two
reasons. First, while some neighborhoods may be larger than others, Theorem 4
assumes that the neighborhoods are of the same order of magnitude, as defined in
Section 2. In other words, the neighborhoods have similar sizes. Second, the conditions of Theorem 4 assume that the model satisfies identifiability and smoothness
conditions for all possible subsets of neighborhoods K ⊆ L. Corollary 4 shows
that, in the special case of curved exponential-family random graphs with withinneighborhood edge and geometrically weighted edgewise shared partner terms, the
identifiability conditions require |A| ≥ 4 for all neighborhoods A ∈ L of the population graph. Thus, no neighborhood can be too small and, at the same time, no
neighborhood can be too large, because all neighborhoods are of the same order
of magnitude. As a consequence, under the stated assumptions, it is impossible to
sample a “bad” subset of neighborhoods K with too small or too large neighborhoods.
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
25
7. Comparison with existing asymptotic consistency results. We focus on
asymptotic consistency results for exponential-family random graphs with dependence. We note that there exist asymptotic consistency results for exponentialfamily random graphs with independence assumptions—see, e.g., Diaconis et al.
[11], Rinaldo et al. [37], Krivitsky and Kolaczyk [29], Yan et al. [51, 50]—but such
independence assumptions may not be satisfied in applications, as discussed in
Section 1. Concerning exponential-family random graphs with dependence, Shalizi and Rinaldo [43] showed that maximum likelihood estimators of exponentialfamily random graphs satisfying strong extendability assumptions are consistent,
but those strong extendability assumptions rule out dependencies induced by transitivity and many other complex topological features of networks. Xiang and Neville
[49] reported asymptotic consistency results under weak dependence assumptions,
but did not give any example of an exponential-family random graph with dependence that satisfies those assumptions. Mukherjee [34] showed that consistent estimation of the so-called two-star model is possible, but those results have not been
extended to other exponential-family random graphs. In addition, Shalizi and Rinaldo [43], Xiang and Neville [49], and Mukherjee [34] focus on asymptotic consistency results for estimators of natural parameter vectors whose dimensions do
not depend on the number of nodes. In contrast, we advance the statistical theory of
exponential-family random graphs by providing the first finite-graph concentration
and consistency results that cover
• a wide range of exponential-family random graphs with dependence induced
by sensible forms of transitivity and other complex topological features of
networks;
• curved exponential-family random graphs with dependence and parameter
vectors whose dimensions can depend on the number of nodes;
• correct and incorrect specifications of exponential-family random graphs;
• M -estimators, including likelihood- and moment-based estimators;
• subgraph-to-graph estimators.
These results demonstrate the importance of additional structure: It is the additional
structure in the form of multilevel structure that facilitates these results.
8. Simulation results. To shed light on the finite-graph properties of maximum likelihood estimators, we generated data from the full and non-full, curved
exponential-family random graphs described in Section 4. We used R package
hergm [41] to generate 1,000 graphs from each model and estimated the datagenerating parameter vector by Monte Carlo maximum likelihood estimators [19].
Figure 1 shows 1,000 estimates of the exponential-family random graph with
within-neighborhood edge and transitive edge terms, where each graph consists
of K = 100, 200, and 300 neighborhoods of size 50 with natural parameter vec-
26
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
K = 100
K = 200
K = 300
0.525
●
0.500
0.475
0.450
0.550
Transitive edge parameter θ2
0.550
Transitive edge parameter θ2
Transitive edge parameter θ2
0.550
0.525
●
0.500
0.475
0.450
−2.04
−2.00
Edge parameter θ1
−1.96
0.525
●
0.500
0.475
0.450
−2.04
−2.00
Edge parameter θ1
−1.96
−2.04
−2.00
Edge parameter θ1
−1.96
1.0
●
0.9
0.8
−2.7
−2.6
−2.5
−2.4
Edge parameter θ1
−2.3
K = 100
−0.15
−0.20
−0.25
●
−0.30
−0.35
−0.40
−0.45
−0.4
−0.3
−0.2
Edge deviation θ3 (50)
−0.1
Transitive edge deviation θ6 (75)
Transitive edge parameter θ2
K = 100
Transitive edge deviation θ4 (50)
F IG 1. 1,000 estimates of the exponential-family random graph with within-neighborhood edge and
transitive edge terms, where each graph consists of K = 100, 200, and 300 neighborhoods of size
50 with natural parameter vectors ηk (θ) = (θ1 , θ2 ). The horizontal and vertical lines indicate the
coordinates of the data-generating parameter vector θ ? = (θ1? , θ2? ).
K = 100
−0.4
−0.5
●
−0.6
−0.7
−1.0
−0.9
−0.8
−0.7
Edge deviation θ5 (75)
−0.6
F IG 2. 1,000 estimates of the exponential-family random graph with within-neighborhood edge
and transitive edge terms, where each graph consists of 33 neighborhoods of size 25 with natural
parameter vectors ηk (θ) = (θ1 , θ2 ), 34 neighborhoods of size 50 with natural parameter vectors
ηk (θ) = (θ1 + θ3 , θ2 + θ4 ), and 33 neighborhoods of size 75 with natural parameter vectors
ηk (θ) = (θ1 + θ5 , θ2 + θ6 ). The horizontal and vertical lines indicate the coordinates of the datagenerating parameter vector θ ? = (θ1? , θ2? , θ3? , θ4? , θ5? , θ6? ).
tors ηk (θ) = (θ1 , θ2 ) (k = 1, . . . , K). The figure suggests that the probability
mass of estimators becomes more and more concentrated in a neighborhood of the
data-generating parameters as the number of neighborhoods K increases from 100
to 300, demonstrating that the finite-graph concentration results in Section 4 are
manifest when K is in the low hundreds and kAk∞ = 50.
Figure 2 sheds light on the performance of a simple form of a size-dependent parameterization that allows small and large neighborhoods to have different parameters. We consider exponential-family random graphs with within-neighborhood
edge and transitive edge terms, where each graph consists of 33 neighborhoods of
size 25 with natural parameter vectors ηk (θ) = (θ1 , θ2 ) (k = 1, . . . , 33), 34 neighborhoods of size 50 with natural parameter vectors ηk (θ) = (θ1 + θ3 , θ2 + θ4 )
(k = 34, . . . , 67), and 33 neighborhoods of size 75 with natural parameter vectors
ηk (θ) = (θ1 + θ5 , θ2 + θ6 ) (k = 68, . . . , 100). Figure 2 demonstrates that the
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
27
25
40
20
30
15
20
10
10
5
0
−3.06
−3.03
−3.00
−2.97
Edge parameter θ1
0
0.18
0.20
0.22
Geometric term parameter θ2
F IG 3. 1,000 estimates of the curved exponential-family random graph with within-neighborhood
edge and geometrically weighted edgewise shared partner terms, where each graph consists of K = 100 neighborhoods of size 50 with natural parameter vectors ηk (θ) =
(θ1 , ηk,1 (θ2 ), . . . , ηk,48 (θ2 )). The vertical lines indicate the coordinates of the data-generating
parameter vector θ ? = (θ1? , θ2? ).
estimates of the baseline edge and transitive edge parameters θ1 and θ2 tend to be
closer to the data-generating parameters than the deviation parameters θ3 , θ4 , θ5 ,
and θ6 , because θ1 and θ2 are estimated from all neighborhoods whereas θ3 , θ4 , θ5 ,
and θ6 are estimated from a subset of neighborhoods.
Figure 3 shows 1,000 estimates of the curved exponential-family random graph
with within-neighborhood edge and geometrically weighted edgewise shared partner terms, where each graph consists of K = 100 neighborhoods of size 50
with natural parameter vectors ηk (θ) = (θ1 , ηk,1 (θ2 ), . . . , ηk,48 (θ2 )), where θ1
is the natural parameter of the edge term and ηk,1 (θ2 ), . . . , ηk,48 (θ2 ) are the natural parameters of the geometrically weighted edgewise shared partner term (k =
1, . . . , 100). The figure shows that the probability mass of the estimators is concentrated in a small neighborhood of the data-generating parameters.
9. Discussion. We have taken constructive steps to demonstrate that statistical
inference for transitivity and other complex topological features of networks is
possible, provided sensible specifications of exponential-family random graphs are
used and additional structure in the form of multilevel structure is available.
The theoretical results reported here demonstrate the importance of additional
structure: It is the additional structure that facilitates these results. A pleasant feature of the general setup of the statistical theory developed here, in particular the
statistical theory in Sections 4 and 5, is that it can be extended to models with other
forms of additional structure, such as neighborhood structures with overlapping
neighborhoods. Such extensions are possible as long as the dependence induced by
the model is sufficiently weak and the model satisfies smoothness and identifiability assumptions.
Last, but not least, it is worth noting that additional structure not only facilitates
28
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
theoretical results, but also facilitates computing: the contributions of neighborhoods to estimating functions—e.g., the expectations of within-neighborhood sufficient statistics—may be computed or approximated by exploiting parallel computing on computing clusters.
Acknowledgements. We acknowledge support from the National Science Foundation (NSF award DMS-1513644).
Supplementary materials. All results are proved in the supplement [42].
References.
[1] Bhamidi, S., Bresler, G., and Sly, A. (2011), “Mixing time of exponential random graphs,” The
Annals of Applied Probability, 21, 2146–2170.
[2] Bhattacharya, B. B., and Mukherjee, S. (2018), “Inference in Ising models,” Bernoulli, 24,
493–525.
[3] Boucheron, S., Lugosi, G., and Massart, P. (2013), Concentration Inequalities: A Nonasymptotic Theory of Independence, Oxford: Oxford University Press.
[4] Brown, L. (1986), Fundamentals of Statistical Exponential Families: With Applications in Statistical Decision Theory, Hayworth, CA, USA: Institute of Mathematical Statistics.
[5] Butts, C. T. (2011), “Bernoulli graph bounds for general random graph models,” Sociological
Methodology, 41, 299–345.
[6] Butts, C. T., and Almquist, Z. W. (2015), “A Flexible Parameterization for Baseline Mean
Degree in Multiple-Network ERGMs,” Journal of Mathematical Sociology, 39, 163–167.
[7] Chatterjee, S. (2007), “Estimation in spin glasses: A first step,” The Annals of Statistics, 35,
1931–1946.
[8] Chatterjee, S., and Diaconis, P. (2013), “Estimating and understanding exponential random
graph models,” The Annals of Statistics, 41, 2428–2461.
[9] Chung, F., and Lu, L. (2006), Complex Graphs and Networks, Providence: American Mathematical Society.
[10] Crane, H., and Dempsey, W. (2015), “A framework for statistical network modeling,” Available
at https://arxiv.org/abs/1509.08185.v4.
[11] Diaconis, P., Chatterjee, S., and Sly, A. (2011), “Random graphs with a given degree sequence,”
The Annals of Applied Probability, 21, 1400–1435.
[12] Efron, B. (1978), “The geometry of exponential families,” The Annals of Statistics, 6, 362–376.
[13] Frank, O., and Strauss, D. (1986), “Markov graphs,” Journal of the American Statistical Association, 81, 832–842.
[14] Handcock, M. S. (2003), “Statistical Models for Social Networks: Inference and Degeneracy,” in Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers,
eds. Breiger, R., Carley, K., and Pattison, P., Washington, D.C.: National Academies Press, pp.
1–12.
[15] Handcock, M. S., and Gile, K. (2010), “Modeling Social Networks from Sampled Data,” The
Annals of Applied Statistics, 4, 5–25.
[16] Holland, P. W., and Leinhardt, S. (1976), “Local Structure in Social Networks,” Sociological
Methodology, 1–45.
[17] Hollway, J., and Koskinen, J. (2016), “Multilevel embeddedness: The case of the global fisheries governance complex,” Social Networks, 44, 281–294.
[18] Hollway, J., Lomi, A., Pallotti, F., and Stadtfeld, C. (2017), “Multilevel social spaces: The
network dynamics of organizational fields,” Network Science, 5, 187–212.
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
29
[19] Hunter, D. R., and Handcock, M. S. (2006), “Inference in curved exponential family models
for networks,” Journal of Computational and Graphical Statistics, 15, 565–583.
[20] Hunter, D. R., Krivitsky, P. N., and Schweinberger, M. (2012), “Computational statistical methods for social network models,” Journal of Computational and Graphical Statistics, 21, 856–
882.
[21] Janson, S., and Rucinski, A. (2002), “The infamous upper tail,” Random Structures and Algorithms, 20, 317–342.
[22] Jonasson, J. (1999), “The random triangle model,” Journal of Applied Probability, 36, 852–876.
[23] Kass, R., and Vos, P. (1997), Geometrical foundations of asymptotic inference, New York:
Wiley.
[24] Kim, J. H., and Vu, V. H. (2004), “Divide and conquer martingales and the number of triangles
in a random graph,” Random Structures & Algorithms, 24, 166–174.
[25] Kolaczyk, E. D. (2009), Statistical Analysis of Network Data: Methods and Models, New York:
Springer-Verlag.
[26] Kontorovich, L., and Ramanan, K. (2008), “Concentration inequalities for dependent random
variables via the martingale method,” The Annals of Probability, 36, 2126–2158.
[27] Krivitsky, P. N. (2012), “Exponential-family models for valued networks,” Electronic Journal
of Statistics, 6, 1100–1128.
[28] Krivitsky, P. N., Handcock, M. S., and Morris, M. (2011), “Adjusting for network size and
composition effects in exponential-family random graph models,” Statistical Methodology, 8,
319–339.
[29] Krivitsky, P. N., and Kolaczyk, E. D. (2015), “On the question of effective sample size in
network modeling: An asymptotic inquiry,” Statistical Science, 30, 184–198.
[30] Lauritzen, S., Rinaldo, A., and Sadeghi, K. (2017), “Random networks, graphical models, and
exchangeability,” Available at https://arxiv:1701.08420v1.
[31] Lazega, E., and Snijders, T. A. B. (eds.) (2016), Multilevel Network Analysis for the Social
Sciences, Switzerland: Springer-Verlag.
[32] Lomi, A., Robins, G., and Tranmer, M. (2016), “Introduction to multilevel social networks,”
Social Networks, 266–268.
[33] Lusher, D., Koskinen, J., and Robins, G. (2013), Exponential Random Graph Models for Social
Networks, Cambridge, UK: Cambridge University Press.
[34] Mukherjee, S. (2013), “Consistent estimation in the two star exponential random graph model,”
Tech. rep., Department of Statistics, Columbia University, arXiv:1310.4526v1.
[35] Nowicki, K., and Snijders, T. A. B. (2001), “Estimation and prediction for stochastic blockstructures,” Journal of the American Statistical Association, 96, 1077–1087.
[36] Ravikumar, P., Wainwright, M. J., and Lafferty, J. (2010), “High-dimensional Ising model selection using `1 -regularized logistic regression,” The Annals of Statistics, 38, 1287–1319.
[37] Rinaldo, A., Petrovic, S., and Fienberg, S. E. (2013), “Maximum likelihood estimation in network models,” The Annals of Statistics, 41, 1085–1110.
[38] Rubin, D. B. (1976), “Inference and missing data,” Biometrika, 63, 581–592.
[39] Schweinberger, M. (2011), “Instability, sensitivity, and degeneracy of discrete exponential families,” Journal of the American Statistical Association, 106, 1361–1370.
[40] Schweinberger, M., and Handcock, M. S. (2015), “Local dependence in random graph models:
characterization, properties and statistical inference,” Journal of the Royal Statistical Society,
Series B, 77, 647–676.
[41] Schweinberger, M., and Luna, P. (2017), “HERGM: Hierarchical exponential-family random
graph models,” Journal of Statistical Software, accepted.
[42] Schweinberger, M., and Stewart, J. (2017), “Supplement: Finite-graph concentration and consistency results for random graphs with complex topological structures,” Tech. rep., Department
of Statistics, Rice University.
30
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
[43] Shalizi, C. R., and Rinaldo, A. (2013), “Consistency under sampling of exponential random
graph models,” The Annals of Statistics, 41, 508–535.
[44] Slaughter, A. J., and Koehly, L. M. (2016), “Multilevel models for social networks: hierarchical
Bayesian approaches to exponential random graph modeling,” Social Networks, 44, 334–345.
[45] Snijders, T. A. B., Pattison, P. E., Robins, G. L., and Handcock, M. S. (2006), “New specifications for exponential random graph models,” Sociological Methodology, 36, 99–153.
[46] Strauss, D. (1986), “On a general class of models for interaction,” SIAM Review, 28, 513–527.
[47] Wang, P., Robins, G., Pattison, P., and Lazega, E. (2013), “Exponential random graph models
for multilevel networks,” Social Networks, 35, 96–115.
[48] Wasserman, S., and Pattison, P. (1996), “Logit models and logistic regression for social networks: I. An introduction to Markov graphs and p∗ ,” Psychometrika, 61, 401–425.
[49] Xiang, R., and Neville, J. (2011), “Relational learning with one network: an asymptotic analysis,” in Proceedings of the 14th International Conference on Artificial Intelligence and Statistics
(AISTATS), pp. 1–10.
[50] Yan, T., Leng, C., and Zhu, J. (2016), “Asymptotics in directed exponential random graph
models with an increasing bi-degree sequence,” The Annals of Statistics, 44, 31–57.
[51] Yan, T., Wang, H., and Qin, H. (2016), “Asymptotics in undirected random graph models parameterized by the strengths of vertices,” Statistica Sinica, 26, 273–293.
[52] Zappa, P., and Lomi, A. (2015), “The analysis of multilevel networks in organizations: models
and empirical tests,” Organizational Research Methods, 18, 542–569.
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
1
SUPPLEMENT:
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
FOR RANDOM GRAPHS WITH COMPLEX TOPOLOGICAL
STRUCTURES
B Y M ICHAEL S CHWEINBERGER A ND J ONATHAN S TEWART
Rice University
We prove the main concentration results for dependent random graphs in Appendix A and the main concentration results for maximum likelihood estimators,
M -estimators, and subgraph-to-graph estimators in Appendices B, C, and D, respectively. Auxiliary lemmas are proved in Appendix E.
APPENDIX A: PROOFS: FINITE-GRAPH CONCENTRATION RESULTS
FOR DEPENDENT RANDOM GRAPHS
We prove the main concentration results of Sections 3 and 5, Propositions 1, 2,
and 3.
Proof of Proposition 1. By assumption, E f (X) exists. We are interested in
deviations of the form |f (X) − E f (X)| ≥ t, where t > 0. Choose any t > 0 and
let X = {x ∈ X : |f (x) − E f (X)| ≥ t}. Since within-neighborhood edges do
not depend on between-neighborhood edges,
P(X ∈ X, Y ∈ Y) = P(X ∈ X).
In the following, we denote by P a probability measure on (X, S) with densities
of the form (1), where S is the power set of the countable set X. Keep in mind
that X = (Xk )K
k=1 denotes the sequence of within-neighborhood edge variables,
where Xk = (Xi,j )i∈Ak < j∈Ak . In an abuse of notation, we denote the elements of
the sequence of edge variables
by X
1 , . . . , Xm with sample spaces X1 , . . . , Xm ,
P X |A
k|
respectively, where m = K
is the number of within-neighborhood edge
k=1
2
variables. Let Xi:j = (Xi , . . . , Xj ) be a subsequence of edge variables with sample space Xi:j , where i ≤ j. By applying Theorem 1.1 of Kontorovich and Ramanan [26] to kf kLip -Lipschitz functions f : X 7→ R defined on the countable set
X,
t2
P(|f (X) − E f (X)| ≥ t) ≤ 2 exp −
,
2 m kΦk2∞ kf k2Lip
2
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where Φ is the m × m-upper triangular matrix with entries
ϕi,j if i < j
φi,j =
1
if i = j
0
if i > j
and
kΦk∞ =
m
X
max 1 +
1≤i≤m
ϕi,j .
j=i+1
The coefficients ϕi,j are known as mixing coefficients and are defined by
ϕi,j ≡
sup
x1:i−1 ∈X1:i−1
(xi , x?i )∈Xi ×Xi
ϕi,j (x1:i−1 , xi , x?i ) =
sup
x1:i−1 ∈X1:i−1
(xi , x?i )∈Xi ×Xi
kπxi − πx?i kTV ,
where kπxi − πx?i kTV is the total variation distance between the distributions πxi
and πx?i given by
πxi ≡ π(xj:m | x1:i−1 , xi ) = P(Xj:m = xj:m | X1:i−1 = x1:i−1 , Xi = xi )
and
πx?i ≡ π(xj:m | x1:i−1 , x?i ) = P(Xj:m = xj:m | X1:i−1 = x1:i−1 , Xi = x?i ).
Since the support of πxi and πx?i is countable,
kπxi − πx?i kTV =
1
2
X
|π(xj:m | x1:i−1 , xi ) − π(xj:m | x1:i−1 , x?i )|.
xj:m ∈Xj:m
An upper bound on kΦk∞ can be obtained by bounding the mixing coefficients ϕi,j
as follows. Consider any pair of edge variables Xi and Xj . If Xi and Xj involve
nodes in more than one neighborhood, the mixing coefficient ϕi,j vanishes by the
local dependence induced by exponential families of the form (1). If the pair of
nodes corresponding to Xi and the pair of nodes corresponding to Xj belong to
the same neighborhood, the mixing coefficient ϕi,j can be bounded as follows:
ϕi,j (x1:i−1 , xi , x?i ) =
≤
1
2
X
xj:m ∈Xj:m
1
2
X
|π(xj:m | x1:i−1 , xi ) − π(xj:m | x1:i−1 , x?i )|
xj:m ∈Xj:m
π(xj:m | x1:i−1 , xi ) +
1
2
X
xj:m ∈Xj:m
π(xj:m | x1:i−1 , x?i ) = 1,
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
3
because πxi and πx?i are conditional probability mass functions with countable
support Xj:m . We note that the upper bound is not sharp, but it has the advantage
that it covers a wide range of dependencies within neighborhoods. Thus,
kΦk∞ =
max 1 +
1≤i≤m
m
X
j=i+1
ϕi,j
kAk∞
,
≤
2
∞
− 1 other edge varibecause each edge variable Xi can depend on at most kAk
2
ables corresponding to pairs of nodes belonging to the same pair of neighborhoods.
Therefore, there exists C > 0 such that, for all t > 0,
!
t2
P(|f (X) − E f (X)| ≥ t) ≤ 2 exp − PK |A |
,
C k=1 2k kAk4∞ kf k2Lip
where kAk∞ ≥ 1 because all neighborhoods Ak are non-empty and kf kLip > 0 by
assumption.
Proof of Proposition 2. By assumption, the neighborhood-dependent natural
parameters ηk,i (θ) are of the form ηk,i (θ) = ηi (θ) (i = 1, . . . , dim(ηk ), k =
1, . . . , K). Therefore, the exponential family can be reduced to an exponential family with natural parameter vector
η(θ) = (η1 (θ), . . . , ηd∞ (θ))
and sufficient statistic vector
s(x) = (s1 (x), . . . , sd∞ (x)) ,
where d∞ = max1≤k≤K dim(ηk ). Here, the sufficient statistics si (x) are sums of
within-neighborhood sufficient statistics sk,i (xk ) (k = 1, . . . , K):
si (x) =
K
X
sk,i (xk ), i = 1, . . . , d∞ .
k=1
\? )) = s(X) and that µ(η(θ ? )) = Eη(θ? ) s(X) = E s(X) ∈
Observe that µ(η(θ
rint(M) exists, because η(θ ? ) ∈ int(N) [4, Theorem 2.2, pp. 34–35]. We bound
the probability of deviations of s(X) from E s(X) in terms of the `∞ - and `2 -norm
below.
4
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
`∞ -norm. For all δ > 0,
\? )) − µ(η(θ ? ))k∞ ≥ δ
P kµ(η(θ
K
X
|Ak | α
!
2
!
K
X
|Ak | α
≥ δ
2
k=1
= P ks(X) − E s(X)k∞
k=1
≤ P
d
∞
[
|si (X) − E si (X)| ≥ δ
i=1
!!
K
X
|Ak | α
2
k=1
.
By condition (2) of Proposition 2, there exists A > 0 such that the Lipschitz coefficient of si : X 7→ R with respect to the Hamming metric d : X × X 7→ R+
0 is
bounded above by ksi kLip ≤ A kAk∞ (i = 1, . . . , d∞ ). Thus, by a union bound
over the d∞ components of s(X) and by applying Proposition 1 to deviations of
PK |Ak |α
, we obtain, for all δ > 0,
size t = δ
k=1
2
P
d
∞
[
|si (X) − E si (X)| ≥ δ
i=1
!!
K
X
|Ak | α
k=1
≤ 2 exp −
δ2
B
K
|Ak | α
k=1
2
P
|Ak |
k=1
2
PK
2
2
kAk4∞ kAk2∞
+ log d∞ ,
where B > 0. Since all neighborhoods are of the same order of magnitude, there
exist C1 > 0 and C2 > 0 such that C1 kAk∞ ≤ |Ak | ≤ C2 kAk∞ . Thus, there
P
|Ak | α 2
) in the numerator of the exponent
exists C3 > 0 such that the term ( K
k=1
2
can be bounded below by
!2
K
X
|Ak | α
k=1
2
≥ C3 K 2 kAk4∞α ,
and there exists C4 > 0 such that the term
exponent can be bounded above by
K
X
|Ak |
k=1
2
PK
k=1
|Ak |
2
in the denominator of the
≤ C4 K kAk2∞ .
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
5
As a result, there exists C > 0 such that, for all δ > 0,
\? )) − µ(η(θ ? ))k∞ ≥ δ
P kµ(η(θ
!
K
X
|Ak | α
k=1
≤ 2 exp −
δ2 C K
4 (2−α)
kAk∞
2
!
+ log d∞
.
`2 -norm. The same argument used above shows that, for all δ > 0,
α !
K
X
|A
|
k
?
\? )) − µ(η(θ ))k2 ≥ δ
P kµ(η(θ
2
k=1
PK |Ak |α !
δ
k=1
2
\? )) − µ(η(θ ? ))k∞ ≥
√
≤ P kµ(η(θ
d∞
!
δ2 C K
≤ 2 exp −
+ log d∞ .
4 (2−α)
d∞ kAk∞
Proof of Proposition 3. Condition [C.1] implies that E b(X) exists. Thus, for
all δ > 0,
!
K
X
|Ak | α
P kb(X) − E b(X)k2 ≥ δ
2
k=1
PK |Ak |α !
δ
pk=1 2
≤ P kb(X) − E b(X)k∞ ≥
dim(b)
PK |Ak |α !
dim(b)
[
δ
.
pk=1 2
≤ P
|bi (X) − E bi (X)| ≥
dim(b)
i=1
By condition [C.1], there exists A > 0 such that the Lipschitz coefficient of bi :
X 7→ R with respect to the Hamming metric d : X × X 7→ R+
0 is bounded above
by kbi kLip ≤ A kAk∞ (i = 1, . . . , dim(b)). Thus, by a union bound over the
dim(b) components of b(X) and by applying Proposition 1 to deviations of size
PK |Ak |α p
t=δ
/ dim(b), we obtain, for all δ > 0,
k=1
2
PK |Ak |α !
dim(b)
[
δ
pk=1 2
P
|bi (X) − E bi (X)| ≥
dim(b)
i=1
P
2
K
|Ak | α
2
δ
k=1
2
≤ 2 exp −
+ log dim(b) ,
PK |Ak |
4
2
B dim(b) k=1 2 kAk∞ kAk∞
6
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where B > 0. We showed
2 that there
C1 > 0
α 2of Proposition
P in the|Aproof
Pexist
K
|Ak |
2 kAk4 α and
k|
and C2 > 0 such that ( K
)
≥
C
K
≤
1
∞
k=1
k=1
2
2
2
C2 K kAk∞ . As a consequence, there exists C > 0 such that, for all δ > 0,
!
K
X
|Ak | α
P kb(X) − E b(X)k2 ≥ δ
2
k=1
!
δ2 C K
+ log dim(b) .
≤ 2 exp −
4 (2−α)
dim(b) kAk∞
APPENDIX B: PROOFS: FINITE-GRAPH CONCENTRATION RESULTS
FOR MAXIMUM LIKELIHOOD ESTIMATORS
We prove the main concentration results of Section 4, Theorems 1 and 2 along
with Corollaries 1 and 2. Auxiliary lemmas are proved in Appendix E.
Proof of Theorems 1 and 2. Theorem 1 can be considered as a special case
of Theorem 2 with η(θ) = θ and d∞ = max1≤k≤K dim(ηk ) = dim(θ). We
note that Theorem 1 makes the same assumptions as Theorem 2 and satisfies the
additional assumption of Theorem 2, which states that the map η : Θ 7→ int(N)
is one-to-one and that, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ),
there exists γ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ), we have η(θ) ∈ N \
B(η(θ ? ), γ()). We therefore prove Theorem 1 by proving Theorem 2.
By assumption, θ ? ∈ int(Θ). Observe that θ ? ∈ int(Θ) implies µ(η(θ ? )) ∈
rint(M), because the maps η : Θ 7→ int(N) and µ : int(N) 7→ rint(M) are oneto-one: the map η : Θ 7→ int(N) is one-to-one by assumption while the map µ :
int(N) 7→ rint(M) is one-to-one by classic exponential-family theory [4, Theorem
3.6, p. 74].
We are interested in estimators of the form
b
θ =
θ ∈ Θ : ks(x) − µ(η(θ))k2 = inf ks(x) − µ(η(θ̇))k2 ,
θ̇∈Θ
where
Θ ⊆ {θ ∈ Rdim(θ) : ψ(η(θ)) < ∞}.
The following proof covers both full and non-full exponential families, including
curved exponential families. We first discuss the existence of θb in full and non-full,
curved exponential families and then bound the probability of event θb ∈ B(θ ? , ).
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
7
b full and non-full, curved exponential families. For any given
Existence of θ:
ω > 0, let
(
)
K
X
|Ak | α
0
0
?
M(ω) =
µ ∈ M : kµ − µ(η(θ ))k2 < ω
2
k=1
be the subset of µ0 ∈ M close to the data-generating mean-value parameter
vector
P
|Ak | α
.
µ(η(θ ? )) ∈ rint(M) in the sense that kµ0 − µ(η(θ ? ))k2 < ω K
k=1
2
Choose ω > 0 small enough so that M(2 ω) ⊆ rint(M) and let
(
)
K
X
|Ak | α
?
G(ω) =
x ∈ X : ks(x) − µ(η(θ ))k2 < ω
2
k=1
be the subset of x ∈ X such that s(x) ∈ M(ω) ⊆ rint(M). In the following, we
show that in the event G(ω) the set θb is non-empty, i.e., θb exists. To see that, let
M(Θ) be the set of mean-value parameter vectors induced by Θ, i.e.,
M(Θ) = {µ0 ∈ rint(M) : there exists θ ∈ Θ such that µ(η(θ)) = µ0 } .
If the exponential family is full, then M(Θ) = rint(M), otherwise M(Θ) ⊂
rint(M), because non-full exponential families exclude some natural parameter
vectors along with the corresponding mean-value parameter vectors. To show that
the set θb is non-empty in the event G(ω), note that, for all x ∈ G(ω) and hence all
b ∈ M(Θ) ⊆
s(x) ∈ M(ω) ⊆ rint(M), there exists at least one element µ(η(θ))
rint(M) such that
b 2 ≤ ks(x) − µ(η(θ ? ))k2 ,
ks(x) − µ(η(θ))k
because the data-generating mean-value parameter vector µ(η(θ ? )) ∈ M(Θ) ⊆
rint(M) is known to be an element of M(Θ) ⊆ rint(M) and any minimizer of
ks(x) − µ(η(θ))k2 cannot be farther from s(x) ∈ M(ω) ⊆ rint(M) than the
data-generating mean-value parameter vector µ(η(θ ? )) ∈ M(Θ) ⊆ rint(M). In
b ∈ M(Θ) ⊆ rint(M), we have
addition, for each element µ(η(θ))
b − µ(η(θ ? ))k2 ≤ ks(x) − µ(η(θ))k
b 2 + ks(x) − µ(η(θ ? ))k2
kµ(η(θ))
≤ 2 ks(x) − µ(η(θ ? ))k2 ,
b ∈ M(Θ) ⊆ rint(M) is contained in
which implies that each element µ(η(θ))
M(2 ω) ⊆ rint(M); note that ω > 0 was chosen small enough so that M(2 ω) ⊆
rint(M). Therefore, for all x ∈ G(ω), the set θb contains at least one element. If
there exists a unique element θ ∈ Θ such that µ(η(θ)) = s(x) ∈ M(ω) ⊆
8
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
rint(M), then the set θb contains one and only one element. In particular, in full
exponential families with η(θ) = θ, for all x ∈ G(ω), there exists a unique element
θ ∈ Θ such that µ(η(θ)) = s(x) ∈ M(ω) ⊆ rint(M) [4, Theorem 3.6, p. 74].
Therefore, in full exponential families θb exists and is unique in the event G(ω),
whereas in non-full exponential families θb exists but may not be unique in the
event G(ω).
Bounding the probability of event θb ∈ B(θ ? , ): full and non-full, curved
exponential families. By the identifiability conditions of Theorem 2, for all > 0
small enough so that B(θ ? , ) ⊆ int(Θ), there exists γ() > 0 such that, for all
θ ∈ Θ \ B(θ ? , ), we have η(θ) ∈ N \ B(η(θ ? ), γ()). In addition, there exists
δ() > 0 such that, for all η(θ) ∈ N \ B(η(θ ? ), γ()),
K
X
|Ak | α
?
kµ(η(θ)) − µ(η(θ ))k2 ≥ δ()
for some 0 ≤ α ≤ 1.
2
k=1
Therefore,
kµ(η(θ)) −
µ(η(θ ? ))k2
< δ()
K
X
|Ak | α
k=1
2
implies η(θ) ∈ B(η(θ ? ), γ()), which in turn implies θ ∈ B(θ ? , ).
To bound the probability of event θb ∈ B(θ ? , ), we exploit the fact that, for all
x ∈ G(ω) and hence all s(x) ∈ M(ω) ⊆ rint(M), the set θb is non-empty and, for
b
each element of the set θ,
b − µ(η(θ ? ))k2 ≤ ks(x) − µ(η(θ))k
b 2 + ks(x) − µ(η(θ ? ))k2
kµ(η(θ))
≤ 2 ks(x) − µ(η(θ ? ))k2 ,
b ∈ M(2 ω) ⊆ rint(M). Thus, the probability of event
which implies that µ(η(θ))
K
X
|Ak | α
?
b
kµ(η(θ)) − µ(η(θ ))k2 < δ()
∩ G(ω)
2
k=1
can be bounded from below by bounding the probability of event
K
X
|Ak | α
?
2 ks(x) − µ(η(θ ))k2 < δ()
∩ G(ω),
2
k=1
which, by definition of event G(ω), is equivalent to bounding the probability of
event
X
K
δ()
|Ak | α
?
ks(x) − µ(η(θ ))k2 < min
.
, ω
2
2
k=1
9
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
These results show that the probability of event θb ∈ B(θ ? , ) can be bounded from
below as follows:
P θb ∈ B(θ ? , ) ⊆ int(Θ) ≥ P θb ∈ B(θ ? , ) ⊆ int(Θ) ∩ G(ω)
!
α
K
X
|A
|
k
?
b − µ(η(θ ))k2 < δ()
≥ P kµ(η(θ))
∩ G(ω)
2
k=1
X
!
K
δ()
|Ak | α
?
≥ P ks(X) − µ(η(θ ))k2 < min
, ω
.
2
2
k=1
To bound the probability of the event on the right-hand side of the inequality above,
note that the smoothness condition (7) of Theorem 2 ensures that the smoothness
condition (2) of Proposition 2 is satisfied. Therefore, Proposition 2 can be invoked
to show that there exists C > 0 such that
X
α !
K
|A
|
δ()
k
, ω
P ks(X) − µ(η(θ ? ))k2 ≥ min
2
2
k=1
!
κ()2 C K
≤ 2 exp −
+ log d∞ ,
4 (2−α)
d∞ kAk∞
where κ() = min(δ() / 2, ω) > 0. Collecting results shows that
P θb ∈
B(θ ? , )
κ()2 C K
⊆ int(Θ)
≥ 1 − 2 exp −
+ log d∞
4 (2−α)
d∞ kAk∞
!
.
Proof of Corollary 1. Corollary 1 follows from Theorem 1, because conditions (4) and (5) of Theorem 1 are satisfied. Condition (4) is satisfied with α = 1
by Lemma 3 in Appendix E.1 provided |Ak | ≥ 3 (k = 1, . . . , K). Condition
(5) is satisfied, because changing an edge cannot change the number of withinneighborhood edges by more than 1 and the number of within-neighborhood transitive edges by more than 2 (kAk∞ − 2) + 1. Thus, Theorem 1 can be invoked to
conclude that, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exist
κ() > 0, C1 > 0, and C2 > 0 such that
!
2 C K
κ()
1
P θb ∈ B(θ ? , ) ⊆ int(Θ) ≥ 1 − 2 exp −
+ log dim(θ)
4 (2−α)
dim(θ) kAk∞
κ()2 C2 K
= 1 − 4 exp −
,
kAk4∞
10
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where we used the fact that dim(θ) = 2 and α = 1.
Proof of Corollary 2. Corollary 2 follows from Theorem 2, because all conditions of Theorem 2 are satisfied. First, by Lemma 4 in Appendix E.2, the map
η : Θ 7→ int(N) is one-to-one and, for all > 0 small enough so that B(θ ? , ) ⊆
int(Θ), there exists γ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ), we have
η(θ) ∈ N \ B(η(θ ? ), γ()). Second, condition (6) of Theorem 2 is satisfied with
α = 3/4 by Lemma 5 in Appendix E.2 provided |Ak | ≥ 4 (k = 1, . . . , K). Last,
but not least, condition (7) of Theorem 2 is satisfied, because changing an edge
cannot change the number of within-neighborhood edges by more than 1 and the
number of within-neighborhood connected pairs of nodes with i shared partners by
more than 2 (kAk∞ − 2) + 1 (i = 1, . . . , |Ak | − 2, k = 1, . . . , K). Thus, by Theorem 2, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exist κ() > 0
and C > 0 such that
!
2CK
κ()
?
P θb ∈ B(θ , ) ⊆ int(Θ)
≥ 1 − 2 exp −
+ log d∞
4 (2−α)
d∞ kAk∞
κ()2 C K
+ log kAk∞ ,
≥ 1 − 2 exp −
kAk6∞
where we used the fact that d∞ = max1≤k≤K dim(ηk ) = kAk∞ −1 and α = 3/4.
APPENDIX C: PROOFS: FINITE-GRAPH CONCENTRATION RESULTS
FOR M -ESTIMATORS
We prove the main concentration result of Section 5, Theorem 3 along with
Corollary 3.
Proof of Theorem 3. Theorem 3 can be proved along the same lines as Theorems 1 and 2.
By condition [C.1], the expectation E b(X) of b(X) under the data-generating
exponential-family distribution exists. Let B be the convex hull of the set {b(x) :
x ∈ X} and, given any ω > 0, let
(
)
K
X
|Ak | α
0
0
B(ω) =
β ∈ B : kβ − E b(X)k2 < ω
2
k=1
0
be the subset of β ∈ B close to the expectation E b(X) of b(X) under the datagenerating exponential-family distribution in the sense that kβ 0 − E b(X)k2 <
P
|Ak | α
ω K
. Choose ω > 0 small enough so that B(2 ω) ⊆ rint(B) and let
k=1
2
(
)
K
X
|Ak | α
G(ω) =
x ∈ X : kb(x) − E b(X)k2 < ω
2
k=1
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
11
be the subset of x ∈ X such that b(x) ∈ B(ω) ⊆ rint(B).
We are interested in estimators of the form
b
θ =
θ ∈ Θ : kb(x) − β(θ)k2 = inf kb(x) − β(θ̇)k2 ,
θ̇∈Θ
where Θ is an open subset of Rdim(θ) and the expectation β(θ) = Eθ b(X) exists
for all θ ∈ Θ by condition [C.2].
In the following, we do not assume that the family of distributions parameterized
by θ ∈ Θ is an exponential family and we do not assume that the set of expectations {Eθ b(X), θ ∈ Θ} covers the whole set rint(B). In other words, the family
may not be able to match all possible b(x) ∈ rint(B) in the sense that there exists
θ ∈ Θ such that Eθ b(X) = b(x) ∈ rint(B). The critical assumption exploited by
the following proof is that the family can match the expectation E b(X) of b(X)
under the data-generating exponential-family distribution in the sense that there
exists one and only one element θ0 ∈ Θ such that Eθ0 b(X) = E b(X) ∈ rint(B),
along with the fact that b(X) falls with high probability into a small ball centered
at Eθ0 b(X) = E b(X) ∈ rint(B) under suitable conditions. We first discuss the
existence of θb and then bound the probability of event θb ∈ B(θ0 , ).
b We show that in the event G(ω) the set θb is non-empty, i.e., θb
Existence of θ.
exists. To show that the set θb is non-empty in the event G(ω), note that, for all
x ∈ G(ω) and hence all b(x) ∈ B(ω) ⊆ rint(B), there exists at least one element
θ ∈ Θ such that
kb(x) − β(θ)k2 ≤ kb(x) − β(θ0 )k2 = kb(x) − E b(X)k2 ,
because, by condition [C.2], there exists θ0 ∈ Θ such that β(θ0 ) = E b(X) ∈
B(ω) ⊆ rint(B) and any minimizer of kb(x) − β(θ)k2 cannot be farther from
b(x) ∈ B(ω) ⊆ rint(B) than β(θ0 ) = E b(X) ∈ B(ω) ⊆ rint(B). Therefore, the
b we
set θb is non-empty in the event G(ω). In addition, for each element of the set θ,
have
b − E b(X)k2 ≤ kb(x) − β(θ)k
b 2 + kb(x) − E b(X)k2
kβ(θ)
≤ 2 kb(x) − E b(X)k2 ,
b ∈ B(2 ω) ⊆ rint(B); note that ω > 0 was chosen small
which implies that β(θ)
enough so that B(2 ω) ⊆ rint(B). Therefore, for all x ∈ G(ω) and hence all b(x) ∈
B(ω) ⊆ rint(B), the set θb contains at least one element. If the set θb contains
b =
more than one element, then all elements of the set θb map to expectations β(θ)
Eθb b(X) ∈ B(2 ω) ⊆ rint(B) that have the same `2 -distance from b(x) ∈ B(ω) ⊆
rint(B) by construction of estimating function kb(x) − β(θ)k2 .
12
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
Bounding the probability of event θb ∈ B(θ0 , ). To bound the probability of
event θb ∈ B(θ0 , ), note that, by condition [C.2], the expectation β(θ) = Eθ b(X)
exists for all θ ∈ Θ and, by condition [C.3], for all > 0 small enough so that
B(θ0 , ) ⊆ Θ, there exists δ() > 0 such that, for all θ ∈ Θ \ B(θ0 , ),
kβ(θ) − β(θ0 )k2 = kβ(θ) − E b(X)k2 ≥ δ()
K
X
|Ak | α
k=1
2
for some 0 ≤ α ≤ 1. Therefore,
kβ(θ) − β(θ0 )k2 = kβ(θ) − E b(X)k2 < δ()
K
X
|Ak | α
k=1
2
implies θ ∈ B(θ0 , ).
To bound the probability of event θb ∈ B(θ0 , ), we exploit the fact that, for all
x ∈ G(ω) and hence all b(x) ∈ B(ω) ⊆ rint(B), the set θb is non-empty and, for
b
each element of the set θ,
b − E b(X)k2 ≤ kb(x) − β(θ)k
b 2 + kb(x) − E b(X)k2
kβ(θ)
≤ 2 kb(x) − E b(X)k2 ,
b ∈ B(2 ω) ⊆ rint(B), as pointed out above. Thus, the
which implies that β(θ)
probability of event
b − E b(X)k2 < δ()
kβ(θ)
K
X
|Ak | α
k=1
2
∩ G(ω)
can be bounded from below by bounding the probability of event
2 kb(x) − E b(X)k2 < δ()
K
X
|Ak | α
k=1
2
∩ G(ω),
which, by definition of event G(ω), is equivalent to bounding the probability of
event
X
K
|Ak | α
δ()
,ω
.
kb(x) − E b(X)k2 < min
2
2
k=1
These results show that the probability of event θb ∈ B(θ0 , ) can be bounded from
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
13
below as follows:
P θb ∈ B(θ0 , ) ⊆ Θ ≥ P θb ∈ B(θ0 , ) ⊆ Θ ∩ G(ω)
!
α
K
X
|A
|
k
b − E b(X)k2 < δ()
≥ P kβ(θ)
∩ G(ω)
2
k=1
X
!
K
δ()
|Ak | α
≥ P kb(X) − E b(X)k2 < min
,ω
.
2
2
k=1
To bound the probability of the event on the right-hand side of the inequality above,
Proposition 3 can be invoked to show that there exists C > 0 such that
X
!
K
δ()
|Ak | α
P kb(X) − E b(X)k2 ≥ min
,ω
2
2
k=1
!
κ()2 C K
≤ 2 exp −
+ log dim(b) ,
4 (2−α)
dim(b) kAk∞
where κ() = min(δ() / 2, ω) > 0. Collecting results shows that
P θb ∈ B(θ0 , ) ⊆ Θ
≥ 1 − 2 exp −
κ()2 C K
4 (2−α)
!
+ log dim(b) .
dim(b) kAk∞
Proof of Corollary 3. Corollary 3 follows from Theorem 3, because all conditions of Theorem 3 are satisfied. Condition [C.1] is satisfied, because E b(X) ∈
rint(B) follows from θ ? ∈ int(Θ? ) and because changing an edge cannot change
the number of within-neighborhood edges by more than 1 and the number of
within-neighborhood transitive edges by more than 2 (kAk∞ − 2) + 1. Conditions [C.2] and [C.3] follow from classic exponential-family theory [4], because
θ ∈ Θ0 is the natural parameter vector and β(θ) ∈ rint(B) is the mean-value
parameter vector of the misspecified exponential family. Condition [C.2] is satisfied, because the parameter space Θ0 = R2 is open, β(θ) = Eθ b(X) exists for all θ ∈ Θ0 = R2 by classic exponential-family theory [4, Theorem
2.2, pp. 34–35], and there exists a unique element θ0 ∈ Θ0 = R2 such that
β(θ0 ) = Eθ0 b(X) = E b(X), because the map β : int(Θ0 ) 7→ rint(B) is one-toone by classic exponential-family theory [e.g., Theorem 3.6, 4, p. 74]. Condition
[C.3] is satisfied with α = 1 by Lemma 6 in Appendix E.3 provided |Ak | ≥ 3
(k = 1, . . . , K). Since all conditions of Theorem 3 are satisfied, we can invoke
Theorem 3 to conclude that, for all > 0 small enough so that B(θ0 , ) ⊆ Θ0 ,
14
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
there exist κ() > 0, C1 > 0, and C2 > 0 such that the M -estimator θb exists, is
unique, and is contained in B(θ0 , ) with probability
P θb ∈ B(θ0 , ) ⊆ Θ0
≥ 1 − 2 exp −
κ()2 C1 K
4 (2−α)
!
+ log dim(b)
dim(b) kAk∞
κ()2 C2 K
= 1 − 4 exp −
,
kAk4∞
where we used the fact that dim(b) = 2 and α = 1. Observe that uniqueness
follows from the fact that there is a unique θ ∈ Θ0 = R2 such that Eθ b(X) =
b(x) ∈ rint(B) for all possible b(x) ∈ rint(B), because θ ∈ Θ0 = R2 is the
natural parameter vector and β(θ) ∈ rint(B) is the mean-value parameter vector of
the misspecified exponential family and therefore the map β : int(Θ0 ) 7→ rint(B)
is one-to-one [e.g., Theorem 3.6, 4, p. 74].
APPENDIX D: PROOFS: FINITE-GRAPH CONCENTRATION RESULTS
FOR SUBGRAPH-TO-GRAPH ESTIMATORS
We prove the main concentration result of Section 6, Theorem 4 along with
Proposition 4 and Corollary 4.
Proof of Proposition 4. By the factorization properties implied by local dependence, we have, for all θ ∈ Θ ⊆ {θ ∈ Rdim(θ) : ψL (η(θ)) < ∞}, all xL ∈ XL ,
and all K ⊆ L,
X
pη(θ) (xL )
xL\K ∈XL\K
=
X
exp (hη(θ), s(xL )i − ψL (η(θ))) νL (xL )
xL\K ∈XL\K
(13)
=
X
Y
exp (hη(θ), s(xA )i − ψA (η(θ))) νA (xA )
xL\K ∈XL\K A∈L
=
Y
exp (hη(θ), s(xA )i − ψA (η(θ))) νA (xA )
A∈K
= exp (hη(θ), s(xK )i − ψK (η(θ))) νK (xK )
= pη(θ) (xK ),
where we exploited the fact that, by local dependence, νL (xK ) satisfies
Y
νL (xL ) =
νA (xA ),
A∈L
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
15
while ψL (η(θ)) satisfies
ψL (η(θ)) =
X
ψA (η(θ))
A∈L
and
ψA (η(θ)) =
X
exp (hη(θ), s(xA )i) νA (xA ), A ∈ L.
xA ∈XA
We note that the natural parameter vector η(θ) takes the form
η(θ) = (η1 (θ), . . . , ηd∞ (θ)),
while the sufficient statistic vector takes the form
s(xK ) = (s1 (xK ), . . . , sd∞ (xK )),
where the sufficient statistics si (xK ) are based on the sufficient statistics sA,i (xA )
of neighborhoods A ∈ K:
X
si (xK ) =
sA,i (xA ), i = 1, . . . , d∞ .
A∈K
In addition, note that
νK (xK ) =
Y
νA (xA )
A∈K
and that
ψK (η(θ)) =
X
ψA (η(θ))
A∈K
P
is finite provided ψL (η(θ)) = A∈L ψA (η(θ)) is finite.
To prove that, for all xK ∈ XK and all yK ∈ YK ,
Pη(θ) (XK = xK , YK = yK , XL\K ∈ XL\K , YL\K ∈ YL\K )
= Pη(θ) (XK = xK , YK = yK ),
observe that the independence of within- and between-neighborhood subgraphs
implies that
(14)
Pη(θ) (XK = xK , YK = yK , XL\K ∈ XL\K , YL\K ∈ YL\K )
= Pη(θ) (XK = xK , XL\K ∈ XL\K ) P(YK = yK , YL\K ∈ YL\K ).
The first term on the right-hand side of (14) can be dealt with by using (13), which
implies that
Pη(θ) (XK = xK , XL\K ∈ XL\K ) = Pη(θ) (XK = xK ).
16
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
The second term on the right-hand side of (14) can be dealt with by using the
assumption that the between-neighborhood subgraphs are independent:
P(YK = yK , YL\K ∈ YL\K ) = P(YK = yK ) P(YL\K ∈ YL\K ) = P(YK = yK ).
Collecting results gives
Pη(θ) (XK = xK , YK = yK , XL\K ∈ XL\K , YL\K ∈ YL\K )
= Pη(θ) (XK = xK ) P(YK = yK ) = Pη(θ) (XK = xK , YK = yK ).
Proof of Theorem 4. By Proposition 4, we have, for all θ ∈ Θ ⊆ {θ ∈
: ψL (η(θ)) < ∞}, all K ⊆ L, and all xK ∈ XK and all yK ∈ YK ,
Rdim(θ)
Pη(θ) (XK = xK , YK = yK , XL\K ∈ XL\K , YL\K ∈ YL\K )
= Pη(θ) (XK = xK , YK = yK )
and, for all xL ∈ XL ,
X
pη(θ) (xL ) = pη(θ) (xK ),
xL\K ∈ XL\K
where
pη(θ) (xK ) = exp (hη(θ), s(xK )i − ψK (η(θ))) νK (xK ).
Here, η(θ) and s(xK ) have dimension d∞ = maxA∈L dim(ηA ). We note that
ψK (η(θ)) satisfies
X
ψK (η(θ)) =
ψA (η(θ)),
A∈K
where
ψA (η(θ)) =
X
exp (hη(θ), s(xA )i) νA (xA ), A ∈ K.
xA ∈XA
P
In
addition,
note
that
ψ
(η(θ))
=
K
A∈K ψA (η(θ)) is finite provided ψL (η(θ)) =
P
A∈L ψA (η(θ)) is finite. In other words, local dependence implies that XK is independent of XL\K and the marginal density of xK ∈ XK induced by K ⊆ L is
an exponential-family density with support XK and local dependence, with natural
parameter vector η(θ) and sufficient statistic vector s(xK ). As a result, Theorems
1 and 2 can be applied to the exponential family with support XK and local dependence, with natural parameter vector η(θ) and sufficient statistic vector s(xK ).
Observe that the conditions of Theorem 4 ensure that all conditions of Theorems 1
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
17
and 2 are satisfied for all K ⊆ L. Therefore, for full exponential families, Theorem
1 shows that, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exist
κ1 () > 0 and C1 > 0 such that θbK exists, is unique, and is contained in B(θ ? , )
with probability
P θbK ∈ B(θ ? , ) ⊆ int(Θ) ≥ 1 − 2 exp −
κ1 ()2 C1 |K|
4 (2−α)
!
+ log dim(θ) ,
dim(θ) kKk∞
where kKk∞ = maxA∈K |A|. For non-full, curved exponential families, Theorem
2 shows that, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exist
κ2 () > 0 and C2 > 0 such that θbK exists and is contained in B(θ ? , ) with
probability
!
κ2 ()2 C2 |K|
?
b
P θK ∈ B(θ , ) ⊆ int(Θ) ≥ 1 − 2 exp −
+ log d∞ .
4 (2−α)
d∞ kKk∞
Thus, there exist κ() = min(κ1 (), κ2 ()) > 0 and C = min(C1 , C2 ) > 0 such
that
!
2 C |K|
κ()
+ log d∞
P θbK ∈ B(θ ? , ) ⊆ int(Θ)
≥ 1 − 2 exp −
4 (2−α)
d∞ kKk∞
!
κ()2 C |K|
+ log d∞ ,
≥ 1 − 2 exp −
4 (2−α)
d∞ kLk∞
where we used the fact that kKk∞ = maxA∈K |A| ≤ maxA∈L |A| = kLk∞
and d∞ = maxA∈L dim(ηA ) = dim(θ) in full exponential families with natural
parameter vectors of the form η(θ) = θ.
Proof of Corollary 4. Corollary 4 follows from Theorem 4, because the conditions of Theorem 4 are satisfied. The proof of Corollary 4 resembles the proof of
Corollary 2 and is therefore omitted.
APPENDIX E: PROOFS: AUXILIARY LEMMAS
We prove auxiliary lemmas that are useful for verifying the conditions of Theorem 1 (Appendix E.1), Theorem 2 (Appendix E.2), and Theorem 3 (Appendix E.3).
We start with two lemmas that are useful for bounding expectations of sufficient
statistics. Throughout, we write
Λ(η) =
min
min
i∈Ak < j∈Ak , k=1,...,K x−i,j ∈ X−i,j
Pη (Xi,j = 1 | X−i,j = x−i,j )
18
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
and
Ω(η) =
max
max
i∈Ak < j∈Ak , k=1,...,K x−i,j ∈ X−i,j
Pη (Xi,j = 1 | X−i,j = x−i,j ),
where η is the natural parameter vector of the exponential family under consideration and x−i,j ∈ X−i,j corresponds to x ∈ X excluding xi,j ∈ Xi,j (i ∈ Ak <
j ∈ Ak , k = 1, . . . , K). Let f : X 7→ R be a function of the random graph. We
denote by EΛ(η) f (X) and EΩ(η) f (X) the expectations of f : X 7→ R under the
Bernoulli random graph model with probabilities Λ(θ) and Ω(θ), respectively, provided EΛ(η) f (X) and EΩ(η) f (X) exist. Here, the Bernoulli random graph model
refers to the exponential-family random graph model which assumes that the edge
variables Xi,j are independent Bernoulli random variables with probabilities Λ(θ)
and Ω(θ), respectively.
Lemma 1. Consider a full or non-full, canonical or curved exponential family with
PK
|Ak |
support X = {0, 1} k=1 ( 2 ) and local dependence and natural parameter vector
η ∈ int(N). Then there exists C(η) ∈ [Λ(η), Ω(η)] such that
K
K
X
X
X
|Ak |
,
Eη
Xi,j = C(η)
2
k=1 i∈Ak < j∈Ak
k=1
where C(η) ∈ [Λ(η), Ω(η)] denotes the probability of an edge under the
Bernoulli(C(η)) random graph model. In addition, if f : X 7→ R is a function
of the random graph that is non-decreasing in the number of edges, then
EΛ(η) f (X) ≤ Eη f (X) ≤ EΩ(η) f (X),
where EΛ(η) f (X) and EΩ(η) f (X) are the expectations of f : X 7→ R under the
Bernoulli random graph model with probabilities Λ(θ) and Ω(θ), respectively. If
f : X 7→ R is a function of the random graph that is non-increasing in the number
of edges, then the inequalities are reversed.
Proof of Lemma 1. To ease the presentation, we consider K = 1 neighborhood and drop the subscript k from all neighborhood-dependent quantities. The
extension to K ≥ 2 neighborhoods is straightforward. Observe that Eη Xi,j =
Pη (Xi,j = 1) can be written as
X
Pη (Xi,j = 1) =
Pη (Xi,j = 1 | X−i,j = x−i,j ) Pη (X−i,j = x−i,j ),
x−i,j ∈ X−i,j
which implies that Eη Xi,j is bounded below by
X
Λ(η) Pη (X−i,j = x−i,j ) = Λ(η)
x−i,j ∈ X−i,j
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
and bounded above by
X
19
Ω(η) Pη (X−i,j = x−i,j ) = Ω(η).
x−i,j ∈ X−i,j
Since the expectation Eη Xi,j is contained in the convex set [Λ(η), Ω(η)], there
exists C(η) ∈ [Λ(η), Ω(η)] such that
X
X
|A|
Eη
Xi,j =
Eη Xi,j = C(η)
.
2
i∈A < j∈A
i∈A < j∈A
In addition, Butts [5, Corollary 1, p. 306] proved that, if f : X 7→ R is a function
of the random graph that is non-decreasing in the number of edges, then
EΛ(η) f (X) ≤ Eη f (X) ≤ EΩ(η) f (X),
where the inequalities are reversed when f : X 7→ R is a function of the random
graph that is non-increasing in the number of edges. The result follows from the
recursive factorization of the probability mass function Pη (X = x). Denote the elements of the sequence of within-neighborhood edge variables X by X 1 , . . . , Xm
and the corresponding sample spaces by X1 , . . . , Xm , where m = |A|
2 . Then, for
all x ∈ X, we have
Pη (X = x) =
m
Y
Pη (Xi = xi | Xj = xj , j = 1, . . . , i − 1),
i=1
where the conditional probabilities Pη (Xi = 1 | Xj = xj , j = 1, . . . , i − 1) are
bounded by
Λ(η) =
min
min
1 ≤ i ≤ m x−i ∈ X−i
Pη (Xi = 1 | X−i = x−i )
≤ Pη (Xi = 1 | Xj = xj , j = 1, . . . , i − 1)
≤
max
max Pη (Xi = 1 | X−i = x−i )
1 ≤ i ≤ m x−i ∈ X−i
= Ω(η),
where x−i ∈ X−i corresponds to x ∈ X excluding xi ∈ Xi (i = 1, . . . , m).
Therefore, Pη (X = x) can be bounded below and above by the probability mass
function of the Bernoulli random graph model with probabilities Λ(η) and Ω(η),
respectively. As a result, the cumulative distribution function and expectation of
any function f : X 7→ R of the random graph that is non-decreasing in the number
of edges can be bounded by the corresponding cumulative distribution functions
and expectations of the Bernoulli random graph model with probabilities Λ(η) and
Ω(η) [5, Corollary 1, p. 306]:
EΛ(η) f (X) ≤ Eη f (X) ≤ EΩ(η) f (X),
20
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where the inequalities are reversed when f : X 7→ R is a function of the random
graph that is non-increasing in the number of edges.
Lemma 2. Consider a full or non-full, canonical or curved exponential family with
PK
|Ak |
support X = {0, 1} k=1 ( 2 ) and local dependence and natural parameter vector
η ∈ int(N). Let f : X 7→ R be the number of within-neighborhood transitive edges
defined by
f (x) =
K
X
X
xi,j
k=1 i∈Ak < j∈Ak
max
h∈Ak , h 6= i,j
xi,h xj,h , x ∈ X,
which may or may not be a sufficient statistic of the exponential family under consideration. If |Ak | ≥ 3 (k = 1, . . . , K), then there exists C(η) > 0 such that
K
X
|Ak |
,
Eη f (X) = C(η)
2
k=1
where C(η) satisfies, for some λ ∈ (0, 1),
0 < λ Λ(η)3 + (1 − λ) Ω(η)3 ≤ C(η) ≤ λ Λ(η) + (1 − λ) Ω(η) < 1.
Proof of Lemma 2. To ease the presentation, we consider K = 1 neighborhood and drop the subscript k from all neighborhood-dependent quantities. The
extension to K ≥ 2 neighborhoods is straightforward. Since the number of transitive edges f : X 7→ R is a function of the random graph that is non-decreasing
in the number of edges, Lemma 1 implies that the expectation Eη f (X) can be
bounded by the expectations EΛ(η) f (X) and EΩ(η) f (X) under the Bernoulli random graph model with edge probabilities Λ(η) and Ω(η), respectively:
EΛ(η) f (X) ≤ Eη f (X) ≤ EΩ(η) f (X),
where Λ(η) and Ω(η) are defined in the introduction of Appendix E and all expectations exist, because the sample space X is finite. The expectation EΛ(η) f (X)
can be written as
X
EΛ(η) f (X) = EΛ(η)
Xi,j
max Xi,h Xj,h
h∈A, h 6= i,j
i∈A < j∈A
= EΛ(η)
X
Xi,j
1
=
i∈A < j∈A
EΛ(η) Xi,j
Xi,h Xj,h ≥ 1
h∈A, h 6= i,j
i∈A < j∈A
X
X
1
X
h∈A, h 6= i,j
Xi,h Xj,h ≥ 1 ,
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
21
P
where 1( h∈A, h 6= i,j xi,h xj,h ≥ 1) is an indicator function, which is 1 if nodes
i and j have one or more shared partners in neighborhood A and is 0 otherwise.
Here, we exploited the fact that the number of transitive edges in neighborhood A
is equal to the number of pairs of nodes with one
P or more edgewise shared partners. To evaluate the expectation EΛ(η) [Xi,j 1( h∈A, h 6= i,j Xi,h Xj,h ≥ 1)], we
take advantage of the independence of edge variables under the Bernoulli(Λ(η))
random graph model, which implies that
X
EΛ(η) Xi,j 1
Xi,h Xj,h ≥ 1
h∈A, h 6= i,j
= EΛ(η) (Xi,j ) EΛ(η) 1
X
Xi,h Xj,h ≥ 1
h∈A, h 6= i,j
= Λ(η) PΛ(η)
X
Xi,h Xj,h ≥ 1 .
h∈A, h 6= i,j
In addition, the independence of edge variables under the Bernoulli(Λ(η)) random graph model
P implies that, for all i ∈ A and all j ∈ A such that i 6= j, the
distribution of h∈A, h 6= i,j Xi,h Xj,h is Binomial(|A| − 2, Λ(η)2 ). Therefore,
X
PΛ(η)
Xi,h Xj,h ≥ 1 = 1 − (1 − Λ(η)2 )|A|−2 ,
h∈A, h 6= i,j
which implies that
EΛ(η) f (X) = Λ(η) [1 − (1 −
Λ(η)2 )|A|−2 ]
|A|
.
2
Along the same lines, it can be shown that the expectation EΩ(η) f (X) is given by
|A|
EΩ(η) f (X) = Ω(η) [1 − (1 − Ω(η)2 )|A|−2 ]
.
2
Let C1 (η) = Λ(η) [1−(1−Λ(η)2 )|A|−2 ] and C2 (η) = Ω(η) [1−(1−Ω(η)2 )|A|−2 ].
Observe that Λ(η) ∈ (0, 1) implies that C1 (η) is bounded below by Λ(η)3 and
bounded above by Λ(η) for all |A| ≥ 3. Likewise, C2 (η) is bounded below by
Ω(η)3 and bounded above by Ω(η) for all |A| ≥ 3. Since the expectation Eη f (X)
|A|
is contained in the convex set [C1 (η) |A|
2 , C2 (η) 2 ], there exists λ ∈ (0, 1)
such that
|A|
|A|
|A|
Eη f (X) = λ C1 (η)
+ (1 − λ) C2 (η)
= C(η)
,
2
2
2
22
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where C(η) satisfies
0 < λ Λ(η)3 + (1 − λ) Ω(η)3 ≤ C(η) ≤ λ Λ(η) + (1 − λ) Ω(η) < 1.
E.1. Auxiliary lemmas: Theorem 1. We prove auxiliary lemmas that are useful for verifying the conditions of Theorem 1.
Lemma 3. Consider an exponential-family random graph of fixed and finite size
with within-neighborhood edge and transitive edge terms as defined in Section 4.1.
Let Θ = R2 and assume that θ ? ∈ int(Θ). Then, for all > 0 small enough so
that B(θ ? , ) ⊆ int(Θ), there exists δ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ),
kµ(η(θ ? ))
− µ(η(θ))k2
K
X
|Ak |
≥ δ()
,
2
k=1
provided |Ak | ≥ 3 (k = 1, . . . , K). Therefore, identifiability condition (4) of Theorem 1 is satisfied with α = 1 provided |Ak | ≥ 3 (k = 1, . . . , K).
Proof of Lemma 3. To ease the presentation, we consider K = 1 neighborhood and drop the subscript k from all neighborhood-dependent quantities. The
extension to K ≥ 2 neighborhoods is straightforward. In the following, we write
µ(η(θ)) = µ(θ) = Eθ s(X) and ψ(η(θ)) = ψ(θ), because the natural parameter
vector of the exponential family is given by η(θ) = θ ∈ Θ = R2 .
To verify identifiability condition (4) of Theorem 1, we need to verify that, for
all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exists δ() > 0 such that,
for all θ ∈ Θ \ B(θ ? , ),
α
|A|
?
kµ(θ ) − µ(θ)k2 ≥ δ()
for some 0 ≤ α ≤ 1.
2
To do so, observe that
kµ(θ ? ) − µ(θ)k2 =
p
(µ1 (θ ? ) − µ1 (θ))2 + (µ2 (θ ? ) − µ2 (θ))2 .
The deviation |µ1 (θ ? ) − µ1 (θ)| can be dealt with by using Lemma 1, which shows
that there exists C1 (θ) ∈ [Λ(θ), Ω(θ)] such that
X
|A|
µ1 (θ) = Eθ
Xi,j = C1 (θ)
,
2
i∈A < j∈A
where Λ(θ) and Ω(θ) are defined in the introduction of Appendix E. Here, Λ(θ)
is given by
• Case θ ∈ R × R− : Λ(θ) = [1 + exp(−θ1 − θ2 (2 |A| − 3))]−1 > 0;
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
23
• Case θ ∈ R × R+ ∪ {0}: Λ(θ) = [1 + exp(−θ1 )]−1 > 0;
while Ω(θ) satisfies 0 < Λ(θ) ≤ Ω(θ) < 1 for all θ ∈ R2 ; note that |A| is
fixed and finite since the size of the random graph is fixed and finite. Therefore, the
deviation |µ1 (θ ? ) − µ1 (θ)| satisfies
|A|
?
?
|µ1 (θ ) − µ1 (θ)| = |C1 (θ ) − C1 (θ)|
.
2
To deal with the deviation |µ2 (θ ? )−µ2 (θ)|, observe that, by Lemma 2, there exists
C2 (θ) > 0 such that
X
|A|
µ2 (θ) = Eθ
Xi,j max Xi,h Xj,h = C2 (θ)
,
h∈A, h6=i,j
2
i∈A < j∈A
where C2 (θ) satisfies
λ Λ(θ)3 + (1 − λ) Ω(θ)3 ≤ C2 (θ) ≤ λ Λ(θ) + (1 − λ) Ω(θ), 0 < λ < 1.
As a result,
|µ2
(θ ? )
− µ2 (θ)| = |C2
(θ ? )
|A|
− C2 (θ)|
.
2
We can hence conclude that, for all θ ∈ Θ \ B(θ ? , ), there exists a function
C(θ, θ ? ) > 0 such that
|A|
?
?
kµ(θ ) − µ(θ)k2 = C(θ , θ)
> 0,
2
where strict positivity follows from the fact that the map µ : Θ 7→ rint(M) is oneto-one by classic exponential-family theory [4, Theorem 3.6, p. 74], which implies
that kµ(θ ? ) − µ(θ)k2 > 0.
Last, but not least, we show that, for all > 0 small enough so that B(θ ? , ) ⊆
int(Θ), there exists δ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ), we have kµ(θ ? ) −
µ(θ)k2 ≥ δ() |A|
> 0. To do so, note that η(θ) = θ is the natural parameter
2
vector of the exponential family. Thus, by classic exponential-family theory, ψ(θ)
is a strictly convex function of the natural parameter vector θ on the interior of the
natural parameter space {θ ∈ Rdim(θ) : ψ(θ) < ∞} = R2 [4, Theorem 1.13, p.
19]. Therefore, the gradient ∇θ ψ(θ) = µ(θ) is a strictly increasing function of θ
and the smallest value of kµ(θ ? ) − µ(θ)k2 , considered as a function of θ for fixed
θ ? ∈ int(Θ), is attained on the boundary of ball B(θ ? , ) ⊆ int(Θ).
24
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
As a result, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exists
δ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ),
|A|
?
kµ(θ ) − µ(θ)k2 ≥ δ()
,
2
provided |A| ≥ 3. Therefore, identifiability condition (4) of Theorem 1 is satisfied
with α = 1 provided |A| ≥ 3.
E.2. Auxiliary lemmas: Theorem 2. We prove auxiliary lemmas that are useful for verifying the conditions of Theorem 2.
Lemma 4. Consider a curved exponential-family random graph of fixed and finite
size with within-neighborhood edge and geometrically weighted edgewise shared
partner terms as defined in Section 4.2. Let Θ = R×(0, 1). Then the map η : Θ 7→
int(N) is one-to-one and, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ),
there exists γ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ), we have η(θ) ∈ N \
B(η(θ ? ), γ()) provided |Ak | ≥ 4 (k = 1, . . . , K).
Proof of Lemma 4. It is straightforward to show that the map η : Θ 7→ int(N)
is one-to-one and that, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there
exists γ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ),
v
u d∞
uX
?
kη(θ) − η(θ )k2 = t (ηi (θ) − ηi (θ ? ))2 ≥ γ().
i=1
To see that, note that d∞ = kAk∞ − 1 ≥ 3 since |Ak | ≥ 4 (k = 1, . . . , K).
Therefore,
v
v
u d∞
u 3
uX
uX
?
2
kη(θ) − η(θ ? )k2 = t (ηi (θ) − ηi (θ )) ≥ t (ηi (θ) − ηi (θ ? ))2 ,
i=1
i=1
where
η1 (θ) = θ1
η2 (θ) = 1
η3 (θ) = 2 − θ2 .
Thus, for all θ ?
∈ int(Θ) = R×(0, 1) and all δ ∈ R2 such that θ ? +δ ∈ int(Θ) =
R × (0, 1),
(η1 (θ ? + δ) − η1 (θ ? ))2 = (θ1? + δ1 − θ1? )2 = δ12
(η2 (θ ? + δ) − η2 (θ ? ))2 = 0
(η3 (θ ? + δ) − η3 (θ ? ))2 = (2 − θ2? − δ2 − 2 + θ2? )2 = δ22 ,
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
25
which implies that
v
u 3
uX
p
?
?
kη(θ + δ) − η(θ )k2 ≥ t (ηi (θ ? + δ) − ηi (θ ? ))2 = δ12 + δ22 = kδk2 .
i=1
Therefore, the `2 -distance of η(θ ? + δ) from η(θ ? ) ∈ int(N) in the natural parameter space N is a strictly increasing function of the `2 -distance = kδk2 of θ ? + δ
from θ ? ∈ int(Θ) in the parameter space Θ. As a result, for all > 0, there exists
γ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ),
v
u d∞
uX
?
kη(θ ) − η(θ)k2 = t (ηi (θ ? ) − ηi (θ))2 ≥ γ().
i=1
Lemma 5. Consider a curved exponential-family random graph of fixed and finite
size with within-neighborhood edge and geometrically weighted edgewise shared
partner terms as defined in Section 4.2. Let Θ = R × (0, 1) and assume that θ ? ∈
int(Θ). Then, for all > 0 small enough so that B(θ ? , ) ⊆ int(Θ), there exists
γ() > 0 such that, for all θ ∈ Θ \ B(θ ? , ), we have η(θ) ∈ N \ B(η(θ ? ), γ()).
In addition, there exists δ() > 0 such that, for all η(θ) ∈ N \ B(η(θ ? ), γ()),
kµ(η(θ ? ))
− µ(η(θ))k2 ≥ δ()
K
X
|Ak | 3/4
k=1
2
,
provided |Ak | ≥ 4 (k = 1, . . . , K). Therefore, identifiability condition (6) of Theorem 2 is satisfied with α = 3/4 provided |Ak | ≥ 4 (k = 1, . . . , K).
Proof of Lemma 5. To ease the presentation, we consider K = 1 neighborhood and drop the subscript k from all neighborhood-dependent quantities. The
extension to K ≥ 2 neighborhoods is straightforward. By Lemma 4, for all > 0
small enough so that B(θ ? , ) ⊆ int(Θ), there exists γ() > 0 such that, for all
θ ∈ Θ \ B(θ ? , ), we have η(θ) ∈ N \ B(η(θ ? ), γ()). Therefore, it is enough to
show that there exists δ() > 0 such that, for all η(θ) ∈ N \ B(η(θ ? ), γ()),
α
|A|
for some 0 ≤ α ≤ 1.
kµ(η(θ ? )) − µ(η(θ))k2 ≥ δ()
2
To do so, observe that
kµ(η(θ ? )) − µ(η(θ))k2
v
u
|A|−1
u
X
= t(µ1 (η(θ ? )) − µ1 (η(θ)))2 +
(µi (η(θ ? )) − µi (η(θ)))2 > 0,
i=2
26
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
where strict positivity follows from the fact that the maps η : Θ 7→ int(N) and
µ : int(N) 7→ rint(M) are one-to-one and hence θ 6= θ ? implies kµ(η(θ ? )) −
µ(η(θ))k2 > 0; note that the map η : Θ 7→ int(N) is one-to-one by Lemma 4
while the map µ : int(N) 7→ rint(M) is one-to-one by classic exponential-family
theory [4, Theorem 3.6, p. 74].
We distinguish two cases, the case µ1 (η(θ ? )) 6= µ1 (η(θ)) and the case
µ1 (η(θ ? )) = µ1 (η(θ)). We note that θ 6= θ ? and µ1 (η(θ ? )) = µ1 (η(θ)) imP|A|−1
ply i=2 (µi (η(θ ? )) − µi (η(θ)))2 > 0, because θ 6= θ ? implies kµ(η(θ ? )) −
µ(η(θ))k2 > 0.
Case µ1 (η(θ ? )) 6= µ1 (η(θ)). We have
kµ(η(θ ? )) − µ(η(θ))k2 ≥ |µ1 (η(θ ? )) − µ1 (η(θ))| > 0.
Observe that µ1 (η(θ)) = Eη(θ) s1 (X) is the expected number of edges in neighborhood A. Therefore, Lemma 1 can be invoked to show that there exist C1 (η(θ ? )) >
0 and C1 (η(θ)) > 0 such that
|A|
?
?
|µ1 (η(θ )) − µ1 (η(θ))| = |C1 (η(θ )) − C1 (η(θ))|
,
2
where Λ(η(θ ? )) ≤ C1 (η(θ ? )) ≤ Ω(η(θ ? )) and Λ(η(θ)) ≤ C1 (η(θ)) ≤ Ω(η(θ))
for all θ ∈ int(Θ) = R × (0, 1). We note that Λ(η(θ)) and Ω(η(θ)) are defined in
the introduction of Appendix E and that Λ(η(θ)) ≥ [1 + exp(−θ1 )]−1 > 0 while
Ω(η(θ)) satisfies 0 < Λ(η(θ)) ≤ Ω(η(θ)) < 1 for all θ ∈ int(Θ) = R × (0, 1).
Case µ1 (η(θ ? )) = µ1 (η(θ)). As pointed out above, θ 6= θ ? and µ1 (η(θ ? )) =
P|A|−1
?
2
µ1 (η(θ)) imply
>
0, where
i=2 (µi (η(θ )) − µi (η(θ)))
µi (η(θ)) = Eη(θ) si (X) is the expected number of pairs of nodes with i − 1
edgewise shared partners in neighborhood A (i = 2, . . . , |A| − 1). Bounding the
P|A|−1
term i=2 (µi (η(θ ? )) − µi (η(θ)))2 is more challenging than bounding the term
(µ1 (η(θ ? )) − µ1 (η(θ)))2 , because the numbers of pairs of nodes with i − 1 edgewise shared partners are neither non-decreasing nor non-increasing functions of
the number of edges (i = 2, . . . , |A| − 1). Therefore, Lemma 1 cannot be applied
to the expectations µi (η(θ)) = Eη(θ) si (X) (i = 2, . . . , |A| − 1). But it turns out
P|A|−1
to be possible to bound i=2 (µi (η(θ ? )) − µi (η(θ)))2 from below in terms of
absolute deviations of the expected numbers of transitive edges under η(θ ? ) and
η(θ). The advantage of doing so is that the number of transitive edges is a nondecreasing function of the number of edges and hence Lemma 1 can be applied
via Lemma 2. To see that the expected numbers of pairs of nodes with one or more
edgewise shared partners is related to the expected number of transitive edges, note
27
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
that the expected number of transitive edges in neighborhood A can be written as
X
Eη(θ)
Xa,b
max Xa,c Xb,c
c∈A, c 6= a,b
a∈A < b∈A
|A|−1
= Eη(θ)
X
X
i=2
a∈A < b∈A
Xa,b
=
i=2
Xa,c Xb,c = i − 1
c∈A, c 6= a,b
|A|−1
X
1
X
X
Eη(θ)
Xa,b
a∈A < b∈A
1
X
Xa,c Xb,c = i − 1
c∈A, c 6= a,b
|A|−1
=
X
µi (η(θ)),
i=2
P|A|−1
which shows that i=2 µi (η(θ)) is equal to the expected number of transiP|A|−1
tive edges under η(θ). To bound i=2 (µi (η(θ ? )) − µi (η(θ)))2 from below
in terms of absolute deviations of the expected numbers of transitive edges under η(θ ? ) and η(θ), it is convenient to operate in the natural parameter space
N = R|A|−1 of the full exponential family with natural parameter vector η and
sufficient statistic vector s(x). Write η ≡ η(θ) and η ? ≡ η(θ ? ). Then, to bound
P|A|−1
the term i=2 (µi (η ? ) − µi (η))2 from below in terms of absolute deviations of
the expected numbers of transitive edges under η ? and η while avoiding the trivial
lower bound of 0, we note that it is possible to choose η̇ ∈ int(N) such that
kµ(η ? ) − µ(η)k2 ≥ kµ(η ? ) − µ(η̇)k2
and such that the expected number of transitive edges under η̇ is not identical to
the expected number of transitive edges under η ? . To see that, note that
|A|−1
kµ(η)k1 =
X
i=1
|A|−1
|µi (η)| = µ1 (η) +
X
µi (η),
i=2
P|A|−1
where µ1 (η) is the expected number of edges while i=2 µi (η) is the expected
number of transitive edges under η. Therefore, if µ1 (η ? ) = µ1 (η) and kµ(η)k1 =
kµ(η ? )k1 , then the expected numbers of transitive edges under η and η ? are identical. Let
T = {µ0 ∈ rint(M) : µ01 = µ1 (η ? ), kµ0 k1 = kµ(η ? )k1 }
be the set of mean-value parameter vectors for which the expected numbers of
edges and the expected numbers of transitive edges are identical. To see that it
28
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
is possible to choose η̇ ∈ int(N) such that µ1 (η̇) = µ1 (η ? ) and kµ(η̇)k1 6=
kµ(η ? )k1 , note that the set T is a subset of the boundary of the `1 -ball with center 0 ∈ R|A|−1 and radius kµ(η ? )k1 . Suppose that the `2 -distance of µ(η) from
µ(η ? ) is equal to ρ1 > 0; note that the `2 -distance is strictly positive because
η 6= η ? implies µ(η) 6= µ(η ? ). Observe that the `2 -ball B(µ(η ? ), ρ1 ) with center µ(η ? ) ∈ rint(M) and radius ρ1 > 0 need not be contained in rint(M), but,
owing to the fact that the set M is convex by construction, it is possible to construct a smaller `2 -ball B(µ(η ? ), ρ2 ) with the same center µ(η ? ) ∈ rint(M) but
smaller radius 0 < ρ2 < ρ1 such that the resulting `2 -ball B(µ(η ? ), ρ2 ) is contained in rint(M). The `1 -ball with center 0 ∈ R|A|−1 and radius kµ(η ? )k1 and
the `2 -ball B(µ(η ? ), ρ2 ) ⊆ rint(M) with center µ(η ? ) ∈ rint(M) and radius ρ2
intersect, but it is not too hard to see that it is possible to choose µ̇ ∈ rint(M)
inside the `2 -ball B(µ(η ? ), ρ2 ) ⊆ rint(M) such that µ̇1 = µ1 (η ? ) and kµ̇k1 6=
kµ(η ? )k1 . In addition, due to the fact that the map µ : int(N) 7→ rint(M) is
a homeomorphism [4, Theorem 3.6, p. 74], there exists η̇ ∈ int(N) such that
µ(η̇) = µ̇ ∈ B(µ(η ? ), ρ2 ) ⊆ rint(M), which implies that µ1 (η̇) = µ1 (η ? )
and kµ(η̇)k1 6= kµ(η ? )k1 . In conclusion, there exists η̇ ∈ int(N) such that
µ1 (η̇) = µ1 (η ? ) and kµ(η̇)k1 6= kµ(η ? )k1 and
|A|−1
kµ(η ? )
− µ(η)k2 ≥
kµ(η ? )
− µ(η̇)k2 =
X
(µi (η ? ) − µi (η̇))2 > 0.
i=2
kµ(η ? )
As a consequence, we can bound
− µ(η̇)k2 from below in terms of absolute deviations of the expected numbers of transitive edges under η ? and η̇ while
avoiding the trivial lower bound of 0.
To do so, we first use the Cauchy-Schwarz inequality to obtain
v
u
|A|−1
u
X
t
?
?
2
kµ(η ) − µ(η̇)k2 =
(µ1 (η ) − µ1 (η̇)) +
(µi (η ? ) − µi (η̇))2
i=2
v
u|A|−1
|A|−1
uX
X
1
?
2
t
|µi (η ? ) − µi (η̇))|
≥
(µi (η ) − µi (η̇)) ≥ p
|A|
−
2
i=2
i=2
and then use the triangle inequality to obtain
|A|−1
|A|−1
X
?
|µi (η ) − µi (η̇)| ≥
i=2
X
(µi (η ? ) − µi (η̇))
i=2
|A|−1
=
X
i=2
|A|−1
?
µi (η ) −
X
i=2
µi (η̇) .
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
29
P|A|−1
P|A|−1
The sums i=2 µi (η ? ) and i=2 µi (η̇) are equal to the expected number of
transitive edges under η ? and η̇, respectively, as shown above. Lemma 2 can hence
be applied to show that, for all η ∈ int(N), there exists C2 (η) > 0 such that
|A|−1
X
X
µi (η) = Eη
i=2
Xa,b
a∈A < b∈A
max
c∈A, c 6= a,b
Xa,c Xb,c
|A|
= C2 (η)
,
2
where C2 (η) satisfies
λ Λ(η)3 + (1 − λ) Ω(η)3 ≤ C2 (θ) ≤ λ Λ(η) + (1 − λ) Ω(η),
provided |A| ≥ 3. Collecting results shows that
v
u
|A|−1
u
X
t
?
?
2
(µ1 (η ) − µ1 (η̇)) +
(µi (η ? ) − µi (η̇))2
kµ(η ) − µ(η̇)k2 =
i=2
≥
=
≥
=
1
p
|A| − 2
1
|A|−1
X
|A|−1
µi (η(θ ? )) −
i=2
p
|C2
|A| − 2
X
µi (η(θ̇))
i=2
(η ? )
− C2 (η̇)|
|A|
2
1
|C2 (η ? ) − C2 (η̇)|
1/4
|A| (|A| − 1)1/4
3/4
1
|A|
?
|C2 (η ) − C2 (η̇)|
,
1/4
2
2
|A|
2
where |C2 (η ? ) − C2 (η̇)| > 0 by the choice of η̇, as explained above.
We can therefore conclude that, for all θ ∈ Θ \ B(θ ? , ), there exists a function
C(θ ? , θ) > 0 such that
kµ(η(θ ? ))
− µ(η(θ))k2 ≥
C(θ ? , θ)
3/4
|A|
.
2
Last, but not least, we show that there exists δ() > 0 such that, for all η(θ) ∈
3/4
N \ B(η(θ ? ), γ()), we have kµ(η(θ ? )) − µ(η(θ))k2 ≥ δ() |A|
> 0. To do
2
so, it is convenient to operate in the natural parameter space N = R|A|−1 of the full
exponential family with natural parameter vector η and sufficient statistic vector
s(x). Write η ≡ η(θ) and η ? ≡ η(θ ? ). By classic exponential-family theory,
ψ(η) is a strictly convex function of the natural parameter vector η on the interior
int(N) of the natural parameter space N, which is a convex set [4, Theorem 1.13,
30
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
p. 19]. Thus, the gradient ∇η ψ(η) = µ(η) is a strictly increasing function of the
natural parameter vector η and kµ(η ? ) − µ(η)k2 , considered as a function of η for
fixed η ? ∈ int(N), attains its minimum on the boundary of the ball B(η ? , γ()).
Therefore, kµ(η ? ) − µ(η)k2 is bounded away from 0 by a function of > 0 for
all η ∈ N \ B(η ? , γ()).
Taken together, these results show that there exists δ() > 0 such that, for all
η(θ) ∈ N \ B(η(θ ? ), γ()),
3/4
|A|
?
kµ(η(θ )) − µ(η(θ))k2 ≥ δ()
,
2
provided |A| ≥ 4. Therefore, identifiability condition (6) of Theorem 2 is satisfied
with α = 3/4 provided |A| ≥ 4.
E.3. Auxiliary lemmas: Theorem 3. We prove an auxiliary lemma that is
useful for verifying the conditions of Theorem 3.
Lemma 6. Consider an exponential-family random graph of fixed and finite size
with within-neighborhood edge, transitive edge, and same-attribute edge terms as
defined in Section 5.2. Then identifiability condition [C.3] of Theorem 3 is satisfied
with α = 1 provided |Ak | ≥ 3 (k = 1, . . . , K).
Proof of Lemma 6. To verify identifiability condition [C.3] of Theorem 3, we
need to verify that, for all > 0 small enough so that B(θ0 , ) ⊆ Θ0 , there exist
δ() > 0 such that, for all θ ∈ Θ0 \ B(θ0 , ),
kEθ? b(X) − Eθ b(X)k2 ≥ δ()
K
X
|Ak | α
k=1
2
for some 0 ≤ α ≤ 1,
where Eθ? b(X) is the expectation of b(X) under the data-generating exponentialfamily distribution with natural parameter vector θ ? ∈ Θ? = R3 and Eθ b(X)
is the expectation of b(X) under the misspecified exponential-family distribution
with natural parameter vector θ ∈ Θ0 = R2 .
Since the expectations Eθ? b(X) and Eθ b(X) are expectations of the number
of within-neighborhood edges and transitive edges, an argument along the lines of
Lemma 3 shows that, for all > 0 small enough so that B(θ0 , ) ⊆ Θ0 , there exist
δ() > 0 such that, for all θ ∈ Θ0 \ B(θ ? , ),
kEθ? b(X) − Eθ b(X)k2 ≥ δ()
K
X
|Ak |
k=1
2
,
provided |Ak | ≥ 3 (k = 1, . . . , K). Therefore, identifiability condition [C.3] of
Theorem 3 is satisfied with α = 1 provided |Ak | ≥ 3 (k = 1, . . . , K).
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
31
REFERENCES
[1] Bhamidi, S., Bresler, G., and Sly, A. (2011), “Mixing time of exponential random graphs,” The
Annals of Applied Probability, 21, 2146–2170.
[2] Bhattacharya, B. B., and Mukherjee, S. (2018), “Inference in Ising models,” Bernoulli, 24,
493–525.
[3] Boucheron, S., Lugosi, G., and Massart, P. (2013), Concentration Inequalities: A Nonasymptotic Theory of Independence, Oxford: Oxford University Press.
[4] Brown, L. (1986), Fundamentals of Statistical Exponential Families: With Applications in Statistical Decision Theory, Hayworth, CA, USA: Institute of Mathematical Statistics.
[5] Butts, C. T. (2011), “Bernoulli graph bounds for general random graph models,” Sociological
Methodology, 41, 299–345.
[6] Butts, C. T., and Almquist, Z. W. (2015), “A Flexible Parameterization for Baseline Mean
Degree in Multiple-Network ERGMs,” Journal of Mathematical Sociology, 39, 163–167.
[7] Chatterjee, S. (2007), “Estimation in spin glasses: A first step,” The Annals of Statistics, 35,
1931–1946.
[8] Chatterjee, S., and Diaconis, P. (2013), “Estimating and understanding exponential random
graph models,” The Annals of Statistics, 41, 2428–2461.
[9] Chung, F., and Lu, L. (2006), Complex Graphs and Networks, Providence: American Mathematical Society.
[10] Crane, H., and Dempsey, W. (2015), “A framework for statistical network modeling,” Available
at https://arxiv.org/abs/1509.08185.v4.
[11] Diaconis, P., Chatterjee, S., and Sly, A. (2011), “Random graphs with a given degree sequence,”
The Annals of Applied Probability, 21, 1400–1435.
[12] Efron, B. (1978), “The geometry of exponential families,” The Annals of Statistics, 6, 362–376.
[13] Frank, O., and Strauss, D. (1986), “Markov graphs,” Journal of the American Statistical Association, 81, 832–842.
[14] Handcock, M. S. (2003), “Statistical Models for Social Networks: Inference and Degeneracy,” in Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers,
eds. Breiger, R., Carley, K., and Pattison, P., Washington, D.C.: National Academies Press, pp.
1–12.
[15] Handcock, M. S., and Gile, K. (2010), “Modeling Social Networks from Sampled Data,” The
Annals of Applied Statistics, 4, 5–25.
[16] Holland, P. W., and Leinhardt, S. (1976), “Local Structure in Social Networks,” Sociological
Methodology, 1–45.
[17] Hollway, J., and Koskinen, J. (2016), “Multilevel embeddedness: The case of the global fisheries governance complex,” Social Networks, 44, 281–294.
[18] Hollway, J., Lomi, A., Pallotti, F., and Stadtfeld, C. (2017), “Multilevel social spaces: The
network dynamics of organizational fields,” Network Science, 5, 187–212.
[19] Hunter, D. R., and Handcock, M. S. (2006), “Inference in curved exponential family models
for networks,” Journal of Computational and Graphical Statistics, 15, 565–583.
[20] Hunter, D. R., Krivitsky, P. N., and Schweinberger, M. (2012), “Computational statistical methods for social network models,” Journal of Computational and Graphical Statistics, 21, 856–
882.
[21] Janson, S., and Rucinski, A. (2002), “The infamous upper tail,” Random Structures and Algorithms, 20, 317–342.
[22] Jonasson, J. (1999), “The random triangle model,” Journal of Applied Probability, 36, 852–876.
[23] Kass, R., and Vos, P. (1997), Geometrical foundations of asymptotic inference, New York:
Wiley.
[24] Kim, J. H., and Vu, V. H. (2004), “Divide and conquer martingales and the number of triangles
32
MICHAEL SCHWEINBERGER AND JONATHAN STEWART
in a random graph,” Random Structures & Algorithms, 24, 166–174.
[25] Kolaczyk, E. D. (2009), Statistical Analysis of Network Data: Methods and Models, New York:
Springer-Verlag.
[26] Kontorovich, L., and Ramanan, K. (2008), “Concentration inequalities for dependent random
variables via the martingale method,” The Annals of Probability, 36, 2126–2158.
[27] Krivitsky, P. N. (2012), “Exponential-family models for valued networks,” Electronic Journal
of Statistics, 6, 1100–1128.
[28] Krivitsky, P. N., Handcock, M. S., and Morris, M. (2011), “Adjusting for network size and
composition effects in exponential-family random graph models,” Statistical Methodology, 8,
319–339.
[29] Krivitsky, P. N., and Kolaczyk, E. D. (2015), “On the question of effective sample size in
network modeling: An asymptotic inquiry,” Statistical Science, 30, 184–198.
[30] Lauritzen, S., Rinaldo, A., and Sadeghi, K. (2017), “Random networks, graphical models, and
exchangeability,” Available at https://arxiv:1701.08420v1.
[31] Lazega, E., and Snijders, T. A. B. (eds.) (2016), Multilevel Network Analysis for the Social
Sciences, Switzerland: Springer-Verlag.
[32] Lomi, A., Robins, G., and Tranmer, M. (2016), “Introduction to multilevel social networks,”
Social Networks, 266–268.
[33] Lusher, D., Koskinen, J., and Robins, G. (2013), Exponential Random Graph Models for Social
Networks, Cambridge, UK: Cambridge University Press.
[34] Mukherjee, S. (2013), “Consistent estimation in the two star exponential random graph model,”
Tech. rep., Department of Statistics, Columbia University, arXiv:1310.4526v1.
[35] Nowicki, K., and Snijders, T. A. B. (2001), “Estimation and prediction for stochastic blockstructures,” Journal of the American Statistical Association, 96, 1077–1087.
[36] Ravikumar, P., Wainwright, M. J., and Lafferty, J. (2010), “High-dimensional Ising model selection using `1 -regularized logistic regression,” The Annals of Statistics, 38, 1287–1319.
[37] Rinaldo, A., Petrovic, S., and Fienberg, S. E. (2013), “Maximum likelihood estimation in network models,” The Annals of Statistics, 41, 1085–1110.
[38] Rubin, D. B. (1976), “Inference and missing data,” Biometrika, 63, 581–592.
[39] Schweinberger, M. (2011), “Instability, sensitivity, and degeneracy of discrete exponential families,” Journal of the American Statistical Association, 106, 1361–1370.
[40] Schweinberger, M., and Handcock, M. S. (2015), “Local dependence in random graph models:
characterization, properties and statistical inference,” Journal of the Royal Statistical Society,
Series B, 77, 647–676.
[41] Schweinberger, M., and Luna, P. (2017), “HERGM: Hierarchical exponential-family random
graph models,” Journal of Statistical Software, accepted.
[42] Schweinberger, M., and Stewart, J. (2017), “Supplement: Finite-graph concentration and consistency results for random graphs with complex topological structures,” Tech. rep., Department
of Statistics, Rice University.
[43] Shalizi, C. R., and Rinaldo, A. (2013), “Consistency under sampling of exponential random
graph models,” The Annals of Statistics, 41, 508–535.
[44] Slaughter, A. J., and Koehly, L. M. (2016), “Multilevel models for social networks: hierarchical
Bayesian approaches to exponential random graph modeling,” Social Networks, 44, 334–345.
[45] Snijders, T. A. B., Pattison, P. E., Robins, G. L., and Handcock, M. S. (2006), “New specifications for exponential random graph models,” Sociological Methodology, 36, 99–153.
[46] Strauss, D. (1986), “On a general class of models for interaction,” SIAM Review, 28, 513–527.
[47] Wang, P., Robins, G., Pattison, P., and Lazega, E. (2013), “Exponential random graph models
for multilevel networks,” Social Networks, 35, 96–115.
[48] Wasserman, S., and Pattison, P. (1996), “Logit models and logistic regression for social networks: I. An introduction to Markov graphs and p∗ ,” Psychometrika, 61, 401–425.
FINITE-GRAPH CONCENTRATION AND CONSISTENCY RESULTS
33
[49] Xiang, R., and Neville, J. (2011), “Relational learning with one network: an asymptotic analysis,” in Proceedings of the 14th International Conference on Artificial Intelligence and Statistics
(AISTATS), pp. 1–10.
[50] Yan, T., Leng, C., and Zhu, J. (2016), “Asymptotics in directed exponential random graph
models with an increasing bi-degree sequence,” The Annals of Statistics, 44, 31–57.
[51] Yan, T., Wang, H., and Qin, H. (2016), “Asymptotics in undirected random graph models parameterized by the strengths of vertices,” Statistica Sinica, 26, 273–293.
[52] Zappa, P., and Lomi, A. (2015), “The analysis of multilevel networks in organizations: models
and empirical tests,” Organizational Research Methods, 18, 542–569.
M ICHAEL S CHWEINBERGER , J ONATHAN S TEWART
D EPARTMENT OF S TATISTICS
R ICE U NIVERSITY
6100 M AIN S T, MS-138
H OUSTON , TX 77005-1827
E- MAIL : M . S @ RICE . EDU
JONATHAN . STEWART @ RICE . EDU
| 10math.ST
|
LOWER BOUND CLUSTER ALGEBRAS:
PRESENTATIONS, COHEN-MACAULAYNESS, AND NORMALITY
arXiv:1508.02314v2 [math.AC] 24 Aug 2015
GREG MULLER, JENNA RAJCHGOT, AND BRADLEY ZYKOSKI
Abstract. We give an explicit presentation for each lower bound cluster algebra. Using this presentation,
we show that each lower bound algebra Gröbner degenerates to the Stanley-Reisner scheme of a vertexdecomposable ball or sphere, and is thus Cohen-Macaulay. Finally, we use Stanley-Reisner combinatorics
and a result of Knutson-Lam-Speyer to show that all lower bound algebras are normal.
1. Introduction and statement of results
Cluster algebras are a family of combinatorially-defined commutative algebras which were introduced by
Fomin and Zelevinsky at the turn of the millennium to axiomatize and generalize patterns appearing in
the study of dual canonical bases in Lie theory [FZ02]. Since their introduction, cluster algebras have been
discovered in the rings of functions on many important spaces, such as semisimple Lie groups, Grassmannians,
flag varieties, and Teichmüller spaces [BFZ05, Sco06, GLS08, GSV05].1
In each of these examples, the cluster algebra is realized as the coordinate ring of a smooth variety. This
makes it all the more surprising that the varieties associated to general cluster algebras can be singular;
in fact, they can possess such nightmarish pathologies as a non-Noetherian singularity [Mul13]. Various
approaches have been introduced to mitigate this.
• Restricting to a subclass of cluster algebras with potentially better behavior: acyclic cluster algebras
[BFZ05], locally acyclic cluster algebras [Mul13, BMRS15], or cluster algebras with a maximal green
sequence [BDP14, Mul15].
• Replacing the cluster algebra by a closely-related algebra with potentially better behavior: upper
cluster algebras [BFZ05, BMRS15], the span of convergent theta functions [GHKK14], or lower
bound algebras [BFZ05].
In this note, we study the algebraic and geometric behavior of lower bound algebras.2
1.1. Lower bound algebras. Lower bound algebras were introduced in [BFZ05] as a kind of ‘lazy approximation’ of a cluster algebra, in the following sense. A cluster algebra is defined to be the subalgebra of a
field of rational functions generated by a (usually infinite) set of cluster variables, produced by a recursive
procedure called mutation. A lower bound algebra is defined to be the subalgebra generated by truncating
this process at a specific finite set of steps. The resulting algebra is contained in the associated cluster
algebra and is manifestly finitely generated.
A lower bound algebra is constructed from an ice quiver Q: this is a quiver (i.e. a finite directed graph)
without loops or directed 2-cycles, in which each vertex is designated unfrozen or frozen. As a matter of
convenience, we assume the vertices of Q have been indexed by the numbers {1, 2, ..., n}. To each unfrozen
−
vertex i, we associate a pair of monomials p+
i , pi ∈ Z[x1 , x2 , ..., xn ] as follows.
Y
Y
(1.1)
p+
xtarget(a) ,
p−
xsource(a)
i :=
i :=
arrows a∈Q
arrows a∈Q
source(a)=i
target(a)=i
2010 Mathematics Subject Classification. Primary 13F60, Secondary 05E40,13F55.
Keywords: Cluster algebras, lower bound cluster algebras, combinatorial commutative algebra, Stanley-Reisner complexes.
The third author was supported by NSF grant DMS-1001764, as an REU project at the University of Michigan in 2015.
1
A more interesting and morally correct statement is that each of these spaces possesses a stratification such that each
stratum naturally has a cluster algebra in its ring of functions.
2More specifically, we consider lower bound algebras defined by a quiver in the body of the paper, and consider the more
general context of geometric type in Appendix B.
1
Each vertex then determines a Laurent polynomial x′i , called the adjacent cluster variable at i, which is
defined by the following formula.3
−1 +
xi (pi + p−
) if i is unfrozen
′
i
(1.2)
xi :=
xi−1
if i is frozen
±1
The lower bound algebra L(Q) defined by Q is the subalgebra of Z[x±1
1 , ..., xn ] generated by the variables
′
′
′
x1 , x2 , ..., xn and the adjacent cluster variables x1 , x2 , ..., xn .
3
1
2
4
Figure 1. An ice quiver (the unique frozen vertex is depicted as a square)
Example 1.3. Consider the ice quiver Q in Figure 1. The four adjacent cluster variables are below.
x′1 =
x22 + 1
,
x1
x′2 =
x3 x4 + x21
1 + x2
, x′3 =
,
x2
x3
x′4 =
1
x4
1.2. Relations in L(Q). We first consider the problem of finding relations among the generators of L(Q).
Each adjacent cluster variable satisfies a defining relation immediately from its definition.
(1.4)
−
∀i unfrozen, x′i xi = (p+
i + pi )
(1.5)
∀i frozen, x′i xi = 1
A more interesting class of relations is given by the following proposition.
Proposition 1.6 (The cycle relations). For each directed cycle of unfrozen vertices v1 → v2 → · · · → vk →
vk+1 = v1 in Q, the following cycle relation holds.
!
k
k
−
Y
Y
Y
X
Y p+
p
p+
p−
v
v
vi
vi
i+1
i
′
|S|
xvi =
+
(1.7)
(−1)
xvi xvi+1
x
x
i=1 vi
i=1 vi
S⊂{1,2,...,k}
i∈S
i6∈S∪(S+1)
S∩(S+1)=∅
We note that the expressions on either side reduce to polynomials in the generators, despite the presence of
fractions. Also note that choosing a different initial vertex v1 in the same directed cycle does not change the
corresponding cycle relation.
1
3
2
Figure 2. An ice quiver (no frozen vertices)
Remark 1.8. A quiver Q is called acyclic if it has no directed cycles of unfrozen vertices. The unifying
theme of this paper is the use of the cycle relations to generalize results about L(Q) which were previously
known only when Q is acyclic (that is, when there are no cycle relations).
3This is an abuse of terminology. Technically speaking, a frozen vertex i should not have an adjacent cluster variable x′ ,
i
and instead we should include x−1
as a generator (though this latter step is a matter of some debate). We are streamlining the
i
process by calling the inverse x−1
‘the adjacent cluster variable at i’.
i
2
Example 1.9. Let Q be the ice quiver in Figure 2. The three adjacent cluster variables are
x2 + x3
x3 + x1
x1 + x2
x′1 =
, x′2 =
, x′3 =
x1
x2
x3
The defining relations here are obtained by clearing the denominators above. A non-trivial directed 3-cycle
starting at any vertex determines the cycle relation
x′1 x′2 x′3 − x′1 − x′2 − x′3 = 2
which may be verified by direct computation.
1.3. A presentation of L(Q). We may ask whether there are other relations in L(Q) that are not an
immediate consequence of the preceding relations; or more concretely, whether the defining relations and the
cycle relations generate the entire ideal of relations among the generators.
Explicitly, we consider the homomorphism of rings4
±1
±1
π : Z[x1 , x2 , ..., xn , y1 , y2 , ...yn ] −→ Z[x±1
1 , x2 , ..., xn ]
∀i, π(xi ) = xi , π(yi ) = x′i
The image of this homomorphism is L(Q), and so KQ := ker(π) is the ideal of relations among the
generators of L(Q), where each adjacent cluster variable x′i has been replaced by the abstract variable yi .
The homomorphism π descends to an isomorphism
∼
Z[x1 , x2 , ..., xn , y1 , y2 , ..., yn ]/KQ −→ L(Q)
We will say a directed cycle v1 → v2 → · · · → vk → vk+1 = v1 is vertex-minimal if no vertex appears
more than once and there is no directed cycle whose vertex set is a proper subset of {v1 , v2 ..., vk }.
Theorem 1.10. The ideal of relations KQ is generated by the following elements.
• For each unfrozen vertex i,
−
yi xi − p+
i − pi
(1.11)
• For each frozen vertex i,
(1.12)
yi xi − 1
• For each vertex-minimal directed cycle of unfrozen vertices v1 → v2 → · · · → vk → vk+1 = v1 ,
!
k
k
−
Y
Y
Y
X
Y p+
p+
p−
vi pvi+1
vi
vi
|S|
yvi −
−
(1.13)
(−1)
xvi xvi+1
x
x
i=1 vi
i=1 vi
S⊂{1,2,...,k}
i∈S
i6∈S∪(S+1)
S∩(S+1)=∅
which simplifies to a polynomial.
The theorem is true without the vertex-minimal condition, which is used here to reduce the set of relations.
Example 1.14. Let Q be the ice quiver in Figure 2. By Theorem 1.10, L(Q) is isomorphic to the quotient
of Z[x1 , x2 , x3 , y1 , y2 , y3 ] by the ideal KQ generated by the following 4 relations.
KQ = hy1 x1 − x2 − x3 , y2 x2 − x3 − x1 , y3 x3 − x1 − x2 , y1 y2 y3 − y1 − y2 − y3 − 2i
1.4. A Gröbner basis for KQ . We prove Theorem 1.10 by means of a stronger result, that the given
generators are a Gröbner basis for the ideal of relations KQ . Recall that, given a polynomial ring with
a monomial order <, a Gröbner basis of an ideal I is a generating set {g1 , g2 , .., gk } of I satisfying the
additional condition that {in< (g1 ), in< (g2 ), ..., in< (gk )} is a generating set of in< (I).
The monomial orders relevant to us are those
are much more expensive than the
P
P in which the y-variables
x-variables, that is xα yβ > xγ yδ whenever i βi is larger than i δi . An example of such a monomial order
is the lexicographical order where the variables are ordered by
y1 > y2 > · · · > yn > x1 > x2 > · · · > xn .
4The y-variables introduced here have no relation to the y-variables or coefficient variables introduced in [FZ07].
3
y1
y2
x1
x2
x3
y3
Figure 3. The Stanley-Reisner complex for hx1 y1 , x2 y2 , x3 y3 , y1 y2 y3 i
Theorem 1.15. For any monomial order of Z[x1 , x2 , ..., xn , y1 , y2 , ..., yn ] in which all of the y-variables are
much more expensive than all of the x-variables, the polynomials given in Theorem 1.10 are a Gröbner basis
for KQ . Consequently, the initial ideal in< KQ is squarefree monomial ideal with generating set
{xi yi | 1 ≤ i ≤ n} ∪ {yv1 yv2 · · · yvk | v1 → v2 → · · · → vk → vk+1 = v1 is a vertex-minimal cycle in Q}.
Remark 1.16. When Q is acyclic, Theorem 1.15 specializes to Corollary 1.17 in [BFZ05]. The proof of
Theorem 1.15 given in Section 2.2 uses [BFZ05, Corollary 1.17] in an essential way, so our proof is not
independent of the original result.
1.5. Simplicial complexes and Cohen-Macaulayness of lower bound algebras. From here on, we
work over a field K, and consider the K-algebra L(Q)⊗Z K. We will still refer to this algebra as a lower bound
algebra and, though a slight abuse of notation, will simply denote it by L(Q). Similarly, we let KQ denote
the associated lower bound ideal, so that it is the kernel of the map π : K[x1 , . . . , xn , y1 . . . , yn ] → L(Q).
To a squarefree monomial ideal I ⊆ K[z1 , . . . , zn ], one can associate a simplicial complex ∆ on the vertex
set {z1 , . . . , zn }. This simplicial complex is called the Stanley-Reisner complex and is defined as follows:
/ I. Observe that the minimal non-faces
{zi1 , . . . zir } is a face of ∆ if and only if the monomial zi1 · · · zir ∈
of ∆ are in one-to-one correspondence with a minimal generating set of I. For further information on
Stanley-Reisner complexes, see the textbook [MS05, Chapter 1].
Whenever the Stanley-Reisner complex is a simplicial ball or a sphere,5 the corresponding face ring
K[z1 , . . . , zn ]/I is a Cohen-Macaulay ring [Mun84]. Furthermore, when I = in< J for some ideal J ⊆
K[z1 , . . . , zn ], we may also conclude that J itself is Cohen-Macaulay (see eg. [BC03, Proposition 3.1]).
Example 1.17. We continue Example 1.14 and observe that the initial ideal in< KQ (for a term order < as
in Theorem 1.15) is hx1 y1 , x2 y2 , x3 y3 , y1 y2 y3 i. The facets (i.e. the maximal faces) of the associated StanleyReisner complex are precisely those {z1 , z2 , z3 } where zi is either xi or yi and at least one of z1 , z2 , z3 is an
xi . This Stanley-Reisner complex is readily seen to be a simplicial ball, and is pictured in Figure 3.
This example generalizes, and we are able to conclude that all lower bound algebras are Cohen-Macaulay.
Theorem 1.18. Let Q be a quiver with n vertices, let KQ ⊆ K[x1 , . . . , xn , y1 , . . . , yn ] be its ideal of relations,
and let < be any monomial order where the y-variables are much more expensive than the x-variables. Let
∆Q be the Stanley-Reisner complex of the squarefree monomial ideal in< KQ .
(i) If Q is acyclic, then ∆Q is a simplicial sphere.
(ii) If Q is not acyclic, then ∆Q is a simplicial ball.
We show additional properties of ∆Q . If Q is acyclic, then ∆Q is the boundary of a cross-polytope. In both
cases, ∆Q satisfies the stronger condition of vertex-decomposibility. Details are in Section 3.
Corollary 1.19. For any Q the lower bound algebra L(Q) over a field K is Cohen-Macaulay.
5More precisely, we mean the geometric realization of the simplicial complex is homeomorphic to a ball or sphere, respectively.
Whenever we refer to a simplicial complex as a topological object, we more precisely mean its geometric realization.
4
Remarks 1.20.
(i) We prove Theorem 1.18 by way of a more general result, which gives a larger class
of simplicial complexes which are automatically vertex-decomposable balls (see Theorem 3.4).
(ii) When Q is acyclic, L(Q) was already known to be Cohen-Macaulay; specifically, [BFZ05, Corollary
1.17] implies that L(Q) is a complete intersection, and, consequently, that it is Cohen-Macaulay.
1.6. Normality of lower bound algebras. Our last main result is the normality of the K-algebra L(Q).
Theorem 1.21. Every lower bound cluster algebra defined by a quiver is normal.
Since L(Q) = K[x1 , ..., xn , y1 , ..., yn ]/KQ is finitely-generated, Serre’s Criterion reduces normality to a pair
of geometric conditions on the variety V(KQ ).
• (R1) The variety V(KQ ) has no codimension-1 singularities.
• (S2) Any regular function on an open subset in V(KQ ) with codimension-2 complement extends to
a regular function on all of V(KQ ).
The Cohen-Macaulay property implies the S2 condition, and so normality of L(Q) reduces to proving there
are no codimension-1 singularities.
As with Cohen-Macaulayness, this geometric question will be reduced to Stanley-Reisner combinatorics,
along with a result of Knutson-Lam-Speyer [KLS14, Proposition 8.1]. See Section 4 for further information.
Remark 1.22. Like our other results, Theorem 1.21 is only new in the case of non-acyclic Q. In the acyclic
setting, L(Q) is equal to its upper cluster algebra6, which is a normal domain by [Mul13, Prop. 2.1].
Structure of paper. Section 2 considers relations in L(Q) and proves the associated results: Proposition
1.6, Theorem 1.10 and Theorem 1.15. Section 3 introduces the relevant combinatorial tools, leading to the
proof of Theorem 1.18. Section 4 addresses normality, proving Theorem 1.21.
The paper concludes with a pair of appendices which frame the scope of the paper. Appendix A considers
the singularities of lower bound algebras directly, and provides an example to suggest this is a difficult
problem. Appendix B explains how the results of the paper can be extended to skew-symmetrizable lower
bound algebras, which are more general but also somewhat less intuitive.
Acknowledgements. This paper is the result of a summer 2015 REU project at the University of Michigan.
This project was supported by Karen Smith’s NSF grant DMS-1001764. We are also grateful to Sergey Fomin
for numerous helpful comments on an early draft of this note.
2. Presentations and Gröbner Bases
2.1. Choice graphs and cycle relations. To every cycle of Q there exists a corresponding relation in
L(Q). Let Q have a directed cycle v1 → v2 → · · · → vk−1 → vk → v1 . By giving an alternate presentation
for the product
(2.1)
k
Y
x′vi
i=1
=
k
Y
−
+
x−1
vi (pvi + pvi ),
i=1
we acquire a nontrivial relation that holds in L(Q). It is our goal to expand the right-hand product as a sum,
and from this, Proposition 1.6 will follow. Each term of the expansion of this product represents a choice,
−
−1 +
−
for each i, of either p+
vi or pvi from xvi (pvi + pvi ). Therefore, to each term, we associate a directed graph
n
with Z/kZ as its vertex set and {(i, i ± 1)}i=1 as its set of arrows, where the sign of ± corresponds to the
−
+
abovementioned choice of p+
vi (corresponding to the positive sign because xvi+1 divides pvi ) or pvi (negative
−
sign, since xvi−1 divides pvi ). Call these graphs the choice graphs of the terms of the expansion of (2.1),
and let C denote the set of all choice graphs. Formally, we write the correspondence between choice graphs
and terms as a function
(2.2)
M : C → {monomials in
±1
Z[x±1
1 , . . . , xk ]},
M (g) =
k
Y
signg (i)
x−1
vi pvi
,
i=1
6This was proven with an additional assumption in [BFZ05, Thm. 1.18], and without said assumption in [Mul14].
5
1
2
6
3
5
4
+
−1 +
−1 −
−1 −
−1 −
−1 −
Figure 4. The choice graph for (x−1
1 p1 )(x2 p2 )(x3 p3 )(x4 p4 )(x5 p5 )(x6 p6 )
where
(2.3)
signg (i) =
(
+
−
if (i, i + 1) ∈ g,
.
if (i, i − 1) ∈ g.
Example 2.4. Let Q be the quiver on Z/6Z whose set of arrows is {(j, j + 1)}6j=1 . The choice graph in
Figure 4 represents the term
+
−1 +
−1 −
−1 −
−1 −
−1 −
−1 +
(x−1
1 p1 )(x2 p2 )(x3 p3 )(x4 p4 )(x5 p5 )(x6 p6 ) = x1 p1
− −
−
p+
2 p3 + p5 p6 −1
p4
x .
x3 x2
x4 x5 6
By our construction of C and M , we have that (2.1) may be written as
k
Y
(2.5)
−
+
x−1
vi (pvi + pvi ) =
i=1
X
M (g),
g∈C
and so it is our goal to expand the sum on the right of (2.5). The following proposition gives us such an
expansion.
Proposition 2.6. We may expand (2.5) as follows:
k
k
Y
Y
p−
p+
vi
vi
−
+
+
+
x−1
(2.7)
vi (pvi + pvi ) =
x
x
v
v
i−1
i+1
i=1
i=1
i=1
k
Y
X
−
Y p+
vi pvi+1
xvi+1 xvi
(−1)|S|+1
i∈S
∅6=S⊂{1,2,...,k}
S∩(S+1)=∅
!
Y
i∈S∪(S+1)
/
x′vi
Proof. Note that a choice graph has a directed 2-cycle if and only if its associated term has a factor of the
p+
vi
p−
v
p+
v
−
−1
+
form (x−1
vi pvi )(xvi+1 pvi+1 ), which is a monomial
i+1
i
xvi+1
p−
vi+1 .
x vi
in the variables xv1 , . . . , xvn since xvi+1 divides
Also notice that no two 2-cycles may share vertices, since each vertex meets the tail
and xvi divides
of one and only one arrow. It follows that a choice graph may have at most ⌊ k2 ⌋ 2-cycles. Finally notice that
a 2-cycle may only exist on adjacent vertices. Now, let Sj denote the collection of subsets S of {1, . . . , k}
S⌊ k2 ⌋
Sj . These subsets S correspond to the “left
of size j ≥ 1 such that S ∩ (S + 1) = ∅, and let S = j=1
endpoints” of the 2-cycles in choice graphs with j 2-cycles.
Let C(S) be the set of all choice graphs with a 2-cycle on every pair {i, i + 1} for i ∈ S ∈ S, and let C0
be the set of choice graphs with no 2-cycles, then we have
!
!
[
[
[
C(S) .
(2.8)
C = C0 ∪
C(S) ∪
C(S) ∪ · · · ∪
S∈S1
S∈S2
S∈S⌊ n ⌋
2
Since every g ∈ C(S) has a pair of arrows (i, i + 1), (i + 1, i) for every i ∈ S, we have
!
−
Y p+
X
Y
vi pvi+1
−
+
M (g) =
(2.9)
x−1
.
vi pvi + pvi
xvi+1 xvi
g∈C(S)
i∈S
i∈S∪(S+1)
/
We also have
X
g∈C0
M (g) =
k
k
Y
Y
p−
p+
vi
vi
+
,
x
x
v
v
i−1
i+1
i=1
i=1
6
since there are precisely two choice graphs with no 2-cycles, corresponding to a consistent choice of either +
or −. We wish use (2.8) to write (2.5) as a sum with summands of the form (2.9). Such a sum must have
precisely one term corresponding to each member of C. However, there is a certain amount of overcounting
involved in (2.8), since for any S ∈ S, we have
(2.10)
C(S) ⊆ C(S r {i}) for every i ∈ S.
We therefore proceed iteratively by way of the inclusion-exclusion principle. We first include a summand
corresponding to C(S) for every S ∈ S1 :
X X
M (g).
T1 =
S∈S1 g∈C(S)
As (2.10) shows, for every S ∈ S2 , the summand T1 contains two terms corresponding to each element of
C(S). Therefore, we now exclude a summand for each C(S), S ∈ S2 :
X X
M (g).
T2 = T1 −
S∈S2 g∈C(S)
Again, (2.10) shows us that, for every S ∈ S3 , the summand T2 excludes one term too many for each element
of C(S). We therefore define T3 accordingly, and continue the process of inclusion and exclusion until we
obtain
⌊k
2⌋
X X
X
(−1)j+1
T⌊ k ⌋ =
M (g) .
2
j=1
S∈Sj g∈C(S)
P
Finally, we must include a term of g∈C0 M (g) corresponding to C0 , and we conclude that
X
X
M (g) =
M (g) + T⌊ k ⌋
2
g∈C
g∈C0
!
⌊k
k
k
2⌋
+ p−
−
+
X
Y
X
Y
Y
Y
p
p
p
vi+1
vi
vi
vi
−
(−1)j+1
x′vi
+
+
(p+
xv−1
vi + pvi ) =
i
x
x
x
x
v
v
v
v
i−1
i+1
i+1
i
j=1
i=1
i=1
i=1
S∈Sj i∈S
i∈S∪(S+1)
/
!
k
k
−
+ p−
Y
Y
X
Y
Y
p
p
p+
vi+1
vi
vi
vi
(−1)|S|+1
x′vi .
+
+
=
x
x
x
v
v
v
i−1
i+1
i+1 xvi
i=1
i=1
k
Y
i∈S
∅6=S⊂{1,2,...,k}
i∈S∪(S+1)
/
S∩(S+1)=∅
Proposition 1.6 follows, since we now have
k
X
Y
−
+
(−1)|S|+1
x−1
vi (pvi + pvi ) −
i=1
∅6=S⊂{1,2,...,k}
S∩(S+1)=∅
X
S⊂{1,2,...,k}
S∩(S+1)=∅
(−1)|S|
−
Y p+
vi pvi+1
xvi+1 xvi
i∈S
−
Y p+
vi pvi+1
xvi+1 xvi
i∈S
!
!
Y
k
k
Y
Y
p+
p−
vi
vi
+
x
x
v
v
i−1
i+1
i=1
i=1
k
k
Y
Y
p+
p−
vi
vi
+
.
x
x
v
v
i−1
i+1
i=1
i=1
x′vi =
i∈S∪(S+1)
/
Y
x′vi =
i∈S∪(S+1)
/
−
2.2. Generators for KQ . Now, in addition to the defining polynomials yi xi − (p+
i + pi ), yi xi − 1 given
by the defining relations (1.11) and (1.12), we have by Proposition 2.6 that the ideal of relations KQ also
contains the cycle polynomials. We define the cycle polynomials in Z[x1 , x2 , . . . , xn , y1 , y2 , . . . , yn ] to be
those polynomials coming from vertex-minimal directed cycles whose images under π vanish by virtue of
(2.7). That is, for every vertex-minimal directed cycle of unfrozen vertices v1 → v2 → · · · → vk → v1 in Q,
we have the cycle polynomial
!
k
k
−
Y p+
Y
Y
X
Y
p
p+
p−
vi+1
vi
vi
vi
(−1)|S|
yvi −
−
.
xvi+1 xvi
x
x
v
v
i−1
i+1
i=1
i=1
S⊂{1,2,...,k}
S∩(S+1)=∅
i∈S
i6∈S∪(S+1)
7
Table 1. Cycle polynomials in several examples
1
y1 y2 y3 − y1 − y2 − y3 − 2
3
2
1
2
y1 y2 y3 y4 −y1 y2 −y1 y4 −y2 y3 −y3 y4
4
3
1
5
y1 y2 y3 y4 y5 − y1 y2 y3 − y1 y2 y5 −
y2 y3 y4 − y3 y4 y5 + y1 + y2 + y3 +
y4 + y5 − 2
2
4
3
1
4
y1 y2 y5 −y1 x3 −y2 −y5 x4 −x3 −x4
and
y2 y3 y4 −y2 −y3 x1 −y4 x5 −x1 −x5
2
5
3
Note again that this expression reduces to a polynomial in the x- and y-variables because each xvi+1 divides
−
p+
vi and each xvi divides pvi+1 . In Table 1, we present the cycle polynomials given by some basic ice quivers.
We now obtain a presentation for the ideal of relations KQ and for the initial ideal in< KQ , where < is a
monomial order in which the y-variables are much more expensive than the x-variables. This presentation
will suffice to prove Theorems 1.10 and 1.15. We first require the following standard lemma.
Lemma 2.11. Let J and L be ideals in a polynomial ring. Suppose that J ⊆ L and in< J = in< L. Then
J = L.
Proof. Let G be a Gröbner basis for J and let f ∈ L. Since in< G generates in< L, dividing f by G gives a
remainder of 0, and so f ∈ J.
Theorem 2.12. Given an ice quiver Q on n vertices, the defining polynomials together with the cycle
polynomials form a Gröbner basis for KQ = ker π, where
±1
±1
π : Z[x1 , x2 , ..., xn , y1 , y2 , ...yn ] −→ Z[x±1
1 , x2 , ..., xn ]
∀i, π(xi ) = xi , π(yi ) = x′i .
Proof. Let J be the ideal of Z[x1 , x2 , . . . , xn , y1 , y2 , . . . , yn ] that is generated by the set G of defining and
cycle polynomials. We know that J ⊆ KQ , and therefore that in< J ⊆ in< KQ . Let M be the monomial
ideal generated by the initial terms of the polynomials in G. We know that M ⊆ in< J ⊆ in< KQ , so we
would like to show that in< KQ ⊆ M .
Assume (for the purpose of contradiction) that there is some f ∈ KQ such that in< (f ) 6∈ M . We may
write (assume all a, b and λ non-zero for simplicity)
(2.13)
a
ai
a
b
b
bj
in< (f ) = λxi1i1 xi2i2 · · · xik k yj1j1 yj2j2 · · · yjℓ ℓ
Note that {j1 , j2 , ..., jℓ } cannot contain the indices of a directed cycle of unfrozen vertices. Otherwise, it
would also contain the indices of a vertex-minimal directed cycle, and in< (f ) would be a multiple of the
initial term of a cycle polynomial, contradicting the assumption that in< (f ) 6∈ M .
8
Let Y ⊂ [n] be the indices of unfrozen vertices which are not in {j1 , j2 , ..., jℓ }, and let Q′ be the ice quiver
obtained by freezing the vertices in Q indexed by Y . By the preceding observation, Q′ is an acyclic quiver.
There is a natural inclusion
L(Q) ֒→ L(Q′ )
±1
induced by inclusions into Z[x±1
1 , ..., xn ]. This inclusion may be lifted to a ring homomorphism
µ : Z[x1 , ..., xn , y1 , ..., yn ] → Z[x1 , ..., xn , y1 , ..., yn ]
yi (p+
+ p−
) if i ∈ Y
i
i
µ(xi ) = xi , µ(yi ) =
yi
otherwise
′
′
with the property that π ◦ µ = π, where π is the map
±1
±1
Z[x1 , x2 , ..., xn , y1 , y2 , ...yn ] −→ Z[x±1
1 , x2 , ..., xn ]
defined by Q′ instead of Q.
Each of the variables appearing in the initial term of f are fixed by µ. In lower-order terms of f , µ may
introduce monomials in x; however, this will never create a term greater than in< (f ). Hence,
in< (f ) = in< (µ(f )) ∈ in< (KQ′ )
′
Since Q is acyclic, it was shown in [BFZ05, Corollary 1.17] that in< (KQ′ ) is generated by {xi yi | i ∈ [n]}.
Hence, in< (f ) is a multiple of xi yi for some i. However, this implies that in< (f ) is a multiple of the initial
term of the ith defining polynomial in KQ , contradicting the assumption that in< (f ) 6∈ M .
It follows that in< (KQ ) ⊂ M . This consequently implies that in< (J) = in< (KQ ) and, by the preceding
lemma, that J = KQ . Furthermore, since in< (G) generates in< (KQ ), the set G is a Gröbner basis for KQ .
3. Simplicial Complexes and Cohen-Macaulayness
Now that we have obtained a generating set for in< KQ , we can explicitly construct the Stanley-Reisner
complex of in< KQ . We first consider a larger class of simplicial complexes, defined as follows:
Definition 3.1. Let S = {1, . . . , n}, let C be a collection {C1 , . . . , Ck } of subsets of S, and let Y ⊆ S.
Define the simplicial complex ∆(S, C , Y ) on the set {xi | i ∈ S} ∪ {yi | i ∈ Y } by the rule7
F ∈ ∆(S, C , Y ) ⇐⇒ ∀i {xi , yi } 6⊆ F and ∀j {yi }i∈Cj 6⊆ F.
Since every facet of ∆(S, C , Y ) is of the form {z1 , . . . , zn }, where zi is either xi or yi , we see that ∆(S, C , Y )
is always a pure simplicial complex. Note that for any quiver Q on vertex set S, where C is the collection
of sets of vertices of vertex-minimal directed cycles on Q, we have by Theorem 2.12 that
+
*
Y
Y
(3.2)
in< KQ = x1 y1 , . . . , xn yn ,
yi , Cj ⊆ S,
yi , . . . ,
i∈C1
i∈Ck
and the Stanley-Reisner complex of in< KQ is precisely ∆(S, C , S).
Remark 3.3. Whenever {i} ∈ C , there is no vertex of the form yi in the simplicial complex ∆(S, C , Y ).
Such confusing notation is necessary for later induction. A vertex in ∆(S, C , Y ) of the form yi will be called
a y-vertex.
We now recall some definitions. Given a simplicial complex ∆ and a vertex v of ∆, the link of v is the set
link∆ (v) := {F ∈ ∆ | F 6∋ v and F ∪ {v} ∈ ∆},
and the deletion of v is the set
/ ∆},
del∆ (v) := {F ∈ ∆ | F ∪ {v} ∈
where the bar denotes closure, so that del∆ (v) is a simplicial complex. We call a vertex v of a simplicial
complex ∆ a shedding vertex of ∆ if no face of link∆ (v) is a facet of del∆ (v). Finally, we recall the
(recursive) notion of vertex-decomposability: a simplicial complex ∆ is vertex-decomposable if it is a
7If a subset C is not contained in Y , we may simply ignore the condition {y }
j
i i∈Cj 6⊆ S, which is vacuously true because yi
is not defined for i 6∈ Y .
9
simplex, or if it has a shedding vertex v such that both link∆ (v) and del∆ (v) are vertex-decomposable (see
[BP79], also [BW97]).
It is our goal to prove the following theorem, from which Theorem 1.18 will follow.
Theorem 3.4. The complex ∆(S, C , Y ) is always homeomorphic to a vertex-decomposable (n−1)-ball, except
when C = ∅ and Y = S, in which case ∆(S, C , Y ) is homeomorphic to a vertex-decomposable (n − 1)-sphere.
Note that the case where C = ∅ and Y = S is precisely the case in which {yj }j∈S is a face of ∆(S, C , Y ).
We first characterize the link and the deletion in ∆(S, C , Y ) for any vertex of the form yi for i ∈ Y .
Proposition 3.5. For y-vertex yi in ∆(S, C , Y ), we have link∆(S,C ,Y ) (yi ) = ∆(S i , C i , Y i ), where S i :=
S r {i}, C i := {Cj ∩ Si | Cj ∈ C }, and Y i := Y ∩ S i .
Proof. We first show Link := link∆(S,C ,Y ) (yi ) ⊆ ∆(S i , C i , Y i ). Since Link is a subcomplex of ∆(S, C , Y ),
no face of Link contains {xj , yj } for any j, since ∆(S, C , Y ) is defined so as never to contain any such
face. Since no face of Link may contain either xi or yi , we see that Link is a simplicial complex on
{xj | j ∈ S i } ∪ {yj | j ∈ Y i }. Finally, were some F ∈ Link to contain {yj }j∈Cℓi for some ℓ, then F ∪ {yi }
would contain {yj }j∈Cℓ , contradicting F ∪ {yi } ∈ ∆(S, C , Y ). We now have that Link ⊆ ∆(S i , C i , Y i ).
We now show ∆(S i , C i , Y i ) ⊆ Link. Consider some F ∈ ∆(S i , C i , Y i ). Clearly F 6∋ yi , so suppose that
F ∪ {yi } ∈
/ ∆(S, C , Y ). Then either {xi , yi } ⊆ F ∪ {yi } or {yj }j∈Cℓ ⊆ F ∪ {yi } for some ℓ. The former case is
impossible since xi ∈
/ S i . The latter case implies {yj }j∈Cℓi ⊆ F , contradicting the definition of ∆(S i , C i , Y i ).
Therefore we must have F ∪ {yi } ∈ ∆(S, C , Y ) for every face F of ∆(S i , C i , Y i ), from which it follows that
∆(S i , C i , Y i ) ⊆ Link. We conclude that ∆(S i , C i , Y i ) = Link.
Proposition 3.6. For y-vertex yi in ∆(S, C , Y ), we have del∆(S,C ,Y ) (yi ) = ∆(S, C , Y i ), where Y i is defined
as above.
Proof. We first show Del := del∆(S,C ,Y ) (yi ) ⊆ ∆(S, C , Y i ). Since no face of Del may contain yi , we see that
Del is a simplicial complex on {xj | j ∈ S} ∪ {yj | j ∈ Y i }. Since Del is a subcomplex of ∆(S, C , Y ), no
face of Del contains either {xj , yj } for any j or {yj }j∈Cℓ for any ℓ. Therefore Del ⊆ ∆(S, C , Y i ).
Since ∆(S, C , Y i ) has xi as a vertex but not yi , we have by the definition of ∆(S, C , Y i ) that every
facet of ∆(S, C , Y i ) contains xi . Consider some arbitrary facet F of ∆(S, C , Y i ). Since xi ∈ F , we cannot
have F ∪ {yi } ∈ ∆(S, C , Y ), and so F ∈ Del. Therefore Del ⊆ ∆(S, C , Y i ), and so we conclude that
Del = ∆(S, C , Y i ).
We may now observe an important relationship between links and deletions that arises in our case. The
following result shows that any vertex yi is a shedding vertex. Note that in the case where C = ∅ and
Y = S, every vertex yi must be a shedding vertex because its link is always empty.
Lemma 3.7. Except in the case where C = ∅ and Y = S, the complex ∆(S i , C i , Y i ) is properly contained
in the boundary complex ∂∆(S, C , Y i ).
Proof. Since ∆(S i , C i , Y i ) ⊆ ∆(S, C , Y i ) and ∆(S, C , Y i ) is pure, it follows that every facet of ∆(S i , C i , Y i )
meets at least one facet of ∆(S, C , Y i ). Now we show that each facet of ∆(S i , C i , Y i ) meets only one
facet of ∆(S, C , Y i ). Observe that the facets of ∂∆(S, C , Y i ) are characterized as the codimension 1 faces
{zj }j6=k , k ∈ {1, . . . , n}, where each zj is either xj or yj , such that exactly one of either {zj }j6=k ∪ {xk } or
{zj }j6=k ∪ {yk } lies in ∆(S, C , Y i ). Every facet of ∆(S i , C i , Y i ) is of the form {zj }j6=i , and it always happens
that {zj }j6=i ∪{xi } ∈ ∆(S, C , Y i ) and {zj }j6=i ∪{yi } ∈
/ ∆(S, C , Y i ). Therefore ∆(S i , C i , Y i ) ⊆ ∂∆(S, C , Y i ).
We now must show that this containment is proper. We have two cases. Either {yj }j6=i is a face of
∆(S, C , Y i ) or it is not. If it is, then it must lie on ∂∆(S, C , Y i ), because {yj }j6=i ∪ {xi } is a face of
∆(S, C , Y i ), while {yj }j6=i ∪ {yi } is not a face of ∆(S, C , Y i ) since either C 6= ∅ or Y 6= S. If {yj }j6=i is not
a face of ∆(S, C , Y i ), then there must be some C ∈ C not containing i such that no other member of C is
i
a subset of C. Then, for any k ∈ C, we have that F = {xj }j ∈C
/ ∪ {yj }j∈Cr{k} is a face of ∆(S, C , Y ). This
face F lies on ∂∆(S, C , Y i ), since F ∪ {xk } ∈ ∆(S, C , Y i ) but F ∪ {yk } ∈
/ ∆(S, C , Y i ). Since both {yj }j6=i
i
i
i
and F contain xi , neither is a face of ∆(S , C , Y ). Therefore there is always an element of ∂∆(S, C , Y i )
that is not in ∆(S i , C i , Y i ), and so the containment ∆(S i , C i , Y i ) ⊆ ∂∆(S, C , Y i ) is proper.
10
By the previous lemma and the remarks above, we see that every ∆(S, C , Y ) is vertex-decomposable,
because the y-vertices are always shedding vertices, and any complex without y-vertices is a simplex. The
remainder of the proof is to strengthen this argument to prove that these simplicial complexes are balls or
spheres, as appropriate.
Proof of Theorem 3.4. First, we prove that ∆(S, C , Y ) is a vertex-decomposable (n − 1)-ball when Y 6= S
or C 6= ∅, by induction on the number of y-vertices.
If there are no y-vertices in ∆(S, C , Y ) (that is, {i} ∈ C for all i ∈ Y ), then ∆(S, C , Y ) is just one simplex
on n vertices, which is homeomorphic to an (n − 1)-ball. Assume the inductive hypothesis holds whenever
there are fewer than m y-vertices, and assume that ∆(S, C , Y ) has m-many y-vertices. Choose a vertex yi
in ∆(S, C , Y ), and define
Link := link∆(S,C ,Y ) (yi ) = ∆(S i , C i , Y i ) and Del := del∆(S,C ,Y ) (yi ) = ∆(S, C , Y i )
We observe that both Link and Del satisfy the inductive hypothesis; this is clear when Y 6= S. If Y = S,
then by assumption C 6= ∅. Since yi is a vertex of ∆(S, C , Y ), we also know that {i} 6∈ C . It follows
that C i 6= ∅, and so Link still satisfies the inductive hypothesis. Therefore, Link is a vertex-decomposable
(n − 2)-ball and Del is a vertex-decomposable (n − 1)-ball.
As a consequence, the cone Cone from yi on link∆(S,C ,Y ) (yi ) is a vertex-decomposable (n − 1)-ball. By
Lemma 3.7, Cone and Del meet at the proper subset Link of ∂Del, which is a vertex-decomposable (n − 2)ball. Therefore, ∆(S, C , Y ) = Cone ∪ Del is a vertex-decomposable (n − 1)-ball, completing the induction.
The remaining case is ∆(S, ∅, S). Consider the mapping {x1 , . . . , xn , y1 , . . . , yn } → Rn given by xi 7→ ei
and yi 7→ −ei , where {e1 , . . . , en } is the standard basis for Rn . This mapping induces a bijection between the
faces of ∆(S, ∅, S) and the faces of the cross-polytope (i.e. orthoplex) on the vertices {e1 , . . . , en , −e1 , . . . , −en }.
Since this figure is homeomorphic to an (n − 1)-sphere, so must be ∆(S, ∅, S).
As noted, by Theorem 2.12 we have that the initial ideal of any lower bound ideal is of the form (3.2).
Therefore, by Theorem 3.4, the Stanley-Reisner complex of the initial ideal of any lower bound ideal is
homeomorphic to either a ball or a sphere. It follows that all lower bound algebras over a field are CohenMacaulay, and so Theorem 1.18 holds.
4. Normality of lower bound algebras
In this section, we prove that all lower bound algebras defined from a quiver are normal. As explained in
the introduction, the case where Q is acyclic follows immediately because L(Q) is equal to its upper cluster
algebra, and is therefore normal. So, for the remainder of the section, we assume that Q contains a cycle.
In this case, our proof of normality relies on a very slight adaptation of [KLS14, Proposition 8.1].
Proposition 4.1. (cf. [KLS14, Proposition 8.1]) Fix a monomial order < on the polynomial ring K[z1 , . . . , zn ].
Let X, and Y1 , . . . , Yr be (reduced and irreducible) closed affine subvarieties of An , where each of Y1 , . . . , Yr
are codimension-1 in X. Assume that, with respect to the term order <, each of X and Y1 , . . . , Yr Gröbner
degenerate to Stanley-Reisner schemes. Then, if
(i) the Stanley-Reisner complex of in< X is a simplicial ball;
(ii) the Stanley-Reisner complex of each in< Yi lies entirely on the boundary sphere ∂∆X ; and
(iii) X r (Y1 ∪ Y2 ∪ · · · ∪ Yr ) is normal,
then X is normal.
We need the following standard result to prove this proposition. It is very similar to [BC03, Proposition
3.1 (b)]; we provide the necessary modifications in the proof below.
Lemma 4.2. Fix a monomial order < on the polynomial ring S := K[z1 , . . . , zn ]. Let X and Y be irreducible
affine subvarieties of An , and assume that Y is codimension-1 in X. Then if in< X is generically regular
along each irreducible component of in< Y then X is generically regular along Y .
Proof. Let X = V(I) and Y P
= V(J) for I, J ⊆ K[z1 , . . . , zn ]. Pick a weight vector λ such that inλ I = in< I
and inλ J = in< J. Let f = i ai mi , where each ai ∈ K, and each mi is a monomial. Let homλ (f ) denote
the λ-homogenization of f inside S[t], that is,
X
homλ (f ) :=
ai mi tλ(f )−λ(mi ) ,
11
where λ(f ) denotes the highest λ-weight of any monomial in f , and λ(mi ) is the λ-weight of the monomial mi .
Let homλ I denote the λ-homogenization of the ideal I, that is, homλ I := hhomλ (f ) | f ∈ Ii. It is a standard
fact that A := S[t]/homλ I is a free K[t]-module and that A/hti ∼
= S/inλ I (see eg. [BC03, Proposition 2.4]
or [Eis95, Theorem 15.17]).
Now, by assumption, in< X is generically regular along each irreducible component of in< Y . That is, the
localization of S/in< I at any minimal prime of in< J is a regular local ring. Thus, by the above facts, we
have that the localization of A/hti at any minimal prime of (homλ J + hti) is a regular local ring.
Observe that A is positively graded. Let m denote the maximal ideal generated by the indeterminates
z1 , ..., zn , t, and let Am denote the localization at m. Because A/hti localized at any minimal prime p of
(homλ J + hti) is regular, so too is Am /hti localized at any non-trivial Am p, and the non-trivial Am p are
precisely the minimal primes of (homλ J + hti) as an ideal in Am /hti.
Now we can use the proof of [BC03, Lemma 3.2] to get that the localization of Am at the height-1 prime
ideal homλ J is regular. The second half of the proof of [BC03, Proposition 3.1 (b)] then gives that the
localization of S/I at J is regular.
We now prove Proposition 4.1.
Proof of Proposition 4.1. We follow the proof given in [KLS14, Proposition 8.1]. To show that X is normal,
we need to show that X is R1 and S2. Since ∆X is a simplicial ball by assumption (i), it follows that X is
Cohen-Macaulay, and hence S2. To show that X is R1, first observe that, by assumption (iii), if p ⊆ S/I is
a prime ideal of height ≤ 1 which is not the generic point of any Yi , then (S/I)p is regular.
The remaining primes in S/I which have height ≤ 1 are the generic points of the various Yi ⊆ X. It
therefore remains to show that X is generically regular along each irreducible subvariety Yi . By assumption
(ii), we have that in< X is generically regular along each irreducible component of each in< Yi . Thus, by the
lemma, we get that X is generically regular along Yi .
To use Proposition 4.1 to prove that all lower bound algebras are normal, we need to show that the
hypotheses (ii) and (iii) always hold for quivers with cycles. We start with (ii). To show the desired result,
we use results of Knutson from [Knu09]8.
Theorem 4.3. (cf. Theorem 2, Lemma 6, Corollary 2 of [Knu09]) Let f ∈ Z[z1 , . . . , zn ] be a polynomial
with the property that, for each prime p, f p−1 (mod p) has a unique term divisible by z1p−1 z2p−1 · · · znp−1 , and
let < be a term order of Z[z1 , . . . , zn ] for which in< f = z1 z2 · · · zn . Denote by J the smallest set of ideals
that contains the ideal hf i and such that
(i) if I1 , I2 ∈ J , then I1 + I2 , I1 ∩ I2 ∈ J ; and
(ii) if I ∈ J and J is a primary component of I then J ∈ J .
Then, over any field K, every ideal J ∈ J is a radical ideal and the initial ideal of every J ∈ J with respect
to < is a squarefree monomial ideal. Furthermore, for any I1 and I2 in J ,
in< (I1 ∩ I2 ) = in< I1 ∩ in< I2 , and in< (I1 + I2 ) = in< I1 + in< I2 .
Qn
−
We will make use of this theorem in the case where f = i=1 (xi yi − p+
i − pi ). Here we take < to be
a weighting of the variables where the y-variables are much more expensive than the x-variables. Observe
that f and < satisfy the assumptions of Theorem 4.3 and the ideal
−
IQ := hxi yi − p+
i − pi | 1 ≤ i ≤ ni.
lies in the collection of ideals J from the theorem. Consequently I is radical. We may then write an
irredundant prime decomposition
(4.4)
IQ = KQ ∩ P1 ∩ · · · ∩ Pr
where each Pi is a minimal prime, and KQ is the lower bound ideal [BMRS15, Lemma 5.7]. Consequently,
each Pi + KQ ∈ J and so each Pi + KQ is radical and degenerates to a squarefree monomial ideal.
8The statement given here is less general than the one that appears in [Knu09, Theorem 2] and [Knu09, Lemma 6]. We also
change the hypotheses of [Knu09, Theorem 2], however, this is harmless as the proof goes through in the exact same way.
12
Proposition 4.5. Let Q be a quiver with a directed cycle, so that there is at least one prime Pi in (4.4).
With respect to a term order where the y-variables are much more expensive than the x-variables, each prime
component of Pi + KQ Gröbner degenerates to the Stanley-Reisner ideal of a sub-simplicial complex of ∂∆KQ .
Furthermore, in< ((P1 ∩ · · · ∩ Pr ) + KQ ) is the Stanley-Reisner ideal of the entire boundary ∂∆KQ .
Proof. By Theorem 4.3, we have that Pi + KQ is radical and Gröbner degenerates to a squarefree monomial
ideal. Let
KQ + Pi = ∩J J
be a decomposition of KQ + Pi into minimal primes. By Theorem 4.3, each in< J is a squarefree monomial
ideal. Applying the second part of Theorem 4.3 and translating into the language of simplicial complexes
yields the equality
∆KQ +Pi = ∆KQ ∩ ∆Pi = ∪J ∆J
where ∆J denotes the Stanley-Reisner complex in< J. To prove the first claim of the proposition, we must
show that every face of each ∆J is contained in the boundary sphere of the simplicial ball ∆KQ . So, suppose
otherwise, and let F be a face of some ∆J which is not contained in the boundary ∂∆KQ . Assume that F is
a maximal such face. We claim that F must be a facet of ∆Pi .
To prove this claim, we first apply Theorem 4.3 to the prime decomposition in equation (4.4) to get
(4.6)
in< I = in< (KQ ) ∩ in< (P1 ) ∩ · · · ∩ in< (Pr )
which, after translating into the language of simplicial complexes, says that ∆KQ and every ∆Pi is contained
in the Stanley-Reisner complex associated to in< I = hxi yi | 1 ≤ i ≤ ni, which can be geometrically realized
as the (n − 1)-dimensional boundary sphere of a cross-polytope on 2n vertices. Decompose this simplicial
sphere into the union of two (n − 1)-dimensional simplicial balls:
∆I = ∆KQ ∪ C, where C := ∆I r ∆KQ .
Observe that, by construction, ∆KQ ∩ C is the boundary sphere of ∆KQ .
Now, suppose that F is not a facet of ∆Pi . Then there is a vertex z such that F ∪ {z} is a face of ∆Pi .
Then, using the decomposition of ∆I , we see that either F ∪ {z} is contained in ∆KQ , or it is contained in
C. If F ∪ {z} ⊆ ∆KQ , we contradict the maximality of F . If F ∪ {z} ∈ C, we contradict that F was not
contained in the boundary of ∆KQ (since ∆KQ and C only intersect along the boundary of ∆KQ ).
Thus, our maximal face F must be a facet of ∆Pi , which, since Pi is prime, must have dimension one less
than the dimension dim(S/Pi ). But this is not possible because dim(S/J) is strictly smaller than dim(S/Pi ).
To obtain the last statement, we translate equality (4.6) into the language of simplicial complexes to see
that the union ∪ri=1 ∆Pi necessarily contains the boundary sphere ∂∆KQ . Thus, so does
∆(P1 ∩···∩Pr )+KQ ) = ∆(P1 ∩···∩Pr ) ∩ ∆KQ = ∪ri=1 (∆Pi ∩ ∆KQ ).
But, as already shown, each ∆Pi ∩ ∆KQ is contained inside of the boundary sphere of ∆KQ and so we are
done.
We next show that (iii) of Proposition 4.1 holds for lower bound algebras.
Proposition 4.7. Let V(KQ ) denote the lower bound variety of a quiver Q. Then V(KQ )rV(P1 ∩P2 ∩· · ·∩Pr )
is normal9.
Proof. Consider q ∈ V(KQ ), and define the (possibly empty) set
Sq := {i ∈ {1, 2, ..., n} | xi (q) = 0}
The vertices indexed by Sq must be unfrozen, since frozen x-variables are invertible.
First, assume Sq does not contain a directed cycle, and consider the open set
Uq := {q ′ ∈ V(KQ ) | ∀i 6∈ Sq , xi (q) 6= 0}
The coordinate ring of Uq is the localization of L(Q) at the set of x-variables which are not in Sq ; hence,
it is isomorphic to the lower bound algebra of the ice quiver Q† obtained by freezing the vertices not in Sq .
This ice quiver is acyclic, and so the lower bound algebra coincides with the upper cluster algebra [BFZ05],
which is normal [Mul13]. Hence, V(KQ ) is normal at q.
9We note that V(P ∩ P ∩ · · · ∩ P ) is empty when Q is acyclic.
r
1
2
13
Next, assume Sq contains a directed cycle, and consider the affine space
Wq := {q ′ ∈ K2n | ∀i, xi (q ′ ) = xi (q), and ∀i 6∈ Sq , yi (q ′ ) = yi (q)}
This contains q and is contained in V(KQ ∩P1 ∩· · ·∩Pr ). Since Sq contains a directed cycle, there is some cycle
polynomial whose leading term is a product of y-variables whose indices are contained in Sq , and hence cannot
vanish on Wq . Hence, Wq 6⊂ V (KQ ). By the irreducibility of Wq , we have that q ∈ Wq ⊂ V(P1 ∩ · · · ∩ Pr ),
and so q 6∈ V(KQ ) r V(P1 ∩ · · · ∩ Pr ).
We can now prove that all lower bound algebras are normal.
Proof of Theorem 1.21. We have already treated the case when Q is acyclic (i.e. L(Q) equals its associated
upper cluster algebra, and is hence normal). So assume that Q contains a cycle. Let KQ be the relevant
−
lower bound ideal, and let P1 , . . . , Pr ⊆ S be minimal primes of hxi yi − p+
i − pi | 1 ≤ i ≤ ni such that
−
hxi yi − p+
i − pi | 1 ≤ i ≤ ni = KQ ∩ P1 ∩ · · · ∩ Pr .
Now apply Proposition 4.1 with X = V(KQ ), and Y1 , . . . , Ys the irreducible components of the various
V(Pj + KQ ), 1 ≤ j ≤ r. Observe that item (i) of Proposition 4.1 follows from Theorem 1.18, item (ii) follows
from Proposition 4.5, and item (iii) holds by Proposition 4.7.
Appendix A. The singular locus of L(Q)
The Cohen-Macaulayness and normality of L(Q) may both be regarded as constraints on the singularities
of the variety V(KQ ). It is then natural to directly consider the singularities of V(KQ ). This appendix
provides an example to demonstrate that, even in the most elementary cases, even the existence of singularities in V(KQ ) can be difficult to predict, suggesting a direct study of of the singularities of V(KQ ) may be
daunting.
Fix an algebraically closed field K. Let Qn denote the ice quiver with vertex set {1, 2, ..., n}, an arrow
from i to i + 1 for each 1 ≤ i < n, and no frozen vertices (see Figure 5).
1
2
3
n
Figure 5. The quiver Qn
Proposition A.1. The K-variety V(KQn ) has a unique singular point when n ≡ 3 (mod 4), and no singularities otherwise.
Proof. Since there are no directed cycles, we have the following presentation of L(Qn ).
L(Qn ) = K[x1 , ..., xn , y1 , .., yn ]/hy1 x1 − x2 − 1, y2 x2 − x3 − x1 , ..., yn−1 xn−1 − xn − xn−2 , yn xn − 1 − xn−1 i
Let p ∈ V(KQn ). First, we observe that xi (p) and xi+1 (p) cannot both be zero. This is clear for i = 1 from
the defining relation for y1 , and the general case follows by induction on i.
The point p is singular if and only if the associated Jacobian matrix Jacp has rank less than n.
y1 (p) −1
0
0
· · · x1 (p)
0
0
0
···
−1 y2 (p) −1
0
···
0
x2 (p)
0
0
···
0
−1
y
(p)
−1
·
·
·
0
0
x
(p)
0
·
··
3
3
Jacp :=
0
0
−1 y4 (p) · · ·
0
0
0
x4 (p)
..
..
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
.
.
.
Equivalently, p is singular if and only if there is a non-trivial linear relation among the rows of Jacp . Clearly,
such a relation can only include the ith row if xi (p) = 0; hence, such a relation can only involve non-adjacent
rows of Jacp . This is only possible when n is odd and when xi (p) = yi (p) = 0 for every odd i; furthermore,
the relation (up to scaling) must be that the alternating sum of the odd rows of Jacp is 0.
14
Returning to the defining relations, the condition that xi (p) = 0 for all odd i is only possible when n
is congruent to 3 mod 4, in which case x2j (p) = (−1)j . Since these are non-zero, it follows that y2i (p) =
x2i−1 (p)+x2i+1 (p)
= 0. Consequently, when n ≡ 3 (mod 4), the unique singular point is given by
x2i(p)
0
if i is odd
∀i ∈ {1, 2, ..., n}, xi (p) =
, yi (p) = 0
i
(−1) 2 if i is even
If n 6≡ 3 (mod 4), there is no singular point.
The family of algebras L(Qn ) is one of the most fundamental and elementary in the theory of cluster
algebras; in this case, the lower bound L(Qn ) coincides with the cluster algebra of Dynkin-type An . This
and other simple examples suggest that that presence of singularities in V(KQ ) is difficult to predict, and is
very sensitive to small changes in the quiver Q.
Appendix B. Skew-symmetrizable lower bound cluster algebras
The body of this paper considered lower bound cluster algebras defined by ice quivers, also called skewsymmetric lower bound cluster algebras. However, the proofs and results can be extended to the larger
generality of skew-symmetrizable lower bound cluster algebras.10 This appendix outlines the necessary modifications. A skew-symmetrizable lower bound algebra is defined by a exchange matrix B: an n × m integer
matrix (for n ≥ m) such that there is an m × m diagonal matrix D with the property that the top m × m
minor of BD is skew-symmetric.
−
To each index i ∈ {1, 2, ..., m}, we associate a pair of monomials p+
i , pi ∈ Z[x1 , x2 , ..., xn ] as follows.
Y
Y
max(−Bji ,0)
max(Bji ,0)
xj
xj
,
p−
(B.1)
p+
i :=
i :=
j∈{1,2,...,n}
j∈{1,2,...,n}
x′i ,
Each index then determines a Laurent polynomial
called the adjacent cluster variable at i, which is
defined by the following formula.11
−1 +
xi (pi + p−
)
if i ∈ {1, 2, ..., m}
′
i
:=
(B.2)
xi
x−1
if i ∈ {m + 1, m + 2, ..., n}
i
±1
The lower bound algebra L(B) defined by B is the subalgebra of Z[x±1
1 , ..., xn ] generated by the variables
′
′
′
x1 , x2 , ..., xn and the adjacent cluster variables x1 , x2 , ..., xn .
Example B.3. Consider the following 3 × 2 matrix B, which is skew-symmetrizable with the given D.
0 3
3 0
B := −2 0 , D :=
0 2
1 2
The three adjacent cluster variables are below.
x′1 =
x3 x2 + 1
x3 + x22
, x′2 = 1 3
, x′3 = x−1
3
x1
x2
To translate prior results into the generality of exchange matrices, we associate an ice quiver Q(B) to B.
The vertex set of Q(B) is the set {1, 2, ..., n}, the frozen vertex set is {m + 1, m + 2, ..., n}, and there is an
arrow from i to j if Bji > 0 or Bij < 0 (and no other arrows). The lower bound algebra L(B) is not the
lower bound algebra of the ice quiver Q(B); the arrows in Q(B) only keep track of the sign of entries in B.
The results and proof in this paper hold verbatim for L(B), using Q(B) in place of Q, for two key reasons.
• Our arguments never use the specific exponent of the monomials p±
i ; rather, we only need to know
±
if a variable xj divides pi , which corresponds to the existence of an arrow in Q(B).
• When Q(B) is acyclic, the presentation of L(B) given in [BFZ05, Corollary 1.17] and used in the
proof of Theorem 1.15 is still valid in the larger generality of skew-symmetrizable cluster algebras.
10We remark that we still work in geometric type, rather than more exotic semifield coordinates.
11The same apology is due as before. When i > m, the element x′ is not technically a cluster variable, but treating it as
i
one streamlines the process.
15
Remark B.4. The body of the paper is not in this larger generality for reasons of clarity and exposition,
rather than any mathematical limitations.
References
[BC03]
[BDP14]
Winfried Bruns and Aldo Conca, Groebner bases, initial ideals and initial algebras, arxiv:0308102 (2003).
Thomas Brüstle, Grégoire Dupont, and Matthieu Pérotin, On maximal green sequences, Int. Math. Res. Not. IMRN
(2014), no. 16, 4547–4586. MR 3250044
[BFZ05]
Arkady Berenstein, Sergey Fomin, and Andrei Zelevinsky, Cluster algebras. III. Upper bounds and double Bruhat
cells, Duke Math. J. 126 (2005), no. 1, 1–52. MR 2110627 (2005i:16065)
[BMRS15] Angélica Benito, Greg Muller, Jenna Rajchgot, and Karen E. Smith, Singularities of locally acyclic cluster algebras,
Algebra Number Theory 9 (2015), no. 4, 913–936. MR 3352824
[BP79]
Louis J. Billera and J. Scott Provan, A decomposition property for simplicial complexes and its relation to diameters
and shellings, Second International Conference on Combinatorial Mathematics (New York, 1978), Ann. New York
Acad. Sci., vol. 319, New York Acad. Sci., New York, 1979, pp. 82–85. MR 556009 (81f:52010)
[BW97]
Anders Björner and Michelle L. Wachs, Shellable nonpure complexes and posets. II, Trans. Amer. Math. Soc. 349
(1997), no. 10, 3945–3975. MR 1401765 (98b:06008)
[Eis95]
David Eisenbud, Commutative algebra, Graduate Texts in Mathematics, vol. 150, Springer-Verlag, New York, 1995,
With a view toward algebraic geometry. MR MR1322960 (97a:13001)
[FZ02]
Sergey Fomin and Andrei Zelevinsky, Cluster algebras. I. Foundations, J. Amer. Math. Soc. 15 (2002), no. 2,
497–529 (electronic). MR 1887642 (2003f:16050)
[FZ07]
, Cluster algebras. IV. Coefficients, Compos. Math. 143 (2007), no. 1, 112–164. MR 2295199 (2008d:16049)
[GHKK14] Mark Gross, Paul Hacking, Sean Keel, and Maxim Kontsevich, Canonical bases for cluster algebras, preprint,
arxiv:1411.1394 (2014).
[GLS08]
Christof Geiss, Bernard Leclerc, and Jan Schröer, Partial flag varieties and preprojective algebras, Ann. Inst. Fourier
(Grenoble) 58 (2008), no. 3, 825–876. MR 2427512 (2009f:14104)
[GSV05]
Michael Gekhtman, Michael Shapiro, and Alek Vainshtein, Cluster algebras and Weil-Petersson forms, Duke Math.
J. 127 (2005), no. 2, 291–311. MR 2130414 (2006d:53103)
[KLS14]
Allen Knutson, Thomas Lam, and David E. Speyer, Projections of Richardson varieties, J. Reine Angew. Math.
687 (2014), 133–157. MR 3176610
[Knu09]
Allen Knutson, Frobenius splitting, point-counting, and degeneration, preprint, arxiv:0911.4941 (2009).
[MS05]
Ezra Miller and Bernd Sturmfels, Combinatorial commutative algebra, Graduate Texts in Mathematics, vol. 227,
Springer-Verlag, New York, 2005. MR 2110098 (2006d:13001)
[Mul13]
Greg Muller, Locally acyclic cluster algebras, Adv. Math. 233 (2013), 207–247. MR 2995670
, A=U for locally acyclic cluster algebras, SIGMA 10 (2014), no. 094, 8.
[Mul14]
, The existence of a maximal green sequence is not invariant under quiver mutation, preprint,
[Mul15]
arxiv:1503.04675 (2015).
[Mun84]
James R. Munkres, Topological results in combinatorics, Michigan Math. J. 31 (1984), no. 1, 113–128. MR 736476
(85k:13022)
[Sco06]
J. S. Scott, Grassmannians and cluster algebras, Proc. London Math. Soc. (3) 92 (2006), no. 2, 345–380.
MR 2205721 (2007e:14078)
(Greg Muller) University of Michigan
E-mail address: [email protected]
(Jenna Rajchgot) University of Michigan
E-mail address: [email protected]
(Bradley Zykoski) University of Virginia
E-mail address: [email protected]
16
| 0math.AC
|
Systems of natural-language-facilitated human-robot
cooperation: A review
Rui Liu, Xiaoli Zhang*
arXiv:1701.08269 [cs.RO] 28 Jan 2017
Department of Mechanical Engineering, Colorado School of Mines, Golden, CO 80401, USA
{rliu, xlzhang}@mines.edu
Abstract: Natural-language-facilitated human-robot cooperation (NLC), in which natural language (NL) is used to
share knowledge between a human and a robot for conducting intuitive human-robot cooperation (HRC), is
continuously developing in the recent decade. Currently, NLC is used in several robotic domains such as
manufacturing, daily assistance and health caregiving. It is necessary to summarize current NLC-based robotic
systems and discuss the future developing trends, providing helpful information for future NLC research. In this
review, we first analyzed the driving forces behind the NLC research. Regarding to a robot’s cognition level during
the cooperation, the NLC implementations then were categorized into four types “NL-based control, NL-based robot
training, NL-based task execution, NL-based social companion” for comparison and discussion. Last based on our
perspective and comprehensive paper review, the future research trends were discussed.
Keywords: robotic systems, natural language, human-robot cooperation, control, task execution, social companion
1. Introduction
1.1 Advantages
Natural-language-facilitated human-robot cooperation (NLC), in which a human uses either spoken or written
instructions to communicate with a robot for task cooperation [1][2][3][4], has received increasing attention in humaninvolved robotics research. By using natural language (NL), human intelligence at high-level mental planning and
robot capability at low-level physical execution are combined to perform an intuitive cooperation [5][6].
Compared with human-robot cooperation (HRC) using tactile indications such as contact location [7] and force
strength [8][9][10], and visual indications such as body pose [11][12][13] and motion [14][15][16], HRC using NL
indications has several advantages. First, NL makes the HRC natural. For the traditional methods mentioned above,
the human involved in HRC needs to be trained to use certain actions/poses for making themselves understandable
[17][18][19][20][21] . While in NLC, even non-expert users without prior training can cooperate with a robot by using
human-like communication [3][22][23] . Second, the cooperation request is described accurately. The traditional
methods using actions/poses only provide limited patterns to roughly describe cooperation requests due to the
information loss in the action/pose simplification (such as using markers to simplify the actions) [24][25][26][27].
While in NLC, cooperation requests related to action, speed, tool and location are already defined in NL expressions
[5][28][29][30]. With these expressions, cooperation requests for various task executions are described accurately.
Third, NL transfers cooperation requests efficiently. The information-transferring method using actions/poses requires
the design of informative patterns for different cooperation requests [24][25][26][27]. While existing languages, such
as English, Chinese and German, already have standard linguistic structures, which contain abundant informative
expressions to serve as patterns [31][32]. NL-based methods do not need to design specific informative patterns for
various cooperation requests, making HRC efficient. Lastly, since the instructions are delivered orally instead of being
physically involved, human hands are set free to perform more important executions [33]. Attracted by the above
advantages, NLC has been widely explored in areas, including daily assistances [1][33][35], medical caregiving
[36][37][38][39], manufacturing [5][40][41], indoor/outdoor navigation [2][42][43], social accompany [44][45][46]
etc. Typical areas using NLC systems are shown in Fig. 1.
From a realization perspective, the pushing forces for recent NLC developments are concluded as natural language
processing (NLP) developments and HRC developments.
1.2. Pushing force one: development of NLP
Recently, supported by machine learning technics in classification [47], clustering [48] and feature extraction
[49], NLP has been developed from simply syntax-driven processing, which builds syntax representations of sentence
___________________________
*Corresponding author: Tel.:+1-303-384-2343; fax:+1-303-273-3602, email:[email protected].
Fig. 1. Typical areas using NLC systems. 1.a[66] and 1.b[1] are NL-based robot daily assistance. 2.a [36] and 2.b [37] are NL-based healthcare.
3.a [41] and 3.b [5] are NL-based intelligence manufacturing. 4.a [42] and 4.b [43] are NL-based indoor/outdoor navigation. 5.a [67] and 5.b
[45] are NL-based accompanying.
structures, to semantically-driven processing, which builds semantic networks for sentence meanings [50].
At the early stage (1950-2000) of NLP [51], a word-based understanding method was developed to enable a naïve
word-symbol understanding by identifying single/multiple keywords [52][53], lexical affinities [54][55], and
word/affinity occurrences [56][57]. The word-based NLP method separately understood word meanings, and sentence
meaning was unknown. For example, with the word-based method, a robot understood when a human mentioned the
word “cup”, but it did not understand the related requests such as “I need a cup of water”[15]. Moreover, this method
relied on training samples. If the available training samples are limited, thereby leading to the ignorance of some
keywords, the meaning understanding will be poor [51]. These two drawbacks limited robots’ understanding to a
shallow level where only symbolic NL expressions were analyzed [58][59].
Compared with the word-based understanding method, which enabled a shallow-level understanding, a conceptbased method allowed a comprehensive understanding of NL expressions. The concept-based method modeled the
meanings for sentences by exploring the embedded concepts, which mainly included implicit NL indications [60][61],
hierarchical ontologies [62][63], and semantic correlations [64][65]. The concept-based method not only understood
explicit facts such as involvements of action/object/events/persons, but also understood the implicit indications such
as action purpose, object usage, event meaning, and human intentions. The concept-based method was widely used in
research such as [5][6][23] for complicated NL expression analysis. Compared with the understanding supported by
the keyword-based method, the concept-based method endowed a robot with a relatively in-depth understanding of
the meanings in NL expressions. However, due to the limited consideration of implicit logic correlations in NL
expressions, the knowledge represented by the concept-based method cannot model a structural knowledge [5][17][40]
This drawback limited robots’ understanding towards task procedures and task-world correlations and further limited
NLC in practical situations in the real world.
To support a practical implementation of NL understanding, a narrative-based method was developed to create a
more sophisticated knowledge representation in a decision-making-focused [68], real-world-aware [69] and humancognition-imitated manner [70][71]. In this method, the mechanisms of human reasoning and planning, knowledgeto-world mapping, and logic-based human understanding & learning were focused on in the NLP process. Supported
by this method, knowledge was practically used in NLC, further improving robot’s knowledge scalability and humanrobot-cooperation flexibility.
The developments of NLP techniques are shown in Fig. 2, with detailed time labels.
1.3. Pushing force two: development of HRC
In an early period (about 1940s’[72]), humans started to cooperate with robots by using remote controllers,
developing an initial HRC, in which the cooperation requirements for action mapping, task goal mapping, and
cooperation naturalness/effectiveness were not considered. As the tasks became complicated, both the robot and the
human in HRC were assigned with different roles, such as leader-follower and cooperator-cooperator, to perform
different parts of a task with considerations of task goal accomplishments, human-robot communications, robot/human
statuses and physical/mental capabilities. Compared with HRI, which focuses on general interactions (detailed HRI
reviews are in [4][73][74][75]) for physical/mental assistances without task-goal constrains, HRC focuses on specific
cooperation for task fulfillment with task-goal constrains such as task planning and adjusting (detailed HRC reviews
are [72][76][77]). In this paper, we emphasized on HRC, specifically exploring the state-of-the-art robotic systems
Fig. 2. The development of natural language processing (NLP) techniques [51]. From the stage of keyword-based method, concept-based
method, to the narrative-based method, the semantic analysis is becoming sophisticated from word detection to meaning modeling.
using NL to facilitate HRC in various research domains.
Comprehensively, HRC has been developing: it began largely from a low-cognition-level action research where
actions were designed and selected according to human instructions, grew to a middle-cognition-level interaction
research where shallow-level understanding of human motions, activities, tasks and safety concerns were conducted,
and is expanding to a high-cognition-level human-centered engagement research where human’s psychological and
cognitive statuses, such as attention, motivation, emotion and user models, are considered to improve the HRC
effectiveness and naturalness. Increasing human involvements in HRC is shown in Fig. 3.
For action-based HRC, which started from motor-control-based action design, a robot followed simple human
instructions to adjust its actions [78][79]. To make the robot actions natural, human-like motion style was then adopted
in robot action design [80][81]. Though robots’ motion behavior was similar to that of a human, robots’ cooperation
performances were still poor due to the limited action-understanding being insufficient to support its adaptations
towards users/environments [82][83] . To improve robot’s adaptability, action interpretations were added for a robot
in HRC. For example, ‘cup manipulation’ in ‘drinking’ activity meant ‘containing liquid’”[40]. Though the
understanding was still limited into action level, robots’ understanding towards human behaviors in HRC was
improved [84]. Overall, action-based HRC was targeting the design of actions in HRC with consideration for
cooperation efficiency and naturalness. Due to the lack of human behavior understanding, action-based HRC was still
at the imitating level and lacked goal/execution motivation interpretations. Therefore, action-based HRC had a lowlevel cognition requirement towards robots and could not support an intuitive cooperation.
For interaction-based HRC, which started from action-understanding-based movement imitation [85][86], a robot
was required to learn from human demonstrations, understand human movements and develop its own movements.
To improve the understanding of human movements, robots were provided with various informative motion data such
as human action trajectories [87], hand/body poses [88], and bio-signals [89]. To further improve robots’ cooperation
performance, robots were trained to explore the mutual influence between a human and his/her surrounding
environment [90][91][92][93][94]. Overall, interaction-based HRC improved the cooperation from the naïve
predefinition stage to the current intuitive interaction stage. Robot understanding toward human behaviors was
improved by using various informative data and considering multiple influential factors from environments. With a
comprehensive understanding, a robot was closely associated with a human by correctly identifying the cooperation
requests and appropriately providing the assistance.
For engagement-based HRC, individual-level factors, such as individual attentions [95][96], personalities
[97][98], emotions [99][100] and safety factors [83][101] , were considered by robots in the cooperation. With the
engagement-based HRC method, robot cognition was further improved by adapting various individuals. This method
is widely used in present-day HRC research. Given the cognition requirements towards robots in HRC, a natural and
efficient cooperation manner such as NLC is in urgent need.
1.4. Systematic overview of NLC research
With cross-disciplinary technique supports, NLC has been developed from low-cognition-level symbol matching
control, such as using “yes/no” for controlling robotic execution, to high-cognition-level task understanding, such as
“go straight and turn left at the second cross”.
As a result of NLC research, a substantial number of projects were launched, including “collaborative research:
Fig. 3. The human is increasingly involved in HRC [72]. From a perspective of a robot’s cognition level, the HRC systems are mainly
categorized into action-based HRC, interaction-based HRC and engagement-based HRC. Between the different categories, there are overlaps,
showing that the HRC developments happen gradually.
Fig. 4. The annual amount of NLC-related publications since the year 2000 according to our comprehensive paper review. In the past 15 years,
the number of NLC publications are steadily increasing and are reaching a history-high level in current time.
jointly learning language and affordances” in Cornell University [102], “robots that learn to communicate with humans
through natural dialog” in the University of Texas at Austin [103], “collaborative research: modeling and verification
of language-based interaction” in MIT [104], “language grounding in robotics” in University of Washington [105],
“semantic systems” in Lund University [106] etc. NLC research is regularly published in international journals such
as IJRR [107], TRO [108], AI [109] and KBS [110], and international conferences such as ICRA [111], IROS [112]
and AAAI [113]. The publication trend is shown in Fig. 4, where the increasing significance of using NL to facilitate
HRC is reflected by the steadily increasing publication numbers in the recent decade.
Compared with the existing review papers about HRC using gesture/pose [114][115], action/motion [116], and
tactile [117], a review paper about NLC is lacking. Given the promising potential and increasing attention of NLC, it
is necessary to make a summary of the state-of-the-art robotic systems in wide-range domains, revealing the current
research progress and signposting future NLC research. The organization of this review paper is shown in Fig. 5.
2. Typical systems
As a human-robot-combined decision-making method, the NLC is widely implemented in various robotic
systems. According to robot cognition level during the cooperation, NLC-based robotic systems are categorized into
four main types: NL-based control, where only the NL-format control symbols are given and comprehensive
instruction understating is not involved; NL-based robot training, where comprehensive instruction understanding is
required and intuitive task execution is not conducted; human-guided task execution, where comprehensive
understanding of human instructions, practical situation conditions and human intentions are required, and intuitive
task execution is conducted; and NL-based social companion, where the understanding of social norms is required in
addition to the NL-based execution conducting. Summary of the typical NLC systems is shown in Table 1.
Fig. 5. Organization of this review paper. This review systematically introduced the NLC implementations, covering topics such as the
motivations of the NLC research, typical NLC systems, typical NLC methods, and emerging trends. The typical NLC systems were summarized
emphatically by categorizing NLC into four categories, including NL-based control, NL-based robot training, NL-based task execution, and
NL-based social companion. In each of the categories, the typical application scenarios, knowledge manners/formats, advantages and
disadvantages were summarized.
Table 1. Comparison analysis for typical NLC systems
NLC
systems
Application
scenarios
Robot
cognition
level
low
NL-based
control
action selection, manipulation pose adjustment,
navigation routine selection
NL-based
robot
training
execution process explanation (assembly task process,
navigation process), action/motion specification
(object identification/grasping), human instruction
disambiguation (industrial tasks, daily assistance),
human-motion imitating
middle
NL-based
task
execution
assembly in dynamic environments with various users,
navigation in unstructured environments, heavy object
delivering
high
NL-based
social
cooperation
restaurant reception, social distance maintenance,
body language learning during speaking, human-like
object manipulation
highest
Humanrobot role
assignments
leaderfollower
Human
involvements
Robot
involvements
all cognitive
burden
all physical
burden
leaderfollower
maximal
cognitive
burden,
minimal
physical
burden
partial
cognitive/phys
ical burden
minimal
cognitive
burden,
maximal
physical
burden
partial
cognitive/phys
ical burden
partial
cognitive/phys
ical burden
partial
cognitive/phys
ical burden
cooperatorcooperator,
leaderfollower
cooperatorcooperator
3. NL-based control
To set human hands free for other tasks and reduce human’s physical burden, NL was initially used to replace
physical control means, such as joysticks and remote controllers, which required the use of human hands [33]. During
the control, NL played a role as information-delivering media, which contains the human-desired robot executions,
Fig. 6. Typical robotic systems using NL-based control. (a) Control using symbolic words. The joints’ motion directions of a robot arm were
controlled by mapping the symbolic vowels from human’s oral instructions [128]. (b). Control using semantic correlations. By mapping the
semantic correlations of the parts from human’s oral instructions, a robot was orally controlled by a human to perform assembling works [129].
(c). Control using logic structures. The cooking task executions were defined in a logic manner. By mapping these logics from human oral
instructions, robot performed kitchen tasks [130]. (d). Control using environment conditions. During the task execution, a robot considered the
practical environmental conditions such as “object availability, objects’ relative sizes” to perform the tasks such as “grab the small brown box”
[33].
shown in Fig. 6. A human user mainly made decisions in the control process, while a robot was controlled by detecting
and following the controlling commands in human NL requests. From a burden-assessment perspective, during the
whole cooperation process, the human mainly took the cognitive burdens, while the robot mainly took the physical
burdens.
3.1. Typical robotic systems
Control using symbolic words. The idea of using human NL for robot control was proposed at the year 1992. NL
was initially used in robot teleoperation [118], in which the correct robot actions were selected by a human to guide it
to a desired destination. The system in paper [119] used NL instructions to plan task-execution procedures for a robot.
The NL-based control for execution procedure planning was a word-to-action mapping process, which was discrete
and the word-action associations were predefined in the robot’s database. In this process, a human was restricted to
give symbolic and simple NL commands to a robot. The robot needed to accurately detect the symbolic words in
human speech and then associate them with the predefined actions or action sequences. Typical NLC systems using
symbolic word control include manipulation control [119][120] , motion trajectory control [121][122], navigation
location & behavior control [88][123]etc. These works addressed the challenges in speech recognition, word
disambiguation and word-action mapping. NL-based control was used to specify the detailed action-related parameters
such as action direction, movement amplitude, motion speed, force intensity, and hand pose status. The NL-based
control for action specification was the word-to-value mapping process, which was continuous and its value ranges
were predefined in robots’ database. During the development of NL-based control, the mapping rules such as fuzzy
[124]/strict [118] mappings were designed.
Control using semantic correlations. To further improve the robustness of NL-based control during HRC,
semantic correlations among the controlling symbols were explored by analyzing the linguistic structures of the NL
commands. Controlling symbols include different types of actions, hand poses, objects etc. By exploring the semantic
correlations, such as “grasp-cup” and “go-To-left” among these symbols [125], human commonsense was initially
involved in robot control process, increasing the naturalness of the robot executions in scenarios such as navigations
and manipulations. Moreover, the execution flexibility of a robot was also improved by extracting the general
cooperation patterns such as “grasp(object)” [119][126] and “goTo(Location)” [2][127], improving robots’
adaptability towards instruction variety.
Control using logic structures. Even though semantic correlation improved the flexibility of NL-execution
mapping, the control performances in dynamic situations was still limited due to the ignorance of control logics among
these semantic correlations. The logic ignorance made a robot incapable of adjusting itself to environmental changes
and incapable of intuitively reasoning about the execution plans. For example, the NL instruction “fill the cup with
water, deliver it to human” includes logics “search cup, then use water pot, last deliver cup”. The ignorance of the
logics in the controlling process will lead to the wrong executions such as “use water pot, then search cup” or will
restrict the correct executions such as “deliver cup, then use water pot” in dynamic environments, which will cause
the removing/adding execution steps. To solve this problem, the logic correlations, including temporal logic, spatial
logic and ontology logic among the controlling symbols, were explored to enhance the adaptability of the NL-based
control method. Typical robotic systems using logic to facilitate robot control include modifying robot movement
trajectory in different situations [127][131], designing robot manipulation post according to fuzzy action type/speed
requirements [88][124], serving meals by considering “foodType–vesselShape” relations [130][132], and assembling
industrial parts by considering the spatial matching relations [129][133] etc.
Control using environmental conditions. Besides the logic relations among the control commands, it is necessary
to consider environmental conditions for a practical NL-based control in real-world situations. The balance among
human commands, robot knowledge and environment conditions should be made for an intuitive NLC. With the NLC
systems, typical applications that consider practical environment conditions include grasping with considering
constraints in safety, temporal relations and human preferences [58][134], navigating by considering location/building
matching [20][135], serving food with consideration of user locations [136], food vessel shapes and path conditions
[119] etc. In these systems, if NL commands and robot knowledge were supported by real world conditions, NLC
tasks are highly likely to be successful, or NLC tasks are likely to fail.
3.2 Open problems
For the NL-based robot control, a human interacted with a robot in a verbal manner. Simple NL control commands
were given separately, or in a hybrid manner where the NL commands were combined with visual/haptic cues. A
human was the only information source, guiding the whole control process. A robot was designed to simply map the
human NL commands to the knowledge structure in robot databases, or to the real-world conditions perceived by
robots’ sensors. With physical/mental work assignments for robots and humans, current efforts in NL-based control
focus on improving control accuracy, decreasing humans’ cognition burdens and increasing a robot’s cognition
burdens. However, the cognition burden of humans in NL-based control was still at a high level and robot cognition
was still at a low level. A human user was required to lead the cooperation and the robot was required to follow the
human with an understanding of control commands. The big cognition-level difference between a human and a robot
restrained the intuitiveness and naturalness of the current NLC systems in the cooperation. Summary of NL-based
control systems is shown in Table 2.
Knowledgeproviding manner
Knowledge format
Control using symbolic
words
predefined
Table 2. Summary of NL-based control systems
Control using semantic Control using logic
correlations
structures
predefined
predefined
Control using environment
conditions
predefined + sensing
linguistic structure
control logic formulas
Typical
applications
symbolic words “yes,
no, go, stop”
object manipulation,
motion trajectory
control, navigation
control
grasping, navigation
sequential action control,
trajectory modification,
hand-pose selection, object
recommendation, assembly
Advantages
concise, accurate
flexible
flexible, adaptable
Disadvantages
inflexible, unnatural,
limited adaptability
unnatural, limited
adaptability on dynamic
situations
implicit, unsafe/risky by
ignoring real-world
conditions
verbal commands + real-world
conditions.
safety concerns in grasping,
human-preferred temporal
action sequence, precise
navigation, user-centered meal
service
flexible, adaptable, consider
real-world conditions
not natural and intelligent for
lacking a meaningful
interpretation
Typical references
[120][121][122]
[125][126][127]
[129][130][131]
[134][135][136]
4. NL-based robot training
To support a robot in complex task planning and intuitive decision making, NL was used to train robots for task
executions with a spoken/written manner. During the training, knowledge was transferred from robot/human experts
to targeted robots, shown in Fig. 7. With consideration of robot physical capabilities such as force/strength/physical
structure/speed, human preferences such as motion/emotion, and real-world conditions such as object
availability/distribution/location, executable knowledge was specified for a robot. With the executable knowledge,
robots’ capabilities in task understanding, environment interpretation and human-request reasoning are improved.
Different from NL-based control, where a robot was not involving in advanced reasoning, in NL-based robot training,
robots were required to reason about human cooperation requests from NL instructions. According to the knowledge
transferring manners, systems using NL-based robot training are mainly categorized into four types: training using
Fig. 7. Typical robotic systems using NL-based robot training. (a) Training using human instructions. By describing the physical properties of
the objects, the robot-related knowledge was transferred from a human to a robot, enabling the robot to recognize objects in the future [143].
(b). Training using human demonstrations. With a human’s physical demonstrations of actions, a robot learned to perform an action by imitating
the motion patterns such as trajectory, action sequence, and motion speed. [150]. (c). Training using proactive robot querying. A robot
proactively detected its missing knowledge and proactively ask human for knowledge support [40]. (d). Training using human feedbacks. During
the execution process, a human proactively interfered with the execution and gives the timely feedbacks to improve a robot’s performances
[153].
human instruction, training using human demonstration, training using human feedback and training using proactive
robot querying. In both the training using human instruction and human demonstration, robots passively learn from a
human. The difference is that in instruction humans only orally describe and do not physically participate, while in
demonstration, humans need to physically participate. In the training both using human feedback and proactive robot
querying, robots proactively learn from a human. The difference is that in the training using human feedback a human
decides what knowledge to learn, while in the training using robot querying a robot decides what knowledge to learn.
4.1. Typical robotic systems
Training using human instruction. Given that human NL instructions are informative and natural, robot training
initially started with using NL instructions to define task execution methods [137]. In the early stage of instruction
training, these instructions were used to perform low-level knowledge grounding, in which short NL expressions given
by humans were mapped into separated knowledge entities in the real-world. With the NLC systems, typical
applications include using the NL instructions to identify an object [132][138], object physical properties [139][140],
and action associations [141][142]. Low-level knowledge grounding endowed robots with a shallow understanding of
the motivations and logics in the instructed task-execution procedures due to which the knowledge was mapped from
a database to the real world in a point-to-point manner. Instead of correlating the execution procedures to form a
semantic network for comprehensive execution understanding, the knowledge correlations in the low-level knowledge
grounding were mutually independent. As information/automation techniques improved, the low-level knowledge
grounding method was then evolved into the high-level knowledge grounding method, in which complex NL
expressions were grounded into hierarchical knowledge structures for motivation/logic understanding. With the
hierarchical knowledge, the instruction-based training for sophisticated task execution was enabled. With the NLC
systems, typical applications include using the NL instructions to describe daily common sense such as precise object
manipulation [143][144], to describe spatial/temporal correlations for intuitive navigations [145][146], and to model
human cognition for daily activity performing [147][148] . During the high-level knowledge grounding, human
cognition processes on task planning and performing were modeled. The advantage of the training using human NL
instructions is that robots’ reasoning mechanisms during NLC is initially developed; the disadvantage is that the
learned execution methods are still abstract in that the sensor-value-level specifications for the NL commands are
lacking, limiting the knowledge implementations in real-world situations.
Training using human demonstration. To improve robot training in practical perceiving and acting, a human
demonstration method was developed to train a robot in real task-execution situations. With training in a humandemonstration manner, theoretical knowledge such as actions, action sequences and object weight/shape/color was
associated with sensor data and data patterns. This theory-practice association enabled a straightforward, sensor-databased interpretation towards the abstract task-related knowledge, improving robots’ understanding and cooperation
by practically implementing the learned knowledge. A general demonstration process was that a human physically
performs a task and meanwhile orally explained the execution intuitions for a robot. The robot was expected to
associate the NL-extracted knowledge with sensor data to specify the task executions. With the NLC systems, typical
applications using the demonstration-based robot training method include: learning object-manipulation methods by
associating human NL expressions with sensor data such as touching force values, object color/shape/size and visual
trajectory [144][149]; learning human-like gestures by associating human hand gesture with speech context
[150][151]; learning object functional usages by simultaneously considering human voice behaviors, motion behaviors
and environment conditions [147]; learning abstract interpretations of environmental conditions by combining human
operations, human NL explanations, and the corresponding sensor data patterns [21][151]; adapting new situations by
replacing NL-instructed knowledge with real-world-available knowledge [36][152] etc. Human demonstration
enabled a robot with a practical understanding of real-world task executions. Compared with robot training using
instructions, robot training using demonstrations specified the abstract theoretical knowledge with the real-world,
making the learned knowledge executable in real-world situations. However, robots’ reasoning capability was not
largely improved since demonstration-based training was actually a sensor-data-level imitation of human behaviors
and ignored the “unobservable human behaviors” such as human’s subjective interpretation towards real-world
conditions, human’s philosophy in execution and human’s cognitive process in decision making.
Training using human feedback. To improve robots’ reasoning capability in NLC, human NL feedback was used
to directly tell the robots “the unobservable human behaviors”. With human NL feedback, robot behaviors in
cooperation were logically modified by adding/removing some operation steps [40][96][153] or subjectively
emphasizing on executions [154][155][156]. With the NLC systems, typical applications include: indicating the
human-desired locations/objects with NL instructions during HRC [153]; assessing robot execution performances and
correcting robots’ undesired manipulation behaviors such as hand poses and object selections by using real-time NL
instructions [96]; analyzing robots’ execution failures and helping a robot to learn from failures with NL conversations
[40]; and emphasizing/ignoring robot behaviors in complex task execution by subjective NL rewards and punishments
such as “joy, anger” [155]. Compared with training using human demonstrations, training using human feedback
proactively and explicitly indicated a robot with operation logics and decision-making mechanisms, enabling better
robot understanding towards human cognitive process in cooperation. Based on both human cognition understanding
and environment perceiving, robots’ surrounding environments in NLC were interpreted as a human-centered
situation. In this human-centered situation, task cooperation was interpreted in a human perspective, improving robots’
reasoning capability in cooperating with a human. However, the feedback-based learning required frequent human
involvements, imposing a heavy cognition burden on a human. Moreover, the knowledge learned by human feedback
was given by a human without considering the robot’s actual knowledge needs, limiting robots’ adaptation to new
environments where robots’ knowledge shortage was waiting to be compensated for successful NLC.
Training using proactive querying. To solve the new situation adaptation problem for further improving a robot’s
reasoning ability, a proactive-querying method was developed for robot training. In the querying process, a robot used
NL to proactively query humans about its missing knowledge about human-intention disambiguation, environment
interpretations, and knowledge-to-world mapping. After the training, a robot was endowed with a more targeted
knowledge to adapt the previously-encountered situations, thereby improving a robot’s environment adaptability. With
the NLC systems, typical applications using querying-based training include: asking for cognitive decisions on
trajectory/action/pose selections in tasks such as “human-robot jointly carrying a bulky bumper” [157]; asking for
knowledge disambiguation of human commands such as confirming the human-attended object “the blue cup” [158];
asking for human physical assistances to deliver a missing objects or to execute robot-incapable actions such as
“deliver a table leg for a robot” [40]; asking for additional information such as “the object is yellow and rectangle”
from a human to assisting robots’ perceiving [153][159] etc. Compared with the training using human instructions,
demonstrations and feedback, in which a robot was trained passively, in the querying method a robot was trained
proactively and knowledge was more situation-specific. Robots were endowed with an advanced self-improving
capability. Supported by a never-ending mechanism, robot performances in NLC were improved in the long term by
continuous knowledge acquiring and refining [91].
4.2. Open problems
During the development of training methods starting from instruction training to querying training, human
cognition burden was gradually decreased and robot cognition level was gradually improved. Robot trainings using
the above-mentioned methods are suffering the shortcomings of robot knowledge scalability and adaptability. The
knowledge scalability problem is caused by limited knowledge from limited information sources such as available
humans/robots experts. The robot knowledge is hard to be largely scaled up with an economic manner. Knowledge
adaptability problems are caused by the current shallow cognition models. It is hard for a robot to adapt to a specific
user/situation and meanwhile adapt to the general types of users/situations. The potential solutions for the knowledge
scalability problem and adaptability problem will be discussed in section 8.3 and 8.4. A summary of NL-based robot
training systems is shown in Table 3.
Table 3. Summary of the NL-based robot training systems
Knowledge
transferring manner
Knowledge format
Human involvement
Robot involvements
Typical applications
Training using human
instruction
human oral descriptions
speech
instructor
follower
object recognition methods,
object manipulation
methods, task execution
procedures
Advantages
completed property/process
definition
Disadvantages
ignoring practical
environment conditions,
limited user/environment
adaptability
Typical references
[139][140][141]
Training using human
demonstration
human physical
demonstration
speech, motion
demonstrator
follower
NL-features
(force/color/size/visual)
association, human-like
gesture learning,
motion-based object
usage learning,
environment condition
interpretation, abstract
human plan execution
make abstract and
ambiguous NL
instructions explicit and
machine-executable.
weak reasoning ability
Training using human
feedbacks
human verb/physical
interference
speech, motion
leader
assistant
human-preferred
execution learning,
human-like
manipulation learning,
robot learns from
failures
Training using proactive
query
robot verbal querying
human preference
consideration, initial
human-like cognitive
process modeling
frequently bother
human, heavy human
cognitive burden
robot get knowledge in
need, relatively strong
environment adaptability
[36][151][152]
[154][155][156]
[157][158][159]
speech
assistant
leader
human-in-the-loop decision
making, human physical
assistance, human intention
disambiguation, humanrobot knowledge
transferring
frequently bother human,
heavy human cognitive
burden
5. NL-based task execution
Different from NL-based robot training systems, in which human NL was helping a robot with its task
understanding, in NL-based robot task execution systems, human NL was helping a robot with its task execution. In
NL-based training, a robot created a structure-completed and execution-specified knowledge representation. But in
NL-based task execution, including understanding the task, the robot was also required to interpret the surrounding
environments, predict human intentions, and make optimal decisions satisfying the constraints from the environment,
the task execution method and the human requirements. Typical robotic systems using NL-based task execution are
shown in Fig. 8. Given that the reasoning was strictly requested in NL-based task execution, robots’ cognition levels
in NL-based task execution were higher than that in NL-based robot training. With respect to whom is leading the
execution, systems of NL-based execution are categorized into human-centered task execution, in which a human is
mentally leading the task executions and a robot is providing appropriate assistances for facilitating human executions,
and robot-centered task execution, where a robot is mentally leading the task executions and the human is providing
oral reminders or physical assistances for facilitating robot executions.
Fig. 8. Typical robotic systems using NL-based task execution. (a) is human-centered execution. A human was performing tasks such as “assemble
a toy”. A robot was standing by and meanwhile prepares to provide help, ensuring the success and smoothness of the human’s task executions. A
robot was expected to infer the human’s ongoing activities, detect human needs timely and proactively provide the appropriate help such as “a toy
part” [160]. (b). Robot-centered execution. A robot was autonomously performing a task. A human was standing by to monitor the robot
executions. If abnormal executions or execution failures occurred, the human provided the timely verbal corrections such as “stop, grasp the top”
or physical assistances such as “delivering the robot-needed object” [161].
5.1. Typical robotic systems
Human-centered task execution. Given the technology limitations in perceiving, reasoning and acting, a robot is
still not fully automatized and human intelligence is still necessary in HRC. To integrate human mental intelligence
and robot physical execution, NL-based execution systems were designed to be human-centered. NL expressions in
task execution deliver information such as explanation of human’s cooperation requests, descriptions of human’s
execution plans and indications of human’s urgent needs. With this information, a robot provides appropriate
assistances timely. With NLC systems, typical applications using human-centered executions include performing tasks
such as “table assembly”, during which the human sets up task goal (assembly a specific part), makes plans (action
steps, pose and tool usages) and partially executes tasks (assemble the parts together), and the robot provides humandesired assistances (tool delivery, part delivery, part holding) [162][163]. During the cooperation, human took both
cognitive and physical responsibilities, and the robot took partial physical responsibilities. Comprehensive humancentered execution was developed so that a human was only burdened with cognitive responsibilities such as
explaining the navigation routine [164][165], describing the needed objects and location/pose [66][166] and guiding
the fine/rough processing [5][167]. Correspondingly, a robot took on only physical responsibilities such as
grasping/transferring the fragile/heavy objects [168][169]. Both the human and the robot performed independent subtasks by considering the same high-level task goal. The robot received less instructions for its tasks and meanwhile
was expected to monitor human’s task processes so that the robot provided appropriate assistances when the human
required it. This cooperation proposed a relatively-high standard towards robot cognition on providing appropriate
assistances at the right location/time. Overall, in human-centered NL-based task execution, the human was leading
the execution at the cognition level, and a robot provided the appropriate assistance for saving the human’s time and
energy, thereby enhancing the human’s physical capability.
Robot-centered task execution. To further improve the cooperation intuitiveness and decrease human’s
mental/physical burdens, systems using robot-centered task execution were developed. Different from humancentered systems in which a human mainly took the cognitive/physical burdens and a robot gave human-needed
assistances to facilitate human execution, in robot-centered systems, a robot mainly took the cognitive/physical
burdens, and a human gives robot-needed assistance physically to facilitate the robot execution. NL expressions in the
robot-centered systems were used for a robot to ask for assistances from a human. With the NLC systems, typical
applications using robot-centered task execution include: robot lead industrial assembly, in which a human enhanced
robot physical capability by providing robot with physical assistances such as grasping [170] and fetching [171]; robot
executed tasks such as heavy object moving and elderly navigation in unstructured outdoor environments, in which a
human analyzed and conquered the environment limitations such as objects/space availability [22][172][173][174].
Compared with robots in human-centered execution, where a robot was required to comprehensively understand
human behaviors, robots in robot-centered applications were required to comprehensively understand the limitations
on robot knowledge, real-world conditions, both humans’ and robots’ physical capabilities etc. The advantage was
that the human was less involved and his/her hands/mind were partially set free.
5.2. Open problems
An open problem for human-centered task execution is that the human’s cognition burden is relatively heavy.
Even though a robot become more autonomous to prove timely assistances, a human is still leading the NLC
mentally/physically. An open problem for robot-centered task execution is that robot-centered applications set a high
standard towards the robot capability on situation analyzing, plan making, assistance asking, and knowledge updating.
While these robot capabilities are still immature [175]. Therefore, a NLC-based system involving human cognitive
process modeling, intelligent robot decision-making, autonomous robotic task-execution, and human/robot physical
capability consideration is in urgent need. Summary of NL-based task execution systems is shown in Table 4.
Activity recognition
requirement
Human involvements
Human’s job
Robot’s job
Human involvement
Robot involvements
Typical applications
Table 4. Summary of the NL-based task execution systems
Human-centered task execution
Robot-centered task execution
robot, high
human, high
high
cognitive burden
physical burden
leader
assistant
assembly, object grasping
Advantages
Disadvantages
accurately assistance
heavy cognitive burdens on human, high requirements
on robot’ reasoning capability on recognizing human
activity and detecting human needs
Typical references
[167][168][169]
low
physical burden
cognitive burden
assistant
leader
daily assistance, industrial assembly, outdoor
navigation, heavy object delivering
strong environment adaptation
human cognitive burden, human disturbance
[172][173][174]
6. NL-based social cooperation
To make NLC natural and intuitive, NL-based social cooperation was developed by involving human’s social
behaviors such as social speeches and social motions in NLC. Different from NL-based task execution systems, which
focused on objective task information such as task goals and task execution procedures, NL-based social cooperation
systems focused on subjective social norms such as social speech or motion behaviors, shown in Fig. 9. With social
norms, robots were naturally integrated with a human in NLC and, meanwhile, robots’ social acceptances were
increased. According to social norm types, robotic systems using NL-based social cooperation are categorized into
social communication in which social NL expressions are used for facilitating communication, and social execution
in which social NL expressions are used for facilitating execution.
6.1. Typical robotic systems
NL-based social communication. In a natural and intuitive NLC, robot-human communication needs to be both
information-correct and social-norm appropriate. Capturing social norms from humans’ NL expressions was helpful
in certain aspects such as detecting human preferences in cooperation, specifying cooperation roles such as “leader,
follower, cooperator”, and increasing social acceptance. NL in social communications served as an information source,
from which both the objective execution methods and the subjective human preferences were extracted. With the NLC
systems, typical applications using NL-based social communication include: a receptionist robot increased its social
acceptance in conference arrangements by using social dialogs with pleasant prosodic contours [180]; cooperative
machine operations used social descriptions considering human action preferences [181][182]; health-caregiving
robots searched and delivered objects by considering user speech confidences, user safety and user roles such as
“primary user, bystander” [183]; adapted unfamiliar users by using NL expressions with fuzzy emotion statuses such
as “fuzzy happiness/sadness/anger” [184]; modeled social NL communications in NLC by defining human-robot
relations such as “love”, “friendship” and “marriage” [185]; designed robotic companion by using friendly NL
conversation [67][186][187] etc.
NL-based social execution. In a natural and intuitive NLC, the communication and the execution need to be
socially appropriate. To introduce social execution norms into NLC, NL-based social execution systems were
designed. NL was used to indicate socially-preferred executions for robots, enhancing robots’ understanding of social
motivations behind task executions and further making robot executions socially-acceptable. With the NLC systems,
Fig. 9. Typical robotic systems using NL-based social cooperation. (a) and (b) are NL-based social communication. A robot learned to nicely
response a human’s request such as “drawing a picture on the paper” [176] and “stop until I touch you” [177]. (c) and (d) are NL-based social
execution. A robot learned to do appropriate body languages during its speaking such as storytelling [178][179].
typical applications using NL-based social executions include: a navigation robot autonomously modified its motion
behaviors (stop, slower, faster) by considering human density (crowded, dense) with the reminding of human NL
instructions (“go ahead to move”, “stop”) [188]; a companion robot moved its head towards the human speaker
according to human’s NL tunes [189]; a storytelling robot depictd stories by mapping NL expressions with human’s
body motion behaviors to catch humans’ attention [178][190].
NL-based social communication and NL-based social execution focused on two different aspects of NLC. To
develop socially intuitive NLC systems, the two aspects need to be focused on simultaneously.
6.2. Open problems
Social norms in both communications and executions are hard to model. First, social norms are implicit. It is
challenging to summarize social norms from human behaviors. Second, social norms vary. Different regions,
countries, cultures, races and personalities form different social norms. A social-norm model that considers the above
influential factors is challenging to create because the representative norms are difficult to extract. Last, social norms
are currently non-evaluable. It is challenging to assess the correctness of social norms because there are no clear
standards to judge the correctness of social norms, and different persons have different levels of social behavior
acceptance/tolerance degree. Summary of NL-based social cooperation systems is shown in Table 5.
Table 5. Summary of NL-based social cooperation systems
NL-based social communication
NL-based social execution
physical
Knowledge
transferring manner
Knowledge format
Human involvement
Robot involvements
Typical applications
verbal
Advantages
more social acceptance, individualized
Disadvantages
Typical references
hard to model user-specific social expressions
physical execution
both cognitive & physical burdens
both cognitive & physical burdens
social distance maintenance, body language learning
in storytelling, social behavior-supported object
carriage, object manipulation using social
expressions
understand/respect human’s social preferences,
understand/respect human expectations
hard to model user-specific social execution manners.
[181][182][183]
[188][189][190]
NL expressions
both cognitive & physical burdens
both cognitive & physical burdens
restaurant receptionist, health caregiving, industrial
cooperation, human cooperation
Fig. 10. Main steps for realizing NLC. The input for a NLC method is the human NL request, the output is the HRC [191]. In the NLC
implementation process, there are mainly three steps, including NL understanding in which the task-related knowledge is extracted from the
human NL requests, knowledge representation in which the extracted knowledge is organized in an algorithm structure, and knowledge-world
mapping where theoretical knowledge is grounded into real-world situations.
7. Methods for realizing NLC
Developing NLC systems are suffering from three main challenges. First, it is challenging to comprehensively
understand human NL instructions. Understanding NL instructions is not about precise speech recognition, but instead
precise semantic analysis, by which the meaning, logic and human cognitive process embedded in NL are extracted.
Second, it is challenging to represent robots’ understanding to support robots’ decision-making in NLC. The
representation of a robot’s understanding is expected to include the task-related knowledge contents and knowledge
implementation manners. With the representations, a robot is able to measure the applicability of its knowledge in a
situation and decide the correct contents and manners of using its knowledge. Third, it is challenging to accurately
map a robot’s knowledge into real-world situations. In the mapping process, the theoretical items in knowledge
databases are expected to be associated with the corresponding practical objects or relations in the real world, and the
incomplete/real-world-inconsistent knowledge is expected to be autonomously corrected.
To address these three challenges, three types of research were conducted: NL understanding, knowledge
representation, and knowledge-world mapping. With the NL understanding research, human verbal instructions were
processed and the task-related knowledge, such as task goal, execution steps, and the execution parameters “speed,
tools, locations, human requirements” were extracted to perform a comprehensive semantic analysis [5][172][192] .
With the knowledge representation research, task knowledge was constructed by algorithms such as Bayesian Network
[91][193], Support Vector Machine (SVM) [194][195], and Hidden Markov Model (HMM) [118][196]. By using
these algorithm structures, the knowledge was capable of supporting robot decision making with various
logic/semantic/spatial/temporal manners. With the knowledge-world mapping research, the theoretical information
patterns were grounded into real-world situations correctly, and the incomplete/inconsistent knowledge gaps were
filled to ensure the success of NLC. In theoretical knowledge grounding, the grounding methods aimed to explore the
semantic correlations among the theoretical knowledge and the real-world by associating the
temporal/spatial/visual/physical features [197][198] . In the gap filling, the methods aimed to detect the missing
knowledge, which was needed in real-world situations but ignored in theoretical knowledge representations, and the
inconsistent knowledge, which was instructed by a human but was not available in the real world [58][199] [200]. The
complete NLC realization process by using the above-mentioned researches is shown in Fig. 10.
8. Emerging trends
NLC has been developed to improve the effectiveness and naturalness of HRC. Due to the limitations of NLCrelated techniques such as NLP, machine learning and robot design, NLC performances in dealing with complex tasks,
various users, and dynamic/unstructured environments still needs to be improved. Based on our comprehensive
review, the future trends for future NLC research are summarized as follows.
8.1. NLP’s contributions to NLC
Comprehensively, NLP is undergoing a deep-neural-network revolution to create sophisticated computational
sematic models, including the word embedding method for comprehensive meaning modeling by adding in extra
meanings such as “cats and dogs are animals” in data preparation stages [201], sequence-to-sequence language
understanding/translation by sequentially outputting the meaning-modeling results based on the previous semantic
context [202], attention-based NL understanding in which the relatively important words/expressions are valued by
increasing the weights of the important expressions and decreasing the weights of the unimportant expressions [203],
unsupervised long-term meaning modeling [204] etc.
With the advantages brought by the NLP revolution, NLC is correspondingly benefited in areas including:
complex cooperation request understanding by adding task-specific knowledge such as manufacturing common sense;
daily assistance common sense and caregiving common sense into the input NL data; real-time and context-aware task
execution by aligning NL expressions with knowledge related to tasks, robots, environments, and humans; humandesired execution priority analysis by analyzing human’s verbal focuses in communication; self-improving of robot
understanding ability during task executions.
8.2. HRC’s contributions to NLC
The current HRC research has two trends: generalization and specification. In the generalization trend, a robot is
endowed with broad commonsense knowledge to support the general NLC under various situations [33][94][205].
Massive commonsense is essentially the general principle during typical task executions. The objective of
generalization is to make a robot adapt to a wide range of tasks/situations/users. To effectively summarize the
representative features shared by various NLC scenarios, effective feature learning methods and accurate knowledgeworld mapping models are required. Developing these methods could be a future direction of NLC research. In the
specification trend, a robot is endowed with delicate knowledge to support the specific types of HRC [206][207]. The
delicate knowledge provides execution details for practical robot execution. The objective is to endow a robot with
high professionality on specific tasks in specific environmental conditions with specific users. It is critical for a robot
to have a broad and delicate knowledge for intuitive NLC.
Both trends are suffering from challenges. For the generalization trend, NLC emphasizes general situation
adaptation by extracting the common-sharing knowledge in various scenarios and ignoring the unique knowledge in
individual scenarios. Caused by the unique-knowledge ignorance, general NLC methods are relatively simple, being
challenged by cold-start phenomenon, which refers the knowledge not being learned in the training stage and causing
execution failures in the implementation stage [208]. For the specification trend, limited by the knowledge coverage
range, the robot-capable task types are within a narrow range. A robot, which is expected to execute a specific type of
task with high professionality, is incapable of executing a wide range of tasks due to intrinsic mental/physical
conflictions between being specific and being general [209].
NL is informative, containing the general execution knowledge such as a typical execution method, and the
specific knowledge such as a user’s emotions, preferences and personalities. HRC using NL is a useful way to balance
the generation and specification trends in robotic research. During NLC, NL transfers the general knowledge and
emphasizes the specific knowledge in the cooperation process. A future trend of NLC could be using advanced NLP
techniques to realize the mutual compensation between the robot generalization and specification.
8.3. Robot knowledge scalability
Scaling up robot knowledge for supporting robot decision making is a critical issue. On one hand, to understand
human NL instructions, represent cooperation tasks, or fill up the knowledge gaps, the effective knowledge scalingup capability is needed to accurately learn a large amount of knowledge. On the other hand, the time/labor cost are
expected to be reduced. Currently, the knowledge-scaling-up research goes in two directions: existed-knowledge
exploitation, and new-knowledge exploration. In existed-knowledge exploitation, the abstract meanings of existed
knowledge are summarized at a high level to increase the knowledge interchangeability, making one type of
knowledge useful in other similar scenarios. The new-knowledge exploration includes: human-based methods, which
query humans for new knowledge, and the big-data-based method, which is an automatic and low-cost information
retrieval method that extracts knowledge from information sources such as the World Wide Web [205], books [210],
machine operation log files [211], and videos [212]. Given the new techniques in computer science and neuro science,
there is still a need to develop efficient, low-cost and large-scale knowledge scaling-up methods.
8.4. Enhancing robot adaptability
Weak robot adaptability is typically caused by failures in execution importance modeling, based on which the
execution priority is made, and execution interchangeability modeling, based on which the execution flexibility is
made. To increase robot adaptability, new research was launched to model the human cognition process [213][214],
which aims to explore humans’ decision-making mechanism for modeling robot execution priority and flexibility.
For execution priority, not all the executions are essential for the execution success. For example, in the task
“assembly”, the procedure “clean the place” is much less important than the procedure “install the screw”. For
interchangeability, a tool request “deliver me a brush” does not necessarily mean the involvement of a specific tool
“brush”, but instead means a practical purpose “cleaning the surface” [15]. By knowing these meanings, the
cooperation plans are flexibly changed by ignoring the trivial execution procedures and focusing on the important
procedures, and replacing the unavailable tools with the other available similar-function tools. Current methods focus
on exploring object affordances (object-action correlation) [215], and lacking the in-depth interpretations of task
cooperation. In the future, NLC research could be methods that interpret robot executions from a human perspective,
improving robot adaptability in unstructured environments and unfamiliar human users.
8.5. Learning from failures
Execution failure causes unnatural task execution or even task failures. The learning-from-failure mechanism has
been implemented in computer science for algorithm efficiency improvement [216] and in material science for new
material discovery [217]. By exploring information in failure experiences, robots’ performance in task execution is
improved by avoiding similar failures in the future. In NLC, learning from failure is involved in a definition-based
manner [40], in which the failure is analyzed by comparing the available knowledge with the defined knowledge,
lacking the analysis of failure causes and recovery mechanism. Therefore, in NLC, learning from failure is also a
promising research direction. From our perspective, the potential research problems could be in-depth failure cause
analysis, concise NL failure explaining to a human, proactive knowledge updating methods for recovering from the
failures, etc.
9. Conclusion
This review summarized the state-of-the-art robotic systems for using natural language (NL) to facilitate humanrobot cooperation (HRC), thereby providing a summary and comparisons of the natural-language-facilitated humanrobot cooperation (NLC) systems. Regarding the robot-cognition levels, NLC systems were categorized into four
types: NL-based control, NL-based robot training, NL-based task execution, and NL-based social companion. Based
on our perspective and comprehensive paper review, the current emerging trends of NLC research were discussed,
providing helpful information for the future of NLC research.
10. Acknowledgements
We would like to thank Mr. Xu Zhou, Ms. Natalie Kalin and Mr. Ian Coberly for helping us with work finalization.
References
[1] J. Baraglia, M. Cakmak, Y. Nagai, R. Rao, and M. Asada, “Initiative in robot assistance during collaborative task execution,” in 2016 11th ACM/IEEE
International Conference on Human-Robot Interaction (HRI), pp. 67-74, 2016.
[2] G. Gemignani, E. Bastianelli, and D. Nardi, “Teaching robots parametrized executable plans through spoken interaction,” in International Conference on
Autonomous Agents and Multiagent Systems, pp. 851-859, 2015.
[3] J. Brooks, C. Lignos, C. Finucane, M. S. Medvedev, I. Perera, V. Raman, H. KressGazit, M. Marcus, and H. A. Yanco, “Make it so: continuous, flexible natural
language interaction with an autonomous robot,” in Workshops at the 26th AAAI Conference on Artificial Intelligence, 2012.
[4] T. Fong, C. Thorpe, and C. Baur, “Collaboration, dialogue, and human-robot interaction,” in Proceedings of International Symposium of Robotics Research, 2001.
[5] R. Liu, J. Webb, and X. Zhang, “Natural-language-instructed industrial task execution,” in ASME International Design Engineering Technical Conferences and
Computers and Information in Engineering Conference, pp. V01BT02A043-V01BT02A04, 2016.
[6] S. A. Tellex, T. F. Kollar, S. R. Dickerson, M. R. Walter, A. Banerjee, S. Teller, and N. Roy, “Understanding natural language commands for robotic navigation
and mobile manipulation,” AAAI Conference on Artificial Intelligence, 2011.
[7] H. Iwata and S. Sugano, “Human-robot-contact-state identification based on tactile recognition,” IEEE Transactions on Industrial Electronics, vol. 52, no. 6, pp.
1468-1477, 2005.
[8] J. Kruger and D. Surdilovic, “Hand force adjustment: robust control of force-coupled human–robot-interaction in assembly processes,” CIRP Annals Manufacturing Technology, vol. 57, no. 1, pp. 41-44, 2008.
[9] E. Colgate and N. Hogan, “An analysis of contact instability in terms of passive physical equivalents,” in IEEE International Conference on Robotics and
Automation, pp. 404-409, 1989.
[10] H. Kjellström, J. Romero, and D. Kragić, “Visual object-action recognition: Inferring object affordances from human demonstration,” Computer Vision and Image
Understanding, vol. 115, no. 1, pp. 81-90, 2011.
[11] S. Kim, J. Jung, S. Kavuri, and M. Lee, “Intention estimation and recommendation system based on attention sharing,” in International Conference on Neural
Information Processing (ICONIP), pp. 395-402, 2013.
[12] N. Hu, G. Englebienne, Z. Lou, and B. Krose, “Latent hierarchical model for activity recognition,” arxiv: 1503.01820v1, 2015.
[13] P. Barattini, C. Morand, and N. Robertson, “A proposed gesture set for the control of industrial collaborative robots,” in IEEE International Symposium on Robot
and Human Interactive Communication (RO-MAN), pp. 132-137, 2012
[14] A. Jain, S. Sharma, T. Joachims, and A. Saxena, “Learning preferences for manipulation tasks from online coactive feedback,” International Journal of Robotics
Research, vol. 34, no. 10, pp. 1296-1313, 2015.
[15] R. Liu and X. Zhang, "Understanding human behaviors with an object functional role perspective for robotics," IEEE Transactions on Cognitive and
Developmental Systems, vol. 8, no. 2, pp. 115-127, 2016.
[16] K. Ramirez-Amaro, M. Beetz, and G. Cheng, “Transferring skills to humanoid robots by extracting semantic representations from observations of human
activities,” Artificial Intelligence, DOI: http://dx.doi.org/10.1016/j.artint.2015.08.009, 2015.
[17] K. Zampogiannis, Y. Yang, C. Fermuller, and Y. Aloimonos, “Learning the spatial semantics of manipulation actions through preposition grounding,” in IEEE
International Conference on Robotics and Automation (ICRA), pp. 1389-1396, 2015.
[18] W. Takano and Y. Nakamura, “Action database for categorizing and inferring human poses from video sequence,” Robotics and Autonomous Systems, vol. 70,
pp. 116-125, 2015.
[19] K. Andrej and F. Li, “Deep visual-semantic alignments for generating image descriptions,” in Computer Vision and Pattern Recognition (CVPR), pp. 3128-3137,
2015.
[20] V. Raman, C. Lignos, C. Finucane, K. C. T. Lee, M. Marcus, and H. Kress-Gazit, “Sorry Dave, I’m afraid I can’t do that: explaining unachievable robot tasks
using natural language,” in Robotics: Science and Systems, vol. 2, no. 1, pp. 2, 2013.
[21] M. R. Walter, S. M. Hemachandra, B. S. Homberg, S. Tellex, and S. Teller, “Learning semantic maps from natural language descriptions,” in Robotics: Science
and Systems, 2013.
[22] F. Duvallet, M. R. Walter, T. Howard, S. Hemachandra, J. Oh, S. Teller, N. Roy, and A. Stentz, “Inferring maps and behaviors from natural language instructions,”
in International Symposium on Experimental Robotics, pp. 373-388, 2014.
[23] C. Matuszek, E. Herbst, L. Zettlemoyer, and D. Fox, “Learning to parse natural language commands to a robot control system,” in 13th International Symposium
on Experimental Robotics, pp. 403-415, 2013.
[24] C. Ott, D. Lee, and Y. Nakamura, “Motion capture based human motion recognition and imitation by direct marker control,” in IEEE-RAS International
Conference on Humanoid Robots, pp. 399-405, 2008.
[25] S. Waldherr, R. Romero, and S. Thrun, “A gesture based interface for human-robot interaction,” Autonomous Robots, vol. 9, no. 2, pp. 151-173, 2000.
[26] R. Dillmann, “Teaching and learning of robot tasks via observation of human performance,” Robotics and Autonomous Systems, vol. 47, no. 2-3, pp. 109-116,
2004.
[27] K. Sakita, K. Ogawara, S. Murakami, K. Kawamura, and K. Ikeuchi, “Flexible cooperation between human and robot by interpreting human intention from gaze
information,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 846-851, 2004.
[28] Y. Yang, Y. Li, C. Fermuller, and Y. Aloimonos, “Robot learning manipulation action plans by ‘watching’ unconstrained videos from the world wide web,” in
Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, pp. 3686-3692, 2015.
[29] J. R. Medina, M. Shelley, D. Lee, W. Takano, and S. Hirche, “Towards interactive physical robotic assistance: parameterizing motion primitives through natural
language,” in IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN), pp. 1097-1102, 2012.
[30] S. Hemachandra and M. R. Walter, “Information-theoretic dialog to improve spatial-semantic representations” in IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), pp. 5115-5121, 2015.
[31] S. Hunston and G. Francis, “Pattern grammar: A corpus-driven approach to the lexical grammar of English,” Computational Linguistics, vol. 27, no. 2, pp. 318320, 2000.
[32] J. L. Bybee and P. J. Hopper, eds., Frequency and the emergence of linguistic structure. Vol. 45. John Benjamins Publishing, 2001.
[33] Y. Cheng, Y. Jia, R. Fang, L. She, N. Xi, and J. Chai, “Modelling and analysis of natural language controlled robotic systems,” IFAC Proceedings Volumes, vol.
47, no. 3, pp. 11767-11772, 2014.
[34] M. Tenorth, D. Nyga, and M. Beetz, “Understanding and executing instructions for everyday manipulation tasks from the world wide web,” in IEEE International
Conference on Robotics and Automation (ICRA), pp. 1486-1491, 2010.
[35] C. Wu, I. Lenz, and A. Saxena, “Hierarchical semantic labeling for task-relevant rgb-d perception,” in Robotics: Science and Systems, 2014.
[36] S. Hemachandra, F. Duvallet, T. M. Howard, N. Roy, A. Stentz, and M. R. Walter, “Learning models for following natural language directions in unknown
environments,” in IEEE International Conference on Robotics and Automation (ICRA), pp. 5608-5615, 2015.
[37] M. Tenorth, A. Perzylo, R. Lafrenz, and M. Beetz, “The RoboEarth language: representing and exchanging knowledge about actions, objects, and environments,”
in IEEE International Conference on Robotics and Automation (ICRA), pp. 1284-1289, 2012.
[38] J. Pineau, R. West, A. Atrash, J. Villemure, and F. Routhier, “On the feasibility of using a standardized test for evaluating a speech-controlled smart wheelchair,”
International Journal of Intelligent Control and System, vol. 16, no. 2, pp. 124-131, 2011.
[39] C. Granata, M. Chetouani, A. Tapus and P. Bidaud, “Voice and graphical -based interfaces for interaction with a robot dedicated to elderly and people with
cognitive disorders,” pp. 785-790, 2010.
[40] R. A. Knepper, S. Tellex, A. Li, N. Roy, and D. Rus, “Recovering from failure by asking for help,” Autonomous Robots, vol. 39, no. 3, pp. 347-362, 2015.
[41] M. Stenmark and J. Malec, “A helping hand: industrial robotics, knowledge and user-oriented services,” in IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), 2013
[42] R. Schulz, B. Talbot, O. Lam, F. Dayoub, P. Corke, B. Upcroft, and G. Wyeth, “Robot navigation using human cues: a robot navigation system for symbolic
goal-directed exploration,” in IEEE International Conference on Robotics and Automation (ICRA), pp. 1100-1105, 2015.
[43] A. Boularias, F. Duvallet, J. Oh, and A. Stentz, “Grounding spatial relations for outdoor robot navigation,” in IEEE International Conference on Robotics and
Automation (ICRA), pp. 1976-1982, 2015.
[44] J. Kory and C. Breazeal, “Storytelling with robots: Learning companions for preschool children’s language development,” in IEEE International Symposium on
Robot and Human Interactive Communication, pp. 643-648, 2014.
[45] M. J. Salvador, S. Silver, and M. H. Mahoor, “An emotion recognition comparative study of autistic and typically-developing children using the zeno robot,” in
IEEE International Conference on Robotics and Automation (ICRA), pp. 6128-6133, 2015.
[46] C. Breazeal, “Social interactions in HRI: The robot view,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 34, no. 2, pp. 181-186, 2004.
[47] S. B. Kotsiantis, I. Zaharakis, and P. Pintelas, “Supervised machine learning: A review of classification techniques,” Informatica, vol. 31, no. 3, pp. 249-268,
2007.
[48] P. Berkhin, “A survey of clustering data mining techniques,” Grouping multidimensional data, vol. 43, no. 1, pp. 25-71, 2006.
[49] S. Ding, H. Zhu, W. Jia, and C. Su, “A survey on feature extraction for pattern recognition,” Artificial Intelligence Review, vol. 37, no. 3, pp. 169-180, 2012.
[50] N. Ranjan, K. Mundada, and K. Phaltane, “A Survey on Techniques in NLP,” International Journal of Computer Applications, vol. 134, no. 8, pp.6-9, 2016.
[51] E. Cambria and B. White, “Jumping NLP curves: A review of natural language processing research,” IEEE Computational Intelligence Magazine, vol. 9, no. 2,
pp. 48-57, 2014.
[52] A. F. Smeaton, “Progress in the application of natural language processing to information retrieval tasks,” Computer journal, vol. 35, no. 3, pp. 268-278, 1992.
[53] A. Ortony, G. Clore, and A. Collins, The cognitive structure of emotions, Cambridge, U.K.: Cambridge Univ. Press, 1988.
[54] N. Bush, “The predictive value of transitional probability for word-boundary palatalization in English,” M.S. thesis, Univ. New Mexico, Albuquerque, NM,
1999.
[55] L. Goldin and D. M. Berry, “AbstFinder, a prototype natural language text abstraction finder for use in requirements elicitation,” Automated Software
Engineering, vol. 4, no. 4, pp. 375-412, 1997.
[56] R. K. Taira and S. G. Soderland, “A statistical natural language processor for medical reports,” in Proceedings of the AMIA Symposium, 1999.
[57] B. B. Rieger, “Semantic relevance and aspect dependency in a given subject domain: contents-driven algorithmic processing of fuzzy word meanings to form
dynamic stereotype representations,” in International Conference on Computational Linguistics, pp. 298-301, 1984.
[58] E. Scioni, G. Borghesan, H. Bruyninckx, and M. Bonfe, “Bridging the gap between discrete symbolic planning and optimization-based robot control,” in IEEE
International Conference on Robotics and Automation (ICRA), pp. 5075-5081, 2015.
[59] N. Dantam and M. Stillman, “The motion grammar: Analysis of a linguistic method for robot control,” IEEE Transactions on Robotics, vol. 29, no. 3, pp. 704718, 2013.
[60] L. Denoyer, H. Zaragoza, and P. Gallinari, “HMM-based passage models for document classification and ranking,” in European Colloquium on Information
Retrieval Research, Darmstadt, pp. 126-135, 2001.
[61] T. Hofmann, “Unsupervised learning by probabilistic latent semantic analysis,” Machine Learning, vol. 42, nos.1–2, pp. 177–196, 2001.
[62] J. E. Busch, A. D. Lin, P. J. Graydon, and M. Caudill, “Ontology-based parser for natural language processing,” U.S. Patent No. 7,027,974, 2006.
[63] H. Alani, S. Kim, D. E. Millard, M. J. Weal, W. Hall, P. H. Lewis, and N. R. Shadbolt, “Automatic ontology-based knowledge extraction from web documents,”
IEEE Intelligent Systems, vol. 18, no. 1, pp. 14-21, 2003.
[64] E. Cambria and A. Hussain, Sentic computing: Techniques, tools, and applications. Dordrecht, The Netherlands: Springer-Verlag, 2012.
[65] D. Olsher, “COGVIEW & INTELNET: Nuanced energy-based knowledge representation and integrated cognitive-conceptual framework for realistic culture,
values, and concept-affected systems simulation,” in IEEE Symposium on Computational Intelligence for Human-like Intelligence (CIHLI), pp. 82–91, 2013.
[66] E. Ovchinnikova, M. Wachter, V. Wittenbeck, and T. Asfour, “Multi-purpose natural language understanding linked to sensorimotor experience in humanoid
robots,” in IEEE-RAS 15th International Conference on Humanoid Robots (Humanoids), pp. 365-372, 2015.
[67] T. Belpaeme, P. Baxter, J. D. Greeff, J. Kennedy, R. Read, R. Looije, M. Neerincx, I. Baroni, and M. C. Zelati, “Child-robot interaction: Perspectives and
challenges,” in International Conference on Social Robotics (ICSR), pp. 452-459, 2013.
[68] R. Young, “Story and discourse: A bipartite model of narrative generation in virtual worlds,” Interaction Studies, vol. 8, no. 2, pp. 177–208, 2007.
[69] E. Mueller, “Modeling space and time in narratives about restaurants,” Literary Linguistic Computing, vol. 22, no.1, pp. 67–84, 2007.
[70] F. Bex, H. Prakken, and B. Verheij, “Formalizing argumentative story-based analysis of evidence,” in International Conference on Artificial Intelligence and
Law, pp. 1-10, 2007.
[71] M. A. Finlayson and P. H. Winston, “Narrative is a key cognitive competency,” in Annual Meeting Biologically Inspired Cognitive Architectures (BICA), pp.
110, 2011.
[72] S. Schaal, “The new robotics—towards human-centered machines,” HFSP journal, vol. 1, no. 2, pp. 115-126, 2007.
[73] T. W. Fong, I. Nourbakhsh, and K. Dautenhahn, “A survey of socially interactive robots,” Robotics and Autonomous Systems, vol. 42, no. 3-4, pp. 143-166, 2002.
[74] S. Kiesler and P. Hinds, “Introduction to this special issue on human-robot interaction,” Human-Computer Interaction, vol. 19, nos. 1-2, pp. 1-8, 2004.
[75] H. A. Yanco and J. L. Drury, “Classifying human-robot interaction: an updated taxonomy,” in IEEE International Conference on Systems, Man & Cybernetics
(SMC), pp. 2841-2846, 2004.
[76] A. Bauer, D. Wollherr, and M. Buss, “Human-robot collaboration: A survey,” International Journal of Humanoid Robotics, vol. 5, no. 1, pp. 47-66, 2008.
[77] M. Asada, K. Hosoda, Y. Kuniyoshi, H. Ishiguro, T. Inui, Y. Yoshikawa, M. Ogino, and C. Yoshida, “Cognitive developmental robotics: A survey,” IEEE
Transactions on Autonomous Mental Development, vol. 1, no. 1, pp. 12-34, 2009.
[78] O. C. Schrempf, U. D. Hanebeck, A. J. Schmid, and H. Worn, “A novel approach to proactive human-robot cooperation,” in IEEE International Workshop on
Robot and Human Interactive Communication, pp. 555-560, 2005.
[79] G. Hoffman and C. Breazeal, “Effects of anticipatory action on human-robot teamwork efficiency, fluency, and perception of team,” in Proceedings of the
ACM/IEEE international conference on Human-robot interaction, pp. 1-8, 2007.
[80] P. Tuffield and H. Elias, “The shadow robot mimics human actions,” Industrial Robot: An International Journal, vol. 30, no. 1, pp. 56-60, 2003.
[81] A. Stenzel, E. Chinellato, M. A. T. Bou, A. P. del Pobil, M. Lappe, and R. Liepelt, “When humanoid robots become human-like interaction partners:
corepresentation of robotic actions,” Journal of Experimental Psychology: Human Perception and Performance, vol. 38, no. 5, pp. 1073, 2012.
[82] A. De Santis, B. Siciliano, A. De Luca, and A. Bicchi, “An atlas of physical human–robot interaction,” Mechanism and Machine Theory, vol. 43, no. 3, pp. 253270, 2008.
[83] D. Kulic and E. A. Croft, “Safe planning for human‐robot interaction,” Journal of Robotic Systems, vol. 22, no. 7, pp. 383-396, 2005.
[84] N. Mitsunaga, C. Smith, T. Kanda, H. Ishiguro, and N. Hagita, “Adapting robot behavior for human--robot interaction,” IEEE Transactions on Robotics, vol. 24,
no. 4, pp. 911-916, 2008.
[85] S. Schaal, “Is imitation learning the route to humanoid robots?” Trends in Cognitive Science, vol. 3, no. 6, pp. 233–242, 1999.
[86] M. Mataric, “Getting humanoids to move and imitate,” IEEE Intelligent Systems, vol. 15, no. 4, pp. 18-24, 2000.
[87] C. D. Frith and D. M. Wolpert, “The neuroscience of social interaction: Decoding, imitating, and influencing the actions of others,” Journal of Consciousness
Studies, vol. 119, no. 4, pp. 664-668, 2004.
[88] R. Stiefelhagen, C. Fugen, R. Gieselmann, H. Holzapfel, K. Nickel, and A. Waibel, “Natural human-robot interaction using speech, head pose and gestures,” in
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2004.
[89] S. Jerritta, M. Murugappan, R. Nagarajan, and K. Wan, “Physiological signals based human emotion recognition: a review,” in IEEE International Colloquium
on Signal Processing and its Applications (CSPA), 2011.
[90] E. Todorov, “Stochastic optimal control and estimation methods adapted to the noise characteristics of the sensorimotor system,” Neural Computation, vol. 17,
no. 5, pp. 1084–1108, 2005.
[91] R. Liu, X. Zhang, J. Webb, and S. Li, “Context-specific intention awareness through web query in robotic caregiving,” in IEEE International Conference on
Robotics and Automation (ICRA), pp. 1962-1967, 2015.
[92] R. Liu, X. Zhang, and S. Li, “Use context to understand user's implicit intentions in activities of daily living,” in IEEE International Conference on Mechatronics
and Automation (ICMA), pp. 1214-1219, 2014.
[93] R. Liu and X. Zhang, “Fuzzy-context-specific intention inference for robotic caregiving,” International Journal of Advanced Robotic Systems, vol. 99, no. 99,
pp. 99, 2016.
[94] R. Liu and X. Zhang, “Context-specific grounding of web natural descriptions to human-centered situations,” Knowledge-Based Systems, vol. 111, pp. 1-16,
2016.
[95] A. Burce, I. Nourbakhsh, and R. Simmons, “The role of expressiveness and attention in human-robot interaction,” in IEEE International Conference on Robotics
and Automation (ICRA), vol. 4, pp. 4138-4142, 2002.
[96] M. Staudte and M. W. Crocker, “Investigating joint attention mechanisms through spoken human–robot interaction,” Cognition, vol. 120, no. 2, pp. 268-291,
2011.
[97] M. L. Walters, K. Dautenhahn, R. Te Boekhorst, K. L. Koay, C. Kaouri, S. Woods, C. Nehaniv, D. Lee, and I. Werry, “The influe nce of subjects' personality
traits on personal spatial zones in a human-robot interaction experiment,” in IEEE International Workshop on Robot and Human Interactive Communication, pp.
347-352, 2005.
[98] L. Takayama and C. Pantofaru, “Influences on proxemic behaviors in human-robot interaction,” in IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS), pp. 5495-5502, 2009.
[99] R. C. Arkin, M. Fujita, T. Takagi, and R. Hasegawa, “An ethological and emotional basis for human–robot interaction,” Robotics and Autonomous Systems, vol.
42, no. 3, pp. 191-201, 2003.
[100] S. Coradeschi, H. Ishiguro, M. Asada, S. C. Shapiro, M. Thielscher, C. Breazeal, M. J. Mataric, and H. Ishida, “Human-inspired robots,” IEEE Intelligent
Systems, vol. 21, no. 4, pp. 74–85, 2006.
[101] G. Tonietti, R. Schiavi, and A. Bicchi, “Design and control of a variable stiffness actuator for safe and fast physical human/robot interaction,” in IEEE
International Conference on Robotics and Automation (ICRA), pp. 526-531, 2015.
[102] B.
Selman.
(2014,
August 1).
NRI: Collaborative
Research: Jointly
Learning Language
and
Affordances
.
Available:
https://www.nsf.gov/awardsearch/showAward?AWD_ID=1426744&HistoricalAwards=false
[103] R. Mooney. (2016, August 1). NRI: Robots that Learn to Communicate through Natural Human Dialog . Available:
https://www.nsf.gov/awardsearch/showAward?AWD_ID=1637736&HistoricalAwards=false
[104] N. Roy. (2014, August 18). NRI: Collaborative Research: Modeling and Verification of Language-based Interaction . Available:
https://www.nsf.gov/awardsearch/showAward?AWD_ID=1427547&HistoricalAwards=false
[105] University of Washington. (2017, January 5). Robotics and State Estimation Lab . Available: http://rse-lab.cs.washington.edu/projects/language-grounding/
[106] Lund University. (2017, January 5). Robotics and State Estimation Lab. Available: http://rss.cs.lth.se/
[107] J. Hollerbach, "The international journal of robotics research," SAGE. [Online]. Available: http://journals.sagepub.com/home/ijr. Accessed: Jan. 27, 2017.
[108] F.
Park,
"IEEE
Xplore:
IEEE
transactions
on
robotics,"
in
IEEE
transactions
on
robotics.
[Online].
Available:
http://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=8860. Accessed: Jan. 27, 2017.
[109] R. Dechter, Artificial intelligence. [Online]. Available: https://www.journals.elsevier.com/artificial-intelligence/. Accessed: Jan. 27, 2017.
[110] H. Fujita and J. Lu, Knowledge-based systems. [Online]. Available: https://www.journals.elsevier.com/knowledge-based-systems. Accessed: Jan. 27, 2017.
[111] "ICRA2015," in ICRA2015. [Online]. Available: http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=7128761. Accessed: Jan. 27, 2017.
[112] "IROS2015," in IROS2015. [Online]. Available: https://www.ieee.org/conferences_events/conferences/conferencedetails/index.html?Conf_ID=33365.
Accessed: Jan. 27, 2017.
[113] "AAAI15: Twenty-Ninth conference on artificial intelligence," in AAAI15, 1995. [Online]. Available: http://www.aaai.org/Conferences/AAAI/aaai15.php.
Accessed: Jan. 27, 2017.
[114] B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,” Robotics and autonomous systems, vol. 57, no. 5,
pp. 469-483, 2009.
[115] S. S. Rautaray and A. Agrawal, “Vision based hand gesture recognition for human computer interaction: a survey,” Artificial Intelligence Review, vol. 43, no. 1,
pp. 1-54, 2015.
[116] C. L. Bethel, K. Salomon, R. R. Murphy, and J. L. Burke, “Survey of psychophysiology measurements applied to human-robot interaction,” in IEEE International
Symposium on Robot and Human Interactive Communication, pp. 732-737, 2007.
[117] B. D. Argall and A. G. Billard, “A survey of tactile human–robot interactions,” Robotics and Autonomous Systems, vol. 58, no. 10, pp. 1159-1176, 2010.
[118] G. Antoniol, R. Cattoni, M. Cettolo, and M. Federico, “Robust speech understanding for robot telecontrol,” in International Conference on Advanced robotics,
pp. 205-209,1993.
[119] S. Chen, Z. Kazi, M. Beitler, M. Salganicoff, D. Chester, and R. Foulds, “Gesture-speech based HMI for a rehabilitation robot,” in Proceedings of Bringing
Together Education, Science and Technology, pp.29-36, 1996.
[120] R. Bischoff and V. Graefe, “Integrating vision, touch and natural language in the control of a situation-oriented behavior-based humanoid robot,” in IEEE
International Conference on Systems, Man, and Cybernetics (SMC), vol.2, pp. 999-1004, 1999.
[121] B. Landau and R. Jackendoff, “Whence and whither in spatial language and spatial cognition?” Behavioral and brain sciences, vol. 16, no. 2, pp. 255-265, 1993.
[122] M. Ferre, J. Macias-Guarasa, R. Aracil, and A. Barrientos, “Voice command generation for teleoperated robot systems,” in IEEE International Workshop on
Robot and Human Communication, pp. 679-685, 1998.
[123] J. Savage, E. Hernandez, G. Vazquez, A. Hernandez, and A. L. Ronzhin, “Control of a mobile robot using spoken commands,” in Conference Speech and
Computer, pp. 333-338, 2004.
[124] C. Jayawardena, K. Watanabe, and K. Izumi, “Posture control of robot manipulators with fuzzy voice commands using a fuzzy coach–player system,” Advanced
Robotics, vol. 21, nos. 3-4, pp. 293-328, 2007.
[125] A. Nordmann, S. Wrede, and J. Steil, “Modeling of movement control architectures based on motion primitives using domain-specific language,” in IEEE
International Conference on Robotics and Automation (ICRA), pp. 5032-5039, 2015.
[126] M. Beetz, F. Balint-Benczedi, N. Blodow, D. Nyga, T. Wiedemeyer, and Z. C. Marton, “ROBOSHERLOCK: Unstructured information processing for robot
perception,” in IEEE International Conference on Robotics and Automation (ICRA), pp. 1549-1556, 2015.
[127] J. Allen, Q. Duong, and C. Thompson, “Natural language service for controlling robots and other agents,” in IEEE Integration of Knowledge Intensive MultiAgent Systems (KIMAS-05), pp. 592-595, 2005.
[128] B. House, J. Malkin, and J. Bilmes, “The VoiceBot: A Voice Controlled Robot Arm,” In Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems, pp. 183-192, 2009.
[129] M. Stenmark and P. Nugues, “Natural Language Programming of Industrial Robots,” International Symposium on Robotics, pp. 1-5, 2013.
[130] D. Jain, L. Mosenlechner, and M. Beetz, “Equipping robot control programs with first-order probabilistic reasoning capabilities,” in IEEE International
Conference on Robotics and Automation (ICRA), pp. 3626-3631, 2009.
[131] G. E. Fainekos, H. Kress-Gazit, and G. J. Pappas, “Temporal logic motion planning for mobile robots,” in IEEE International Conference on Robotics and
Automation (ICRA), pp. 2020-2025, 2005.
[132] T. Oates, Z. Eyler-Walker, and P. R. Cohen, “Using syntax to learn semantics: An experiment in language acquisition with a mobile robot,” Walker, pp. 1-6,
1999.
[133] M. Stenmark, J. Malec, K. Nilsson, and A. Robertsson, “On distributed knowledge bases for robotized small-batch assembly,” IEEE Transactions on Automation
Science and Engineering, vol. 12, no. 2, pp. 1-10, 2015.
[134] J. M. G. Romano, E. F. Camacho, J. G. Ortega, and M. T. Bonilla, “A generic natural language interface for task planning application to a mobile robot,” Control
Engineering Practice, vol. 8, no. 10, pp. 1119-1133, 2000.
[135] C. Finucane, G. Jing, and H. Kress-Gazit, “LTLMoP: Experimenting with language, temporal logic and robot control,” in IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS), pp. 1988-1993, 2010.
[136] M. Bollini, S. Tellex, T. Thompson, N. Roy, and D. Rus, “Interpreting and executing recipes with a cooking robot,” in 13th International Symposium on
Experimental Robotics, pp. 481-495, 2013.
[137] H. Cuayahuitl, “Robot learning from verbal interaction: A brief survey,” in International Symposium on New Frontiers in Human-Robot Interaction, 2015.
[138] Y. Chen and D. H. Ballard, “On the integration of grounding language and learning objects,” in National Conference on Artificial Intelligence, vol. 4, pp. 2,
2004.
[139] M. N. Nicolescu and M. J. Matari, “Task learning through imitation and human-robot interaction,” in Models and Mechanisms of Imitation and Social Learning
in Robots, Humans and Animals: Behavioural, Social and Communicative Dimension, pp. 407-424, 2005.
[140] D. Roy, “Learning visually grounded words and syntax of natural spoken language,” Evolution of Communication, vol. 4, no. 1, pp. 33-56, 2002.
[141] S. Lauria, G. Bugmann, T. Kyriacou, and J. Bos, “Personal Robot Training via Natural-Language Instructions,” IEEE Intelligent Systems, vol. 16, no.5, pp. 3845, 2001.
[142] M. N. Nicolescu and M. J. Mataric, “Natural methods for robot task learning: instructive demonstrations, generalization and practice,” in Proceedings of the
second international joint conference on Autonomous agents and multiagent systems, pp. 241-248, 2003.
[143] G. M. Kruijff, J. D. Kelleher, G. Berginc, and A. Leonardis, “Structural descriptions in human-assisted robot visual learning,” in ACM SIGCHI/SIGART
conference on Human-robot interaction (HRI), pp. 343-344, 2006.
[144] K. Sugiura and N. Iwahashi, “Learning object-manipulation verbs for human-robot communication,” in Proceedings of the 2007 workshop on Multimodal
interfaces in semantic interaction (WMISI), pp. 32-38, 2007.
[145] P. Kordjamshidi, J. Hois, M. V. Otterlo, and M. Moens, “Learning to interpret spatial natural language in terms of qualitative spatial relations,” Representing
space in cognition: interrelations of behavior, language, and formal models, 2012.
[146] N. Iwahashi, “Robots that learn language: A developmental approach to situated human-robot conversations,” in Human Robot Interaction, Nilanjan Sarkar (Ed.),
InTech, 2007.
[147] K. F. Uyanik, Y. Caliskan, A. K. Bozcuoglu, O. Yuruten, S. Kalkan, and E. Sahin, “Learning social affordances and using them for planning,” in 35th Annual
Meeting of the Cognitive Science Society (CogSci), 2013.
[148] A. Holroyd and C. Rich, “Using the behavior markup language for human-robot interaction,” in Proceedings of the seventh annual ACM/IEEE international
conference on Human-Robot Interaction, pp. 147-148, 2012.
[149] L. Montesano, M. Lopes, A. Bernardino, and J. Santos-Victor, “Modeling affordances using Bayesian networks,” in IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), pp. 4102-4107, 2007.
[150] M. Salem, S. Kopp, I. Wachsmuth, and F. Joublin, “Towards an integrated model of speech and gesture production for multi-modal robot behavior,” in IEEE
International Symposium on Robot and Human Interactive Communication (RO-MAN), pp. 614-619, 2010.
[151] M. Forbes, M. Chung, M. Cakmak, and L. Zettlemoyer, “Grounding antonym adjective pairs through interaction,” HRI, 2014.
[152] E. Krause, M. Zillich, T. Williams, and M. Scheutz, “Learning to recognize novel objects in one shot through human-robot interactions in natural language
dialogues,” in Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, pp. 2796-2802, 2014.
[153] H. Dindo and D. Zambuto, “A probabilistic approach to learning a visually grounded language model through human-robot interaction,” in IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), pp. 790-796, 2010.
[154] T. Williams, G. Briggs, B. Oosterveld, and M. Scheutz, “Going beyond literal command-based instructions: Extending robotic natural language interaction
capabilities,” in AAAI Conference on Artificial Intelligence, pp. 1387-1393, 2015.
[155] A. Bannat, J. Blume, J. T. Geiger, T. Rehrl, and F. Wallhoff, C. Mayer, B. Radig, S. Sosnowski, and K. Kuhnlenz, “A multimodal human-robot-dialog applying
emotional feedbacks,” in Proceedings of Second International Conference on Social Robotics (ICSR), 2010.
[156] A. L. Thomaz and C. Breazeal, “Teachable robots: Understanding human teaching behavior to build more effective robot learners,” Artificial Intelligence, vol.
172, nos. 6-7, pp. 716-737, 2008.
[157] J. R. Medina, M. Lawitzky, A. Mortl, and D. Lee, “An experience-driven robotic assistant acquiring human knowledge to improve haptic cooperation,” in
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2011.
[158] K. Sugiura, N. Iwahashi, H. Kawai, and S. Nakamura, “Situated spoken dialogue with robots using active learning,” Advanced Robotics, vol. 25, no. 17, pp.
2207-2232, 2011.
[159] D. K. Misra, J. Sung, K. Lee, and A. Saxena, “Tell me Dave: Context-sensitive grounding of natural language to manipulation instructions,” in Robotics: Science
and Systems, vol.35, no. 1, pp. 281-300, 2014.
[160] E. Bicho, L. Louro, W. Erlhagen, “Integrating verbal and nonverbal communication in a dynamic neural field architecture for human–robot interaction,” Frontiers
in neurorobotics, vol. 4, no. 5, 2010.
[161] A. Broad, J. Arkin, N. Ratliff, T. Howard, and B. Argall, “Towards Real-Time Natural Language Corrections for Assistive Robots,” RSS Workshop on Model
Learning for Human-Robot Communication, 2016.
[162] P. F. Dominey, A. Mallet, and E. Yoshida, “Progress in programming the hrp-2 humanoid using spoken language,” in IEEE International Conference on Robotics
and Automation (ICRA), pp. 2169-2174, 2007.
[163] S. Profanter, A. Perzylo, N. Somani, M. Rickert, and A. Knoll, “Analysis and semantic modeling of modality preferences in industrial human-robot interaction,”
in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1812-1818, 2015.
[164] D. Lu, F. Wu, and X. Chen, “Understanding User Instructions by Utilizing Open Knowledge for Service Robots,” arXiv preprint arXiv:1606.02877, 2016.
[165] B. Thomas, and O. Jenkins, “RoboFrameNet: Verb-Centric Semantics for Actions in Robot Middleware,” IEEE International Conference on Robotics and
Automation, pp. 4750-4755, 2012.
[166] B. Burger, I. Ferrane, F. Lerasle, and G. Infantes, “Two-handed gesture recognition and fusion with speech to command a robot,” Autonomous Robots, vol. 32,
no. 2, pp. 129-147, 2012.
[167] T. Fong, I. Nourbakhsh, C Kunz, L. Fluckiger and J. Schreiner, “The Peer-to-Peer Human-Robot Interaction Project,” Space 2005, pp. 6750, 2005.
[168] R. Deits, S. Tellex, P. Thaker, D. Simeonov, T. Kolloar, and N. Roy, “Clarifying commands with information-theoretic human-robot dialog,” Journal of HumanRobot Interaction, vol. 1, no. 1, pp. 78-95, 2012.
[169] P. Rybski, J. Stolarz, K. Yoon and M. Veloso, “Using dialog and human observations to dictate tasks to a learning robot assistant,” Intelligent Service Robotics,
vol. 1, no. 2, pp. 159-167, 2008.
[170] R. Bischoff and V. Graefe, “Dependable multimodal communication and interaction with robotic assistants,” in IEEE International Workshop on Robot and
Human Interactive Communication (RO-MAN), pp. 300-305, 2002.
[171] A. Clodic, R. Alami, V. Montreuil, and S. Li, “A study of interaction between dialog and decision for human-robot collaborative task achievement,” in
International Symposium on Robot and Human interactive Communication (RO-MAN), pp. 913-918, 2007.
[172] S. S. Ghidary, Y. Nakata, H. Saito, M. Hattori, and T. Takamori, “Multi-Modal human robot interaction for map generation,” in IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), vol. 4, pp. 2246-2251, 2001.
[173] T. Kollar, S. Tellex, D. Roy, and N. Roy, “Grounding verbs of motion in natural language commands to robots,” in 12th International Symposium on Experimental
Robotics, pp. 31-47, 2014.
[174] J. Bos, “Applying automated deduction to natural language understanding,” Journal of Applied Logic, vol. 7, no. 1, pp. 100-112, 2009.
[175] M. Mason and M. Lopes, “Robot self-initiative and personalization by learning through repeated interactions,” In Proceedings of the 6th international conference
on Human-robot interaction, pp. 433-440, 2011.
[176] M. Strait, P. Briggs, and M. Scheutz, “Gender, more so than Age, Modulates Positive Perceptions of Language-Based Human-Robot Interactions,” International
Symposium on New Frontiers in Human-Robot Interaction, 2015.
[177] J. F. Gorostiza and M. A. Salichs, “Natural programming of a social robot by dialogs,” in AAAI Fall Symposium, 2010.
[178] B. Mutlu, J. Forlizzi and J. Hodgins, “A Storytelling Robot: Modeling and Evaluation of Human-like Gaze Behavior,” IEEE-RAS International conference on
humanoid robotics, pp. 518-523, 2006.
[179] W. Wang, G. Athanasopoulos, S. Yilmazyildiz, G. Patsis, V. Enescu, H. Sahli, W. Verhelst, A. Hiolle, M. Lewis, and L. Canamero, “Natural emotion elicitation
for emotion modeling in child-robot interactions,” in 4th Workshop on Child Computer Interaction (WOCCI), pp. 51-56, 2014.
[180] C. Breazeal and L. Aryananda, “Recognition of affective communicative intent in robot-directed speech,” Autonomous Robots, vol. 12, no. 1, pp. 83-104, 2002.
[181] A. Lockerd and C. Breazeal, “Tutelage and socially guided robot learning,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS),
vol. 4, pp. 3475-3480, 2004.
[182] C. Breazeal, “Toward sociable robots,” Robotics and Autonomous Systems, vol. 42, no. 3, pp. 167-175, 2003.
[183] K. Severinson-Eklundh, A. Green, and H. Huttenrauch, “Social and collaborative aspects of interaction with a service robot,” Robotics and Autonomous Systems,
vol. 42, nos. 3-4, pp. 223-234, 2003.
[184] A. Austermann, N. Esau, L. Kleinjohann, and B. Kleinjohann, “Fuzzy emotion recognition in natural speech dialogue,” in International Workshop on Robot and
Human interactive Communication (RO-MAN), pp. 317-332, 2005.
[185] M. Coeckelbergh, “You, robot: on the linguistic construction of artificial others,” AI & Society, vol. 26, no. 1, pp. 61-69, 2011.
[186] R. Read and T. Belpaeme, “How to use non-linguistic utterances to convey emotion in child-robot interaction,” in 7th ACM/IEEE International Conference
on Human-Robot Interaction (HRI), pp. 219-220, 2012.
[187] I. Kruijff-Korbayova, I. Baroni, and M. Nalin, “Children’s Turn-Taking Behavior Adaptation in Multi-Session Interactions with a Humanoid Robot,” In IEEE
workshop on timing in human-robot interaction, 2014.
[188] S. Sabanovic, M. P. Michalowski, and R. Simmons, “Robots in the wild: Observing human-robot social interaction outside the lab,” in 9th IEEE International
Workshop on Advanced Motion Control, pp. 596-601, 2006.
[189] H. G. Okuno, K. Nakadai, and H. Kitano, “Social interaction of humanoid robot based on audio-visual tracking,” in International Conference on Industrial and
Engineering Applications of Artificial Intelligence and Expert Systems, pp. 725-735, 2002.
[190] A. Chella, R. E. Barone, G. Pilato, and R. Sorbello, “An emotional storyteller robot,” in AAAI Spring Symposium on Emotion, Personality, and Social Behavior,
pp. 17-22, 2008
[191] T. Kollar, S. Tellex, M. R. Walter, A. Huang, A. Bachrach, S. Hemachandra, E. Brunskill, A. Banerjee, D. Roy, S. Teller, and N. Roy, “Generalized grounding
graphs: A probabilistic framework for understanding grounded language,” Journal of Artificial Intelligence Research, pp. 1-35, 2013.
[192] M. Forbes, R. P. N. Rao, L. Zettlemoyer, and M. Cakmak, “Robot programming by demonstration with situated spatial language understanding,” in IEEE
International Conference on Robotics and Automation (ICRA), pp. 2014-2020, 2015.
[193] G. Salvi, L. Montesano, A. Bernardino, and J. Santos-Victor, “Language bootstrapping: Learning word meanings from perception–action association,” IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 42, no. 3, pp. 660-671, 2011.
[194] C. M. Huang and B. Mutlu, “Anticipatory robot control for efficient human-robot collaboration,” in 11th ACM/IEEE International Conference on Human-Robot
Interaction (HRI), pp. 83-90, 2016.
[195] M. Johnson-Roberson, J. Bohg, G. Skantze, J. Gustafson, R. Carlson, B. Rasolzadeh, and D. Kragic, “Enhanced visual scene understanding through human-robot
dialog,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3342-3348, 2011.
[196] F. Doshi and N. Roy, “Spoken language interaction with model uncertainty: an adaptive human–robot interaction system,” Connection Science, vol. 20, no. 4,
pp. 299-318, 2008.
[197] S. Lauria, G. Bugmann, T. Kyriacou, and E. Klein, “Mobile robot programming using natural language,” Robotics and Autonomous Systems, vol. 38, no. 3-4, pp.
171-181, 2002.
[198] S. Hemachandra, M. R. Walter, S. Tellex, and S. Teller, “Learning spatial-semantic representations from natural language descriptions and scene classifications,”
in IEEE International Conference on Robotics and Automation (ICRA), pp. 2623-2630, 2015.
[199] M. Hanheide, M. Gobelbecker, G. S. Horn, A. Pronobis, K. Sjoo, A. Aydemir, P. Jensfelt, C. Gretton, R. Dearden, M. Janicek, H. Zender, G. Kruijff, N. Hawes,
and J. L. Wyatt, “Robot task planning and explanation in open and uncertain worlds,” Artificial Intelligence, 2015.
[200] X. Chen, J. Xie, J. Ji, and Z. Sui, “Toward open knowledge enabling for Human-Robot interaction,” Journal of Human-Robot Interaction, vol. 1, no. 2, pp. 100117, 2012.
[201] J. Weston, F. Ratle, and R. Collobert, “Deep learning via semi-supervised embedding,” in Proceedings of the 25th international conference on Machine learning
(ICML), pp. 639-655, 2008.
[202] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence learning with neural networks,” Advances in neural information processing systems, vol. 4, pp.
3104-3112, 2014.
[203] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and translate,” in International Conference on Learning
Representation (ICLR), 2015.
[204] X. Glorot, A. Bordes, and Y. Bengio, “Domain adaptation for large-scale sentiment classification: A deep learning approach,” in International Conference on
Machine learning (ICML), pp. 513-520, 2011.
[205] M. Samadi, T. Kollar, and M. Veloso, “Using the web to interactively learn to find objects,” in Proceedings of the Twenty-Sixth AAAI Conference on Artificial
Intelligence, 2012.
[206] G. Gordon, S. Spaulding, J. K. Westlund, J. L. Jin, L. Plummer, M. Martinez, M. Das, and C. Breazeal, “Affective personalization of a social robot tutor for
children's second language skills,” in Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pp. 3951-3957, 2016.
[207] J. Saunders, D. S. Syrdal, K. L. Koay, N. Burke, and K. Dautenhahn, “Teach me -- show me” -- end-user personalization of a smart home and companion robot,”
IEEE Transactions on Human-Machine Systems, vol. 46, no. 1, pp. 27-40, 2016.
[208] M. Slamani, A. Nubiola, and I. Bonev, “Assessment of the positioning performance of an industrial robot,” Industrial Robot, vol. 39, no. 1, pp. 57-68, 2012.
[209] A. Antunes, L. Jamone, G. Saponaro, A. Bernardino, and R. Ventura, “From human instructions to robot actions: Formulation of goals, affordances and
probabilistic planning,” in IEEE International Conference on Robotics and Automation (ICRA), pp. 5449-5454, 2016.
[210] G. Gordon and C. Breazeal, “Bayesian active learning-based robot tutor for children’s word-reading skills,” in Proceedings of the Twenty-Ninth AAAI Conference
on Artificial Intelligence, pp. 1343-1349, 2015.
[211] Q. Zeng, S. Sun, H. Duan, C. Liu, and H. Wang, “Cross-organizational collaborative workflow mining from a multi-source log,” Decision Support Systems vol.
54, no. 3, pp. 1280-1301, 2013.
[212] R. Liu, X. Zhang, and H. Zhang, “Web-video-mining-supported workflow modeling for laparoscopic surgeries,” Artificial Intelligence in Medicine, vol. 74, pp.
9-20, 2016.
[213] R. Liu and X. Zhang, “Generating machine-executable plans from end-user’s natural-language-instructions,” arXiv preprint arXiv:1611.06468, 2016.
[214] N. Legany, G. Toldi, N. Megyes, C. Orban, L. Kovacs, and A. Balog, “ACT-R/E: An embodied cognitive architecture for human-robot interaction,” Journal of
Human-Robot Interaction, vol. 2, no. 1, pp. 30-55, 2013.
[215] H. Koppula and A. Saxena, “Anticipating human activities using object affordances for reactive robotic response,” in Robotics: Science and Systems, vol. 38, no.
1, pp. 14-29, 2013.
[216] M. L. Littman, “Reinforcement learning improves behaviour from evaluative feedback,” Nature, vol. 521, no. 7553, pp. 445-451, 2015.
[217] P. Raccuglia, K. C. Elbert, P. D. F. Adler, C. Falk, M. B. Wenny, A. Mollo, M. Zeller, S. A. Friedler, J. Schrier, and A. J. Norquist, “Machine-learning-assisted
materials discovery using failed experiments,” Nature, vol. 533, no. 7601, pp.73-76, 2016.
Rui Liu received the M.S. degree in active vibration control from Shanghai Jiao Tong University,
Shanghai, China in 2013. Since January, 2014, he has been a Ph.D. student in Mechanical Engineering at
Colorado School of Mines, Golden, Colorado, USA.
From June to December 2013, he was a Research Engineer at the Chinese Academy of Science,
Chengdu, China. His current research interests include robotics, Knowledge-based Systems, NLP and
machine learning.
All his papers could be found at:
Google Scholar: https://scholar.google.com/citations?user=X5X0IyQAAAAJ&hl=en&authuser=1
Research Gate: https://www.researchgate.net/profile/Rui_Liu99/contributions
Xiaoli Zhang received the B.S. degree in Mechanical Engineering from Xi’an Jiaotong University, Xi’an,
ShanXi, China in 2003, the M.S. degree in Mechatronics Engineering from Xi’an Jiaotong University in
2006, the Ph.D. degree in Biomedical Engineering from the University of Nebraska Lincoln, Lincoln,
USA, in 2009.
Since 2013, she has been an Assistant Professor in the Mechanical Engineering Department, Colorado
School of Mines, Golden, CO. She is the author of more than 50 articles, and 5 inventions. Her research
interests include intelligent human-robot interaction, human intention awareness, robotics system design
and control, haptics, and their applications in healthcare fields.
| 2cs.AI
|
Magnetic-Visual Sensor Fusion based Medical SLAM
for Endoscopic Capsule Robot
arXiv:1705.06196v2 [cs.CV] 6 Nov 2017
Mehmet Turana,b , Yasin Almaliogluc , Hunter Gilbertd , Helder Araujoe , Ender
Konukoglub , Metin Sittia
a Max
Planck Institute for Intelligent Systems, Stuttgart, Germany
Vision Laboratory, ETH Zurich, Switzerland
c Computer Engineering Department, Bogazici University, Turkey
d Louisiana State University, Baton Rouge, LA, USA
e Robotics Laboratory, University of Coimbra, Portugal
b Computer
Abstract
A reliable, real-time simultaneous localization and mapping (SLAM) method
is crucial for the navigation of actively controlled capsule endoscopy robots.
These robots are an emerging, minimally invasive diagnostic and therapeutic
technology for use in the gastrointestinal (GI) tract. In this study, we propose
a dense, non-rigidly deformable, and real-time map fusion approach for actively
controlled endoscopic capsule robot applications. The method combines magnetic and vision based localization, and makes use of frame-to-model fusion and
model-to-model loop closure. The performance of the method is demonstrated
using an ex-vivo porcine stomach model. Across four trajectories of varying
speed and complexity, and across three cameras, the root mean square localization errors range from 0.42 to 1.92 cm, and the root mean square surface
reconstruction errors range from 1.23 to 2.39 cm.
Keywords: Endoscopic Capsule Robot, Magnetic-Visual Sensor Fusion,
Simultaneous Localization and Mapping, Frame-to-Model Fusion, Non-rigid
Map Reconstruction
Email addresses: [email protected] (Mehmet Turan),
[email protected] (Yasin Almalioglu), [email protected] (Hunter Gilbert),
[email protected] (Helder Araujo), [email protected] (Ender
Konukoglu), [email protected] (Metin Sitti)
Preprint submitted to Elsevier
November 7, 2017
1. Introduction
Gastrointestinal diseases are the primary diagnosis for about 28 million patient visits per year in the United States[1]. In many cases, endoscopy is an
effective diagnostic and therapeutic tool, and as a result about 7 million upper and 11.5 million lower endoscopies are carried out each year in the U.S.
[2]. Wireless capsule endoscopy (WCE), introduced in 2000 by Given Imaging
Ltd., has revolutionized patient care by enabling inspection of regions of the GI
tract that are inaccessible with traditional endoscopes, and also by reducing the
pain associated with traditional endoscopy [3]. Going beyond passive inspection, researchers are striving to create capsules that perform active locomotion
and intervention [4]. With the integration of further functionalities, e.g. remote
control, biopsy, and embedded therapeutic modules, WCE can become a key
technology for GI diagnosis and treatment in near future.
Several research groups have recently proposed active, remotely controllable
robotic capsule endoscope prototypes equipped with additional operational functionalities, such as highly localized drug delivery, biopsy, and other medical
functions [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]. To facilitate effective navigation and intervention, the robot must be accurately localized and must also
accurately perceive the surrouding tissues. Three-dimensional intraoperative
SLAM algorithms will therefore be an indispensable component of future active
capsule systems.
Several localization methods have been proposed for robotic capsule endoscopes such as fluoroscopy [17], ultrasonic imaging [18], positron emission tomography (PET) [17, 18], magnetic resonance imaging (MRI) [17], radio transmitter based techniques, and magnetic field-based techniques [18, 19, 20, 21]. It
has been proposed that combinations of sensors, such as RF range estimation
and visual odometry, may improve the estimation accuracy [22, 23]. Morover,
solutions that incorporate vision are attractive because a camera is already
present on capsule endoscopes, and vision algorithms have been widely applied
for robotic localization and map reconstruction.
2
Feature-based SLAM methods have been applied on endoscopic type of image sequences in the past e.g [24, 25, 26, 27]. As improvements to accomodate
the flexibility of the GI tract, [28] suggested a motion compensation model to
deal with peristaltic motions, whereas [29] proposed a learning algorithm to
deal with them. [30] adapted parallel tracking and mapping techniques to a
stereo-endoscope to obtain reconstructed 3D maps that were denser when compared to monoscopic camera methods. [31] has applied ORB features to track
the camera and proposed a method to densify the reconstructed 3D map, but
pose estimation and map reconstruction are still not accurate enough. All of
these methods can fail to produce accurate results in cases of low-texture areas,
motion blur, specular highlights, and sensor noise – all of which are typically
present during endoscopy. In this paper, we propose that a non-rigid and dense
RGB Depth fusion method, which combines magnetic localization and visual
pose estimation using particle filtering and recurrent neural network-based motion model estimation, can provide real-time, accurate localization and mapping
for wireless capsule robots. We demonstrate the system in a pig stomach model
by measuring the system performance in terms of both capsule localization and
surface mapping accuracy.
2. System Overview and Analysis
The system architecture of the method is depicted in Figure 1. Alternating
between localization and mapping, our approach performs frame-to-model 3D
map reconstruction in real-time. Below we summarize key steps of the proposed
system:
• Perform offline inter-sensor calibration between magnetic hall sensor array
and capsule camera system;
• Estimate endoscopic capsule robot pose in 5-DoF using magnetic localization system;
3
Figure 1: System overview
• Estimate endoscopic capsule robot pose in 6-DoF using visual joint photometricgeometric frame-to-model pose optimization;
• Fuse magnetic and visual pose estimations using back propagation based
particle filtering and recurrent neural network based motion model estimation;
• Perform non-rigid frame-to-model registration making use of particle filter
based pose estimation and deformation constraints defined by the graph
equations;
• To perform loop closure, in case there exists an intersection of the active model with inactive model within the current frame, fuse intersecting
regions and deform the entire model non-rigidly.
3. Magnetically Actuated Soft Capsule Endoscopes (MASCE)
Our capsule prototype is a magnetically actuated soft capsule endoscope
(MASCE) designed for disease detection, drug delivery and biopsy operations in
4
the upper GI-tract. The prototype is composed of a RGB camera, a permanent
magnet, an empty space for drug chamber and a biopsy tool (see Fig 3 for
visual reference). The magnet exerts magnetic force and torque on the robot
in response to a controlled external magnetic field. The magnetic torque and
forces are used to actuate the capsule robot and to release drug. Magnetic fields
from the electromagnets generate the magnetic force and torque on the magnet
inside MASCE so that the robot moves inside the workspace. Sixty-four threeaxis magnetic sensors are placed on the top, and nine electromagnets are placed
in the bottom.
4. Methods
4.1. Magnetic Localization System
Our 5-DoF magnetic localization system is designed for the position and orientation estimation of untethered mesoscale magnetic robots [19]. The system
uses an external magnetic sensor system and electromagnets for the localization of the magnetic capsule robot. A 2D-Hall-effect sensor array measures the
component of the magnetic field from the permanent magnet inside the capsule robot at several locations outside of the robotic workspace. Additionally, a
computer-controlled magnetic coil array consisting of nine electromagnets generates the magnetic field for actuation. The core idea of our localization technique
is the separation of the capsule’s magnetic field component from the actuator’s
magnetic field component. For that purpose, the actuator’s magnetic field is
subtracted from the magnetic field data which is acquired by a Hall-effect sensor array. As a further step, second-order directional differentiation is applied
to reduce the localization error.
4.2. Visual Localization
We propose the use of a direct surfel map fusion method for actively controllable endoscopic capsule robots. The core algorithm is the ElasticFusion
method originally described by Whelan et al. [32], which uses a dense map and
5
Figure 2: Actuation system of the MASCE
non-rigid model deformation to account for changing environments. It performs
joint volumetric and photometric alignment, frame-to-model predictive tracking, and dense model-to-model loop closure with non-rigid space deformation.
Prior to using endoscopic video with such a method, the images must first be
prepared.
4.2.1. Multi-scale vessel enhancement and depth image creation
Endoscopic images have mostly homogeneous and poorly textured areas.
To prepare the camera frames for input into the ElasticFusion pipeline, our
framework starts with a vessel enhancement operation inspired from [33]. Our
approach enhances blood vessels by analyzing the multiscale second order local
structure of an image. First, we extract the Hessian matrix :
Ixx Ixy
H=
Iyx Iyy
6
(1)
Figure 3: Experimental setup
where I is the input image, and Ixx , Ixy , Iyx , Iyy the second order derivatives,
respectively. Secondly, eigenvalues |λ1 | ≤ |λ2 | and principal directions u1 , u2
of the Hessian matrix are extracted. The eigenvalues and principal directions
are then ordered and analyzed to decide whether the region belongs to a vessel.
To identify vessels in different scales and sizes, multiple scales are created by
convolving the input image and the final output is taken as the maximum of the
vessel filtered image across all scales. For further details of the mathematical
equations, the reader is referred to the original paper of [33]. Figure 4 shows
input RGB images, vessel detection and vessel enhancement results for four
different frames.
To create the depth channel from input RGB data, we implemented a realtime version of the perspective shape from shading under realistic conditions
[34] by reformulating the complex inverse problem for shading into a highly
7
Figure 4: Input RGB image, vessel detection, vessel enhancement and depth image creation
for sample frames.
parallelized non-linear optimization problem, which we solve efficiently using
GPU programming and a Gauss-Newton solver. Figure 4 shows samples of
input RGB images and depth images created from them.
4.2.2. Joint photometric-geometric pose estimation
The vision-based localization system operates on the principle of optimizing
both a photometric measure of error and a 3D geometric measure of error. The
camera pose of the endoscopic capsule robot is described by a transformation
matrix Pt :
Rt
tt
Pt =
01×3
∈ SE3 .
(2)
1
Geometric pose estimation is performed by minimizing the energy cost function Eicp between the current depth frame, Dtl , and the active depth model,
a
:
D̂t−1
Eicp =
X
t
k 2
ˆ
((vk − exp (ξ)Tv
k) · n )
k
8
(3)
where vkt is the back-projection of the k-th vertex in Dtl , vk and nk are the
corresponding vertex and normal from the previous frame. T is the estimated
ˆ is the
transformation from the previous to the current robot pose and exp (ξ)
exponential mapping function from Lie algebra se3 to Lie group SE3 , which represents small changes The photometric pose ξ between the current surfel-based
a
reconstructed RGB image Ctl and the active RGB model Cˆt−1
is determined by
minimizing the photometric energy cost function:
Ergb =
2
X
a
ˆ
I(u, Ctl ) − I(π(K exp(ξ)Tp(u,
Dtl )), Cˆt−1
)
(4)
u∈Ω
where as above T is the estimated transformation from previous to the current
camera pose. The joint photometric-geometric pose optimization is defined by
the cost function:
Etrack = Eicp + wrgb Ergb ,
(5)
with wrgb = 0.13, which was determined experimentally to yield good performance for our datasets. For the minimization of this cost function in real-time,
the Gauss-Newton method is employed. At each iteration of the method, the
ˆ
transformation T is updated as T → exp(ξ)T.
4.3. Particle Filtering based Magnetic-Visual Sensor Fusion
Particle filter is a statistical Bayesian filtering method that computes the posterior probability density function of sequentially obtained state vectors xt ∈ X
which are suggested by sensor measurements. We have implemented the method
of Caron et al., which provides robustness against sensor failure through the introduction of latent variables characterizing the sensor’s reliability as either
normal or failing, which are estimated along with the system state [35]. The
method is briefly described in what follows.
The state xt composes the 6-DoF pose for the capsule robot, which is assumed to propagate in time according to a transition model:
xt = f (xt−1 , vt )
9
(6)
Figure 5: The switching state-space model. The double rectangles denote observable variables
and the gray rectangles denote hyper-parameters.
where f is a non-linear state transition function and vt is a white noise. t is
the index of a time sequence, t ∈ {1, 2, 3, ...}.
6-DoF pose state estimation with a high precision often requires multi-sensor
input or sequential observations. The endoscopic capsule is equipped with two
sensor systems, one being a 5-DoF magnetic sensor array and the other one being
an endoscopic monocular RGB camera. Observations of the pose are produced
by n sensors zk,t (k = 1, ..., n) in general, where the probability distribution
p(zk,t |xt ) is known for each sensor.
We estimate the 6-DoF pose states which rely on latent (hidden) variables by
using the Bayesian filtering approach. The relations between all of the variables
are shown in the probabilistic graphical model, Figure 5. The hidden variables
of sensor states are denoted as sk,t , which we call switch variables, where sk,t ∈
{0, ..., dk } for k = 1, ..., n. dk is the number of possible observation models,
e.g., failure and nominal sensor states. The observation model for zk,t can be
10
described as:
zk,t = hk,sk,t ,t (xt ) + wk,sk,t ,t
(7)
where hk,sk,t ,t (xt ) is the non-linear observation function and wk,sk,t ,t is the
observation noise. The latent variable of the switch parameter sk,t is defined
to be 0 if the sensor is in a failure state, which means that observation zk,t is
statistically independent of xt , and 1 if the sensor k is in its nominal state of
work. The prior probability for the switch parameter sk,t being in a given state
j, is denoted as αk,j,t and it is the probability for each sensor to be in a given
state:
P r(sk,t = j) = αk,j,t ,
where αk,j,t ≥ 0 and
Pdk
j=0
0 ≤ j ≤ dk
(8)
αk,j,t = 1 with a Markov evolution property. The
objective posterior density function p(x0:t , s1:t , α0:t |z1:t ) and the marginal posterior probability p(xt |z1:t ) , in general, cannot be determined in a closed form
due to its complex shape. However, sequential Monte Carlo methods (particle
filters) provide a numerical approximation of the posterior density function with
a set of samples (particles) weighted by the kinematics and observation models.
4.3.1. Sensor Failure Detection and Handling
The proposed multi-sensor fusion approach is able to detect the sensor failure
periods and to handle the failures, accordingly. As seen in Figure 6), The
posterior probabilities of the switch parameters sk,t and the minimum mean
square error (MMSE) estimates of αk,t indicate an accurate detection of sensor
failure states. Visual localization failed between seconds 14-36 due to very fast
frame-to-frame motions and magnetic sensor failed between seconds 57-76 due
to increased distance of the ringmagnet to the sensor array. Thanks to the
switching option ability from one observation model to another in case of a
sensor failure so that RMSE is kept low during sensor failure as seen in Figure
7. In our sensor failure model, we do not make a Markovian assumption for the
11
Figure 6: Top figures: Posterior probability of sk,t parameter for endoscopic RGB camera
(left) and for magnetic localization system (right). Bottom figures: The minimum mean
square error (MMSE) of αk,t for endoscopic RGB camera (left) and for magnetic localization
system (right). The switch parameter, sk,t , and the confidence parameter αk,t reflect the
failure times accurately: Visual localization fails between 14-36 seconds and magnetic sensor
fails between 57-76 seconds. Both failures are detected confidentially.
switch variable sk,t but we do for its prior αk,t , resulting in a priori dependent on
the past trajectory sections, which is more likely for the incremental endoscopic
capsule robot motions. The model thus introduces a memory over the past
sensor states rather than simply considering the last state. The length of the
α
memory is tuned by the hyper-parameters σk,t
, leading to a long memory for
large values and vice-versa. This is of particular interest when considering
sensor failures. Our system detects automatically failure states. Thus, the
confidence in the RGB sensor decrease when visual localization fails recently
due to occlusions, fast-frame-to frame changes etc. On the other hand, the
confidence in magnetic sensor decreases if the magnetic localization fails due to
noise interferences from environment or if the ringmagnet has a big distance to
the magnetic sensor array.
12
Figure 7: Translational (top) and rotational (bottom) RMSEs for multi-sensor fusion, visual
localization and magnetic localization.
4.3.2. LSTM-based Kinematic Model
Existing sensor fusion methods based on traditional particle filter and Kalman
filter approaches have limited functionality in non-linear dynamic systems. It is
assumed that, in the Kalman filter and extended Kalman filter, the underlying
dynamic process is well-modelled by linear equations or that these equations
can be linearised without a major loss of fidelity. On the other hand, particle
filters accommodate a wide variety of dynamic models, allowing for highly complex and non-linear dynamics in the state variables with a realistic non-linear
kinematics. Unlike traditional feed-forward and just-in-moment artificial neural
networks, LSTMs are very suitable for modelling the dependencies across time
sequences and for creating a temporal motion model since it has a memory of
hidden states over time and has directed cycles among hidden units, enabling
the current hidden state to be a function of arbitrary sequences of inputs. Thus,
using an LSTM, the pose estimation of the current time step benefits from information encapsulated in previous time sequences and is suitable to formulate
the state transition function f in Equation 6. In particular, information about
the most recent velocities and accelerations become available to the model.
The training data is divided into input sequences of length 50 and the slices
are passed into the LSTM modules with the expectation that it predicts the next
6-DoF pose value, i.e. the 51st pose measurement, which was used to compute
the cost function for training. The LSTM module was trained using Keras
13
library with GPU programming and Theano back-end. Using back-propagationthrough-time method, the weights of hidden units were trained for up to 200
epochs with an initial learning rate of 0.001. Overfitting was prevented using
dropout and early stopping techniques.
4.4. Coupling of Magnetic and Vision Systems
To couple the magnetic actuation and localization system which is seen in
Fig. 2 with the proposed vision based SLAM system, an eye-in-hand calibration
problem must be solved. The vision system measures the pose of the camera,
and the magnetic localization system measures the 5D pose of the magnet on
the MASCE. The transformation between the coordinate frames attached at the
ring magnet and at the camera origin must be known, because the particle filter
assumes that the two systems make measurements on the same system state,
which in this case is a single rigid body pose associated with the capsule.
There does not exist a standard method for determining the transformation
matrix when one of the measurement systems is 5 DoF. We applied the standard
theory for two 6 DoF sensors by assuming a value for the remaining rotational
DoF in the magnetic sensor data. We then used the dual quaternion-based
algorithm to determine the unknown transformation from the data [36].
When the calibration is known, then the actuation and vision systems are
automatically registered. This will facilitate the closing of the perception-action
loop in future work by a planner and/or human teleoperator that uses the map
to generate desired capsule trajectories in the magnetic coordinate system.
4.5. Scene representation, Deformation graph and Loop Closure
Due to strict real-time concerns of the approach, instead of using every pixel
for reconstruction, the method uses surfels, where each surfel has a position,
normal, color, weight, radius indicating the local surface area around a given
point, initialization timestamp and last updated timestamp. The ElasticFusion method uses a deformation graph to non-rigidly deform the model during
mapping. Non-rigid deformation information is embedded in the surfels that
14
comprise the map, and loop closures are applied frequently. The importance of
accounting for non-rigid deformation is critical in the soft environment of the GI
tract, where the map may need to be updated not only due to previous error in
measurement but also due to real changes in the geometry of the environment.
To ensure globally consistent surface reconstruction, our framework closes
loops with the existing map as those areas are revisited by fusing reactivated
parts into the active model area and simultaneously deactivating surfels which
have not been seen in a period of time. Inactive surfels are not used for either
frame tracking or map fusion until a loop closure happens which activates them
again. This has the advantage of continuous frame-to-model and model-tomodel tracking which provides viewpoint-invariance. Moreover, frame-to-model
tracking has benefits over frame-to-frame tracking once the capsule robot encounters drifts and needs to relocalize. For a detailed description of the method,
the reader is referred to the original source [32].
5. EXPERIMENTS AND RESULTS
We evaluate the performance of our system both quantitatively and qualitatively in terms of trajectory estimation, surface reconstruction accuracy and
computational performance.
5.1. Dataset
Three different endoscopic cameras were used to capture the endoscopic
videos; i.e AWAIBA Naneye, Misumi-V3506-2ES and POTENSIC camera. We
mounted endoscopic cameras on our magnetically activated soft capsule endoscope (MASCE) systems. The dataset was recorded on an open real pig stomach.
We scanned the stomach using the 3D Artec Space Spider image scanner. This
3D scan served as the ground truth for the evaluation of the 3D map reconstruction. To obtain the ground truth for 6-DoF camera pose, an OptiTrack motion
tracking system consisting of eight infrared cameras and tracking software was
utilized. Figure 3 illustrates our experimental setup. A total of 15 minutes of
15
(a) Trajectory 1
(b) Trajectory 2
(c) Trajectory 3
(d) Trajectory 4
Figure 8: Sample ground truth trajectories and estimated trajectories.
stomach video was recorded containing over 10000 frames. Some sample frames
of the dataset are shown in Figure 4.
5.2. Trajectory Estimation
Table 1 demonstrates the results of the trajectory estimation for four different trajectories that present varying levels of difficulty for the system. Trajectory 1 is an uncomplicated path with slow translations and rotations. Trajectory
2 follows a comprehensive scan of the stomach with many local loop closures.
Trajectory 3 contains an extensive scan of the stomach with more complicated
local loop closures. Trajectory 4 consists of more challenging motions including faster rotational and translational movements. Some qualitative tracking
16
Table 1: Trajectory lengths and RMSE results in centimeters
Trajectory ID
POTENSIC
AWAIBA
MISUMI
LENGTH
1
0.51
0.73
0.42
41.4
2
0.64
0.81
0.48
51.3
3
1.77
1.86
1.62
43.2
4
1.92
1.95
1.73
47.8
Table 2: Absolute depth and 3D surface reconstruction RMSE results in centimeters
Trajectory ID
POTENSIC
MISUMI
AWAIBA
Length
1
1.23
1.25
1.28
31.4
2
1.25
1.27
1.35
41.3
3
2.22
2.26
2.37
53.2
4
2.29
2.31
2.39
47.8
results of our proposed system and corresponding ground truth trajectories are
demonstrated in Figure 8. For the quantitative analysis, we used the absolute
trajectory (ATE) root-mean-square error between estimated pose values and
the ground truth acquired by Optitrack system. As seen in Table 1, the system
provides robust and accurate tracking in all of the tested conditions, and is not
adversely affected by sudden movements, blur, noise or spectral reflections. For
the duration of the dataset, the Bayesian filtering algorithm estimated that the
camera remained in the normally functioning sensor state at all times.
5.3. Surface Reconstruction
We scanned the non-rigid five different real porcine stomachs to obtain the
ground truth for 3D map reconstruction. Reconstructed 3D surface and ground
truth scan data were aligned using iterative closest point algorithm (ICP) and
map reconstruction error is calculated using the absolute trajectory (ATE)
17
Figure 9: 3D reconstructed soft stomach simulator surface
RMSE between estimated depth values and the depth values of the scan data.
We evaluated the surface reconstruction on the same four trajectories. The detailed RMSE results in Table 2 prove that our system reconstructs 3D organ
surface precisely. A sample 3D reconstruction process is shown in Figure 9 for
visual reference.
5.4. Computational Performance
To analyze the computational performance of the system, we observed the
average frame processing time across the sequences of Trajectory 1 to 4. The
test platform was a desktop PC with an Intel Xeon E5-1660v3-CPU at 3.00
GHz, 8 cores, 32GB of RAM and an NVIDIA Quadro K1200 GPU with 4GB
of memory. The execution time of the system is depended on the number of
surfels in the map, with an overall average of 58 ms per frame scaling to a peak
average of 67 ms implying a worst case processing frequency of 15 Hz.
18
6. CONCLUSION
In this paper, we have presented a magnetic-RGB Depth fusion based SLAM
method for endoscopic capsule robots. Our system makes use of surfel-based
dense reconstruction in combination with particle filter based fusion of magnetic
and visual localization information and sensor failure detection. The proposed
system is able to produce a highly accurate 3D map of the explored inner organ
tissue and is able to stay close to the ground truth endoscopic capsule robot
trajectory even for challenging robot trajectories. In the future, in vivo testing
is required to validate the accuracy and robustness of the SLAM system in the
challenging conditions of the GI tract. We also intend to extend our work into
stereo capsule endoscopy applications to achieve even more accurate localization
and mapping. In addition, the inclusion of more sophisticated methods for the
hand-eye calibration may result in improved accuracy.
References
References
[1] National Center for Health Statistics, National ambulatory medical care
survey: 2014 state and national summary tables, U.S. Centers for Disease
Control and Prevension.
[2] A. F. Peery, E. S. Dellon, J. Lund, S. D. Crockett, C. E. McGowan, W. J.
Bulsiewicz, L. M. Gangarosa, M. T. Thiny, K. Stizenberg, D. R. Morgan,
et al., Burden of gastrointestinal disease in the united states: 2012 update,
Gastroenterology 143 (5) (2012) 1179–1187.
[3] G. Iddan, G. Meron, A. Glukhovsky, P. Swain, Wireless capsule endoscopy,
Nature 405 (6785) (2000) 417–418.
[4] A. Moglia, A. Menciassi, M. O. Schurr, P. Dario, Wireless capsule endoscopy: from diagnostic devices to multipurpose robotic systems, Biomedical microdevices 9 (2) (2007) 235–243.
19
[5] M. Sitti, H. Ceylan, W. Hu, J. Giltinan, M. Turan, S. Yim, E. Diller,
Biomedical applications of untethered mobile milli/microrobots, Proceedings of the IEEE 103 (2) (2015) 205–224.
[6] S. Yim, K. Goyal, M. Sitti, Magnetically actuated soft capsule with the
multimodal drug release function, IEEE/ASME Transactions on Mechatronics 18 (4) (2013) 1413–1418.
[7] S. Yim, M. Sitti, Shape-programmable soft capsule robots for semiimplantable drug delivery, IEEE Transactions on Robotics 28 (5) (2012)
1198–1202.
[8] S. Yim, M. Sitti, Design and rolling locomotion of a magnetically actuated
soft capsule endoscope, IEEE Transactions on Robotics 28 (1) (2012) 183–
194.
[9] M. K. Goenka, S. Majumder, U. Goenka, Capsule endoscopy: Present status and future expectation, World J Gastroenterol 20 (29) (2014) 10024–
10037.
[10] T. Nakamura, A. Terano, Capsule endoscopy: past, present, and future,
Journal of gastroenterology 43 (2) (2008) 93–99.
[11] F. Munoz, G. Alici, W. Li, A review of drug delivery systems for capsule
endoscopy, Advanced drug delivery reviews 71 (2014) 77–85.
[12] F. Carpi, N. Kastelein, M. Talcott, C. Pappone, Magnetically controllable
gastrointestinal steering of video capsules, IEEE Transactions on Biomedical Engineering 58 (2) (2011) 231–234.
[13] H. Keller, A. Juloski, H. Kawano, M. Bechtold, A. Kimura, H. Takizawa,
R. Kuth, Method for navigation and control of a magnetically guided
capsule endoscope in the human stomach, in: 2012 4th IEEE RAS &
EMBS International Conference on Biomedical Robotics and Biomechatronics (BioRob), IEEE, 2012, pp. 859–865.
20
[14] A. W. Mahoney, S. E. Wright, J. J. Abbott, Managing the attractive magnetic force between an untethered magnetically actuated tool and a rotating permanent magnet, in: Robotics and Automation (ICRA), 2013 IEEE
International Conference on, IEEE, 2013, pp. 5366–5371.
[15] S. Yim, E. Gultepe, D. H. Gracias, M. Sitti, Biopsy using a magnetic capsule endoscope carrying, releasing, and retrieving untethered microgrippers,
IEEE Transactions on Biomedical Engineering 61 (2) (2014) 513–521.
[16] A. J. Petruska, J. J. Abbott, An omnidirectional electromagnet for remote
manipulation, in: Robotics and Automation (ICRA), 2013 IEEE International Conference on, IEEE, 2013, pp. 822–827.
[17] T. D. Than, G. Alici, H. Zhou, W. Li, A review of localization systems for
robotic endoscopic capsules, IEEE Transactions on Biomedical Engineering
59 (9) (2012) 2387–2399.
[18] S. Yim, M. Sitti, 3-d localization method for a magnetically actuated soft
capsule endoscope and its applications, IEEE Transactions on Robotics
29 (5) (2013) 1139–1151.
[19] D. Son, S. Yim, M. Sitti, A 5-d localization method for a magnetically
manipulated untethered robot using a 2-d array of hall-effect sensors,
IEEE/ASME Transactions on Mechatronics 21 (2) (2016) 708–716.
[20] K. M. Popek, A. W. Mahoney, J. J. Abbott, Localization method for a
magnetic capsule endoscope propelled by a rotating magnetic dipole field,
in: Robotics and Automation (ICRA), 2013 IEEE International Conference
on, IEEE, 2013, pp. 5348–5353.
[21] C. Di Natali, M. Beccani, P. Valdastri, Real-time pose detection for magnetic medical devices, IEEE Transactions on Magnetics 49 (7) (2013) 3524–
3527.
21
[22] Y. Geng, K. Pahlavan, On the accuracy of rf and image processing based
hybrid localization for wireless capsule endoscopy, in: Wireless Communications and Networking Conference (WCNC), 2015 IEEE, 2015, pp. 452–457.
[23] G. Bao, K. Pahlavan, L. Mi, Hybrid localization of microrobotic endoscopic
capsule inside small intestine by data fusion of vision and rf sensors, IEEE
Sensors Journal 15 (5) (2015) 2669–2678.
[24] P. Mountney, G.-Z. Yang, Dynamic view expansion for minimally invasive
surgery using simultaneous localization and mapping, in: Engineering in
Medicine and Biology Society, 2009. EMBC 2009. Annual International
Conference of the IEEE, IEEE, 2009, pp. 1184–1187.
[25] O. G. Grasa, E. Bernal, S. Casado, I. Gil, J. Montiel, Visual slam for handheld monocular endoscope, IEEE transactions on medical imaging 33 (1)
(2014) 135–146.
[26] D. Stoyanov, M. V. Scarzanella, P. Pratt, G.-Z. Yang, Real-time stereo
reconstruction in robotically assisted minimally invasive surgery, in: International Conference on Medical Image Computing and Computer-Assisted
Intervention, Springer, 2010, pp. 275–282.
[27] L. Liu, C. Hu, W. Cai, M. Q.-H. Meng, Capsule endoscope localization
based on computer vision technique, in: Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the
IEEE, IEEE, 2009, pp. 3711–3714.
[28] P. Mountney, G.-Z. Yang, Motion compensated slam for image guided
surgery, Medical Image Computing and Computer-Assisted Intervention–
MICCAI 2010 (2010) 496–504.
[29] P. Mountney, D. Stoyanov, A. Davison, G.-Z. Yang, Simultaneous stereoscope localization and soft-tissue mapping for minimal invasive surgery,
in: International Conference on Medical Image Computing and ComputerAssisted Intervention, Springer, 2006, pp. 347–354.
22
[30] B. Lin, A. Johnson, X. Qian, J. Sanchez, Y. Sun, Simultaneous tracking,
3d reconstruction and deforming point detection for stereoscope guided
surgery, in: Augmented Reality Environments for Medical Imaging and
Computer-Assisted Interventions, Springer, 2013, pp. 35–44.
[31] N. Mahmoud, I. Cirauqui, A. Hostettler, C. Doignon, L. Soler,
J. Marescaux, J. Montiel, Orbslam-based endoscope tracking and 3d reconstruction, arXiv preprint arXiv:1608.08149.
[32] T. Whelan, R. F. Salas-Moreno, B. Glocker, A. J. Davison, S. Leutenegger, Elasticfusion: Real-time dense slam and light source estimation, The
International Journal of Robotics Research (2016) 1697–1716.
[33] A. F. Frangi, W. J. Niessen, K. L. Vincken, M. A. Viergever, Multiscale
vessel enhancement filtering, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, 1998, pp.
130–137.
[34] M. Visentini-Scarzanella, D. Stoyanov, G.-Z. Yang, Metric depth recovery
from monocular images using shape-from-shading and specularities, IEEE
International Conference on Image Processing (ICIP).
[35] F. Caron, M. Davy, E. Duflos, P. Vanheeghe, Particle filtering for multisensor data fusion with switching observation models: Application to land
vehicle positioning, IEEE Transactions on Signal Processing 55 (6) (2007)
2703–2719.
[36] K. Daniilidis, Dual quaternions for absolute orientation and hand-eye calibration, Advances in Computer Vision. Advances in Computing Science.
23
| 1cs.CV
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.